# 110

Question 110

pts

Describe the various problems afflicting the Uniform Crime Reports

Flag question: Question 2

**Question 210 pts**

Explain why it is possible to generate categorical variables from continuous data but not possible to obtain continuous data from categorical variables.

Flag question: Question 3

**Question 310 pts**

Briefly explain the essential differences between bar charts and histograms.

Flag question: Question 4

**Question 410 pts**

A professor has recently completed her grading for the final exam. The scores can be seen in the data set below. Unfortunately, the professor has noticed the mean is extremely low. She is perplexed because she was certain the class had performed extraordinarily well as there were several scores in the 90s and two perfect exams. Take a look at the grade distribution below, calculate the mean, examine the scores, and figure out why the mean was so low.

Grade

99

97

100

100

64

55

40

52

63

96

50

65

60

52

Flag question: Question 5

**Question 510 pts**

What is the purpose of inferential statistics?

Flag question: Question 6

**Question 610 pts**

Explain what is meant by ‘sampling error’.

Flag question: Question 7

**Question 710 pts**

A researcher has a data set of homicides occurring in large Southern metropolitan areas consisting of 304 cases with a mean of 25.68 and a standard deviation of 11.26. The Professor has established α = .05. Calculate the resulting confidence interval for this set of data.

Flag question: Question 8

**Question 810 pts**

Briefly explain the difference between a Type I and Type II error.

Flag question: Question 9

**Question 910 pts**

Very briefly explain what is meant by the term ‘non-directional test.’

Flag question: Question 10

**Question 1010 pts**

Explain why a researcher would opt for an ANOVA instead of a series of t-tests.

Statistics for Criminology and Criminal Justice

Third Edition

2

3

Statistics for Criminology and Criminal Justice

Third Edition

Jacinta M. Gau

University of Central Florida

4

FOR INFORMATION:

SAGE Publications, Inc.

2455 Teller Road

Thousand Oaks, California 91320

E-mail: order@sagepub.com

SAGE Publications Ltd.

1 Oliver’s Yard

55 City Road

London EC1Y 1SP

United Kingdom

SAGE Publications India Pvt. Ltd.

B 1/I 1 Mohan Cooperative Industrial Area

Mathura Road, New Delhi 110 044

India

SAGE Publications Asia-Pacific Pte. Ltd.

3 Church Street

#10–04 Samsung Hub

Singapore 049483

Copyright © 2019 by SAGE Publications, Inc.

All rights reserved. No part of this book may be reproduced or utilized in any form or by any means,

electronic or mechanical, including photocopying, recording, or by any information storage and retrieval

system, without permission in writing from the publisher.

Printed in the United States of America

Library of Congress Cataloging-in-Publication Data

Names: Gau, Jacinta M., author.

Title: Statistics for criminology and criminal justice / Jacinta M. Gau, University of Central Florida.

Description: Third edition. | Los Angeles : SAGE, [2019] | Includes bibliographical references and index.

Identifiers: LCCN 2017045048 | ISBN 9781506391786 (pbk. : alk. paper)

Subjects: LCSH: Criminal statistics. | Statistical methods.

Classification: LCC HV7415 .G38 2019 | DDC 519.5—dc23 LC record available at https://lccn.loc.gov/2017045048

All trademarks depicted within this book, including trademarks appearing as part of a screenshot, figure, or other image are included solely for

the purpose of illustration and are the property of their respective holders. The use of the trademarks in no way indicates any relationship with,

or endorsement by, the holders of said trademarks. SPSS is a registered trademark of International Business Machines Corporation.

This book is printed on acid-free paper.

Acquisitions Editor: Jessica Miller

Editorial Assistant: Rebecca Lee

e-Learning Editor: Laura Kirkhuff

Production Editor: Karen Wiley

Copy Editor: Alison Hope

5

https://lccn.loc.gov/2017045048

Typesetter: C&M Digitals (P) Ltd.

Proofreader: Wendy Jo Dymond

Indexer: Beth Nauman-Montana

Cover Designer: Janet Kiesel

Marketing Manager: Jillian Oelsen

6

Brief Contents

Preface to the Third Edition

Acknowledgments

About the Author

Part I Descriptive Statistics

Chapter 1 Introduction to the Use of Statistics in Criminal Justice and Criminology

Chapter 2 Types of Variables and Levels of Measurement

Chapter 3 Organizing, Displaying, and Presenting Data

Chapter 4 Measures of Central Tendency

Chapter 5 Measures of Dispersion

Part II Probability and Distributions

Chapter 6 Probability

Chapter 7 Population, Sample, and Sampling Distributions

Chapter 8 Point Estimates and Confidence Intervals

Part III Hypothesis Testing

Chapter 9 Hypothesis Testing: A Conceptual Introduction

Chapter 10 Hypothesis Testing With Two Categorical Variables: Chi-Square

Chapter 11 Hypothesis Testing With Two Population Means or Proportions

Chapter 12 Hypothesis Testing With Three or More Population Means: Analysis of Variance

Chapter 13 Hypothesis Testing With Two Continuous Variables: Correlation

Chapter 14 Introduction to Regression Analysis

Appendix A Review of Basic Mathematical Techniques

Appendix B Standard Normal (z) Distribution

Appendix C t Distribution

Appendix D Chi-Square (χ²) Distribution

Appendix E F Distribution

Glossary

Answers to Learning Checks

Answers to Review Problems

References

Index

7

Detailed Contents

Preface to the Third Edition

Acknowledgments

About the Author

Part I Descriptive Statistics

Chapter 1 Introduction to the Use of Statistics in Criminal Justice and Criminology

▶ Research Example 1.1: What Do Criminal Justice and Criminology Researchers Study?

▶ Data Sources 1.1: The Uniform Crime Reports

▶ Data Sources 1.2: The National Crime Victimization Survey

Science: Basic Terms and Concepts

Types of Scientific Research in Criminal Justice and Criminology

Software Packages for Statistical Analysis

Organization of the Book

Review Problems

Chapter 2 Types of Variables and Levels of Measurement

Units of Analysis

Independent Variables and Dependent Variables

▶ Research Example 2.1: Choosing Variables for a Study on Police Use of Conductive

Energy Devices

▶ Research Example 2.2: Units of Analysis

Relationships Between Variables: A Cautionary Note

▶ Research Example 2.3: The Problem of Omitted Variables

Levels of Measurement

The Categorical Level of Measurement: Nominal and Ordinal Variables

▶ Data Sources 2.1: The Police–Public Contact Survey

▶ Data Sources 2.2: The General Social Survey

The Continuous Level of Measurement: Interval and Ratio Variables

▶ Data Sources 2.3: The Bureau of Justice Statistics

Chapter Summary

Review Problems

Chapter 3 Organizing, Displaying, and Presenting Data

Data Distributions

Univariate Displays: Frequencies, Proportions, and Percentages

Univariate Displays: Rates

Bivariate Displays: Contingency Tables

▶ Data Sources 3.1: The Census of Jails

▶ Research Example 3.1: Does Sexual-Assault Victimization Differ Between Female and

Male Jail Inmates? Do Victim Impact Statements Influence Jurors’ Likelihood of

8

Sentencing Murder Defendants to Death?

Graphs and Charts

Categorical Variables: Pie Charts

▶ Data Sources 3.2: The Law Enforcement Management and Administrative Statistics

Survey

Categorical Variables: Bar Graphs

Continuous Variables: Histograms

▶ Research Example 3.2: Are Women’s Violent-Crime Commission Rates Rising?

Continuous Variables: Frequency Polygons

Longitudinal Variables: Line Charts

Grouped Data

▶ Data Sources 3.3: CQ Press’s State Factfinder Series

SPSS

Chapter Summary

Review Problems

Chapter 4 Measures of Central Tendency

The Mode

▶ Research Example 4.1: Are People Convicted of Homicide More Violent in Prison Than

People Convicted of Other Types of Offenses? Do Latino Drug Traffickers’ National

Origin and Immigration Status Affect the Sentences They Receive?

The Median

The Mean

▶ Research Example 4.2: How Do Offenders’ Criminal Trajectories Impact the

Effectiveness or Incarceration? Can Good Parenting Practices Reduce the Criminogenic

Impact of Youths’ Time Spent in Unstructured Activities?

Using the Mean and Median to Determine Distribution Shape

Deviation Scores and the Mean as the Midpoint of the Magnitudes

SPSS

Chapter Summary

Review Problems

Chapter 5 Measures of Dispersion

The Variation Ratio

The Range

The Variance

The Standard Deviation

The Standard Deviation and the Normal Curve

▶ Research Example 5.1: Does the South Have a Culture of Honor That Increases Gun

Violence? Do Neighborhoods With Higher Immigrant Concentrations Experience More

Crime?

▶ Research Example 5.2: Why Does Punishment Often Increase—Rather Than Reduce—

9

Criminal Offending?

SPSS

Chapter Summary

Review Problems

Part II Probability and Distributions

Chapter 6 Probability

Discrete Probability: The Binomial Probability Distribution

▶ Research Example 6.1: Are Police Officers Less Likely to Arrest an Assault Suspect When

the Suspect and the Alleged Victim Are Intimate Partners?

Successes and Sample Size: N and r

The Number of Ways r Can Occur, Given N: The Combination

The Probability of Success and the Probability of Failure: p and q

Putting It All Together: Using the Binomial Coefficient to Construct the Binomial

Probability Distribution

Continuous Probability: The Standard Normal Curve

▶ Research Example 6.2: What Predicts Correctional Officers’ Job Stress and Job

Satisfaction?

The z Table and Area Under the Standard Normal Curve

Chapter Summary

Review Problems

Chapter 7 Population, Sample, and Sampling Distributions

Empirical Distributions: Population and Sample Distributions

Theoretical Distributions: Sampling Distributions

Sample Size and the Sampling Distribution: The z and t Distributions

Chapter Summary

Review Problems

Chapter 8 Point Estimates and Confidence Intervals

The Level of Confidence: The Probability of Being Correct

Confidence Intervals for Means With Large Samples

Confidence Intervals for Means With Small Samples

▶ Research Example 8.1: Do Criminal Trials Retraumatize Victims of Violent Crimes?

▶ Data Sources 8.1: The Firearm Injury Surveillance Study, 1993–2013

Confidence Intervals With Proportions and Percentages

▶ Research Example 8.2: What Factors Influence Repeat Offenders’ Completion of a

“Driving Under the Influence” Court Program? How Extensively Do News Media Stories

Distort Public Perceptions About Racial Minorities’ Criminal Involvement?

▶ Research Example 8.3: Is There a Relationship Between Unintended Pregnancy and

Intimate Partner Violence?

Why Do Suspects Confess to Police?

Chapter Summary

10

Review Problems

Part III Hypothesis Testing

Chapter 9 Hypothesis Testing: A Conceptual Introduction

Sample Statistics and Population Parameters: Sampling Error or True Difference?

Null and Alternative Hypotheses

Chapter Summary

Review Problems

Chapter 10 Hypothesis Testing With Two Categorical Variables: Chi-Square

▶ Research Example 10.1: How Do Criminologists’ and Criminal Justice Researchers’

Attitudes About the Criminal Justice System Compare to the Public’s Attitudes?

Conceptual Basis of the Chi-Square Test: Statistical Dependence and Independence

The Chi-Square Test of Independence

▶ Research Example 10.2: Do Victim or Offender Race Influence the Probability That a

Homicide Will Be Cleared and That a Case Will Be Tried as Death-Eligible?

Measures of Association

SPSS

Chapter Summary

Review Problems

Chapter 11 Hypothesis Testing With Two Population Means or Proportions

▶ Research Example 11.1: Do Multiple Homicide Offenders Specialize in Killing?

Two-Population Tests for Differences Between Means: t Tests

Independent-Samples t Tests

▶ Data Sources 11.1: Juvenile Defendants in Criminal Courts

Dependent-Samples t Tests

▶ Research Example 11.2: Do Mentally Ill Offenders’ Crimes Cost More?

▶ Research Example 11.3: Do Targeted Interventions Reduce Crime?

Two-Population Tests for Differences Between Proportions

▶ Research Example 11.4: Does the Gender Gap in Offending Rates Differ Between Male

and Female Drug Abusers?

SPSS

Chapter Summary

Review Problems

Chapter 12 Hypothesis Testing With Three or More Population Means: Analysis of Variance

ANOVA: Different Types of Variances

▶ Research Example 12.1: Do Asian Defendants Benefit From a “Model Minority”

Stereotype?

▶ Research Example 12.2: Are Juveniles Who Are Transferred to Adult Courts Seen as

More Threatening?

When the Null Is Rejected: A Measure of Association and Post Hoc Tests

▶ Research Example 12.3: Does Crime Vary Spatially and Temporally in Accordance With

11

Routine Activities Theory?

SPSS

Chapter Summary

Review Problems

Chapter 13 Hypothesis Testing With Two Continuous Variables: Correlation

▶ Research Example 13.1: Part 1: Is Perceived Risk of Internet Fraud Victimization Related

to Online Purchases?

▶ Research Example 13.2: Do Prisoners’ Criminal Thinking Patterns Predict Misconduct?

Do Good Recruits Make Good Cops?

Beyond Statistical Significance: Sign, Magnitude, and Coefficient of Determination

SPSS

▶ Research Example 13.1, Continued: Part 2: Is Perceived Risk of Internet Fraud

Victimization Related to Online Purchases?

Chapter Summary

Review Problems

Chapter 14 Introduction to Regression Analysis

One Independent Variable and One Dependent Variable: Bivariate Regression

Inferential Regression Analysis: Testing for the Significance of b

Beyond Statistical Significance: How Well Does the Independent Variable Perform as

a Predictor of the Dependent Variable?

Standardized Slope Coefficients: Beta Weights

The Quality of Prediction: The Coefficient of Determination

Adding More Independent Variables: Multiple Regression

▶ Research Example 14.1: Does Childhood Intelligence Predict the Emergence of Self-

Control?

Ordinary Least Squares Regression in SPSS

▶ Research Example 14.2: Does Having a Close Black Friend Reduce Whites’ Concerns

About Crime?

▶ Research Example 14.3: Do Multiple Homicide Offenders Specialize in Killing?

Alternatives to Ordinary Least Squares Regression

▶ Research Example 14.4: Is Police Academy Performance a Predictor of Effectiveness on

the Job?

Chapter Summary

Review Problems

Appendix A Review of Basic Mathematical Techniques

Appendix B Standard Normal (z) Distribution

Appendix C t Distribution

Appendix D Chi-Square (c²) Distribution

Appendix E F Distribution

Glossary

12

Answers to Learning Checks

Answers to Review Problems

References

Index

13

Preface to the Third Edition

In 2002, James Comey, the newly appointed U.S. attorney for the Southern District of New York who would

later become the director of the Federal Bureau of Investigation, entered a room filled with high-powered

criminal prosecutors. He asked the members of the group to raise their hands if they had never lost a case.

Proud, eager prosecutors across the room threw their hands into the air, expecting a pat on the back. Comey’s

response befuddled them. Instead of praising them, he called them chickens (that is not quite the term he

used, but close enough) and told them the only reason they had never lost is that the cases they selected to

prosecute were too easy.1 The group was startled at the rebuke, but they really should not have been. Numbers

can take on various meanings and interpretations and are sometimes used in ways that conceal useful

information rather than revealing it.

1. Eisinger, J. (2017). The chickens**t club: Why the Justice Department fails to prosecute executives. New York:

Simon & Schuster.

This book enters its third edition at a time when the demand for an educated, knowledgeable workforce has

never been greater. This is as true in criminal justice and criminology as in any other university major and

occupational field. Education is the hallmark of a professional. Education is not just about knowing facts,

though—it is about thinking critically and treating incoming information with a healthy dose of skepticism.

All information must pass certain tests before being treated as true. Even if it passes those tests, the possibility

remains that additional information exists that, if discovered, would alter our understanding of the world.

People who critically examine the trustworthiness of information and are open to new knowledge that

challenges their preexisting notions about what is true and false are actively using their education, rather than

merely possessing it.

At first glance, statistics seems like a topic of dubious relevance to everyday life. Convincing criminology and

criminal justice students that they should care about statistics is no small task. Most students approach the

class with apprehension because math is daunting, but many also express frustration and impatience. The

thought, “But I’m going to be a [police officer, lawyer, federal agent, etc.], so what do I need this class for?” is

on many students’ minds as they walk through the door or log in to the learning management system on the

first day. The answer is surprisingly simple: Statistics form a fundamental part of what we know about the

world. Practitioners in the criminal justice field rely on statistics. A police chief who alters a department’s

deployment plan so as to allocate resources to crime hot spots trusts that the researchers who analyzed the

spatial distribution of crime did so correctly. A prison warden seeking to classify inmates according to the risk

they pose to staff and other inmates needs assessment instruments that accurately predict each person’s

likelihood of engaging in behavior that threatens institutional security. A chief prosecutor must recognize that

a high conviction rate might not be testament to assistant prosecutors’ skill level but, rather, evidence that they

only try simple cases and never take on challenges.

Statistics matter because what unites all practitioners in the criminology and criminal justice occupations and

14

professions is the need for valid, reliable data and the ability to critically examine numbers that are set before

them. Students with aspirations for graduate school have to understand statistical concepts because they will

be expected to produce knowledge using these techniques. Those planning to enter the workforce as

practitioners must be equipped with the background necessary to appraise incoming information and evaluate

its accuracy and usefulness. Statistics, therefore, is just as important to information consumers as it is to

producers.

The third edition of Statistics for Criminology and Criminal Justice, like its two predecessors, balances quantity

and complexity with user-friendliness. A book that skimps on information can be as confusing as one

overloaded with it. The sacrificed details frequently pertain to the underlying theory and logic that drive

statistical analyses. The pedagogical techniques employed in this text draw from the scholarship of teaching

and learning, wherein researchers have demonstrated that students learn best when they understand logical

connections within and across concepts, rather than merely memorizing key terms or steps to solving

equations. In statistics, students are at an advantage if they first understand the overarching goal of the

techniques they are learning before they begin working with formulas and numbers.

This book also emphasizes the application of new knowledge. Students can follow along in the step-by-step

instructions that illustrate plugging numbers into formulas and solving them. Additional practice examples are

embedded within the chapters, and chapter review problems allow students to test themselves (the answers to

the odd-numbered problems are located in the back of the book), as well as offering instructors convenient

homework templates using the even-numbered questions.

Real data and research also further the goal of encouraging students to apply concepts and showing them the

relevance of statistics to practical problems in the criminal justice and criminology fields. Chapters contain

Data Sources boxes that describe some common, publicly available data sets such as the Uniform Crime

Reports, National Crime Victimization Survey, General Social Survey, and others. Most in-text examples and

end-of-chapter review problems use data drawn from the sources highlighted in the book. The goal is to lend

a practical, tangible bent to this often-abstract topic. Students get to work with the data their professors use.

They get to see how elegant statistics can be at times and how messy they can be at others, how analyses can

sometimes lead to clear conclusions and other times to ambiguity.

The Research Example boxes embedded throughout the chapters illustrate criminal justice and criminology

research in action and are meant to stimulate students’ interest. They highlight that even though the math

might not be exciting, the act of scientific inquiry most definitely is, and the results have important

implications for policy and practice. In the third edition, the examples have been expanded to include

additional contemporary criminal justice and criminology studies. Most of the examples contained in the first

and second editions were retained in order to enhance diversity and allow students to see firsthand the rich

variety of research that has been taking place over time. The full texts of all articles are available on the SAGE

companion site (http://www.sagepub.com/gau) and can be downloaded online by users with institutional

access to the SAGE journals in which the articles appear.

This edition retains the Learning Check boxes. These are scattered throughout the text and function as mini-

15

http://www.sagepub.com/gau

quizzes that test students’ comprehension of certain concepts. They are short so that students can complete

them without disrupting their learning process. Students can use each Learning Check to make sure they are on

the right track in their understanding of the material, and instructors can use them for in-class discussion. The

answer key is in the back of the book.

Where relevant to the subject matter, chapters end with a section on IBM® SPSS® Statistics2 and come with

one or more shortened versions of a major data set in SPSS file format. Students can download these data sets

to answer the review questions presented at the end of the chapter. The full data sets are all available from the

Inter-University Consortium for Political and Social Research at www.icpsr.umich.edu/icpsrweb/ICPSR/ and

other websites as reported in the text. If desired, instructors can download the original data sets to create

supplementary examples and practice problems for hand calculations or SPSS analyses.

2 SPSS is a registered trademark of International Business Machines Corporation.

The third edition features the debut of Thinking Critically sections. These two-question sections appear at the

end of each chapter. The questions are open-ended and designed to inspire students to think about the

nuances of science and statistics. Instructors can assign them as homework problems or use them to initiate

class debates.

The book is presented in three parts. Part I covers descriptive statistics. It starts with the basics of levels of

measurement and moves on to frequency distributions, graphs and charts, and proportions and percentages.

Students learn how to select the correct type(s) of data display based on a variable’s level of measurement and

then construct that diagram or table. They then learn about measures of central tendency and measures of

dispersion and variability. These chapters also introduce the normal curve.

Part II focuses on probability theory and sampling distributions. This part lays out the logic that forms the

basis of hypothesis testing. It emphasizes the variability in sample statistics that precludes direct inference to

population parameters. Part II ends with confidence intervals, which is students’ first foray into inferential

statistics.

Part III begins with an introduction to bivariate hypothesis testing. The intention is to ease students into

inferential tests by explaining what these tests do and what they are for. This helps transition students from

the theoretical concepts covered in Part II to the application of those logical principles. The remaining

chapters include chi-square tests, t tests and tests for differences between proportions, analysis of variance

(ANOVA), correlation, and ordinary least squares (OLS) regression. The sequence is designed such that

some topics flow logically into others. Chi-square tests are presented first because they are the only

nonparametric test type covered here. Two-population t tests then segue into ANOVA. Correlation, likewise,

supplies the groundwork for regression. Bivariate regression advances from correlation and transitions into the

multivariate framework. The book ends with the fundamentals of interpreting OLS regression models.

This book provides the foundation for a successful statistics course that combines theory, research, and

practical application for a holistic, effective approach to teaching and learning. Students will exit the course

ready to put their education into action as they prepare to enter their chosen occupation, be that in academia,

16

http://www.icpsr.umich.edu/icpsrweb/ICPSR/

law, or the field. Learning statistics is not a painless process, but the hardest classes are the ones with the

greatest potential to leave lasting impressions. Students will meet obstacles, struggle with them, and ultimately

surmount them so that in the end, they will look back and say that the challenge was worth it.

17

Acknowledgments

The third edition of this book came about with input and assistance from multiple people. With regard to the

development and preparation of this manuscript, I wish to thank Jessica Miller and the staff at SAGE for

their support and encouragement, as well as Alison Hope for her excellent copyediting assistance. You guys

are the best! I owe gratitude to my family and friends who graciously tolerate me when I am in “stats mode”

and a tad antisocial. Numerous reviewers supplied advice, recommendations, and critiques that helped shape

this book. Reviewers for the third edition are listed in alphabetical order here. Of course, any errors contained

in this text are mine alone.

Calli M. Cain, University of Nebraska at Omaha

Kyleigh Clark, University of Massachusetts, Lowell

Jane C. Daquin, Georgia State University

Courtney Feldscher, University of Massachusetts Boston

Albert M. Kopak, Western Carolina University

Bonny Mhlanga, Western Illinois University

Elias Nader, University of Massachusetts Lowell

Tyler J. Vaughan, Texas State University

Egbert Zavala, The University of Texas at El Paso

Reviewers for the second edition:

Jeb A. Booth, Salem State University

Ayana Conway, Virginia State University

Matthew D. Fetzer, Shippensburg University

Anthony W. Hoskin, University of Texas of the Permian Basin

Shelly A. McGrath, University of Alabama at Birmingham

Bonny Mhlanga, Western Illinois University

Carlos E. Posadas, New Mexico State University

Scott Senjo, Weber State University

Nicole L. Smolter, California State University, Los Angeles

Brian Stults, Florida State University

George Thomas, Albany State University

18

About the Author

Jacinta M. Gau, Ph.D.,

is an associate professor in the Department of Criminal Justice at the University of Central Florida. She

received her Ph.D. from Washington State University. Her primary areas of research are policing and

criminal justice policy, and she has a strong quantitative background. Dr. Gau’s work has appeared in

journals such as Justice Quarterly, British Journal of Criminology, Criminal Justice and Behavior, Crime &

Delinquency, Criminology & Public Policy, Police Quarterly, Policing: An International Journal of Police

Strategies & Management, and the Journal of Criminal Justice Education. In addition to Statistics for

Criminology and Criminal Justice, she is author of Criminal Justice Policy: Origins and Effectiveness (Oxford

University Press) and coauthor of Key Ideas in Criminology and Criminal Justice (SAGE). Additionally,

she coedits Race and Justice: An International Journal, published by SAGE.

19

Part I Descriptive Statistics

Chapter 1 Introduction to the Use of Statistics in Criminal Justice and Criminology

Chapter 2 Types of Variables and Levels of Measurement

Chapter 3 Organizing, Displaying, and Presenting Data

Chapter 4 Measures of Central Tendency

Chapter 5 Measures of Dispersion

20

Chapter 1 Introduction to the Use of Statistics in Criminal Justice

and Criminology

21

Learning Objectives

Explain how data collected using scientific methods are different from anecdotes and other nonscientific information.

List and describe the types of research in criminal justice and criminology.

Explain the difference between the research methods and statistical analysis.

Define samples and populations.

Describe probability sampling.

List and describe the three major statistics software packages.

You might be thinking, “What do statistics have to do with criminal justice or criminology?” It is reasonable

for you to question the requirement that you spend an entire term poring over a book about statistics instead

of one about policing, courts, corrections, or criminological theory. Many criminology and criminal justice

undergraduates wonder, “Why am I here?” In this context, the question is not so much existential as it is

practical. Luckily, the answer is equally practical.

You are “here” (in a statistics course) because the answer to the question of what statistics have to do with

criminal justice and criminology is “Everything!” Statistical methods are the backbone of criminal justice and

criminology as fields of scientific inquiry. Statistics enable the construction and expansion of knowledge about

criminality and the criminal justice system. Research that tests theories or examines criminal justice

phenomena and is published in academic journals and books is the basis for most of what we know about

criminal offending and the system that has been designed to deal with it. The majority of this research would

not be possible without statistics.

Statistics can be abstract, so this book uses two techniques to add a realistic, pragmatic dimension to the

subject. The first technique is the use of examples of statistics in criminal justice and criminology research.

These summaries are contained in the Research Example boxes embedded in each chapter. They are meant to

give you a glimpse into the types of questions that are asked in this field of research and the ways in which

specific statistical techniques are used to answer those questions. You will see firsthand how lively and diverse

criminal justice and criminology research is. Research Example 1.1 summarizes seven studies. Take a moment

now to read through them.

The second technique to add a realistic, pragmatic dimension to the subject of this book is the use of real data

from reputable and widely used sources such as the Bureau of Justice Statistics (BJS). The BJS is housed

within the U.S. Department of Justice and is responsible for gathering, maintaining, and analyzing data on

various criminal justice topics at the county, state, and national levels. Visit http://bjs.ojp.usdoj.gov/ to

familiarize yourself with the BJS. The purpose behind the use of real data is to give you the type of hands-on

experience that you cannot get from fictional numbers. You will come away from this book having worked

with some of the same data that criminal justice and criminology researchers use. Two sources of data that will

be used in upcoming chapters are the Uniform Crime Reports (UCR) and the National Crime Victimization

Survey (NCVS). See Data Sources 1.1 and 1.2 for information about these commonly used measures of

criminal incidents and victimization, respectively. All the data sets used in this book are publicly available and

22

http://bjs.ojp.usdoj.gov/

were downloaded from governmental websites and the archive maintained by the Inter-University

Consortium for Political and Social Research at www.icpsr.umich.edu.

Research Example 1.1 What Do Criminal Justice and Criminology Researchers Study?

Researchers in the field of criminology and criminal justice examine a wide variety of issues pertaining to the criminal justice system

and theories of offending. Included are topics such as prosecutorial charging decisions, racial and gender disparities in sentencing,

police use of force, drug and domestic violence courts, and recidivism. The following are examples of studies that have been

conducted and published. You can find the full text of each of these articles and of all those presented in the following chapters at

(www.sagepub.com/gau).

1. Can an anticrime strategy that has been effective at reducing certain types of violence also be used to combat open-air drug markets?

The “pulling levers” approach involves deterring repeat offenders from crime by targeting them for enhanced prosecution

while also encouraging them to change their behavior by offering them access to social services. This strategy has been shown

to hold promise with gang members and others at risk for committing violence. The Rockford (Illinois) Police Department

(RPD) decided to find out if they could use a pulling levers approach to tackle open-air drug markets and the crime problems

caused by these nuisance areas. After the RPD implemented the pulling levers intervention, Corsaro, Brunson, and

McGarrell (2013) used official crime data from before and after the intervention to determine whether this approach had

been effective. They found that although there was no reduction in violent crime, nonviolent crime (e.g., drug offenses,

vandalism, and disorderly conduct) declined noticeably after the intervention. This indicated that the RPD’s efforts had

worked, because drug and disorder offenses were exactly what the police were trying to reduce.

2. Are prisoners with low self-control at heightened risk of victimizing, or being victimized by, other inmates? Research has

consistently shown that low self-control is related to criminal offending. Some studies have also indicated that this trait is a

risk factor for victimization, in that people with low self-control might place themselves in dangerous situations. One of the

central tenets of this theory is that self-control is stable and acts in a uniform manner regardless of context. Kerley,

Hochstetler, and Copes (2009) tested this theory by examining whether the link between self-control and both offending and

victimization held true within the prison environment. Using data gathered from surveys of prison inmates, the researchers

discovered that low self-control was only slightly related to in-prison offending and victimization. This result could challenge

the assumption that low self-control operates uniformly in all contexts. To the contrary, something about prisoners

themselves, the prison environment, or the interaction between the two might change the dynamics of low self-control.

3. Does school racial composition affect how severely schools punish black and Latino students relative to white ones? Debates about the

so-called school-to-prison pipeline emphasize the long-term effects of school disciplinary actions such as suspension,

expulsion, and arrest or court referral. Youth who experience these negative outcomes are at elevated risk for dropping out of

school and getting involved in delinquency and, eventually, crime. Hughes, Warren, Stewart, Tomaskovic-Devey, and Mears

(2017) set out to discover whether schools’ and school boards’ racial composition affects the treatment of black, Latino, and

white students. The researchers drew from two theoretical perspectives: The racial threat perspective argues that minorities

are at higher risk for punitive sanctions when minority populations are higher, because whites could perceive minority groups

as a threat to their place in society. On the other hand, the intergroup contact perspective suggests that racial and ethnic

diversity reduces the harshness of sanctions for minorities, because having contact with members of other racial and ethnic

groups diminishes prejudice. Hughes and colleagues used data from the Florida Department of Education, the U.S. Census

Bureau, and the Uniform Crime Reports. Statistical results provided support for both perspectives. Increases in the size of

the black and Hispanic student populations led to higher rates of suspension for students of these groups. On the other hand,

intergroup contact among school board members of different races reduced suspensions for all students. The researchers

concluded that interracial contact among school board members equalized disciplinary practices and reduced discriminatory

disciplinary practices.

4. What factors influence police agencies’ ability to identify and investigate human trafficking? Human trafficking has been

recognized as a transnational crisis. Frequently, local police are the first ones who encounter victims or notice signs

suggesting the presence of trafficking. In the United States, however, many local police agencies do not devote systematic

attention to methods that would enable them to detect and investigate suspected traffickers. Farrell (2014) sought to learn

more about U.S. police agencies’ antitrafficking efforts. Using data from two national surveys of medium-to-large municipal

police departments, Farrell found that 40% of departments trained their personnel on human trafficking, 17% had written

policies pertaining to this crime, and 13% dedicated personnel to it. Twenty-eight percent had investigated at least one

23

http://www.icpsr.umich.edu

http://www.sagepub.com/gau

trafficking incident in the previous six years. Larger departments were more likely to have formalized responses (training,

policies, and dedicated personnel), and departments that instituted these responses were more likely to have engaged in

trafficking investigations. These results show a need for departments to continue improving their antitrafficking efforts.

Departments that are more responsive to local problems and more open to change will be more effective at combating this

crime.

5. How safe and effective are conducted energy devices as used by police officers? Conducted energy devices (CEDs) have proliferated

in recent years. Their widespread use and the occasional high-profile instances of misuse have generated controversy over

whether these devices are safe for suspects and officers alike. Paoline, Terrill, and Ingram (2012) collected use-of-force data

from six police agencies nationwide and attempted to determine whether officers who deployed CEDs against suspects were

more or less likely to sustain injuries themselves. The authors’ statistical analysis suggested a lower probability of officer

injury when only CEDs were used. When CEDs were used in combination with other forms of force, however, the

probability of officer injury increased. The results suggest that CEDs can enhance officer safety, but they are not a panacea

that uniformly protects officers in all situations.

6. How prevalent is victim precipitation in intimate partner violence? A substantial number of violent crimes are initiated by the

person who ultimately becomes the victim in an incident. Muftić, Bouffard, and Bouffard (2007) explored the role of victim

precipitation in instances of intimate partner violence (IPV). They gleaned data from IPV arrest reports and found that

victim precipitation was present in cases of both male and female arrestees but that it was slightly more common in instances

where the woman was the one arrested. This suggests that some women (and, indeed, some men) arrested for IPV might be

responding to violence initiated by their partners rather than themselves being the original aggressors. The researchers also

discovered that victim precipitation was a large driving force behind dual arrests (cases in which both parties are arrested),

because police could either see clearly that both parties were at fault or, alternatively, were unable to determine which party

was the primary aggressor. Victim precipitation and the use of dual arrests, then, could be contributing factors behind the

recent rise in the number of women arrested for IPV against male partners.

7. What are the risk factors in a confrontational arrest that are most commonly associated with the death of the suspect? There have been

several high-profile instances of suspects dying during physical confrontations with police wherein the officers deployed

CEDs against these suspects. White and colleagues (2013) collected data on arrest-related deaths (ARDs) that involved

CEDs and gained media attention. The researchers triangulated the data using information from medical-examiner reports.

They found that in ARDs, suspects were often intoxicated and extremely physically combative with police. Officers, for their

part, had used several other types of force before or after trying to solve the situation using CEDs. Medical examiners most

frequently attributed these deaths to drugs, heart problems, and excited delirium. These results suggest that police

departments should craft policies to guide officers’ use of CEDs against suspects who are physically and mentally

incapacitated.

In this book, emphasis is placed on both the production and interpretation of statistics. Every statistical

analysis has a producer (someone who runs the analysis) and a consumer (someone to whom an analysis is

being presented). Regardless of which role you play in any given situation, it is vital for you to be sufficiently

versed in quantitative methods that you can identify the proper statistical technique and correctly interpret the

results. When you are in the consumer role, you must also be ready to question the methods used by the

producer so that you can determine for yourself how trustworthy the results are. Critical thinking skills are an

enormous component of statistics. You are not a blank slate standing idly by, waiting to be written on—you

are an active agent in your acquisition of knowledge about criminal justice, criminology, and the world in

general. Be critical, be skeptical, and never hesitate to ask for more information.

Data Sources 1.1 The Uniform Crime Reports

The Federal Bureau of Investigation (FBI) collects annual data on crimes reported to police agencies nationwide and maintains the

UCR. Crimes are sorted into eight index offenses: homicide, rape, robbery, aggravated assault, burglary, larceny-theft, motor vehicle

theft, and arson. An important aspect of this data set is that it includes only those crimes that come to the attention of police—

crimes that are not reported or otherwise detected by police are not counted. The UCR also conforms to the hierarchy rule, which

24

mandates that in multiple-crime incidents only the most serious offense ends up in the UCR. If, for example, someone breaks into a

residence with intent to commit a crime inside the dwelling and while there, kills the homeowner and then sets fire to the structure

to hide the crime, he has committed burglary, murder, and arson. Because of the hierarchy rule, though, only the murder would be

reported to the FBI—it would be as if the burglary and arson had never occurred. Because of underreporting by victims and the

hierarchy rule, the UCR undercounts the amount of crime in the United States. It nonetheless offers valuable information and is

widely used. You can explore this data source at www.fbi.gov/about-us/cjis/ucr/ucr.

Data Sources 1.2 The National Crime Victimization Survey

The U.S. Census Bureau conducts the periodic NCVS under the auspices of the BJS to estimate the number of criminal incidents

that transpire each year and to collect information about crime victims. Multistage cluster sampling is used to select a random sample

of households, and each member of that household who is 12 years or older is asked to participate in an interview. Those who agree

to be interviewed are asked over the phone or in person about any and all criminal victimizations that transpired in the 6 months

prior to the interview. The survey employs a rotating panel design, so respondents are called at 6-month intervals for a total of 3

years, and then new respondents are selected (BJS, 2006). The benefit of the NCVS over the UCR is that NCVS respondents might

disclose victimizations to interviewers that they did not report to police, thus making the NCVS a better estimation of the total

volume of crime in the United States. The NCVS, though, suffers from the weakness of being based entirely on victims’ memory

and honesty about the timing and circumstances surrounding criminal incidents. The NCVS also excludes children younger than 12

years, institutionalized populations (e.g., persons in prisons, nursing homes, and hospitals), and the homeless. Despite these

problems, the NCVS is useful because it facilitates research into the characteristics of crime victims. The 2015 wave of the NCVS is

the most recent version available as of this writing.

25

http://www.fbi.gov/about-us/cjis/ucr/ucr

Science: Basic Terms and Concepts

There are a few terms and concepts that you must know before you get into the substance of the book.

Statistics are a tool in the larger enterprise of scientific inquiry. Science is the process of systematically

collecting reliable information and developing knowledge using techniques and procedures that are accepted

by other scientists in a discipline. Science is grounded in methods—research results are trustworthy only when

the procedures used to arrive at them are considered correct by experts in the scientific community.

Nonscientific information is that which is collected informally or without regard for correct methods.

Anecdotes are a form of nonscientific information. If you ask one person why he or she committed a crime,

that person’s response will be an anecdote; it cannot be assumed to be broadly true of other offenders. If you

use scientific methods to gather a large group of offenders and you survey all of them about their motivations,

you will have data that you can analyze using statistics and that can be used to draw general conclusions.

Science: The process of gathering and analyzing data in a systematic and controlled way using procedures that are generally accepted

by others in the discipline.

Methods: The procedures used to gather and analyze scientific data.

In scientific research, samples are drawn from populations using scientific techniques designed to ensure that

samples are representative of populations. For instance, if the population is 50% male, then the sample should

also be approximately 50% male. A sample that is only 15% male is not representative of the population.

Research-methods courses instruct students on the proper ways to gather representative samples. In a statistics

course, the focus is on techniques used to analyze the data to look for patterns and test for relationships.

Together, proper methods of gathering and analyzing data form the groundwork for scientific inquiry. If there

is a flaw in either the gathering or the analyzing of data, then the results might not be trustworthy. Garbage

in, garbage out (GIGO) is the mantra of statistics. Data gathered with the best of methods can be rendered

worthless if the wrong statistical analysis is applied to them; likewise, the most sophisticated, cutting-edge

statistical technique cannot salvage improperly collected data. When the data or the statistics are defective, the

results are likewise deficient and cannot be trusted. Studies using unscientific data or flawed statistical analyses

do not contribute to theory and research or to policy and practice because their findings are unreliable and

could be erroneous.

Sample: A subset pulled from a population with the goal of ultimately using the people, objects, or places in the sample as a way to

generalize to the population.

Population: The universe of people, objects, or locations that researchers wish to study. These groups are often very large.

26

Learning Check 1.1

Identify whether each of the following is a sample or a population.

1. A group of 100 police officers pulled from a department with 300 total officers

2. Fifty prisons selected at random from all prisons nationwide

3. All persons residing in the state of Wisconsin

4. A selection of 10% of the defendants processed through a local criminal court in 1 year

Everybody who conducts a study has an obligation to be clear and open about the methods they used. You

should expect detailed reports on the procedures used so that you can evaluate whether they followed proper

scientific methods. When the methods used to collect and analyze data are sound, it is not appropriate to

question scientific results on the basis of a moral, emotional, or opinionated objection to them. On the other

hand, it is entirely correct (and is necessary, in fact) to question results when methodological or statistical

procedures are shoddy or inadequate. Remember GIGO!

A key aspect of science is the importance of replication. No single study ever proves something definitively;

quite to the contrary, much testing must be done before firm conclusions can be drawn. Replication is

important because there are times when a study is flawed and needs to be redone or when the original study is

methodologically sound but needs to be tested on new populations and samples. For example, a correctional

treatment program that reduces recidivism rates among adults might or might not have similar positive results

with juveniles. Replicating the treatment and evaluation with a sample of juvenile offenders would provide

information about whether the program is helpful to both adults and juveniles or is only appropriate for

adults. The scientific method’s requirement that all researchers divulge the steps they took to gather and

analyze data allows other researchers and members of the public to examine those steps and, when warranted,

to undertake replications.

Replication: The repetition of a particular study that is conducted for purposes of determining whether the original study’s results

hold when new samples or measures are employed.

27

Types of Scientific Research in Criminal Justice and Criminology

Criminal justice and criminology research is diverse in nature and purpose. Much of it involves theory testing.

Theories are proposed explanations for certain events. Hypotheses are small “pieces” of theories that must be

true in order for the entire theory to hold up. You can think of a theory as a chain and hypotheses as the links

forming that chain. Research Example 1.1 discusses a test of the general theory of crime conducted by Kerley

et al. (2009). The general theory holds that low self-control is a static predictor of offending and

victimization, regardless of context. From this proposition, the researchers deduced the hypothesis that the

relationship between low self-control and both offending and victimization must hold true in the prison

environment. Their results showed an overall lack of support for the hypothesis that low self-control operates

uniformly in all contexts, thus calling that aspect of the general theory of crime into question. This is an

example of a study designed to test a theory.

Theory: A set of proposed and testable explanations about reality that are bound together by logic and evidence.

Hypothesis: A single proposition, deduced from a theory, that must hold true in order for the theory itself to be considered valid.

Evaluation research is also common in criminal justice and criminology. In Research Example 1.1, the article

by Corsaro et al. (2013) is an example of evaluation research. This type of study is undertaken when a new

policy, program, or intervention is put into place and researchers want to know whether the intervention

accomplished its intended purpose. In this study, the RPD implemented a pulling levers approach to combat

drug and nuisance offending. After the program had been put into place, the researchers analyzed crime data

to find out whether the approach was effective.

Evaluation research: Studies intended to assess the results of programs or interventions for purposes of discovering whether those

programs or interventions appear to be effective.

Exploratory research occurs when there is limited knowledge about a certain phenomenon; researchers

essentially embark into unfamiliar territory when they attempt to study this social event. The study by Muftić

et al. (2007) in Research Example 1.1 was exploratory in nature because so little is known about victim

precipitation, particularly in the realm of IPV. It is often dangerous to venture into new areas of study when

the theoretical guidance is spotty; however, exploratory studies have the potential to open new areas of

research that have been neglected but that provide rich information that expands the overall body of

knowledge.

Exploratory research: Studies that address issues that have not been examined much or at all in prior research and that therefore

might lack firm theoretical and empirical grounding.

Finally, some research is descriptive in nature. White et al.’s (2013) analysis of CED-involved deaths

illustrates a descriptive study. White and colleagues did not set out to test a theory or to explore a new area of

research—they merely offered basic descriptive information about the suspects, officers, and situations

involved in instances where CED use was associated with a suspect’s death. In descriptive research, no

28

generalizations are made to larger groups; the conclusions drawn from these studies are specific to the objects,

events, or people being analyzed. This type of research can be very informative when knowledge about a

particular phenomenon is scant.

Descriptive research: Studies done solely for the purpose of describing a particular phenomenon as it occurs in a sample.

With the exception of purely descriptive research, the ultimate goal in most statistical analyses is to generalize

from a sample to a population. A population is the entire set of people, places, or objects that a researcher

wishes to study. Populations, though, are usually very large. Consider, for instance, a researcher trying to

estimate attitudes about capital punishment in the general U.S. population. That is a population of more than

300 million! It would be impossible to measure everyone directly. Researchers thus draw samples from

populations and study the samples instead. Probability sampling helps ensure that a sample mirrors the

population from which it was drawn (e.g., a sample of people should contain a breakdown of race, gender, and

age similar to that found in the population). Samples are smaller than populations, and researchers are

therefore able to measure and analyze them. The results found in the sample are then generalized to the

population.

Probability sampling: A sampling technique in which all people, objects, or areas in a population have a known chance of being

selected into the sample.

29

Learning Check 1.2

For each of the following scenarios, identify the type of research being conducted.

1. A researcher wants to know more about female serial killers. He gathers news articles that report on female serial killers and

records information about each killer’s life history and the type of victim she preyed on.

2. A researcher wants to know whether a new in-prison treatment program is effective at reducing recidivism. She collects a sample

of inmates that participated in the program and a sample that did not go through the program. She then gathers recidivism data

for each group to see if those who participated had lower recidivism rates than those who did not.

3. The theory of collective efficacy predicts that social ties between neighbors, coupled with neighbors’ willingness to intervene when

a disorderly or criminal event occurs in the area, protect the area from violent crime. A researcher gathers a sample of

neighborhoods and records the level of collective efficacy and violent crime in each one to determine whether those with higher

collective efficacy have lower crime rates.

4. A researcher notes that relatively little research has been conducted on the possible effects of military service on later crime

commission. She collects a sample of people who served in the military and a sample of people that did not and compares them to

determine whether the military group differs from the nonmilitary group in terms of the numbers or types of crimes committed.

30

Software Packages for Statistical Analysis

Hand computations are the foundation of this book because seeing the numbers and working with the

formulas facilitates an understanding of statistical analyses. In the real world, however, statistical analysis is

generally conducted using a software program. Microsoft Excel contains some rudimentary statistical

functions and is commonly used in situations requiring only basic descriptive analyses; however, this program’s

usefulness is exhausted quickly because researchers usually want far more than descriptives. Many statistical

packages are available. The most common in criminal justice and criminology research are SPSS, Stata, and

SAS. Each of these packages has strengths and weaknesses. Simplicity and ease of use makes SPSS a good

place to start for people new to statistical analysis. Stata is a powerful program excellent for regression

modeling. The SAS package is the best one for extremely large data sets.

This book incorporates SPSS into each chapter. This allows you to get a sense for what data look like when

displayed in their raw format and permits you to run particular analyses and read and interpret program

output. Where relevant, the chapters offer SPSS practice problems and accompanying data sets that are

available for download from www.sagepub.com/gau. This offers a practical, hands-on lesson about the way

that criminal justice and criminology researchers use statistics.

31

http://www.sagepub.com/gau

Organization of the Book

This book is divided into three parts. Part I (Chapters 1 through 5) covers descriptive statistics. Chapter 2

provides a basic overview of types of variables and levels of measurement. Some of this material will be review

for students who have taken a methods course. Chapter 3 delves into charts and graphs as means of

graphically displaying data. Measures of central tendency are the topic of Chapter 4. These are descriptive

statistics that let you get a feel for where the data are clustered. Chapter 5 discusses measures of dispersion.

Measures of dispersion complement measures of central tendency by offering information about whether the

data tend to cluster tightly around the center or, conversely, whether they are very spread out.

Part II (Chapters 6 through 8) describes the theoretical basis for statistics in criminal justice and criminology:

probability and probability distributions. Part I of the book can be thought of as the nuts-and-bolts of the

mathematical concepts used in statistics, and Part II can be seen as the theory behind the math. Chapter 6

introduces probability theory. Binomial and continuous probability distributions are discussed. In Chapter 7,

you will learn about population, sample, and sampling distributions. Chapter 8 provides the book’s first

introduction to inferential statistics with its coverage of point estimates and confidence intervals. The

introduction of inferential statistics at this juncture is designed to help ease you into Part III.

Part III (Chapters 9 through 14) of the book merges the concepts learned in Parts I and II to form the

discussion on inferential hypothesis testing. Chapter 9 offers a conceptual introduction to this framework,

including a description of the five steps of hypothesis testing that will be used in every proceeding chapter. In

Chapter 10, you will encounter your first bivariate statistical technique: chi-square. Chapter 11 describes two-

population t tests and tests for differences between proportions. Chapter 12 covers analysis of variance, which

is an extension of the two-population t test. In Chapter 13, you will learn about correlations. Finally, Chapter

14 wraps up the book with an introduction to bivariate and multiple regression.

The prerequisite that is indispensable to success in this course is a solid background in algebra. You absolutely

must be comfortable with basic techniques such as adding, subtracting, multiplying, and dividing. You also

need to understand the difference between positive and negative numbers. You will be required to plug

numbers into equations and solve those equations. You should not have a problem with this as long as you

remember the lessons you learned in your high school and college algebra courses. Appendix A offers an

overview of the basic mathematical techniques you will need to know, so look those over and make sure that

you are ready to take this course. If necessary, use them to brush up on your skills.

Statistics are cumulative in that many of the concepts you learn at the beginning form the building blocks for

more-complex techniques that you will learn about as the course progresses. Means, proportions, and standard

deviations, for instance, are concepts you will learn about in Part I, but they will remain relevant throughout

the remainder of the book. You must, therefore, learn these fundamental calculations well and you must

remember them.

Repetition is the key to learning statistics. Practice, practice, practice! There is no substitute for doing and

32

redoing the end-of-chapter review problems and any other problems your instructor might provide. You can

also use the in-text examples as problems if you just copy down the numbers and do the calculations on your

own without looking at the book. Remember, even the most advanced statisticians started off knowing

nothing about statistics. Everyone has to go through the learning process. You will complete this process

successfully as long as you have basic algebra skills and are willing to put in the time and effort it takes to

succeed.

Thinking Critically

1. Media outlets and other agencies frequently conduct opinion polls to try to capture information about the public’s thoughts

on contemporary events, controversies, or political candidates. Poll data are faster and easier to collect than survey data are,

because they do not require adherence to scientific sampling methods and questionnaire design. Agencies conducting polls

often do not have the time or resources to engage in full-scale survey projects. Debate the merits of poll data from a policy

standpoint. Is having low-quality information better than having none at all? Or is there no place in public discussions for

data that fall short of scientific standards? Explain your answer.

2. Suppose you tell a friend that you are taking a statistics course, and your friend reacts with surprise that a criminology or

criminal justice degree program would require students to take this class. Your friend argues that although it is necessary for

people whose careers are dedicated to research to have a good understanding of statistics, this area of knowledge is not useful

for people with practitioner jobs, such as police and corrections officers. Construct a response to this assertion. Identify ways

in which people in practical settings benefit from possessing an understanding of statistical concepts and techniques.

Review Problems

1. Define science and explain the role of methods in the production of scientific knowledge.

2. What is a population? Why are researchers usually unable to study populations directly?

3. What is a sample? Why do researchers draw samples?

4. Explain the role of replication in science.

5. List and briefly describe the different types of research in criminal justice and criminology.

6. Identify three theories that you have encountered in your criminal justice or criminology classes. For each one, write one

hypothesis for which you could collect data in order to test that hypothesis.

7. Think of three types of programs or policies you have heard about or read about in your criminal justice or criminology

classes. For each one, suggest a possible way to evaluate that program’s or policy’s effectiveness.

8. If a researcher were conducting a study on a topic about which very little is known and the researcher does not have theory or

prior evidence to make predictions about what she will find in her study, what kind of research would she be doing? Explain

your answer.

9. If a researcher were solely interested in finding out more about a particular phenomenon and focused entirely on a sample

without trying to make inference to a population, what kind of research would he be doing? Explain your answer.

10. What does GIGO stand for? What does this cautionary concept mean in the context of statistical analyses?

33

Key Terms

Science 8

Methods 8

Sample 8

Population 9

Replication 9

Theory 10

Hypothesis 10

Evaluation research 10

Exploratory research 10

Descriptive research 10

Probability sampling 11

34

Chapter 2 Types of Variables and Levels of Measurement

35

Learning Objectives

Define variables and constants.

Define unit of analysis and be able to identify the unit of analysis in any given study.

Define independent and dependent variables and be able to identify each in a study.

Explain the difference between empirical associations and causation.

List and describe the four levels of measurement, including similarities and differences between them, and be able to identify the

level of measurement of different variables.

The first thing you must be familiar with in statistics is the concept of a variable. A variable is, quite simply,

something that varies. It is a coding scheme used to measure a particular characteristic of interest. For

instance, asking all of your statistics classmates, “How many classes are you taking this term?” would yield

many different answers. This would be a variable. Variables sit in contrast to constants, which are

characteristics that assume only one value in a sample. It would be pointless for you to ask all your classmates

whether they are taking statistics this term because of course the answer they would all provide is “yes.”

Variable: A characteristic that describes people, objects, or places and takes on multiple values in a sample or population.

Constant: A characteristic that describes people, objects, or places and takes on only one value in a sample or population.

36

Units of Analysis

It seems rather self-evident, but nonetheless bears explicit mention, that every scientific study contains

something that the researcher conducting the study gathers and examines. These “somethings” can be objects

or entities such as rocks, people, molecules, or prisons. This is called the unit of analysis, and it is, essentially,

whatever the sample under study consists of. In criminal justice and criminology research, individual people

are often the units of analysis. These individuals might be probationers, police officers, criminal defendants, or

judges. Prisons, police departments, criminal incidents, or court records can also be units of analysis. Larger

units are also popular; for example, many studies focus on census tracks, block groups, cities, states, or even

countries. Research Example 2.2 describes the methodological setup of a selection of criminal justice studies,

each of which employed a different unit of analysis.

Unit of analysis: The object or target of a research study.

37

Independent Variables and Dependent Variables

Researchers in criminal justice and criminology typically seek to examine relationships between two or more

variables. Observed or empirical phenomena give rise to questions about the underlying forces driving them.

Take homicide as an example. Homicide events and city-level rates are empirical phenomena. It is worthy of

note that Washington, D.C., has a higher homicide rate than Portland, Oregon. Researchers usually want to

do more than merely note empirical findings, however—they want to know why things are the way they are.

They might, then, attempt to identify the criminogenic (crime-producing) factors that are present in

Washington but absent in Portland or, conversely, the protective factors possessed by Portland and lacked by

Washington.

Empirical: Having the qualities of being measurable, observable, or tangible. Empirical phenomena are detectable with senses such as

sight, hearing, or touch.

Research Example 2.1 Choosing Variables for a Study on Police Use of Conducted Energy Devices

Conducted energy devices (CEDs) such as the Taser have garnered national—indeed, international—attention in the past few years.

Police practitioners contend that CEDs are invaluable tools that minimize injuries to both officers and suspects during contentious

confrontations, whereas critics argue that police sometimes use CEDs in situations where such a high level of force is not warranted.

Do police seem to be using CEDs appropriately? Gau, Mosher, and Pratt (2010) addressed this question. They sought to determine

whether suspects’ race or ethnicity influenced the likelihood that police officers would deploy or threaten to deploy CEDs against

those suspects. In an analysis of this sort, it is important to account for other variables that might be related to police use of CEDs or

other types of force; therefore, the researchers included suspects’ age, sex, and resistance level. They also measured officers’ age, sex,

and race. Finally, they included a variable indicating whether it was light or dark outside at the time of the encounter. The

researchers found that police use of CEDs was driven primarily by the type and intensity of suspect resistance but that even

controlling for resistance, Latino suspects faced an elevated probability of having CEDs either drawn or deployed against them.

Research Example 2.2 Units of Analysis

Each of the following studies used a different unit of analysis.

1. Do prison inmates incarcerated in facilities far from their homes commit more misconduct than those housed in facilities closer to home?

Lindsey, Mears, Cochran, Bales, and Stults (2017) used data from the Florida Department of Corrections to find out

whether distally placed inmates (i.e., those sent to facilities far from their homes) engaged in more in-prison misbehavior,

and, if so, whether this effect was particularly pronounced for younger inmates. Individual prisoners were the units of analysis

in this study. The findings revealed a curvilinear relationship between distance and misconduct: Prisoners’ misconduct

increased along with distance up to approximately 350 miles, but then the relationship inverted such that further increases in

distance were associated with less misconduct. As predicted, this pattern was strongest among younger inmates. Visitation

helped offset the negative impact of distance but did not eliminate it. The researchers concluded that family visitation might

have mixed effects on inmates. Inmates might be less inclined to commit misconduct if they fear losing visitation privileges,

but receiving visits might induce embarrassment and shame when their family sees them confined in the prison environment.

This strain, in turn, could prompt them to act out. Those who do not see their families much or at all do not experience this

unpleasant emotional reaction.

2. Is the individual choice to keep a firearm in the home affected by local levels of crime and police strength? Kleck and Kovandzic

(2009), using individual-level data from the General Social Survey (GSS) and city-level data from the FBI, set out to

determine whether city-level homicide rates and the number of police per 100,000 city residents affected GSS respondents’

likelihood of owning a firearm. There were two units of analysis in this study: individuals and cities. The statistical models

indicated that high homicide rates and low police levels both modestly increased the likelihood that a given person would

own a handgun; however, the relationship between city homicide rate and individual gun ownership decreased markedly

38

when the authors controlled for whites’ and other nonblacks’ racist attitudes toward African Americans. It thus appeared that

the homicide–gun ownership relationship was explained in part by the fact that those who harbored racist sentiments against

blacks were more likely to own firearms regardless of the local homicide rate.

3. How consistent are use-of-force policies across police agencies? The U.S. Supreme Court case Graham v. Connor (1989) requires

that police officers use only the amount of force necessary to subdue a resistant suspect; force exceeding that minimum is

considered excessive. The Court left it up to police agencies to establish force policies to guide officers’ use of physical

coercion. Terrill and Paoline (2013) sought to determine what these policies look like and how consistent they are across

agencies. The researchers mailed surveys to a sample of 1,083 municipal police departments and county sheriff offices

nationwide, making the agency the unit of analysis. Results showed that 80% of agencies used a force continuum as part of

their written use-of-force policies, suggesting some predictability in the way in which agencies organize their policies.

However, there was substantial variation in policy restrictiveness and the placement of different techniques and weapons.

Most agencies placed officer presence and verbal commands at the lowest end, and deadly force at the highest, but between

those extremes there was variability in the placement of soft and hard hand tactics, chemical sprays, impact weapons, CEDs,

and other methods commonly used to subdue noncompliant suspects. These findings show how localized force policies are,

and how inconsistent they are across agencies.

4. Does gentrification reduce gang homicide? Gentrification is the process by which distressed inner-city areas are transformed by

an influx of new businesses or higher-income residents. Gentrification advocates argue that the economic boost will revitalize

the area, provide new opportunities, and reduce crime. Is this assertion true? Smith (2014) collected data from 1994 to 2005

on all 342 neighborhoods in Chicago with the intention of determining whether gentrification over time reduces gang-

motivated homicide. Smith measured gentrification in three ways: Recent increases in neighborhood residents’

socioeconomic statuses, increases in coffee shops, and demolition of public housing. The author predicted that the first two

would suppress gang homicide and that the last one would increase it; even though public-housing demolition is supposed to

reduce crime, it can also create turmoil, residential displacement, and conflict among former public-housing residents and

residents of surrounding properties. Smith found support for all three hypotheses. Socioeconomic-status increases were

strongly related to reductions in gang-motivated homicides, coffee-shop presence was weakly related to reductions, and

public-housing demolition was robustly associated with increases. These results suggest that certain forms of gentrification

might be beneficial to troubled inner-city neighborhoods but that demolishing public housing might cause more problems

than it solves, at least in the short term.

Researchers undertaking quantitative studies must specify dependent variables (DVs) and independent

variables (IVs). Dependent variables are the empirical events that a researcher is attempting to explain.

Homicide rates, property crime rates, recidivism among recently released prisoners, and judicial sentencing

decisions are examples of DVs. Researchers seek to identify variables that help predict or explain these events.

Independent variables are factors a researcher believes might affect the DV. It might be predicted, for

instance, that prisoners released into economically and socially distressed neighborhoods and given little

support during the reentry process will recidivate more frequently than those who receive transitional housing

and employment assistance. Different variables—crime rates, for instance—can be used as both IVs and DVs

across different studies. The designation of a certain phenomenon as an IV or a DV depends on the nature of

the research study.

Dependent variable: The phenomenon that a researcher wishes to study, explain, or predict.

Independent variable: A factor or characteristic that is used to try to explain or predict a dependent variable.

39

Relationships Between Variables: A Cautionary Note

It is vital to understand that independent and dependent are not synonymous with cause and effect, respectively.

A particular IV might be related to a certain DV, but this is far from definitive proof that the former is the

cause of the latter. To establish causality, researchers must demonstrate that their studies meet three criteria.

First is temporal ordering, meaning that the IV must occur prior to the DV. It would be illogical, for

instance, to predict that adolescents’ participation in delinquency will impact their gender; conversely, it does

make sense to predict that adolescents’ gender affects the likelihood they will commit delinquent acts. The

second causality requirement is that there be an empirical relationship between the IV and the DV. This is a

basic necessity—it does not make sense to try to delve into the nuances of a nonexistent connection between

two variables. For example, if a researcher predicts that people living in high-crime areas are more likely to

own handguns for self-protection, but then finds no relationship between neighborhood-level crime rates and

handgun ownership, the study cannot proceed.

Temporal ordering: The causality requirement holding that an independent variable must precede a dependent variable.

Empirical relationship: The causality requirement holding that the independent and dependent variables possess an observed

relationship with one another.

The last requirement is that the relationship between the IV and the DV be nonspurious. This third criterion

is frequently the hardest to overcome in criminology and criminal justice research (indeed, all social sciences)

because human behavior is complicated, and each action a person engages in has multiple causes.

Disentangling these causal factors can be difficult or impossible.

Nonspuriousness: The causality requirement holding that the relationship between the independent variable and dependent variable

not be the product of a third variable that has been erroneously omitted from the analysis.

The reason spuriousness is a problem is that there could be a third variable that explains the DV as well as, or

even better than, the IV does. This third variable might partially or fully account for the relationship between

the IV and DV. The inadvertent exclusion of one or more important variables can result in erroneous

conclusions because the researcher might mistakenly believe that the IV strongly predicts the DV when, in

fact, the relationship is actually partially or entirely due to intervening factors. Another term for this problem

is omitted variable bias. When omitted variable bias (i.e., spuriousness) is present in an IV–DV relationship

but erroneously goes unrecognized, people can reach the wrong conclusion about a phenomenon. Research

Example 2.3 offers examples of the problem of omitted variables.

Omitted variable bias: An error that occurs as a result of unrecognized spuriousness and a failure to include important third variables

in an analysis, leading to incorrect conclusions about the relationship between the independent and dependent variables.

A final caution with respect to causality is that statistical analyses are examinations of aggregate trends.

Uncovering an association between an IV and a DV means only that the presence of the IV has the tendency to

be related to either an increase or a reduction in the DV in the sample as a whole—it is not an indication that

40

the IV–DV link inevitably holds true for every single person or object in the sample. For example, victims of

early childhood trauma are more likely than nonvictims to develop substance abuse disorders later in life (see

Dass-Brailsford & Myrick, 2010). Does this mean that every person who was victimized as a child has

substance abuse problems as an adult? Certainly not! Many people who suffer childhood abuse do not become

addicted to alcohol or other drugs. Early trauma is a risk factor that elevates the risk of substance abuse, but it

is not a guarantee of this outcome. Associations present in a large group are not uniformly true of all members

of that group.

Research Example 2.3 The Problem of Omitted Variables

In the 1980s and 1990s a media and political frenzy propelled the “crack baby” panic to the top of the national conversation. The

allegations were that “crack mothers” were abusing the drug while pregnant and were doing irreparable damage to their unborn

children. Stories of low-birth-weight, neurologically impaired newborns abounded. What often got overlooked, though, was the fact

that women who use crack cocaine while pregnant are also likely to use drugs such as tobacco and alcohol, which are known to harm

fetuses. These women also frequently have little or no access to prenatal nutrition and medical care. Finally, if a woman abuses crack

—or any other drug—while pregnant, she could also be at risk for mistreating her child after its birth (see Logan, 1999, for a

review). She might be socially isolated, as well, and have no support from her partner or family. There are many factors that affect

fetal and neonatal development, some under mothers’ control and some not; trying to tie children’s outcomes definitively to a single

drug consumed during mothers’ pregnancies is inherently problematic.

In the 1980s, policymakers and the public became increasingly concerned about domestic violence. This type of violence had

historically been treated as a private affair, and police tended to take a hands-off approach that left victims stranded and vulnerable.

The widely publicized results of the Minneapolis Domestic Violence Experiment suggested that arrest effectively deterred abusers,

leading to lower rates of recidivism. Even though the study’s authors said that more research was needed, states scrambled to enact

mandatory arrest laws requiring officers to make arrests in all substantiated cases of domestic violence. Subsequent experiments and

more detailed analyses of the Minneapolis data, however, called the effectiveness of arrest into question. It turns out that arrest has

no effect on some offenders and even increases recidivism among certain groups. Offenders’ employment status, in particular,

emerged as an important predictor of whether arrest deterred future offending. Additionally, the initial reduction in violence

following arrest frequently wore off over time, putting victims back at risk. Pervasive problems collecting valid, reliable data also

hampered researchers’ ability to reach trustworthy conclusions about the true impact of arrest (see Schmidt & Sherman, 1993, for a

review). The causes of domestic violence are numerous and varied, so it is unwise to assume that arrest will be uniformly

advantageous.

In sum, you should always be cautious when interpreting IV–DV relationships. It is better to think of IVs as

predictors and DVs as outcomes rather than to view them as causes and effects. As the adage goes, correlation

does not mean causation. Variables of all kinds are related to each other, but it is important not to leap

carelessly to causal conclusions on the basis of statistical associations.

41

Levels of Measurement

Every variable possesses a level of measurement. Levels of measurement are ways of classifying or describing

variable types. There are two overarching classes of variables: categorical (also sometimes called qualitative)

and continuous (also sometimes referred to as quantitative). Categorical variables comprise groups or

classifications that are represented with labels, whereas continuous variables are made of numbers that

measure how much of a particular characteristic a person or object possesses. Each of these variable types

contains two subtypes. This two-tiered classification system is diagrammed in Figure 2.1 and discussed in the

following sections.

Level of measurement: A variable’s specific type or classification. There are four types: nominal, ordinal, interval, and ratio.

Categorical variable: A variable that classifies people or objects into groups. There are two types: nominal and ordinal.

Continuous variable: A variable that numerically measures the presence of a particular characteristic. There are two types: interval

and ratio.

42

The Categorical Level of Measurement: Nominal and Ordinal Variables

Categorical variables are made up of categories. They represent ways of divvying up people and objects

according to some characteristic. Categorical variables are subdivided into two types: nominal and ordinal. The

nominal level of measurement is the most rudimentary of all the levels. It is the least descriptive and

sometimes the least informative. Race is an example of a nominal-level variable. See Tables 2.1 and 2.2 for

examples of nominal variables (see also Data Sources 2.1 for a description of the data set used in these tables).

The variable in Table 2.1 comes from a question on the survey asking respondents whether or not they

personally know a police officer assigned to their neighborhood. This variable is nominal because respondents

said “yes” or “no” in response and so can be grouped accordingly. In Table 2.2, the variable representing the

races of stopped drivers is nominal because races are groups into which people are placed. The labels offer

descriptive information about the people or objects within each category. Data are from the Police–Public

Contact Survey (PPCS).

Nominal variable: A classification that places people or objects into different groups according to a particular characteristic that

cannot be ranked in terms of quantity.

Figure 2.1 Levels of Measurement

43

Data Sources 2.1 The Police–Public Contact Survey

The Bureau of Justice Statistics (BJS; see Data Sources 2.3) conducts the Police–Public Contact Survey (PPCS) periodically as a

supplement to the National Crime Victimization Survey (NCVS; see Data Sources 1.2). Interviews are conducted in English only.

NCVS respondents aged 16 and older are asked about recent experiences they might have had with police. Variables include

respondent demographics, the reason for respondents’ most recent contact with police, whether the police used or threatened force

against the respondents, the number of officers present at the scene, whether the police asked to search respondents’ vehicles, and so

on (BJS, 2011). This data set is used by BJS statisticians to estimate the number of police–citizen contacts that take place each year

and is used by researchers to study suspect, officer, and situational characteristics of police–public contacts. The 2011 wave of the

PPCS is the most current one available at this time.

Gender is another example of a nominal variable. Table 2.3 displays the gender breakdown among people who

reported that they had sought help from the police within the past year.

Much information is missing from the nominal variables in Tables 2.1 through 2.3. For instance, the question

about knowing a local police officer does not tell us how often respondents talk to the officers they know or

whether they provide these officers with information about the area. Similarly, the race variable provides fairly

basic information. This is why the nominal level of measurement is lowest in terms of descriptiveness and

utility. These classifications represent only differences; there is no way to arrange the categories in any

meaningful rank or order. Nobody in one racial group can be said to have “more race” or “less race” than

someone in another category—they are merely of different races. The same applies to gender. Most people

identify as being either female or male, but members of one gender group do not have more or less gender

relative to members of the other group.

One property that nominal variables possess (and share with other levels) is that the categories within any

given variable are mutually exclusive and exhaustive. They are mutually exclusive because each unit in the data

set (person, place, and so on) can fall into only one category. They are exhaustive because all units have a

category that applies to them. For example, a variable measuring survey respondents’ criminal histories that

asks them if they have been arrested “0–1 time” or “1–2 times” would not be mutually exclusive because a

respondent who has been arrested once could circle both answer options. This variable would also violate the

principle of exhaustiveness because someone who has been arrested three or more times cannot circle any

available option because neither is applicable. To correct these problems, the answer options could be changed

to, for instance, “no arrests,” “1–2 arrests,” and “3 or more arrests.” Everyone filling out the survey would have

one, and only one, answer option that accurately reflected their experiences.

Mutually exclusive: A property of all levels of measurement whereby there is no overlap between the categories within a variable.

Exhaustive: A property of all levels of measurement whereby the categories or range within a variable capture all possible values.

44

Ordinal variables are one step up from nominal variables in terms of descriptiveness because they can be

ranked according to the quantity of a characteristic possessed by each person or object in a sample. University

students’ class level is an ordinal variable because freshmen, sophomores, juniors, and seniors can be rank-

ordered according to how many credits they have earned. Numbers can also be represented as ordinal

classifications when the numbers have been grouped into ranges like those in Table 2.3, where the income

categories of respondents to the General Social Survey (GSS; see Data Sources 2.2) are shown. Table 2.4

displays another variable from the PPCS. This survey question queried respondents about how often they

drive. Respondents were offered categories and selected the one that most accurately described them.

Ordinal variable: A classification that places people or objects into different groups according to a particular characteristic that can be

ranked in terms of quantity.

Ordinal variables are useful because they allow people or objects to be ranked in a meaningful order. Ordinal

variables are limited, though, by the fact that no algebraic techniques can be applied to them. This includes

ordinal variables made from numbers such as those in Table 2.3. It is impossible, for instance, to subtract

<$1,000 from $15,000–$19,999. This eliminates the ability to determine exactly how far apart two
respondents are in their income levels. The difference between someone in the $20,000–$24,999 group and
the ≥$25,000 group might only be $1 if the former makes $24,999 and the latter makes $25,000. The
difference could be enormous, however, if the person in the ≥$25,000 group has an annual family income of
$500,000 per year. There is no way to figure this out from an assortment of categories like those in Table 2.3.
The same limitation applies to the variable in Table 2.4. A glance at the table reveals general information
about the frequency of driving, but there is no way to add, subtract, multiply, or divide the categories to obtain
a specific measurement of how more or less one person in the sample drives relative to another.
45
Learning Check 2.1
In a study of people incarcerated in prison, the variable “offense type” captures the crime that each person was convicted of and
imprisoned for. This variable could be coded as either a nominal or an ordinal variable. Explain why this is. Give an example of each type
of measurement approach.
Data Sources 2.2 The General Social Survey
The National Opinion Research Center has conducted the GSS annually or every 2 years since 1972. Respondents are selected using
a multistage clustering sample design. First, cities and counties are randomly selected. Second, block groups or districts are selected
from those cities and counties. Trained researchers then canvass each block group or district on foot and interview people in person.
Interviews are offered in both English and Spanish. The GSS contains a large number of variables. Some of these variables are asked
in every wave of the survey, whereas others are asked only once. The variables include respondents’ attitudes about religion, politics,
abortion, the death penalty, gays and lesbians, persons of racial groups other than respondents’ own, free speech, marijuana
legalization, and a host of other topics (Davis & Smith, 2009). The most current wave of the GSS available at this time is the one
conducted in 2014.
Table 2.6 contains another example of an ordinal variable. This comes from the GSS and measures female
respondents’ educational attainment. You can see that this variable is categorical and ranked. Someone who
46
has attended junior college has a higher educational attainment than someone who did not complete high
school. As before, though, no math is possible. Someone whose highest education level is junior college might
be in her first semester or quarter, putting her barely above someone in the high-school-only category, or she
might be on the verge of completing her associate’s degree, which places her nearly two years above.
47
The Continuous Level of Measurement: Interval and Ratio Variables
Continuous variables differ from categorical ones in that the former are represented not by categories but
rather by numbers. Interval variables are numerical scales in which there are equal distances between all
adjacent points on those scales. Ambient temperature is a classic example of an interval variable. This scale is
measured using numbers representing degrees, and every point on the scale is exactly one degree away from
the nearest points on each side. Twenty degrees Fahrenheit, for instance, is exactly 1 degree cooler than 21
degrees and exactly 4 degrees warmer than 16 degrees.
Interval variable: A quantitative variable that numerically measures the extent to which a particular characteristic is present or absent
and does not have a true zero point.
An example of an interval-level variable is the GSS’s scoring of respondents’ occupational prestige. The GSS
uses a ranking system to assign each person a number representing how prestigious his or her occupation is.
Figure 2.2 displays the results. The scores range from 16 to 80 and so are presented as a chart rather than a
table.
As can be seen in Figure 2.2, the prestige scores are numerical. This sets them apart from the categories seen
in nominal and ordinal variables. The scores can be subtracted from one another; for instance, a respondent
with a prestige score of 47 has 9 points fewer than someone whose score is 56. Importantly, however, the
scores cannot be multiplied or divided. It does not make sense to say that someone with a score of 60 has
twice the occupational prestige as someone with a 30. This is because interval data, such as this scale, do not
have a true zero point. The numbers, although useful and informative, are ultimately arbitrary. An
occupational prestige score of 30, for instance, could have been any other number if a different coding system
had been employed.
Table 2.7 contains another interval variable. This one measures GSS respondents’ political views. The scale
ranges from 1 (extremely liberal) to 7 (extremely conservative). This variable in Table 2.7 is interval because it
is a scale (as opposed to distinctly separate categories), but lacks a zero point and therefore is an arbitrary
numbering system. The 1-to-7 numbering system could be replaced by any other 7-point consecutive
sequence without affecting the meaning of the scale. These are the hallmark identifiers of interval-level data.
Figure 2.2 Respondents’ Occupational Prestige Scores (GSS)
48
Another attitudinal variable measured at the interval level is shown in Table 2.8. This is a scale tapping into
how punitive GSS respondents feel toward people convicted of crimes. The scale ranges from the lower range
of punitiveness to the upper range. The idea behind attitudinal scores being interval (as opposed to
categorical) is that attitudes are best viewed as a continuum. There is a lot of gray area that prevents attitudes
from being accurately portrayed as ordinal; chunking attitudes up into categories introduces artificial
separation between the people falling into adjacent positions on the scale.
Ratio variables are the other subtype within the continuous level of measurement. The ratio level resembles
the interval level in that ratio, too, is numerical and has equal and known distance between adjacent points.
The difference is that ratio-level scales, unlike interval ones, have meaningful zero points that represent the
absence of a given characteristic. Temperature, for instance, is not ratio level because the zeros in the various
temperature scales are just placeholders. Zero does not signify an absence of temperature. Likewise, the data
49
presented in Figure 2.2 and Tables 2.7 and 2.8, as discussed previously, do not have meaningful zero points
and therefore cannot be multiplied or divided. You could not, for instance, say that someone who scores a 4
on the punitiveness scale is twice as punitive as a person with a 2 on this measure. Ratio-level data permit this
higher level of detail.
Ratio variable: A quantitative variable that numerically measures the extent to which a particular characteristic is present or absent
and has a true zero point.
Criminal justice and criminology researchers deal with many ratio-level variables. Age is one example.
Although it is strange to think of someone as having zero age, age can be traced close enough to zero to make
analytically reasonable to think of this variable as ratio. A 40-year-old person is twice as old as someone who is
20.
Table 2.9 displays the number of state prisoners who were executed in 2013. These data come from the BJS
(Snell, 2014; see Data Sources 2.3). The left-hand column of the table displays the number of people executed
per state, and the right-hand column shows the frequency with which each of those state-level numbers
occurred (note that the frequency column sums to 50 to represent all of the states in the country). Can you
explain why number of persons executed is a ratio-level variable?
In Table 2.10, the number of children GSS respondents report having is shown. Since the zero is meaningful
here, this qualifies as ratio level.
50
Learning Check 2.2
Zip codes are five-digit sequences that numerically identify certain locations. What is the level of measurement of zip codes? Explain your
answer.
Another example of the ratio level of measurement is offered in Figure 2.3. These data come from a question
on the PPCS asking respondents who had been the subject of a traffic or pedestrian stop in the past 12
months the number of minutes the stop lasted. Time is ratio level because, in theory, it can be reduced down
to zero, even if nobody in the sample actually reported zero as their answer.
Figure 2.3 Length of Stop, in Minutes (PPCS)
51
Data Sources 2.3 The Bureau of Justice Statistics
The BJS is the U.S. Department of Justice’s repository for statistical information on criminal justice–related topics. The BJS offers
downloadable data and periodic reports on various topics that summarize the data and present them in a user-friendly format.
Researchers, practitioners, and students all rely on the BJS for accurate, timely information about crime, victims, sentences,
prisoners, and more. Visit http://bjs.ojp.usdoj.gov/ and explore this valuable information source.
In the real world of statistical analysis, the terms interval and ratio variables are often used interchangeably. It
is the overarching categorical-versus-continuous distinction that usually matters most when it comes to
statistical analyses. When a researcher is collecting data and has a choice about level of measurement, the best
strategy is to always use the highest level possible. A continuous variable can always be made categorical later,
but a categorical variable can never be made continuous.
Level of measurement is a very important concept. It might be difficult to grasp if this is the first time you
have been exposed to this idea; however, it is imperative that you gain a firm understanding because level of
measurement determines what analyses can and cannot be conducted. This fundamental point will form an
underlying theme of this entire book, so be sure you understand it. Do not proceed with the book until you
can readily identify a given variable’s level of measurement. Table 2.11 summarizes the basic characteristics
that define each level and distinguish it from the others.
52
http://bjs.ojp.usdoj.gov/
Chapter Summary
This chapter discussed the concept of a variable. You also read about units of analysis, independent variables, dependent variables,
and about the importance of not drawing strict causal conclusions about statistical relationships. It is important to consider whether
meaningful variables have been left out and to keep in mind that empirical associations do not imply that one variable causes the
other.
This chapter also described the two overarching levels of measurement: categorical and continuous. Categorical variables are
qualitative groupings or classifications into which people or objects are placed on the basis of some characteristic. The two subtypes
of categorical variables are nominal and ordinal. These two kinds of variables are quite similar in appearance, with the distinguishing
feature being that nominal variables cannot be rank-ordered, whereas ordinal variables can be. Continuous variables are quantitative
measurements of the presence or absence of a certain characteristic in a group of people or objects. Interval and ratio variables are
both continuous. The difference between them is that ratio-level variables possess true zero points and interval-level variables do not.
You must understand this concept and be able to identify the level of measurement of any given variable because, in statistics, the
level at which a variable is measured is one of the most important determinants of the graphing and analytic techniques that can be
employed. In other words, each type of graph or statistical analysis can be used with some levels of measurement and cannot be used
with others. Using the wrong statistical procedure can produce wildly inaccurate results and conclusions. You must therefore possess
an understanding of level of measurement before leaving this chapter.
Thinking Critically
1. Suppose you have been contacted by a reporter from the local newspaper who came across data showing that men tend to be
sentenced more harshly than women (e.g., more likely to be sent to prison, given longer sentence lengths). The reporter
believes this to be a clear case of discrimination and asks you for comment. What is your response? Do you agree that gender
discrimination has been demonstrated here, or do you need more information? If the latter, what additional data would you
need before you could arrive at a conclusion?
2. Many researchers have tried to determine whether capital punishment deters murder. Suppose a new study has been
published analyzing how death-sentence rates in one year relate to murder rates the following year. The researchers who
conducted this study included only the 32 states that authorize the death penalty, and excluded the remaining states. Do you
think this is a justifiable approach to studying the possible deterrent effects of the death penalty? Would you trust the results
of the analysis and the conclusions the researchers reach on the basis of those results? Explain your answer.
Review Problems
1. A researcher wishes to test the hypothesis that low education affects crime. She gathers a sample of people aged 25 and older.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
2. A researcher wishes to test the hypothesis that arrest deters recidivism. She gathers a sample of people who have been
arrested.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
3. A researcher wishes to test the hypothesis that poverty affects violent crime. He gathers a sample of neighborhoods.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
4. A researcher wishes to test the hypothesis that prison architectural design affects the number of inmate-on-inmate assaults
that take place inside a facility. He gathers a sample of prisons.
1. What is the independent variable?
2. What is the dependent variable?
53
3. What is the unit of analysis?
5. A researcher wishes to test the hypothesis that the amount of money a country spends on education, health, and welfare
affects the level of violent crime in that country. She gathers a sample of countries.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
6. A researcher wishes to test the hypothesis that police officers’ job satisfaction affects the length of time they stay in their jobs.
He gathers a sample of police officers.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
7. A researcher wishes to test the hypothesis that the location of a police department in either a rural or an urban area affects
starting pay for entry-level police officers. She gathers a sample of police departments.
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
8. A researcher wishes to test the hypothesis that the level of urbanization in a city or town affects residents’ social cohesion.
She gathers a sample of municipal jurisdictions (cities and towns).
1. What is the independent variable?
2. What is the dependent variable?
3. What is the unit of analysis?
9. Suppose that a researcher found a statistical relationship between ice cream sales and crime—during months when a lot of ice
cream is purchased, crime rates are higher. The researcher concludes that ice cream causes crime. What has the researcher
done wrong?
10. Suppose that in a random sample of adults a researcher found a statistical relationship between parental incarceration and a
person’s own involvement in crime. Does this mean that every person who had a parent in prison committed crime? Explain
your answer.
11. Identify the level of measurement of each of the following variables:
1. Suspects’ race measured as white, black, Latino, and other
2. The age at which an offender was arrested for the first time
3. The sentences received by convicted defendants, measured as jail, prison, probation, fine, and other
4. The total number of status offenses that adult offenders reported having committed as juveniles
5. The amount of money, in dollars, that a police department collects annually from drug asset forfeitures
6. Prison security level, measured as minimum, medium, and maximum
7. Trial judges’ gender
12. Identify the level of measurement of each of the following variables:
1. The amount of resistance a suspect displays toward the police, measured as not resistant, somewhat resistant, or very
resistant
2. The number of times someone has shoplifted in her or his life
3. The number of times someone has shoplifted, measured as 0–2, 3–5, or 6 or more
4. The type of attorney a criminal defendant has at trial, measured as privately retained or publicly funded
5. In a sample of juvenile delinquents, whether or not those juveniles have substance abuse disorders
6. Prosecutors’ charging decisions, measured as filed charges and did not file charges
7. In a sample of offenders sentenced to prison, the number of days in their sentences
13. If a researcher is conducting a survey and wants to ask respondents about their self-reported involvement in shoplifting, there
are a few different ways he could phrase this question.
1. Identify the level of measurement that each type of phrasing shown below would produce.
2. Explain which of the three possible phrasings would be the best one to choose and why this is.
Possible phrasing 1: How many times have you taken small items from stores without paying for those items?
Please write in: ______
Possible phrasing 2: How many times have you taken small items from stores without paying for those items?
Please circle one of the following:
54
Never 1–2 times 3–4 times 5+ times
Possible phrasing 3: Have you ever taken small items from stores without paying for those items? Please circle
one of the following:
Yes No
14. If a researcher is conducting a survey and wants to ask respondents about the number of times each of them has been
arrested, there are a few different ways she could phrase this question. Write down the three possible phrasing methods.
15. The following table contains BJS data on the number of prisoners under sentence of death, by region. Use the table to do the
following
1. Identify the level of measurement of the variable region.
2. Identify the level of measurement of number of prisoners under sentence of death.
16. The following table contains data from the 2012 National Crime Victimization Survey showing the number of victimization
incidents and whether or not those crimes were reported to the police. The data are broken down by victims’ household
income level. Use the table to do the following:
1. Identify the level of measurement of the variable income.
2. Identify the level of measurement of the variable victimization reported.
17. Haynes (2011) conducted an analysis to determine whether victim advocacy affects offender sentencing. She gathered a
sample of courts and measured victim advocacy as a yes/no variable indicating whether or not there was a victim witness
office located inside each courthouse. She measured sentencing as the number of months of incarceration imposed on
convicted offenders in the courts.
1. Identify the independent variable in this study.
2. Identify the level of measurement of the independent variable.
3. Identify the dependent variable in this study.
4. Identify the level of measurement of the dependent variable.
5. Identify the unit of analysis.
18. Bouffard and Piquero (2010) wanted to know whether arrested suspects’ perceptions of the way police treated them during
the encounter affected the likelihood that those suspects would commit more crimes in the future. Their sample consisted of
males who had been arrested at least once during their lives. They measured suspects’ perceptions of police behavior as fair or
55
unfair. They measured recidivism as the number of times suspects came into contact with police after that initial arrest.
1. Identify the independent variable in this study.
2. Identify the level of measurement of the independent variable.
3. Identify the dependent variable in this study.
4. Identify the level of measurement of the dependent variable.
5. Identify the unit of analysis.
19. Kleck and Kovandzic (2009; see Research Example 2.2) examined whether the level of homicide in a particular city affected
the likelihood that people in that city would own firearms. They measured homicide as the number of homicides that took
place in the city in 1 year divided by the total city population. They measured handgun ownership as whether survey
respondents said they did or did not own a gun.
1. Identify the independent variable used in this study.
2. Identify the level of measurement of the independent variable.
3. Identify the dependent variable in this study.
4. Identify the level of measurement of the dependent variable.
5. Identify the unit of analysis. (Hint: This study has two!)
20. Gau et al. (2010; see Research Example 2.1) examined whether suspects’ race or ethnicity influenced the likelihood that
police would brandish or deploy Tasers against them. They measured race as white, Hispanic, black, or other. They measured
Taser usage as Taser used or some other type of force used.
1. Identify the independent variable used in this study.
2. Identify the level of measurement of the independent variable.
3. Identify the dependent variable in this study.
4. Identify the level of measurement of the dependent variable.
5. Identify the unit of analysis.
56
Key Terms
Variable 15
Constant 15
Unit of analysis 15
Empirical 18
Dependent variable 18
Independent variable 18
Temporal ordering 19
Empirical relationship 19
Nonspuriousness 19
Omitted variable bias 19
Level of measurement 21
Categorical variable 21
Continuous variable 21
Nominal variable 21
Mutually exclusive 23
Exhaustive 23
Ordinal variable 23
Interval variable 26
Ratio variable 28
57
Chapter 3 Organizing, Displaying, and Presenting Data
58
Learning Objectives
Define univariate and bivariate.
Identify the data displays available for each of the four levels of measurement.
Identify the computations available for univariate displays, and be able to calculate each one.
Construct univariate and bivariate numerical and graphical displays.
Create numerical displays, charts and graphs, and rate variables in SPSS.
Data are usually stored in electronic files for use with software programs designed to conduct statistical
analyses. The program SPSS is one of the most common data software programs in criminal justice and
criminology. The SPSS layout is a data spreadsheet. You might be familiar with Microsoft Excel; if so, then
the SPSS data window should look familiar. Figure 3.1 depicts what a typical SPSS file might look like.
The data in Figure 3.1 are from the Bureau of Justice 2011 Police–Public Contact Survey (PPCS; see Data
Sources 2.1). Each row (horizontal line) in the grid represents one respondent, and each column (vertical line)
represents one variable. Where any given row and column meet is a cell containing a given person’s response
to a particular question.
Cell: The place in a table or spreadsheet where a row and a column meet.
Your thoughts as you gaze on the data screen in Figure 3.1 can probably be aptly summarized as “Huh?” That
is a very appropriate response, because there is no way to make sense of the data when they are in this raw
format. This brings us to the topic of this chapter: methods for organizing, displaying, and presenting data.
You can see from the figure that something has to be done to the data set to get it into a useful format. This
chapter will teach you how to do just that.
Chapter 2 introduced levels of measurement (nominal, ordinal, interval, and ratio). A variable’s level of
measurement determines which graphs or charts are and are not appropriate for that variable. Various data
displays exist, and many of them can be used only with variables of particular types. As you read this chapter,
take notes on two main concepts: (1) the proper construction of each type of data display and (2) the level of
measurement for which each display type is applicable.
59
Data Distributions
60
Univariate Displays: Frequencies, Proportions, and Percentages
Perhaps the most straightforward type of pictorial display is the univariate (one variable) frequency
distribution. A frequency is simply a raw count; it is the number of times a particular characteristic appears in
a data set. A frequency distribution is a tabular display of frequencies. Table 3.1 shows the frequency
distribution for the variable respondent gender in the 2011 PPCS. Let us pause and consider two new symbols.
The first is the f that sits atop the right-hand column in Table 3.1. This stands for frequency. There are 25,078
males in this sample. An alternative way of phrasing this is that the characteristic male occurs 25,078 times. A
more formal way to write this would be fmale = 25,078; for females, ffemale = 27,451. The second new symbol is
the N found in the bottom right-hand cell. This represents the total sample size; here, fmale + ffemale = N.
Numerically, 25,078 + 27,451 = 52,529.
Univariate: Involving one variable.
Frequency: A raw count of the number of times a particular characteristic appears in a data set.
Figure 3.1 SPSS Data File
Raw frequencies are of limited use in graphical displays because it is often difficult to interpret them and they
do not offer much information about the variable being examined. What is needed is a way to standardize the
numbers to enhance interpretability. Proportions do this. Proportions are defined as the number of times a
particular characteristic appears in a sample relative to the total sample size. Formulaically,
where
61
p = proportion
f = raw frequency
N = total sample size
Proportion: A standardized form of a frequency that ranges from 0.00 to 1.00.
Proportions range from 0.00 to 1.00. A proportion of exactly 0.00 indicates a complete absence of a given
characteristic. If there were no males in the PPCS, their proportion would be
Conversely, a trait with a proportion of 1.00 would be the only characteristic present in the sample. If the
PPCS contained only men, then the proportion of the sample that was male would be
Another useful technique is to convert frequencies into percentages (abbreviated pct). Percentages are a
variation on proportions and convey the same information, but percentages offer the advantage of being more
readily interpretable by the public. Percentages are computed similarly to proportions, with the added step of
multiplying by 100:
Percentage: A standardized form of a frequency that ranges from 0.00 to 100.00.
Proportions and percentages can be used in conjunction with frequencies to form a fuller, more informative
display like that in Table 3.2. Note that the two proportions (pmales and pfemales) sum to 1.00 because the two
categories contain all respondents in the sample; the percentage column sums to 100.00 for the same reason.
The “Σ” symbol in the table is the Greek letter sigma and is a summation sign. It instructs you to add up
everything that is to the right of the symbol. The number to the right of the equal sign in “Σ =” is the summed
total. As long as all cases in a sample have been counted once and only once, proportions will sum to 1.00 and
percentages to 100.00 or within rounding error of these totals. Significant deviation from 1.00 or 100.00 that
is greater than what could be explained by rounding error suggests a counting or calculation error.
Another useful technique for frequency distributions is the computation of cumulative measures. Cumulative
frequencies, cumulative proportions, and cumulative percentages can facilitate meaningful interpretation of
distributions, especially when data are continuous. Consider Table 3.3, which contains data from identity
theft victims who discovered that someone had filed a fraudulent tax return in their name. The variable in the
table reflects the dollar amount victims personally lost from the crime. The data are part of the larger National
Crime Victimization Survey (NCVS; see Data Sources 1.2) and come from a 2014 supplement tapping into
62
respondents’ experiences with various forms of identity theft. The dollar amounts reported by victims and the
frequency of each response are located in the two left-most columns, respectively. To their right is a column
labeled cf, which stands for cumulative frequency. The cp and cpct columns contain cumulative proportions and
percentages, respectively.
63
Learning Check 3.1
If you were summing proportions or percentages across the different levels of a categorical variable and you arrived at a result that was
greater than the maximum of 1.00 or 100.00, respectively, what potential errors might have occurred in the calculations? What about if
the result were less than 1.00 or 100.00?
Cumulative columns are constructed by summing the f, p, and pct columns successively from row to row. The
arrows in the table are intended to help you visualize this process. In the second row of Table 3.3’s cf column,
for instance, 48 is the sum of 45 and 3. In the cp column, likewise, .87 is the sum of .82 + .05. Cumulatives
allow assessments of whether the data are clustered at one end of the scale or spread fairly equally throughout.
In Table 3.3, it can be readily concluded that the data cluster at the low end of the scale, because .82 (or
81.82%) of victims did not incur any personal financial costs, and .95 (or 94.55%) spent $70 or less. This
suggests that whatever profits offenders enjoy from income-tax fraud, victims themselves suffer minimal out-
of-pocket expense.
Cumulative: A frequency, proportion, or percentage obtained by adding a given number to all numbers below it.
64
Learning Check 3.2
What is the level of measurement of the variable state? What is the level of measurement of the variable property crime rate? Refer to
Chapter 2 if needed.
65
Univariate Displays: Rates
Suppose someone informed you that 1,579,527 burglaries were reported to the police in 2015. What would
you make of this number? Nothing, probably, because raw numbers of this sort are simply not very useful.
They lack a vital component—a denominator. The question that would leap to your mind immediately is
“1,579,527 out of what?” You would want to know if this number was derived from a single city, from a single
state or from the United States as a whole. This is where rates come in. A rate is a method of standardization
that involves dividing the number of events of interest (e.g., burglaries) by the total population:
Table 3.4 contains data from the 2015 Uniform Crime Report (UCR; see Data Sources 1.1). The column
titled Rate displays the rate per capita that is obtained by employing Formula 3(3).
Note: 2012 U.S. Population = 321,418,820
Note how tiny the numbers in the rate column are. Rates per capita do not make sense in the context of low-
frequency events like crime because they end up being so small. It is, therefore, customary to multiply rates by
a certain factor. This factor is usually 1,000, 10,000, or 100,000. You should select the multiplier that makes
the most sense with the data at hand. In Table 3.4, the 10,000 multiplier has been used to form the Rate per
10,000 column. Multiplying in this fashion lends clarity to rates because now it is no longer the number of
crimes per person but, rather, the number of crimes per 10,000 people. If you randomly selected a sample of
10,000 people from the population, you would expect 49.14 of them to have been the victim of burglary in the
past year and 23.78 to have experienced aggravated assault. These numbers and their interpretation are more
real and more tangible than those derived using Formula 3(3) without a multiplier.
Sometimes rates must be calculated using multiple denominators, in contrast to Table 3.4 where there was
only one. Table 3.5 shows a sample of states and the number of property crimes they reported in 2015.
States have different populations, so rates have to be calculated using each state’s unique denominator. Table
3.6 expands upon the counts in Table 3.5 by including the property-crime rates per 10,000.
66
67
Learning Check 3.3
Rates and percentages have similar computational steps but very different meanings. Explain the differences between them, including the
additional information needed to calculate a rate that is not needed for a percentage, the reason rates do not sum to 100 the way
percentages do, and the substantive meaning of each one (i.e., the information that is provided by each type of number).
68
Bivariate Displays: Contingency Tables
Researchers are often interested not just in the frequency distribution of a single variable (a univariate display)
but, rather, in the overlap between two variables. The Census of Jails (COJ; see Data Sources 3.1) collects
data from all jails in the United States that hold inmates past arraignment (i.e., excluding lockups like
stationhouse holding cells where inmates are confined for only a brief period before being transferred
elsewhere).
Data Sources 3.1 The Census of Jails
The BJS has conducted the COJ periodically since 1970. Every 5 years, the BJS sends surveys to all jails operated by federal, state,
and local governments, as well as by private corporations. The surveys capture institution-level data such as the total inmate
population, the number of correctional staff, the number of inmate assaults against staff, whether the facility offers services
(vocational, educational, or mental-health), and so on. The 2013 wave of the COJ is the most recent version available, but the 2006
version of the survey contains information on jail facilities that is not found in the 2013 version, so the examples in this book draw
from both data sets.
Researchers studying jails might be interested in knowing whether facilities of various sizes differ in terms of
whether they provide inmates with the opportunity to take secondary education courses that advance them
toward the acquisition of a GED certificate. One might predict, for example, that larger jails will be busier
and more crowded and that institutional security will therefore take precedence over the provision of services
like GED courses. A contingency table (also sometimes called crosstabs) allows us to see the overlap between
these two variables. Table 3.7 contains this display. This is a bivariate analysis, meaning it contains two
variables. You can see from Table 3.7 that 155 small jails, 315 medium-sized ones, and 518 large ones provide
inmates the opportunity to take GED classes while incarcerated.
Contingency table: A table showing the overlap between two variables.
Bivariate: An analysis containing two variables. Usually, one is designated the independent variable and the other the dependent
variable.
Raw frequencies such as those shown in Table 3.7 offer a basic picture of bivariate overlap, but are not as
informative as they could be. It is not immediately apparent from Table 3.7 whether the provision of GED
courses differs across small, medium, and large facilities, since each of these groups is a different size (i.e.,
there are 778 small jails, 786 medium ones, and 807 large ones). To organize the data into a more readily
interpretable format, proportions—or, more commonly, percentages—can be computed and entered into the
contingency table in place of frequencies.
There are two types of proportions and percentages that can be computed in a bivariate contingency table: row
and column. Row proportions and percentages are computed using the row marginals in the denominator,
whereas column proportions and percentages employ the column marginals. There is no rule about which
variable to place in the rows and which one in the columns, or whether you should compute row or column
marginal (or both). These decisions should always be made based on the variables at hand and the point the
researcher is trying to make with the data. Here, we want to discover the percentage of each facility type that
69
offers GED courses. It is common (though, again, not required) to place the independent variable in the rows
and the dependent variable in the columns. Since we are predicting that facility size influences GED
offerings, size is the IV and GED course availability is the DV. Hence, we will place size in the rows and
calculate row percentages. Table 3.8 shows the percentage distribution.
Row proportions and percentages: In a contingency (crosstabs) table, the proportions and percentages that are calculated using row
marginals as the denominators. Each row sums to 1.00 and 100.00, respectively.
Column proportions and percentages: In a contingency (crosstabs) table, the proportions and percentages that are calculated using
column marginals as the denominators. Each column sums to 1.00 and 100.00, respectively.
Interesting results emerge from the row percentages in Table 3.8. It appears that small jails are the least likely
to offer GED classes (19.92%) and that large ones are the most likely (64.19%), with medium-sized facilities
falling in the middle (40.08%). The prediction that larger jails would be too preoccupied with security matters
to concern themselves with providing educational services is not empirically supported. Perhaps the existence
of a GED program inside a jail is more dependent on available resources, and smaller jails might lack the
money and personnel needed to offer this type of benefit to inmates.
70
Learning Check 3.4
Table 3.8 displays row percentages (i.e., percentages computed on the basis of row marginal). This shows the percentage of each type of
facility that offers GED classes. How would column percentages be interpreted in this instance? As practice, calculate the column
percentages for Table 3.8 and explain their meaning.
Research Example 3.1 Does Sexual-Assault Victimization Differ Between Female and Male Jail Inmates?
Research has shown that criminal offenders experience victimization at higher rates than the general population. This victimization
might have predated the beginning of their criminal involvement, or it might have occurred because of the risky lifestyles that many
offenders lead. Female offenders, in particular, experience high levels of sexual abuse and assault. Lane and Fox (2013) gathered data
on a sample of jail inmates and asked them about their victimization histories and their fear of future victimization. (The numbers in
the table are averages, with higher scores indicating greater levels of fear.) The following table displays the results, broken down by
gender.
More than half of female inmates (51%) reported having been sexually assaulted; this number was 7% for men. Women were also
more worried about future victimization of all types, although their scores on the fear-of-victimization variables showed that their
fear of sexual assault outweighed their fear of other crimes. It has been argued in the past that sexual assault takes a unique physical
and psychological toll on women—even those who have never actually experienced it—because it is an ever-present threat. Lane and
Fox’s results confirm the poignant impact that sexual assault has on female offenders.
71
Do Victim Impact Statements Influence Jurors’ Likelihood of
Sentencing Murder Defendants to Death?
Victim impact statements (VIS) are controversial. On the one hand, victims’ rights advocates claim that speaking out in court is
cathartic for victims and ensures they play a role in the process of punishing defendants who have hurt them or their loved ones.
Critics, however, fear that these statements bias juries by injecting emotional pleas into what should be a neutral, detached review of
the evidence presented during the sentencing phase of trial. Nuñez, Myers, Wilkowski, and Schweitzer (2017) sought a greater
understanding of the effects of VIS on juror decision making in capital sentencing trials. In particular, they wanted to find out
whether VIS that were sad in tone differed from those couched in anger. Nuñez and colleagues gathered a sample of people eligible
to serve on juries and randomly assigned them to view one of six videotaped mock trials. In two trials, the murder victim’s wife
(played by an actor) read impact statements infused with either anger or sadness. The third trial contained no VIS. Each VIS
condition was repeated across trial scenarios in which the mitigating evidence favoring the defendant was either weak or strong. The
researchers then asked the participants whether they would sentence the defendant to life or to death. The table shows the
sentencing decisions made by mock jurors in each of the three VIS conditions, broken down by whether the mitigating evidence in
the case was weak or strong.
The table reveals several interesting findings, foremost of which is that angry VIS do indeed seem to appeal to jurors’ emotions and
lead them to hand down death sentences more often; this conclusion flows from the fact that the participants in the “angry VIS”
condition voted for death more often than those in either other VIS condition. The impact of an angry VIS is especially pronounced
when mitigating evidence is weak. The effect of sad VIS is less pronounced; the presence of a sad victim does not substantially alter
sentencing outcomes compared to having no victim present during sentencing. This study suggests that jurors making capital
sentencing decisions are emotionally swayed by angry victims and might rest their sentencing decisions partially upon this stirring up
of anger.
72
Graphs and Charts
Frequency, proportion, and percentage distributions are helpful ways of summarizing data; however, they are
rather dull to look at. It is sometimes desirable to arrange data in a more attractive format. If you were giving a
presentation to a local police department or district attorney’s office, for instance, you would not want to
throw numbers at your audience for 45 minutes. The monotony would be boring. Additionally, discerning a
table’s meaning typically requires close examination, which taxes an audience’s ability to follow along with
ease. Presentations can be diversified by the introduction of charts and graphs, of which there are many
different types. Charts and graphs inject variety, color, and interest value into written and oral presentations of
quantitative results. This chapter concentrates on five of the most common: pie charts, bar graphs, histograms,
frequency polygons, and line graphs. Some of these charts and graphs are limited to specific levels of
measurement, while others are useful with multiple types of data.
73
Categorical Variables: Pie Charts
Pie charts can be used only with categorical data, and they are most appropriate for variables that have
relatively few classes (i.e., categories or groups) because pie charts get messy fast. A good general rule is to use
a pie chart only when a variable contains five or fewer classes. Pie charts are based on percentages: The entire
circle represents 100%, and the slices are sized according to the size of their contribution to that total.
Classes: The categories or groups within a nominal or ordinal variable.
The variable that will be used here to illustrate a pie chart is the race and ethnicity of stopped drivers from
Table 2.2 (see page 22). The first step is to transform the raw frequencies into percentages using Formula
3(2). Once percentages have been computed, the pie chart can be built by dividing 100% into its constituent
parts. Figure 3.2 contains the pie chart. Flip back to Table 2.2 and compare this pie chart to the raw frequency
distribution to note the dramatic difference between the two presentation methods.
Figure 3.2 Race of Stopped Drivers (Percentages)
A pie chart can also be used to display data from the 2013 Law Enforcement Management and
Administrative Statistics (LEMAS; see Data Sources 3.2) survey. This survey captures data on the types of
police agencies nationwide (i.e., their primary jurisdiction). Figure 3.3 contains the percentages of agencies
that are operated at the municipal (city or town police), county (sheriffs’ offices), state, and tribal level. As you
can see, the vast majority of police agencies are municipally operated.
Figure 3.3 Police Agency Type (Percentages)
74
75
Learning Check 3.5
Pie charts like the one in Figures 3.2 and 3.3 are made of percentages. Rates cannot be used for pie charts. Why is this true?
Data Sources 3.2 The Law Enforcement Management and Administrative Statistics Survey
The BJS conducts the Law Enforcement Management and Administrative Statistics (LEMAS) survey every 3 to 4 years. The
sampling design involves two stages. First, all police agencies with 100 or more sworn personnel are included. Second, the BJS pulls
a random sample of agencies with fewer than 100 officers. The agencies in the sample are sent surveys to fill out. The surveys capture
agency-level data such as the number of sworn law-enforcement personnel an organization employs, the number of civilian
personnel, whether the agency participates in community policing, whether the agency has specialized units, and so on. At this time,
the 2013 LEMAS survey is the most recent wave available.
76
Categorical Variables: Bar Graphs
Like pie charts, bar graphs are meant to be used with categorical data; unlike pie charts, though, bar graphs
can accommodate variables with many classes without damage to the charts’ readability. Bar graphs are thus
more flexible than pie charts. For variables with five or fewer classes, pie charts and bar graphs might be
equally appropriate; when there are six or more classes, bar graphs should be used.
In Chapter 1, you learned that one of the reasons for the discrepancy between crime prevalence as reported by
the UCR and the NCVS is that a substantial portion of crime victims do not report the incident to police.
Truman and Morgan (2016) analyzed NCVS data and reported the percentage of people victimized by
different crime types who reported their victimization to police. Figure 3.4 contains a bar graph illustrating
the percentage of victims who contacted the police to report the crime. Bar graphs provide ease of
visualization and interpretation. It is simple to see from Figure 3.4 that substantial portions of all types of
victimizations are not reported to the police and that motor vehicle theft is the most reliably reported crime.
Figure 3.4 Percentage of Victimizations Reported to Police in 2015
Rates can also be presented as bar graphs. Figure 3.5 is a bar graph of the rates in Table 3.6.
Figure 3.5 2015 Property Crime Rates (per 10,000) in Four States
77
A useful feature of bar graphs is that they can also be used to show the overlap between two variables. Figure
3.6 draws from LEMAS and the agency-type variable used in Figure 3.3. In the figure, the bars represent the
number of agencies within each category that do and do not allow members of the public to report crimes
through email or text message. Many agencies are adopting these electronic methods in hopes of encouraging
people to report crimes by making it easier and more convenient for them to do so.
The frequencies in Figure 3.6 can also be turned into percentages and graphed like Figure 3.7. Now, instead
of representing frequencies, the bars mark the percentage of agencies of each type that allow (or do not allow)
people to report crimes through text or email.
78
Continuous Variables: Histograms
Histograms are for use with continuous data. Histograms resemble bar charts, with the exception that in
histograms, the bars touch one another. In bar charts, the separation of the bars signals that each category is
distinct from the others; in histograms, the absence of space symbolizes the underlying continuous nature of
the data. Figure 3.8 contains a histogram showing the ages of Hispanic respondents to the 2011 PPCS (see
Data Sources 2.1) who reported that they had called the police for help within the past 24 months.
Figure 3.6 Police Agencies’ use of Email and Text for Crime Reporting, by Agency Type
Figure 3.7 Police Agencies’ use of Email and Text for Crime Reporting, by Agency Type
Figure 3.8 Age of Hispanic Respondents Who Called the Police in the Past 24 Months
79
Research Example 3.2 Are Women’s Violent-Crime Commission Rates Rising?
Women have historically experienced a very small arrest rate, much smaller than men’s rate. Over the past few years, arrest rates for
women have risen dramatically, even as crime itself has been declining or remaining stable. Some observers believe that women
commit more crime today than they did in the past. Critics, however, argue that the change in the arrest rate is not caused by actual
increases in women’s criminality; instead, they say, it is because tough-on-crime policies have eroded the leniency that women used
to receive from police. Who is right? Has women’s violent-crime involvement truly risen? Or are women just more likely to be
arrested today? Steffensmeier, Zhong, Ackerman, Schwartz, and Agha (2006) set out to answer these questions. The researchers
used two data sources. First, they relied on the UCR to track the arrest rates for women and men across different types of violent
crime from 1980 through 2003. Second, they used the NCVS to measure violent victimizations perpetrated by women and men
during this same period. The authors knew that the UCR would show a rise in women’s arrest rates, at least for certain kinds of
offenses. The real question was whether the NCVS would also show a rise. The authors displayed their results in histograms.
The top histogram displays UCR arrest data and the bottom one shows NCVS victimization data. The shorter bars in each one
represent the percentage of arrestees and offenders, respectively, who were female. Together, these graphs show that even though
female arrests for assault have risen, their participation in criminal assaults has not; assaults perpetrated by women have remained
constant. The authors concluded that women are not committing violent crimes at higher rates now than in the past; their increasing
arrest rates are caused by criminal-justice policies that lead police to arrest women in situations where, a few decades ago, they would
have shown them leniency.
Note: Includes aggravated and simple assaults.
80
Note: Includes aggravated and simple assaults.
Histograms can also feature percentages instead of frequencies. For example, we can use a histogram to
display the average daily population of small jails using data from the COJ facilities (Data Sources 3.1). Figure
3.9 shows the results. The horizontal axis lists the numbers of inmates each jail reported containing, and the
vertical axis represents the percent of all facilities that housed each quantity. A glance at the histogram allows
us to see that while small jails are fairly diverse in size, they are somewhat clustered toward the smaller end.
Figure 3.9 Number of Inmates in Small Jails
81
Continuous Variables: Frequency Polygons
Frequency polygons are an alternative to histograms. There is no right or wrong choice when it comes to
deciding whether to use a histogram or a frequency polygon with a particular continuous variable; the best
strategy is to mix it up a bit so that you are not using the same chart type repeatedly. Figure 3.10 contains the
frequency polygon for data similar to that used in Figure 3.8 except this time it includes the ages of non-
Hispanic respondents instead of Hispanic respondents. Frequency polygons are created by placing a dot in the
places where the tops of the bars would be in a histogram and then connecting those dots with a line.
Another type of continuous variable that could be graphed using a frequency polygon is the number of female
sergeants in police departments and sheriffs’ offices. Nationwide, police agencies are seeking not merely to
recruit and retain female officers but also to promote them to ranks with supervisory and management
responsibilities. Figure 3.11 shows the number of female sergeants in sheriffs’ offices, according to the
LEMAS survey (Data Sources 3.2).
Figure 3.10 Age of Non-Hispanic Respondents Who Called the Police in the Past 24 Months
Figure 3.11 Number of Female Sergeants in Sheriffs’ Offices
82
Longitudinal Variables: Line Charts
People who work with criminal justice and criminology data often encounter longitudinal variables.
Longitudinal variables are measured repeatedly over time. Crime rates are often presented longitudinally as a
means of determining trends. Line graphs can make it easy to discern trends.
Longitudinal variables: Variables measured repeatedly over time.
Trends: Patterns that indicate whether something is increasing, decreasing, or staying the same over time.
Figure 3.12 shows a line graph of data from the Uniform Crime Reports measuring the annual number of
hate-crime incidents from 1996 to 2015. Figure 3.13 shows the annual percentage of all hate crimes that are
motivated by sexual-orientation bias. Together, these two line charts show two trends. First, total hate-crime
incidents have declined slightly in recent years. Second, despite the downward trend in total incidents, the
percentage of incidents that are based on the victims’ sexual orientation has risen steadily, albeit with a small
drop from 2011 to 2012.
Figure 3.12 Annual Number of Hate-Crime Incidents, 1996–2015
Line charts can employ percentages, as well. We might ask not merely the number of hate-crime incidents
that occur annually, but the percentage of these incidents motivated by sexual orientation bias. Figure 3.13
shows this graph.
Together, Figures 3.12 and 3.13 reveal that although hate crimes have declined over the past 20 years, there
has been a slight rise in the percentage of these crimes arising from prejudice against members of the gay and
lesbian individuals.
Figure 3.13 Percentage of Hate-Crime Incidents Motivated by Sexual Orientation Bias, 1996–2015
83
84
Grouped Data
The overarching purpose of a frequency distribution, chart, or graph is to display data in an accessible, readily
understandable format. Sometimes, though, continuous variables do not lend themselves to tidy displays.
Consider Table 3.9’s frequency distribution for the amount of money, in dollars per person, that local
governments in each state spent on criminal justice operations (Morgan, Morgan, & Boba, 2010; Data
Sources 3.3). Figure 3.14 displays a histogram of the data. You can see that neither the frequency distribution
nor the histogram is useful; there are too many values, and most of the values occur only once in the data set.
There is no way to discern patterns or draw any meaningful conclusion from these data displays.
Data Sources 3.3 CQ Press’s State Factfinder Series
The Factfinder Series’ Crime Rankings are compilations of various crime and crime-related statistics from the state and local levels.
These volumes are comprehensive reports containing data derived from the Federal Bureau of Investigation (FBI), BJS, U.S. Census
Bureau, and U.S. Drug Enforcement Administration. The data used here come from Morgan et al. (2010).
Figure 3.14 Per Capita Local Government Expenditures on Criminal Justice, Ungrouped
85
Grouping the data can provide a solution to this problem by transforming a continuous variable (either
interval or ratio) into an ordinal one. There are several steps to grouping. First, find the range in the data by
subtracting the smallest number from the highest. Second, select the number of intervals you want to use.
This step is more art than science; it might take you a bit of trial and error to determine the number of
intervals that is best for your data.
The ultimate goal is to find a middle ground between having too few and too many intervals—too few can
leave your data display flat and uninformative, whereas too many will defeat the whole purpose of grouping.
Third, determine interval width by dividing the range by the number of intervals. This step is probably best
illustrated in formulaic terms:
This will often produce a number with a decimal, so round up or down depending on your reasoned judgment
as to the optimum interval width for your data. Fourth, construct the stated class limits by starting with the
smallest number in the data set and creating intervals of the width determined in Step 3 until you run out of
numbers. Finally, make a new frequency (f) column by counting the number of people or objects within each
stated class interval.
Let us group the legal expenditure data in Table 3.9. First, we need the range:
Range = 623 – 123 = 500
Now we choose the number of intervals we want to use. With a range as large as 500, it is advisable to select
relatively few intervals so that each interval will encompass enough raw values to make it meaningful. We will
start with 10 intervals. The next step is to compute the interval width. Using the formula from above,
Each interval will contain 50 raw scores. Now the stated class limits can be constructed. Take a look at the
left-hand column in Table 3.10. There are three main points to keep in mind when building stated class
limits. The stated limits must be inclusive—in this example, the first interval contains the number 123, the
86
number 172, and everything in between. They also must be mutually exclusive and exhaustive. Once the stated
class limits have been determined, the frequency for each interval is calculated by summing the number of raw
data points that fall into each stated class interval. The sum of the frequency column in a grouped distribution
should equal the sum of the frequencies in the ungrouped distribution.
You can see that Table 3.10 is much neater and more concise than Table 3.9. It is more condensed and easier
to read. Where you will really see the difference, though, is in the histogram. Compare Figure 3.15 to Figure
3.14. Quite an improvement! It has a real shape now. This demonstrates the utility of data grouping.
Figure 3.15 Per Capita Local Government Expenditures on Criminal Justice, Grouped
87
Learning Check 3.6
The data displayed in Figure 3.14 clearly need to be grouped, for the reasons described in the text. Refer back to the age variables shown
in Figures 3.8 and 3.10. Do you think these variables should be grouped for ease of interpretation? Explain your answer.
Another variable that could benefit from grouping is the daily population of small jails. Although the
histogram in Figure 3.9 is not too bad to look at, it is rather busy. Grouping could clean the distribution up
and give it a more streamlined appearance. This would also be helpful if you were writing a report and had
space constraints.
We will group the data in Table 3.11 using the three-step process described earlier. First, the range is 36 – 0 =
36. Second, with a range of 36, let us try 8 intervals. As noted, there are no specific rules guiding the selection
of intervals. Eight seems like a good number here because it will create enough intervals so that the
distribution does not lose its shape (a danger when there are too few intervals in the grouped distribution) but
will significantly reduce the complexity of the ungrouped data and make for a clean grouped histogram. Third,
the width is which we will round to 5. Each interval will contain five numbers.
The stated class limits will start with zero and continue until the highest number (here, 36) has been included
88
in an interval. The left column in Table 3.12 shows the limits. To calculate the frequency for each group, we
sum the frequencies for each of the numbers in each one. For example, the frequency for the “0 – 4” group is 4
+ 51 + 42 + 44 + 38 = 179. The frequency for the “5 – 9” group is 29 + 40 + 32 + 27 + 14 = 142. We complete
this process until all the ungrouped data have been shrunk down into the stated class limits.
After finishing the frequency table (Table 3.12), we can display the grouped data in a histogram. Refer to
Figure 3.16. A comparison between Figures 3.16 and 3.9 reveals the benefits of grouping. The bars in Figure
3.9 jump around a lot, but the bars in Figure 3.16 show a distinct and constant decline that quickly conveys
information about inmate populations in small jails. This figure would allow an audience to easily see that
small jails tend to hold only a few inmates.
Figure 3.16 Number of Inmates Housed in Small Jails, Grouped
89
SPSS
This is our first encounter with SPSS. There are a few preliminary things you should know before you start
working with data. First, GIGO alert! Recall that when garbage goes in, the output is also garbage. Using the
wrong statistical technique will produce unreliable and potentially misleading results. Statistical software
programs generally do not alert you to errors of this sort; they will give you output even if that output is wrong
and useless. It is your responsibility to ensure that you are using the program correctly. Second, pay attention
to the SPSS file extension. The .sav extension signifies an SPSS data file: Anytime you see this extension, you
know the file contains data in SPSS format. SPSS is the only program that will open a file with the .sav
extension, so make sure you are working on a computer equipped with SPSS.
To obtain a frequency distribution, click on Analyze → Descriptive Statistics → Frequencies, as shown in Figure
3.17. Select the variable you want from the list on the left side and either drag it to the right or click the arrow
to move it over. For this illustration, we will use the LEMAS data displayed in the bar chart in Figure 3.3.
This variable captures the number (and corresponding percentage) of each type of policing agency in the
United States. Following the steps depicted in Figure 3.17 will produce the table shown in Figure 3.18.
The SPSS program can also be used to produce graphs and charts. The Chart Builder (accessible from the
Graphs drop-down menu) allows you to select a chart type and then choose the variable you want to use. The
SPSS Chart Builder requires that the level of measurement for each variable be set properly. SPSS will not
permit certain charts to be used with some levels of measurement. Before constructing graphs or charts, visit
the Measure column in the Variable View and make sure that continuous variables are marked as Scale and
that nominal and ordinal variables are designated as such. To begin the chart or graph, click Graphs → Chart
Builder. This will produce the dialog boxes shown in Figure 3.19, where a pie chart has been selected from the
list on the bottom left. In the Element Properties box on the right, you can change the metric used in the
chart. Counts are the default, but for reasons we covered in this chapter percentages are often preferable to
raw counts.
Figure 3.17 Running Frequencies in SPSS
Figure 3.18 SPSS Frequency Output
90
Figure 3.19 Using the SPSS Chart Builder to Create a Pie Chart
The Chart Builder can be used for bar graphs, too. Figure 3.20 displays the dialog box you will see if you
select Bar from the list on the lower left. For this graph, we will stick with the count default. Figures 3.21 and
3.22 show the pie chart and bar graph output, respectively.
Figure 3.20 Using the SPSS Chart Builder to Create a Bar Graph
Figure 3.21 SPSS Pie Chart
91
Figure 3.22 SPSS Bar Graph
Bivariate contingency tables are available in SPSS, too. Locate this option through Analyze → Descriptive
Statistics → Crosstabs. Select the variable that you would like to put in the rows of the table and the one you
would like to place in the column. (Recall that there are no rules governing this choice, although it is common
to put the variable you conceptualize as the independent variable or predictor in the rows and the one you
think is a dependent variable or outcome in the columns.) For this example, we will again work with the
LEMAS variable capturing police agency types. The second variable we will use comes from the same
LEMAS survey and measures whether each agency offers incentive pay (i.e., raises for educational attainment
intended to encourage officers to obtain college degrees). It makes more sense to think that agency type would
92
affect the offering of education incentives than to think the offering of educational incentives would affect an
agency’s jurisdiction (an unrealistic scenario). Therefore, the agency type variable will go into the rows and the
incentive variable will form the columns. We obtain percentages by clicking the Cells button and selecting row
or column percentages. For present purposes, we will opt for row percentages so SPSS will display the
percentage of each type of agency that does (and does not) offer educational incentives. Figure 3.23 shows the
output.
Finally, SPSS can be used to transform raw numbers into rates. To demonstrate this, we will keep working
with LEMAS and compute the number of municipal police officers per 1,000 residents for each of the city
police departments in the sample. Clicking on the Transform button at the very top of the SPSS data screen
and then clicking Compute will produce the Compute Variable box pictured in Figure 3.24.
Figure 3.23 SPSS Contingency Table Output (With Row Percentages)
Figure 3.24 Creating Rates in SPSS
In the Target Variable box (see Figure 3.24), type the name you want to give your new variable; in the present
93
example, the variable will be called policeper1000. In the “Numeric Expression” area, type the equation you
wish SPSS to follow to create the new variable. Here, the portion of the equation reading
“(police/population)” tells the program to begin the computation by dividing the total number of police in a
jurisdiction by that jurisdiction’s population size, and the “*1000” instructs SPSS to then multiply the results
by 1,000. Click OK, and a new variable will appear at the very end of the data set. This is your rate variable.
94
Learning Check 3.7
There is a good reason for using rates (such as the number of officers per 1,000 residents) to determine the size of a police agency rather
than simply using the number of officers employed by that organization. Can you explain why rates are better than raw counts in this
instance?
The chapter review questions contain directions for accessing data sets that you can use to practice
constructing charts and transforming variables into rates in SPSS. Play around with the data! Familiarize
yourself with SPSS; we will be visiting it regularly throughout the book, so the more comfortable you are with
it, the better prepared you will be. You do not have to worry about ruining the data set—if you make a
mistake, just click the “Undo” button or hit “Don’t save” when you exit the program, and the file will be as
good as new.
Chapter Summary
This chapter discussed some of the most common types of graphs and charts. Frequency distributions offer basic information about
the number of times certain characteristics appear in a data set. Frequencies are informative and can convey valuable information;
however, numbers are often difficult to interpret when they are in a raw format. Proportions and percentages offer a way to
standardize frequencies and make it easy to determine which characteristics occur more often and which less often. Rates are another
option for enhancing the interpretability of frequencies. Rates are generally multiplied by a number such as 1,000, 10,000, or
100,000.
Graphs and charts portray this same information—frequencies, proportions, percentages, and rates—using pictures rather than
numbers. Pictorial representations are more engaging than their numerical counterparts and can capture audiences’ interest more
effectively. Pie charts can be used with categorical variables that have five or fewer classes and are in percentage format. Bar graphs
are useful for categorical variables with any number of classes. They can be made from frequencies, proportions, percentages, or rates.
For continuous data, histograms and frequency polygons can be used to graph frequencies, proportions, or percentages. Line graphs
are useful for longitudinal data. Finally, some continuous variables that do not have a clear shape and are difficult to interpret in their
raw form can be grouped. Grouping transforms continuous variables into ordinal ones. Histograms can be used to display data that
have been grouped.
It is good to diversify a presentation by using a mix of pie charts, bar graphs, histograms, frequency polygons, and line charts.
Simplicity and variety are the keys to a good presentation. Simplicity ensures that your audience can make sense of your data display
quickly and easily. Variety helps keep your audience engaged and interested. Good data displays are key to summarizing data so that
you and others can get a sense for what is going on in a data set.
Thinking Critically
1. Suppose that in the past year, there were 15 incidents of violence committed by inmates against correctional staff in a nearby
prison. An administrator from the prison calls you to ask for an assessment of the severity of the problem of violence within
the prison (i.e., how prevalent inmate-on-staff violence is within this facility). What additional data will you request from the
prison administrator and what calculations will you perform? Think about rates, percentages, and longitudinal trends, as well
as any additional analyses that would help you understand this issue.
2. Consider three settings in which a researcher would potentially give presentations that summarize the results of data analyses:
academic conferences, presentations to practitioners (such as administrators from police agencies or jails), and presentations
to the community (such as community-based agencies or city councils). Based on the different audiences in each of these
three settings, explain what types of data displays you would use. In other words, how would you tailor your presentation to
make sure it was appropriate for your audience? Identify the displays you would use and explain your choices.
95
Review Problems
1. The following table contains data from the BJS’s State Court Processing Statistics, which includes data on felony defendants
in large urban counties in 2009 (Reaves, 2013). The variable is most serious arrest charge, which captures the most severe of the
offenses for which defendants were arrested.
1. Construct columns for proportion, percentage, cumulative frequencies, cumulative proportions, and cumulative
percentages.
2. Identify the types of charts or graphs that could be used to display this variable.
3. Based on your answer to (b), construct a graph or chart for this variable using percentages.
2. The following table contains data from the State Court Processing Statistics showing the number of urban felony defendants
sentenced to prison time after conviction for different types of property offenses (Reaves, 2013).
1. Construct columns for proportion, percentage, cumulative frequencies, cumulative proportions, and cumulative
percentages.
2. Identify the types of charts or graphs that could be used to display this variable.
3. Based on your answer to (b), construct a graph or chart for this variable using percentages.
3. The following table contains data from the COJ facilities showing the number of inmates on work release in medium-sized
facilities that offer this program for inmates.
1. Choose the appropriate graph type for this variable and construct that graph using frequencies.
2. Group this variable using 10 intervals.
3. Choose an appropriate graph type for the grouped variable and construct that graph using frequencies.
96
4. The following table contains PPCS data on the ages of respondents who said that the police had requested consent to search
their car during their most recent traffic stop.
97
1. Choose the appropriate graph type for this variable and construct that graph using frequencies.
2. Group this variable using six intervals.
3. Choose an appropriate graph type for the grouped variable and construct that graph using frequencies.
5. The General Social Survey (GSS) asks respondents whether they think that marijuana should be legalized. The following
table contains the percentage of respondents who supported legalization in each wave of the GSS from 1990 to 2014.
Construct a line graph of these data, and then interpret the longitudinal trend. Does support for marijuana legalization
appear to be increasing, decreasing, or staying the same over time?
98
6. The following table displays data from the official website of the U.S. courts (www.uscourts.gov) on the number of wiretap
authorizations issued by state and federal judges every year from 1997 to 2016. Construct a line graph of the data, and then
interpret the longitudinal trend. Have wiretap authorizations been increasing, decreasing, or staying the same over time?
7. The following table contains data on the number of violent crimes that occurred in six cities during 2015. The table also
displays each city’s population.
1. Compute the rate of violent crime per 1,000 city residents in each city.
2. Select the appropriate graph type for this variable and construct that graph using rates.
8. The following table contains data on the number of property crimes that occurred in six cities during 2015. The table also
displays each city’s population.
99
http://www.uscourts.gov
1. Compute the rate of property crime per 1,000 city residents in each city.
2. Select the appropriate graph type for this variable and construct that graph using rates.
9. The website for this chapter (http://www.sagepub.com/gau) contains a data set called Work Release for Chapter 3.sav. These
are the data from Review Problem 3. Use the SPSS Chart Builder to construct a frequency histogram showing the number of
inmates on work release in facilities offering this service.
10. The file City Police for Chapter 3.sav contains data from the 2013 LEMAS, narrowed to municipal (town and city)
departments. This data file contains the variable forfeiture, which indicates whether each agency includes asset forfeiture
revenues as a funding source in its operating budget. Run a frequency analysis to find out the percentage of departments that
do and do not include forfeiture in their operating budgets.
11. In the City Police for Chapter 3.sav file, there is a variable called sworn that displays each police department’s number of full-
time sworn officers and a variable called population that records the city population served by each department. Use these two
variables and the compute function to calculate the number of officers per 1,000 population for each department.
12. The website (www.sagepub.com/gau) also features a data file called Hate Crimes for Chapter 3.sav. This is the same data set
used in the in-text demonstration of line graphs. Use the SPSS Chart Builder to construct a line graph mirroring the one in
the text.
13. The website (www.sagepub.com/gau) contains a data set called Crime Attitudes for Chapter 3.sav. This is a trimmed version
of the 2014 GSS. One of the variables contained in the data set is courts, which measures respondents’ opinions about how
well courts punish people who commit crimes. Select an appropriate graph or chart and create it using SPSS.
14. The Crime Attitudes for Chapter 3.sav file also contains the variable marijuana, which captures respondents’ beliefs about
whether marijuana should be made legal. Use SPSS to construct a contingency table with courts in the rows and marijuana in
the columns. Include row percentages in your table. Provide a brief written explanation of the percentages and apparent
overlap between these two variables.
100
http://www.sagepub.com/gau
http://www.sagepub.com/gau
http://www.sagepub.com/gau
Key Terms
Cell 37
Univariate 37
Frequency 37
Proportion 38
Percentage 38
Cumulative 40
Contingency table 43
Bivariate 43
Classes 47
Longitudinal variables 56
Trends 37
Glossary of Symbols and Abbreviations Introduced in This Chapter
101
Chapter 4 Measures of Central Tendency
102
Learning Objectives
Define the three types of data distribution shapes.
Identify the shape of a distribution based on a comparison of the mean and median.
Describe the mode, the median, and the mean.
Explain which level(s) of measurement each measure of central tendency can be used with.
Identify, locate, or calculate the mode, the median, and the mean for a variety of variables and variable types.
Explain the difference between the two mean formulas and correctly use each one on the appropriate types of data.
Explain deviation scores and their relationship to the mean.
Use SPSS to produce the mode, the median, and the mean.
Criminal justice and criminology researchers and practitioners are often interested in averages. Averages offer
information about the centers or middles of distribution. They indicate where data points tend to cluster. This
is important to know. Consider the following questions that might be of interest to a researcher or
practitioner:
1. What is the most common level of educational attainment among police officers?
2. How does the median income for people living in a socioeconomically disadvantaged area of a certain
city compare to that for all people in the city?
3. What is the average violent crime rate across all cities and towns in a particular state?
4. Do prison inmates, in general, have a lower average reading level compared to the general population?
All of these questions make some reference to an average, a middle point, or, to use a more technical term, a
measure of central tendency. Measures of central tendency offer information about where the bulk of the
scores in a particular data set are located. A person who is computing a measure of central tendency is, in
essence, asking, “Where is the middle?”
Measures of central tendency: Descriptive statistics that offer information about where the scores in a particular data set tend to
cluster. Examples include the mode, the median, and the mean.
Averages offer information about the normal or typical person, object, or place in a sample. A group of people
with an average age of 22, for instance, probably looks different from a group averaging 70 years of age.
Group averages help us predict the score for any individual within that group. Suppose in two samples of
people, the only information you have is that one group’s average weight is 145 pounds and that the other’s is
200 pounds. If someone asked you, “How much does an individual person in the first group weigh?” your
response would be, “About 145 pounds.” If you were asked, “Who weighs more, a person randomly selected
from the first group or from the second group?” you would respond that the person from the second group is
probably the heavier of the two. Of course, you do not know for sure that you are correct; there might be
people in the first group who are heavier than some people in the second group. The average, nonetheless,
gives you predictive capability. It allows you to draw general conclusions and to form a basic level of
understanding about a set of objects, places, or people.
Measures of central tendency speak to the matter of distribution shape. Data distributions come in many
103
different shapes and sizes. Figure 4.1 contains data from the Police–Public Contact Survey (PPCS; see Data
Sources 2.1) showing the ages of non-Hispanic respondents who reported having called the police for help
within the past 24 months. This is the same variable used in the frequency polygon shown in Figure 3.10 in
the previous chapter. The shape this variable assumes is called a normal distribution. The normal curve
represents an even distribution of scores. The most frequently occurring values are in the middle of the curve,
and frequencies drop off as one traces the number line to the left or right. Normal distributions are ideal in
research because the average is truly the best predictor of the scores for each case in the sample, since the
scores cluster around that value.
Normal distribution: A set of scores that clusters in the center and tapers off to the left (negative) and right (positive) sides of the
number line.
Standing in contrast to normal curves are skewed distributions. Skew can be either positive or negative. The
distribution in Figure 4.2 contains the data from the Census of Jails (COJ; see Data Sources 3.1) showing the
number of inmate-on-staff assaults each jail reported experiencing in the past year. The distribution in Figure
4.2 manifests what is called a positive skew. Positively skewed data cluster on the left-hand side of the
distribution, with extreme values in the right-hand portion that pull the tail out toward the positive side of the
number line. Positively skewed data are common in criminal justice and criminology research.
Positive skew: A clustering of scores in the left-hand side of a distribution with some relatively large scores that pull the tail toward
the positive side of the number line.
Figure 4.1 Ages of Non-Hispanic Respondents Who Called the Police in the Past 24 Months
104
Learning Check 4.1
Skew type (positive versus negative) is determined by the location of the elongated tail of a skewed distribution. Positively skewed
distributions are those in which the tail extends toward the positive side of the number line; likewise, negative skew is signaled by a tail
extending toward negative infinity. Set aside your book and draw one of each type of distribution from memory.
Figure 4.2 Number of Inmate Assaults on Jail Staff
Figure 4.3 shows 2014 General Social Survey (GSS; see Data Sources 2.2) respondents’ annual family
incomes. This distribution has a negative skew: Scores are sparse on the left-hand side, and they increase in
frequency on the right side of the distribution.
Negative skew: A clustering of scores in the right-hand side of a distribution with some relatively small scores that pull the tail
toward the negative side of the number line.
Knowing whether a given distribution of data points is normal or skewed is vital in criminal justice and
criminology research. The average is an excellent predictor of individual scores when the curve is normal.
When a distribution departs from normality, however, the average becomes less useful and, in extreme cases,
can be misleading. For example, the mean number of inmate-on-staff assaults is 5.5 per jail, but you can see in
Figure 4.2 that the vast majority of jails had four or fewer assaults; in fact, a full two-thirds experienced just
two, one, or even zero assaults. A statement such as “Jails had an average of 5.5 inmate-on-staff assaults in
2013” would be technically correct, but would be very misleading because the typical jail has fewer incidents,
and many have a great deal more than that as well. Because this distribution is so skewed, the mean loses its
usefulness as a description of the middle of the data. Distribution shape plays a key role in more-complicated
statistical analyses (more on this in Parts II and III) and is important in a descriptive sense so that information
conveyed to academic, practitioner, or lay audiences is fully accurate and transparent. You must always know
where the middle of your data set is located; measures of central tendency give you that information.
Figure 4.3 GSS Respondents’ Annual Family Incomes
105
106
The Mode
The mode is the simplest of the three measures of central tendency covered in this chapter. It requires no
mathematical computations and can be employed with any level of measurement. It is the only measure of
central tendency available for use with nominal data. The mode is simply the most frequently occurring
category or value. Table 4.1 contains data from the 2011 PPCS (Data Sources 2.1). Interviewers asked PPCS
respondents whether they had been stopped by police while driving a vehicle. The people who answered yes
were then asked to report the reason for that stop. This is a nominal-level variable. Table 4.1 presents the
distribution of responses that participants gave for their stop. The mode is speeding because that is the stop
reason that occurs most frequently (i.e., 2,040 people said that this is the violation for which they were pulled
over by police).
Mode: The most frequently occurring category or value in a set of scores.
A frequency bar graph of the same data is shown in Figure 4.4. The mode is easily identifiable as the category
accompanied by the highest bar.
The mode can also be used with continuous variables. Instead of identifying the most frequently occurring
category as with nominal or ordinal data, you will identify the most common value. Figure 4.5 shows a
frequency histogram for the variable from the PPCS that asks respondents how many face-to-face contacts
they had with the police in the past 12 months. The sample has been narrowed to include only female
respondents who were 21 or younger at the time of the survey. Can you identify the modal number of
contacts? If you answered “1,” you are correct!
107
Learning Check 4.2
Flip back to Table 3.7 in the previous chapter and identify the modal jail size. (Hint: Use the row totals.)
Figure 4.4 Among Stopped Drivers, Reason for the Stop
Figure 4.5 Number of Police Contacts in Past 12 Months Among Females Age 21 and Younger
Research Example 4.1 Are People Convicted of Homicide More Violent in Prison Than People Convicted of Other Types of
Offenses?
Sorensen and Cunningham (2010) analyzed the institutional conduct records of all inmates incarcerated in Florida state correctional
facilities in 2003, along with the records of inmates who entered prison that same year. They divided inmates into three groups. The
stock population consisted of all people incarcerated in Florida prisons during 2003, regardless of the year they were admitted into
prison. The new persons admitted into prison during 2002 and serving all of 2003 composed the admissions cohort. The close
custody group was a subset of the admissions cohort and was made of the inmates who were considered to be especially high threats
to institutional security. The table below contains descriptive information about the three samples. For each group, can you identify
the modal custody level and modal conviction offense type? (Hint: These are column percentages.) Visit www.sagepub.com/gau to
view the full article and see the results of this study.
108
http://www.sagepub.com/gau
Source: Adapted from Table 1 in Sorensen and Cunningham (2010).
109
Do Latino Drug Traffickers’ National Origin and Immigration Status
Affect the Sentences They Receive?
The United States is experiencing significant demographic shifts, one of the most notable being the steady increase in the percentage
of the population that is Latino. Immigration (both legal and illegal) is a significant contributor to this trend. The influx has inspired
many native-born U.S. citizens to react with a get-tough, law enforcement–oriented mind-set toward immigrants who commit
crimes in this country. Mexican immigrants, in particular, are frequently seen as a threat, and this belief could translate into harsher
treatment for those immigrants who commit crimes. Logue (2017) examined whether Latino drug traffickers’ countries of origin and
their immigration status impacted the severity of the sentences they receive in federal court. One of the variables she included in her
analysis of sentencing outcomes was the type of drug that defendants were alleged to have trafficked. The table shows the breakdown
of drug types across defendants of different origins and immigration statuses. Can you identify the modal drug type for each of the
four defendant groups? (Hint: These are column percentages.) Visit www.sagepub.com/gau to view the full article and see the results
of this study.
Source: Adapted from Table 1 in Logue (2017).
110
http://www.sagepub.com/gau
Learning Check 4.3
Remember that the mode is the category or value that occurs the most frequently—it is not the frequency itself. Check your ability to tell
the difference between the mode and its frequency by looking back at Figure 4.2. Identify the modal number of assaults and the
approximate frequency for this number.
The previous two examples highlight the relative usefulness of the mode for categorical data as opposed to
continuous data. This measure of central tendency can be informative for the former but is usually not all that
interesting or useful for the latter. This is because there are other, more-sophisticated measures that can be
calculated with continuous data.
The strengths of the mode include the fact that this measure is simple to identify and understand. It also, as
mentioned previously, is the only measure of central tendency that can be used with nominal variables. The
mode’s major weakness is actually the flipside of its primary strength: Its simplicity means that it is usually too
superficial to be of much use. It accounts for only one category or value in the data set and ignores the rest. It
also cannot be used in more-complex computations, which greatly limits its usefulness in statistics. The mode,
then, can be an informative measure for nominal variables (and sometimes ordinal, as well) and is useful for
audiences who are not schooled in statistics, but its utility is restricted, especially with continuous variables.
111
The Median
The median (Md) is another measure of central tendency. The median can be used with continuous and
ordinal data; it cannot be used with nominal data, however, because it requires that the variable under
examination be rank orderable, which nominal variables are not.
Median: The score that cuts a distribution in half such that 50% of the scores are above that value and 50% are below it.
The median is the value that splits a data set exactly in half such that 50% of the data points are below it and
50% are above it. For this reason, the median is also sometimes called the 50th percentile. The median is a
positional measure, which means that it is not so much calculated as it is located. Finding the median is a
three-step process. First, the categories or scores need to be rank ordered. Second, the median position (MP)
can be computed using the formula
where
N = total sample size
The median position tells you where the median is located within the ranked data set. The third step is to use
the median position to identify the median. When N is odd, the median will be a value in the data set. When
N is even, the median will have to be computed by averaging two values.
Let us figure out the median violent crime rate among the five Colorado cities listed in Table 4.2. The
variable violent crime rate is continuous (specifically, ratio level), and the median is therefore an applicable
measure of central tendency. The numbers in Table 4.2 are derived from the 2015 Uniform Crime Reports
(UCR; see Data Sources 1.1).
The first step is to rank the rates in either ascending or descending order. Ranked in ascending order, they
look like this:
8.771 15.302 22.743 43.834 67.395
112
Superscripts have been inserted in the ranked list to help emphasize the median’s nature as a positional
measure that is dependent on the location of data points rather than these points’ actual values. The
superscripts represent each number’s position in the data set now that the values have been rank ordered.
Second, the formula for the median position will tell you where to look to find the median. Here,
This means that the median is in position 3. Remember that the MP is not the median; rather, it is a “map”
that tells you where to look to find the median.
Finally, use the MP to identify the median. Since the median is in position 3, we can determine that Md =
22.74. This group of five Colorado cities has a median violent crime rate of 22.74 per 10,000.
In this example, the sample had five cases. When there is an even number of cases, finding the median is
slightly more complex and requires averaging two numbers together. To demonstrate this, we will use 2015
property crime rates in six North Dakota cities (Table 4.3).
First, rank the values: 38.501 130.372 149.673 239.094 297.245 325.036.
Second, find the median position:
Notice that the MP has a decimal this time—this is what happens when the number of cases is even rather
than odd. What this means is that the median is halfway between positions 3 and 4, so finding Md requires
averaging these two numbers. The median is
This sample of six North Dakota cities has a median property crime rate of 194.38 per 10,000 residents. Of
these cities, 50% have rates that are lower than this, and 50% have rates that are higher.
For another example of locating the median in an even-numbered sample size, we will use state-level
113
homicide rates (per 100,000 population) from the 2015 UCR. Table 4.4 shows the data. To increase the
complexity a bit, we will use eight states.
Following the three steps, we first rank order the values:
1.331 1.882 1.933 4.404 5.635 5.796 6.087 7.168
Next, we use the MP formula to locate the position of the median in the data:
The MP of 4.5 tells us the median is halfway between the numbers located in positions 4 and 5, so we take
these two numbers and find their average. The number in position 4 is 4.40 and the number in position 5 is
5.63. Their average is
The median homicide rate (per 100,000) in this sample of eight states is 5.02.
114
Learning Check 4.4
Add the homicide rate for the state of Maryland (8.59) to the sample of states in Table 4.4 and locate the median using the three steps.
How much did the median change with the inclusion of this state?
Medians can also be found in ordinal-level variables, though the median of an ordinal variable is less precise
than that of a continuous variable because the former is a category rather than a number. To demonstrate, we
will use an item from the PPCS that measures the driving frequency among respondents who said that they
had been pulled over by police for a traffic stop within the past 6 months, either as a driver or as a passenger in
the stopped vehicle. Table 4.5 displays the data.
The first two steps to finding the Md of ordinal data mirror those for continuous data. First, the median
position must be calculated using Formula 4(1). For ordinal data, the total sample size (N) is found by
summing the frequencies. In Table 4.5, N = 1,544 and so the MP is calculated as
The median position in this case is a person—the person who is in position 772.5 is the median.
The second step involves identifying the category in which the MP is located. Instead of ranking the
categories according to frequencies as we did with continuous data, we are now going to arrange them in
either ascending or descending order according to the categories themselves. In other words, the internal
ranking system of the categories themselves is used to structure the sequence of the list. In Table 4.5, the
categories are arranged from the most-frequent driving habits (Every Day or Almost Every Day) to the least-
frequent ones (Never). As such, the categories are already in descending order and do not need to be
rearranged.
Next, add cumulative frequencies of the rank-ordered categories until the sum meets or exceeds the MP.
Table 4.6 illustrates this process. Here, 751 + 217 = 968, so the median is located in the A Few Days a Week
category. In other words, if you lined up all 1,544 people in this sample in the order in which they answered
the question, labeling the people in each group accordingly, and then counted them until you reached 772.5,
115
the person in that position would be standing in the A Few Days a Week group. Therefore, we now know that
half the people in this sample drive a few days a week or more, and half drive a few days a week or less.
Note how much less informative the median is for ordinal variables as compared to continuous ones. For the
crime rates in Tables 4.2 through 4.4, we were able to identify the specific, numerical median; for the driving-
frequency variable in Table 4.5, we are able to say only that the median case is contained within the A Few
Days a Week category. This is a rough estimate that paints a limited picture.
116
Learning Check 4.5
Rearrange the frequencies from Table 4.5 so that they are in descending order of driving frequency, rather than in ascending order as was
the case in the demonstration. Complete the cumulative-frequencies exercise to locate the median. What is your conclusion?
The median has advantages and disadvantages. Its advantages are that it uses more information than the mode
does, so it offers a more-descriptive, more-informative picture of the data. It can be used with ordinal
variables, which is advantageous because, as we will see, the mean cannot be.
A key advantage of the median is that it is not sensitive to extreme values or outliers. To understand this
concept, revisit Table 4.3 and replace Minot’s rate of 325.03 with 600.00; then relocate the median. It did not
change, despite a near doubling of this city’s crime rate! That is because the median does not get pushed and
pulled in various directions when there are extraordinarily high or low values in the data set. As we will see,
this feature of the median gives it an edge over the mean, the latter of which is sensitive to extremely high or
extremely low values and does shift accordingly.
The median has the disadvantage of not fully using all available data points. The median offers more
information than the mode does, but it still does not account for the entire array of data. This shortfall of the
median can be seen by going back to the previous example regarding Minot’s property crime rate. The fact
that the median did not change when the endpoint of the distribution was noticeably altered demonstrates
how the median fails to offer a comprehensive picture of the entire data set. Another disadvantage of the
median is that it usually cannot be employed in further statistical computations. There are limited exceptions
to this rule, but, generally speaking, the median cannot be plugged into statistical formulas for purposes of
performing more-complex analyses.
117
The Mean
This brings us to the third measure of central tendency that we will cover: the mean. The mean is the
arithmetic average of a data set. Unlike locating the median, calculating the mean requires using every raw
score in a data set. Each individual point exerts a separate and independent effect on the value of the mean.
The mean can be calculated only with continuous (interval or ratio) data; it cannot be used to describe
categorical variables.
Mean: The arithmetic average of a set of data.
There are two formulas for the computation of the mean, each of which is for use with a particular type of
data distribution. The first formula is one with which you are likely familiar from college or high school math
classes. The formula is
where
(x bar) = the sample mean,
Σ (sigma) = a summation sign directing you to sum all numbers or terms to the right of it,
x = values in a given data set, and
N = the sample size.
This formula tells you that to compute the mean, you must first add all the values in the data set together and
then divide that sum by the total number of values. Division is required because, all else being equal, larger
data sets will produce larger sums, so it is vital to account for sample size when attempting to construct a
composite measure such as the mean.
For the example concerning computation of the mean, we can reuse the Colorado violent crime rate data from
Table 4.2:
In 2015, these five cities had a mean violent crime rate of 31.61 per 10,000 residents. Let us try one more
example using data from Table 4.4 (homicide rates in 8 states). The mean is calculated as
118
Learning Check 4.6
Practice calculating the mean using the property crime rates in Table 4.3.
The second formula for the mean is used for large data sets that are organized in tabular format using both an
x column that contains the raw scores and a frequency (f) column that conveys information about how many
times each x value occurs in the data set. Table 4.7 shows data from the Bureau of Justice Statistics (BJS) on
the number of death-sentenced prisoners received per state in 2013 (Snell, 2014). Note that the f column
sums to 36 rather than 50 because in 2013, 14 states did not authorize the death penalty and are thus excluded
from the analysis. (This number climbed to 15 that same year when Maryland abolished capital punishment
and the governor commuted the sentences of the four people remaining on death row.) Table 4.7 contains the
numbers that states reported receiving and the frequency of each number. For instance, 20 states that
authorize the death penalty did not admit any new prisoners to death row, while 5 admitted one new inmate
each.
To calculate the mean using frequencies, first add a new column to the table. This column—titled fx—is the
product of the x and f columns. The results of these calculations for the death row data are located in the
right-hand column of Table 4.8. The fx column saves time by using multiplication as a shortcut and thereby
avoiding cumbersome addition. Using the conventional mean formula would require extensive addition
because you would have to sum 36 numbers (i.e., 20 zeroes plus 5 ones plus 4 twos, and so on). This process is
unwieldy and impractical, particularly with very large data sets. Instead, merely multiply each value by its
frequency and then sum these products to find the total sum of all x values. You can see from Table 4.8 that,
in 2013, states received 81 new death-sentenced inmates.
Once the fx column is complete and has been summed, the mean can be calculated using a formula slightly
different from the one presented in Formula 4(2). The mean formula for large data sets is
119
where
f = the frequency associated with each raw score x and
fx = the product of x and f.
The process of computing this mean can be broken down into three steps: (1) Multiply each x by its f, (2) sum
the resulting fx products, and (3) divide by the sample size N. Plugging the numbers from Table 4.8 into the
formula, it can be seen that the mean is
In 2013, the 36 states that authorized use of the death penalty each received a mean of 2.25 new death-
sentenced prisoners.
120
Learning Check 4.7
Anytime you need to compute a mean, you will have to choose between Formulas 4(2) and 4(3). This is a simple enough choice if you just
consider that in order to use the formula with an f in it, there must be an f column in the table. If there is no f column, use the formula
that does not have an f. Refer back to Table 2.9 in Chapter 2. Which formula would be used to calculate the mean? Explain your answer.
As practice, use the formula and compute the mean number of prisoners executed per state.
Note that Table 2.9 contains all 50 states, including those states that have abolished capital punishment, whereas Tables 4.7 and 4.8
contain only those 36 states that allowed the death penalty at the time the data were collected. What would happen to the mean
calculated based on Table 4.8 if the 14 states that did not allow the death penalty (and thus had zero admissions) were added to the table
and to the computation of the mean? Alternatively, what would happen to the mean you calculated on Table 2.9 if the 14 states without
capital punishment (which therefore had no executions) were removed from the calculation of the mean? Think about these questions
theoretically and make a prediction about whether the mean would increase or decrease. Then make the change to each set of data and
redo the means. Were your predictions correct?
Research Example 4.2 How Do Offenders’ Criminal Trajectories Impact the Effectiveness of Incarceration?
It is well known that some offenders commit a multitude of crimes over their life and others commit only a few, but the intersection
of offense volume (offending rate) and time (the length of a criminal career) has received little attention from criminal
justice/criminology researchers. Piquero, Sullivan, and Farrington (2010) used a longitudinal data set of males in South London who
demonstrated delinquent behavior early in life and were thereafter tracked by a research team who interviewed them and looked up
their official conviction records. The researchers were interested in finding out whether males who committed a lot of crimes in a
short amount of time (the short-term, high-rate [STHR] offenders) differed significantly from those who committed crimes at a
lower rate over a longer time (the long-term, low-rate [LTLR] offenders) on criminal justice outcomes. The researchers gathered the
following descriptive statistics. The numbers not in parentheses are means. The numbers in parentheses are standard deviations,
which we will learn about in the next chapter.
You can see from the table that the LTLR offenders differed from the STHR offenders on a number of dimensions. They were,
overall, older at the time of their first arrest and had a longer mean career length. They committed many fewer crimes per year and
were much less likely to have been sentenced to prison.
Piquero et al.’s (2010) analysis reveals a dilemma about what should be done about these groups of offenders with respect to
sentencing; namely, it shows how complicated the question of imprisonment is. The STHR offenders might appear to be the best
candidates for incarceration based on their volume of criminal activity, but these offenders’ criminal careers are quite short. It makes
no sense from a policy and budgetary perspective to imprison people who would not be committing crimes if they were free in
society. The STHR offenders also tended to commit property offenses rather than violent ones. The LTLR offenders, by contrast,
committed a disproportionate number of violent offenses despite the fact that their overall number of lifetime offenses was lower
than that for the STHR group. Again, though, the question of the utility of imprisonment arises: Is it worth incarcerating someone
who, though he might still have many years left in his criminal career, will commit very few crimes during that career? The dilemma
of sentencing involves the balance between public safety and the need to be very careful in the allotting of scarce correctional
resources.
121
Source: Adapted from Table 1 in Piquero et al. (2010).
122
Can Good Parenting Practices Reduce the Criminogenic Impact of
Youths’ Time Spent in Unstructured Activities?
Youths spending time with peers, away from parents and other adults, are at elevated risk for engaging in delinquency either
individually or in a group setting. The chances of unstructured, unsupervised activities leading to antisocial acts increases in
disadvantaged urban settings, where opportunities for deviance are higher and youths who have never been in trouble before are
more likely to encounter delinquent peers. Janssen, Weerman, and Eichelsheim (2017) posited that parents can reduce the chances
that their children’s unstructured time will result in deviance or delinquency. The researchers hypothesized that strong bonds
between parents and children mitigate the criminogenic impacts of time spent in urban environments with peer groups, as does the
extent to which parents monitor their children and set boundaries. They used a longitudinal research design, wherein a random
sample of adolescents was interviewed twice over a span of two years. Data were collected on how active parents were in their
children’s lives, the quality of each parent–child relationship, time spent in unstructured activities within disorderly environments,
and the number of delinquent acts each child had committed. The table contains the means and the standard deviations for each
wave of data collection.
The authors found that all three measures of positive parenting practices significantly reduced children’s delinquency. There was no
indication that positive parenting buffers children against the deleterious impacts of criminogenic environments; however, good
parenting and strong parent–child bonds did negate the effects of increases in the amount of time adolescents spent in these
environments. These results suggest that although parenting practices alone do not fully combat the bad influence of unstructured
time spent in disorderly neighborhoods, they can offset the effects of an increase in time spent in this manner. Parents have an
important role in preventing their children from slipping into deviance and delinquency.
Source: Adapted from Table 1 in Janssen et al. (2017).
123
Learning Check 4.8
Explain why the mean cannot be calculated on ordinal data. Look back at the driving frequency variable in Table 4.6 and identify the
information that is missing and prevents you from being able to calculate a mean.
The mean is sensitive to extreme values and outliers, which gives it both an advantage and a disadvantage
relative to the median. The advantage is that the mean uses the entire data set and accounts for “weird” values
that sometimes appear at the high or low ends of the distribution. The median, by contrast, ignores these
values. The disadvantage is that the mean’s sensitivity to extreme values makes this measure somewhat
unstable; it is vulnerable to the disproportionate impact that a small number of extreme values can exert on the
data set.
To illustrate this property of the mean, consider Table 4.9, which contains the 2015 homicide rates for six
cities in California. Trace the changes in the mean homicide rates from left to right. Do you notice how the
rate increases with the successive introductions of Los Angeles, Soledad, and San Bernardino? Los Angeles
pulls the mean up from 2.17 to 3.41, and Soledad tugs it to 5.11. The most dramatic increase occurs when
San Bernardino is added: The mean shoots up to 7.65. The successive addition of higher-rate cities caused, in
total, more than a threefold increase from the original mean across the three low-crime cities.
This demonstrates how the inclusion of extreme values can cause the mean to move in the direction of those
values. A score that is noticeably greater than the others in the sample can draw the mean upward, while a
value that is markedly lower than the rest can drive the mean downward. There is a good reason why, for
example, average income in the United States is reported as a median rather than a mean—a mean would
lump extremely poor people who are barely scraping by in with multibillionaires. That would not be accurate
at all! The apparent “average” income in the United States would be huge. Finding the point at which 50% of
households sit below that particular annual income and 50% above it is more useful and accurate. Because of
its sensitivity to extreme values, the mean is most informative when a distribution is normally distributed; the
accuracy of the mean as a true measure of central tendency is reduced in distributions that are positively or
negatively skewed, and the mean can actually become not merely inaccurate but downright misleading. For
instance, a severely economically divided city could have a mean income in the hundreds of thousands of
124
dollars, even if a significant portion of the local population is impoverished.
Another implication of the mean’s sensitivity to extreme values is that the mean and the median can be
compared to determine the shape of a distribution, as described in the following section.
125
Using the Mean and the Median to Determine Distribution Shape
Given that the median is invulnerable to extreme values but the mean is not, the best strategy is to report both
of these measures when describing data distributions. The mean and the median can, in fact, be compared to
form a judgment about whether the data are normally distributed, positively skewed, or negatively skewed. In
normal distributions, the mean and median will be approximately equal. Positively skewed distributions will
have means markedly greater than their medians. This is because extremely high values in positively skewed
distributions pull the mean up but do not affect the location of the median. Negatively skewed distributions,
on the other hand, will have medians that are noticeably larger than their means because extremely low
numbers tug the mean downward but do not alter the median’s value. Figure 4.6 illustrates this conceptually.
To give a full picture of a data distribution, then, it is best to make a habit of reporting both the mean and the
median.
The mean—unlike the mode or the median—forms the basis for further computations; in fact, the mean is an
analytical staple of many inferential hypothesis tests. The reason that the mean can be used in this manner is
that the mean is the midpoint of the magnitudes. This point is important and merits its own section.
Midpoint of the magnitudes: The property of the mean that causes all deviation scores based on the mean to sum to zero.
Figure 4.6 The Mean and Median as Indicators of Distribution Shape
126
Deviation Scores and the Mean as the Midpoint of the Magnitudes
The mean possesses a vital property that enables its use in complex statistical formulas. To understand this, we
must first discuss deviation scores. A deviation score is a given data point’s distance from its group mean. The
formula for a deviation score is based on simple subtraction:
Deviation score: The distance between the mean of a data set and any given raw score in that set.
where
di = the deviation score for a given data point xi, and
xi = a given data point,
= the sample mean.
Suppose, for instance, that a group’s mean is = 24. If a certain raw score xi is 22, then dx=22 = 22 – 24 = –2.
A raw score of 25, by contrast, would have a deviation score of dx=25 = 25 – 24 = 1.
A deviation score conveys two pieces of information. The first is the absolute value of the score or, in other
words, how far from the mean a particular raw score is. Data points that are exactly equal to the mean will
have deviation scores of 0; therefore, deviation scores with larger absolute values are farther away from the
mean, while deviation scores with smaller absolute values are closer to it.
The second piece of information that a deviation score conveys is whether the raw score associated with that
deviation score is greater than or less than the mean. A deviation score’s sign (positive or negative)
communicates this information. Positive deviation scores represent raw scores that are greater than the mean
and negative deviation scores signify raw numbers that are less than the mean. You can thus discern two
characteristics of the raw score xi that a given deviation score di represents: (1) the distance between xi and
and (2) whether xi is above or below it. Notice that you would not even need to know the actual value of xi
or in order to effectively interpret a deviation score. Deviation scores convey information about the position
of a given data point with respect to its group mean; that is, deviation scores offer information about raw
scores’ relative, rather than absolute, positions within their group. Figure 4.7 illustrates this.
What lends the mean its title as the midpoint of the magnitudes is the fact that deviation scores computed
using the mean (as opposed to the mode or median) always sum to zero (or within rounding error of it). The
mean is the value in the data set at which all values below it balance out with all values above it. For an
example of this, try summing the deviation scores in Figure 4.7. What is the result?
To demonstrate this concept more concretely, the raw homicide counts that were used as rates in Table 4.9
are listed in Table 4.11. These cities had a mean of 87.25 homicides in 2015. The d column contains the
deviation score for each homicide count.
127
Illustrative of the mean as the fulcrum of the data set, the positive and negative deviation scores ultimately
cancel each other out, as can be seen by the sum of zero at the bottom of the deviation-score column. This
represents the mean’s property of being the midpoint of the magnitudes—it is the value that perfectly balances
all of the raw scores. This characteristic is what makes the mean a central component in more-complex
statistical analyses. You will see in later chapters that the mean features prominently in many calculations.
Figure 4.7 Deviation Scores in a Set of Data With a Mean of 24
128
Learning Check 4.9
To test your comprehension of the concept of the mean as the midpoint of the magnitudes, go back to Table 4.3. You calculated the
mean property-crime rate in Learning Check 4.6. Use that mean to compute each city’s deviation score, and then sum the scores.
129
SPSS
Criminal justice and criminology researchers generally work with large data sets, so computing measures of
central tendency by hand is not feasible; luckily, it is not necessary, either, because statistical programs such as
SPSS can be used instead. There are two different ways to obtain central tendency output. Under the Analyze
→ Descriptive Statistics menu, SPSS offers the options Descriptives and Frequencies. Both of these functions
will produce central tendency analyses, but the Frequencies option offers a broader array of descriptive statistics
and even some charts and graphs. For this reason, we will use Frequencies rather than Descriptives. Once you
have opened the Frequencies box, click on the Statistics button to open a menu of options for measures of
central tendency. Select Mean, Median, and Mode, as shown in Figure 4.8. Then click OK, and the output
displayed in Figure 4.9 will appear. The variable used in this example comes from Table 4.7, measuring the
number of prisoners received under sentence of death in 2015.
You can see from Figure 4.9 that the mean is identical to the result we arrived at by hand earlier. The mode is
zero, which you can verify by looking at Table 4.7. The median is zero, meaning half the states did not receive
any new death-row inmates and half received one or more. We can also compare the mean and the median to
determine the shape of the distribution. With a mean of 2.25 and a median of zero, do you think that this
distribution is normally distributed, positively skewed, or negatively skewed? If you said positively skewed, you
are correct!
Figure 4.8 Running Measures of Central Tendency in SPSS
Figure 4.9 SPSS Output
130
For another example, we will use 2015 homicide rates (per 100,000) in all California cities. Figure 4.10 shows
the mean, the median, and the mode for this variable.
Follow along with this example of using SPSS to obtain measures of central tendency by downloading the file California Homicides
for Chapter 4.sav at www.sagepub.com/gau.
Figure 4.10 Homicide Rates in California Cities
The mode is zero, but because homicide rates vary so widely across this large sample of cities (N = 460), the
mode is not a useful or informative measure. More interesting are the mean and the median. The mean (6.58)
is much larger than the median (1.08), indicating significant positive skew. This can be verified by using SPSS
to produce a histogram of the data using the Chart Builder; refer back to Chapter 3 if you need a refresher on
the use of the Chart Builder. Figure 4.11 shows the histogram.
Figure 4.11 Histogram of Homicide Rates in California Cities
131
http://www.sagepub.com/gau
Figure 4.11 confirms the severe skew in this variable: The data cluster toward the lower end so much that the
values out in the tail are barely visible in the histogram.
132
Learning Check 4.10
Flip back a few pages to Figure 4.1. Recall that this is an example of a normal distribution. Based on the distribution’s shape, how close or
far apart do you think the mean and the median are? In other words, do you think they are close together, or do you predict that one is a
lot bigger (or smaller) than the other? Explain your answer.
There is a GIGO alert relevant here. It is your responsibility to ensure that you use the correct measure(s) of
central tendency given the level of measurement of the variable with which you are working. The SPSS
program will not produce an error message if you make a mistake by, for instance, telling it to give you the
mean of a nominal or ordinal variable. You will get a mean, just the same as you get when you correctly ask for
the mean of a continuous variable. To illustrate this, Figure 4.12 contains output from the National Crime
Victimization Survey (NCVS) showing respondents’ marital status. Although this is a nominal variable—
making the median and the mean inappropriate—SPSS went ahead with the calculations anyway and
produced results. Of course, the mean and the median of a variable measuring whether someone is married,
separated, divorced, and so on are nonsense, but SPSS does not know that. This statistical program is not a
substitute for knowing which techniques are appropriate for which data types.
Figure 4.12 Respondent Marital Status
Chapter Summary
This chapter introduced you to three measures of central tendency: mode, median, and mean. These statistics offer summary
information about the middle or average score in a data set. The mode is the most frequently occurring category or value in a data
set. The mode can be used with variables of any measurement type (nominal, ordinal, interval, or ratio) and is the only measure that
can be used with nominal variables. Its main weakness is in its simplicity and superficiality—it is generally not all that useful.
The median is a better measure than the mode for data measured at the ordinal or continuous level. The median is the value that
splits a data set exactly in half. Since it is a positional measure, the median’s value is not affected by the presence of extreme values;
this makes the median a better reflection of the center of a distribution the mean is when a distribution is highly skewed. The
median, though, does not take into account all data points in a distribution, which makes it less informative than the mean.
The mean is the arithmetic average of the data and is used with continuous variables only. The mean accounts for all values in a data
set, which is good because no data are omitted; the flipside, however, is that the mean is susceptible to being pushed and pulled by
extreme values.
It is good to report both the mean and the median because they can be compared to determine the shape of a distribution. In a
normal distribution, these two statistics will be approximately equal; in a positively skewed distribution, the mean will be markedly
greater than the median; and in a negatively skewed distribution, the mean will be noticeably smaller than the median. Reporting
both of them provides your audience with much more information than they would have if you just reported one or the other.
133
The mode, the median, and the mean can all be obtained in SPSS using the Analyze → Descriptive Statistics → Frequencies sequence.
As always, GIGO! When you order SPSS to produce a measure of central tendency, it is your responsibility to ensure that the
measure you choose is appropriate to the variable’s level of measurement. If you err, SPSS will probably not alert you to the mistake
—you will get output that looks fine but is actually garbage. Be careful!
Thinking Critically
1. According to the Law Enforcement Management and Administrative Statistics (LEMAS) survey, the mean number of
officers per police agency (of all types) is 163.92. Do you trust that this mean is an accurate representation of the middle of
the distribution of police agency size? Why or why not? If not, what additional information would you need in order to gain
an accurate understanding of this distribution’s shape and central tendency?
2. The COJ reports that the modal number of inmates per jail is 1. This value occurs more frequently than any other population
value (51 times among 2,371 jails). Use this example to discuss the limitations and drawbacks of using the mode to describe
the central tendency of a continuous variable. Then identify the measure(s) you would use instead of the mode, and explain
why.
Review Problems
1. A survey item asks respondents, “How many times have you shoplifted?” and allows them to fill in the appropriate number.
1. What level of measurement is this variable?
2. What measure or measures of central tendency can be computed on this variable?
2. A survey item asks respondents, “How many times have you shoplifted?” and gives them the answer options: 0, 1–3, 4–6, 7 or
more.
1. What level of measurement is this variable?
2. What measure or measures of central tendency can be computed on this variable?
3. A survey item asks respondents, “Have you ever shoplifted?” and tells them to circle yes or no.
1. What level of measurement is this variable?
2. What measure or measures of central tendency can be computed on this variable?
4. Explain what an extreme value is. Include in your answer (1) the effect extreme values have on the median, if any, and (2) the
effect extreme values have on the mean, if any.
5. Explain why the mean is the midpoint of the magnitudes. Include in your answer (1) what deviation scores are and how they
are calculated and (2) what deviation scores always sum to.
6. In a negatively skewed distribution . . .
1. the mean is less than the median.
2. the mean is greater than the median.
3. the mean and the median are approximately equal.
7. In a normal distribution . . .
1. the mean is less than the median.
2. the mean is greater than the median.
3. the mean and the median are approximately equal.
8. In a positively skewed distribution . . .
1. the mean is less than the median.
2. the mean is greater than the median.
3. the mean and the median are approximately equal.
9. In a positively skewed distribution, the tail extends toward _____ of the number line.
1. the positive side
2. both sides
3. the negative side
10. In a negatively skewed distribution, the tail extends toward _____ of the number line.
1. the positive side
2. both sides
3. the negative side
134
11. The following table contains 2015 UCR data on the relationship between murder victims and their killers, among those
crimes for which the relationship status is known.
1. Identify this variable’s level of measurement and, based on that, state the appropriate measure or measures of central
tendency.
2. Determine or calculate the measure or measures of central tendency that you identified in part (a).
12. The frequency distribution in the following table shows rates of violent victimization, per victim racial group, in 2014
according to the NCVS (Truman & Langton, 2015). Use this table to do the tasks following the table.
1. Identify the median victimization rate using all three steps.
2. Compute the mean victimization rate across all racial groups.
13. Morgan, Morgan, and Boba (2010) report state and local government expenditures, by state, for police protection in 2007.
The data in the following table contain a random sample of states and the dollars spent per capita in each state for police
services.
1. Identify the median dollar amount using all three steps.
2. Calculate the mean dollar amount.
14. The following frequency distribution shows LEMAS data on the number of American Indian officers employed by state
police agencies.
135
1. Identify the modal number of American Indian officers in this sample of agencies.
2. Compute the mean number of American Indian officers in this sample.
3. The median number of American Indian officers is 6.00. Based on this median and the mean you calculated, would
you say that this distribution is normally distributed, positively skewed, or negatively skewed? Explain your answer.
15. The following frequency distribution shows a variable from the PPCS measuring, among female respondents who had been
stopped by police while walking or riding a bike, the number of minutes those stops lasted.
1. Compute the mean number of minutes per stop.
2. The median number of minutes was 5.00. Based on this median and the mean you calculated earlier, would you say
that this distribution is normally distributed, positively skewed, or negatively skewed? Explain your answer.
16. The following table shows UCR data on the number of aggravated assaults that occurred in five Wyoming cities and towns
in 2015.
136
1. Identify the median number of assaults in this sample.
2. Calculate the mean number of assaults.
3. Calculate each city’s deviation score, and sum the scores.
17. The following table displays the number of juveniles arrested for arson in select states in 2015, according to the UCR.
1. Identify the median number of arson arrests per state.
2. Calculate the mean number of arrests in this sample.
3. Calculate each state’s deviation score and sum the scores.
18. The data set NCVS for Chapter 4.sav (www.sagepub.com/gau) contains the ages of respondents to the 2015 NCVS. Run an
SPSS analysis to determine the mode, the median, and the mean of this variable. Summarize the results in words.
Helpful Hint: When running measures of central tendency on large data sets in SPSS, deselect the “Display frequency
tables” option in the “Frequencies” dialog box. This will not alter the analyses you are running but will make your output
cleaner and simpler to examine.
19. The data set NCVS for Chapter 4.sav (www.sagepub.com/gau) also contains portions of the Identity Theft Supplement
survey. The variable “purchases” asked respondents how many times they had made online purchases in the past year. Run an
SPSS analysis to determine the mode, the median, and the mean of this variable. Summarize the results in words.
20. The data file Staff Ratio for Chapter 4.sav (www.sagepub.com/gau) contains a variable from the 2006 COJ showing the
ratio of inmates to security staff per institution. Run an SPSS analysis to determine the mode, the median, and the mean of
this variable. Summarize the results in words.
137
http://www.sagepub.com/gau
http://www.sagepub.com/gau
http://www.sagepub.com/gau
Key Terms
Measures of central tendency 76
Normal distribution 76
Positive skew 77
Negative skew 77
Mode 78
Median 83
Mean 88
Midpoint of the magnitudes 96
Deviation score 97
Glossary of Symbols and Abbreviations Introduced in This Chapter
138
Chapter 5 Measures of Dispersion
139
Learning Objectives
Explain the difference between measures of dispersion and measures of central tendency.
Explain why measures of dispersion must be reported in addition to measures of central tendency.
Define kurtosis, leptokurtosis, and platykurtosis.
Explain and know the relevant formulas for variation ratio, range, variance, and standard deviation.
Apply the correct measure(s) of dispersion to any given variable based on that variable’s level of measurement.
Explain the normal curve and the concept behind two-thirds of cases being within one standard deviation of the mean.
Calculate the upper and lower boundaries of the typical range of scores in a normal distribution based on the mean and the
standard deviation.
Consider this question: Do you think that there is more variability in the physical size of housecats or in that
of dogs? In other words, if I gathered a random sample of housecats and a random sample of dogs, would I
find a greater diversity of sizes and weights in the cat sample or in the dog sample? In the dog sample, of
course! Dogs range from puny things you might lose in your sofa cushions all the way up to behemoths that
might eat your sofa. Cats, on the other hand, are . . . well, pretty much just cat-sized. If I were to draw
separate size distributions for dogs and housecats, they might look something like Figure 5.1.
The dog distribution would be wider and flatter than the cat distribution, while the cat distribution would
appear somewhat tall and narrow by comparison. This is because dogs’ sizes vary more than cats’ sizes do.
Now, let us consider a situation in which two distributions have the same mean. (The dog and cat
distributions would, of course, have different means.) Consider the hypothetical raw data for variables X1 and
X2 in Table 5.1.
These distributions have the same mean, so if this were the only piece of information you had about them,
you might be tempted to conclude that they are similar to one another. This conclusion would be quite wrong,
though. Look at Figure 5.2, which displays a line chart for these two variables. Which series do you think
represents Sample 1? Sample 2? If you said that the stars are Sample 1 and diamonds are Sample 2, you are
correct. You can see that the raw scores of Sample 1 cluster quite tightly around the mean, while Sample 2’s
scores are scattered about in a much less cohesive manner.
The previous examples highlight the importance of the subject of this chapter: measures of dispersion.
Dispersion (sometimes called variability) is the amount of “spread” present in a set of raw scores. Dogs have
more dispersion in physical size than cats have, just as Sample 2 has more dispersion than Sample 1 has.
Measures of dispersion are vital from an informational standpoint for the reason exemplified by the thought
experiment just presented—measures of central tendency convey only so much information about a
distribution and can actually be misleading, because it is possible for two very discrepant distributions to have
similar means. It is also necessary to know the shape of a data distribution so that you can assess whether it
appears fairly normal or whether it deviates from normality. We talked about skew in Chapter 3: Skew occurs
when values cluster at one end of the distribution or the other. The dispersion analog of skew is kurtosis.
There are two types of kurtosis: Leptokurtosis happens when values cluster together very tightly, and
140
platykurtosis is evident when values are markedly spread out. In Figure 5.1, the dog-size distribution is
platykurtic (representing a spread-out smattering of values), and the cat-size curve is leptokurtic (indicating
that values are highly clustered and have minimal variability).
Dispersion: The amount of spread or variability among the scores in a distribution.
Kurtosis: A measure of how much a distribution curve’s width departs from normality.
Leptokurtosis: A measure of how peaked or clustered a distribution is.
Platykurtosis: A measure of how flat or spread out a distribution is.
Figure 5.1 Hypothetical Distributions of Dog and Housecat Sizes
For all these reasons, measures of dispersion are necessary pieces of information about any distribution. They
go hand in hand with measures of central tendency, and the two types of descriptive statistics are usually
presented alongside one another. This chapter discusses four of the most common types of measures of
dispersion: the variation ratio, the range, the variance, and the standard deviation. Like measures of central
tendency, each measure of dispersion is suitable for variables of only certain levels of measurement.
Figure 5.2 Line Chart for Variables X1 and X2
141
142
The Variation Ratio
The variation ratio (VR) is the only measure of dispersion discussed in this chapter that can be used with
categorical (nominal and ordinal) data; the remainder of the measures that will be covered are restricted to
continuous data (interval or ratio). The VR is based on the mode (see Chapter 4). It measures the proportion
of cases that are not in the modal category. Recall from the discussion of proportions in Chapter 3 that the
proportion of cases that are in a certain category can be found using Formula 3(1): It is easy to take
this formula a step farther in order to calculate the proportion that is outside any given category.
Variation ratio: A measure of dispersion for variables of any level of measurement that is calculated as the proportion of cases located
outside the modal category. Symbolized as VR.
For finding the proportion of cases outside a specific category, we rely on the bounding rule, which states that
proportions always range from 0.00 to 1.00. The bounding rule leads to the rule of the complement. This rule
states that the proportion of cases in a certain category (call this category A) and the proportion located in
other categories (call this Not A) always sum to 1.00. The proportion of cases in A is written as “p(A)” and the
proportion of cases outside category A is “p(Not A).” Formally,
p(A) + p(Not A) = 1.00
Bounding rule: The rule stating that all proportions range from 0.00 to 1.00.
Rule of the complement: Based on the bounding rule, the rule stating that the proportion of cases that are not in a certain category
can be found by subtracting the proportion that are in that category from 1.00.
If p(A) is known, the formula can be reconfigured thusly in order to calculate p(Not A):
The variation ratio is a spinoff of the rule of the complement:
where
fmode = the number of cases in the modal category and
N = the sample size.
To illustrate use of the VR, consider Table 5.2, which contains data from the 2007 Census of State Court
Prosecutors (CSCP; see Data Sources 5.1). The variable used here is the size of the population served by full-
time prosecutorial offices (as summarized by Perry & Banks, 2011).
143
To compute the VR, first identify the mode and its associated frequency. Here, the mode is 99,999 or fewer,
and its frequency is 1,389. Now, plug the numbers into Formula 5(2):
The variation ratio is .30, which means that .30 (30%) of the cases fall outside the modal category. This is a
fairly limited amount of variation—70% of offices serve relatively small populations, and just 30% serve
midsized or large jurisdictions.
Data Sources 5.1 Census of State Court Prosecutors
Every 4 to 5 years, the Bureau of Justice Statistics (BJS) sends surveys to chief prosecutors in state courts. Some of the waves use
random samples of all offices nationwide, and some waves employ censuses (i.e., contain the total population of offices). The most
recent available wave is that for 2007, which was a census. The survey is delivered by phone, Internet, or mail to chief prosecutors,
who are asked to report on various aspects of their offices’ organization and operation. The data set contains information such as the
number of attorneys on staff, the number of felony cases closed in the prior year, and whether the office regularly makes use of DNA
analysis.
Table 5.3 displays another variable from the CSCP. This survey item asked prosecutors whether, in the past year, their offices had
pursued criminal actions against defendants suspected of elder abuse.
The mode is Yes, and the variation ratio is
This VR of .45 is close to .50, indicating that the offices were split nearly in half with respect to whether or not they had prosecuted
cases of elder abuse.
In sum, the variation ratio offers information about whether the data tend to cluster inside the modal category or whether a fair
number of the cases are in other categories. This statistic can be used with nominal and ordinal variables, which makes it unique
relative to the range, the variance, and the standard deviation. The problem with the VR is that it uses only a small portion of the
data available in a distribution. It indicates the proportion of cases not in the modal category, but it does not actually show where
those cases are located. It would be nice to know where the data are rather than merely where they are not.
144
145
Learning Check 5.1
The bounding rule and the rule of the complement are going to resurface in later chapters, so it is a good idea to make sure that you fully
understand them before moving on to the next section of this chapter. To check your comprehension of this idea, quickly fill in the
following blanks.
55% of a sample is female, and ____% is male.
96% of defendants in a sample were convicted, and ____% were not.
A judge choosing from three punishment options sentenced 70% of defendants to jail, 20% to probation, and the remaining ____% to pay
fines.
146
The Range
The range (R) is the simplest measure of dispersion for continuous-level variables. It measures the span of the
data or, in other words, the distance between the smallest and largest values. The range is very easy to
compute:
Range: A measure of dispersion for continuous variables that is calculated by subtracting the smallest score from the largest.
Symbolized as R.
The first step in computing the range is identification of the maximum and minimum values. The minimum
is then subtracted from the maximum, as shown in Formula 5(3). That is all there is to it!
For an example, we can use data from the Uniform Crime Reports (UCR; Data Sources 1.1) on the number
of juveniles arrested on suspicion of homicide in 2015 in eight states. Table 5.4 shows the data.
To calculate the range, first identify the maximum and minimum values. The maximum is 20 and the
minimum is 0. The range, then, is
R = 19 − 0 = 19
Another example of the calculation of the range can be found using another variable from the CSCP. This
variable collects information about the salary of the chief prosecutor in each office. Table 5.5 shows these data
for the eight districts in Maine.
The largest value in Table 5.5 is 98,000, and the smallest is 66,511.06, so the range is
R = 98,000 – 66,511.06 = 31,488.94.
147
There is a difference of nearly $32,000 between the lowest-paid and the highest-paid chief prosecutors in
Maine.
The range has advantages and disadvantages. First, it is both simple and straightforward. It offers useful
information and is easy to calculate and understand. This same feature, however, makes this measure of
dispersion too simplistic to be of much use. The range is superficial and offers minimal information about a
variable. It is silent as to the distribution of the data, such as whether they are normally distributed or whether
there is kurtosis present. The range can be misleading, too: Table 5.5 shows clearly that salaries cluster at the
upper end of the distribution (i.e., around the $90,000 mark) and that the salary of $66,511.06 is an unusually
low value. The range of $31,488.94 is mathematically correct but, in a practical sense, overstates the amount
of variation in salaries. Finally, because the range does not use all of the available data, it has no place in
further computations.
148
Learning Check 5.2
If most prosecutors’ salaries nationwide were between $90,000 and $100,000, with very few values that are higher or lower, would this
distribution be normally distributed, leptokurtic, or platykurtic? Explain your answer.
149
The Variance
Like the range, the variance can be used only with continuous data; unlike the range, the variance uses every
number in a data set. The variance and its offshoot, the standard deviation (discussed next), are the
quintessential measures of dispersion. You will see them repeatedly throughout the remainder of the book.
For this reason, it is crucial that you develop a comprehensive understanding of them both formulaically and
conceptually.
Variance: A measure of dispersion calculated as the mean of the squared deviation scores. Notated as s2.
The formula for the variance can be rather intimidating, so let us work our way up to it piece by piece and
gradually construct it. We will use the juvenile homicide arrest data from Table 5.4 to illustrate the
computation as we go. First, recall the concept of deviation scores that was discussed in the last chapter. Go
back and review if necessary. A deviation score (symbolized di) is the difference between a raw score in a
distribution and that distribution’s mean. Recall Formula 4(4):
where
xi = a given data point and
= the sample mean
The variance is constructed from mean-based deviation scores. The first step in computing the variance is to
compute the mean, and the second is to find the deviation score for each raw value in a data set. The first
piece of the variance formula, therefore, is (i.e., a deviation score for each raw value in the data).
Table 5.6a shows the original data from Table 5.4 along with a deviation score column to the right of the raw
scores. The mean of the data set has been calculated ( = 4.50). In the deviation score column, the mean has
been subtracted from each raw score to produce a variety of positive and negative deviation scores. Recall that
the sigma symbol (Σ) is a summation sign.
Deviation scores are a good first step, but what we end up with is an array of numbers. A table full of
deviation scores is no more informative than a table full of raw scores. What we need is a single number that
combines and represents all of the individual deviation scores. The most obvious measure is the sum—sums
are good ways of packaging multiple numbers into a single numerical term. The problem with this approach,
though, should be obvious: Deviation scores sum to zero. Since adding the deviation-score column will always
produce a zero (or within rounding error of it), the sum is useless as a measure of variance.
150
Learning Check 5.3
Test your recall of the mean as the midpoint of the magnitudes by creating an x-axis (horizontal line) with numbered tick marks and
drawing a vertical line representing the mean of the data in Table 5.6a. Then plot each raw score to get a chart similar to that in Figure
4.7.
We need to find a way to get rid of those pesky negative signs. If all of the numbers were positive, it would be
impossible for the sum to be zero. Squaring is used to accomplish this objective. Squaring each deviation score
eliminates the negative signs (because negative numbers always become positive when squared), and, as long
as the squaring is applied to all the scores, it is not a problematic transformation of the numbers. Table 5.6b
contains a new right-hand column showing the squared version of each deviation score.
Now the numbers can be summed. We can write the sum of the right-hand column in Table 5.6b as
, and we can compute the answer as such:
= 6.25 + 12.25 + 20.25 + 2.25 + 20.25 + 12.25 + 30.25 + 210.25 = 314.00
This sum represents the total squared deviations from the mean. We are not quite done yet, though, because
sample size must be taken into consideration in order to control for the number of scores present in the
variance computation. You would not be able to compare the sum of squared deviation scores across two data
sets with different sample sizes. The sum of the squared deviation scores must be divided by the sample size.
151
Learning Check 5.4
Standardizing sums by dividing them by the sample size is common in statistics, as later chapters will continue to demonstrate. To
understand the need for this type of division, imagine two samples that each sum to 100. The first sample contains 10 cases and the
second one contains 50. Calculate the mean for each of these samples, and explain why they are different even though the sums are the
same.
There is a bit of a hiccup, however: The variance formula for samples tends to produce an estimate of the
population variance that is downwardly biased (i.e., too small) because samples are littler than populations and
therefore typically contain less variability. This problem is especially evident in very small samples, such as
when N < 50.
The way to correct for this bias in sample-based estimates of population variances is to subtract 1.00 from the
sample size in the formula for the variance. Reducing the sample size by 1.00 shrinks the denominator and
thereby slightly increases the number you get when you perform the division. The sample variance is
symbolized s2, whereas the population variance is symbolized σ2, which is a lowercase Greek sigma. The
reason for the exponent is that, as described earlier, the variance is made up of squared deviation scores. The
symbol s2 communicates the squared nature of this statistic.
We can now assemble the entire s2 formula and compute the variance of the juvenile homicide arrest data.
The formula for the sample variance is
Plugging in the numbers and solving yields
152
The variance of the juvenile homicide arrest data is 44.86. This is the average squared deviation from the
mean in this data set.
Let’s try a second example of the variance. The CSCP captures the number of felony cases each office tried
before a jury in the past year. The vast majority of criminal defendants nationwide plead guilty, with a small
portion opting to exercise their right to trial by jury. Table 5.7a shows the number of felony jury trials handled
in the past year by a random sample of seven offices serving populations of 50,000 or less. (The letters in the
left column represent the offices.) Table 5.7b displays the deviation scores and squared deviation scores.
Applying Formula 5(4),
These offices’ jury-trial variance is 150.62. In other words, the average squared deviation from the mean is
150.62 units.
153
The variance is preferable to the variation ratio and range because it uses all of the raw scores in a data set and
is therefore a more informative measure of dispersion. It offers information about how far the scores are from
the mean. Every case in the sample is used when the variance is calculated. When the data are continuous, the
variance is better than either the VR or the range.
154
Learning Check 5.5
Variances and squared deviation scores are always positive; it is impossible to do the calculations correctly and arrive at a negative answer.
Why is this?
155
The Standard Deviation
Despite its usefulness, the variance has an unfortunate hitch. When we squared the deviations (refer to Table
5.7b), by definition we also squared the units in which those deviations were measured. The variance of the
juvenile homicide arrest data, then, is 44.86 arrests squared, and the variance for prosecutors’ offices is 150.62
trials squared. This obviously makes no sense. The variance produces oddities such as crimes squared and years
squared. The variance is also difficult to interpret—the concept of the “average squared deviation score” is not
intuitive and does not provide a readily understandable description of the data. Something needs to be done to
correct this. Luckily, a solution is at hand. Since the problem was created by squaring, it can be solved by
doing the opposite—taking the square root. The square root of the variance (symbolized s, or sometimes sd) is
the standard deviation:
Standard deviation: Computed as the square root of the variance, a measure of dispersion that is the mean of the deviation scores.
Symbolized as s or sd.
The standard deviation of the juvenile homicide arrest data is
And s for the prosecutors’ jury trials is
The square root transformation restores the original units; we are back to arrests and trials now and have
solved the problem of impossible and nonsensical squares. Note that the standard deviation—just like the
squared deviation scores and the variance—can never be negative.
Substantively interpreted, the standard deviation is the mean of the deviation scores. In other words, it is the
average distance between the individual raw scores and the distribution mean. It indicates the general spread
of the data by conveying information as to whether the raw values cluster close to the mean (thereby
producing a relatively small standard deviation) or are more dispersed (producing a relatively large standard
deviation). The standard deviation is generally presented in conjunction with the mean in the description of a
continuous variable. We can say, then, that in 2015, the eight states considered here had a mean of 4.50
juvenile arrests for homicides, with a standard deviation of 6.70. Similarly, this sample of prosecutors’ offices
handled a mean of 14.57 felony jury trials, with a standard deviation of 12.27.
To put the entire process together (from calculating the deviation scores all the way through to computing the
standard deviation), we will use data from the Census of Jails. Table 5.8a contains a random sample of 10 jails
and shows the number of juvenile (i.e., younger than age 18) girls held in each one.
156
Applying Formula 5(4),
Now we know the variance is 340.28. To find the standard deviation, we take the square root:
The standard deviation is reported more often than the variance is, for the reasons explained earlier. It is a
good idea to present the standard deviation alongside any mean that you report (also expect others to do the
same!). We will be using the standard deviation a lot in later chapters; it is a fundamental descriptive statistic
critical to many formulas. The standard deviation also has another useful property: It can be employed to
determine the upper and lower boundaries of the “typical” range in a normal distribution. When you know the
157
mean and the standard deviation for a sample, you can find out some important information about that
distribution, as discussed in the next section.
158
The Standard Deviation and the Normal Curve
Recall that the standard deviation is the mean of the deviation scores; in other words, it is the mean deviation
between each raw score and the distribution mean. Larger standard deviations represent greater variability,
whereas smaller ones suggest less dispersion. Figure 5.3 displays the relationship between the mean and the
standard deviation for populations and samples.
Figure 5.3 Mean and Standard Deviation for a Normally Distributed, Continuous Variable
159
Learning Check 5.6
Like the mean, the standard deviation plays a key role in many statistical analyses. We will be using both the mean and the standard
deviation a lot in the following chapters; therefore, be sure that you have a solid grasp on the calculation and the interpretation of the
standard deviation. As practice, revisit Table 5.5 and try calculating the standard deviation of chief prosecutors’ salaries.
Research Example 5.1 Does the South Have a Culture of Honor That Increases Gun Violence?
Scholars have frequently noted that the South leads the nation in rates of violence and that gun violence is particularly prevalent in
this region. This has led to the formation and proposal of multiple theories attempting to explain southern states’ disproportionate
involvement in gun violence. One of these theories is the “culture of honor” thesis that predicts that white male southerners are more
likely than their counterparts in other regions of the country to react violently when they feel that they have been disrespected or that
they, their family, or their property has been threatened. Copes, Kovandzic, Miller, and Williamson (2014) tested this theory using
data from a large, nationally representative survey of adults’ reported gun ownership and defensive gun use (the use of a gun to ward
off a perceived attacker). A primary independent variable (IV) was whether respondents themselves currently lived in the South. The
main dependent variable (DV) was the number of times respondents had used a firearm (either fired or merely brandished) to defend
themselves or their property against a perceived human threat within the past 5 years. The researchers reported the descriptive
statistics shown in the table below.
The authors ran a statistical analysis to determine if living in the South or in a state where the majority of the population was born in
the South was related to defensive gun use. They found that it was not: Neither currently living in the South nor living in a state
populated primarily with southerners increased the likelihood that respondents had used a gun defensively in the past 5 years. These
findings refuted the southern culture of honor thesis by suggesting that southern white males are no more likely than white males in
other areas of the country to resort to firearms to defend themselves.
Source: Adapted from Table 1 in Copes et al. (2014).
160
Do Neighborhoods With Higher Immigrant Concentrations
Experience More Crime?
Many studies have examined the possible link between immigration and crime. Popular assumptions that immigrants commit crimes
at high rates have been proven false by numerous studies uncovering no relationship between immigrant concentrations and crime
rates, or even negative relationships wherein crime rates were lower in cities with larger immigrant populations than in those with
fewer immigrants. Most of the research into immigration and crime, however, has been based in the United States. Sydes (2017)
extended this empirical inquiry to Australia. She collected data on neighborhoods in two Australian cities. Means and standard
deviations for the DV and several of the IVs are shown in the table. After running sophisticated regression models, Sydes found no
relationship between immigrant concentrations and violent crime within neighborhoods in any given year, but also discovered that
increases in immigrant populations from one year to the next were associated with statistically significant reductions in violence.
These results are similar to most of the research conducted in the United States and lend support to the conclusion that immigration
does not increase crime and might even reduce it.
Source: Adapted from Tables 1 and 2 in Sydes (2017).
A distinguishing characteristic of normal distributions is that approximately two-thirds of the scores in a
distribution are located within one standard deviation below and one standard deviation above the mean.
Figure 5.4 depicts this important property of normal distributions. This distance between one standard
deviation below and one standard deviation above the mean constitutes the “normal” or “typical” range of
values. Most of the scores fall within this range, with a smaller number being higher or lower.
To illustrate this, let us suppose we have a sample of housecats with a mean weight of 11 pounds and a
standard deviation of 2 pounds (see Figure 5.5). We can find the range of weights that fall within the “normal
zone” using subtraction and addition. Two-thirds of the cats would be between 11 – 2 = 9 pounds (i.e., one
standard deviation below the mean) and 11 + 2 = 13 pounds (one sd above the mean). The remaining one-
third would weigh less than 9 pounds (i.e., they would be more than one sd below the mean) or more than 13
pounds (greater than one sd above the mean). These extreme values would indeed occur but with relatively low
frequency. People, objects, and places tend to cluster around their group means, with extreme values being
infrequent and improbable. Remember this point; we will come back to it in later chapters.
Whenever you know the mean and the standard deviation of a normally distributed, continuous variable, you
161
can find the two values between which two-thirds of the cases lie (i.e., the outer boundaries of the “typical”
area). In a distribution with a mean of 100 and a standard deviation of 15, two-thirds of cases will be between
85 and 115.
162
Learning Check 5.7
The concept regarding two-thirds of cases lying between –1 sd and +1 sd will form a fundamental aspect of later discussions. You will need
to understand this concept both mathematically and conceptually. Practice computing the upper and lower limits using the means and the
standard deviations listed in the first table in Research Example 5.1.
Figure 5.4 In a Normal Distribution, Approximately Two-Thirds of the Scores Lie Within One Standard
Deviation of the Mean
Figure 5.5 Hypothetical Distribution of Housecats’ Weights
Research Example 5.2 Why Does Punishment Often Increase—Rather Than Reduce—Criminal Offending?
Deterrence is perhaps the most pervasive and ingrained punishment philosophy in the United States and, indeed, in much of the
world. It is commonly assumed that punishing someone for a criminal transgression will lessen the likelihood of that person
reoffending in the future. Several studies have noted, however, that the chance of offending actually increases after someone has been
punished. Offenders who have been caught, moreover, tend to believe that their likelihood of being caught again are very low
because they think that it is improbable that the same event would happen to them twice. This flawed probabilistic reasoning is
called the “gambler’s bias” or “gambler’s fallacy.”
Pogarsky and Piquero (2003) attempted to learn more about the way that the gambler’s bias could distort people’s perceptions of the
likelihood of getting caught for criminal wrongdoing. They distributed a survey with a drunk driving scenario to a sample of
university students and asked the respondents about two DVs: (a) On a scale of 0 to 100, how likely they would be to drive under the
influence in this hypothetical situation, and (b) on a scale of 0 to 100, how likely it is that they would be caught by police if they did
drive while intoxicated. The researchers also gathered data on several IVs, such as respondents’ criminal histories (which were used
to create a risk index scale), levels of impulsivity in decision making, and ability to correctly identify the probability of a flipped coin
landing on tails after having landed heads side up four times in a row. The coin-flip question tapped into respondents’ ability to use
probabilistic reasoning correctly; those who said that the coin is more likely to land tails up were coded as engaging in the type of
logical fallacy embodied by the gambler’s bias. The researchers obtained the means and standard deviations shown in the table on
163
page 126.
Pogarsky and Piquero (2003) divided the sample into those at high risk of offending and those at low risk and then analyzed the
relationship between the gambler’s fallacy and perceived certainty of punishment within each group. They found that the high-risk
respondents’ perceptions of certainty were not affected by the gambler’s bias. Even though 26% of people in this group engaged in
flawed assessments of probabilities, this fallacious reasoning did not impact their perceptions of the certainty of punishment.
Among low-risk respondents, however, there was a tendency for those who engaged in flawed probabilistic reasoning to believe that
they stood a very low chance of detection relative to those low-risk respondents who accurately assessed the probability of a coin flip
landing tails side up. The researchers concluded that people who are at high risk of offending might not even stop to ponder their
likelihood of being caught and will proceed with a criminal act when they feel so inclined. Those at low risk, on the other hand,
attempt to employ probabilities to predict their chances of being apprehended and punished. The gambler’s bias, therefore, might
operate only among relatively naive offenders who attempt to use probabilistic reasoning when making a decision about whether or
not to commit a criminal offense.
Source: Adapted from the appendix in Pogarsky and Piquero (2003).
164
SPSS
SPSS offers ranges, variances, and standard deviations. It will not provide you with variation ratios, but those
are easy to calculate by hand. The process for obtaining measures of dispersion in SPSS is very similar to that
for measures of central tendency. The juvenile homicide data set from Table 5.4 will be used to illustrate
SPSS. First, use the Analyze → Descriptive Statistics → Frequencies sequence to produce the main dialog box.
Then click the Statistics button in the upper right to produce the box displayed in Figure 5.6. Select Std.
deviation, Variance, and Range to obtain these three statistics.
Figure 5.6 Using SPSS to Obtain Measures of Dispersion
After you click Continue and OK, the output in Figure 5.7 will appear.
You can see in Figure 5.7 that all the numbers generated by SPSS match those that we obtained by hand.
Follow along with this example of using SPSS to obtain measures of central tendency by downloading the file Juvenile Arrests for
Chapter 5.sav at www.sagepub.com/gau.
Figure 5.7 SPSS Output for Measures of Dispersion
Chapter Summary
This chapter introduced four measures of dispersion: variation ratio, range, variance, and standard deviation. The VR tells you the
proportion of cases not located in the modal category. VR can be computed on data of any level of measurement; however, it is the
only measure of dispersion covered in this chapter that is available for use with categorical data. The range, the variance, and the
standard deviation, conversely, can be used only with continuous variables.
The range is the distance between the lowest and highest value on a variable. The range provides useful information and so should
be reported in order to give audiences comprehensive information about a variable. This measure’s usefulness is severely limited by
165
http://www.sagepub.com/gau
the fact that it accounts for only the two most extreme numbers in a distribution. It ignores everything that is going on between
those two endpoints.
The variance improves on the range by using all of the raw scores on a variable. The variance is based on deviation scores and is an
informative measure of dispersion. This measure, though, has a conceptual problem: Because computation of the variance requires all
of the deviation scores to be squared, the units in which the raw scores are measured also end up getting squared. The variance thus
suffers from a lack of interpretability.
The solution to the conceptual problem with the variance is to take the square root of the variance. The square root of the variance is
the standard deviation. The standard deviation is the mean of the deviation scores. Raw scores have a mean, and so do deviation
scores. The former is the distribution mean, and the latter is the standard deviation.
The mean and the standard deviation together are a staple set of descriptive statistics for continuous variables. The standard
deviation will be key to many of the concepts that we will be discussing in the next few chapters. Approximately two-thirds of the
cases in any normal distribution are located between one standard deviation below and one standard deviation above the mean.
These scores are within the normal or typical range; scores that are more than one standard deviation below or above the mean are
relatively uncommon.
Thinking Critically
1. It is common practice to present the mean and standard deviation together (as seen in the research examples in this chapter).
Why is this? In other words, explain (1) the reasons why a measure of central tendency and a measure of dispersion would be
reported together to describe the same set of data, and (2) why the mean and standard deviation are the measures typically
selected for this. Refer back to Chapter 4 if needed.
2. Suppose you are reading a news article discussing bank-fraud victimization. The report features the story of one victim who
was 20 years old when she discovered her bank account had been compromised and her money had been stolen. Because you
are familiar with the National Crime Victimization Survey’s Identity Theft Supplement, you know that the mean age of
bank-fraud victims is 45.52 (s = 15.26). Based on this, is the victim in the report typical of bank-fraud victims in terms of
age? How do you know? Do you trust that her experiences are representative of other bank-fraud victims? Explain your
answer.
Review Problems
1. Explain the reason why measures of dispersion are necessary in addition to measures of central tendency. That is, what
information is given by measures of dispersion that is not provided by measures of central tendency?
2. A data distribution that was very narrow and peaked would be considered
1. normal.
2. platykurtic.
3. leptokurtic.
3. In a normal curve, approximately _____ of values are between one standard deviation above and one standard deviation
below the mean.
4. A data distribution that was very flat and spread out would be considered
1. normal.
2. platykurtic.
3. leptokurtic.
5. The following table contains data from the Census of Jails. Compute the variation ratio.
6. The following table contains data from the PPCS. Compute the variation ratio.
166
7. The following table contains data from the Census of Jails. Compute the variation ratio.
8. The table below contains UCR data showing the 2015 rate of officer assaults (per 100 officers) in each of the four regions of
the country. Use this table to do the following:
1. Calculate the range.
2. Calculate the mean assault rate.
3. Calculate the variance.
4. Calculate the standard deviation.
9. The following table shows data from the BJS on the number of death-row inmates housed by states in the Midwest in 2013.
Use this table to do the following:
1. Calculate the range.
2. Calculate the mean number of inmates per state.
3. Calculate the variance.
4. Calculate the standard deviation.
10. The following table contains UCR data on the number of hate crimes in 2015. Use this table to do the following:
167
1. Calculate the range.
2. Calculate the mean number of firearm-perpetrated murders in these states.
3. Calculate the variance.
4. Calculate the standard deviation.
11. The following table shows UCR data on the percentage of burglaries cleared by arrest in 2015, as broken down by region.
Use this table to do the following:
1. Calculate the range.
2. Calculate the mean clearance rate per region.
3. Calculate the variance.
4. Calculate the standard deviation.
12. For each of the following means and standard deviations, calculate the upper and lower limits of the middle two-thirds of the
distribution.
1. = 6.00, sd = 1.50
2. = 14.00, sd = 3.40
168
3. = 109.32, sd = 14.98
13. For each of the following means and standard deviations, calculate the upper and lower limits of the middle two-thirds of the
distribution.
1. = 63.10, sd = 18.97
2. = 1.75, sd = .35
3. = 450.62, sd = 36.48
14. Explain the conceptual problem with the variance that is the reason why the standard deviation is generally used instead.
15. Explain the concept behind the standard deviation; that is, what does the standard deviation represent substantively?
16. The companion website (www.sagepub.com/gau) has an SPSS file called Firearm Murders for Chapter 5.sav that contains
2015 UCR data showing the percentage of murders, per state, that were committed with firearms. (There are 49 states in
this file because Florida did not submit UCR data. In addition, murders for which no weapon data were submitted are
excluded from these counts.) Use SPSS to obtain the range, the variance, and the standard deviation for the variable percent.
17. There is an SPSS file called Census of State Prosecutors for Chapter 5.sav on the website (www.sagepub.com/gau) that contains
data from the 2007 CSCP. This data set contains three variables. For the variable yearsinoffice, which measures the number
of years chief prosecutors have held their positions, use SPSS to obtain the range, the standard deviation, and the variance.
18. Using Census of State Prosecutors for Chapter 5.sav, use SPSS to obtain the range, standard deviation, and variance of the
variable assistants, which shows the number of full-time assistant prosecutors employed by each office.
19. Using Census of State Prosecutors for Chapter 5.sav and the variable felclosed, which captures the number of felony cases each
office closed in 2007, use SPSS to obtain the range, the standard deviation, and the variance.
169
http://www.sagepub.com/gau
http://www.sagepub.com/gau
Key Terms
Dispersion 107
Kurtosis 108
Leptokurtosis 108
Platykurtosis 108
Variation ratio 109
Bounding rule 109
Rule of the complement 109
Range 112
Variance 113
Standard deviation 118
Glossary of Symbols and Abbreviations Introduced in This Chapter
170
Part II Probability and Distributions
Chapter 6 Probability
Chapter 7 Population, Sample, and Sampling Distributions
Chapter 8 Point Estimates and Confidence Intervals
Part I of this book introduced you to descriptive statistics. You learned the mathematical and conceptual
underpinnings of proportions, means, standard deviations, and other statistics that describe various aspects of
data distributions. Many times, though, researchers want to do more than merely describe a sample—they
want to run a statistical test to analyze relationships between two or more variables. The problem is there is a
gap between a sample and the population from which it was drawn. Sampling procedures produce a subset of
the population, and it is not correct to assume that this subset’s descriptive statistics (such as its mean) are
equivalent to those of the population. Going back to our housecat example in the previous chapter, suppose
you go door to door in your neighborhood recording the weights of all the cats, and you find a mean weight of
9.50 pounds. Would it be safe to conclude that if you weighed all housecats in the world (i.e., the entire
population), the mean would be exactly 9.50? Definitely not! In the process of pulling your sample, you might
have picked up a few cats that are atypically small or large. We saw in the previous chapter how extreme values
can pull the sample mean up or down; if you got an extremely large or extremely small cat in your sample, the
mean would be thrown off as a result. There is always a possibility that any given sample contains certain
values that cause the mean, the proportion, or other descriptive statistic to be higher or lower than that for the
entire population from which the sample was derived. For this reason, sample statistics cannot be
automatically generalized to populations. We need a way of bridging the gap.
Inferential statistics (also called hypothesis testing, the subject of Part III of this book) provide this bridge
between a descriptive statistic and the overarching population. The purpose of inferential statistics is to permit
a descriptive statistic to be used in a manner such that the researcher can draw an inference about the larger
population. This procedure is grounded in probability theory. Probability forms the theoretical foundation for
statistical tests and is therefore the subject of Part II of this book. You can think of Part I as having
established the foundational mathematical and formulaic concepts necessary for inferential tests and of Part II
as laying out the theory behind the strategic use of those descriptive statistics. Part III is where these two areas
of knowledge converge.
Inferential statistics: The field of statistics in which a descriptive statistic derived from a sample is employed
probabilistically to make a generalization or inference about the population from which the sample was drawn.
Probability theory: Logical premises that form a set of predictions about the likelihood of certain events or the
empirical results that one would expect to see in an infinite set of trials.
Part II is heavily grounded in theory. You will not see SPSS sections or much use of research examples. This
is because probability is largely concealed from view in criminal justice and criminology research; probability is
the “man behind the curtain” who is pulling the levers and making the machine run but who usually remains
171
hidden. Although there will be formulas and calculations that you will need to understand, your primary task
in Part II is to form conceptual comprehension of the topics presented. When you understand the logic
behind inferential statistics, you will be ready for Part III.
172
Chapter 6 Probability
173
Learning Objectives
Explain the relationship between proportions and probabilities.
Explain the differences between theoretical predictions and empirical outcomes.
Define the binomial distribution and the normal distribution and explain the types of data for which each one is used.
Construct the binomial distribution using the binomial coefficient for a given probability and sample size.
Explain and diagram out the relationship among raw scores, z scores, and areas.
Use the z -score formula and z table to calculate unknown raw scores, z scores, and areas.
A probability is the likelihood that a particular event will occur. We all use probabilistic reasoning every day.
If I buy a lottery ticket, what are my chances of winning? What is the likelihood that I will get a promotion if
I put in extra effort at work? What is the probability that I will get pulled over if I drive 5 miles per hour over
the speed limit? These musings involve predictions about the likelihood that a certain event will (or will not)
occur. We use past experiences and what we know (or what we assume to be true) about the world to inform
our judgments about the chance of occurrence.
Probability: The likelihood that a certain event will occur.
Probabilities are linked intricately with proportions; in fact, the probability formula is a spinoff of Formula
3(1). Flip back to this formula right now for a refresher. Let us call a particular event of interest A. Events are
phenomena of interest that are being studied. The probability that event A will occur can be symbolized as p
(A) (pronounced “p of A ”) and written formulaically as
Coin flips are a classic example of probabilities. Any two-sided, fair coin will land on either heads or tails
when flipped, so the denominator in the probability formula is 2 (i.e., there are two possible outcomes of the
coin flip). There is one tails side on a coin, so the numerator is 1 (i.e., there is only one option for the coin to
land tails-side up). Probabilities, like proportions, are expressed as decimals. The probability of the coin
landing tails side up, then, is
Any time you flip a fair coin, there is a .50 probability that the coin will land on tails. Of course, the
probability that it will land heads side up is also .50. Note that the two probabilities together sum to 1.00:
p(tails) + p(heads) = .50 + .50 = 1.00
The probabilities sum to 1.00 because heads and tails are the only two possible results and thus constitute an
exhaustive list of outcomes. The coin, moreover, must land (it will not hover in midair or fly around the
174
room), so the probability sums to 1.00 because the landing of the coin is inevitable.
Think about rolling one fair die. A die has six sides, so any given side has
of being the one that lands faceup. If you were asked, “What is the
probability of obtaining a 3 on a single roll of a fair die?” your answer would be “.17.” There are 52 cards in a
standard deck and only one of each number and suit, so during any random draw from the deck, each card has
probability of being the card that is selected. If someone asked you,
“What is the probability of selecting the two of hearts from a full deck?” you would respond, “.02.” Note, too,
that in the instance of a die and a deck of cards—just as with the flipped coin—the sum of all events’
individual probabilities is 1.00 (or within rounding error of it). This is a reflection of the bounding rule and
the rule of the complement that we discussed in Chapter 5. A die has six sides, so there are six total possible
events: The die can land on 1, 2, 3, 4, 5, or 6. The probability that it will land on a specific predicted value
(say, 5, e.g.) is .17, whereas the probability that it will land on any of its six sides is .17 + .17 + .17 + .17 + .17
+ .17 = 1.02 (technically, this sum is zero, but the numbers used in the example contain rounding error).
Likewise, the probability of pulling a predetermined playing card (for instance, the four of clubs) is .02, but
the chance that you will, in fact, retrieve some card from the deck when you pull one out is .02(52) = 1.04
(exceeding 1.00 due to rounding error).
Probabilities can be used to make specific predictions about how likely it is that an event will occur. Maybe
you want to predict that you will draw any card from the diamonds suit. Decks contain four suits, each with
13 cards, meaning there are 13 diamond cards. The probability of drawing any one of them is
This is different from the previous example using the two of hearts
because here we do not care what number the card is as long as it belongs to the diamond suit. Likewise, there
are four cards of each number (one of each suit), so the probability that your random draw would produce a
nine of any suit would be Again, this is different from .02 because the prediction
that a draw will yield a nine of any suit opens up more options (four, to be exact) than when the suit is
restricted (resulting in just one option).
175
Learning Check 6.1
Use probability to answer the following two questions.
1. When you predict that a die will land on 2, what is the probability that it will actually land on 3 instead?
2. When you predict that a card will be the jack of diamonds, what is the probability that you will actually pull the ace of hearts
instead?
Based on what you know about the bounding rule and the rule of the complement, answer the following questions.
1. When you predict that a die will land on 1, what is the probability that it will land on any other value except 1?
2. When you predict that a card will be the nine of spades, what is the probability that the card you actually pull will be anything
except the nine of spades?
The major difference between proportions and probabilities is that proportions are purely descriptive, whereas
probabilities represent predictions. Consider Table 6.1, which shows the proportion of the respondents to the
Police–Public Contact Survey (PPCS; see Data Sources 2.1) that is male and the proportion that is female. If
you threw all 52,529 people into a gigantic hat and randomly drew one, what is the probability that the person
you selected would be female? It is .52! The probability that the person would be male? The answer is .48.
This is the relationship between proportions and probabilities.
Let us try another example. Table 6.2 shows data from the Uniform Crime Reports (UCR; see Data Sources
1.1) showing the percentage of crimes cleared by arrest in 2015.
We can use the percentages in Table 6.2 just as if they were proportions. (Flip back to Formulas 3[1] and 3[2]
if you need a refresher on this concept.) What is the probability that a robbery reported to police will result in
an arrest of the perpetrator? You can see that 29.30% of these crimes are cleared by arrest, which means that
176
any given robbery has a .29 chance of resulting in an arrest. Try this exercise with the other crime types listed
in the table.
177
Learning Check 6.2
In Table 6.2, the bounding rule and rule of the complement play out a little differently compared to the examples using coins, dice, and
cards. In the clearance data, these two rules must be applied separately for each individual crime type. There are two possible outcomes
anytime a crime occurs: that the police identify the perpetrator or that they do not. As such, there is a 100.00% that one of these two
outcomes will occur. We can use clearance rates and simple subtraction to find the chance that a particular crime will not be cleared by
arrest. For example, if 12.90% of burglaries result in arrest of the suspect, then 100.00% – 12.90% = 87.10% of burglaries are not cleared
(i.e., there is a .87 probability that any given burglary will not result in the arrest of a suspect). For each of the remaining crime types in
Table 6.2, calculate the percentage of crimes not cleared, and identify the corresponding probability.
Probability theory is grounded in assumptions about infinite trials; in other words, probabilities concern what
is expected over the long run. Think back to the coin flip example. Since p (tails) on any given flip is .50, then
over the course of a long-term flipping session, half of the flips will yield tails and the other half will land on
heads. In other words, if you spent several hours flipping one coin and recording each outcome, eventually you
would end up with a 50/50 split.
There is a distinct difference, though, between theoretical predictions and empirical outcomes. (Empirical
outcomes can also be called observed outcomes ; we will use the terms interchangeably.) Quite simply, and as
you have undoubtedly already learned in life, you do not always get what you expect. You can see this in
practice. First, imagine that you flipped a coin six times. How many of those flips would you expect to land
tails side up? Three, of course, because every flip has p (tails) = .50 and so you would expect half of all flips to
result in tails. Now, grab a real coin, flip it six times, and record each outcome. How many tails did you get?
Try 20 flips. Now how many tails? Transfer the following chart to a piece of paper and fill it in with your
numbers.
6 flips proportion tails =
20 flips proportion tails =
Theoretical prediction: A prediction, grounded in logic, about whether or not a certain event will occur.
Empirical outcome: A numerical result from a sample, such as a mean or frequency. Also called observed outcomes.
You might have found in your six-flip exercise that the number of tails departed noticeably from three; you
might have gotten one, five, or even six or zero tails. On the other hand, the number of tails yielded in the 20-
flip exercise should be approximately 10 (with a little bit of error; it might have been 9 or 11). Why is this? It
is because you increased the number of trials and thereby allowed the underlying probability to emerge in the
empirical outcomes. The knowledge that half of coin flips will produce tails over time is a theoretical
178
prediction based on the concept of infinite coin flipping (nobody has actually spent infinity flipping a coin),
whereas the flip experiment you just conducted is an empirical test and outcome. Theory guides us in
outlining what we expect to see (e.g., if you are asked how many times you would expect to see tails in a series
of six flips, your best guess would be three), but sometimes empirical results do not match expectations (e.g.,
you might see one tail, or possibly all six will be tails).
The relationship between theoretical predictions and empirical findings is at the heart of statistical analysis.
Researchers constantly compare observations to expectations to determine if empirical outcomes conform to
theory-based predictions about those outcomes. The extent of the match (or mismatch) between what we
expect (theory) and what we see (reality) is what leads us to make certain conclusions about both theory and
reality.
179
Learning Check 6.3
If you are finding the material in this chapter to be about as clear as mud, then you are experiencing the perfectly normal learning process
for this topic. These ideas take time to sink in. A highly recommended study technique is to spend time simply thinking about these
concepts. Work them over in your mind. Take breaks as you read so that you can digest concepts before continuing. Practice additional
probability calculations or try concrete examples like drawing cards, flipping a coin, or rolling a die.
Now, let us add another layer to the discussion. We have thus far covered four examples of situations in which
probability can be used to make predictions: coin tosses, die rolls, card selections, and clearance rates. You
might have noticed that clearance rates stand apart from the other three examples—unlike coins, dice, and
cards, the probability of clearance is not equal across the different crime types. Whereas a rolled die offers
equal probability of landing on two versus six (each one is .17), a crime will vary in its clearance probability
based on the type of crime that it is. Motor vehicle theft has a .13 clearance probability, and homicide has a
.62 chance of being cleared. Unlike each of the six sides of a die, any given crime does not have an equal
probability of clearance; instead, a crime’s probability of being solved depends in large part on the type of
crime it is.
This leads us to an important point: In most cases, the different possible outcomes have unequal probabilities
of occurring. The existence of outcomes that have greater or lesser theoretical probabilities of being the one
that actually occurs forms the basis for everything else we are going to discuss in this chapter. Put simply,
some outcomes are more likely than others.
180
Learning Check 6.4
The general population is roughly 50% female, meaning that approximately 50% of newborns are female. Suppose there are two hospitals,
Hospital A and Hospital B. In Hospital A, four babies are born each day. In Hospital B, 20 babies are born per day. In which hospital
would you expect to see roughly half of the babies born each day be female? That is, in which hospital would the sex of the babies born
each day be representative of the entire population? Explain your answer. What about over the course of a year? If you tracked the sex of
babies born for a year, would there be a difference between the two hospitals’ female-baby percentages? Why or why not?
A probability distribution is a table or graph showing the full array of theoretical probabilities for any given
variable. These probabilities represent not what we actually see (i.e., tangible empirical outcomes) but, rather,
the gamut of potential empirical outcomes and each outcome’s probability of being the one that actually
happens. Probability distributions are theoretical. A probability distribution is constructed on the basis of an
underlying parameter or statistic (such as a proportion or a mean) and represents the probability associated
with each possible outcome. Two types of probability distributions are discussed in this chapter: binomial and
continuous.
Probability distribution: A table or graph showing the entire set of probabilities associated with every possible empirical outcome.
181
Discrete Probability: The Binomial Probability Distribution
A trial is a particular act with multiple different possible outcomes (e.g., rolling a die, where the die will land
on any one of its six sides). Binomials are trials that have exactly two possible outcomes; this type of trial is
also called dichotomous or binary. Coin flips are binomials because coins have two sides. Research Example 6.1
describes two types of binomials that criminal justice and criminology researchers have examined. Binomials
are used to construct binomial probability distributions. The binomial probability distribution is a list of
expected probabilities; it contains all possible results over a set of trials and lists the probability of each result.
Trial: An act that has several different possible outcomes.
Binomial: A trial with two possible outcomes. Also called a dichotomous or binary variable empirical outcome.
Binomial probability distribution: A numerical or graphical display showing the probability associated with each possible outcome of
a trial.
So, how do we go about building the binomial probability distribution? The distribution is constructed using
the binomial coefficient. The formula for this coefficient is a bit intimidating, but each component of the
coefficient will be discussed individually in the following pages so that when you are done reading, you will
understand the coefficient and how to use it. The binomial coefficient is given by the formula
where
p (r) = the probability of r ,
r = the number of successes,
N = the number of trials or sample size,
p = the probability that a given event will occur, and
q = the probability that a given event will not occur.
Binomial coefficient: The formula used to calculate the probability for each possible outcome of a trial and to create the binomial
probability distribution.
Research Example 6.1 Are Police Officers Less Likely to Arrest an Assault Suspect When the Suspect and the Alleged Victim Are
Intimate Partners?
Critics of the police response to intimate partner violence have accused police of being “soft” on offenders who abuse intimates.
Klinger (1995) used a variable measuring whether or not police made an arrest when responding to an assault of any type. The
variable was coded as arrest/no arrest. He then examined whether the probability of arrest was lower when the perpetrator and victim
were intimates as compared to assaults between strangers or non-intimate acquaintances. The results indicated that police were
unlikely to make arrests in all types of assault, regardless of the victim–offender relationship and that they were not less likely to
arrest offenders who victimized intimate partners relative to those who victimized strangers or acquaintances.
182
183
Are ex-prisoners returning to society less likely to recidivate if they are
given reentry services?
Researchers, corrections agencies, and community groups constantly seek programming that reduces recidivism among recently
released prisoners by helping them transition back into society. The reentry process is difficult—newly released ex-prisoners struggle
to find housing, transportation, and employment. Many have ongoing problems with substance dependence. Ray, Grommon,
Buchanan, Brown, and Watson (2017) evaluated recidivism outcomes for clients of a federally funded program implemented across
five organizations in Indiana that provided reentry services to adult ex-prisoners with histories of drug or alcohol abuse. Recidivism
was coded as yes/no to indicate whether each person was returned to prison after completing the program. Approximately one-third
recidivated. There were significant differences in recidivism rates across the five agencies, suggesting that better-funded agencies
combining multiple different therapeutic and social-support services produce lower recidivism rates compared to those with fewer
resources that offer a narrow range of services.
The ultimate goal of binomials is to find the value of p(r) for every possible value of r. The resulting list of p(r)
values is the binomial probability distribution.
Before we get into the math, let’s first consider a conceptual example using the clearance data. Table 6.2
shows that 61.50% of homicides result in arrest, meaning that any given homicide has a .62 probability of
clearance. Suppose you gathered a random sample of 10 homicide cases. Out of these 10, there are 11 separate
possible (i.e., theoretical) outcomes: Anywhere from none of them to all 10 could have been cleared by arrest.
Each of these individual outcomes has a certain probability of being the one that occurs in reality. You might
have already guessed that the most likely outcome is that 6 of them would be cleared (since 61.50% of 10 is
6.25), but what are the other possible outcomes’ probabilities of occurrence? This is what we use the binomial
probability distribution for.
184
Successes and Sample Size: N and r
The binomial coefficient formula contains the variables N and r. The N represents sample size or the total
number of trials. As an example, we will use a hypothetical study of jail inmates who have been booked on
arrest and are awaiting word as to whether they will be released on bail or whether they will remain confined
while they await criminal-court proceedings. The Bureau of Justice Statistics (BJS) reports that 62% of felony
defendants are released from jail prior to the final disposition of their case (Reaves, 2013), so p = .62. Let’s say
we draw a random sample of five recently arrested jail inmates. This means that for the binomial coefficient
formula, N = 5.
In the binomial coefficient, r represents the number of successes or, in other words, the number of times the
outcome of interest happens over N trials. A researcher decides what “success” will mean in a given study. In
the current example, we will define success as a defendant obtaining pretrial release. There are multiple values
that r takes on in any given study, since there are multiple possible successes. If we wanted to find the
probability that three of the five defendants would be released, we would input r = 3 into the binomial
coefficient formula; if we wanted the probability that all five would be released, then r = 5.
Success: The outcome of interest in a trial.
185
The Number of Ways r Can Occur, Given N : The Combination
In our sample of five defendants, there are a lot of possibilities for any given value of r. For instance, if one
defendant was released (i.e., r = 1), then it could be that Defendant 1 was released and the remaining four
were detained. Alternatively, Defendant 4 might have been the one who made bail. The point is, a given value
of r can occur in multiple different ways. If a release is symbolized r , then, for an illustration, let us call a
detention d. (Remember that release and detention are binary; each defendant receives one of these two
possible outcomes.) Consider the following sets of possible arrangements of outcomes for one release and four
detentions:
{r , d , d , d , d } {d , r , d , d , d } {d , d , r , d , d } {d , d , d , r , d } {d , d , d , d , r }
What these sets tell you is that there are five different ways for r = 1 (i.e., one success) to occur over N = 5
trials. The same holds true for any number of successes—there are many different ways that r = 2, r = 3, and
so on can occur in terms of which defendants are the successes and which are the failures.
The total number of ways that r can occur in a sample of size N is called a combination and is calculated as
where
N = the total number of trials or total sample size,
r = the number of successes, and
! = factorial.
Combination: The total number of ways that a success r can occur over N trials.
186
Learning Check 6.5
In this and subsequent chapters, you will see various notations representing multiplication. The most popular symbol, × , will not be used;
in statistics, x represents raw data values, and so using this symbol to also represent multiplication would be confusing. Instead, we will
rely on three other indicators of multiplication. First are parentheses; numbers separated by parentheses should be multiplied together.
This could appear as 2(3) = 6, or (2)(3) = 6. Second, sometimes no operand is used, and it is merely the placement of two numbers or
symbols right next to each other that signals multiplication. An example of this is xy , where whatever numbers x and y symbolized,
respectively, would be multiplied together. If x = 2 and y = 3, then xy = 2(3) = 6. Third is the centered dot that connects the numbers
being multiplied, such as 2 ⋅ 3 = 6.
187
Learning Check 6.6
Most calculators meant for use in math classes (which excludes the calculator that came free with that new wallet you bought) will
compute factorials and combinations for you. The factorial function is labeled as !. The combination formula is usually represented by nCr
or sometimes just C. Depending on the type of calculator you have, these functions are probably either located on the face of the calculator
and accessed using a 2nd or alpha key, or can be found in a menu accessed using the math, stat , or prob buttons, or a certain combination
of these functions. Take a few minutes right now to find these functions on your calculator.
Also, note that 1 factorial and 0 factorial both equal 1. Try this out on your calculator to see for yourself. Since 0! = 1, you do not need to
worry if you end up with a zero in the denominator of your fractions.
1! = 1
0! = 1
The factorial symbol ! tells you to multiply together the series of descending whole numbers starting with the
number to the left of the symbol all the way down to 1.00. If r = 3, then r ! = 3 ⋅ 2 ⋅ 1 = 6. If N = 5, then N ! =
5 ⋅ 4 ⋅ 3 ⋅ 2 ⋅ 1 = 120.
Factorial: Symbolized ! , the mathematical function whereby the first number in a sequence is multiplied successively by all numbers
below it down to 1.00.
There is also shorthand notation for the combination formula that saves us from having to write the whole
thing out. The shorthand notation is , which is pronounced “N choose r.” Be very careful—this is not
a fraction! Do not mistake it for “N divided by r .”
The combination formula can be used to replicate the previously presented longhand demonstration wherein
we were interested in the number of ways that one release and four detentions can occur. Plugging the
numbers into the combination formula yields
When we wrote out the possible ways for r = 1 to occur, we concluded that there were five options; we have
now confirmed this mathematically using the combination formula. There are five different ways for one
person to be released and the remaining four to be detained.
We will do one more example calculation before moving on. Suppose that three of the five defendants were
released. How many combinations of three are there in a sample of five? Plug the numbers into the
combination formula to find out:
188
There are 10 combinations of three in a sample of five. In the context of the present example, there are 10
different ways for three defendants to be released and two to be detained. This has to be accounted for in the
computation of the probability of each possible result, which is why you see the combination formula included
in the binomial coefficient.
189
The Probability of Success and the Probability of Failure: p and q
The probability of success (symbolized p) is at the heart of the binomial probability distribution. This number
is obtained on the basis of prior knowledge or theory. In the present example pertaining to pretrial release, we
know that 62% of felony defendants obtain pretrial release. This means that each defendant’s probability of
release is p = .62.
We know that release is not the only possible outcome, though—defendants can be detained, too. The
opposite of a success is a failure. Because we are dealing with events that have two potential outcomes, we
need to know the probability of failure in addition to that of success. The probability of failure is represented
by the letter q ; to compute the value of q , the bound rule and the rule of the complement must be invoked. In
Chapter 5, you learned that
p (A) + p (Not A) = 1.00; therefore,
p (Not A) = 1.00 – p (A) .
Failure: Any outcome other than success or the event of interest.
What we are doing now is exactly the same thing, with the small change that p (A) is being changed to simply
p , and p (Not A) will now be represented by the letter q. So,
With q = .62, the probability that a felony defendant will be detained prior to trial (i.e., will not obtain pretrial
release) is
q = 1.00 – .62 = .38
190
Putting It All Together: Using the Binomial Coefficient to Construct the
Binomial Probability Distribution
Using p and q , the probability of various combinations of successes and failures can be computed. When there
are r successes over N trials, then there are N – r failures over that same set of trials. There is a formula called
the restricted multiplication rule for independent events that guarantees that the probability of r successes and
N – r failures is the product of p and q. In the present example, there are three successes (each with probability
p = .62) and two failures (each with probability q = .38). You might also recall from prior math classes that
exponents can be used as shorthand for multiplication when a particular number is multiplied by itself many
times. Finally, we also have to account for the combination of N and r. The probability of three successes is
thus
= 10(.24)(.14)
= .34
Restricted multiplication rule for independent events: A rule of multiplication that allows the probability that two events will both
occur to be calculated as the product of each event’s probability of occurrence: That is, p (A and B) = p (A) · p (B).
This result means that given a population probability of .62, there is a .34 probability that if we pulled a
sample of five felony defendants, three would obtain pretrial release.
To create the binomial probability distribution, repeat this procedure for all possible values of r. The most
straightforward way to do this is to construct a table like Table 6.3. Every row in the table uses a different r
value, whereas the values of N , p , and q are fixed. The rightmost column, p(r) , is obtained by multiplying the
191
, p r, and q N –r terms.
The values in the p(r) column tell you the probability of each possible outcome being the one that actually
occurs in any given random sample of five defendants. The probability that two of the five defendants will be
released pending trial is .19, and the probability that all five will be released is .09. Probabilities can also be
added together. The probability of four or more defendants being released is p (4) + p (5) = .29 + .09 = .38. The
chance that two or fewer would be released is p (2) + p (1) + p (0) = .19 + .06 + .01 = .26.
Our original question was, “With a population probability of .62, what is the probability that three of five
defendants would be released?” You can see from the table that the answer is .34. This, as it happens, is the
highest probability in the table, meaning that it is the outcome that we would most expect to see in any given
random sample. Another way to think of it is to imagine that someone who was conducting this study asked
you to predict—based on a population probability of .62 and N = 5 defendants—how many defendants would
be released. Your “best guess” answer would be three, because this is the outcome with the highest likelihood
of occurring.
Contrast this to the conclusion we would draw if none of the five defendants in the sample had been released.
This is an extremely improbable event with only p (0) = .01 likelihood of occurring. It would be rather
surprising to find this empirical result, and it might lead us to wonder if there was something unusual about
our sample. We might investigate the possibility that the county we drew the sample from had particularly
strict policies regulating pretrial release or that we happened to draw a sample of defendants charged with
especially serious crimes.
192
There are two neat and important characteristics of the binomial probability distribution. The first is that it
can be graphed using a bar chart (Figure 6.1) so as to form a visual display of the numbers in Table 6.3. The
horizontal axis contains the r values and the bar height (vertical axis) is determined by p(r) . The bar chart
makes it easy to determine at a glance which outcomes are most and least likely to occur.
The second important thing about the binomial probability distribution is that the p(r) column sums to 1.00.
This is because the binomial distribution is exhaustive; that is, all possible values of r are included in it. Any
time an exhaustive list of probabilities is summed, the result will be 1.00. In Table 6.3, we covered all of the
values that r can assume (i.e., zero to five); therefore, all possible probabilities are included, and the sum is
1.00. Memorize this point! It is applicable in the context of continuous probability distributions, too, and will
be revisited shortly.
Figure 6.1 The Binominal Probability Distribution for Pretrial Release in a Sample of Five Defendants, With
p = .62
193
Learning Check 6.7
Recall from Table 6.2 and the discussion that 61.50% of homicides are cleared by arrest. Use this clearance rate to construct the binomial
probability distribution for a hypothetical sample of six homicide cases.
194
Continuous Probability: The Standard Normal Curve
The binomial probability distribution is applicable for trials with two potential outcomes; in other words, it is
used for dichotomous categorical variables. Criminal justice and criminology researchers, however, often use
continuous variables. The binomial probability distribution has no applicability in this context. Continuous
variables are represented by a theoretical distribution called the normal curve. Figure 6.2 shows this curve’s
familiar bell shape.
Normal curve: A distribution of raw scores from a sample or population that is symmetric and unimodal and has an area of 1.00.
Normal curves are expressed in raw units and differ from one another in metrics, means, and standard deviations.
Figure 6.2 The Normal Curve for Continuous Variables
195
Learning Check 6.8
We have already visited the concept of a normal distribution and have encountered four ways that a curve can deviate from normality. Can
you remember those four deviation shapes? Revisit Chapters 4 and 5, if needed.
The normal curve is a unimodal , symmetric curve with an area of 1.00. It is unimodal because it peaks once
and only once. In other words, it has one modal value. It is symmetric because the two halves (split by the
mean) are identical to one another. Additionally, it has an area of 1.00 because it encompasses all possible
values of the variable in question. Just as the binomial probability distribution’s p (r) column always sums to
1.00 because all values that r could possibly take on are contained within the table, so too the normal curve’s
tails stretch out to negative and positive infinity. This might sound impossible, but remember that this is a
theoretical distribution. This curve is built on probabilities, not on actual data.
Research Example 6.2 What Predicts Correctional Officers’ Job Stress and Job Satisfaction?
Correctional officers work in high-stress environments, and their job performance has implications for the quality and safety of the
correctional institution. High stress levels can reduce these workers’ performance levels and can increase their likelihood of physical
injury; likewise, low job satisfaction can lead to constant staff turnover and to burnout, which also dampens the overall effectiveness
of the correctional institution.
Paoline, Lambert, and Hogan (2006) sought to uncover the predictors of correctional officers’ attitudes toward their jobs. They
gathered a sample of jail security staff and administered surveys that asked these respondents several questions about the levels of
stress they experience at work and the amount of satisfaction they derive from their job. There were six stress variables and five
satisfaction variables. Each set of variables was summed to form a single, overarching score on each index for each respondent. The
indexes were continuous. Jail staff members with higher scores on the stress index experienced greater work anxiety and tension;
likewise, those with lower scores on the satisfaction index had relatively poor feelings about their job.
Paoline et al. (2006) found that the most consistent predictors of both stress and satisfaction were organizational factors specific to
the jail itself; in particular, officers who felt that that the jail policies were clear and fair and who had positive views toward their
coworkers experienced significantly less stress and greater satisfaction as compared to those officers who were not happy about the
policies and their coworkers. These findings suggest that jail managers who seek to foster a positive work environment for their
employees should ensure clear, fair policies and should promote harmony and teamwork among jail staff.
The characteristics that determine a normal curve’s location on the number line and its shape are its mean and
standard deviation, respectively. When expressed in raw units, normal curves are scattered about the number
line and take on a variety of shapes. This is a product of variation in metrics, means, and standard deviations.
Figure 6.3 depicts this variation.
This inconsistency in locations and dimensions of normal curves can pose a problem in statistics. It is
impossible to determine the probability of a certain empirical result when every curve differs from the rest.
What is needed is a way to standardiz e the normal curve so that all variables can be represented by a single
curve. This widely applicable single curve is constructed by converting all of a distribution’s raw scores to z
scores. The z -score transformation is straightforward:
196
where
zx = the z score for a given raw score x ,
x = a given raw score,
x̄ = the distribution mean, and
s = the distribution standard deviation.
z score: A standardized version of a raw score that offers two pieces of information about the raw score: (1) how close it is to the
distribution mean and (2) whether it is greater than or less than the mean.
Figure 6.3 Variation in Normal Curves: Different Means and Standard Deviations
A z score conveys two pieces of information about the raw score on which the z score is based. First, the
absolute value of the z score reveals the location of the raw score in relation to the distribution mean. Z scores
are expressed in standard deviation units. A z score of .25, for example, tells you that the underlying raw score
is one-fourth of one standard deviation away from the mean. A z score of –1.50, likewise, signifies a raw score
that is one-and-one-half standard deviations away from the mean.
197
Learning Check 6.9
It is very important that you understand standard deviations; z scores will not make much sense if you did not fully grasp this earlier
concept from Chapter 5. If necessary, return to that chapter for a review. What is a standard deviation, conceptually? What two pieces of
complementary information are given by the mean and the standard deviation?
The second piece of information is given by the sign of the z score. Although standard deviations are always
positive, z scores can be negative. A z score’s sign indicates whether the raw score that the z score represents is
greater than the mean (producing a positive z score) or is less than the mean (producing a negative z score). A
z score of .25 is above the mean, while a score of –1.50 is below it. Figure 6.4 shows the relationship between
raw scores and their z -score counterparts. You can see that every raw score has a corresponding z score. The z
score tells you the distance between the mean and an individual raw score, as well as whether that raw score is
greater than or less than the mean.
Figure 6.4 Raw Scores and z Scores
When all of the raw scores in a distribution have been transformed to z scores and plotted, the result is the
standard normal curve. The z -score transformation dispenses with the original, raw values of a variable and
replaces them with numbers representing their position relative to the distribution mean. A normal curve,
then, is a curve constructed of raw scores, while the standard normal curve is composed of z scores.
Standard normal curve: A distribution of z scores. The curve is symmetric and unimodal and has a mean of zero, a standard
deviation of 1.00, and an area of 1.00.
Like ordinary normal curves, the standard normal curve is symmetric, unimodal, and has an area of 1.00.
Unlike regular normal curves, though, the standard normal curve’s mean and standard deviation are fixed at 0
and 1, respectively. They remain constant regardless of the units in which the variable is measured or the
original distribution’s mean and standard deviation. This allows probabilities to be computed.
To understand the process of using the standard normal curve to find probabilities, it is necessary to
comprehend that, in this curve, area is the same as proportion and probability. A given area of the curve (such
as the area between two raw scores) represents the proportion of scores that are between those two raw values.
Figures 6.5 and 6.6 display the relationship between z scores and areas.
In Chapter 5, you learned that approximately two-thirds of the scores in any normal distribution lie between
198
one standard deviation below and one standard deviation above the mean for that set of scores. (Refer back to
Figure 5.4.) In Figure 6.5, you can see that .3413 (or 34.13%) of the scores are located between the mean and
one standard deviation. If you add the proportion of cases (i.e., area under the curve) that is one standard
deviation below the mean to the proportion or area that is one standard deviation above, you get .3413 +
.3413 = .6826, or 68.26%. This is just over two-thirds! The bulk of scores in a normal distribution cluster
fairly close to the center, and those scores that are within one standard deviation of the mean (i.e., z scores
that have absolute values of 1.00 or less) are considered the typical or normal scores. This confirms what we
saw in Chapter 5 when we talked about the “normal” range being between 1 sd above and 1 sd below the
mean.
Z scores that are greater than 1.00 or less than –1.00 are relatively rare, and they get increasingly rare as you
trace the number line away from zero in either direction. These very large z scores do happen, but they are
improbable, and some of them are incredibly unlikely. In Figure 6.6, you can see that a full 95% of scores (or
.9544, to be exact) are within two standard deviations above and below the mean. In other words, only about
5% of scores in a normal distribution will be either greater than 2 sd above the mean (i.e., have a z value
greater than 2.00) or more than 2 sd below the mean (a z value less than –2.00).
This is all getting very abstract, so let us get an example going. According to the Census of Jails (COJ; see
Data Sources 3.1), medium-sized jails have a mean of 23.54 full-time correctional officers. The standard
deviation is 17.66.
Suppose that a facility has 19 officers. To find this jail’s z score, plug the numbers into Formula 6(5):
This facility is approximately one-fourth of one standard deviation below the group mean on staffing.
Substantively speaking, this facility is inside the typical zone of ±1 sd .
Figure 6.5 Standard Normal Curve: Area Between the Mean and One Standard Deviation
Figure 6.6 Standard Normal Curve: Area Between the Mean and Two Standard Deviations
199
Now for another one. One institution employed 55 correctional officers, so its z score is
This institution is roughly one-and-three-fourths standard deviations above the mean, well outside the typical
zone.
One more example. Suppose a facility has 80 officers. Its z score is thus
This jail’s raw score is three-and-one-fifths standard deviations greater than the sample mean, which is very
far out in the right-hand tail of the distribution.
Let’s draw a hypothetical plot that offers a visualization of the three raw and z scores used in these examples
and each one’s relationship to the mean. You can see in Figure 6.7 each pair of x ’s and z ’s and how close to
or far from the mean they are, as well as whether they are above or below it. The larger the z score is, the
farther away from the mean that score is.
200
The z Table and Area Under the Standard Normal Curve
It is also possible to use z scores to find the area (i.e., proportion of values) between the mean and a particular
score, or even the area that is beyond that score (i.e., the area that is in the tail of the distribution).
Figure 6.7 Raw and z Scores in Relation to the Sample Mean
To do this, the z table is used. The z table is a chart containing the area of the curve that is between the mean
and a given z score. Appendix B contains the z table. The area associated with a particular z score is found by
decomposing the score into an x.x and .0x format such that the first half of the decomposed score contains the
digit and the number in the 10ths position, and the second half contains a zero in the 10ths place and the
number that is in the 100ths position. In the three previous examples, we calculated z scores of –.26, 1.78, and
3.20. The first z score would be broken down as –.2 + .06. Note that there are no negative numbers in the z
table, but this does not matter because the standard normal curve is symmetric, and, therefore, the z table is
used the same way irrespective of an individual score’s sign. Go to the z table and locate the .2 row and .06
column; then trace them to their intersection. The area is .1026. (Area cannot be negative, so the area
between the mean and any given z score is positive even if the score itself is negative.) This means that
approximately .10 or 10% of the scores in the distribution sit between the mean and a raw score of 19.
z table: A table containing a list of z scores and the area of the curve that is between the distribution mean and each individual z
score.
The z score of 1.78 would decompose as 1.7 + .08. Locating the 1.7 row and .08 column, we can see that the
area is .4625. This tells us that roughly .46 or 46% of scores lie between the mean and a raw score of 55.
Lastly, to find the area for a z score of 3.20, we look at the 3.2 row and .00 column. Here we hit a bump
because the z table does not have a row for 3.2; it has only 3.0 and 3.5. We will select 3.0 because 3.2 is closer
to that number than it is to 3.5. Using 3.0 and .00 yields an area of .4987. In other words, approximately .50
or 50% of scores in the distribution sit between the mean and a raw score of 80. This makes sense because
there is a very large gap between the mean (23.54) and 80, so lots of raw scores will lie within this range.
Figure 6.8 shows the areas between the mean and each of our three z scores. The area between the mean and
z = 3.20 is displayed with a bracket because this area encompasses the area between the mean and z = 1.78.
201
Figure 6.8 Areas Between the Mean and z
Finding the area between the mean and a z score is informative, but what is generally of interest to
criminology and criminal justice researchers is the area beyond that z score; that is, researchers usually want to
know how big or small a z score would have to be in order to fall into the very tip of either the positive or
negative tail of the distribution. To do this, we rely on a simple fact: Because the entire area of the standard
normal curve is 1.00, the mean splits the curve exactly in half so that 50% of scores are below it and 50% are
above it. As such, we have two pieces of information. First, we know that the total area on each side is .50.
Second, for any individual z score, we know the area between the mean and that score. We are looking for the
third, unknown number, which is the area beyond that z score. How do you think we can find that number? If
you said, “Subtraction,” you are right! We subtract the known area from .50 to figure out what’s left over.
Let us find the area beyond z for z = –.26. The area between the mean and z is .1026 and the total area on the
left-hand side is .50, so we use subtraction, as such:
.50 – .1026 = .3974
Thus, .3974 (or approximately 40%, using rounding) of the raw scores in the standard normal distribution for
staff in medium-sized jails are less than 19 (alternatively, less than z = –.26).
How about for a staff size of 55 officers (z = 1.78)? We found that the area between the mean and z is .4625,
so
.50 – .4625 = .0375
Approximately .04 (or 4%) of raw scores are greater than 55. This is a small proportion of scores because 55 is
so far away from the mean (nearly 2 standard deviations), which leaves very little room for scores exceeding
202
that number. Figure 6.9 depicts each of these calculated z scores’ location on the standard normal curve.
203
Learning Check 6.10
Before moving on, double-check your understanding of areas between the mean and z and of areas beyond z by finding both of these
areas for each of the following z scores:
1. z = 1.38
2. z = –.65
3. z = 2.46
4. z = –3.09
Figure 6.9 Areas Between the Mean and z , Beyond z , and Total Area for Two z Scores
Another handy feature of the standard normal curve is that just as z scores can be used to find areas, areas can
likewise be used to find z scores. Basically, you just use the z table backward. We might, for instance, want to
know the z scores that lie in the upper 5% (.05) or 1% (.01) of the distribution. These are very unlikely values
and are of interest because of their relative rarity. How large would a z score have to be such that only .05 or
.01 of scores is above it?
The process of finding these z scores employs similar logic about the area of each side of the curve. First, we
know that the z table provides the area between z and the mean; it does not tell you about the area beyond z ,
so it cannot be used yet. This problem is surmounted using subtraction, as we have already seen. Let us start
with the area of .01. The area between the mean and the z score that cuts the distribution at .01 is
.50 – .01 = .49
Thus, the area between the mean and the z score we are searching for is .49. Since we know this, we can now
use the z table.
To do this, scan the areas listed in the body of the z table to find the one that is closest to .49. The closest area
is .4901. Third, find the z score associated with the identified area. Instead of tracing the two elements of a z
score inward to locate the area, as you did before, now start at the area and trace outward along the row and
column. The area of .4901 is in the 2.3 row and .03 column, so the z score is
2.3 + .03 = 2.33
204
And that is the answer! We now know that .01 (or 1%) of scores in the standard normal distribution area have
z scores greater than 2.33.
205
Learning Check 6.11
The text uses the example of finding the z scores that are in the top 1% of the standard normal distribution. What z scores are, likewise,
in the bottom 1%? Explain how you arrived at your answer. (Hint: You do not need to do any math to answer this question; the
calculations have already been done in the text.)
Now let us find the z score associated with an area of .05, but with a twist: We will place this area in the lower
(i.e., left-hand) tail of the distribution. Recall that we were working in the upper (right-hand) tail when we
did the previous example using an area of .01. The first and second steps of the process are the same no matter
which side of the distribution is being analyzed. Subtraction shows that .50 – .05 = .45 of the scores is
between the mean and z. Going to the table, you can see that there are actually two areas that are closest to
.45. They are .4495 and .4505. The z score for the former is 1.64 and for the latter is 1.65; however , since we
are on the left (or negative) side of the curve, these z scores are actually –1.64 and –1.65. Be aware of the sign
of your z score! Scores on the left side of the standard normal curve are negative, so you have to add a negative
sign to the score once you locate it using the table.
Finally, since there are two z scores in this instance, they must be averaged:
That is the answer! In the standard normal distribution, z scores less than –1.65 have a .05 or less probability
of occurring. In other words, these extremely small scores happen 5% of the time or less. That is unlikely
indeed.
206
Learning Check 6.12
There are two points you should be very clear on before leaving this chapter. First, although z scores can be negative, areas are always
positive. When you are working on the left side of the standard normal z curve, scores will be negative but areas will not. Second,
areas/probabilities and z scores are two ways of expressing the same idea. Large z scores are associated with small probabilities, and small
z scores represent large probabilities. The larger the absolute value of a z score is, the smaller the likelihood of observing that score will be.
Z scores near z ero (i.e., near the mean of the standard normal curve) are not unlikely or unusual, whereas scores that are very far away
from z ero are relatively rare; they are out in the far left or far right tail. Take a moment now to graph the four z values listed in Learning
Check 6.10.
Chapter Summary
Probability is the basis of inferential statistics. This chapter introduced two of the major probability distributions: binomial and
standard normal. The binomial distribution is for categorical variables that have two potential outcomes (dichotomous or binary
variables), and the standard normal curve applies to continuous variables. The binomial probability distribution is constructed using
an underlying probability derived from research or theory. This distribution shows the probability associated with each possible
outcome, given an overarching probability of success and a predetermined number of cases or trials.
The standard normal curve consists of z scores, which are scores associated with known areas or probabilities. Raw scores can be
transformed to z scores using a simple conversion formula that employs the raw score, the mean, and the standard deviation. Because
the area under the standard normal curve is a constant 1.00 (i.e., .50 on each side), areas can be added and subtracted, thus allowing
probabilities to be determined on the basis of z scores and vice versa.
Both distributions are theoretical, which means that they are constructed on the basis of logic and mathematical theory. They can be
contrasted to empirical distributions, which are distributions made from actual, observed raw scores in a sample or population.
Empirical distributions are tangible; they can be manipulated and analyzed. Theoretical distributions exist only in the abstract.
Thinking Critically
1. Suppose that in a particular state, female offenders sentenced to prison received a mean sentence length of 30 months (s = 12)
and male offenders’ mean sentence length was 40 months (s = 10). If you randomly selected a female prison inmate and
found that she had been sentenced to 24 months, would you be surprised? What about if she was serving a 45-month
sentence? If you randomly selected a male inmate who had a 60-month sentence, would you be surprised to find this result?
What about if his sentence was 31 months? Explain all your answers.
2. Earlier in the chapter, it was shown that in a 52-card deck, any given card has a chance of being selected
in a random draw. How would this probability change after a card was drawn and not replaced? Calculate the successive
changes in probability as one, two, three, four, and five cards are taken from the deck. Does the probability for each
remaining card increase or decrease? Why?
Review Problems
1. Eight police officers are being randomly assigned to two-person teams.
1. Identify the value of N.
2. Identify the value of r.
3. How many combinations of r are possible in this scenario? Do the combination by hand first, and then check your
answer using the combination function on your calculator.
2. Correctional staff are randomly assigning nine jail inmates to three-person cells.
207
1. Identify the value of N.
2. Identify the value of r.
3. How many combinations of r are possible in this scenario? Do the combination by hand first, and then check your
answer using the combination function on your calculator.
3. In a sample of seven parolees, three are rearrested within 2 years of release.
1. Identify the value of N.
2. Identify the value of r.
3. How many combinations of r are possible in this scenario? Do the combination by hand first, and then check your
answer using the combination function on your calculator.
4. Out of six persons recently convicted of felonies, five are sentenced to prison.
1. Identify the value of N.
2. Identify the value of r.
3. How many combinations of r are possible in this scenario? Do the combination by hand first, and then check your
answer using the combination function on your calculator.
5. Four judges in a sample of eight do not believe that the law provides them with sufficient sanction options when sentencing
persons convicted of crimes.
1. Identify the value of N.
2. Identify the value of r.
3. How many combinations of r are possible in this scenario? Do the combination by hand first, and then check your
answer using the combination function on your calculator.
6. For each of the following variables, identify the distribution—either binomial or standard normal—that would be the
appropriate theoretical probability distribution to represent that variable. Remember that this is based on the variable’s level
of measurement.
1. Defendants’ completion of a drug court program, measured as success or failure
2. The total lifetime number of times someone has been arrested
3. Crime victims’ reporting of their victimization to police, measured as reported or did not report
7. For each of the following variables, identify the distribution—either binomial or standard normal—that would be the
appropriate theoretical probability distribution to represent that variable. Remember that this is based on the variable’s level
of measurement.
1. The number of months of probation received by juveniles adjudicated guilty on delinquency charges
2. City crime rates
3. Prosecutorial charging decisions, measured as filed charges or did not file charges
8. According to the UCR, 56% of aggravated assaults reported to police are cleared by arrest. Convert this percentage to a
proportion and use it as your value of p to do the following:
1. Compute the binomial probability distribution for a random sample of five aggravated assaults, with r defined as the
number of assaults that are cleared.
2. Based on the distribution, what is the outcome (i.e., number of successes) you would most expect to see?
3. Based on the distribution, what is the outcome (i.e., number of successes) you would least expect to see?
4. What is the probability that two or fewer aggravated assaults would be cleared?
5. What is the probability that three or more would be cleared?
9. According to the BJS, of all criminal charges filed against defendants for domestic violence, 62% are for aggravated assault.
Convert this percentage to a proportion and use it as your value of p to do the following:
1. Compute the binomial probability distribution for a random sample of four domestic-violence cases, with r defined as
the number of cases charged as aggravated assault.
2. Based on the distribution, what is the outcome (i.e., number of successes) you would most expect to see?
3. Based on the distribution, what is the outcome (i.e., number of successes) you would least expect to see?
4. What is the probability that two or fewer of the charges would be for aggravated assault?
5. What is the probability that three or more of them would be for assault?
10. According to the UCR, 49% of the hate crimes that were reported to police are racially motivated. Convert this percentage to
a proportion and use it as your value of p to do the following:
1. Compute the binomial probability distribution for a random sample of six hate crimes, with r defined as the number
that are racially motivated.
208
2. Based on the distribution, what is the outcome (i.e., number of successes) you would most expect to see?
3. Based on the distribution, what is the outcome (i.e., number of successes) you would least expect to see?
4. What is the probability that two or fewer of the hate crimes were motivated by race?
5. What is the probability that three or more were racially motivated?
11. According to the UCR, 61% of murders are committed with firearms. Convert this percentage to a proportion and use it as
your value of p to do the following:
1. Compute the binomial probability distribution for a random sample of five murders, with r defined as the number
that are committed with firearms.
2. Based on the distribution, what is the outcome (i.e., number of successes) you would most expect to see?
3. Based on the distribution, what is the outcome (i.e., number of successes) you would least expect to see?
4. What is the probability that one or fewer murders were committed with firearms?
5. What is the probability that four or more murders were committed with firearms?
The Law Enforcement Management and Administrative Statistics (LEMAS) survey reports that the mean number
of municipal police per 1,000 local residents in U.S. cities with populations of 100,000 or more was 1.99 (s = .84).
Use this information to answer questions 12 through 15.
12. One department had 2.46 police per 1,000 residents.
1. Convert this raw score to a z score.
2. Find the area between the mean and z.
3. Find the area in the tail of the distribution beyond z.
13. One department had 4.28 police per 1,000 residents.
1. Convert this raw score to a z score.
2. Find the area between the mean and z.
3. Find the area in the tail of the distribution beyond z.
14. One department had 1.51 police per 1,000 residents.
1. Convert this raw score to a z score.
2. Find the area between the mean and z
3. Find the area in the tail of the distribution beyond z.
15. One department had 1.29 officers per 1,000 residents.
1. Convert this raw score to a z score.
2. Find the area between the mean and z.
3. Find the area in the tail of the distribution beyond z.
16. What z scores fall into the upper .15 of the distribution?
17. What z scores fall into the upper .03 of the distribution?
18. What z scores fall into the lower .02 of the distribution?
19. What z scores fall into the lower .10 of the distribution?
20. What z scores fall into the lower .015 of the distribution?
209
Key Terms
Inferential statistics 134
Probability theory 134
Probability 135
Theoretical prediction 138
Empirical outcome 138
Probability distribution 140
Trial 140
Binomial 140
Binomial probability distribution 140
Binomial coefficient 141
Success 142
Combination 144
Factorial 144
Failure 144
Restricted multiplication rule for independent events 145
Normal curve 148
z score 150
Standard normal curve 151
z table 154
Glossary of Symbols and Abbreviations Introduced in This Chapter
210
Chapter 7 Population, Sample, and Sampling Distributions
211
Learning Objectives
Explain the difference between empirical and theoretical distributions.
Define population, sample, and sampling distributions and identify each as either empirical or theoretical.
Explain the difference between statistics and parameters.
Define sampling error and explain how it affects efforts to generalize from a sample to a population.
Define the central limit theorem.
Describe the z and t distributions, including whether these are empirical or theoretical distributions and which one is appropriate
depending on sample size.
A population is the entire universe of objects, people, places, or other units of analysis that a researcher wishes
to study. Criminal justice and criminology researchers use all manner of populations. Bouffard (2010), for
instance, examined the relationship between men’s military service during the Vietnam era and their criminal
offending later in life. Rosenfeld and Fornango (2014) assessed how effective order maintenance policing is at
reducing robbery and burglary at the neighborhood level. Morris and Worrall (2010) investigated whether
prison architectural design influences inmate misconduct. These are three examples of populations—male
Vietnam veterans, neighborhoods, prison inmates—that can form the basis for study.
The problem is that populations are usually far too large for researchers to examine directly. There are
millions of male military veterans in the United States, thousands of communities nationwide, and
approximately 1.5 million prison inmates. Nobody can possibly study any of these populations in its entirety.
Samples are thus drawn from populations of interest. Samples are subsets of populations. Morris and Worrall
(2010), for example, drew a random sample of 2,500 inmates. This sample, unlike its overarching population,
was of manageable size and could be analyzed directly.
Populations and samples give rise to three types of distributions: population, sample, and sampling. A
population distribution contains all values in the entire population, while a sample distribution shows the
shape and form of the values in a sample pulled from a population. Population and sample distributions are
both empirical. They are made of raw scores derived from actual people or objects. Sampling distributions, by
contrast, are theoretical arrays of sample statistics. Each of these is discussed in turn in this chapter.
Population distribution: An empirical distribution made of raw scores from a population.
Sample distribution: An empirical distribution made of raw scores from a sample.
Sampling distribution: A theoretical distribution made out of infinite sample statistics.
212
Empirical Distributions: Population and Sample Distributions
Population and sample distributions are both empirical because they exist in reality; every person or object in
the population or sample has a value on a given variable that can be measured and plotted on a graph. To
illustrate these two types of distributions, we can use the 2006 Census of Jails (COJ; see Data Sources 3.1).
This data set contains information on all jails that house inmates for extended periods either awaiting trial or
serving sentences (N = 2,949). No sampling was done; every jail in the United States was asked to provide
information. This makes the COJ a population data set.
Figure 7.1 is a histogram made in SPSS showing the population distribution for the variable total number of
inmates , which is a measure of the number of people housed in each facility. Every facility’s inmate count was
plotted to form this histogram. You should be able to recognize immediately that this distribution has an
extreme positive skew.
What might the distribution look like for a sample pulled from this population? The SPSS program can be
commanded to select a random sample of cases from a data set, so this function will be used to simulate
sampling. A random sample of 50 facilities was pulled from the COJ and the variable total number of inmates
plotted to form Figure 7.2. The sample distribution looks somewhat similar to the population distribution in
that they are both positively skewed; however, you can see that there are clear differences between them. This
is because there are only 50 cases in this sample, and 50 is a very small subset of 2,949.
Let’s try increasing the sample size to 100. Telling SPSS to randomly select 100 jails and graph their sizes
yields the histogram in Figure 7.3. The shape is closer to that of the population, since 100 is a better
representation of the full 2,949 than 50 is. Of course, differences linger.
Figure 7.1 Population Distribution for Total Inmates per Correctional Facility (N = 2,371)
Figure 7.2 Sample Distribution (N = 50)
213
We can try the sampling exercise again and this time pull 500 cases randomly. Figure 7.3 shows the total
inmates histogram for this third sample. This histogram is a close match to that for the entire population,
which is a function of the larger sample size. We have thus demonstrated a fact that will be critical to the
understanding of empirical distributions: Among samples drawn at random from populations, larger samples
are more-accurate reflections of those populations. The technical term for this is representativeness. A
representative sample matches the population on key characteristics. A random sample of adults from a
population that is 50% female should also be roughly 50% female if it is representative. A sample that was
12% or 70% female would not be representative. Sample size is not the only determinant of how
representative a sample will be of its population, but it is an important factor.
Representativeness: How closely the characteristics of a sample match those of the population from which the sample was drawn.
Before we continue, we should stop and consider some new terms and their definitions. First, although we
have encountered the word statistic several times up to this point, it is necessary now to offer a formal
definition of this concept. A statistic is a number that describes a sample. This might be a mean, proportion,
or standard deviation. The second term is parameter. A parameter is just like a statistic except that it describes
a population. Populations, like samples, have means, proportions, standard deviations, and so on. Statistics are
estimates of parameters. Table 7.1 displays the symbols for some common statistics and their corresponding
parameters. The statistic notations for the mean and the standard deviation are familiar from previous
chapters, but this is the first time we have considered the population symbols. They are Greek letters. The
population mean is the lowercase version of the letter mu (µ ; pronounced “mew”), and the standard deviation
is a lowercase sigma (σ) . The population proportion is a less exciting uppercase P. We have previously
represented sample proportions with a lowercase p , but now that we are differentiating between samples and
populations, we are going to change the letter to , which is pronounced “p hat.”
Statistic: A number that describes a sample that has been drawn from a larger population.
Parameter: A number that describes a population from which samples might be drawn.
214
Figure 7.3 Sample Distribution (N = 100)
Figure 7.4 Sample Distribution (N = 500)
215
Learning Check 7.1
For each of the following, identify whether, based on the information provided, you believe the sample is representative of the population.
1. In a sample of people, the mean age is and in the population, µ = 41.
2. In a sample of prosecutors’ offices, the proportion of criminal cases in each office that are felonies (compared to misdemeanors and
other types of offenses) is . In the population, felonies make up P = .20 of cases.
3. In a sample of cities, the mean violent crime rate (per 100,000) is , compared to µ = 373. In this same sample, the
mean property crime rate (per 100,000) is compared to µ = 2,487.
Often, criminal justice and criminology researchers want to make a statement about a population, but what
they actually have in front of them to work with is a sample (because samples are typically more feasible to
collect and work with than populations are). This creates a conundrum because sample statistics cannot simply
be generalized to population parameters. Statistics are estimates of population parameters, and, moreover,
they are estimates that contain error (as demonstrated by Figures 7.1 through 7.4). Population parameters are
fixed. This means that they have only one mean and standard deviation on any given measure (e.g., people’s
ages), one percentage on other measures (e.g., the percentage that is female), and so on. Sample statistics, by
contrast, vary from sample to sample because of sampling error. Sampling error arises from the fact that
multiple (theoretically, infinite) random samples can be drawn from any population. Any given sample that a
researcher actually draws is only one of a multitude that he or she could have drawn. Figure 7.5 depicts this.
Every potential sample has its own distribution and set of descriptive statistics. In any given sample, these
statistics might be exactly equal to, roughly equal to, or completely different from their corresponding
parameters.
Sampling error: The uncertainty introduced into a sample statistic by the fact that any given sample is one of many samples that
could have been drawn from that population.
Figure 7.5 Multiple Random Samples Can Be Drawn From a Population
The COJ can be used to illustrate the effects of sampling error. As described earlier, this data set is a
216
population; therefore, samples can be drawn from it. We will continue using the variable total number of
inmates housed. The population means is µ = 269.63, and the standard deviation is σ = 802.13. Drawing five
random samples of 500 facilities each and computing the mean and the standard deviation of each sample
produces the data in Table 7.2.
Figure 7.6 The Relationship Between Samples, Sampling Error, and Populations
Look how the means and standard deviations vary—this is sampling error! Sample 1’s mean is 282.80, while
Sample 3’s is a very different 200.57. There is a substantial amount of variation among the samples, and each
one differs from the true population mean by a smaller or larger amount. In this example, we have the benefit
of knowing the true population mean and the standard deviation, but that is usually not the case in criminal
justice and criminology research. What researchers generally have is one sample and no direct information
about the population as a whole. Imagine, for instance, that you drew Sample 1 in Table 7.2. This sample’s
mean and standard deviation are reasonable approximations of—though clearly not equivalent to—their
corresponding population values, but you would not know that. Now picture Sample 3 being the one you
pulled for a particular study. This mean and standard deviation are markedly discrepant from the population
parameters, but, again, you would be unaware of that.
This example illustrates the chasm between samples and populations that is created by sampling error and
prevents inferences from being made directly. It would be a mistake to draw a sample, compute its mean, and
automatically assume that the population mean must be equal or close to the sample mean. As displayed
pictorially in Figure 7.5, sampling error prevents direct inference from a sample to the larger population from
which it was drawn. What is needed is a bridge between samples and populations so that inferences can be
reliably drawn. This bridge is the sampling distribution.
217
Theoretical Distributions: Sampling Distributions
Sampling distributions, unlike population and sample distributions, are theoretical; that is, they do not exist as
empirical realities. We have already worked with theoretical distributions in the form of the binomial and
standard normal distributions. Sampling distributions are theoretical because they are based on the notion of
multiple (even infinite) samples being drawn from a single population. What sets sampling distributions apart
from empirical distributions is that sampling distributions are created not from raw scores but, rather, from
sample statistics. These descriptors can be means, proportions, or any other statistic. Imagine plotting the
means in Table 7.2 to form a histogram, like Figure 7.7.
Figure 7.7 Histogram of the Five Sample Means in Table 7.2
Not terribly impressive, is it? Definitely leaves something to be desired. That is because there are only five
samples. Sampling distributions start to take shape only when many samples have been drawn. If we continue
the iterative process of drawing a sample, computing and plotting mean, throwing the sample back, and
pulling a new sample, the distribution in Figure 7.7 gradually starts looking something like the curve in Figure
7.8.
Now the distribution has some shape! It looks much better. It is, moreover, not just any old shape—it is a
normal curve. What you have just seen is the central limit theorem (CLT) in action. The CLT states that any
time descriptive statistics are computed from an infinite number of large samples, the resulting sampling
distribution will be normally distributed. The sampling distribution clusters around the true population mean
(here, 269.63), and if you were to compute the mean of the sampling distribution (i.e., the mean of means),
the answer you obtained would match the true population mean. Its standard deviation (called the standard
error) is smaller than the population standard deviation because there is less dispersion, or variability, in
means than in raw scores. This produces a narrower distribution, particularly as the size of the samples
increases. The mean of the sampling distribution is symbolized as and the standard error is represented as
.
218
Central limit theorem (CLT): The property of the sampling distribution that guarantees that this curve will be normally distributed
when infinite samples of large size have been drawn.
Standard error: The standard deviation of the sampling distribution.
The CLT is integral to statistics because of its guarantee that the sampling distribution will be normal when a
sample is large. Criminal justice and criminology researchers work with many variables that show signs of
skew or kurtosis. The CLT saves the day by ensuring that even skewed or kurtotic variables will produce
normal sampling distributions. The inmate count variable demonstrates this. Compare Figure 7.8 to Figure
7.4: Figure 7.4 is derived from a single sample (N = 500) and is highly skewed, yet the sampling distribution
in Figure 7.8 is normal. That is because even when raw values produce skew, sample statistics will still hover
around the population parameter and fall symmetrically on each side of it. Some statistics will be greater than
the parameter and some will be smaller, but the majority will be close approximations of (or even precisely
equal to) the true population value.
Figure 7.8 Distribution of Sample Means
All descriptive statistics have sampling distributions to which the CLT applies. In Chapter 6, you learned that
nationwide, .62 (or 62%) of felony defendants are granted pretrial release. Figure 7.9 sketches what the
sampling distribution of proportions for the pretrial release variable might look like. You can see that this
curve is roughly normal in shape, too, just like the sampling distributions of means in Figure 7.8.
219
Learning Check 7.2
In your own words, explain why the sampling distribution takes on a normal shape. That is, why do the bulk of scores cluster in one area
and taper off in each tail? Provide this explanation in simple language that someone who was not taking a statistics class could understand.
Figure 7.9 Sampling Distribution of Proportions
220
Sample Size and the Sampling Distribution: The z and t Distributions
The key benefit of the sampling distribution being normally distributed is that the standard normal curve can
be used. Everything we did in Chapter 6 with respect to using raw scores to find z scores, z scores to find
areas, and areas to find z scores can be done with the sampling distribution. The applicability of z , though, is
contingent on N being large. “Large” is a vague adjective in statistics because there is no formal rule specifying
the dividing line between small and large samples. Generally speaking, large samples are those containing at
least 100 cases. When N ≥ 100, the sampling distribution is assumed to be normally distributed, and the
standard normal curve can be used. This requirement reduces the overall usefulness of the z distribution; it
turns out that the standard normal curve makes somewhat rigid demands, and many real-world data sets fall
short of these expectations.
Although it is generally not advisable to work with samples smaller than 100, there are times when it is
unavoidable. In these situations, the z distribution cannot be employed and researchers turn to the t
distribution instead. The t distribution—like the z curve—is symmetric, unimodal, and has a constant area of
1.00. The key difference between the two is that t is a family of several different curves rather than one fixed,
single curve like z. The t distribution changes shape depending on the size of the sample. When the sample is
small, the curve is wide and flattish; as the sample size increases, the t curve becomes more and more normal
until it looks identical to the z curve. See Figure 7.10.
t distribution: A family of curves whose shapes are determined by the size of the sample. All t curves are unimodal and symmetric
and have an area of 1.00.
Figure 7.10 The Family of t Curves
This phenomenon can also be demonstrated using hypothetical examples of random samples from the Census
of State and Federal Adult Correctional Facilities (CSFACF) data set. Figures 7.11 and 7.12 demonstrate how
the t curve would change shape depending on the size of the samples being randomly selected from this
population. The curve in Figure 7.11 is based on a sample size of 75, and that in Figure 7.12 is premised on N
= 25. The first curve is taller and thinner. It is normal, but contrasting it with Figure 7.8, you can see that
Figure 7.11 does not cluster around the population mean as tightly. Scores are more spread out. This trend is
even more pronounced in Figure 7.12. This is because there is more variability in smaller samples—it is
difficult to get an accurate estimate of the true population parameter when there are only a handful of cases. A
sample of N = 75 is not ideal, but it is an improvement over 25.
221
Figure 7.11 The t Distribution for Inmate Population at N = 75
Figure 7.12 The t Distribution for Inmate Population at N = 25
The t distribution’s flexibility allows it to accommodate samples of various sizes. It is an important theoretical
probability distribution because it allows researchers to do much more than they would be able to if z were
their only option. All else being equal, large samples are better than small ones, and it is always advisable to
work with large samples when possible. When a small sample must be used, though, t is a trustworthy
alternative. We will use both the z and t distributions in later chapters.
Chapter Summary
Criminal justice and criminology researchers often seek information about populations. Populations, though, are usually too large to
analyze directly. Samples are therefore pulled from them and statistical analyses are applied to these samples instead. Population and
sample distributions are made of raw scores that have been plotted. These are empirical distributions.
Sampling error, though, introduces an element of uncertainty into sample statistics. Any given sample that is drawn is only one of a
multitude of samples that could have been drawn. Because of sampling error, statistics are merely estimates of their corresponding
population parameters and cannot be interpreted as matching them exactly. In this way, there is a gap between samples and
populations.
The sampling distribution links samples to populations. Sampling distributions are theoretical curves made out of infinite sample
statistics. All descriptive statistics have sampling distributions. The CLT ensures that sampling distributions are normal when
sample sizes are large (i.e., N ≥ 100). When this is the case, the z distribution can be used. When samples are small, though, the
sampling distribution cannot be assumed to be normal. The t distribution solves this problem because t is a family of curves that
change shape depending on sample size. The t distribution is more flexible than z and must be used any time N ≤ 99, though it can
be used with large samples as well.
222
Thinking Critically
1. Suppose you read a research article in which the study authors stated that they had collected a sample of 150 adolescents
from a single city and found that 22 of them reported being in a gang. What is the population being studied here? What
additional pieces of information would you need to know about this sample in order to decide whether to trust this finding as
being truly reflective of that population? What characteristics would the sample need to have in order for it to be
trustworthy?
2. In your own words, explain (1) why sample size is important to ensuring that samples are representative of the populations
from which they are derived and (2) why the sampling distribution for any given statistic is normally distributed even when
the population and sample distributions are skewed.
Review Problems
1. Population distributions are . ..
1. empirical.
2. theoretical.
2. Sample distributions are . ..
1. empirical.
2. theoretical.
3. Sampling distributions are . ..
1. empirical.
2. theoretical.
4. The ______________ distribution is made from the raw scores in a sample.
5. The ______________ distribution is made from statistics calculated on multiple or infinite samples.
6. The ______________ distribution is made from the raw scores in a population.
7. The CLT guarantees that as long as certain conditions are met, a sampling distribution will be . ..
1. positively skewed.
2. normally distributed.
3. negatively skewed.
8. For the CLT’s promise of distribution shape to hold true, samples must be . ..
1. large.
2. small.
9. When a sample contains 100 or more cases, the correct probability distribution to use is the ________ distribution.
10. When a sample contains 99 or fewer cases, the correct probability distribution to use is the _________ distribution.
11. A researcher gathers a sample of 200 people, asks each one how many times he or she has been arrested, and then plots each
person’s response. From the list below, select the type of distribution that this researcher has created.
1. A sample distribution with a large N
2. A population distribution with a small N
3. A sampling distribution with a large N
4. A sample distribution with a small N
12. A researcher gathers a sample of 49 police departments, finds out how many officers were fired for misconduct in each
department over a 2-year time span, and plots each department’s score. From the list below, select the type of distribution
that this researcher has created.
1. A population distribution with a small N
2. A population distribution with a large N
3. A sampling distribution with a small N
4. A sample distribution with a small N
13. A researcher gathers a sample of 20 cities, calculates each city’s mean homicide rate, and plots that mean. Then the
researcher puts that sample back into the population and draws a new sample of 20 cities and computes and plots the mean
homicide rate. The researcher does this repeatedly. From the list below, select the type of distribution that this researcher has
created.
1. A sampling distribution with a large N
223
2. A population distribution with a small N
3. A sampling distribution with a small N
4. A population distribution with a large N
14. A researcher has data on each of the nearly 2,000 adult correctional facilities in the United States and uses them to plot the
number of inmate-on-inmate assaults that took place inside each prison in a 1-year span. From the list below, select the type
of distribution that this researcher has created.
1. A sample distribution with a large N
2. A population distribution with a small N
3. A sampling distribution with a large N
4. A population distribution with a large N
15. A researcher gathers a sample of 132 people and computes the mean number of times the people in that sample have
shoplifted. The researcher then puts this sample back into the population and draws a new sample of 132 people, for whom
the researcher computes the mean number of times shoplifted. The researcher does this repeatedly. From the list below,
select the type of distribution that this researcher has created.
1. A sample distribution with a large N
2. A sampling distribution with a large N
3. A sample distribution with a small N
4. A population distribution with a large N
224
Key Terms
Population distribution 163
Sample distribution 163
Sampling distribution 164
Representativeness 165
Statistic 165
Parameter 165
Sampling error 168
Central limit theorem 170
Standard error 170
t distribution 172
Glossary of Symbols and Abbreviations Introduced in This Chapter
225
Chapter 8 Point Estimates and Confidence Intervals
226
Learning Objectives
Explain the problems with point estimates and the need for confidence intervals.
Explain the trade-off between confidence and precision.
Identify the key elements of the t curve, and use sample size to select the correct distribution (z or t).
Use a specified confidence level to find alpha and the critical value of z or t.
Identify the correct distribution and formula for a given sample statistic and sample size.
Calculate and interpret confidence intervals for means and proportions with different confidence levels and sample sizes.
Any given sample that is drawn from a population is only one of a multitude of samples that could have been
drawn. Every sample that is drawn (and every sample that could potentially be drawn) has its own descriptive
statistics, such as a mean or proportion. This phenomenon, as you learned in Chapter 7, is called sampling
error. The inherent variation in sample statistics such as means and proportions prevents direct inference from
a sample to a population. It cannot be assumed that a mean or proportion in a sample is an exact match to the
mean or proportion in the population because sometimes sample statistics are very similar to their
corresponding population parameters and other times they are quite dissimilar. Flip back to Table 7.2 for an
illustration of this concept. Basically, there is always an element of uncertainty in a sample statistic, or what
can also be called a point estimate.
Point estimate: A sample statistic, such as a mean or proportion.
Fortunately, though, a procedure exists for calculating a range of values within which the parameter of interest
is predicted to be. This range stretches out on each side of the point estimate and is called a confidence
interval (CI) . The confidence interval acts as a sort of “bubble” that introduces flexibility into the estimate. It
is much more likely that an estimate of the value of a population parameter is accurate when the estimate is a
range of values rather than one single value.
Confidence interval: A range of values spanning a point estimate that is calculated so as to have a certain probability of containing
the population parameter.
Try thinking about it this way: Suppose I guessed that you are originally from Chicago. This is a very precise
prediction! Of all the cities and towns in the world, I narrowed my guess down to a single area. Given its
precision, though, this prediction is very likely to be wrong; there are more than 7.4 billion people in the
world, and only about 2.7 million of them live in Chicago. My point estimate (Chicago) is probably incorrect.
But what if I instead guessed that you are from the state of Illinois? There are 12.9 million people in Illinois,
so I have increased my chances of being correct because I have broadened the scope of my estimate. If I went
up another step and predicted that you are from the Midwest—without specifying a city or state—I have
further increased my probability of being right, since this is a much larger geographic region and contains
more than 66 million residents. It is still possible that I am wrong, of course, but I am far more likely to guess
your place of origin correctly when I guess a large geographical area, such as a region, than when I guess a
much smaller one, such as a city.
227
This is, conceptually, what a confidence interval does. The interval serves as a buffer zone that allows for
greater confidence in the accuracy of a prediction. It also allows us to determine the probability that the
prediction we are making is correct. Using distributions—specifically, the z and t probability curves—we can
figure out how likely it is that our confidence interval truly does contain the true population parameter. The
probability that the interval contains the parameter is called the level of confidence.
Level of confidence: The probability that a confidence interval contains the population parameter. Commonly set at 95% or 99%.
228
The Level of Confidence: The Probability of Being Correct
In the construction of confidence intervals, you get to choose your level of confidence (i.e., the probability that
your confidence interval accurately estimates the population parameter). This might sound great at first blush
—why not just choose 100% confidence and be done with it, right?—but confidence is actually the classic
double-edged sword because there is a trade-off between it and precision. Think back to the
Chicago/Illinois/Midwest example. The Chicago guess has a very low probability of being correct (we could
say that there is a low level of confidence in this prediction), but it has the benefit of being a very precise
estimate because it is just one city. The Illinois guess carries an improvement in confidence because it is a
bigger geographical territory; however, because it is bigger, it is also less precise. If I guess that you are from
Illinois and I am right, I am still left with many unknown pieces of information about you. I would not know
which part of the state you are from, whether you hail from a rural farming community or a large urban
center, what the socioeconomic characteristics of your place of origin are, and so on.
The problem gets worse if all I guess is that you are from the Midwest—now I would not even know which
state you are from, much less which city! If I want to be 100% sure that I will guess your place of origin
correctly, I have to put forth “planet Earth” as my prediction. That is a terrible estimate. If you want greater
confidence in your estimate, then, you pay the price of reduced precision and, therefore, a diminished amount
of useful information.
Confidence levels are expressed in percentages. Although there is no “right” or “wrong” level of confidence in
a technical sense (i.e., there is nothing mathematically preventing you from opting for a 55% or 72%
confidence level [CI ]), 95% and 99% have become conventional in criminal justice and criminology research.
Because of the trade-off between confidence and precision, a 99% CI has a greater chance than a 95% one of
being correct, but the 99% one will be wider and less precise. A 95% CI will carry a slightly higher likelihood
of error but will yield a more informative estimate. The 99% level would be akin to the Midwest guess in the
previous example, whereas the 95% would be like the Chicago guess. You should select your level of
confidence by deciding whether it is more important that your estimate be correct or that it be precise.
Confidence levels are set a priori (in advance), which means that you must decide whether you are going to
use 95% or 99% before you begin constructing the interval. The reason for this is that the level of confidence
affects the calculation of the interval. You will see this when we get to the CI formula.
Since we are dealing with probabilities, we have to face the unpleasant reality that our prediction could be
incorrect. The flipside of the probability of being right (i.e., your confidence level) is the probability of being
wrong. Consider these statements:
100% − 95% = 5%
100% − 99% = 1%
229
Each of these traditional levels of confidence carries a corresponding probability that a confidence interval
does not contain the true population parameter. If the 95% level is selected, then there is a 5% chance that the
CI will not contain the parameter; a 99% level of confidence generates a 1% chance of an inaccurate CI .
You will, in all likelihood, never know whether the sample you have in front of you is one of the 95% or 99%
that is correct, or whether it is one of the 5% or 1% that is not. There is no way to tell; you just have to
compute the confidence interval and hope for the best. This is an intractable problem in statistics because of
the reliance on probability.
Three types of confidence intervals will be discussed in this chapter: CIs for means with large samples (N ≥
100), for means with small samples (N ≤ 99), and for proportions and percentages. All three types of CIs are
meant to improve the accuracy of point estimates by providing a range of values that most likely contains the
true population parameter.
230
Confidence Intervals for Means With Large Samples
When a sample is of large size (N ≥ 100), the z distribution (the standard normal curve) can be used to
construct a CI around a sample mean. Confidence intervals for means with large samples are computed as
where
= the sample mean,
z⍺ = the z score associated with a given alpha level (i.e., the critical value of z) ,
⍺ = the probability of being wrong (the alpha level),
s = the sample standard deviation, and
N = the sample size.
This formula might appear intimidating, but we can read it left to right and break down the different
elements. The starting point is the sample mean. A certain value will be added to and subtracted from the
mean to form the interval. That value is the end result of the term on the right side of the ± operator. The z
score’s subscript ⍺ (this is the Greek letter alpha) represents the probability that the CI does not contain the
true population parameter. Recall that every level of confidence carries a certain probability of inaccuracy; this
probability of inaccuracy is ⍺ or, more formally, the alpha level. Alpha is computed as
Alpha level: The opposite of the confidence level; that is, the probability that a confidence interval does not contain the true
population parameter. Symbolized ⍺.
For 95% and 99% confidence, first convert the percentages to proportions:
⍺95% = 1 – .95 = .05
⍺99% = 1 – .99 = .01
Alpha, itself, is not inserted into the CI formula; rather, ⍺ is used to find the critical value of z , and you enter
that critical value into the formula (this is the z⍺ term in Formula 8[1]). The critical value of z is the z score
associated with a particular area on the curve. In other words, it is the score beyond which a certain area (this
being alpha) is out in the tail. We will see later that the t curve has critical values, too.
Critical value: The value of z or t associated with a given alpha level. Symbolized z⍺ or t⍺ .
There is another piece of information that you need to know about CIs: They are always two-tailed. This is
because the normal curve has two halves that are split by the mean, with the result being (pessimistically) that
231
there are two ways to be wrong with a confidence interval. The first option is for the interval to be wholly
above the mean and miss it by being too far out in the positive tail, and the second possibility is that it lands
entirely below the mean and misses it by being too far out into the negative side. Since either error is possible
—and since we have no control over which type of error we might end up with—we must use both sides of
the curve. This is a two-tailed test. In such a test, the alpha level is split in half and placed in each of the two
tails of the distribution, as pictured in Figure 8.1. Two-tailed tests, then, actually have two critical values. The
absolute value of these critical values is the same because the curve is symmetric, but one value is negative and
the other is positive. Confidence intervals, by definition, use both of the critical values for any alpha level.
Two-tailed test: A statistical test in which alpha is split in half and placed into both tails of the z or t distribution.
Figure 8.1 The Alpha Level and Critical Values of z
So, what are these critical values? For the standard normal curve, we can find them using the techniques we
learned in Chapter 6. Recall that in the z -score exercises, we figured out how to find z scores using specified
areas: Essentially, the z table (see Appendix B) is used backward. Let us begin with the 95% confidence
interval. First, we have to find ⍺, which is 5%. Second, because this is a two-tailed test, we divide ⍺ in half, as
shown in Figure 8.1. Doing this shows that the critical value of z is going to be the value that leaves 2.5% of
cases in each tail of the distribution. Of course, we cannot work with percentages in the z table, so we have to
convert this to .025. Third, the z table is used to find the critical value of z . We have done this before, in
Chapter 6. What we are asking here is “If the area in the tail is .025, what is z ?” The tail must first be
subtracted from .50:
.50 – .025 = .4750
Next, go to the z table and locate the area that is equal to .4750 or, if there is no area that is exactly .4750, the
area that is closest. In this instance, the value we are seeking is actually located in the table. Trace away from
.4750 upward along the column and to the left along the row, and record each of the numbers you arrive at.
Here,
z = 1.9 + .06 = 1.96
We are not quite done! Remember that there are two critical values, not just one. Since the standard normal
curve is symmetric and .025 is the area in each tail, these two z scores will take on the same absolute value but
232
will have opposite signs. Thus, z⍺ =.05 = ±1.96. The critical value of z for the top .025 (the positive side of the
curve) is 1.96, and the value for the bottom .025 (the negative side) is –1.96. Figure 8.2 displays the two
critical values of z for a 95% confidence level.
The same process is used to find z⍺ for a 99% confidence level. First, alpha is 100% – 99% = 1%. Second, 1%
divided in half is .5% and converted to a proportion is .005. Third, .50 – .005 = .4950. This exact value does
not appear in the z table, so the z scores associated with the two closest values (.4949 and .4951) must be
averaged. These scores are 2.57 and 2.58. Therefore,
Figure 8.2 Critical Values of z for a 95% Level of Confidence
The critical value is 2.58. Again, remember that there are two values, so z is actually ±2.58. This is displayed
graphically in Figure 8.3.
Figure 8.3 Critical Values of z for a 99% Level of Confidence
The standard normal curve is handy because it is fixed; therefore, the critical values of z for 95% and 99%
confidence intervals will always be ±1.96 and ±2.58, respectively. Later on, you will see that this is not the case
when the t distribution is employed; in that situation, the critical value of t will have to be located each time
you calculate a confidence interval. For now, though, we can rely on these two critical values.
233
Learning Check 8.1
Now that you have seen the formula for calculating confidence intervals and you know that the critical value of z for 95% confidence is
±1.96 and that for 99% is ±2.58, which level of confidence do you think will produce a wider confidence interval (i.e., a less precise
estimate)? Write down your prediction so that when we get to the calculations, you can see if you were correct.
Now let’s consider an example from the 2013 Census of Jails (COJ; see Data Sources 3.1). This data set is
ideal for present purposes because it contains population data (i.e., it is a census, not a sample, of jails), so we
can compute the actual population mean and then pull a random sample, compute its mean and confidence
interval, and see if the computed interval contains the true population value. Remember that researchers are
not ordinarily able to access information on the population parameters (and, obviously, if you knew the
population mean, there would be no point in attempting to estimate it using a confidence interval). What we
are doing here is for demonstration purposes only and is not typical of the reality of most criminal justice and
criminology research.
For this example we will use the COJ variable in which each facility reported the number of inmates it
currently housed who had been convicted and were awaiting sentencing. The confidence level will be set at
95%.
In the population, the mean number of convicted individuals who have not yet been sentenced is µ = 12.06 (σ
= 56.04). Let us use a sample size of N = 183, which is large enough to permit use of the z distribution. The
SPSS random-sample generator produces a sample with a mean of and a standard deviation
of s = 97.52. Since the confidence level is 95%, z⍺ = ±1.96. Plugging values for N , , s , and z⍺ into Formula
8(1) yields
= 18.18 ± 1.96 (7.23)
= 18.18 ± 14.17.
Let’s pause for a moment and consider what we have calculated thus far. We have the sample mean (18.18) as
the midpoint of the interval, and now we know that the buffer is going to extend 14.17 units above and 14.17
units below the mean. Picture these two extensions of the mean as forming the full width of the interval.
The next step is to compute the lower limit (LL) and the upper limit (UL) of the interval, which requires the
234
± operation to be carried out, as such:
LL = 18.18 – 14.17 = 4.01
UL = 18.18 + 14.17 = 32.35
Finally, the full interval can be written out:
95% CI : 4.01 ≤ µ ≤ 32.35
The interpretation of this interval is that there is a 95% chance that the true population mean is 4.01, 32.35,
or some number in between. More formally, “There is a 95% chance that the interval 4.01 to 32.35, inclusive,
contains the true value of µ.” Of course, this also means that there is a 5% chance that the true population
mean is not in this range. In the present example, we know that µ = 12.06, so we can see that here the
confidence interval does indeed contain the population mean (i.e., 12.06 is greater than 4.01 and less than
32.35). This is good! This sample was one of the 95% that produces accurate confidence intervals rather than
one of the 5% that does not. Remember, though, that knowing the value of the true population mean is a
luxury that researchers generally do not have; ordinarily, there is no way to check the accuracy of a sample-
based confidence interval.
Let us repeat the previous example using a 99% confidence level. This set of calculations would proceed as
such:
= 18.18 ± 2.58 (7.23)
= 18.18 ± 18.65.
And computing the lower and upper limits yields
LL = 18.18 – 18.65 = –.47
UL = 18.18 + 18.65 = 36.83
99% CI : –.47 ≤ µ ≤ 36.83.
235
There is a 99% chance that the interval –.47 to 36.83, inclusive, contains the population mean µ. (Of course,
the negative values are nonsensical because it is impossible for a facility to have fewer than zero unsentenced
inmates. From a practical standpoint, the negative values are treated as zero; this CI indicates that it is possible
the true population mean is zero.) Again, we know that µ = 12.06, so this interval does, indeed, contain the
parameter.
Take a look at the difference in width between the 95% and 99% intervals. The 95% interval ranges from 4.01
to 32.35, and the 99% one spans –.47 to 36.83. The 99% interval is much wider than the 95% one. Can you
explain the reason for this? If you said that it is because with a 99% confidence level we sacrifice precision in
the estimate, you are correct! When we enhanced our confidence, we increased z from 1.96 to 2.58, thus
causing the interval to expand. This demonstrates the trade-off between confidence and precision.
For a third example, we will use the Police–Public Contact Survey (PPCS; see Data Sources 2.1). This is not
a population: The survey was administered to a random sample of U.S. residents. We must, therefore, use
sample statistics to estimate the true population values. Let us consider the ages of respondents who reported
having experienced multiple (i.e., more than one) contact with police officers during the past year and
construct a 99% CI to estimate the mean age for the population. The respondents with multiple contacts (N =
2,304) had a mean age of 45.23 with a standard deviation of 16.80. The CI is
= 45.23 ± 2.58 (.35)
= 45.23 ± .90
Now calculate the lower and upper limits:
LL = 45.23 – .90 = 44.33
UL = 45.23 + .90 = 46.13
Finally, construct the interval:
99% CI: 44.33 ≤ µ ≤ 46.13
We can say with 99% confidence that the interval 44.33 and 46.13, inclusive, contains µ. This is a very precise
interval even though we used the 99% level of confidence. The primary reason for this is the sample size. The
236
larger the sample, the larger the denominator in the portion of the CI equation will be, thus
producing a very small number after the division is complete. Of course, this also depends on the size of the
standard deviation (and the quality of the estimate is predicated on the sample being representative of the
population), but, generally speaking, larger samples yield narrower confidence intervals compared with smaller
samples.
237
Confidence Intervals for Means With Small Samples
The z distribution can be used to construct confidence intervals when N ≥ 100 because the sampling
distribution of means can be safely assumed to be normal in shape. When N ≤ 99, though, this assumption
breaks down and a different distribution is needed. This alternative distribution is the t curve. The CI formula
for means with small samples is nearly identical to that for large samples; the only difference is that the critical
value of t is used instead of the critical value of z . The formula is
where t⍺ = the critical value of t at a given alpha level.
Research Example 8.1 Do Criminal Trials Retraumatize Victims of Violent Crimes?
Violent crimes frequently leave victims with significant negative psychological and emotional consequences. Some victims develop
long-term problems, and many are diagnosed with posttraumatic stress disorder. There is widespread concern about the negative
impacts trials can exert on victims already suffering emotional and psychological trauma. During trials, victims have to recount their
experiences in open court and undergo cross-examination by defense attorneys. Victims might feel shame, embarrassment, and fear.
It is not known with certainty, however, whether trials do indeed have these effects; limited research has been conducted and so it is
unclear whether the concerns for victims’ wellbeing are empirically founded. Orth and Maercker (2004) addressed this gap in the
research. Using a sample of violent-crime victims in Germany, the authors calculated participants’ mean levels of posttraumatic stress
reactions prior to and then again after their attackers’ trials. The researchers subtracted the two means to obtain a measure of change
over time. The table summarizes the results.
The results revealed a slight decrease in posttraumatic stress reactions over time, as indicated by the negative difference scores. This
decline was very small and was indistinguishable from zero, as indicated by the fact that the 95% CIs contained zero (that is, it was
possible that the true population differences were zero). These results call into question the common assumption that trials
retraumatize victims, since there was no evidence in this study of victims experiencing an increase in emotional distress after trial.
More research is needed before firm conclusions can be drawn, but these findings suggest that trials are not significantly painful
events for most victims.
Source: Adapted from Table 3 in Orth and Maercker (2004).
The critical value of t (i.e., t⍺) is found using the t table (see Appendix C). We have not used this table yet, so
take a few minutes now to familiarize yourself with it. There are three pieces of information you need in order
to locate t⍺. The first is the number of tails. As described previously, confidence intervals are always two-
tailed, so this is the option that is always used with this type of test. The second determinant of the value of t⍺
238
is the alpha level, the computation of which was shown in Formula 8(2). Finally, finding t⍺ requires you to
first compute the degrees of freedom (df) . With the t distribution, df are related to sample size, as such:
239
Learning Check 8.2
We will be using the t distribution in later chapters, so it is a good idea to memorize the three pieces of information you always need in
order to locate t⍺ on the t table:
Take a moment now to practice using the t table. Find the critical value of t for each of the following:
1. A two-tailed test with ⍺ = .05 and df = 10
2. A two-tailed test with ⍺ = .10 and df = 20
3. A two-tailed test with ⍺ = .01 and df = 60
The df values are located in the rows of the t table. Note that not all of the possible values that df might
assume are included on the table. When the number you are looking for is not there, it is customary to use the
table df that is closest to the sample-derived df . If the derived df is halfway between two numbers in the table,
select the smaller one. For instance, if your sample size were 36, your df would be 36 – 1 = 35. Your options in
using the table would be df = 30 and df = 40, and you would elect to use df = 30. The rationale is that using the
lower of the two makes for a more conservative estimate (i.e., a wider confidence interval). In statistics,
researchers usually err on the side of caution—if you have to choose between inflating your calculated CI and
shrinking it, you should opt for inflation because that is in line with custom in criminology and criminal
justice research. (The inflation this causes is generally miniscule and has a trivial impact on the CI .)
240
Learning Check 8.3
You have seen now that the value of t varies depending on the sample size. This is unlike z , which remains constant across samples of all
sizes. Why is this? What is it about the t distribution that causes the critical value of t to change as the sample size changes? If you need a
hint, flip back to Chapter 7 and, in particular, Figures 7.10 through 7.12.
For an example of CIs with means and small samples, the Firearm Injury Surveillance Study (FISS; see Data
Sources 8.1) will be used. This data set contains information about a sample of patients treated in emergency
departments (EDs) for gunshot wounds between the years 1993 and 2013. We will narrow the file down to
the most recent year of data available. For the current example, we will analyze the mean age of Hispanic
female gunshot victims. There were 40 such victims, and their mean age was 26.18 (s = 11.55). We will set
confidence at 95%.
Data Sources 8.1 The Firearm Injury Surveillance Study, 1993–2013
The Centers for Disease Control and Prevention (CDC) maintain the National Electronic Injury Surveillance System, of which the
FISS is a part. The data are from a nationally representative sample of hospitals stratified by size. Detailed incident data include only
those patients who did not die in the ED in which they were treated; those who died prior to arrival are assigned dead on arrival
(DOA) status, and those who died after being transferred from the ED to some other hospital unit are coded as transfers (U.S.
Department of Health and Human Services, n.d.). This is a limitation of the FISS that circumscribes its usefulness in criminal
justice research, but this data set is nonetheless valuable as one of the few major studies that systematically tracks gun-related
injuries. Data include patient age, sex, and race, as well as incident characteristics such as whether the shooting was intentional or
unintentional and what the relationship was between the shooter and the victim. The most recent version of the FISS was collected
in 2013.
The first step we must take to construct the CI is to find t⍺ . We know that this is a two-tailed test with ⍺ = 1
−.95 = .05 and df = 40 − 1 = 39. You can see that df = 39 is not located on the table, so we must use the df = 40
row instead. The critical value of t is ±2.021.
Next, plug all the numbers into Formula 8(2):
= 26.18 ± 2.021 (1.85)
= 26.18 ± 3.74
Compute the lower and upper limits:
241
LL = 26.18 – 3.74 = 22.44
UL = 26.18 + 3.74 = 29.92
Finally, assemble the full confidence interval:
95% CI : 22.44 ≤ µ ≤ 29.92
There is a 95% chance that the interval 22.44 to 29.92, inclusive, contains the population mean.
Let’s try one more example. We will again use the FISS and this time analyze the mean age of males whose
firearm injuries occurred in the course of a drug-related crime (the FISS does not contain information about
the role gunshot victims played in these transactions). In the 2013 data, there were 124 males shot during
drug-related crimes, and their mean age was 30.06 (s = 10.08). The confidence level will be 99%.
The three pieces of information needed for finding the critical value of t are (1) this is a two-tailed test, (2)
alpha is .01 (since 1 – .99 = .01), and (3) the degrees of freedom are 124 – 1 = 123. With these criteria, t⍺ =
±2.617 (using the df = 120 line of the t table). Plugging all the numbers into Formula 8(2) yields
= 30.06 ± 2.617 (.91)
= 30.06 ± 2.38
The lower and upper limits are
LL = 30.06 – 2.38 = 27.68
UL = 30.06 + 2.38 = 32.44
The interval is
99% CI: 27.68 ≤ µ ≤ 32.44
There is a 99% probability that the interval 27.68 to 32.44, inclusive, contains µ .
One more t example can be derived from the PPCS. Respondents who had been stopped by police while
driving or walking were asked how long that stop lasted. The 208 people who answered the question reported
242
a mean stop length of 13.30 minutes (s = 21.19). We will opt for a 95% confidence interval, and the df = 208 –
1 = 207. Looking at the t table, you can see that there is no df = 207, so we will have to choose between the df
= 120 and df = ∞ (infinity) values. Sticking to our principle of always opting for the more conservative (i.e.,
wider) confidence interval, we will select df = 120. This makes our critical value ±1.980. Now we can plug in
the numbers:
= 13.30 ± 1.980 (1.47)
= 13.30 ± 2.91
The lower and upper limits are
LL = 13.30 – 2.91 = 10.39
UL = 13.30 + 2.91 = 16.21
The interval is
99% CI: 10.39 ≤ µ ≤ 16.21
We can be 95% confident in our prediction that the population mean is between 10.39 and 16.21, inclusive.
243
Learning Check 8.4
The interval calculated in the police stop example spans nearly 6 minutes (16.21 – 10.39 = 5.82). Suppose a researcher wants a more
precise estimate of µ. Use what you have learned thus far in the chapter to identify two actions that could be taken to shrink the interval
and increase precision.
244
Confidence Intervals With Proportions and Percentages
The principles that guide the construction of confidence intervals around means also apply to confidence
intervals around proportions and percentages. There is no difference in the procedure used for proportions
versus that for percentages—percentages simply have to be converted to proportions before they are plugged
into the CI formula. For this reason, we will speak in terms of proportions for the remainder of the discussion.
There are sampling distributions for proportions just as there are for means. Because of sampling error, a
sample proportion (symbolized , pronounced “p hat”) cannot be assumed to equal the population proportion
(symbolized as an uppercase P) , so confidence intervals must be created in order to estimate the population
values with a certain level of probability.
Research Example 8.2 What Factors Influence Repeat Offenders’ Completion of a “Driving Under the Influence” Court Program?
Specialized courts are an increasingly popular way for dealing with low-level offenders, especially those who have drug or mental-
health problems. The rationale is that these people should not be incarcerated in jail or prison and should instead be allowed to
remain in the community and complete one or more treatment programs to help them with their problems. Judges are responsible
for supervising the defendants in these courts. Defendants typically appear before the judge every month or two, and the judge
praises them when they have done well in their program and reprimands them when they have failed to follow through on an
assignment. Usually, charges are dropped when defendants complete the program successfully, and those who drop out or are
removed for noncompliance get sentenced to a previously agreed-on penalty (such as a specified jail or probation term).
In recent years, courts dedicated to handling people convicted of driving under the influence (DUI) of drugs or alcohol have
appeared as a new incarnation of specialized courts. The success of these courts at reducing recidivism hinges on their ability to keep
defendants in the program; defendants who drop out have to be sentenced to jail or probation (which is more expensive) and will not
experience the benefits of the treatment regimen. Saum, Hiller, and Nolan (2013) sought to identify the factors associated with
treatment completion versus dropout. They gathered records on 141 third-time DUI offenders who went through a DUI court
program in Wisconsin. Most of the defendants (114) completed the program, but 27 did not. Overall, the researchers found very few
differences between the groups. Having a mental-health problem did not affect the odds of completion, as shown by the wide
confidence interval for this variable’s predictive capability (.17 to 1.6). The only variable that emerged as significant was the number
of days in the jail or work-release sentence; those who had been threatened with more-severe sentences were more likely to drop out.
This variable’s confidence interval was very small (.95 to 1.00), suggesting that is was a good predictor of program completion versus
program dropout. It would appear that more-serious DUI offenders need enhanced supervision and greater incentives to stay in and
successfully complete DUI court programs.
245
How Extensively Do News Media Stories Distort Public Perceptions
About Racial Minorities’ Criminal Involvement?
Race is inextricably linked with crime in the public’s mind. Numerous stereotypes paint blacks and Latinos, in particular, as being
responsible for the majority of drug trafficking and violence that takes place in the United States. The media both reflect and fuel
these stereotypes, and have a significant influence on the impression the public forms about the “color” of crime. The media do this
by choosing which crimes to cover and how to portray the victim and offender in any given violent event. Dixon (2017) sought to
examine media (mis-)representation of offender and victim race. Dixon’s study added to the existing work on the subject by
including Spanish-language television stations, which previous studies had not done. He drew a sample of 117 news programs from
the population of programs that aired between 2008 and 2012 in Los Angeles. Each program was coded according to crime type,
victim race, perpetrator race, and the race of responding police officers. To figure out how closely news reporting matched real crime
characteristics, Dixon compared the news programs to official crime data. The tables show the results, presented as percentages,
including the 95% CI for each estimate derived from the news programs. The differentials indicate whether people of each race are
underrepresented or overrepresented as homicide perpetrators, victims, or police officers.
Source: Adapted from Table 2 in Dixon (2017).
Source: Adapted from Table 3 in Dixon (2017).
Source: Adapted from Table 4 in Dixon (2017).
Contrary to much previous research, Dixon’s (2017) analysis did not find that black perpetrators were significantly overrepresented
in news reports; in fact, white perpetrators were overrepresented more often (9 percentage points relative to actual arrest rates) than
black perpetrators were (4 points). Interestingly, Latino perpetrators were notably underrepresented (a 12-percentage point
differential). A very different pattern manifested in the analyses of victims and police officers. Black victims were slightly
246
overrepresented (5 percentage points), while white victims were significantly overrepresented (22 percentage points) and Latino
victims were strikingly underrepresented (a differential of 40). The findings for police officers followed a similar trend. Black officers
were slightly underrepresented (3 percentage points) and Latino officers even more so (14 points), while white officers were
markedly overrepresented (a differential of 20 points). The main conclusions from these results is that news media do not
significantly distort perpetrators’ races, but they heavily emphasize whites in positions of sympathy (victims) or bravery (police
officers). Latinos receive less reporting in all three capacities, particularly as homicide victims. Thus, media might be feeding into
race–crime stereotypes less by emphasizing blacks and Latinos as predators than by consistently portraying whites as innocent
victims and brave police officers.
Confidence intervals for proportions employ the z distribution. The normality of the sampling distribution of
sample proportions is a bit iffy, but, generally speaking, z is a safe bet as long as the sample is large (i.e., N ≥
100) and contains at least five successes and at least five failures (more on this shortly). The formula for
confidence intervals with proportions is
where = the sample proportion.
To illustrate CIs for proportions with large samples, we will again use the 2013 FISS. This time, we will
examine the involvement of handguns in violent altercations. According to the FISS, handguns were the
firearm used in 91.30% of gunshots arising from fights in which the type of gun used is known (i.e., excluding
the cases in which gun type cannot be determined). The sample size is N = 173, which meets the criterion that
the sample have at least 100 cases. Now we have to consider whether the requirement that there be at least
five successes and at least five failures. Within the context of this problem, we will define “success” as the use
of a handgun and “failure” as the use of any other gun type. Handguns were used in 158 instances and other
guns in the remaining 15, so this variable meets the criteria regarding sample size being large enough and
there being more than five successes and more than five failures.
Confidence will be set at 99%, which means z⍺ = ±2.58.
First, the sample percentage needs to be converted to a proportion; dividing 91.30 by 100 and rounding to
two decimal places yields .91.
Next, plug the numbers into Formula 8(5) and solve:
247
= .91 ± 2.58 (0.02)
= .91 ± .05.
The calculation of the lower and upper limits and the formal statement of the confidence interval proceeds
along the same lines for proportions as for means. In the current example,
LL = .91 – .05 = .86
UL = .91 + .05 = .96
and
99%CI : .86 <=P<=.96
We can say with 99% confidence that the interval .86 to .96, inclusive, contains the true population mean P .
248
Learning Check 8.5
Redo the confidence interval for the analysis of handgun usage in fights, this time using a 95% confidence level. What happened to the
width of the interval? Why did this happen?
For a second example, we can again use the FISS and this time analyze the relationship between shooters and
victims in arguments. Among those gunshots resulting from arguments (N = 139), 28.8% of victims were shot
by friends or acquaintances (again, excluding cases in which the victim–shooter relationship was unknown).
This variable meets the sample size requirement; since there were 40 friend/acquaintance shootings (successes)
and 99 cases involving different relationships (failures), the requirement that there be a minimum of 5
successes and failures each is also met. We can calculate a confidence interval. The proportion is 28.8/100 =
.288, which rounds to .29. The confidence level will be set at 95%, making the critical value of z ±1.96. Using
Formula 8(4),
= .29 ± 1.96 (.04)
= .29 ± .08.
The lower and upper limits are
LL = .29 – .08 = .21
UL = .29 + .08 = .37
Finally, the confidence interval is
95% CI : .21 ≤ P ≤ .37
There is a 95% chance that the interval .21 to .37, inclusive, contains the true population proportion P .
Research Example 8.3 Is There a Relationship Between Unintended Pregnancy and Intimate Partner Violence?
Intimate partner violence (IPV) perpetrated by a man against a female intimate is often associated with not just physical violence but
249
also a multifaceted web of control in which the woman becomes trapped and isolated. One of the possible consequences is that
abused women might not have access to reliable birth control and might therefore be susceptible to unintended pregnancies. These
pregnancies, moreover, might worsen the IPV situation because of the emotional and financial burden of pregnancy and
childbearing. Martin and Garcia (2011) sought to explore the relationship between IPV and unintended pregnancy in a sample of
Latina women in Los Angeles. Latina immigrants might be especially vulnerable to both IPV and unintended pregnancy because of
the social isolation faced by those who have not assimilated into mainstream U.S. society. Staff at various prenatal clinics in Los
Angeles distributed surveys to their Latina patients. The surveys asked women several questions pertaining to whether their
pregnancy had been intentional, whether they had experienced emotional or physical abuse by their partner before or during
pregnancy, and their level of assimilation into the United States (as compared to feeling isolated or alienated). They used a statistic
called an odds ratio . An odds ratio measures the extent to which consideration of certain independent variables (IVs) changes the
likelihood that the dependent variable (DV) will occur. An odds ratio of 1.00 means that the IV does not change the likelihood of
the DV. Odds ratios greater than 1.00 mean that an IV increases the probability that the DV will occur, and those less than 1.00
indicate that an IV reduces the chances of the DV happening.
In Martin and Garcia’s (2011) study, the DVs in the first analysis were physical and emotional abuse and the DV in the second
analysis was physical abuse during pregnancy. The researchers found the following odds ratios and confidence intervals.
Prepregnancy physical abuse was not related to the chances that a woman would become pregnant accidentally, as indicated by the
odds ratio of .92, which is very close to 1.00. Emotional abuse, surprisingly, significantly reduced the odds of accidental pregnancy
(odds ratio = .50). This was an unexpected finding because the researchers predicted that abuse would increase the odds of
pregnancy. They theorized that perhaps some emotionally abused women try to get pregnant out of a hope that having a child will
improve the domestic situation.
Turning to the second analysis, it can be seen that unintended pregnancy substantially increased women’s odds of experiencing
physical abuse during pregnancy. Contrary to expectations, women’s level of acculturation did not alter the odds of abuse or
pregnancy, once factors such as a woman’s age and level of education were accounted for. The findings indicated that the
relationship between IPV and unintended pregnancy is complex and deserving of further study in order to identify risk factors for
both.
250
Why Do Suspects Confess to Police?
The Fifth Amendment to the U.S. Constitution that provides protection from compelled self-incrimination ensures that persons
suspected of having committed criminal offenses can refuse to answer questions about the crime during police interrogations.
Despite this right, a good number of suspects do talk and do provide incriminating statements. As the wry saying goes, “Everyone
has the right to remain silent, but not everyone has the ability to do so.” So what makes suspects confess? Deslauriers-Varin,
Beauregard, and Wong (2011) sought to identify some of the contextual factors that make suspects more likely to confess, even when
those suspects initially indicated that they wished to remain silent. The researchers obtained a sample of 211 convicted male
offenders from a Canadian prison and gathered extensive information about each participant. The researchers analyzed odds ratios.
Recall that an odds ratio of 1.00 means that the IV does not alter the odds that the DV will occur. Odds ratios less than 1.00 mean
the IV makes the DV less likely, whereas odds ratios greater than 1.00 indicate that the IV makes the DV more likely to happen.
The researchers found the following odds ratios and confidence intervals.
The numbers in the following table that are flagged with asterisks are those that are statistically significant, meaning that the IV
exerted a noteworthy impact on suspects’ decision to remain silent rather than confessing. The initial decision to not confess was the
strongest predictor of suspects’ ultimate refusal to provide a confession; those who initially resisted confessing were likely to stick to
that decision. Criminal history was also related—having only one or two priors was not related to nonconfession, but suspects with
three or more prior convictions were substantially more likely to remain silent. This might be because these suspects were concerned
about being sentenced harshly as habitual offenders. The presence of an accomplice also made nonconfession more likely, as did a
lawyer’s advice to not confess. Those accused of drug-related crimes were more likely to not confess, though crime type was not
significant for suspects accused of other types of offenses. Finally, the strength of police evidence was a factor in suspects’ decisions.
Strong police evidence was related to a significant reduction in the odds of nonconfession (i.e., strong evidence resulted in a greater
chance that the suspect would confess).
It appeared, then, that there are many factors that impact suspects’ choice regarding confession. Most of these factors appear to be
out of the control of police interrogators; however, the researchers did not include variables measuring police behavior during
interrogation, so there might well be techniques police can use to elicit confessions even from those suspects who are disinclined to
offer information. One policy implication from these results involves the importance of the initial decision in the final decision—
79% of offenders stuck with their initial decision concerning whether to confess. Police might, therefore, benefit from focusing their
efforts on influencing the initial decision rather than allowing a suspect to formulate a decision first and then applying interrogation
tactics.
Source: Adapted from Table 2 in Deslauriers-Varin et al. (2011).
251
Statistically significant.
We can also construct a confidence interval using the COJ to estimate the proportion of jail inmates being
held for felonies. In a random sample of 177 facilities, 60.7% of inmates were incarcerated for felonies (either
awaiting trial or after having been convicted). The sample size is large enough to permit the use of the z
distribution, and there are far more than five successes (people confined for felonies) and failures (people held
for misdemeanor and other types of offenses). The threshold criteria having been satisfied, we can proceed to
the calculations. We will use the 95% confidence level. Dividing 60.7 by 100 and rounding yields a proportion
of .61; the computation is
= .61 ± 1.96 (.03)
= .61 ± .06
The lower and upper limits are
LL = .61 – .06 = .55
UL = .61 + .06 = .67
Finally, the confidence interval is
95% CI : .55 ≤ P ≤ .67
There is a 95% chance that the interval .55 to .67, inclusive, contains the true population proportion P . Since
this is a census, we can check to see if our interval actually does contain the population proportion. The
proportion of inmates held for felonies across the entire sample is 60.10%. Not only does the calculated
interval contain the mean; it also turns out that the sample mean is nearly identical to the population mean!
Of course, this is information researchers typically would not have at their disposal.
Chapter Summary
Confidence intervals are a way for researchers to use sample statistics to form conclusions about the probable values of population
parameters. Confidence intervals entail the construction of ranges of values predicted to contain the true population parameter. The
researcher sets the level of confidence (probability of correctness) according to her or his judgment about the relative costs of a loss of
252
confidence versus compromised precision. The conventional confidence levels in criminal justice and criminology research are 95%
and 99%. The decision about level of confidence must be made with consideration to the trade-off between confidence and
precision: As confidence increases, the quality of the estimate diminishes. All confidence intervals are two-tailed, which means that
alpha is divided in half and placed in both tails of the distribution. This creates two critical values. The critical values have the same
absolute value, but one is negative and one is positive.
There are two types of CIs for means: large sample and small sample. Confidence intervals for means with large samples employ the
z distribution, whereas those for means with small samples use the t curve. When the t distribution is used, it is necessary to calculate
degrees of freedom in order to locate the critical value on the t table.
Confidence intervals can be constructed on the basis of proportions, providing that two criteria are met: First, the sample must
contain at least 100 cases. Second, there must be at least five successes and five failures in the sample. These two conditions help
ensure the normality of the sampling distribution and, thus, the applicability of the z curve.
Thinking Critically
1. Two researchers are arguing about confidence intervals. One of them contends that precision is more important than the
level of confidence, and the other claims that confidence is more important than precision. What would you say to them to
help resolve this disagreement? Identify a research scenario in social science or other sciences (medicine, biology, etc.) in
which precision would be of utmost concern and one in which confidence would be researchers’ main goal. Explain your
rationale for each.
2. Suppose you read a research report that claimed to have uncovered a finding with significant implications for policy. The
researchers summarize the results of a multicity homicide reduction program that they claim was markedly effective. They
support this claim by pointing out that, during the study period, there was a mean reduction of 1.80 homicides per month
across the cities in the sample. Along with that mean, they report a 95% confidence interval of –.40 ≤ µ ≤ 4 (i.e., 1.80 ± 2.20).
What is wrong with the researchers’ claim that this program was effective? Identify the flaw in their logic and explain why
that flaw makes their claim unsupportable.
Review Problems
Answer the following questions with regard to confidence intervals .
1. How many cases must be in a sample for that sample to be considered “large”?
2. “Small” samples are those that have ____ or fewer cases.
3. Which distribution is used with large samples?
4. Which distribution is used with small samples?
5. Why can the distribution that is used with large samples not also be used with small ones?
6. Explain the trade-off between confidence and precision.
7. The Law Enforcement Management and Administrative Statistics (LEMAS) survey asks agencies to report the number of
sworn personnel who are designated school resource officers (SROs). In Florida municipal police departments serving
populations of 1 million or more, the agencies sampled in LEMAS (N = 18) reported a mean of 13.00 SROs (s = 12.10).
Construct a 95% confidence interval around this sample mean, and interpret the interval in words.
8. The sheriff’s offices sampled in LEMAS (N = 827) reported that their agencies require new recruits to complete a mean of
599 hours of academy training (s = 226). Construct a 99% confidence interval around this mean and interpret the interval in
words.
9. The PPCS asks respondents who have been stopped by the police while driving a vehicle how many officers were on the
scene during the stop. Female stopped drivers (N = 2,033) reported a mean of 1.14 (s = .36) officers. Construct a 95%
confidence interval around this sample mean, and interpret the interval in words.
10. In the PPCS, respondents who say that they have been stopped by police while driving vehicles are asked to report the reason
why they were stopped and the total length of time that the stop took. Among female drivers stopped for illegal use of
cellphones while driving (N = 42), stops lasted a mean of 9.00 minutes (s = 5.48). Construct a 95% confidence interval
around this sample mean and interpret the interval in words.
11. The General Social Survey (GSS) contains an item asking respondents about their TV habits. Female respondents (N = 707)
253
said they watched a mean of 3.06 (s = 2.62) hours of TV per day. Construct a 99% confidence interval around this sample
value and interpret the interval in words.
12. In a random sample of jails (N = 110) from the COJ, the mean number of male juveniles confined per facility was 1.17 (s =
6.21). Construct a 99% confidence interval around this sample value and interpret the interval in words.
13. In a random sample of prisons (N = 30) from the COJ, the mean number of inmates per security-staff member was 4.34 (s =
3.94). Construct a 95% confidence interval around this sample value and interpret the interval in words.
14. Respondents to the GSS (N = 1,166) worked a mean of 40.27 hours per week (s = 15.54). Construct a 99% confidence
interval around this sample value, and interpret the interval in words.
15. The GSS asks respondents whether they keep a gun in their homes. In this survey, 34% of respondents (N = 1,281) said that
they have at least one firearm. Construct a 95% confidence interval around this sample value and interpret the interval in
words.
16. In the LEMAS survey, 45% of sampled sheriff’s offices (N = 823) reported that their agencies’ formal mission statements do
not include a community-policing component. Construct a 95% confidence interval around this sample value, and interpret
the interval in words.
17. According to LEMAS, 31% of municipal law-enforcement agencies (N = 1,967) use computerized statistics to identify high-
crime hot spots. Construct a 95% confidence interval around this sample value and interpret the interval in words.
18. The FISS captures information on whether drugs were involved in the altercation that led up to a shooting. In 2013, victims
were male in 87.3% of the 142 cases that involved drugs. Construct a 99% confidence interval around this sample value, and
interpret the interval in words.
19. The FISS reports that of the 329 female shooting victims whose relationship with their shooter was known, 28.3% were shot
by strangers. Construct a 95% confidence interval around this sample value, and interpret the interval in words.
20. The GSS found that 74% of male respondents (N = 573) believe that people suffering from incurable diseases should be
permitted to die if that is their choice. Construct a 99% confidence interval around this sample value and interpret the
interval in words.
254
Key Terms
Point estimate 177
Confidence interval 177
Level of confidence 178
Alpha level 180
Critical value 180
Two-tailed test 180
Glossary of Symbols and Abbreviations Introduced in This Chapter
255
Part III Hypothesis Testing
Chapter 9 Hypothesis Testing: A Conceptual Introduction
Chapter 10 Hypothesis Testing With Two Categorical Variables: Chi-Square
Chapter 11 Hypothesis Testing With Two Population Means or Proportions
Chapter 12 Hypothesis Testing With Three or More Population Means: Analysis of Variance
Chapter 13 Hypothesis Testing With Two Continuous Variables: Correlation
Chapter 14 Introduction to Regression Analysis
You have now learned about descriptive statistics (Part I) and the theories of probability and distributions that
form the foundation of statistics (Part II). Part III brings all of this together to form what most criminal
justice and criminology researchers consider to be the high point of statistics: inferential analyses or what is
also called hypothesis testing. Hypothesis testing involves using a sample to arrive at a conclusion about a
population. Samples are vehicles that allow you to make generalizations or predictions about what you believe
is happening in the population as a whole. The problem, as we discussed in Chapter 7, is sampling error: It is
erroneous to conclude that a sample statistic is an accurate reflection of the population parameter, because the
sample is merely one of a multitude of samples that could have been drawn from the population and,
therefore, there is an unknown amount of error. Refer back to Figures 7.5 and 7.6 for an illustration.
Inferential analysis: The process of generalizing from a sample to a population; the use of a sample statistic to
estimate a population parameter. Also called hypothesis testing.
To account for sampling error, inferential statistics use sampling distributions to make probabilistic
predictions about the sample statistic that is being analyzed. The basic strategy is a two-step process: First, a
sample statistic is computed. Second, a probability distribution is used to find out whether this statistic has a
low or high probability of occurrence. This process should sound very familiar—we already followed these
steps when we worked with z scores and areas under the standard normal curve. Hypothesis testing is an
expansion of this underlying idea. A sample statistic (such as a mean) can be used to find out whether this
value is close to the center of the distribution (i.e., has a high probability of occurrence) or far out in the tail
(low probability of occurrence). It is this probability assessment that guides researchers in making decisions
and reaching conclusions.
Part III covers some of the bivariate (i.e., involving two variables) inferential tests commonly used in
criminology and criminal justice research: chi-square tests of independence, two-population tests for
differences between means and between proportions, analyses of variance, and correlations. Part III ends with
an introduction to bivariate and multiple regression.
Levels of Measurement
256
The proper test to use in a given hypothesis-testing situation is determined by the level of measurement of the
variables with which you are working. If your memory of levels of measurement has become a bit fuzzy, go
back to Chapter 2 now and review this important topic. You will not be able to select the correct analysis
unless you can identify your variables’ levels of measurement. The figure here is a reproduction of Figure 2.1
showing the levels of measurement. For purposes of hypothesis testing, the most important distinction is that
between categorical and continuous variables. Be sure you can accurately identify any given variable’s level of
measurement before you proceed.
There is no sense denying that you might find many of the concepts presented in the following chapters
confusing at first. This is normal! Remember that the process of learning statistics hinges on repetition. Read
and reread the chapters, study your lecture notes, and do the chapter review problems at least one time each—
things will start to sink in. Terms, formulas, and ideas that initially seemed incomprehensible will gradually
solidify in your mind and begin making sense. Remember that most criminology and criminal justice
researchers started off in a position just like yours! There was a point when they knew very little about
statistics and had to study hard to develop a knowledge base. Commit the time and effort and you will be
pleasantly surprised by how well you do.
Learning Check
Take a moment now to test your memory of levels of measurement. Identify the level of measurement of each
of the following variables:
1. The survey item that asks, “How many times have you been arrested in your life?” and has respondents
write in the answer
2. The survey item that asks, “How many times have you been arrested in your life?” and has respondents
circle never, 1–3 times, or 4 or more times
3. Whether a convicted defendant’s sentence was jail, probation, or drug treatment
4. Defendants’ annual household income, measured as $0–$9,999, $10,000–$19,999, or $20,000 or more
5. The method by which a defendant was convicted, measured as guilty plea, jury trial, or bench trial
6. In a sample of people responding to a survey, attitudes about the court system ranging from “1 = too
harsh” to “5 = too lenient.”
257
7. Now you try it! Create four variables, one representing each level of measurement.
258
Chapter 9 Hypothesis Testing A Conceptual Introduction
259
Learning Objectives
Explain the difference between expected and observed outcomes.
Identify the two possible explanations for differences between observed and expected outcomes and explain how probability is used
to determine which of these potential explanations is the true explanation.
Define the null hypothesis and the alternative hypothesis and write each out in both words and symbols.
Summarize the logic behind the assumption that the null is true.
Define Type I and Type II errors and explain the trade-off between them.
List the four types of bivariate inferential tests, and identify the correct one to use for any given independent variable and
dependent variable combination depending on level of measurement.
The purpose of this chapter is to provide a clear conceptual foundation explaining the nature and purpose of
hypothesis testing. It is worth developing a solid understanding of the substance of inferential statistics before
approaching the specific types of hypothesis tests covered in the chapters that follow. This chapter will help
you grasp the overarching logic behind these sorts of tests so that you can approach the “trees” (specific tests)
with a clear picture of what the “forest” (underlying conceptual framework) looks like.
By now, you should be very familiar with the idea that researchers are usually interested in populations but,
because populations are so large, samples must suffice as substitutes. Researchers draw random samples using
a variety of methods. A researcher distributing surveys by mail might use a police department’s address
database to electronically pull the addresses of 10% of the residences in a particular city. Someone conducting
phone surveys might use random-digit dialing to contact 200 respondents. The ultimate goal in statistical
research is to generalize from the sample to the population. Hypothesis testing is the process of making this
generalization. (Note that throughout the following discussion, we are going to assume that samples are
simple and random. When either of these two criteria is not true in a given sample, adjustments sometimes
have to be made to the statistics used to analyze them. For present purposes, we are going to assume we are
working with simple, random samples.)
What we are really talking about in inferential statistics is the probability of empirical outcomes. There are
many (even infinite) random samples that can be drawn from any given population and, therefore, are
numerous possible values for sample statistics to take on. A population with a mean of 10, for instance, can
produce samples with means of 9, 11, 10, 7, and so on. When we have a sample statistic that we wish to use
inferentially, the question asked is, “Out of all the samples and sample statistics possible, what is the
probability that I would draw this one?”
260
Learning Check 9.1
Remember the difference between expected and observed or empirical outcomes. Expected outcomes are the results you anticipate seeing
on the basis of probability theory. In Chapter 6, you constructed a table of expected outcomes in the context of binomials. You did not use
any data; you made this table entirely on the basis of probability calculations. Observed outcomes, by contrast, are what you actually see.
These results might or might not mirror expectations. A coin-flip exercise will help refresh your memory. Write down the probability of
any given coin flip resulting in heads. Now flip a coin six times and record each outcome; then tally up the total number of heads and tails.
Did you see what you expected, on the basis of the underlying probability, or were you surprised at the outcome? Try again with 10 flips.
Did the observed outcome match the one you expected?
261
Sample Statistics and Population Parameters: Sampling Error or True
Difference?
Any time a sample statistic is not equal to a population parameter, there are two potential explanations for the
difference. (Well, technically there are three, since a mismatch can result from mistakes in the sampling
process. For our purposes, though, as mentioned earlier, we are assuming correct research methods and
simple, random samples.) First, the inequality could be the product of inherent random fluctuations in sample
statistics (i.e., sampling error). In other words, the disparity might simply be a meaningless fluke. If you
flipped a fair coin six times, you would expect the coin to land tails side up three times. If, instead, you got
four tails, you would not think that there was anything weird happening; the next set of six trials might result
in two tails. This is sampling error—variation that is like white noise in the background.
The second possible explanation for the difference is that there is a genuine discrepancy between the sample
statistic and the population parameter. In other words, the disparity could represent a bona fide statistical
effect. If you flipped a coin 20 times and got 19 tails, you would suspect there was something wrong with the
coin because this is an extremely unlikely outcome. Perhaps the coin is weighted on one side, which would
mean that it is different from the ordinary quarter or dime you might have in your pocket. Large discrepancies
between observed and expected outcomes are sufficiently improbable to lead us to conclude that there is
something genuinely unique about the empirical outcome we have in front of us.
When researchers first approach an empirical finding, they do not know which of the two possible
explanations accounts for the observed or empirical result. In the earlier coin-flip example, we knew the
underlying population probability (.50), and we knew the number of trials that had been conducted (six in the
first and 20 in the second). If either of those pieces of information is omitted, then it becomes difficult to
make sense of the results. If your friend told you that he flipped a coin and it landed on tails seven times, but
he did not tell you the total number of times he flipped it, then you would not know how to interpret his
report about seven tails. Similarly, if you did not know that every flip has a .50 probability of tails (and that, by
extension, roughly half of a string of flips will be tails), then your friend might say that he flipped a coin 14
times and got seven tails and you would not have a clue as to whether this result is normal or atypical.
In the real world of criminal justice and criminology research, there are missing bits of information that
prevent us from being able to readily discriminate between sampling error and true difference. The
overarching purpose of hypothesis testing is to determine which of them appears to be the more valid of the
two, based on a calculated judgment. This is where probabilities come in. Researchers identify the probability
of observing a particular empirical result and then use that probability to make a decision about which
explanation seems to be correct.
This is pretty abstract, so an example is in order. We will use the Law Enforcement Management and
Administrative Statistics (LEMAS; see Data Sources 3.2) survey. Suppose we are investigating whether
agency type (municipal, county, tribal, or state) affects the ratio of officers to residents within a given
jurisdiction. (This is a measure of agency size relative to the size of the population served.) We might predict
262
that municipal departments have a higher officer-to-resident ratio than county sheriff’s offices do. This is
because city and town police departments typically serve more-centralized populations, and the concentration
is linked with higher crime rates. By contrast, sheriffs’ offices generally cover larger territories and populations
that are more spread out, with lower crime rates. Using LEMAS, we can calculate the mean ratio of officers
per 1,000 residents for each type of agency in the sample. Municipal departments’ mean is 2.20, and sheriffs’
offices mean is 1.08.
263
Learning Check 9.2
In the example using police agencies, we set up a hypothetical study exploring the possible effect that agency type has on officer-to-
resident ratio. Identify the two variables in this study, and state which is the independent variable (IV) and which is the dependent
variable (DV).
You can see that these means are unequal, and you might be tempted to conclude that our hypothesis has
been supported (i.e., that municipal departments are indeed larger than sheriffs’ offices, relative to the size of
the population served). However, recall that there are two potential reasons for this disparity. Their inequality
might be meaningless and the differences between the numbers purely the product of chance; in other words,
the finding might be a fluke. This is a random sample of agencies, so we cannot rule out the possibility that
sampling error manufactured an apparent difference where there actually is none. On the other hand, the
means might be unequal because officer-to-resident ratios really do vary across these two types of agencies. In
other words, there might be a real difference between the means. These two competing potential explanations
can be framed as hypotheses. We will use the LEMAS numbers to guide us through a discussion of these
competing hypotheses.
264
Null and Alternative Hypotheses
There are two hypotheses used to state predictions about whether or not a statistic is an accurate estimate of a
parameter. The first is called the null hypothesis. The null (symbolized H0) represents the prediction that the
difference between the two samples is merely sampling error or, in other words, chance variation in the data.
You can use the word null as its own mnemonic device because this word means “nothing.” Something that is
null is devoid of meaning. In the context of the present example, the null predicts that agency type is not
related to officer-to-resident ratio and that the observed difference between the means is just white noise.
Null hypothesis: In an inferential test, the hypothesis predicting that there is no relationship between the independent and
dependent variables. Symbolized H0.
The second possible explanation is that municipal departments really do have a higher ratio than sheriffs’
offices do. This prediction is spelled out in the alternative hypothesis (symbolized H1). The alternative
hypothesis is sometimes also called the research hypothesis. The alternative or research hypothesis is,
essentially, the opposite of the null: The null predicts that there is no relationship between the two variables
being examined, and the alternative predicts that they are related.
Alternative hypothesis: In an inferential test, the hypothesis predicting that there is a relationship between the independent and
dependent variables. Symbolized H1. Also referred to as a research hypothesis.
In the context of the present example, the null and alternative hypotheses can be written as
H0: Municipal departments and sheriffs’ agencies have the same officer-to-resident ratio; that is, the two means
are equal, and there is no relationship between agency type and ratio.
H1: Municipal departments have a higher officer-to-resident ratio than sheriffs’ agencies do; that is, the means
are unequal and there is a relationship between agency type and ratio.
More common than writing the hypotheses in words is to use symbols to represent the ideas embodied in the
longhand versions. Transforming these concepts into such symbols turns the null and alternative hypotheses
into
H0: µ1 = µ2
H1: µ1 > µ2,

where µ1 = the mean ratio in municipal departments, and

µ2 = the mean ratio in sheriffs’ offices.

It might seem strange to write the hypotheses using the symbol for the population mean (µ) rather than the

265

sample mean ( ), but remember that in inferential statistics, it is the population parameter that is of interest.

We use the sample means to make a determination about the population mean(s). Basically, there are two

options: There might be one population from which the samples derive (sampling error), or each sample

might represent its own population (true difference). If the null is true and there is, in fact, no relationship

between agency type and officer-to-resident ratio, then we would conclude that all the agencies come from the

same population. If, instead, the alternative is true, then there are actually two populations at play here—

municipal and county. Figure 9.1 shows this idea pictorially.

Figure 9.1 One Population or Two?

Another reason a researcher might conduct a hypothesis test is to determine if men and women differ in terms

of how punitive they feel toward people who have been convicted of crimes. The General Social Survey (GSS;

see Data Sources 2.2) asks people whether they favor or oppose capital punishment for people convicted of

murder. Among men, 31% oppose capital punishment; this number is 40% among women. A researcher

might want to know whether this difference represents a genuine “gender effect” or whether it is merely

chance variation. The null and alternative could be set up as such:

H0: Men and women oppose the death penalty equally; the proportions are equal and there is no relationship

between gender and death penalty attitudes.

H1: Men are less likely to oppose the death penalty than women are; the proportions are unequal and there is a

relationship between gender and death penalty attitudes.

Formally stated using symbols, the null and alternative are written as

H0: P1 = P2

H1: P1 < P2 where P1 = the proportion of men who oppose capital punishment and P2 = the proportion of women who oppose capital punishment. As with the example regarding police agency type and size, the hypotheses pertaining to gender and death 266 penalty attitudes are phrased in terms of the population proportion (P) rather than the sample proportion ( ) because we are using a sample estimate to draw an inference about a population parameter. Until we conduct a full hypothesis test, we will not know whether women and men truly do differ in their opposition to capital punishment. The assumption going into an inferential analysis is that the null is the true state of affairs. In other words, the default assumption is that there is no relationship between the two variables under examination. The goal in conducting the test is to decide whether to retain the null (concluding that there is no relationship) or to reject the null (concluding that there is, in fact, a relationship between the variables). The null can be rejected only if there is solid, compelling evidence that leads you to decide that this hypothesis is inaccurate. A good analogy to the logic behind hypothesis testing is the presumption of innocence in a criminal trial. At the outset of a trial, the jury must consider the defendant to be legally innocent of the crime of which she or he is accused. The “null” here is innocence and the “alternative” is guilt. If the prosecutor fails to convincingly show guilt, then the innocence assumption stands, and the defendant must be acquitted. If, however, the prosecutor presents sufficient incriminating evidence, then the jury rejects the assumption of innocence and renders a guilty verdict. Of course, the prosecutor does not have to leave the jury with complete certainty that the defendant is guilty; rather, the prosecutor’s job is to overcome a reasonable doubt. In other words, the jury can convict when the probability of the defendant’s guilt far outweighs the probability of his innocence. There are good reasons for the null being the default in both criminal trials and scientific research. The presumption of innocence helps prevent wrongful convictions. Likewise, in clinical trials testing new pharmaceutical drugs, it is of utmost importance that a drug be demonstrated to be effective and safe before it is approved and put on the market. Medical researchers err on the side of caution. If there is a .50 probability that a drug will help the people who take it achieve better health, then this might be sufficient to send that drug to market. On the other hand, if a drug has a .60 probability of making people who take it feel better and a .15 probability of killing them, researchers should not let it be released for public consumption. In criminal justice and criminology research, important questions about theory and policy hang in the balance. Like medical researchers, social scientists err on the side of caution. A criminologist testing a hypothesis about a theory of crime causation must tread carefully. Failing to reject the null hypothesis could lead to erroneous conclusions about the accuracy of the theory; however, the test can be repeated to determine whether there was a mistake in the research design that corrupted the results. Rejecting a null hypothesis that is true, however, could take this criminologist and others doing similar research down a completely wrong path in the study of crime causation. Criminal justice researchers, likewise, often deal with policy questions. They handle matters such as whether or not a particular rehabilitation program reduces recidivism, whether a given policing strategy deters violent crime, and whether school-based prevention programs prevent drug use and gang involvement among youth. For each of these examples, you can see that neither of the two possible mistakes is harmless but that retaining the null (and, possibly, conducting further research into the issue) is safer than leaping to a conclusion that might be incorrect. Erroneously concluding that a rehabilitation program, police strategy, or school-based prevention program works could result in their widespread adoption 267 across jurisdictions, even though they are ineffective. In inferential statistics, researchers construct a probability framework based on the assumption of a true null. The question they are trying to answer is, “What is the probability of observing the empirical result that I see in front of me if the null hypothesis is correct ?” If the probability of the null being true is extremely low, then the null is rejected because it strains the imagination to think that something with such a low likelihood of being correct is the right explanation for an empirical phenomenon. The alternative would, thus, be taken as being the more likely version of reality. If the probability of the null being true is not low, then the null is considered to be a viable explanation for the results, and it is retained. Think back to the study of police agency type and the ratio of officers per 1,000 local residents. Recall that municipal departments’ mean officer-to-resident ratio is 2.20 and sheriff’s offices is 1.08. Let us say, for the sake of example, that we determine that the probability of these two means being two parts of the same population is .70. That is a pretty high probability! We would conclude that the null is likely correct and there is only one population. What if we found that the probability was .20? This is a much smaller probability than .70, but it still means that there is a 20% chance that the null is true, which is substantial. A jury in a criminal trial should not convict a defendant if there is a 20% chance he is innocent. If, on the other hand, we found a probability of .01, meaning that there is only a 1% chance that the two samples are from the same population, then rejecting the null would be warranted because it is extremely unlikely to be the true state of affairs (i.e., there is a .99 or 99% chance that the null is false). At a probability of .01, it is highly likely that municipal police departments and county sheriff’s agencies are two separate populations with different means. Of course, as you have probably already figured out, there is always a chance that a researcher’s decision regarding whether to reject or retain the null is wrong. We saw when we worked with confidence intervals that the flipside of the probability of being right is the probability of being wrong—any time you make a decision about the null, you are either right or wrong, so the two probabilities sum to 100% or 1.00. There are two types of errors that can be made in this regard. A Type I error occurs when a true null is erroneously rejected, whereas a Type II error happens when a false null is inaccurately retained. Type I errors are like false positives, and Type II errors are like false negatives. Wrongfully convicting an innocent defendant or approving an ineffective drug for the market is a Type I error, while incorrectly acquitting a guilty defendant or concluding that an effective drug does not work is a Type II error. Type I error: The erroneous rejection of a true null hypothesis. Symbolized ⍺. Type II error: The erroneous retention of a false null hypothesis. Symbolized β . Type I errors are often symbolized using ⍺, which we have seen before. Recall from the confidence interval lesson in Chapter 8 that alpha is the probability of being wrong about the interval containing the true population mean or proportion. The interpretation of alpha in inferential statistics is a little different from confidence intervals because of the involvement of null and alternative hypotheses; however, the underlying logic is the same. The symbol for a Type II error is the uppercase Greek letter beta (β) . Researchers can often minimize the probability that they are wrong about a decision, but they can never eliminate it. For this reason, 268 you should always be circumspect as both a producer and a consumer of statistical information. Never rush haphazardly to conclusions. Any time you or anyone else runs a statistical analysis and makes a decision about the null hypothesis, there is a probability—however minute—that the decision is wrong. There is a trade-off between Type I and Type II error rates. The Type I error rate (⍺) is set a priori (in advance) of the start of the hypothesis test. A researcher who is worried about making a Type I error could help minimize the chance of this mistake occurring by setting alpha very low, which increases the difficulty of rejecting the null hypothesis. The flipside, however, is that making it harder to reject a true null also makes it hard to reject a false one. By reducing the chance of a Type I error, the researcher has increased the risk of making a Type II error. The following chapters will cover several different types of hypothesis testing procedures in the bivariate (i.e., two variables) context. The choice between the different tests is made on the basis of each variable’s level of measurement. You must identify the levels of measurement of the IVs and DVs and then select the proper test for those measurement types. This book covers four types of bivariate inferential tests. The first is chi-square, which is the statistical procedure used to test for an association between two categorical (nominal or ordinal) variables. If, for example, you had a sample of criminal defendants and wanted to find out whether there was a relationship between the type of crime a defendant was charged with (violent or property) and the disposition method that the defendant chose (guilty plea, jury trial, or bench trial), you would use a chi-square test. The second type of analysis is a t test. The t test is used when the DV of interest is continuous (interval or ratio) and the IV is categorical with two classes (such as gender measured as male or female). The t test is a test for differences between two means. In the officer-to-resident ratio example, a t test would be the analysis we select to determine whether or not to reject the null hypothesis (since we had two types of police agencies and the ratio variable is continuous). The third type of test is the analysis of variance (ANOVA). The ANOVA is an extension of the t test and is used when the DV is continuous and the IV is categorical with three or more classes. If we added state police agencies to our ratio analysis, we would use an ANOVA instead of a t test. The rationale behind this is that conducting multiple t tests is time consuming and cumbersome, and creates statistical problems. The ANOVA streamlines the process by using a single analysis across all classes of the IV. The final bivariate inferential test we will discuss is correlation. This test is used when both the DV and the IV are continuous. Correlations are tests for linear relationships between two variables. You might predict, for instance, that the level of correctional officer staffing in a prison (i.e., the inmate-to-staff ratio) affects the assault rate—it stands to reason that higher staffing results in better supervision and, thus, fewer assaults. You would test this prediction using a correlation analysis. Table 9.2 is a handy chart that you should study closely and refer to repeatedly throughout the next few chapters. 269 If you choose the wrong test, you will arrive at an incorrect answer. This is true in both hand calculations and SPSS programming. SPSS rarely gives error messages and will usually run analyses even when they are deeply flawed. Remember, GIGO! When garbage is entered into an analysis, the output is also garbage. You must be knowledgeable about the proper use of these statistical techniques, or you risk becoming either a purveyor or a consumer of erroneous results. Before we leave this chapter and dive into inferential analyses, let’s introduce the steps of hypothesis testing. It is useful to outline a framework you can use consistently as you learn the different types of analyses. This lends structure to the learning process and allows you to see the similarities between various techniques. Hypothesis testing is broken down into five steps, as follows: Step 1. State the null (H0) and alternative (H1) hypotheses. The two competing hypotheses that will be tested are laid out. Step 2. Identify the distribution and compute the degrees of freedom (df) . Each type of statistical analysis uses a certain probability distribution. You have already seen the z and t distributions, and more will be introduced in later chapters. You have encountered the concept of degrees of freedom (df) in the context of the t distribution. Other distributions also require the computation of df. Step 3. Identify the critical value of the test statistic and state the decision rule. The critical value is based on probability. The critical value is the number that the obtained value (which will be derived in Step 4) must exceed in order for the null to be rejected. The decision rule is an a priori statement formally laying out the criteria that must be met for the null to be rejected. The decision rule is useful because it makes it very clear what must happen in order for the null to be rejected. You will return to the decision rule in Step 5 after computing the critical value in Step 4. Step 4. Compute the obtained value of the test statistic. This is the analytical heart of the hypothesis test. You will select the appropriate formula, plug in the relevant numbers, and solve. The outcome is the obtained value of the test statistic. Step 5. Make a decision about the null and state the substantive conclusion. You will revisit your decision rule from Step 3 and decide whether to reject or retain the null based on the comparison between the critical and obtained values of the test statistic. Then you will render a substantive conclusion. Researchers have the responsibility to interpret their statistical findings and draw substantive conclusions that make sense to other researchers and to the public. 270 Chapter Summary This chapter provided an overview of the nature, purpose, and logic of hypothesis testing. The goal of statistics in criminology and criminal justice is usually generalization from a sample to a population. This is accomplished by first finding a sample statistic and then determining the probability that that statistic would be observed by chance alone. If the probability of the result being attributable solely to chance is exceedingly low, then the researcher concludes that the finding is not due to chance and is, instead, a genuine effect. When a researcher has identified two variables that might be related, there are two possible true or correct states of affairs. The first possibility is that the variables are actually not related. This possibility is embodied by the null hypothesis, symbolized H0. The second possibility is that they are in fact related to one another. This is the alternative hypothesis, H1, which is sometimes also called a research hypothesis. The null hypothesis is always assumed to be the one that is true, and formal hypothesis testing employing probabilities and a sampling distribution is conducted to determine whether there is sufficient evidence to overrule the null and opt for the alternative instead. A hypothesis test using the five steps outlined in this chapter will ultimately result in the null being either rejected or retained, and the researcher concluding that the variables are, or are not, related to each other. Thinking Critically 1. You have been summoned for jury duty and decide at the outset of the trial that you will not vote to convict unless you are 99.9% sure of the defendant’s guilt. Which type of error does this decision minimize? What does this decision do to the probability that you will make the other type of error? If you reduce your certainty threshold to 75%, how have you altered the chances of making each error? 2. A friend of yours is excited about the results of an evaluation of a new rehabilitation program for adult prison inmates. She compared a sample of released prisoners who completed the program while they were incarcerated to a sample who did not, and she found that 21% of the treatment group was rearrested within one year of release, compared to 36% of the no- treatment group. She asserts that this is definitive proof that the program works. What should you tell your friend? Craft a response using the relevant concepts from this chapter. If you plan to recommend a statistical test, which one in Table 9.2 will you suggest? Review Problems 1. Suppose a researcher was studying gender differences in sentencing. She found that males sentenced to jail received a mean of 6.45 months and females sentenced to jail received a mean of 5.82 months. Using what you learned in this chapter, describe the two possible reasons for the differences between these two means. 2. Write the symbol for the null hypothesis, and explain what this hypothesis predicts. 3. Write the symbol for the alternative hypothesis, and explain what this hypothesis predicts. 4. You learned in this chapter that the null is assumed to be true unless very compelling evidence suggests that the alternative hypothesis is actually the correct one. Why is this? That is, what is the rationale for the null being the default? 5. Explain what a Type I error is. 6. Explain what a Type II error is. 7. Explain the trade-off between Type I and Type II error rates. 8. Define the word bivariate. 9. List and describe each of the five steps for hypothesis tests. 10. If you computed an empirical result, identified the probability of observing that result, and found that the probability was high . .. 1. would you conclude that this is the product of sampling error, or would you think that it is a true effect? 2. would you reject or retain the null? 11. If you computed an empirical result, identified the probability of observing that result, and found that the probability was very low . .. 1. would you conclude that this is the product of sampling error, or would you think that it is a true effect? 2. would you reject or retain the null? 271 12. Which inferential statistical analysis would be used if the IV was criminal defendants’ ages at sentencing (measured in years) and the DV length of their terms of confinement (measured in months)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 13. Which inferential statistical analysis would be used if the IV was criminal defendants’ gender (measured as male or female) and the DV was the length of their terms of confinement (measured in months)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 14. Which inferential statistical analysis would be used if the IV was criminal defendants’ gender (measured as male or female) and the DV was whether they obtained pretrial release (measured as yes or no)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 15. Which inferential statistical analysis would be used if the IV was police force size (measured as the number of officers per 1,000 residents) and the DV was crime rates (measured as the number of crimes per 10,000 residents)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 16. Which inferential statistical analysis would be used if the IV was assault victims’ race (measured as white, black, Latino, or other) and the DV was the length of prison terms given to their attackers (measured in months)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 17. Which inferential statistical analysis would be used if the IV was murder victims’ race (measured as white, black, Latino, or other) and the DV was whether the killers were sentenced to death (measured as yes or no)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 18. Which inferential statistical analysis would be used if the IV was assault victims’ gender (measured as male or female) and the DV was the length of prison terms given to their attackers (measured in months)? 1. Chi-square 2. t test 3. ANOVA 4. Correlation 272 Key Terms Inferential analysis 203 Null hypothesis 210 Alternative hypothesis 210 Type I error 213 Type II error 213 Glossary of Symbols and Abbreviations Introduced in This Chapter 273 Chapter 10 Hypothesis Testing With Two Categorical Variables Chi-Square 274 Learning Objectives Identify the levels of measurement of variables used with a chi-square test of independence. Explain the difference between parametric and nonparametric statistics. Conduct a five-step hypothesis test for a contingency table of any size. Explain what statistical significance means and how it differs from practical significance. Identify the correct measure of association used with a particular chi-square test, and interpret those measures. Use SPSS to produce crosstabs tables, chi-square tests, and measures of association. Interpret SPSS chi-square output. The chi-square test of independence is used when the independent variable (IV) and dependent variable (DV) are both categorical (nominal or ordinal). The chi-square test is a member of the family of nonparametric statistics, which are used when sampling distributions cannot be assumed to be normally distributed, as is the case when a DV is categorical. Chi-square thus sits in contrast to parametric statistics, which are used when DVs are continuous (interval or ratio) and sampling distributions are safely assumed to be normal. The t test, analysis of variance, and correlation are all parametric. Because they have continuous DVs, they can rely on normal (or at least relatively normal) sampling distributions such as the t curve. (There are exceptions to the use of parametric statistics on continuous data, such as when the data are severely skewed, but that is beyond our scope here. For present purposes, we will distinguish the two classes of statistics on the basis of level of measurement.) Before going into the theory and math behind the chi-square statistic, Research Example 10.1 for an illustration of a type of situation in which a criminal justice or criminology researcher would turn to the chi-square test. Chi-square test of independence: The hypothesis-testing procedure appropriate when both the independent and dependent variables are categorical. Nonparametric statistics: The class of statistical tests used when dependent variables are categorical and the sampling distribution cannot be assumed to approximate normality. Parametric statistics: The class of statistical tests used when variables are continuous and normally distributed and the sampling distribution can be assumed to approximate normality. A researcher might be interested in finding out whether male and female offenders receive different sentences. In a study like this, gender would be nominal (male or female) and sentence might be measured nominally as well (such as jail, probation , or fine) . The question the researcher would be asking is, “Are these two variables related? In other words, does knowing an offender’s gender help me predict what type of sentence he or she receives?” Answering this question requires the use of the chi-square test of independence because both the IV and the DV are categorical. 275 Conceptual Basis of the Chi-Square Test: Statistical Independence and Dependence Two variables that are not related to one another are said to possess statistical independence. When two variables are related, they have statistical dependence. Statistical independence means that knowing which category an object falls into on the IV does not help predict its placement on the DV. Conversely, if two variables are statistically dependent, the independent does have predictive power over the outcome variable. Statistical independence: The condition in which two variables are not related to one another; that is, knowing what class persons or objects fall into on the independent variable does not help predict which class they will fall into on the dependent variable. Statistical dependence: The condition in which two variables are related to one another; that is, knowing what class persons or objects fall into on the independent variable helps predict which class they will fall into on the dependent variable. Research Example 10.1 How Do Criminologists’ and Criminal Justice Researchers’ Attitudes About the Criminal Justice System Compare to the Public’s Attitudes? Many criminal justice and criminology researchers study public opinion about crime and the justice system. What is less commonly known, however, is how these academics themselves perceive crime and society’s response to it. Griffin, Pason, Wiecko, and Brace (2016) compared the opinions of a sample of researchers to the opinions expressed by a sample of nonacademics from the general population. The table displays some of their findings. (The percentages do not sum to 100% because the people responding “I don’t know” to each question have been omitted from the table.) Substantial differences emerge between the two groups, with strong majorities of researchers opposing capital punishment and supporting marijuana legalization. The public was more likely to express favorable attitudes toward the death penalty and was more split on the issue of whether marijuana should be legalized. Source: Adapted from Table 1 in Griffin et al. (2016). The researchers ran chi-square tests to determine whether there were differences between researchers and the public. All the tests showed that academics and nonacademics have noticeably divergent opinions about these issues. In Research Example 10.1, the IV was whether survey respondents were either criminology or criminal justice researchers, or members of the public. There were four DVs (each coded yes/no) measuring respondents’ attitudes about the death penalty and about marijuana legalization. If these two variables are statistically independent, the two groups (researchers and the public) will be largely similar in their attitudes. In other words, knowing whether someone is or is not a researcher will not help us predict that person’s opinion on any 276 of these four dimensions. If they are statistically dependent, then we would gain predictive power by knowing which of the two groups a given individual is part of. When we want to know whether there is a relationship between two categorical variables, we turn to the chi-square test of independence. 277 The Chi-Square Test of Independence Let us work slowly through an example of a chi-square hypothesis test using the five steps described in Chapter 9 and discuss each step in detail along the way. For this example, we will turn to the 2014 General Social Survey (GSS; see Data Sources 2.2) and the issue of gender differences in attitudes about crime and punishment. Theory is somewhat conflicting as to whether women tend to be more forgiving of transgressions and to prefer leniency in punishment or whether they generally prefer harsher penalties out of the belief that offenders pose a threat to community safety. The GSS contains data on the sex of respondents and these persons’ attitudes toward the death penalty. The joint frequency distribution is displayed in Table 10.1. We will test for a relationship between gender (the IV) and death-penalty attitudes (the DV). Both of these variables are nominal, so the chi-square test of independence is the correct analysis. Note that we are setting gender as the IV and death-penalty support as the DV. There is no mathematical requirement pertaining to the placement of the variables in the rows versus the columns, but it is customary to place the IV in the rows of the table and the DV in the columns. You have seen tables like 10.1 before—it is a contingency (or crosstabs) table just like the ones we worked with in Chapter 3! Each cell of the table displays the number of people who fall into particular classes on each variable. (Ignore the superscripts for now; we will come back to these later.) For example, 511 GSS respondents are female and oppose the death penalty, and 752 are male and favor it. We can perform a cursory assessment of the possible relationship between gender and death-penalty support by calculating row percentages for each cell of the table (we use row percentages because the IV is in the rows). Approximately 60% of women favor the death penalty , and 40% oppose it . Among men, there is 69% favorability and 31% opposition. Judging by the 8 percentage-point difference between men and women, it appears that there is a relationship between gender and attitudes about capital punishment, with women expressing less support for it. Recall from the previous chapter, however, that it would be erroneous to conclude on the basis of these percentages alone that there is a true difference between men and women in terms of their death-penalty attitudes—we have not yet ruled out the possibility that this discrepancy is the product of chance variation (i.e., sampling error). A formal hypothesis test is required before we reach a conclusion about whether these variables are related. 278 Figure 10.1 The Chi-Square Probability Distribution, ⍺, χ ²crit, and χ ²obt Step 1. State the null (H0) and alternative (H1) hypotheses . The null hypothesis (H0) in chi-square tests is that there is no relationship between the IV and DV. The chi- square test statistic is χ2 (χ is the Greek letter chi and is pronounced “kye”). A χ2 value of zero means that the variables are unrelated, so the null is formally written as H0: χ2 = 0 The alternative hypothesis (H1), on the other hand, predicts that there is a relationship. The chi-square statistic gets larger as the overlap or relationship between the IV and DV increases. The chi-square statistic has its own sampling distribution, and the distribution contains only positive values; it is bounded at zero and has no negative side. This is because the statistic is a squared measure and, therefore, cannot take on negative values. As such, the alternative hypothesis is always expressed as H1: χ2 > 0

Step 2. Identify the distribution and compute the degrees of freedom .

As mentioned, the χ ² statistic has its own theoretical probability distribution—it is called the χ 2 distribution.

The χ ² table of critical values is located in Appendix D. Like the t curve, the χ ² distribution is a family of

differently shaped curves, and each curve’s shape is determined by degrees of freedom (df) . At small df values,

279

the distribution is extremely nonnormal; as the df increases, the distribution gradually normalizes somewhat,

but remains markedly different from a normal curve. Unlike the t curve, df for χ ² are based not on sample size

but, rather, on the size of the crosstabs table (i.e., the number of rows and columns). Looking at Table 10.1,

you can see that there are two rows (female and male) and two columns (favor and oppose) . The marginals (row

and column totals) are not included in the df calculation. The formula for degrees of freedom in a χ ²

distribution is

where

r = the number of rows, excluding the marginal and

c = the number of columns, excluding the marginal.

χ ² distribution: The sampling or probability distribution for chi-square tests. This curve is nonnormal and contains only positive

values. Its shape depends on the size of the crosstabs table.

Table 10.1 has two rows and two columns. Inserting these into the formula, the result is

df = (2 – 1)(2 – 1) = (1)(1) = 1

Step 3. Identify the critical value of the test statistic and state the decision rule .

Remember in Chapter 8 when we used the ⍺ (alpha) level to find a particular value of z or t to plug into a

confidence interval formula? We talked about ⍺ being the proportion of cases in the distribution that are out

in the tail beyond a particular value of z or t. You learned that the critical value is the number that cuts ⍺ off

the tail of the distribution. Alpha is the probability that a certain value will fall in the tail beyond the critical

value. If ⍺ = .05, for instance, then the values of the test statistic that are out in the tail beyond the critical

value constitute just 5% of the entire distribution. In other words, these values have a .05 or less probability of

occurring if, indeed, there is no relationship between the two variables being analyzed. These values, then,

represent observed outcomes that are extremely unlikely if the null hypothesis is true.

The process of finding the critical value of χ ² (symbolized χ ²crit) employs the same logic as that for finding

critical values of z or t. The value of χ ²crit depends on two considerations: the ⍺ level and the df. Alpha must

be set a priori so that the critical value can be determined before the test is run. Alpha can technically be set at

any number, but .05 and .01 are the most commonly used ⍺ levels in criminal justice and criminology.

For the present example, we will choose ⍺ = .05. Using Appendix D and finding the number at the

intersection of ⍺ = .05 and df = 1, it can be seen that χ ²crit = 3.841. This is the value that cuts .05 (i.e., 5%) of

the cases off the tail of the χ ² distribution. The obtained value of χ ² (symbolized χ ²obt) that is calculated in

Step 4 must exceed the critical value in order for the null to be rejected. Figure 10.1 illustrates this concept.

Obtained value: The value of the test statistic arrived at using the mathematical formulas specific to a particular test. The obtained

280

value is the final product of Step 4 of a hypothesis test.

The decision rule is the a priori statement regarding the action you will take with respect to the null

hypothesis based on the results of the statistical analysis that you are going to do in Step 4. The final product

of Step 4 will be the obtained value of the test statistic. The null hypothesis will be rejected if the obtained

value exceeds the critical value. If χ ²obt > χ ²crit, then the probability of obtaining this particular χ ²obt value by

chance alone is less than .05. Another way to think about it is that the probability of H0 being true is less than

.05. This is unlikely indeed! This would lead us to reject the null in favor of the alternative. The decision rule

for the current test is the following: If χ ²obt > 3.841 , H0 will be rejected.

Step 4. Compute the obtained value of the test statistic .

Now that we know the critical value, it is time to complete the analytical portion of the hypothesis test. Step 4

will culminate in the production of the obtained value, or χ ²obt. In substantive terms, χ ²obt is a measure of the

difference between observed frequencies (fo) and expected frequencies (fe). Observed frequencies are the

empirical values that appear in the crosstabs table produced from the sample-derived data set. Expected

frequencies are the frequencies that would appear if the two variables under examination were unrelated to one

another. In other words, the expected frequencies are what you would see if the null hypothesis were true. The

question is whether observed equals expected (indicating that the null is true and the variables are unrelated)

or whether there is marked discrepancy between them (indicating that the null should be rejected because

there is a relationship).

Observed frequencies: The empirical results seen in a contingency table derived from sample data. Symbolized fo.

Expected frequencies: The theoretical results that would be seen if the null were true, that is, if the two variables were, in fact,

unrelated. Symbolized fe.

Let’s talk about observed and expected frequencies a little more before moving on. Table 10.2 is a crosstabs

table for two hypothetical variables that are totally unrelated to one another. The 100 cases are spread evenly

across the four cells of the table. The result is that knowing which class a given case falls into on the IV offers

no information about which class that case is in on the DV. For instance, if you were faced with the question,

“Who is more likely to fall into category Y on the DV, someone in category A or in category B?” your answer

would be that both options are equally likely. The distribution in Table 10.2 illustrates the null hypothesis in a

chi-square test—the null predicts that the IV does not help us understand the DV.

Table 10.3 shows a distribution of hypothetical observed frequencies. There is a clear difference between this

distribution and that in Table 10.2. In Table 10.2, it is clear that knowing what category a person is in on the

IV does help predict their membership in a particular category on the DV. Someone in category A is more

likely to be in category Y than in category X , whereas someone in category B is more likely to be in X than in

Y. If you had a distribution like that in Table 10.2 and someone asked you to predict whether someone in

category A was in X or Y , you would have a 50/50 shot at being right; in other words, you would simply have

to guess. You would be wrong half the time (25 out of 50 guesses). On the other hand, if you were looking at

Table 10.3 and someone asked you the same question, you would predict the person to be in Y. You would

281

still be wrong occasionally, but the frequency of incorrect guesses would diminish from 50% to 20% (10 out of

50 guesses).

The chi-square analysis is, therefore, premised on a comparison of the frequencies that are observed in the

data and the frequencies that would be expected, theoretically, if there were no relationship between the two

variables. If there is minimal difference between observed and expected, then the null will be retained. If the

difference is large, the null must be rejected. We already know the observed frequencies, so the first task in

Step 4 is to calculate the expected frequencies. This must be done for each cell of the crosstabs table.

The formula for an expected frequency count is

where

= the expected frequency for cell i ,

rmi = the row marginal of cell i ,

cmi = the column marginal of cell i , and

N = the total sample size.

Since the expected frequency calculations must be done for each cell, it is a good idea to label them as a way to

keep track. This is the reason why the numbers in Table 10.1 are accompanied by superscripts. The letters A

through D identify the cells. Using Formula 10(2) for each cell,

282

Once the expected frequencies have been calculated, χ2

obt can be computed using the formula

where

= the observed frequency of cell i and

= the expected frequency of cell i.

Formula 10(3) looks intimidating, but it is actually just a sequence of arithmetic. First, each cell’s expected

value will be subtracted from its observed frequency. Second, each of these new terms will be squared and

divided by the expected frequency. Finally, these terms will be summed. Recall that the uppercase sigma (Σ) is

a symbol directing you to sum whatever is to the right of it.

The easiest way to complete the steps for Formula 10(3) is by using a table. We will rearrange the values from

Table 10.1 into a format allowing for calculation of χ2

obt. Table 10.4 shows this.

The obtained value of the test statistic is found by summing the final column of the table, as such:

χ2

obt = 3.14 + 5.65 + 3.71 + 6.68 = 19.18

There it is! The obtained value of the test statistic is 19.18.

Step 5. Make a decision about the null and state the substantive conclusion .

It is time to decide whether to retain or reject the null. To do this, revisit the decision rule laid out in Step 3.

It was stated that the null would be rejected if the obtained value of the test statistic exceeded 3.841. The

obtained value turned out to be 19.18, so χ2

obt > χ2

crit and we therefore reject the null. The alternative

hypothesis is what we take as being the true state of affairs. The technical term for this is statistical

significance. A statistically significant result is one in which the obtained value exceeds the critical value and

the variables are determined to be statistically related to one another.

Statistical significance: When the obtained value of a test statistic exceeds the critical value and the null is rejected.

283

The final stage of hypothesis testing is to interpret the results. People who conduct statistical analyses are

responsible for communicating their findings in a manner that effectively resonates with their audience,

whether it is an audience comprising scholars, practitioners, the public, or the media. It is especially important

when discussing statistical findings with lay audiences that clear explanations be provided about what a set of

quantitative results actually means in a substantive, practical sense. This makes findings accessible to a wide

array of audiences who might find criminological results interesting and useful.

In the context of the present example, rejecting the null leads to the conclusion that the IV and the DV are

statistically related; that is, there is a statistically significant relationship between gender and death-penalty

attitudes. Another way of saying this is that there is a statistically significant difference between men and

women in their attitudes toward capital punishment. Note that the chi-square test does not tell us about the

precise nature of that difference. Nothing in χ2

obt conveys information about which gender is more supportive

or more opposed than the other. This is not a big problem with two-class IVs. Referring back to the

percentages reported earlier, we know that a higher percentage of women than men oppose capital

punishment, so we can conclude that women are significantly less supportive of the death penalty (40%

oppose) compared to men (31% oppose). We will see later, when we use IVs that have more than two classes,

that we are not able to so easily identify the location of the difference.

Note, as well, the language used in the conclusion—it is phrased as an association and there is no cause-and-

effect assertion being advanced. This is because the relationship that seems to be present in this bivariate

analysis could actually be the result of unmeasured omitted variables that are the real driving force behind the

gender differences (recall from Chapter 2 that this is the problem of spuriousness and its counterpart, the

omitted variable bias). We have not, for instance, measured age, race, political beliefs, or religiosity, all of

which might relate to people’s beliefs about the effectiveness and morality of capital punishment. If women

differ from men systematically on any of these characteristics, then the gender–attitude relationship might be

spurious, meaning it is the product of another variable that has not been accounted for in the analysis. It is

best to keep your language toned down and to use words like relationship and association rather than cause or

effect.

Research Example 10.2 Do Victim or Offender Race Influence the Probability That a Homicide Will Be Cleared and That a Case

284

Will Be Tried as Death-Eligible?

A substantial amount of research has been conducted examining the impact of race on the use of the death penalty. This research

shows that among murder defendants, blacks have a higher likelihood of being charged as death-eligible (i.e., the prosecutor files

notice that he or she intends to seek the death penalty). The real impact of race, however, is not on the defendants’ part but, rather,

on the victims’: People who kill whites are more likely than people who kill blacks to be prosecuted as death-eligible. Blacks accused

of killing whites are the group most likely to face a death sentence, even controlling for relevant legal factors. There are open

questions, though, about what happens prior to prosecutors’ decisions about whether or not to seek the death penalty. In particular,

it is not clear what effect police investigations and clearance rates have in shaping the composition of cases that reach prosecutors’

desks. Petersen (2017) used data from Los Angeles County, California, to examine two stages in the justice-system response to

homicide: clearance and the decision to seek death. The table shows the racial breakdown of victims and defendants across these

categories.

Source: Adapted from Table 1 in Petersen (2017).

The contingency table shows no overt discrepancies for black victims (i.e., their representation in all three categories remains at

roughly one-third), but murders involving Latino victims (which make up 50% of all murders) are slightly less likely to be cleared

(48%) and much less likely to be prosecuted as death-eligible (38%). White victims, by contrast, make up only 15% of victims but

30% of victims in death-eligible trials. Looking at defendant race, blacks constitute 41% of the people arrested for homicide and 48%

of death-eligible defendants, whites likewise are somewhat overrepresented as defendants (13% versus 19%), while Latinos are

markedly underrepresented among defendants (46% compared to 33%). Of course, these relationships are bivariate and do not

account for legal factors (i.e., aggravating and mitigating circumstances) that might increase or reduce a prosecutor’s inclination to

seek the death penalty.

To thoroughly examine the relationship between race and the probability of death-eligible charges being filed, Petersen (2017)

estimated a series of predicted probabilities, which are displayed in the figure. These probabilities show the interaction between

victim and defendant race and are adjusted to control for case characteristics. The findings mirror previous research showing that

black and Latino defendants are more likely to face death-eligible charges when victims are white. White defendants are least likely

to face death when victims are Latino and most likely when they are black, but these differences were not statistically significant.

285

Source: Adapted from Figure 1 in Petersen (2017).

For the second example, let’s use the GSS again and this time test for a relationship between education level

and death-penalty attitudes. To make it interesting, we will split the data by gender and analyze males and

females in two separate tests. We will start with males (see Table 10.5). Using an alpha level of .01, we will

test for a relationship between education level (the IV) and death-penalty attitudes (the DV). All five steps

will be used.

Step 1. State the null (H0) and alternative (H1) hypotheses .

H0 : χ2 = 0

H1 : χ2 > 0

Step 2. Identify the distribution and compute the degrees of freedom .

The distribution is χ2 and the df = (r – 1)(c – 1) = (3 – 1)(2 – 1) = (2)(1) = 2.

Step 3. Identify the critical value of the test statistic and state the decision rule .

With ⍺ =.01 and df = 2, χ2

crit = 9.210. The decision rule is that if χ2 obt > 9.210 , H0 will be rejected.

286

Step 4. Compute the obtained value of the test statistic .

First, we need to calculate the expected frequencies using Formula 10(3). The frequencies for the first three

cells (labeled A, B, and C, left to right) are as follows:

Next, the computational table is used to calculate χ2

obt (Table 10.6). As you can see in the summation cell in

the last column, the obtained value of the test statistic is 19.74.

Before moving to Step 5, take note of a couple points about the chi-square calculation table. Both of these

features will help you check your math as you work through the computation. First, the expected-frequency

column always sums to the sample size. This is because we have not altered the number of cases in the sample:

We have merely redistributed them throughout the table. After calculating the expected frequencies, sum

them to make sure they add up to N. Second, the column created by subtracting the expected frequencies

from the observed frequencies will always sum to zero (or within rounding error of it). The reason for this is,

again, that no cases have been added to or removed from the sample. There are some cells that have observed

frequencies that are less than expected and others where is greater than . In the end, these variations

cancel each other out. Always sum both of these columns as you progress through a chi-square calculation.

This will help you make sure your math is correct.

287

Step 5. Make a decision about the null and state the substantive conclusion .

The decision rule stated that the null would be rejected if the obtained value exceeded 9.210. Since χ2

obt ended

up being greater than the critical value (i.e., 19.74 > 9.210), the null is rejected. There is a statistically

significant relationship between education and death-penalty attitudes among male respondents. Calculating

row percentages from the data in Table 10.5 shows that approximately 27% of men with high school diplomas

or less oppose the death penalty, roughly 22% with some college (no degree) are in opposition, and 40% with

a bachelor’s degree or higher do not support it. It seems that men with college educations that include at least

a bachelor’s degree stand out from the other two educational groups in their level of opposition to capital

punishment. We are not able to say with certainty, however, whether all three groups are statistically

significantly different from the others or whether only one of them stands apart. You can roughly estimate

differences using row percentages, but you have to be cautious in your interpretation. The chi-square test tells

you only that at least one group is statistically significantly different from at least one other group.

Let’s repeat the same analysis for female respondents. Again, we will set alpha at .01 and proceed through the

five steps. The data are in Table 10.7.

Step 1. State the null (H0) and alternative (H1) hypotheses .

H0: χ2 = 0

H1: χ2 > 0

Step 2. Identify the distribution and compute the degrees of freedom .

The distribution is χ2 and df = (3 – 1)(2 – 1) = 2.

Step 3. Identify the critical value of the test statistic and state the decision rule .

With ⍺ =.01 and df = 2, χ2

crit = 9.210. The decision rule is that if χ2

obt > 9.210 , H0 will be rejected.

288

289

Learning Check 10.1

In the third example, the calculations are not shown. Check your mastery of the computation of expected frequencies by doing the

calculations yourself and making sure you arrive at the same answers shown in Table 10.8.

Step 4. Compute the obtained value of the test statistic .

Step 5. Make a decision about the null and state the substantive conclusion .

The decision rule stated that the null would be rejected if the obtained value exceeded 9.210. Since χ2

obt =

11.23, the null is rejected. There is a statistically significant relationship between education and death-penalty

attitudes among female respondents; it appears that women’s likelihood of favoring or opposing capital

punishment changes with their education level. Another way to phrase this is that there are significant

differences between women of varying levels of education. As we did with male respondents, we can use row

percentages to gain a sense of the pattern. Opposition to the death penalty is approximately 40%, 26%, and

44% among women with high school diploma or less, some college, or a bachelor’s degree or higher,

respectively. This is similar to the pattern of opposition seen among men in that those with some college were

the most supportive of capital punishment and those with a college degree were the least supportive, but

among women the group with a high school diploma or less were nearly as likely as those with college degrees

to oppose this penalty.

290

Learning Check 10.2

One criticism of the chi-square test for independence is that this statistic is sensitive to sample size. The problem lies in the way that

χ2obt is calculated. Sample size can cause a test statistic to be significant or not significant, apart from the actual distribution of observed

values. For instance, a crosstabs table with a fairly even distribution of scores might yield a statistically significant χ2obt if N is large.

Similarly, a distribution that looks decidedly uneven (i.e., where there is an apparent relationship between the IV and the DV) can

produce nonsignificant χ2obt if N is small. To see this for yourself, recalculate χ2obt using the death-penalty and gender data but with a

sample size of 88 instead of 2,379. Make a decision about the null hypothesis, recalling that χ ²crit = 3.841.

Are you surprised by the results? This demonstrates the importance of being cautious when you interpret statistical significance. Do not

leap to hasty conclusions; be aware that there are factors (such as sample size) that can impact the results of a statistical test irrespective of

the relationship between the IVs and the DVs.

291

Measures of Association

The chi-square test alerts you when there is a statistically significant relationship between two variables, but it

is silent as to the strength or magnitude of that relationship. We know from the previous two examples, for

instance, that gender and education are related to attitudes toward capital punishment, but we do not know

the magnitudes of these associations: They could be strong, moderate, or weak. This question is an important

one because a trivial relationship—even if statistically significant in a technical sense—is not of much

substantive or practical importance. Robust relationships are more meaningful. To illustrate this, suppose an

evaluation of a gang-prevention program for youth was declared a success after researchers found a statistically

significant difference in gang membership rates among youth who did and did not participate in the program.

Digging deeper, however, you learn that 9% of the youth who went through the program ended up joining

gangs, compared to 12% of those who did not participate. While any program that keeps kids out of gangs is

laudable, a reduction of three percentage points can hardly be considered a resounding success. We would

probably want to continue searching for a more effective way to prevent gang involvement. Measures of

association offer insight into the magnitude of the differences between groups so that we can figure out how

strong the overlap is.

Measures of association: Procedures for determining the strength or magnitude of a relationship after a chi-square test has revealed a

statistically significant association between two variables.

There are several measures of association, and this chapter covers four of them. The level of measurement of

the IV and the DV dictate which measures are appropriate for a given analysis. Measures of association are

computed only when the null hypothesis has been rejected—if the null is not rejected and you conclude that

there is no relationship between the IV and the DV, then it makes no sense to go on and try to interpret an

association you just said does not exist. The following discussion will introduce four tests, and the next section

will show you how to use SPSS to compute χ2

obt and accompanying measures of association.

Cramer’s V can be used when both of the variables are nominal or when one is ordinal and the other is

nominal. It is symmetric, meaning that it always takes on the same value regardless of which variable is

posited as the independent and which the dependent. This statistic ranges from 0.00 to 1.00, with higher

values indicative of stronger relationships and values closer to 0.00 suggestive of weaker associations. Cramer’s

V is computed as

where

χ2

obt = the obtained value of the test statistic,

N = the total sample size, and

m = the smaller of either (r – 1) or (c – 1).

292

Cramer’s V : A symmetric measure of association for χ2 when the variables are nominal or one is ordinal and the other is nominal. V

ranges from 0.00 to 1.00 and indicates the strength of the relationship. Higher values represent stronger relationships. Identical to

phi in 2 × 2 tables.

In the first example we saw in this chapter, where we found a statistically significant relationship between

gender and death-penalty attitudes, χ2

obt = 19.18, N = 2,379, and there were two rows and two columns, so m

= 2 – 1 = 1. Cramer’s V is thus

This value of V suggests a weak relationship. This demonstrates how statistical significance alone is not

indicative of genuine importance or meaning—a relationship might be significant in a technical sense but still

very in significant in practical terms. This is due in no small part to the chi-square test’s sensitivity to sample

size, as discussed earlier. Think back to the percentages we calculated for this table. Approximately 60% of

women and 69% of men favored the death penalty. This is a difference, to be sure, but it is not striking. If any

given respondent were randomly selected out of this sample, there would be a roughly two-thirds likelihood

that the person would support capital punishment, irrespective of her or his gender. It is wise, then, to be

cautious in interpreting statistically significant results—statistical significance does not always translate into

practical significance.

When both variables under examination are nominal, lambda is an option. Like Cramer’s V , lambda ranges

from 0.00 to 1.00. Unlike Cramer’s V , lambda is asymmetric, meaning that it requires that one of the

variables be clearly identified as the independent and the other as the dependent. This is because lambda is a

proportionate reduction in error measure.

Lambda: An asymmetric measure of association for χ2 when the variables are nominal. Lambda ranges from 0.00 to 1.00 and is a

proportionate reduction in error measure.

Proportionate reduction in error (PRE) refers to the extent to which knowing a person’s or object’s placement

on an IV helps predict that person’s or object’s classification on the dependent measure. Referring back to

Table 10.1, if you were trying to predict a given individual’s attitude toward capital punishment and the only

piece of information you had was the frequency distribution of this DV (i.e., you knew that 1,530 people in

the sample support capital punishment and 849 oppose it), then your best bet would be to guess the modal

category (mode = support) because this would produce the fewest prediction errors. There would, though, be a

substantial number of these errors—849, to be exact!

Now, suppose that you know a given person’s gender or education level, both of which we found to be

significantly related to capital punishment attitudes. The next logical inquiry is the extent to which this

knowledge improves our accuracy when predicting whether that person opposes or favors the death penalty. In

other words, know that we would make 849 errors if we simply guessed the mode for each person in the

sample, and now we want to know how many fewer mistakes we would make if we knew each person’s gender

293

or education level. This is the idea behind PRE measures like lambda. Let’s do an example using the

relationship between education and death penalty attitudes among women.

Lambda is symbolized as λ (the Greek lowercase letter lambda) and is calculated as

where

E1 = Ntotal – NDV mode and

This equation and its different components looks strange, but, basically, E1 represents the number of

prediction errors made when the IV is ignored (i.e., predictions based entirely on the mode of the DV), and

E2 reflects the number of errors made when the IV is taken into account. Using the education and death-

penalty data from Table 10.7, we can first calculate E1 and E2:

E1 = 1,289 – 778 = 511,

E2 = (795 – 479) + (115 – 85) + (379 – 214) = 511

and lambda is

Lambda is most easily interpreted by transforming it to a percentage. A lambda of zero shows us that knowing

women’s education levels does not reduce prediction errors. This makes sense, as you can see in Table 10.7

that “favor” is the mode across all three IV categories. This lack of variation accounts for the overall zero

impact of the IV on the DV. Again, we see a statistically significant but substantively trivial association

between variables. This makes sense, because we would not expect a single personal characteristic (like

education) to have an enormous influence on someone’s attitudes. As noted previously, we have not measured

women’s religiosity, age, social and political views, and other demographic and background factors that are

closely linked with people’s opinions about capital punishment.

There is a third measure for nominal data that bears brief mention, and that is phi. Phi can only be used on 2

× 2 tables (i.e., two rows and two columns) with nominal variables. It is calculated and interpreted just like

Cramer’s V with the exception that phi does not account for the number of rows or columns in the crosstabs

table: since it can only be applied to 2 × 2 tables, m will always be equal to 1.00. For 2 × 2 tables, Cramer’s V is

identical to phi, but since Cramer’s V can be used for tables of any size, it is more useful than phi is.

294

Phi: A symmetric measure of association for χ2 with nominal variables and a 2 × 2 table. Identical to Cramer’s V.

When both variables are ordinal or when one is ordinal and the other is dichotomous (i.e., has two classes),

Goodman and Kruskal’s gamma is an option. Gamma is a PRE measure but is symmetric—unlike lambda—

and ranges from –1.00 to +1.00, with zero meaning no relationship, –1.00 indicating a perfect negative

relationship (as one variable increases, the other decreases), and 1.00 representing a perfect positive

relationship (as one increases, so does the other). Generally speaking, gamma values between 0 and ±.19 are

considered weak, between ±.20 and ±.39 moderate, ±.40 to ±.59 strong, and ±.60 to ±1.00 very strong.

Goodman and Kruskal’s gamma: A symmetric measure of association used when both variables are ordinal or one is ordinal and the

other is dichotomous. Ranges from –1.00 to +1.00.

Two other measures available when both variables are ordinal are Kendall’s taub and Kendall’s tauc. Both are

symmetric. Taub is used when the crosstabs table has an equal number of rows and columns, and tauc is used

when they are unequal. Both tau statistics range from –1.00 to +1.00. They measure the extent to which the

order of the observations in the IV match the order in the DV; in other words, as cases increase in value on

the IV, what happens to their scores on the DV? If their scores on the dependent measure decrease, tau will

be negative; if they increase, tau will be positive; and if they do not display a clear pattern (i.e., the two

variables have very little dependency), tau will be close to zero. Similar to the tau measures is Somers’ d . This

measure of association is asymmetric and used when both variables are ordinal. Its range and interpretation

mirror those of tau . The calculations of gamma, tau, and d are complicated, so we will refrain from doing

them by hand and will instead use SPSS to generate these values.

Kendall’s taub: A symmetric measure of association for two ordinal variables when the number of rows and columns in the crosstabs

table are equal. Ranges from –1.00 to +1.00.

Kendall’s tauc: A symmetric measure of association for two ordinal variables when the number of rows and columns in the crosstabs

table are unequal. Ranges from –1.00 to +1.00.

Somers’ d : An asymmetric measure of association for two ordinal variables. Ranges from –1.00 to +1.00.

None of the measures of association discussed here is perfect; each has limitations and weaknesses. The best

strategy is to examine two or more measures for each analysis and use them to gain a comprehensive picture of

the strength of the association. There will likely be variation among them, but the differences should not be

wild, and all measures should lean in a particular direction. If they are all weak or are all strong, then you can

safely arrive at a conclusion about the level of dependency between the two variables.

295

SPSS

The SPSS program can be used to generate χ2

obt , determine statistical significance, and produce measures of

association. The chi-square analysis is found via the sequence Analyze → Descriptive Statistics → Crosstabs. Let

us first consider the gender and capital punishment example from earlier in the chapter. Figure 10.2 shows the

dialog boxes involved in running this analysis in SPSS. Note that you must check the box labeled Chi-square

in order to get a chi-square analysis; if you do not check this box, SPSS will merely give you a crosstabs table.

This box is opened by clicking Statistics in the crosstabs window. By default, SPSS provides only observed

frequencies in the crosstabs table. If you want expected frequencies or percentages (row or column), you can

go into Cells and request them. Since both of these variables are nominal, lambda and Cramer’s V are the

appropriate measures of association. Figure 10.3 shows the output for the chi-square test, and Figure 10.4

displays the measures of association.

The obtained value of the χ2

obt statistic is located on the line labeled Pearson Chi-Square. You can see in

Figure 10.3 that χ2

obt = 19.182, which is identical to the value we obtained by hand. The output also tells you

whether or not the null should be rejected, but it does so in a way that we have not seen before. The SPSS

program gives you what is called a p value. The p value tells you the exact probability of the obtained value of

the test statistic: The smaller p is, the more unlikely the χ2

obt is if the null is true, and, therefore, the

probability that the null is, indeed, correct. The p value in SPSS χ2 output is the number located at the

intersection of the Asymp. Sig. (2-sided) column and the Pearson Chi-Square row. Here, p = .000. What you do

is compare p to ⍺. If p is less than ⍺, it means that the obtained value of the test statistic exceeded the critical

value, and the null is rejected; if p is greater than ⍺, the null is retained. Since in this problem ⍺ was set at .05,

the null hypothesis is rejected because .000 < .05. There is a statistically significant relationship between
gender and death-penalty attitudes. Of course, as we have seen, rejection of the null hypothesis is only part of
the story because the χ2 statistic does not offer information about the magnitude or strength of the
relationship between the variables. For this, we turn to measures of association.
p value: In SPSS output, the probability associated with the obtained value of the test statistic. When p < ⍺, the null hypothesis is
rejected.
Figure 10.2 Running a Chi-Square Test and Measures of Association in SPSS
296
Figure 10.3 Chi-Square Output
Figure 10.4 Measures of Association
Judging by both Cramer’s V and lambda, this relationship is very weak. We actually already knew this because
we calculated V by hand and arrived at .10, which is within rounding error of the .09 SPSS produces. Lambda
is zero, which means that knowing people’s gender does not reduce the number of errors made in predicting
their death-penalty attitudes. As noted earlier, using multiple tests of association helps provide confidence in
the conclusion that the association between these two variables, while statistically significant, is tenuous in a
297
substantive sense. In other words, knowing someone’s gender only marginally improves our ability to predict
that person’s attitudes about capital punishment.
Chapter Summary
This chapter introduced the chi-square test of independence, which is the hypothesis-testing procedure appropriate when both of
the variables under examination are categorical. The key elements of the χ2 test are observed frequencies and expected frequencies.
Observed frequencies are the empirical results seen in the sample, and expected frequencies are those that would appear if the null
hypothesis were true and the two variables unrelated. The obtained value of chi-square is a measure of the difference between
observed and expected, and comparing χ2obt to χ2crit for a set ⍺ level allows for a determination of whether the null hypothesis
should be retained or rejected.
When the null is retained (i.e., when χ2obt < χ2crit ), the substantive conclusion is that the two variables are not related. When the
null is rejected (when χ2obt >χ2crit ), the conclusion is that there is a relationship between them. Statistical significance, though, is

only a necessary and not a sufficient condition for practical significance. The chi-square statistic does not offer information about the

strength of a relationship and how substantively meaningful this association is.

For this, measures of association are turned to when the null has been rejected. Cramer’s V , lambda, Goodman and Kruskal’s

gamma, Kendall’s taua and taub, and Somers’ d are appropriate in any given situation depending on the variables’ levels of

measurement and the size of the crosstabs table. SPSS can be used to obtain chi-square tests, p values for determining statistical

significance, and measures of association. When p < ⍺, the null is rejected, and when p > ⍺, it is retained. You should always generate

measures of association when you run χ2 tests yourself, and you should always expect them from other people who run these analyses

and present you with the results. Statistical significance is important, but the magnitude of the relationship tells you how meaningful

the association is in practical terms.

Thinking Critically

1. Suppose you read a report claiming that children raised in families with low socioeconomic status are less likely to go to

college compared to children raised in families with middle and upper income levels. The news story cites college

participation rates of 20%, 35%, and 60% among low, middle, and upper socioeconomic statuses, respectively, and explains

these differences as meaning that children raised in poor families are less intelligent or less ambitious than those from better-

off families. Do you trust this conclusion? Why or why not? If you do not, what more do you need to know about these data

before you can make a decision about the findings, and what they mean for the relationship between family income and

children’s college attendance?

2. Two researchers are arguing about statistical findings. One of them believes that any statistically significant result is

important, irrespective of the magnitude of the association between the IVs and the DVs. The other one contends that

statistical significance is meaningless if the association is weak. Who is correct? Explain your answer. Offer a hypothetical or

real-world example to illustrate your point.

Review Problems

1. A researcher wants to test for a relationship between the number of citizen complaints that a police officer receives and

whether that officer commits serious misconduct. He gathers a sample of officers and records the number of complaints that

have been lodged against them (0 – 2, 3 – 5, 6+) and whether they have ever been written up for misconduct (yes or no). Can

he use a chi-square to test for a relationship between these two variables? Why or why not?

2. A researcher wishes to test for a relationship between age and criminal offending. She gathers a sample and for each person,

she collects his or her age (in years) and whether that person has ever committed a crime (yes or no). Can she use a chi-

square to test for a relationship between these two variables? Why or why not?

3. A researcher is interested in finding out whether people who drive vehicles that are in bad condition are more likely than

those driving better cars to get pulled over by police. She collects a sample and codes each person’s vehicle’s condition (good,

298

fair, poor) and the number of times that person has been pulled over (measured by respondents writing in the correct

number). Can she use a chi-square to test for a relationship between these two variables? Why or why not?

4. A researcher is studying the effectiveness of an in-prison treatment program in reducing post-release recidivism. He gathers a

sample of recently released prisoners and records, for each person, whether he or she participated in a treatment program

while incarcerated (yes or no) and whether that person committed a new crime within 6 months of release (yes or no). Can

he use a chi-square to test for a relationship between these two variables? Why or why not?

5. Is a criminal defendant’s gender related to the type of sentence she or he receives? A researcher collects data on defendants’

gender (male or female) and sentence (jail, probation, fine).

1. Which of these variables is the IV, and which is the DV?

2. Identify each variable’s level of measurement.

3. How many rows and columns would the crosstabs table have?

6. Is the value of the goods stolen during a burglary related to the likelihood that the offender will be arrested? A researcher

collects data on the value of stolen goods ($299 or less, $300–$599, $600 and more) and on whether the police arrested

someone for the offense (yes or no).

1. Which of these variables is the IV, and which is the DV?

2. Identify each variable’s level of measurement.

3. How many rows and columns would the crosstabs table have?

7. Is the crime for which a person is convicted related to the length of the prison sentence she or he receives? A research gathers

data on crime type (violent, property, drug) and sentence length (18 months or less, 19–30 months, 31 or more months).

1. Which of these variables is the IV, and which is the DV?

2. Identify each variable’s level of measurement.

3. How many rows and columns would the crosstabs table have?

8. Is a victim’s gender related to whether or not the offender will be convicted for the crime? A researcher collects data on

victim gender (male or female) and whether the offender was convicted (yes or no).

2. Identify each variable’s level of measurement.

3. How many rows and columns would the crosstabs table have?

9. It might be expected that jails that offer alcohol treatment programs to inmates also offer psychiatric counseling services,

since alcohol abuse is frequently a symptom of an underlying psychological problem. The following table displays data from a

random sample from the Census of Jails (COJ). With an alpha level of .01, conduct a five-step chi-square hypothesis test to

determine whether the two variables are independent.

10. Is there an association between the circumstances surrounding a violent altercation that results in a shooting and the type of

firearm used? The Firearm Injury Surveillance Study (FISS) records whether the shooting arose out of a fight and the type of

firearm used to cause the injury (here, handguns vs. rifles and shotguns). With an alpha of .01, conduct a five-step hypothesis

test to determine if the variables are independent.

299

11. Continuing with an examination of gunshots resulting from fights, we can analyze FISS data to determine whether there is a

relationship between victims’ genders and whether their injuries were the result of fights. With an alpha of .05, conduct a

five-step hypothesis test to determine if the variables are independent.

12. In the chapter, we saw that there was a statistically significant difference between men and women in terms of their attitudes

about capital punishment. We can extend that line of inquiry and find out whether there is a gender difference in general

attitudes about crime and punishment. The GSS asks respondents whether they think courts are too harsh, about right, or

not harsh enough in dealing with criminal offenders. The following table contains the data. With an alpha level of .05,

conduct a five-step chi-square hypothesis test to determine whether the two variables are independent.

13. Do men and women differ on their attitudes toward drug laws? The GSS asks respondents to report whether they think

marijuana should be legalized. The following table shows the frequencies, by gender, among black respondents. With an

alpha level of .05, conduct a five-step chi-square hypothesis test to determine whether the two variables are independent.

14. The following table shows the support for marijuana legalization, by race, among male respondents. With an alpha level of

.05, conduct a five-step chi-square hypothesis test to determine whether the two variables are independent.

15. There is some concern that people of lower-income statuses are more likely to come in contact with the police as compared

to higher-income individuals. The following table contains Police–Public Contact Survey (PPCS) data on income and police

contacts among respondents who were 21 years of age or younger. With an alpha of .01, conduct a five-step hypothesis test

300

to determine if the variables are independent.

16. One criticism of racial profiling studies is that people’s driving frequency is often unaccounted for. This is a problem because,

all else being equal, people who spend more time on the road are more likely to get pulled over eventually. The following

table contains PPCS data narrowed down to black male respondents. The variables measure driving frequency and whether

these respondents had been stopped by police for traffic offenses within the past 12 months. With an alpha of .01, conduct a

five-step hypothesis test to determine if the variables are independent.

17. The companion website (www.sagepub.com/gau) contains the SPSS data file GSS for Chapter 10.sav. This is a portion of the

2014 GSS. Two of the variables in this file are race and courts , which capture respondents’ race and their attitudes about

courts’ harshness, respectively. Run a chi-square analysis to determine if people’s attitudes (the DV) vary by race (the IV).

Then do the following.

1. Identify the obtained value of the chi-square statistic.

2. Make a decision about whether you would reject the null hypothesis of independence at an alpha level of .05 and

explain how you arrived at that decision.

3. State the conclusion that you draw from the results of each of these analyses in terms of whether there is a

relationship between the two variables.

4. If you rejected the null hypothesis , interpret row percentages and applicable measures of association. How strong is the

relationship? Would you say that this is a substantively meaningful relationship?

18. Using GSS for Chapter 10.sav (www.sagepub.com/gau), run a chi-square analysis to determine whether there is a relationship

between the candidate people voted for in the 2012 presidential election (Barack Obama or Mitt Romney) and their opinions

about how well elected officials are doing at controlling crime rates than do the following:

1. Identify the obtained value of the chi-square statistic.

2. Make a decision about whether you would reject the null hypothesis of independence at an alpha level of .01 and

explain how you arrived at that decision.

3. State the conclusion that you draw from the results of each of these analyses in terms of whether there is a

relationship between the two variables.

4. If you rejected the null hypothesis , interpret row percentages and applicable measures of association. How strong is the

relationship? Would you say that this is a substantively meaningful relationship?

19. A consistent finding in research on police–community relations is that there are racial differences in attitudes toward police.

Although all racial groups express positive views of police overall, the level of support is highest for whites and tends to

dwindle among persons of color. The companion website (www.sagepub.com/gau) contains variables from the PPCS (PPCS

301

http://www.sagepub.com/gau

http://www.sagepub.com/gau

http://www.sagepub.com/gau

for Chapter 10.sav) . The sample has been narrowed to males who were stopped by the police while driving a car and were

issued a traffic ticket. There are three variables in this data set: race , income , and legitimacy. The legitimacy variable measures

whether respondents believed that the officer who pulled them over had a credible reason for doing so. Use SPSS to run a

chi-square analysis to determine whether legitimacy judgments (the DV) differ by race (the IV). Based on the variables’ level

of measurement, select appropriate measures of association. Then do the following:

1. Identify the obtained value of the chi-square statistic.

2. Make a decision about whether you would reject the null hypothesis of independence at an alpha level of .01 and

explain how you arrived at that decision.

3. State the conclusion that you draw from the results of each of these analyses as to whether or not there is a difference

between private and public prisons in terms of offering vocational training.

4. If you rejected the null hypothesis , interpret row percentages and applicable measures of association. How strong is the

relationship? Would you say that this is a substantively meaningful relationship?

20. Using the PPCS for Chapter 10.sav file again (www.sagepub.com/gau), run a chi-square test to determine whether

respondents’ perceptions of stop legitimacy (the DV) vary across income levels (the IV). Based on the variables’ level of

measurement, select appropriate measures of association. Then do the following:

1. Identify the obtained value of the chi-square statistic.

2. Make a decision about whether you would reject the null hypothesis of independence at an alpha level of .01 and

explain how you arrived at that decision.

3. State the conclusion that you draw from the results of each of these analyses in terms of whether there is a difference

between private and public prisons in terms of offering vocational training.

4. If you rejected the null hypothesis , interpret row percentages and applicable measures of association. How strong is the

relationship? Would you say that this is a substantively meaningful relationship?

302

http://www.sagepub.com/gau

Key Terms

Chi-square test of independence 219

Nonparametric statistics 219

Parametric statistics 220

Statistical independence 221

Statistical dependence 221

χ² distribution 222

Obtained value 223

Observed frequencies 224

Expected frequencies 224

Statistical significance 227

Measures of association 234

Cramer’s V 235

Lambda 235

Phi 237

Goodman and Kruskal’s gamma 237

Kendall’s taub 237

Kendall’s tauc 237

Somers’ d 237

p value 238

Glossary of Symbols and Abbreviations Introduced in This Chapter

303

Chapter 11 Hypothesis Testing With Two Population Means or

Proportions

304

Learning Objectives

Identify situations in which, based on the levels of measurement of the independent and dependent variables, t tests are

appropriate.

Explain the logic behind two-population tests for differences between means and proportions.

Explain what the null hypothesis predicts and construct an alternative or research hypothesis appropriate to a particular research

question.

For tests of means, identify the correct type of test (dependent or independent samples).

For tests of means with independent samples, identify the correct variance formula (pooled or separate).

Select the correct equations for a given test type, and use them to conduct five-step hypothesis tests.

In SPSS, identify the correct type of test, run that analysis, and interpret the output.

There are many situations in which criminal justice and criminology researchers work with categorical

independent variables (IVs) and continuous dependent variables (DVs). They might want to know, for

instance, whether male and female police officers differ in the number of arrests they make, whether criminal

offenders who have children are given shorter jail sentences compared to those who do not, or whether

prisoners who successfully complete a psychological rehabilitation program have a significant reduction in

antisocial thinking. Research Example 11.1 illustrates another instance of a categorical IV and a continuous

DV.

In Research Example 11.1, Wright, Pratt, and DeLisi (2008) had two groups (multiple homicide offenders

[MHOs] and single homicide offenders [SHOs]), each with its own mean and standard deviation; the goal

was to find out whether the groups’ means differ significantly from one another. A significant difference

would indicate that MHOs and SHOs do indeed differ in the variety of crimes they commit, whereas rough

equivalency in the means would imply that these two types of homicide offenders are equally diverse in

offending. What should the researchers do to find out whether MHOs and SHOs have significantly different

diversity indices?

The answer is that they should conduct a two-population test for differences between means, or what is

commonly referred to as a t test. As you probably figured out, these tests rely on the t distribution. We will

also cover two-population tests for differences between proportions, which are conceptually similar to t tests

but employ the z distribution.

t test: The test used with a two-class, categorical independent variable and a continuous dependent variable.

Tests for differences between two means or two proportions are appropriate when the IV is categorical with

two classes or groups and the DV is expressed as one mean or proportion per group. Examples of two-class,

categorical IVs include gender (male or female) and political orientation (liberal or conservative) . Examples of

DVs appropriate for two-population tests are the mean number of times people in a sample report that they

drove while intoxicated, or the proportion of people in the sample who have been arrested for driving while

under the influence of alcohol. The diversity index that Wright et al. (2008) used is continuous, which is why

the researchers computed a mean and a standard deviation, and the reason that a t test is the appropriate

305

analytic strategy. In the review problems at the end of the chapter, you will conduct a t test to find out

whether MHOs and SHOs differ significantly in offending diversity.

Research Example 11.1 Do Multiple Homicide Offenders Specialize in Killing?

Serial killers and mass murderers capture the public’s curiosity and imagination. Who can resist some voyeuristic gawking at a killer

who periodically snuffs out innocent victims while outwardly appearing to be a regular guy or at the tormented soul whose troubled

life ultimately explodes in an episode of wanton slaughter? Popular portrayals of multiple homicide offenders (MHOs) lend the

impression that these killers are fundamentally different from more ordinary criminals and from single homicide offenders (SHOs)

in that they only commit homicide and lead otherwise crime-free lives. But is this popular conception true?

Wright, Pratt, and DeLisi (2008) decided to find out. They constructed an index measuring diversity of offending within a sample

of homicide offenders. This index captured the extent to which homicide offenders committed only homicide and no other crimes

versus the extent to which they engaged in various types of illegal acts. The researchers divided the sample into MHOs and SHOs

and calculated each group’s mean and standard deviation on the diversity index. They found the statistics located in the table.

Source: Adapted from Table 1 in Wright et al. (2008).

306

Two-Population Tests for Differences Between Means: t Tests

There are many situations in which people working in criminal justice and criminology would want to test for

differences between two means: Someone might be interested in finding out whether offenders who are

sentenced to prison receive significantly different mean sentence lengths depending on whether they are male

or female. A municipal police department might implement an innovative new policing strategy and want to

know whether the program significantly reduced mean crime rates in the city. These types of studies require t

tests.

In Chapter 7, you learned the difference between sample, sampling, and population distributions. Recall that

sampling distributions are theoretical curves created when multiple or infinite samples are drawn from a single

population and a statistic is computed and plotted for each sample. Over time, with repeated drawing,

calculating, plotting, throwing back, and drawing again, the distribution of sample statistics builds up and, if

the size of each sample is large (meaning N ≥ 100), the statistics form a normal curve. There are also sampling

distributions for differences between means. See Figure 11.1. You have seen in prior chapters how sampling

distributions’ midpoint is the population mean (symbolized μ) . Sampling distributions for differences

between means are similar in that they center on the true population difference, μ1 − μ2. A sampling

distribution of differences between means is created by pulling infinite pairs of samples, rather than single

samples. Imagine drawing two samples, computing both means, subtracting one mean from the other to form

a difference score, and then plotting that difference score. Over time, the difference scores build up. If N ≥

100, the sampling distribution of differences between means is normal; if N ≤ 99, then the distribution is more

like “normalish” because it tends to be wide and flat. The t distribution, being flexible and able to

accommodate various sample sizes, is the probability distribution of choice for tests of differences between

means.

There are two general types of t tests, one for independent samples and one for dependent samples. The

difference between them pertains to the method used to select the two samples under examination. In

independent sampling designs, the selection of cases into one sample in no way affects, or is affected by, the

selection of cases into the other sample. If a researcher is interested in the length of prison sentences received

by male and female defendants, then that researcher would draw a sample of females and a sample of males. A

researcher investigating the effects of judicial selection type on judges’ sentencing decisions might draw a

sample of judges who were elected and a sample of those who were appointed to their posts. In neither of

these instances does the selection of one person into one sample have bearing on the selection of another into

the other sample. They are independent because they have no influence on each other.

Independent samples: Pairs of samples in which the selection of people or objects into one sample in no way affected, or was affected

by, the selection of people or objects into the other sample.

Dependent samples: Pairs of samples in which the selection of people or objects into one sample directly affected, or was directly

affected by, the selection of people or objects into the other sample. The most common types are matched pairs and repeated

measures.

307

In dependent-samples designs, by contrast, the two samples are related to each other in some way. The two

major types of dependent-samples designs are matched pairs and repeated measures. Matched-pairs designs

are used when researchers need an experimental group and a control group but are unable to use random

assignment to create the groups. They therefore gather a sample from a treatment group and then construct a

control group via the deliberate selection of cases that did not receive the treatment but that are similar to the

treatment group cases on key characteristics. If the unit of analysis is people, participants in the control group

might be matched to the treatment group on race, gender, age, and criminal history.

Matched-pairs design: A research strategy where a second sample is created on the basis of each case’s similarity to a case in an

existing sample.

Figure 11.1 The Sampling Distribution of Differences Between Means

Repeated-measures designs are commonly used to evaluate program impacts. These are before-and-after

designs wherein the treatment group is measured prior to the intervention of interest and then again afterward

to determine whether the post-intervention scores differ significantly from the pre-intervention scores. In

repeated measures, then, the “two” samples are actually the same people or objects measured twice.

Repeated-measures design: A research strategy used to measure the effectiveness of an intervention by comparing two sets of scores

(pre and post) from the same sample.

The first step in deciding what kind of t test to use is to figure out whether the samples are independent or

dependent. It is sometimes easier to identify dependent designs than independent ones. The biggest clue to

look for is a description of the research methods. If the samples were collected by matching individual cases

on the basis of similarities between them, or if an intervention was being evaluated by collecting data before

and after a particular event, then the samples are dependent and the dependent-samples t test is appropriate.

If the methods do not detail a process of matching or of repeated measurement, if all that is said is that two

samples were collected or a single sample was divided on the basis of a certain characteristic to form two

subsamples, then you are probably dealing with independent samples and should use the independent-samples

t test.

There is one more wrinkle. There are two types of independent-samples t tests: pooled variances and separate

variances. The former is used when the two population variances are similar to one another, whereas the latter

is used when the variances are significantly disparate. The rationale for having these two options is that when

308

two samples’ variances are similar, they can safely be combined (pooled) into a single estimate of the

population variance. When they are markedly unequal, however, they must be mathematically manipulated

before being combined. You will not be able to tell merely by looking at two samples’ variances whether you

should use a pooled-variance or separate-variance approach, but that is fine. In this book, you will always be

told which one to use. When we get to SPSS, you will see that this program produces results from both of

these tests, along with a criterion to use for deciding between them. (More on this later.) You will, therefore,

always be able to figure out which type of test to use.

Pooled variances: The type of t test appropriate when the samples are independent and the population variances are equal.

Separate variances: The type of t test appropriate when the samples are independent and the population variances are unequal.

309

Learning Check 11.1

Be very careful about order of operations! The formulas we will encounter in this chapter require multiple steps, and you have to do those

steps in proper sequence or you will arrive at an erroneous result. Remember “P lease E xcuse M y D ear A unt S ally”? This mnemonic

device reminds you to use the order parentheses, exponents, multiplication, division, addition, subtraction. Your calculator automatically

employs proper order of operations, so you need to insert parentheses where appropriate so that you can direct the sequence. To illustrate

this, type the following into your calculator: 3 + 4/2 and (3 + 4) /2. What answers did your calculator produce? Now try −32 and (−3) 2.

What are the results?

This has probably all gotten somewhat murky. Luckily, there is a simple set of steps you can follow anytime

you encounter a two-population test that will help you determine which type of t test to use. This mental

sequence is depicted in Figure 11.2.

In this chapter, we will encounter something we have touched on before but have not addressed in detail: one-

tailed tests versus two-tailed tests. We discussed the t distribution in Chapters 7 and 8. The t distribution is

symmetric and has positive and negative sides. In two-tailed tests, there are two critical values, one positive

and one negative. You learned in Chapter 8 that confidence intervals are always two-tailed. In t tests, by

contrast, some analyses will be two-tailed and some will be one-tailed. Two-tailed tests split alpha (⍺) in half

such that half is in each tail of the distribution. The critical value associated with that ⍺ is both positive and

negative. One-tailed tests, by contrast, place all ⍺ into a single tail. One-tailed tests have ⍺ in either the upper

(or positive) tail or lower (or negative) tail, depending on the specific question under investigation. The critical

value of a one-tailed test is either positive or negative.

One-tailed tests: Hypothesis tests in which the entire alpha is placed in either the upper (positive) or lower (negative) tail such that

there is only one critical value of the test statistic. Also called directional tests.

The choice of a one-tailed test versus a two-tailed test is generally made on a case-by-case basis. It depends on

whether a researcher has a good reason to believe that the relationship under examination should be positive

or negative. Suppose that you are studying an in-prison treatment program that focuses on improving

participants’ literacy skills. You would measure their reading levels before the program began and then again

after it had ended, and would expect to see an increase—you have good reason to predict that post-

intervention literacy skills would be greater than pre-intervention ones. In this case, a one-tailed test would be

in order (these are also called directional tests, since a prediction is being made about the direction of a

relationship). Now suppose you want to know whether that literacy program works better for men or for

women. You do not have any particular reason for thinking that it would be more effective for one group than

the other, so you set out merely to test for any difference at all, regardless of direction. This would be cause for

using a two-tailed test (also called a nondirectional test). Let us work our way through some examples and

discuss one-tailed and two-tailed tests as we go.

Figure 11.2 Steps for Deciding Which t Test to Use

310

311

Independent-Samples t Tests

A Two-Tailed Test With Equal Population Variances: Transferred Female

Juveniles’ Ethnicity and Mean Age of Arrest

There is a theoretical and empirical connection between how old people are when they start committing

delinquent offenses (this is called the age of onset) and their likelihood of continuing law-breaking behavior in

adulthood. All else being equal, younger ages of onset are associated with greater risks of adult criminal

activity. Using the Juvenile Defendants in Criminal Court (JDCC) data set (Data Sources 11.1), we can test

for a significant difference in the mean age at which juveniles were arrested for the offense that led to them

being transferred to adult court. To address questions about gender and ethnicity, the sample is narrowed to

females, and we will test for an age difference between Hispanics and non-Hispanic whites in this subsample.

Among Hispanic female juveniles in the JDCC sample (N = 44), the mean age of arrest was 15.89 years (s =

1.45). Among non-Hispanic whites (N = 31), the mean age of arrest was 16.57 years (s = 1.11). Since the IV is

race (Hispanic; white) and the DV is age at arrest (years) , a t test is the proper analytic strategy. We will use an

⍺ level of .05, a presumption that the population variances are equal, and the five steps of hypothesis testing.

Data Sources 11.1 Juvenile Defendants in Criminal Courts

The JDCC is a subset of the Bureau of Justice Statistics’ (BJS) State Court Processing series that gathers information on defendants

convicted of felonies in large, urban counties. BJS researchers pulled information about juveniles charged with felonies in 40 of these

counties in May 1998. Each case was tracked through disposition. Information about the juveniles’ demographics, court processes,

final dispositions, and sentences was recorded. Due to issues with access to and acquisition of data in some of the counties, the

JDCC is a nonprobability sample, and conclusions drawn from it should therefore be interpreted cautiously (BJS, 1998).

It is useful in an independent-samples t test to first make a table that lays out the relevant pieces of

information that you will need for the test. Table 11.1 shows these numbers. It does not matter which sample

you designate Sample 1 and which you call Sample 2 as long as you stick with your original designation

throughout the course of the hypothesis test. Since it is easy to simply designate the samples in the order in

which they appear in the problem, let us call Hispanic females Sample 1 and white females Sample 2.

We will use a two-tailed test because we have no solid theoretical reason for thinking that non-Hispanic

whites’ mean would be greater than Hispanics’ or vice versa. The alternative hypothesis will merely specify a

difference (i.e., an inequality) between the means, with no prediction about which one is greater than or less

than the other.

312

Step 1. State the null (H0) and alternative (H1) hypotheses .

In t tests, the null (H0) and alternative (H1) are phrased in terms of the population means. Recall that

population means are symbolized μ (the Greek letter mu , pronounced “mew”). We use the population

symbols rather than the sample symbols because the goal is to make a statement about the relationship, or lack

thereof, between two variables in the population. The null hypothesis for a t test is that the means are equal:

H0: μ1 = μ2

Equivalence in the means suggests that the IV is not exerting an impact on the DV. Another way of thinking

about this is that H0 predicts that the two samples came from the same population. In the context of the

present example, retaining the null would indicate that ethnicity does not affect female juveniles’ age of arrest

(i.e., that all female juveniles are part of the same population).

The alternative or research hypothesis is that there is a significant difference between the population means or,

in other words, that there are two separate populations, each with its own mean:

H1: µ1 ≠ µ2

Rejecting the null would lead to the conclusion that the IV does affect the DV; here, it would mean that

ethnicity does appear related to age of arrest. Note that this phrasing of the alternative hypothesis is specific to

two-tailed tests—the “not equal” sign implies no prediction about the direction of the difference. The

alternative hypothesis will be phrased slightly differently for one-tailed or directional tests.

Step 2. Identify the distribution and compute the degrees of freedom .

As mentioned earlier, two-population tests for differences between means employ the t distribution. The t

distribution, you should recall, is a family of curves that changes shape depending on degrees of freedom (df) .

The t distribution is normal at large df and gets wider and flatter as df declines.

The df formula differs across the three types of t tests, so you have to identify the proper test before you can

compute the df. Using the sequence depicted in Figure 11.2, we know (1) that the samples are independent

because this is a random sample divided into two groups and (2) that the population variances are equal. This

leads us to choose the pooled-variances t test. The df formula is

where

N1 = the size of the first sample and

N2 = the size of the second sample.

Pulling the sample sizes from Table 11.1,

313

df = 44 + 31 − 2 = 73

Step 3. Identify the critical value and state the decision rule .

Three pieces of information are required to find the critical value of t (tcrit) using the t table: the number of

tails in the test, the alpha level, and the df. The exact df value of 73 does not appear on the table, so we use the

value that is closest to it, which is 60. With two tails, an ⍺ of .05, and 73 degrees of freedom, we see that tcrit

is 2.000.

This is not the end of finding the critical value, though, because we still have to figure out the sign or signs;

that is, we need to determine whether tcrit is positive, negative, or both. A one-tailed test has only one critical

value and it is either positive or negative. In two-tailed tests, there are always two critical values. Their

absolute values are the same, but one is negative and one is positive. Figure 11.3 illustrates this.

Given that there are two tails in this current test and, therefore, two critical values, tcrit = ±2.000. The decision

rule is stated thus: If tobt is either greater than 2.000 or less than −2.000 , H0 will be rejected. The decision rule

has to be stated as an “either/or” proposition because of the presence of two critical values. There are, in

essence, two ways for the null to be rejected: tobt could be out in the right tail beyond 2.000 or out in the left

tail beyond −2.000.

Figure 11.3 The Critical Values for a Two-Tailed Test With ⍺ = .05 and df = 73

Step 4. Calculate the obtained value of the test statistic .

The formulas for the obtained value of t (tobt) vary across the three different types of t tests; however, the

common thread is to have (1) a measure of the difference between means in the numerator and (2) an estimate

of the standard error of the sampling distribution of differences between means in the denominator

(remember that the standard error is the standard deviation of a sampling distribution). The estimated

standard error is symbolized and the formula for estimating it with pooled variances is

This formula might look a bit daunting, but keep in mind that it comprises only sample sizes and standard

deviations, both of which are numbers you are accustomed to working with. The most important thing is to

314

work through the formula carefully. Plug the numbers in correctly, use proper equation-solving techniques

(including order of operations), and round correctly. Entering the numbers from our example yields

=1.32 (.22)

= .29

This is our estimate of the standard error (i.e., the standard deviation of the sampling distribution). Recall that

this is not tobt! Be careful. The next step is to plug the standard error into the tobt formula. This formula is

Using our numbers, we perform the calculation:

This is the final answer! The obtained value of t is −2.34. Step 4 is done.

Step 5. Make a decision about the null, and state the substantive conclusion .

We said in the decision rule that if tobt was either greater than 2.000 or less than −2.000, the null would be

rejected. So, what will we do? If you said, “Reject the null,” you are correct. Since tobt is less than −2.000, the

null is rejected. The conclusion is that non-Hispanic white and Hispanic female juveniles transferred to adult

court differ significantly in terms of mean age of arrest for their current offense. Another way to think about it

is that there is a relationship between ethnicity and mean age at arrest. Looking at the means, it is clear that

Hispanic youths were younger, on average (mean = 15.89), than non-Hispanic white youths were (mean =

16.57).

The interpretation of a significant result in a two-tailed test is complicated and must be done carefully. In the

present example, we used a two-tailed test because we did not have sufficient theory or prior empirical

evidence to make a defensible prediction about the direction of the difference. The fact that we ultimately

315

found that Hispanics’ mean age was lower is not enough to arrive at a conclusion about the reason for this.

Never let your data drive your thinking—do not construct a reality around an empirical finding. In an

exploratory analysis, it is best to avoid speculating about the larger implications of your results. The more

scientifically sound approach is to continue this line of research to uncover potential reasons why Hispanic

females’ age of arrest might be lower than non-Hispanic white females’ and then collect more data to test this

prediction.

A One-Tailed Test With Unequal Population Variances: Trial Types and

Sentence Lengths

For the second t -test example, we will again use the JDCC data on juveniles diverted to adult criminal courts.

Let’s consider whether the length of time it takes for a juvenile’s case to be disposed of is affected by the type

of attorney the juvenile has. The prediction will be that juvenile defendants who retain private counsel

experience significantly longer time-to-disposition compared to those who use the services of public

defenders. This, theoretically, is because private attorneys might file more pretrial motions and spend more

time negotiating with the prosecutor and the judge. The sample is narrowed to juveniles charged with violent

offenses who were not released pending trial and who were ultimately sentenced to prison. Those with private

attorneys (N = 36) had a mean of 7.93 months to disposition (s = 4.53), whereas those represented by public

defenders (N = 234) experienced a mean of 6.36 months (s = 3.66). We will call the defendants who retained

private attorneys Sample 1 and those who were represented by public defenders Sample 2. Using an alpha

level of .01 and the assumption that the population variances are unequal, we will conduct a five-step

hypothesis test to determine whether persons convicted by juries receive significantly longer prison sentences.

Table 11.2 shows the numbers we will need for the analysis.

Step 1. State the null (H0) and (H1) alternative hypotheses .

The null hypothesis is the same as that used above (H0: µ1 = µ2) and reflects the prediction that the two means

do not differ. The alternative hypothesis used in the previous example, however, does not apply in the present

context because this time, we are making a prediction about which mean will be greater than the other. The

nondirectional sign (≠) must therefore be replaced by a sign that indicates a specific direction. This will either

be a greater than (>) or less than (<) sign. We are predicting that jury trials will result in significantly longer
mean sentences, so we can conceptualize the hypothesis as private attorney disposition time > public defender

disposition time. Since defendants represented by private attorneys are Sample 1 and those by public defenders

Sample 2, the alternative hypothesis is

H1: μ1 > μ2

316

Step 2. Identify the distribution and compute the degrees of freedom .

The distribution is still t , but the df equation for unequal population variances differs sharply from that for

equal variances because the situation of unequal variances mandates the use of the separate-variances t test.

The df formula is obnoxious, but as with the prior formulas we have encountered, you have everything you

need to solve it correctly—just take care to plug in the right numbers and use proper order of operations.

Plugging in the correct numbers from the current example,

317

= 40

Step 2 is complete; df = 40. We can use this df to locate the critical value of the test statistic.

Step 3. Identify the critical value and state the decision rule .

With one tail, an ⍺ of .01, and 40 degrees of freedom, tcrit = 2.423. The sign of the critical value is positive

because the alternative hypothesis predicts that > . Revisit Figure 11.1 for an illustration. When the

alternative predicts that , the critical value will be on the left (negative) side of the distribution and when

the alternative is that > , tcrit will be on the right (positive) side. The decision rule is that if tobt is

greater than 2.423 , H0 will be rejected.

Step 4. Compute the obtained value of the test statistic .

As before, the first step is to obtain an estimate of the standard error of the sampling distribution. Since the

population variances are unequal, the separate variances version of independent-samples t must be used. The

standard error formula for the difference between means in a separate-variance t test is

Plugging in the numbers from the present example yields

Now, the standard error estimate can be entered into the same tobt formula used with the pooled-variances t

test. Using Formula 11(3),

Step 5. Make a decision about the null and state the substantive conclusion .

The decision rule stated that the null would be rejected if tobt exceeded 2.423. Since tobt ended up being 1.94,

the null is retained. Juveniles who had privately retained attorneys did not experience a statistically significant

increase in the amount of time it took for their cases to be resolved, compared to juveniles who had public

defenders. Another way to say this is that there is no relationship between attorney type and disposition time.

318

Dependent-Samples t Tests

The foregoing discussion centered on the situation in which a researcher is working with two independently

selected samples; however, as described earlier, there are times when the samples under examination are not

independent. The main types of dependent samples are matched pairs and repeated measures. Dependent

samples require a t formula different from that used when the study samples are independent because of the

manipulation entailed in selecting dependent samples. With dependent-samples t , the sample size (N) is not

the total number of people or objects in the sample but, rather, the number of pairs being examined. We will

go through an example now to demonstrate the use of this t test.

Research Example 11.2 Do Mentally Ill Offenders’ Crimes Cost More?

There is ongoing debate about the precise role of mental illness (MI) in offending. Some individuals with MI are unstable or

antisocial, but it would be wrong to stigmatize an entire group based on the behavior of a few members. One question that could be

asked is whether the crimes of offenders with MI exact a particularly heavy toll on taxpayers relative to offenders who do not have

MI. Ostermann and Matejkowski (2014) collected data on all persons released from New Jersey prisons in 2006. The data included

the number of times each ex-prisoner was rearrested within a 3-year follow-up period. First, the researchers divided the group

according to whether or not each person had received a MI diagnosis and then calculated the average cost of each group’s recidivism.

The results showed that MI offenders’ crimes were nearly three times more expensive compared to non-MI offenders. Next, the

authors matched the sample of MI offenders to a subsample of non-MI offenders on the basis of each person’s demographic

characteristics and offense histories and then recalculated each group’s average cost. The results changed dramatically. After the one-

to-one matching procedure, the non-MI group’s average cost was more than double that of the MI group. It turns out that the initial

results—the ones suggesting that MI offenders’ crimes are much more expensive—are misleading. It is not good policy to use the

mere existence of MI as cause to enhance supervision or restrictions on ex-prisoners. What should be done instead is to focus on the

risk factors that are associated with recidivisms among both MI and non-MI offenders. This policy focus would cut costs and create

higher levels of social justice.

Dependent-Samples t Test: Female Correctional Officers and Institutional

Security

The traditionally male-dominated field of correctional security is gradually being opened to women who wish

to work in jails and prisons, yet there are lingering concerns regarding how well female correctional officers

can maintain order in male institutions. Critics claim that women are not as capable as men when it comes to

controlling inmates, which threatens the internal safety and security of the prison environment. Let us test the

hypothesis that facilities with relatively small percentages of female security staff will have lower inmate-on-

staff assault rates relative to those institutions with high percentages of female security staff because security in

the latter will be compromised. We will use data from the Census of Jails (COJ; Data Sources 3.1) and an

alpha level of .05. The first sample consists of five jails with below-average percentages of female security staff,

and the second sample contains five prisons selected on the basis of each one’s similarity to a prison in the first

sample (i.e., the second sample’s prisons are all male, state-run, maximum-security facilities in Texas with

inmate totals similar to those of the first sample). The difference between the samples is that the second

sample has above-average percentages of female security staff. Table 11.3 contains the raw data.

Step 1. State the null (N0) and (N1) alternative hypotheses .

319

The null, as is always the case with t tests, is H0: µ1 = µ2. It is being suggested in this problem that low-

percentage female jails (i.e., jails with greater percentages of male staff) should have lower assault rates than

high-percentage female (low male percentage) jails do; hence, the alternative in words is low < high and in
formal symbols is H1: µ1 < µ2.
Step 2. Identify the distribution and compute the degrees of freedom .
The distribution is t. The degrees of freedom calculation for dependent-samples tests is different from those
earlier—for dependent samples, the df is based on the number of pairs rather than on the size of each sample.
The df for pairs is
Using the data from Example 3,
df = 5 − 1 = 4
Step 3. Identify the critical value and state the decision rule .
With an alpha of .05, a one-tailed test, and df = 4, the absolute value derived using the t table is 2.132. To
determine whether this critical value is positive or negative, refer back to Figure 11.1. The phrasing of our
alternative hypothesis makes the critical value negative, so tcrit = −2.132. The decision rule states that if tobt <
−2.132, H0 will be rejected.
Step 4. Compute the obtained value of the test statistic .
The formulas required to calculate tobt for dependent samples looks quite different from those for independent
samples, but the logic is the same. The numerator contains a measure of the difference between means, and
the denominator is an estimate of the standard error. Finding the standard error requires two steps. First, the
standard deviation of the differences between means (symbolized sD) is needed. In the independent-samples t
tests we conducted earlier, we knew each sample’s standard deviation and so we could plug those numbers
directly into the standard-error formula. We do not yet know the standard deviation for the matched-pairs
sample, however, so we have to obtain that number before tackling the standard error. The standard deviation
for matched pairs is computed using the formula
320
where
xD = the difference scores and
= the mean of the difference scores.
The standard deviation is then used to find the standard error of the sampling distribution, as follows:
Finally, the obtained value of the test statistic is calculated as
This computation process is substantially simplified by the use of Table 11.4. This table contains the raw
scores and three additional columns. The first column to the right of the raw scores contains the difference
scores (xD), which are computed by subtracting each x2 score from its corresponding x1 value. The mean
difference score ( ) is then needed and is calculated thus:
The mean difference score is subtracted from each individual difference score (this is the
column) to form deviation scores. Finally, each of these scores is squared and the last column summed. The
final product of Table 11.4 is the sum of the squared deviation scores, located in the lower-right-hand corner.
This number (here, .38) gets entered into the sD formula and the calculations proceed through tobt as such:
And Step 4 is (finally) done! The obtained value of t is 1.29.
321
Step 5. Make a decision about the null and state the substantive conclusion.
The decision rule stated that the null would be rejected if the obtained value was less than −2.132. With tobt =
1.29, the null is retained. The conclusion is that there is no relationship between a jail’s level of female staffing
and its rate of inmate-on-staff assaults. There is no support for the notion that female correctional officers
compromise institutional security.
Research Example 11.3 Do Targeted Interventions Reduce Crime?
The past two decades have seen a shift in criminal justice operations toward an emphasis on multiagency efforts designed to address
the specific crime problems plaguing individual neighborhoods and communities. One strategy that has been devised is the “pulling
levers” approach that entails identifying high-risk offenders and offering them a choice between tough prosecution and reform.
Police and prosecutors explain the penalties these offenders would face if they continue their antisocial behavior, and community
social service providers offer treatment and educational opportunities for those who want to put their lives on a better track. In 2007,
the Rockford (Illinois) Police Department (RPD) implemented a pulling levers approach in an attempt to reduce violent crime in
the city. Corsaro, Brunson, and McGarrell (2013) evaluated the RPD intervention. They measured the mean number of violent
crimes that occurred in several areas of the city per month before and after the program. The following table shows these results. The
target neighborhood is the one in which the pulling levers strategy was implemented.
Using a statistical analysis, the researchers found that the target neighborhood experienced a significant reduction in nonviolent
crime as a result of the pulling levers program. The reduction in violent crime, though, was not statistically significant, so it appeared
that the intervention did not affect these types of offenses. The researchers were able to attribute the target area’s decline in
nonviolent offenses to the pulling levers strategy because crime fell only in the target zone and nowhere else in the city. It thus
appears from this analysis that pulling levers is a promising strategy for reducing nonviolent crimes such as drug, property, and
nuisance offenses but is perhaps less useful with respect to violent crime.
322
323
Two-Population Tests for Differences Between Proportions
Two-population tests for differences between proportions follow the same logic as those for differences
between means. The IV is still a two-class, categorical measure; however, the DV is a proportion rather than a
mean. Differences between proportions have their own sampling distribution, which looks very much like that
for differences between means and can be drawn as Figure 11.4.
Population proportions are symbolized with the letter P and sample proportions with (pronounced “p hat”).
Like confidence intervals for proportions, two-population tests employ the z distribution. Thus, the sample
size must be at least 100; these tests cannot be carried out on samples smaller than that. All the other
fundamentals for proportion tests are the same as those for tests of differences between means, so we will dive
right into an example.
Research Example 11.4 Does the Gender Gap in Offending Rates Differ Between Male and Female Drug Abusers?
It is well known that males have higher rates of criminal involvement than females do, but it is less clear whether this gender
differential exists among people who have criminal histories. Kaskela and Pitkänen (2016) examined the gender gap among
substance abusers with criminal histories. Using data from a national database in Finland, the researchers compared substance
abusers with and without criminal records to determine the magnitude of the differences in offending between men and women in
each group. The table displays information about the people in these four groups.
Several significant differences emerged between men and women in each group, as indicated by the p values of .05 or less. Females of
both groups had lower educational attainment and lower employment rates. Men and women did not differ in rates of drug abuse (as
opposed to having problems with alcohol only). The gender gap was apparent in both groups, as well, with men significantly more
likely than women to have committed new crimes during the study period. Interestingly, the women with criminal histories
committed crimes at higher rates (68.5%) than did men without records (49.3%). The gap between the men and women with
records was 8.3 percentage points and that between men and women without prior convictions was 13.6 percentage points. These
findings suggest that the gender gap narrows among people who have committed criminal offenses in the past and that women with
previous run-ins with the law are more likely than men without such histories to engage in future lawbreaking.
Figure 11.4 The Sampling Distribution of Differences Between Proportions
324
Community-Based Sex-Offender Treatment Programs and Recidivism
Sex offenders are perhaps the most reviled class of criminal offenders, yet imprisoning low-level sex offenders
is not necessarily the best option in terms of social and economic policy. These people might be good
candidates for community-based treatment, which can be effective at reducing the likelihood of recidivism
and is less expensive than imprisonment. Washington State has a program called the Special Sex Offender
Sentencing Alternative (SSOSA) that separates low-risk from high-risk sex offenders and sentences the low-
risk offenders to community-based supervision and treatment. Participants must maintain standards of good
behavior, comply with imposed conditions, and adhere to the required treatment regimen. Transgressors are
removed from SSOSA and incarcerated. This program has raised questions about how well the SSOSA
program reduces recidivism. It has also generated concerns among critics who fear that allowing convicted sex
offenders to remain in the community jeopardizes public safety.
To address the issue of sex offender recidivism and the effectiveness of the SSOSA program, researchers from
the Washington State Institute for Public Policy compared recidivism across sex offenders who were
sentenced to SSOSA and those who were sentenced to jail or prison instead of to this community-based
treatment program. Recidivism was measured as a conviction for a new crime within 5 years of release from
prison or the program. Among the SSOSA group (N = 1,097), 10% were convicted of new crimes; among the
non-SSOSA group that was incarcerated (N = 2,994), 30% were reconvicted (Barnoski, 2005). Using an alpha
level of .05, let us test the null hypothesis that there is no difference between the population proportions
against the alternative hypothesis that the SSOSA group’s recidivism rate was significantly lower than the
incarceration group’s. The use of a directional hypothesis is justified because the SSOSA program is intended
to lower recidivism.
Step 1. State the null (H0) and (H1) alternative hypotheses .
The null hypothesis for a two-population test for differences between proportions represents the prediction
that the two proportions do not differ or, in other words, that the two population proportions are equal. The
alternative hypothesis can take on the three forms discussed in the preceding section regarding two-
population tests for means. Here, calling the SSOSA group Sample 1 and the incarceration group Sample 2,
the hypotheses are
H0: P1 = P2
H1: P1 < P2
325
Step 2. Identify the distribution and compute the degrees of freedom .
The proper distribution for two-population tests of proportions is the z curve. It is important to note, though,
that this distribution can be used only when the samples are large and independent; violation of either of these
assumptions can be fatal to a test of this type. You should always have enough knowledge about the research
design that produced the data you are working with to determine whether the independence assumption has
been met. In the present case, it has, because the SSOSA and incarceration groups were unrelated to each
other. Remember that the proportion test can be conducted only when N ≥ 100; there is an additional
requirement that both the proportion of cases that fall into the category of interest ( ) and the proportion
that does not (symbolized ) are both greater than or equal to five. Formally stated,
and
where
= the sample proportion for each of Sample 1 and Sample 2 and
for each of Sample 1 and Sample 2.
In the current example, the SSOSA group meets the large sample criterion because , which means
that . Therefore,
1,097(.10) = 109.70
1,097(.90) = 987.30
The incarceration group likewise succeeds, because and , so
2,994(.30) = 898.20
2,994(.70) = 2,095.80
The z distribution can be used. There is no need to compute degrees of freedom because they are not
applicable to z.
Step 3. Identify the critical value and state the decision rule .
326
It has been a while since we used the z distribution, but recall that a z value can be found using a known area.
Here, alpha is that area. Since ⍺ = .05 and the test is one-tailed, go to the z table and find the area closest to
.50 − .05 = .45. There are actually two areas that fit this description (.4495 and .4505), so
. The critical value is negative because the alternative
hypothesis (P1 < P2 ) tells us that we are working on the left (negative side) of the distribution (see Figure
11.5). The decision rule is thus: If zobt is less than −1.65, H0 will be rejected.
Step 4. Compute the obtained value of the test statistic .
For this portion of the test, you need to be very careful about symbols because there are a few that are similar
to one another but actually represent very different numbers. Pay close attention to the following:
P = population proportion
= pooled sample proportions as an estimate of the population proportion
= 1.00 − the pooled sample proportions
= Sample 1 proportion
= Sample 2 proportion
There are a few analytical steps in the buildup to the zobt formula. First, the sample proportions have to be
pooled in order to form a single estimate of the proposed population proportion:
The complement of the pooled proportion ( ) is also needed. This is done using the formula
Next, the standard error of the sampling distribution of differences between proportions must be estimated
using the formula
Finally, the obtained value of the test statistic is calculated as
327
Now, we will plug in the numbers from the current example and solve the formulas all the way up to zobt :
Step 5. Make a decision about the null and state the substantive conclusion .
Since zobt = −20.00, which is far less than the critical value of −1.65, the null is rejected. The sex-offender
group that went through the SSOSA program had a significantly lower recidivism rate compared to the group
sentenced to jail or prison. Of course, we cannot say for sure that this is because the SSOSA program is
effective, because the offenders selected to go through this program might be those who were at the lowest
risk for reoffending anyway. What does seem clear is that, at the very least, the SSOSA program is not
harming public safety, could be reducing recidivism rates, and, therefore, might be a useful cost-savings
program.
328
SPSS
The SPSS program can run all of the t tests discussed in this chapter, though it cannot run tests for
differences between proportions, so you would have to use SPSS to derive the proportions and then do the
analysis by hand. Independent-samples t tests are located under the Analyze menu in the SPSS data screen. In
this menu, find Compare Means and then Independent-Samples T Test. To demonstrate the use of SPSS, some
of the examples we did by hand will be replicated. Note that the final answers obtained in SPSS might depart
somewhat from those that we calculated by hand because we use only two decimal places and SPSS uses far
more than that, so our hand-derived answers have rounding errors.
First, let us run an independent-samples t test using the ethnicity and age at arrest variables that were under
study earlier in this chapter. Figure 11.5 shows how to select these variables for an analysis. The IV is
designated as the Grouping Variable and the DV goes into the Test Variable(s) space. You have to specify the
values of the IV, as Figure 11.5 depicts. In the JDCC data set, the ethnicity variable is measured as 1 =
Hispanic and 2 = non–Hispanic white. SPSS asks you to define groups in case the IV has more than two
categories. If your IV has only two categories, then this step is redundant; if it has three or more, this is where
you tell SPSS which two to use in the t test. Click OK to obtain the output shown in Figure 11.6.
The first box in Figure 11.6 shows the descriptive statistics for each category. The second box contains the t
test results. You will notice that there are two values reported for tobt , along with two p values. This is because
SPSS produces results for both pooled- and separate-variances tests. Remember at the beginning of the
chapter when we said that SPSS would help you decide which of these tests you should use? Well, this is it!
Levene’s test for equality of variances is an analysis that SPSS runs automatically. Levene’s F statistic is a
hypothesis test. The null is that the variances are equal (meaning that they can be pooled). The “F” column
displays the obtained value of the F statistic and the “Sig.” column shows the p value for F. It is customary to
use an alpha level of .05 when determining whether or not to reject the null of equivalence. If p <.05, the null
is rejected, which means the variances are unequal and the “Equal variances not assumed” row is the one
containing the correct t test results. If p >.05, the null is retained and the “Equal variances assumed” line is the

one to look at. In Figure 11.6, F = .126 and p = .724, so the null is retained and the variances are equal. The

obtained value of t is −2.203, which is close to the value we arrived at by hand (−2.34). The p value is .031,

which is less than .05, so judging by this output, the null should be rejected. (Typically, the alpha level .05 is

used when interpreting SPSS output. Unless there is good reason to do otherwise, researchers generally

consider any results with p < .05 to be statistically significant.) There is a relationship between ethnicity and
age at arrest or, in other words, there is a significant difference between non-Hispanic white and Hispanic
girls in the age at which they were arrested.
Figure 11.5 Running a Dependent-Samples t Test in SPSS
329
Figure 11.6 SPSS Output for Independent-Samples t Test
The procedure for dependent-samples t tests is a little different. The SPSS file has to be set up so that the
rows represent pairs and two columns contain the two sets of scores for each pair. When you go to Analyze →
Compare Means → Paired-Samples T Test , you will see a variable list on the left and will need to select the two
sets of scores that you wish to compare and move them to the right. Figure 11.7 demonstrates this.
Once you have moved the variables over, click OK and output like that in Figure 11.8 will appear.
Figure 11.8 contains the results of the matched sample of male, maximum-security prisons that we used for
hand calculations in Example 3. The obtained value of t is .964, which differs only slightly from the value of
.95 that our hand computations yielded. The p value is .390. This is well above the alpha level of .05, so the
null is retained. There is no relationship between female staffing and inmate assault rates.
SPSS can be used to run t tests. You should use your knowledge of the research design to determine whether
the samples are independent or dependent. If they are independent, Levene’s F statistic is used to determine
whether the population variances can be assumed equal. If F is not statistically significant (generally at an
alpha of .05), then the null of equality is retained and you should use the equal/pooled variances t ; if the null
is rejected at .05, the unequal/separate variances test is the one to look at. Tests for differences in proportions
must be done by hand.
Figure 11.7 Running a Dependent-Samples t Test in SPSS
330
Figure 11.8 SPSS Output for Dependent (Paired) Samples t Test
As always, GIGO! It is your responsibility to make sure that you select the proper test and run that test
correctly. SPSS is pretty unforgiving—it will not alert you to errors unless those errors are so serious that the
requested analysis simply cannot be run at all. If you make a mistake, such as using dependent samples when
you should use independent samples, SPSS will give you a result that resembles the serial killers (mentioned in
Research Example 11.1) that opened this chapter: It looks normal on the surface, but it is actually
untrustworthy and potentially dangerous.
Chapter Summary
In this chapter, you learned several types of analyses that can be conducted to test for differences between two populations. These
can be used to test for differences between means or for differences between proportions; each type of difference has its own
sampling distribution. The t distribution can be used for means tests and the z for proportions tests.
When you approach a hypothesis-testing question involving a two-class, categorical IV and a continuous DV, you first have to ask
yourself whether the two samples are independent. If they are not, the dependent-samples t test must be used. This would be the
case if the sample consists of matched pairs or is a before-and-after design. If the samples are not matched pairs or repeated
measures, and are instead independent, the next step is to decide whether the population variances are equal. When you use SPSS,
the program automatically produces Levene’s test for equality of variances, so you will know whether you should look to the results
of the t test assuming equality or that in which the assumption of equality has been rejected. If the DV under examination is a
proportion, then the two-population tests for differences in proportions should be used. As long as certain criteria are met, the z
distribution can be employed to determine if two proportions are significantly different from one another.
Thinking Critically
331
1. Suppose your friend ran a t test comparing inmate-on-staff assault rates in prisons with high and low percentages of female
staff and found that prisons with higher levels of female staff also had higher assault rates. Does this mean that female
correctional staff compromise security? Why or why not? If not, what is the flaw in this conclusion? Refer back to Chapter 2
if needed.
2. A researcher evaluating a policing strategy intended to reduce burglary computes the mean rate of burglary in the city before
the strategy was implemented and the mean rate after the program has been in place for six months, then runs an
independent-samples t test to determine if there is a significant difference. What mistake has this researcher made? How
should the researcher collect the data instead, and what kind of t test is appropriate?
Review Problems
1. A researcher wants to test the hypothesis that defendants who plead guilty are sentenced more leniently than those who insist
on going to trial. The researcher measures the plea decision as guilty plea or trial and the sentence as the number of months of
incarceration to which defendants were sentenced.
1. What is the independent variable?
2. What is the level of measurement of the independent variable?
3. What is the dependent variable?
4. What is the level of measurement of the dependent variable?
2. A researcher wants to test the hypothesis that defendants who are released on bail are more likely to be convicted than those
who are unable to post bail or are denied it. The researcher measures pretrial release as released or detained and calculates the
proportion within each group that is convicted.
1. What is the independent variable?
2. What is the level of measurement of the independent variable?
3. What is the dependent variable?
4. What is the level of measurement of the dependent variable?
3. A researcher wants to test the hypothesis that male and female judges differ in the severity of the sentences they hand down.
The researcher gathers a sample of judges and measures gender as male or female and sentence severity as the number of
prison sentences they issue in 1 year.
1. What is the independent variable?
2. What is the level of measurement of the independent variable?
3. What is the dependent variable?
4. What is the level of measurement of the dependent variable?
4. A researcher wants to test the hypothesis that police are less likely to arrest a suspect if the suspect has a full-time job. The
researcher gathers a sample and measures job status as employed or unemployed and records the proportion of each group that
was arrested.
1. What is the independent variable?
2. What is the level of measurement of the independent variable?
3. What is the dependent variable?
4. What is the level of measurement of the dependent variable?
5. In Question 1, the researcher studying guilty pleas and sentence severity would use what kind of statistical test?
1. Test for differences between means
2. Test for differences between proportions
6. In Question 2, the researcher studying pretrial release and conviction likelihood would use what kind of statistical test?
1. Test for differences between means
2. Test for differences between proportions
7. In Question 3, the researcher studying judges’ gender and sentencing behavior would use what kind of statistical test?
1. Test for differences between means
2. Test for differences between proportions
8. In Question 4, the researcher studying suspect unemployment status and arrest would use what kind of statistical test?
1. Test for differences between means
2. Test for differences between proportions
9. Tests for differences between two population means use the ___ distribution, and tests for differences between population
332
proportions use the ___ distribution.
10. A researcher wishes to find out whether a new drug court program appears to be effective at reducing drug use among
participants. He gathers a random sample of drug defendants who are about to enter the drug court’s treatment regimen and
measures the number of times per month that they use drugs. After the participants finish the program, the researcher again
measures their monthly drug use. Which type of t test would be appropriate for analyzing the data?
1. Independent samples, pooled variances
2. Independent samples, separate variances
3. Dependent samples
11. A researcher is investigating the relationship between the restrictiveness of gun laws and gun-crime rates. She gathers a
sample of states and divides them into two groups: strict gun laws or lax gun laws. She then calculates the gun crime rate in
each state. She finds that the two groups have unequal variances. Which type of t test would be appropriate for analyzing the
data?
1. Independent samples, pooled variances
2. Independent samples, separate variances
3. Dependent samples
12. A researcher wishes to test the hypothesis that attorney type affects the severity of the sentence a defendant receives. He
gathers a sample of defendants and records attorney type (publicly funded; privately retained) and the number of days in the jail
or prison sentence. He finds that the groups have equal variances. Which type of t test would be appropriate for analyzing
the data?
1. Independent samples, pooled variances
2. Independent samples, separate variances
3. Dependent samples
13. In Research Example 11.1, you learned of a study by Wright et al. (2008) in which the researchers set out to determine
whether MHOs were diverse in the number and types of crimes they commit or whether, instead, they tend to specialize in
killing. The researchers compared MHOs to SHOs for purposes of this study. MHOs (Sample 1; N = 155) had a mean
diversity index score of .36 (s = .32) and SHOs (Sample 2; N = 463) had a mean of .37 (s = .33). Using an alpha level of .05,
test the hypothesis that the two groups’ means are significantly different. Assume equal population variances. Use all five
steps.
14. Are juvenile defendants who are released pending adjudication processed more slowly than those who are detained? The
JDCC data set records whether a juvenile was released and the number of months it took for that juvenile’s case to reach a
disposition. The sample will be narrowed to black female youth. Released juveniles (Sample 1) had a mean of 4.30 months to
disposition (s = 3.86, N = 123) and detained juveniles (Sample 2) had a mean of 3.57 (s = 4.20, N = 64). Using an alpha level
of .01, test the hypothesis that released juveniles’ mean time-to-adjudication is significantly greater than detained juveniles’
mean. Assume equal population variances. Use all five steps.
15. Do juveniles transferred to adult court get treated more, or less, harshly depending on their age? The JDCC data set contains
information about juveniles’ age at arrest and the sentence length received by those sent to jail pursuant to conviction. Those
male juveniles who were younger than 16 at the time of arrest (N = 85) received a mean of 68.84 days in jail (s = 125.59) and
those who were older than 16 at arrest (N = 741) had a jail sentence mean of 95.24 days (s = 146.91). Using an alpha level of
.05, test the hypothesis that the over-16 group received a significantly longer mean sentence compared to those younger than
16 at the time of arrest. Assume unequal population variances. Use all five steps.
16. One critique of sentencing research is that it generally focuses on the differences between white defendants and defendants of
color and rarely examines differences between minority groups. The JDCC data set can be used to test for differences
between the sentences received by black and Hispanic juveniles convicted of property crimes. Black juveniles (Sample 1; N =
229) were sentenced to a mean of 31.09 months of probation (s = 15.42), and Hispanic juveniles (Sample 2; N = 118)
received a mean of 40.84 months (s = 16.45). Using an alpha level of .01, test the hypothesis that there is a statistically
significant difference between the two group means. Assume unequal population variances. Use all five steps.
17. Do property and drug offenders get sentenced differently? The JDCC data show that among male juveniles convicted and
fined, those convicted of drug offenses (Sample 1) had a mean fine of $674.78 (s = 867.94; N = 160) and that those convicted
of property offenses had a mean fine of $344.91 (s = 251.91; N = 181). Using an alpha level of .05, test the hypothesis that
there is a statistically significant difference between the two groups’ means. Assume equal population variances. Use all five
steps.
18. In Question 16, you tested for a relationship between black and Hispanic juveniles in terms of the length of probation
333
sentences received. Now let’s find out whether there is a between-group difference among those sentenced to community
service. The following table shows pairs of youths matched on gender (male) and offense type (weapons). Using an alpha
level of .05, test the hypothesis that there is a statistically significant difference between the groups’ means. Use all five steps.
19. One of the most obvious potential contributors to the problem of assaults against police officers is exposure—all else being
equal, jurisdictions wherein officers make more arrests might have elevated rates of officer assaults relative to lower-arrest
jurisdictions. The Uniform Crime Reports (UCR) offer state-level information on arrest rates and officer assault rates. The
states in the two samples in the following table were selected based on key similarities; that is, they are all in the western
region of the country, have similar statewide violent crime rates, and have similar populations. The difference is that the
states in the first sample have relatively low arrest rates and those in the second sample have relatively high arrest rates. The
table shows the officer assault rate (number of officers assaulted per 1,000 officers) in each pair of states. Using an alpha level
of .05, test the hypothesis that there is a statistically significant difference between the groups’ means. Use all five steps.
20. Among juvenile victims of gunshots, are males older than females? The Firearm Injury Surveillance Study (FISS) can be used
to derive a set of pairs matched on juvenile status (younger than 18 years), firearm type (handgun), and circumstances of the
incident (intentional assault involving drugs). The table lists the ages of each pair. Using an alpha level of .01, test the
hypothesis that male victims are significantly older than female victims. Use all five steps.
21. The American Bar Association recommends that defendants who obtain pretrial release should have their cases disposed of
within 180 days of their first court appearance. One question with regard to the speed of case processing is whether attorney
type matters. Some evidence suggests that publicly appointed attorneys move cases faster than their privately retained
counterparts do, other evidence points toward the opposite conclusion, and some studies find no difference. The JDCC data
set contains information on attorney type, pretrial release, and days to adjudication. The sample consists of juveniles charged
with drug felonies who were granted preadjudication release. Among those juveniles represented by public attorneys (Sample
1; N = 509), .64 had their cases disposed of in 180 days or less, while .32 of the juveniles who retained private attorneys
334
(Sample 2; N = 73) were adjudicated within 180 days. Using an alpha level of .05, test the hypothesis that there is a
significant difference between the proportions. Use all five steps.
22. Are male victims of crime-involved shootings more likely than female victims to be shot by a stranger? The FISS captures
data on the circumstances and on victims of shootings that took place in the course of a criminal event. Among males
(Sample 1; N = 390), 72% were shot by strangers, and among females (Sample 2; N = 84), 56% were shot by strangers. Using
an alpha level of .05, test the hypothesis that males are significantly more likely than females to be shot by strangers. Use all
five steps.
23. Do pedestrian stops vary in duration depending on the time of day at which they take place? The data set from the Police–
Public Contact Survey (PPCS) PPCS Independent Samples for Chapter 11.sav (www.sagepub.com/gau) contains the variables
time and minutes , which measure whether a stop took place during the day or at night and the number of minutes the stop
lasted. Run an independent-samples t test to determine whether these two variables are related.
1. At an alpha level of .05, will you use the results for the pooled/equal variances t or that for separate/unequal
variances? How did you make this decision?
2. What is the obtained value of t ?
3. Would you reject the null at an alpha level of .01? Why or why not?
4. What is your substantive conclusion? In other words, are there significant differences between the two IV groups?
Are the IV and the DV related?
24. Do juveniles facing charges for violent crimes have more total charges filed against them compared to juveniles charged with
property crimes? The JDCC data set contains information on charge type and total charges. The data file JDCC Independent
Samples for Chapter 11.sav (www.sagepub.com/gau) contains the variables offense and charges. Run an independent-samples t
test to determine whether offense type appears to be related to the total number of charges.
1. At an alpha level of .05, will you use the results for the pooled/equal variances t or that for separate/unequal
variances? How did you make this decision?
2. What is the obtained value of t ?
3. Would you reject the null at an alpha level of .05? Why or why not?
4. What is your substantive conclusion? In other words, are there significant differences between the two IV groups?
Are the IV and the DV related?
25. Critics of the use of DNA evidence in criminal trials sometimes argue that DNA usage would clog up courts due to the delay
caused by waiting for labs to return test results. Supporters claim, though, that any such delays would be justified by the
improvement in the accuracy of felony convictions and acquittals. The Census of State Court Prosecutors is contained in the
file CSCP Independent Samples for Chapter 11.sav. The variable DNA measures whether a prosecutor’s office uses DNA
evidence in plea negotiations or in criminal trials, and the variable convictions displays the number of felony convictions the
office obtained within the past year. Run an independent-samples t test.
1. At an alpha level of .05, will you use the results for the pooled/equal variances t or that for separate/unequal
variances? How did you make this decision?
2. What is the obtained value of t ?
3. Would you reject the null at an alpha level of .01? Why or why not?
4. What is your substantive conclusion? In other words, are there significant differences between the two IV groups?
Are the IV and DV related?
26. In Question 23, we did not account for PPCS respondents’ demographic characteristics. We can ask the same question—
whether stop duration varies across day and night—and this time narrow the sample down by gender and race to create a
subsample of white males. This strategy removes any potentially confounding effects of gender and race, and allows us to
isolate the effects of the day/night variable. Using the data set PPCS Matched Pairs for Chapter 11.sav
(www.sagepub.com/gau), run a dependent-samples t test using the day and night variables.
1. What is the obtained value of t ?
2. Would you reject the null at an alpha level of .05? Why or why not?
3. What is your substantive conclusion? In other words, are there significant differences between the two IV groups?
Are the IV and the DV related?
Note : Answers to review problems in this and subsequent chapters might vary depending on the number of steps used and whether
rounding is employed during the calculations. The answers provided in this book’s key were derived using the procedures illustrated
in the main text.
335
http://www.sagepub.com/gau
http://www.sagepub.com/gau
http://www.sagepub.com/gau
336
Key Terms
t test 247
Independent samples 249
Dependent samples 249
Matched-pairs design 250
Repeated-measures design 250
Pooled variances 250
Separate variances 250
One-tailed tests 251
Glossary of Symbols and Abbreviations Introduced in This Chapter
337
Chapter 12 Hypothesis Testing With Three or More Population
Means Analysis of Variance
338
Learning Objectives
Identify situations in which, based on the levels of measurement of the independent and dependent variables, analysis of variance
is appropriate.
Explain between- and within-group variances and how they can be compared to make a judgment about the presence or absence
of group effects.
Explain the F statistic conceptually.
Explain what the null and alternative hypotheses predict.
Use raw data to solve equations and conduct five-step hypothesis tests.
Explain measures of association and why they are necessary.
Use SPSS to run analysis of variance and interpret the output.
In Chapter 11, you learned how to determine whether a two-class categorical variable exerts an impact on a
continuous outcome measure: This is a case in which a two-population t test for differences between means is
appropriate. In many situations, though, a categorical independent variable (IV) has more than two classes.
The proper hypothesis-testing technique to use when the IV is categorical with three or more classes and the
dependent variable (DV) is continuous is analysis of variance (ANOVA).
Analysis of variance (ANOVA): The analytic technique appropriate when an independent variable is categorical with three or more
classes and a dependent variable is continuous.
As its name suggests, ANOVA is premised on variance. Why do we care about variance when testing for
differences between means? Consider the hypothetical distributions displayed in Figure 12.1. The
distributions have the same mean but markedly disparate variances—one curve is wide and flat, indicating
substantial variance, whereas the other is tall and thin, indicating relatively little variance.
In any analysis of differences between means, the variance associated with each mean must be accounted for.
This is what ANOVA does. It combines means and variances into a single test for significant differences
between means. A rejected null indicates the presence of a relationship between an IV and a DV.
You might be wondering why, if we have a categorical IV and a continuous DV, we do not just use a series of
t tests to find out if one or more of the means are different from the others. Familywise error is the primary
reason that this is not a viable analytic strategy. Every time that you run a t test, there is a certain probability
that the null is true (i.e., that there is no relationship between the IV and the DV) but will be rejected
erroneously. This probability, as we saw in Chapter 9, is alpha, and the mistake is called a Type I error. Alpha
(the probability of incorrectly rejecting a true null) attaches to each t test, so, in a series of t tests, the Type I
error rate increases exponentially until the likelihood of mistake reaches an unacceptable level. This is the
familywise error rate, and it is the reason that you should not run multiple t tests on a single sample.
Familywise error: The increase in the likelihood of a Type I error (i.e., erroneous rejection of a true null hypothesis) that results from
running repeated statistical tests on a single sample.
Another problem is that multiple t tests get messy. Imagine a categorical IV with classes A, B, C, and D. You
339
would have to run a separate t test for each combination (AB, AC, AD, BC, BD, CD). That is a lot of t tests!
The results would be cumbersome and difficult to interpret.
Figure 12.1 Hypothetical Distributions With the Same Mean and Different Variances
The ANOVA test solves the problems of familywise error and overly complicated output because ANOVA
analyzes all classes on the IV simultaneously. One test is all it takes. This simplifies the process and makes for
cleaner results.
340
ANOVA: Different Types of Variances
There are two types of variance analyzed in ANOVA. Both are based on the idea of groups, which are the
classes on the IV. If an IV was political orientation measured as liberal, moderate , or conservative , then liberals
would be a group, moderates would be a group, and conservatives would be a group. Groups are central to
ANOVA.
The first type of variance is between-group variance. This is a measure of the similarity among or difference
between the groups. It assesses whether groups are markedly different from one another or whether the
differences are trivial and meaningless. This is a measure of true group effect. Figure 12.2 illustrates the
concept of between-group variance. The groups on the left cluster closely together, while those on the right
are distinctly different from one another.
Between-group variance: The extent to each group or classes is similar to or different from the others in a sample. This is a measure
of true group effect, or a relationship between the independent and dependent variables.
The second kind of variance is within-group variance and measures the extent to which people or objects
differ from their fellow group members. Within-group variance is driven by random variations between
people or objects and is a measure of error. Figure 12.3 depicts the conceptual idea behind within-group
variance. The cases in the group on the left cluster tightly around their group’s mean, whereas the cases in the
right-hand group are scattered widely around their mean. The left-hand group, then, would be said to have
much smaller within-group variability than the right-hand group.
Within-group variance: The amount of diversity that exists among the people or objects in a single group or class. This is a measure
of random fluctuation, or error.
The ANOVA test statistic—called the F statistic because the theoretical probability distribution for ANOVA
is the F distribution—is a ratio that compares the amount of variance between groups to that within groups.
When true differences between groups substantially outweigh the random fluctuations present within each
group, the F statistic will be large and the null hypothesis that there is no IV–DV relationship will be rejected
in favor of the alternative hypothesis that there is an association between the two variables. When between-
group variance is small relative to within-group variance, the F statistic will be small, and the null will be
retained.
F statistic: The statistic used in ANOVA; a ratio of the amount of between-group variance present in a sample relative to the
amount of within-group variance.
F distribution: The sampling distribution for ANOVA. The distribution is bounded at zero on the left and extends to positive
infinity; all values in the F distribution are positive.
Figure 12.2 Small and Large Between-Group Variability
341
Figure 12.3 Small and Large Within-Group Variability
An example might help illustrate the concept behind the F statistic. Suppose we wanted to test the
effectiveness of a mental-health treatment program on recidivism rates in a sample of probationers. We gather
three samples: treatment-program completers, people who started the program and dropped out, and those
who did not participate in the program at all. Our DV is the number of times each person is rearrested within
2 years of the end of the probation sentence. There will be some random fluctuations within each group; not
everybody is going to have the same recidivism score. This is white noise or, more formally, within-group
variance—in any sample of people, places, or objects, there will be variation. What we are attempting to
discern is whether the difference between the groups outweighs the random variance among the people within
each group. If the program is effective, then the treatment-completion group should have significantly lower
recidivism scores than the other two groups. The impact of the program should be large relative to the
random white noise. We might even expect the dropout group’s recidivism to be significantly less than the no-
treatment group (though probably not as low as the treatment completers). Figure 12.4 diagrams the two
possible sets of results.
Figure 12.4 Recidivism Scores
The x’ s in the figure represent recidivism scores under two possible scenarios: that within-group variance
trumps between-group variance, and that between-group variance is stronger than that within groups. The
overlap depicted on the left side suggests that the treatment program was ineffective, since it failed to pull one
or two groups away from the others. On the right side, the separation between the groups indicates that they
are truly different; this implies that the treatment program did work and that those who completed or started
342
and dropped out are significantly different from each other and from the group that did not participate. An
ANOVA test for the left side would yield a small F statistic, because the between-group variance is minimal
compared to the within-group variance. An ANOVA for the right side, though, would produce a large
(statistically significant) F because the ratio of between-to-within is high.
The F distribution is bounded on the left at zero, meaning it does not have a negative side. As a result, all
critical and obtained values of F are positive; it is impossible for a correctly calculated F to be negative. This is
because F is based on variance and variance cannot be negative.
Take a moment now to read Research Example 12.1, which describes a situation in which researchers would
use ANOVA to test for a difference between groups or, in other words, would attempt to determine whether
there is a relationship between a multiple- class IV and a continuous DV.
Franklin and Fearn’s (2010) IV (race coded as white; black; Hispanic; Asian) was a four-class, categorical
variable. Their DV (sentence length , measured in months) was continuous. ANOVA is the correct bivariate
analysis in this situation.
Let’s get into an example to see the ANOVA steps and calculations in action. We will use the Juvenile
Defendants in Criminal Courts (JDCC; see Data Sources 11.1). We can examine whether attorney type
(measured as public defender, assigned counsel, or private attorney) affects the jail sentences received by male
youth convicted of weapons offenses. Table 12.1 shows the youths’ sentences in months.
Research Example 12.1 Do Asian Defendants Benefit From a “Model Minority” Stereotype?
Numerous studies have found racially based sentencing disparities that are not attributable to differences in defendants’ prior records
or the severity of their instant offenses. Most such studies have focused on white, black, and Hispanic/Latino defendants. One area
of the race-and-sentencing research that has received very little scholarly attention is the effect of race on sentencing among Asians.
Franklin and Fearn (2010) set out to determine whether Asian defendants are treated differently from those of other races. They
predicted that Asians would be sentenced more leniently due to the stereotype in the United States that Asians are a “model
minority,” in that they are widely presumed to be an economically, academically, and socially productive group.
To test the hypothesis that Asian defendants are given lighter sentences relative to similarly situated defendants of other races,
Franklin and Fearn’s (2010) DV was sentence length , which was coded as the number of months of incarceration imposed on
offenders sentenced to jail or prison. The researchers reported the statistics shown in the table with respect to the mean sentence
length across race in this sample.
So, what did the researchers find? It turned out that there were no statistically significant differences between the groups. Franklin
and Fearn (2010) retained the null hypothesis that there is no relationship between race and sentencing, and concluded that Asian
defendants do not, in fact, receive significantly shorter jail or prison sentences relative to other racial groups once relevant legal
factors (e.g., offense type) are taken into account
Source: Adapted from Table 1 in Franklin and Fearn (2010).
343
We will conduct a five-step hypothesis test to determine whether defendants’ conviction histories affect their
sentences. Alpha will be set at .01.
Step 1. State the null (H0) and alternative (H1) hypotheses.
The null hypothesis in ANOVA is very similar to that in t tests. The difference is that now there are more
than two means. The null is phrased as
H0: µ1 = µ2 = µ3
The structure of the null is dependent on the number of groups—if there were four groups, there would be a
µ4 as well, and five groups would require the addition of a µ5 .
The alternative hypothesis in ANOVA is a bit different from what we have seen before because the only
information offered by this test is whether at least one group is significantly different from at least one other
group. The F statistic indicates neither the number of differences nor the specific group or groups that stand
out from the others. The alternative hypothesis is, accordingly, rather nondescript. It is phrased as
H1: some µi ≠ some µj
If the null is rejected in an ANOVA test, the only conclusion possible is that at least one group is markedly
different from at least one other group—there is no way to tell which group is different or how many
between-group differences there are. This is the reason for the existence of post hoc tests, which will be
covered later in the chapter.
Post hoc tests: Analyses conducted when the null is rejected in ANOVA in order to determine the number and location of
differences between groups.
Step 2. Identify the distribution, and compute the degrees of freedom.
344
As aforementioned, ANOVA relies on the F distribution. This distribution is bounded on the left at zero
(meaning it has only positive values) and is a family of curves whose shapes are determined by alpha and the
degrees of freedom (df) . There are two types of degrees of freedom in ANOVA: between-group (dfB) and
within-groups (dfW ). They are computed as
where
N = the total sample size across all groups and
k = the number of groups.
The total sample size N is derived by summing the number of cases in each group, the latter of which are
called group sample sizes and are symbolized nk . In the present example, there are three groups (k = 3) and N =
n1 + n2 + n3 = 5 + 5 + 5 = 15. The degrees of freedom are therefore
dfB = 3 − 1 = 2
dfW = 15 − 3 = 12
Step 3. Identify the critical value and state the decision rule.
The F distribution is located in Appendix E. There are different distributions for different alpha levels, so
take care to ensure that you are looking at the correct one! You will find the between-group df across the top
of the table and the within-group df down the side. The critical value is located at the intersection of the
proper column and row. With ⍺ = .01, dfB = 2, and dfW = 12, Fcrit = 6.93. The decision rule is that if Fobt >

6.93, the null will be rejected. The decision rule in ANOVA is always phrased using a greater than inequality

because the F distribution contains only positive values, so the critical region is always in the right-hand tail.

Step 4. Compute the obtained value of the test statistic.

Step 4 entails a variety of symbols and abbreviations, all of which are listed and defined in Table 12.2. Stop for

345

a moment and study this chart. You will need to know these symbols and what they mean in order to

understand the concepts and formulas about to come.

You already know that each group has a sample size (nk ) and that the entire sample has a total sample size

(N) . Each group also has its own mean ( ), and the entire sample has a grand mean ( ). These sample

sizes and means, along with other numbers that will be discussed shortly, are used to calculate the three types

of sums of squares. The sums of squares are then used to compute mean squares, which, in turn, are used to

derive the obtained value of F. We will first take a look at the formulas for the three types of sums of squares:

total (SST ), between-group (SSB) , and within-group (SSW ).

where

= the sum of all scores i in group k ,

= the sum of each group total across all groups in the sample,

x = the raw scores, and

N = the total sample size across all groups.

where

nk = the number of cases in group k ,

= the mean of group k , and

= the grand mean across all groups.

The double summation signs in the SST formula are instructions to sum sums. The i subscript denotes

individual scores and k signifies groups, so the double sigmas direct you to first sum the scores within each

group and to then add up all the group sums to form a single sum representing the entire sample.

Sums of squares are measures of variation. They calculate the amount of variation that exists within and

between the groups’ raw scores, squared scores, and means. The SSB formula should look somewhat familiar

—in Chapters 4 and 5, we calculated deviation scores by subtracting the sample mean from each raw score.

Here, we are going to subtract the grand mean from each group mean. See the connection? This strategy

346

produces a measure of variation. The sums of squares provide information about the level of variability within

each group and between the groups.

The easiest way to compute the sums of squares is to use a table. What we ultimately want from the table are

(a) the sums of the raw scores for each group, (b) the sums of each group’s squared raw scores, and (c) each

group’s mean. All of these numbers are displayed in Table 12.3.

We also need the grand mean, which is computed by summing all of the raw scores across groups and dividing

by the total sample size N , as such:

Here,

With all of this information, we are ready to compute the three types of sums of squares, as follows. The

process begins with SST :

= 715 – 528.07

= 186.93

347

Then it is time for the between-groups sums of squares:

= 5(4.20 − 5.93)2 = 4(3.75 − 5.93)2 + 6(8.83 − 5.93)2

=5(−1.73)2 + 4(−2.18)2 + 6(2.90)2

= 5(2.99) + 4(4.75) + 6(8.41)

= 14.95 + 19.00 + 50.46

= 84.41

Next, we calculate the within-groups sums of squares:

SSW = SST − SSB = 186.93 − 84.41 = 102.52

348

Learning Check 12.1

A great way to help you check your math as you go through Step 4 of ANOVA is to remember that the final answers for any of the sums

of squares, mean squares, or Fobt will never be negative. If you get a negative number for any of your final answers in Step 4, you will

know immediately that you made a calculation error, and you should go back and locate the mistake. Can you identify the reason why all

final answers are positive? Hint: The answer is in the formulas.

We now have what we need to compute the mean squares (symbolized MS) . Mean squares transform sums of

squares (measures of variation) into variances by dividing SSB and SSw by their respective degrees of freedom,

dfB and dfw . This is a method of standardization. The mean squares formulas are

Plugging in our numbers,

We now have what we need to calculate Fobt . The F statistic is the ratio of between-group variance to within-

group variance and is computed as

Inserting the numbers from the present example,

Step 4 is done! Fobt = 4.94.

Step 5. Make a decision about the null hypothesis and state the substantive conclusion.

The decision rule stated that if the obtained value exceeded 6.93, the null would be rejected. With an Fobt of

4.94, the null is retained. The substantive conclusion is that there is no significant difference between the

groups in terms of sentence length received. In other words, male juvenile weapons offenders’ jail sentences do

not vary as a function of the type of attorney they had. That is, attorney type does not influence jail sentences.

This finding makes sense. Research is mixed with regard to whether privately retained attorneys (who cost

349

defendants a lot of money) really are better than publicly funded defense attorneys (who are provided to

indigent defendants for free). While there is a popular assumption that privately retained attorneys are better,

the reality is that publicly funded attorneys are frequently as or even more skilled than private ones are.

We will go through another ANOVA example. If you are not already using your calculator to work through

the steps as you read and make sure you can replicate the results obtained here in the book, start doing so.

This is an excellent way to learn the material.

For the second example, we will study handguns and murder rates. Handguns are a prevalent murder weapon

and, in some locations, they account for more deaths than all other modalities combined. In criminal justice

and criminology researchers’ ongoing efforts to learn about violent crime, the question arises as to whether

there are geographical differences in handgun-involved murders. Uniform Crime Report (UCR) data can be

used to find out whether there are significant regional differences in handgun murder rates (calculated as the

number of murders by handgun per 100,000 residents in each state). A random sample of states was drawn,

and the selected states were divided by region. Table 12.4 contains the data in the format that will be used for

computations. Alpha will be set at .05.

Step 1. State the null (H0) and alternative (H1) hypotheses.

H0: µ1 = µ2 = µ3 = µ4

H1: some µi ≠ some µj

Step 2. Identify the distribution and compute the degrees of freedom.

This being an ANOVA, the F distribution will be employed. There are four groups, so k = 4. The total

sample size is N = 5 + 5 + 7 + 6 = 23. Using Formulas 12(1) and 12(2), the degrees of freedom are

dfB = 4 − 1 = 3

dfW = 23 − 4 = 19

Step 3. Identify the critical value and state the decision rule.

With ⍺ = .05 and the earlier derived df values, Fcrit = 3.13. The decision rule states that if Fobt > 3.13 , H0 will

be rejected.

350

Step 4. Calculate the obtained value of the test statistic.

We begin by calculating the total sums of squares:

= 78.75 − 49.32

= 29.43

Before computing the between-groups sums of squares, we need the grand mean:

Now SSB can be calculated:

SSB = 5( 93 −1 46) 2 + 5( 88 −1 46) 2 + 7(2 67 −1 46) 2 + 6(1 00 −1 46) 2

=5(−.53)2 + 5(−.58)2 + 7(1.21)2 + 6(−.46)2

= 5(.28) + 5(.34) + 7(1.46) + 6(.21)

= 1.40 + 1.70 + 10.22 + 1.26

351

= 14.58

Next, we calculate the within-groups sums of squares:

SSw = 29.43 −14.58 = 14.85

Plugging our numbers into Formulas 12(7) and 12(8) for mean squares gives

Finally, using Formula 12(9) to derive Fobt ,

This is the obtained value of the test statistic. Fobt = 6.23, and Step 4 is complete.

Step 5. Make a decision about the null and state the substantive conclusion .

In Step 3, the decision rule stated that if Fobt turned out to be greater than 3.13, the null would be rejected.

Since Fobt ended up being 6.23, the null is indeed rejected. The substantive interpretation is that there is a

significant difference across regions in the handgun-murder rate.

Research Example 12.2 Are Juveniles Who Are Transferred to Adult Courts Seen as More Threatening?

Recent decades have seen a shift in juvenile-delinquency policy. There has been an increasing zero tolerance sentiment with respect

to juveniles who commit serious offenses. The reaction by most states has been to make it easier for juveniles to be tried as adults,

which allows their sentences to be more severe than they would be in juvenile court. The potential problem with this strategy is that

there is a prevalent stereotype about juveniles who get transferred or waived to adult court: They are often viewed as vicious, cold-

hearted predators. Judges, prosecutors, and jurors might be biased against transferred juveniles, simply because they got transferred.

This means that a juvenile and an adult could commit the same offense and yet be treated very differently by the court, potentially

even ending up with different sentences.

Tang, Nuñez, and Bourgeois (2009) tested mock jurors’ perceptions about the dangerousness of 16-year-olds who were transferred

to adult court, 16-year-olds who were kept in the juvenile justice system, and 19-year-olds in adult court. They found that mock

jurors rated transferred 16-year-olds as committing more serious crimes, being more dangerous, and having a greater likelihood of

chronic offending relative to non-transferred juveniles and to 19-year-olds. The following table shows the means, standard

deviations, and F tests.

Source: Adapted from Table 1 in Tang et al. (2009).

As you can see, all the F statistics were large; the null was rejected for each test. The transferred juveniles’ means are higher than the

352

other two groups’ means for all measures. These results suggest that transferring juveniles to adult court could have serious

implications for fairness. In some cases, prosecutors have discretion in deciding whether to waive a juvenile over to adult court,

which means that two juveniles guilty of similar crimes could end up being treated very differently. Even more concerning is the

disparity between transferred youths and 19-year-olds—it appears that juveniles who are tried in adult court could face harsher

penalties than adults, even when their crimes are the same.

As another example, we will analyze data from the Firearm Injury Surveillance Study (FISS; Data Sources

8.2) to find out whether victim age varies significantly across the different victim–offender relationships.

There are four relationship categories, and a total sample size of 22. Table 12.5 shows the data and

calculations of the numbers needed to complete the hypothesis test. We will proceed using the five steps.

Alpha will be set at .05.

Step 1. State the null (H0) and alternative (H1) hypotheses.

H0: µ1 = µ2 = µ3 = µ4

H1: some µi ≠ some µj

Step 2. Identify the distribution and compute the degrees of freedom.

This being an ANOVA, the F distribution will be employed. There are four groups, so k = 4. The total

sample size is N = 7 + 6 + 4 + 5 = 22. Using Formulas 12(1) and 12(2), the degrees of freedom are

dfB = 4 − 1 = 3

dfW = 22 − 4 = 18

Step 3. Identify the critical value and state the decision rule.

With ⍺ =.05 and the earlier derived df values, Fcrit = 3.16. The decision rule states that if Fobt > 3.16 , H0 will

be rejected.

353

Step 4. Calculate the obtained value of the test statistic.

The total sums of squares for the data in Table 12.5 is

= 18,472 − 17360.18

= 1,111.82

Next, we need the grand mean:

Now SSB can be calculated:

= 7(29 43 − 28 09) 2 + 6(24 17 − 28 09) 2 + 4(40 00 − 28 09) 2 + 5 (21 40 – 28.09) 2

354

=7(1.34)2 + 6(−3.92)2 + 4(11.91)2 + 5(−6.69)2

= 7(1.80) + 6(15.37) + 4(141.85) + 5(44.76)

= 12.60 + 92.22 + 567.40 + 223.80

= 896.02

Next, we calculate the within-groups sums of squares:

SSw = 1,111.82 − 896.02 = 215.80

And the mean squares are

Finally, Fobt is calculated as

And Step 4 is done. Fobt = 24.91.

Step 5. Make a decision about the null and state the substantive conclusion.

In Step 3, the decision rule stated that if Fobt turned out to be greater than 3.16, the null would be rejected.

Since Fobt is 24.91, we reject the null. It appears that victim age does vary across the different victim–offender

relationship categories.

After finding a significant F indicating that at least one group stands out from at least one other one, the

obvious question is, “Which group or groups are different?” We might want to know which region or regions

have a significantly higher or lower rate than the others or which victim–offender relationship or relationships

contain significantly younger or older victims. The F statistic is silent with respect to the location and number

of differences, so post hoc tests are used to get this information. The next section covers post hoc tests and

measures of association that can be used to gauge relationship strength.

355

When the Null Is Rejected: A Measure of Association and Post Hoc Tests

If the null is not rejected in ANOVA, then the analysis stops because the conclusion is that the IVs and DV

are not related. If the null is rejected, however, it is customary to explore the statistically significant results in

more detail using measures of association (MAs) and post hoc tests. Measures of association permit an

assessment of the strength of the relationship between the IV and the DV, and post hoc tests allow

researchers to determine which groups are significantly different from which other ones. The MA that will be

discussed here is fairly easy to calculate by hand, but the post hoc tests will be discussed and then

demonstrated in the SPSS section, because they are computationally intensive.

Omega squared (ω2) is an MA for ANOVA that is expressed as the proportion of the total variability in the

sample that is due to between-group differences. Omega squared can be left as a proportion or multiplied by

100 to form a percentage. Larger values of ω2 indicate stronger IV–DV relationships, whereas smaller values

signal weaker associations. Omega squared is computed as

Omega squared: A measure of association used in ANOVA when the null has been rejected in order to assess the magnitude of the

relationship between the independent and dependent variables. This measure shows the proportion of the total variability in the

sample that is attributable to between-group differences.

Earlier, we found a statistically significant relationship between region and handgun murder rates. Now we

can calculate how strong the relationship is. Using ω2 ,

Omega squared shows that 41% of the total variability in the states’ handgun-murder rates is a function of

regional characteristics. Region appears to be a very important determinate of the prevalence of handgun

murders.

We can do the same for the test showing significant differences in victims’ ages across four different types of

victim–offender relationships. Plugging the relevant numbers into Formula 12(10) yields

This means that 77% of the variability in victims’ ages is attributable to the relationship between the victim

and the shooter. This points to age being a function of situational characteristics. Younger people are more at

risk of firearm injuries in certain types of situations, while older people face greater risk in other

circumstances. Of course, we still do not know which group or groups are significantly different from which

other group or groups. For this, post hoc tests are needed.

356

There are many different types of post hoc tests, so two of the most popular ones are presented here. The first

is Tukey’s honest significant difference (HSD). Tukey’s test compares each group to all the others in a series

of two-variable hypothesis tests. The null hypothesis in each comparison is that both group means are equal;

rejection of the null means that there is a significant difference between them. In this way, Tukey’s is

conceptually similar to a series of t tests, though the HSD method sidesteps the problem of familywise error.

Tukey’s honest significant difference: A widely used post hoc test that identifies the number and location(s) of differences between

groups.

Bonferroni is another commonly used test and owes its popularity primarily to the fact that it is fairly

conservative. This means that it minimizes Type I error (erroneously rejecting a true null) at the cost of

increasing the likelihood of a Type II error (erroneously retaining a false null). The Bonferroni, though, has

been criticized for being too conservative. In the end, the best method is to select both Tukey’s and

Bonferroni in order to garner a holistic picture of your data and make an informed judgment.

Bonferroni: A widely used and relatively conservative post hoc test that identifies the number and location(s) of differences between

groups.

The computations of both post hoc tests are complex, so we will not attempt them by hand and will instead

demonstrate their use in SPSS.

357

Learning Check 12.2

Would it be appropriate to compute omega squared and post hoc tests for the ANOVA in the example pertaining to juvenile defendants’

attorneys and sentences? Why or why not?

Research Example 12.3 Does Crime Vary Spatially and Temporally in Accordance With Routine Activities Theory?

Crime varies across space and time; in other words, there are places and times it is more (or less) likely to occur. Routine activities

theory has emerged as one of the most prominent explanations for this variation. Numerous studies have shown that the

characteristics of places can attract or prevent crime and that large-scale patterns of human behavior shape the way crime occurs. For

instance, a tavern in which negligent bartenders frequently sell patrons too much alcohol might generate alcohol-related fights, car

crashes, and so on. Likewise, when schools let out for summer break, cities experience a rise in the number of unsupervised juveniles,

many of whom get into mischief. Most of this research, however, has been conducted in Western nations. De Melo, Pereira,

Andresen, and Matias (2017) extended the study of spatial and temporal variation in crime rates to Campinas, Brazil, to find out if

crime appears to vary along these two dimensions. They broke crime down by type and ran ANOVAs to test for temporal variation

across different units of time (season, month, day of week, hour of day). The table displays the results for the ANOVAs that were

statistically significant. (Nonsignificant findings have been omitted.)

Source: Adapted from Table 1 in De Melo et al. (2017).

As the table shows, homicide rates vary somewhat across month. Post hoc tests showed that the summer months experienced spikes

in homicide, likely because people are outdoors more often when the weather is nice, which increases the risk for violent

victimization and interpersonal conflicts. None of the variation across month was statistically significant (which is why there are no

rows for this unit of time in the table). There was significant temporal variation across weeks and hours. Post hoc tests revealed

interesting findings across crime type. For example, homicides are more likely to occur on weekends (since people are out and about

358

more during weekends than during weekdays), while burglaries are more likely to happen on weekdays (since people are at work).

The variation across hours of the day was also significant for all crime types, but the pattern was different within each one. For

instance, crimes of violence were more common in late evenings and into the night, while burglary was most likely to occur during

the daytime hours.

359

SPSS

Let us revisit the question asked in Example 2 regarding whether handgun murder rates vary by region. To

run an ANOVA in SPSS, follow the steps depicted in Figure 12.5. Use the Analyze → Compare Means →

One-Way ANOVA sequence to bring up the dialog box on the left side in Figure 12.5 and then select the

variables you want to use. Move the IV to the Factor space and the DV to the Dependent List. Then click Post

Hoc and select the Bonferroni and Tukey tests. Click Continue and OK to produce the output shown in Figure

12.6.

The first box of the output shows the results of the hypothesis test. You can see the sums of squares, df , and

mean squares for within groups and between groups. There are also total sums of squares and total degrees of

freedom. The number in the F column is Fobt . Here, you can see that Fobt = 6.329. When we did the

calculations by hand, we got 6.23. Our hand calculations had some rounding error, but this did not affect the

final decision regarding the null because you can also see that the significance value (the p value) is .004,

which is less than .05, the value at which ⍺ was set. The null hypothesis is rejected in the SPSS context just

like it was in the hand calculations.

The next box in the output shows the Tukey and Bonferroni post hoc tests. The difference between these tests

is in the p values in the Sig. column. In the present case, those differences are immaterial because the results

are the same across both types of tests. Based on the asterisks that flag significant results and the fact that the

p values associated with the flagged numbers are less than .05, it is apparent that the South is the region that

stands out from the others. Its mean is significantly greater than all three of the other regions’ means. The

Northeast, West, and Midwest do not differ significantly from one another, as evidenced by the fact that all of

their p values are greater than .05.

Figure 12.5 Running an ANOVA in SPSS

In Figure 12.7, you can see that the Fobt SPSS produces (24.719) is nearly identical to the 12.91 we arrived at

by hand. Looking at Tukey’s and Bonferroni, it appears that the categories “relative” and

“friend/acquaintance” are the only ones that do not differ significantly from one another. In the full data set,

the mean age of victims shot by relatives is 21.73 and that for the ones shot by friends and acquaintances is

24.05. These means are not significantly different from each other, but they are both distinct from the means

for stranger-perpetrated shootings (mean age of 29.58) and intimate-partner shootings (39.12).

Figure 12.6 ANOVA Output

360

*The mean difference is significant at the 0.05 level.

We can also use SPSS and the full FISS to reproduce the analysis we did by hand using a sample of

cases. Figure 12.7 shows the ANOVA and post hoc tests.

Figure 12.7 ANOVA Output

361

*The mean difference is significant at the 0.05 level.

Chapter Summary

This chapter taught you what to do when you have a categorical IV with three or more classes and a continuous DV. A series of t

tests in such a situation is not viable because of the familywise error rate. In an analysis of variance, the researcher conducts multiple

between-group comparisons in a single analysis. The F statistic compares between-group variance to within-group variance to

determine whether between-group variance (a measure of true effect) substantially outweighs within-group variance (a measure of

error). If it does, the null is rejected; if it does not, the null is retained.

The ANOVA F , though, does not indicate the size of the effect, so this chapter introduced you to an MA that allows for a

determination of the strength of a relationship. This measure is omega squared (w2 ), and it is used only when the null has been

rejected—there is no sense in examining the strength of an IV–DV relationship that you just said does not exist! Omega squared is

interpreted as the proportion of the variability in the DV that is attributable to the IV. It can be multiplied by 100 to be interpreted

as a percentage.

The F statistic also does not offer information about the location or number of differences between groups. When the null is

retained, this is not a problem because a retained null means that there are no differences between groups; however, when the null is

rejected, it is desirable to gather more information about which group or groups differ from which others. This is the reason for the

existence of post hoc tests. This chapter covered Tukey’s HSD and Bonferroni, which are two of the most commonly used post hoc

tests in criminal justice and criminology research. Bonferroni is a conservative test, meaning that it is more difficult to reject the null

362

hypothesis of no difference between groups. It is a good idea to run both tests and, if they produce discrepant information, make a

reasoned judgment based on your knowledge of the subject matter and data. Together, MAs and post hoc tests can help you glean a

comprehensive and informative picture of the relationship between the independent and dependent variables.

Thinking Critically

1. What implications does the relationship between shooting victims’ ages and these victims’ relationships with their shooters

have for efforts to prevent firearm violence? For each of the four categories of victim–offender relationship, consider the

mean age of victims and devise a strategy that could be used to reach people of this age group and help them lower their risks

of firearm victimization.

2. A researcher is evaluating the effectiveness of a substance abuse treatment program for jail inmates. The researcher

categorizes inmates into three groups: those who completed the program, those who started it and dropped out, and those

who never participated at all. He follows up with all people in the sample six months after their release from jail and asks

them whether or not they have used drugs since being out. He codes drug use as 0 = no and 1 = yes. He plans to analyze the

data using an ANOVA. Is this the correct analytic approach? Explain your answer.

Review Problems

1. A researcher wants to know whether judges’ gender (measured as male; female) affects the severity of sentences they impose

on convicted defendants (measured as months of incarceration) . Answer the following questions:

1. What is the independent variable?

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

2. A researcher wants to know whether judges’ gender (measured as male; female) affects the types of sentences they impose on

convicted criminal defendants (measured as jail; prison; probation; fine; other) . Answer the following questions:

1. What is the independent variable?

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

3. A researcher wishes to find out whether arrest deters domestic violence offenders from committing future acts of violence

against intimate partners. The researcher measures arrest as arrest; mediation; separation; no action and recidivism as number of

arrests for domestic violence within the next 3 years. Answer the following questions:

1. What is the independent variable?

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

4. A researcher wishes to find out whether arrest deters domestic violence offenders from committing future acts of violence

against intimate partners. The researcher measures arrest as arrest; mediation; separation; no action and recidivism as whether

these offenders were arrested for domestic violence within the next 2 years (measured as arrested; not arrested) . Answer the

following questions:

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

5. A researcher wants to know whether poverty affects crime. The researcher codes neighborhoods as being lower-class, middle-

class , or upper-class and obtains the crime rate for each area (measured as the number of index offenses per 10,000 residents).

Answer the following questions:

363

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

6. A researcher wants to know whether the prevalence of liquor-selling establishments (such as bars and convenience stores) in

neighborhoods affects crime in those areas. The researcher codes neighborhoods as having 0–1 , 2–3 , 4–5 , or 6 + liquor-

selling establishments. The researcher also obtains the crime rate for each area (measured as the number of index offenses per

10,000 residents). Answer the following questions:

2. What is the level of measurement of the independent variable?

3. What is the dependent variable?

4. What is the level of measurement of the dependent variable?

5. What type of hypothesis test should the researcher use?

7. Explain within-groups variance and between-groups variance. What does each of these concepts represent or measure?

8. Explain the F statistic in conceptual terms. What does it measure? Under what circumstances will F be small? Large?

9. Explain why the F statistic can never be negative.

10. When the null hypothesis in an ANOVA test is rejected, why are MA and post hoc tests necessary?

11. The Omnibus Crime Control and Safe Streets Act of 1968 requires state and federal courts to report information on all

wiretaps sought by and authorized for law enforcement agencies (Duff, 2010). One question of interest to someone studying

wiretaps is whether wiretap use varies by crime type; that is, we might want to know whether law enforcement agents use

wiretaps with greater frequency in certain types of investigations than in other types. The following table contains data from

the U.S. courts website (www.uscourts.gov/Statistics.aspx) on the number of wiretaps sought by law enforcement agencies in

a sample of states. The wiretaps are broken down by offense type, meaning that each number in the table represents the

number of wiretap authorizations received by a particular state for a particular offense. Using an alpha level of .05, test the

null hypothesis of no difference between the group means against the alternative hypothesis that at least one group mean is

significantly different from at least one other. Use all five steps. If appropriate, compute and interpret omega squared.

364

http://www.uscourts.gov/Statistics.aspx

12. Some studies have found that people become more punitive as they age, such that older people, as a group, hold harsher

attitudes toward people who commit crimes. The General Social Survey (GSS) asks people for their opinions about courts’

handling of criminal defendants. This survey also records respondents’ ages. Use the data below and an alpha level of .05 to

test the null hypothesis of no difference between the group means against the alternative hypothesis that at least one group

mean is significantly different from at least one other. Use all five steps. If appropriate, compute and interpret omega

squared.

13. In the ongoing effort to reduce police injuries and fatalities resulting from assaults, one issue is the technology of violence

against officers or, in other words, the type of implements offenders use when attacking police. Like other social events,

weapon use might vary across regions. The UCRs collect information on weapons used in officer assaults. These data can be

used to find out whether the percentage of officer assaults committed with firearms varies by region. The following table

contains the data. Using an alpha level of .01, test the null of no difference between means against the alternative that at least

one region is significantly different from at least one other. Use all five steps. If appropriate, compute and interpret omega

squared.

365

14. An ongoing source of question and controversy in the criminal court system are the possible advantages that wealthier

defendants might have over poorer ones, largely as a result of the fact that the former can pay to hire their own attorneys,

whereas the latter must accept the services of court-appointed counsel. There is a common perception that privately retained

attorneys are more skilled and dedicated than their publicly appointed counterparts. Let us examine this issue using a sample

of property defendants from the JDCC data set. The IV is attorney type and the DV is days to pretrial release , which measures

the number of days between arrest and pretrial release for those rape defendants who were released pending trial. (Those who

did not make bail or were denied bail are not included.) Using an alpha level of .05, test the null of no difference between

means against the alternative that at least one region is significantly different from at least one other. Use all five steps. If

appropriate, compute and interpret omega squared.

15. In Research Example 12.1, we read about a study that examined whether Asian defendants were sentenced more leniently

than offenders of other races. Let us run a similar test using data from the JDCC. The following table contains a sample of

juveniles convicted of property offenses and sentenced to probation. The IV is race , and the DV is each person’s probation

sentence in months. Using an alpha level of .01, test the null of no difference between means against the alternative that at least

one region is significantly different from at least one other. Use all five steps. If appropriate, compute and interpret omega

squared.

366

16. Across police agencies of different types, is there significant variation in the prevalence of bachelor’s degrees among sworn

personnel? The table contains Law Enforcement Management and Administrative Statistics (LEMAS) data showing a

sample of agencies broken down by type. The numbers represent the percentage of sworn personnel that has a bachelor’s

degree or higher. Using an alpha level of .01, test the null of no difference between means against the alternative that at least

one facility type is significantly different from at least one other. Use all five steps. If appropriate, compute and interpret

omega squared.

17. Let’s continue using the LEMAS survey and exploring differences across agencies of varying types. Problem-oriented

policing has been an important innovation in the police approach to reducing disorder and crime. This approach encourages

officers to investigate ongoing problems, identify their source, and craft creative solutions. The LEMAS survey asks agency

top managers whether they encourage patrol officers to engage in problem solving and, if they do, what percentage of their

patrol officers are encouraged to do this type of activity. Using an alpha level of .05, test the null of no difference between

means against the alternative that at least one agency type is significantly different from at least one other. Use all five steps.

If appropriate, compute and interpret omega squared.

367

18. Do the number of contacts people have with police officers vary by race? The Police–Public Contact Survey (PPCS) asks

respondents to report their race and the total number of face-to-face contacts they have had with officers in the past year.

The following table shows the data. Using an alpha level of .05, test the null of no difference between means against the

alternative that at least one facility type is significantly different from at least one other. Use all five steps. If appropriate,

compute and interpret omega squared.

19. Are there race differences among juvenile defendants with respect to the length of time it takes them to acquire pretrial

release? The data set JDCC for Chapter 12.sav (www.sagepub.com/gau) can be used to test for whether time-to-release varies

by race for juveniles accused of property crimes. The variables are race and days. Using SPSS, run an ANOVA with race as

the IV and days as the DV. Select the appropriate post hoc tests.

1. Identify the obtained value of F.

2. Would you reject the null at an alpha of .01? Why or why not?

3. State your substantive conclusion about whether there is a relationship between race and days to release for juvenile

property defendants.

4. If appropriate, interpret the post hoc tests to identify the location and total number of significant differences.

5. If appropriate, compute and interpret omega squared.

20. Are juvenile property offenders sentenced differently depending on the file mechanism used to waive them to adult court?

The data set JDCC for Chapter 12.sav (www.sagepub.com/gau) contains the variables file and jail , which measure the

mechanism used to transfer each juvenile to adult court (discretionary, direct file, or statutory) and the number of months in

the sentences of those sent to jail on conviction. Using SPSS, run an ANOVA with file as the IV and jail as the DV. Select

the appropriate post hoc tests.

1. Identify the obtained value of F.

2. Would you reject the null at an alpha of .05? Why or why not?

3. State your substantive conclusion about whether there is a relationship between attorney type and days to release for

juvenile defendants.

4. If appropriate, interpret the post hoc tests to identify the location and total number of significant differences.

5. If appropriate, compute and interpret omega squared.

21. The data set FISS for Chapter 12.sav (www.sagepub.com/gau) contains the FISS variables capturing shooters’ intentions

368

http://www.sagepub.com/gau

http://www.sagepub.com/gau

http://www.sagepub.com/gau

(accident, assault, and police involved) and victims’ ages. Using SPSS, run an ANOVA with intent as the IV and age as the

DV. Select the appropriate post hoc tests.

1. Identify the obtained value of F.

2. Would you reject the null at an alpha of .05? Why or why not?

3. State your substantive conclusion about whether victim age appears to be related to shooters’ intentions.

4. If appropriate, interpret the post hoc tests to identify the location and total number of significant differences.

5. If appropriate, compute and interpret omega squared.

369

Key Terms

Analysis of variance (ANOVA) 281

Familywise error 281

Between-group variance 282

Within-group variance 282

F statistic 282

F distribution 282

Post hoc tests 286

Omega squared 297

Tukey’s honest significant difference (HSD) 298

Bonferroni 298

Glossary of Symbols and Abbreviations Introduced in This Chapter

370

Chapter 13 Hypothesis Testing With Two Continuous Variables

Correlation

371

Learning Objectives

Identify situations in which, based on the levels of measurement of the independent and dependent variables, correlation is

appropriate.

Define positive and negative correlations.

Use graphs or hypotheses to determine whether a bivariate relationship is positive or negative.

Explain the difference between linear and nonlinear relationships.

Explain the r statistic conceptually.

Explain what the null and alternative hypotheses predict about the population correlation.

Use raw data to solve equations and conduct five-step hypothesis tests.

Explain the sign, magnitude, and coefficient of determination and use them in the correct situations.

Use SPSS to run correlation analyses and interpret the output.

Thus far, we have learned the hypothesis tests used when the two variables under examination are both

categorical (chi-square) when the independent variable (IV) is categorical and the dependent variable (DV) is

a proportion (two-population z test for proportions), when the IV is a two-class categorical measure and the

DV is continuous (t tests), and when the IV is categorical with three or more classes and the DV is continuous

(analysis of variance, or ANOVA). In the current chapter, we will address the technique that is proper when

both of the variables are continuous. This technique is Pearson’s correlation (sometimes also called Pearson’s

r) , because it was developed by Karl Pearson, who was instrumental in advancing the field of statistics.

Pearson’s correlation: The bivariate statistical analysis used when both independent and dependent variables are continuous.

The question asked in a correlation analysis is, “When the IV increases by one unit, what happens to the

DV?” The DV might increase (a positive correlation), it might decrease (a negative correlation), or it might

do nothing at all (no correlation). Figure 13.1 depicts these possibilities.

Positive correlation: When a one-unit increase in the independent variable is associated with an increase in the dependent variable.

Negative correlation: When a one-unit increase in the independent variable is associated with a decrease in the dependent variable.

A positive correlation might be found between variables such as drug use and violence in neighborhoods:

Since drug markets often fuel violence, it would be expected that neighborhoods with high levels of drug

activity would be more likely to also display elevated rates of violent crime (i.e., as drug activity increases, so

does violence). A negative correlation would be anticipated between the amount of collective efficacy in an

area and the crime rate. Researchers have found that neighborhoods where residents know one another and

are willing to take action to protect their areas from disorderly conditions have lower crime rates. Higher rates

of collective efficacy should correspond to lower crime rates because of collective efficacy’s protective capacity.

The bivariate associations represented by correlations are linear relationships. This means that the amount of

change in the DV that is associated with an increase in the IV remains constant across all levels of the IV and

is always in the same direction (positive or negative). Linear relationships can be contrasted to nonlinear or

curvilinear relationships such as those displayed in Figure 13.2. You can see in this figure how an increase in

372

the IV is associated with varying changes in the DV. Sometimes the DV increases, sometimes it decreases,

and sometimes it does nothing at all. These nonlinear relationships cannot be modeled using correlational

analyses.

Linear relationship: A relationship wherein the change in the dependent variable associated with a one-unit increase in the

independent variable remains static or constant at all levels of the independent variable.

Figure 13.1 Three Types of Correlations Between a Continuous IV and a Continuous DV

The statistic representing correlations is called the r coefficient. This coefficient ranges from −1.00 to +1.00.

The population correlation coefficient is ρ , which is the Greek letter rho (pronounced “row”). A correlation of

±1.00 signals a perfect relationship where a one-unit increase in x (the IV) is always associated with exactly the

same unit increase of x , across all values of both variables. Correlations of zero indicate that there is no

relationship between the two variables. Coefficients less than zero signify negative relationships, whereas

coefficients greater than zero represent positive relationships. Figure 13.3 depicts the sampling distribution for

r.

r coefficient: The test statistic in a correlation analysis.

Figures 13.4, 13.5, and 13.6 exemplify perfect, strong, and weak positive relationships, respectively. These

scatterplots all show that as y increases, so does x , but you can see how the association breaks down from one

figure to the next. Each scatterplot contains what is called a line of best fit—this is the line that minimizes the

distance between itself and each value in the data. In other words, no line would come closer to all the data

points than this one. The more tightly the data points cluster around the line, the better the line represents

the data, and the stronger the r coefficient will be. When the data points are scattered, there is a lot of error

(i.e., distance between the line and the data points), and the r coefficient will be smaller.

Figure 13.2 Examples of Nonlinear Relationships

Figure 13.3 The Sampling Distribution of Correlation Coefficients

373

Figure 13.4 A Perfect Linear, Positive Relationship Between x and y

There are no strict rules regarding what constitutes a “strong” or “weak” value of r. Researchers use general

guidelines to assess magnitudes. In criminal justice and criminology research, values between 0 and ±.29 are

generally considered weak, from about ±.30 to ±.49 are moderate, ±.50 to ±.69 are strong, and anything

beyond ±.70 is very strong. We will use these general guidelines throughout the chapter when assessing the

magnitude of the relationship suggested by a certain r value.

Figure 13.5 A Strong Linear, Positive Relationship Between x and y

Figure 13.6 A Weak Linear, Positive Relationship Between x and y

374

As always, it must be remembered that correlation is not causation. A statistically significant correlation

between two variables means that there is an empirical association between them (one of the criteria necessary

for proving causation, as discussed in Chapter 2), but by itself this is not evidence that the IV causes the DV.

There could be another variable that accounts for the DV better than the IV does but that has been omitted

from the analysis. It could also be the case that both the IV and the DV are caused by a third, omitted

variable. For instance, crime frequently increases during the summer months, meaning that in any given city,

ambient temperature might correlate positively with crime rates. Does this mean heat causes crime? It is

possible that hot temperatures make people extra cranky, but a more likely explanation is that crime increases

in the summer because people are outdoors and there are more adolescents engaged in unsupervised,

unstructured activities. This illustrates the need to think critically about statistically significant relationships

between IVs and DVs. Proceed with caution when interpreting correlation coefficients, and keep in mind that

there might be more going on than what is captured by these two variables.

Research Example 13.1 Part 1: Is Perceived Risk of Internet Fraud Victimization Related to Online Purchases?

Many researchers have addressed the issue of perceived risk with regard to people’s behavioral adaptations. Perceived risk has

important consequences at both the individual level and the community level, because people who believe their likelihood of

victimization to be high are less likely to connect with their neighbors, less likely to use public spaces in their communities, and more

likely to stay indoors to. What has not been addressed with much vigor in the criminal justice and criminology literature is the issue

of perceived risk of Internet theft victimization. Given how integral the Internet is to American life and the enormous volume of

commerce that takes place online every year, it is important to study the online shopping environment as an arena ripe for theft and

fraud.

Reisig, Pratt, and Holtfreter (2009) examined this issue in the context of Internet theft victimization. They used self-report data

from a survey administered to a random sample of citizens. Their research question was whether perceived risk of Internet theft

victimization would dampen people’s tendency to shop online because of the vulnerability created when credit cards are used to make

Internet purchases. They also examined whether people’s financial impulsivity (the tendency to spend money rather than save it and

to possibly spend more than one’s income provides for) affected perceived risk. The researchers ran a correlation analysis. Was their

hypothesis supported? We will revisit this study later in the chapter to find out.

Correlation analyses employ the t distribution because this probability distribution adequately mirrors the

sampling distribution of r at small and large sample sizes (see Figure 13.3). The method for conducting a

correlation analysis is to first calculate r and then test for the statistical significance of r by comparing tcrit and

tobt. Keep this two-step procedure in mind so that you understand the analytic technique in Step 4.

375

For our first example, we will use the Police–Public Contact Survey (PPCS; see Data Sources 2.1) to find out

if there is a relationship between respondents’ ages and the number of face-to-face contacts they have had

with police officers in the past year. Table 13.1 shows the data for a random sample of seven prisons. We will

set alpha at .05.

Step 1. State the null (H0) and alternative (H1) hypotheses.

In a correlation analysis, the null hypothesis is that there is no correlation between the two variables. The null

is phrased in terms of ρ , the population correlation coefficient. This is the Greek letter rho (pronounced

“roe”). Recall that a correlation coefficient of zero signifies an absence of a relationship between two variables;

therefore, the null is

H0: ρ = 0

Three options are available for the phrasing of the alternative hypothesis. Since correlations use the t

distribution, these three options are the same as those in t tests. There is a two-tailed option (written H1: ρ ≠

0) that predicts a correlation of unspecified direction. This is the option used when a researcher does not wish

to make an a priori prediction about whether the correlation is positive or negative. There are also two one-

tailed options. The first predicts that the correlation is negative (H1: ρ < 0) and the second predicts that it is
positive (H1: ρ > 0).

In the present example, we have no a priori reason for predicting a direction. Past research suggests that young

people have more involuntary contacts (such as traffic stops) whereas older people have more voluntary ones

(such as calling the police for help). Since the contact variable used here includes all types of experiences, a

prediction cannot be made about how age might correlate with contacts; therefore, we will use a two-tailed

test. The alternative hypothesis is

H1: ρ ≠ 0

376

Step 2. Identify the distribution and compute the degrees of freedom.

The t distribution is the probability curve used in correlation analyses. Recall that this curve is symmetric and,

unlike the χ2 and F distributions, has both a positive side and a negative side. The degrees of freedom (df) in

correlation are computed as

In the present example, there are five people in the sample, so

df = 7 − 2 = 5

Step 3. Identify the critical value, and state the decision rule.

With a two-tailed test, ⍺ = .05, and df = 5, the value of tcrit is 2.571. Since this is a two-tailed test, there are

two critical values because half of ⍺ is in each tail. This means tcrit = ±2.571. The decision rule is that if tobt is

either > 2.571 or < −2.571 , H0 will be rejected. Step 4. Compute the obtained value of the test statistic. There are two parts to Step 4 in correlation analyses: First, the correlation coefficient r is calculated, and second, the statistical significance of r is tested by plugging r into the tobt formula. The formula for r looks complex, but we will solve it step by step. The formula requires several different sums. These sums must be calculated and then entered into the equation. The easiest way to obtain these sums is to use a table. Table 13.2 reproduces the raw data from Table 13.1 and adds three columns to the right that allow us to compute the needed sums. The sums from Table 13.2 can be entered into Formula 13(2): 377 = −.13. The first part of Step 4 is thus complete. We now know that r = −.13. This is a low value, so we can see already that there is not much of a correlation between the variables, if indeed there is any correlation at all. It is possible this r value is merely a chance finding. Until we either reject or retain the null, we cannot reach any conclusions about r. To make a decision about the null, the following formula is used: Plugging our numbers in, = – .13(2.26) = −.29 Step 4 is complete! We know that r = −.13 and tobt = −.29. 378 379 Learning Check 13.1 The calculated value of r and obtained value of t will always have the same sign: If r is negative, t will be negative, and if r is positive, t will be positive. Can you explain the reason for this? If you need a hint, refer back to Figure 13.3. Step 5. Make a decision about the null and state the substantive conclusion. In the decision rule, we said that the null would be rejected if tobt ended up being > 2.571 or < −2.571. Since the tobt we calculated is only −.29, the null must be retained. We conclude that there is no relationship between people’s ages and the number of contacts they have with police officers. For a second example of correlation testing, we turn to the Law Enforcement Management and Administrative Statistics (LEMAS; see Data Sources 3.2) survey. We will also pull from the Uniform Crime Reports (UCR; see Data Sources 1.1). The LEMAS data set contains a variable measuring the ratio of officers per 1,000 citizens, which indicates the size of the police force relative to the local population. From the UCR, we can calculate each city’s violent crime rate (per 1,000). We will test the hypothesis that police force size correlates negatively with the crime rate (i.e., cities with larger police departments have less violent crime and vice versa). Alpha will be set at .01. Table 13.3 contains the data and the sums we will need for the r formula. Step 1. State the null (H0) and alternative (H1) hypotheses. H0: ρ = 0 H1: ρ < 0 Step 2. Identify the distribution and compute the degrees of freedom. The distribution is t , and the df are computed using Formula 13(1): df = 8 − 2 = 6 Step 3. Identify the critical value and state the decision rule. With a one-tailed test, an alpha of .01, and df = 6, tcrit = −3.143. The critical value is negative because the alternative hypothesis predicts that the correlation is less than zero. The decision rule is that if tobt < −3.143, H0 will be rejected. 380 Step 4. Compute the obtained value of the test statistic. Using Formula 13(2) and the sums from Table 13.3, = −.29 This is a modest r value, suggesting that there is an overlap between these two variable, but it is not a substantial amount. Nonetheless, r might be statistically significant, so we proceed to finding tobt using Formula 13(3): 381 = – .29 (2.55) = −.74 And Step 4 is done! tobt = −.74. Step 5. Make a decision about the null and state the substantive conclusion. The decision rule stated that the null would be rejected if tobt was less than −2.896. Since −.74 does not meet this requirement, the null must be retained. There is not a statistically significant correlation between the size of a city’s police force and the violent crime rate in that area. Research Example 13.2 Do Prisoners’ Criminal Thinking Patterns Predict Misconduct? Institutional security inside prisons depends in part on officials’ ability to accurately classify prisoners as low, medium, or high risk. Reliable classification decisions also benefit inmates so that, for instance, inmates with relatively minimal criminal tendencies are not housed with hardened, dangerous offenders. Static risk factors (such as criminal history) can help officials make accurate decisions. Static risk factors, though, are limited because they capture only general characteristics about inmates and do not tap into those individuals’ thoughts and attitudes. Walters (2015) proposed that incorporating dynamic risk factors would improve the accuracy of prison officials’ classification decisions. He collected data on all inmates entering a medium-security, men’s prison between 2003 and 2010. Each inmate was administered an inventory measuring the extent of his criminal thinking and antisocial attitudes. Walters then collected data on the number of disciplinary infractions each person was found guilty of during the study period. The table shows the correlations between the DVs (infractions) and IVs (criminal thinking and static risk factors). p < .05; p < .001. Source: Adapted from Table 1 in Walters (2015). Walters’ analyses showed that, as predicted, criminal thinking was significantly and positively related to prison disciplinary infractions. In other words, inmates scoring higher on this measure of antisocial attitudes also misbehaved more often. This correlation was fairly weak, however, suggesting that although knowing an inmate’s score on the criminal thinking inventory helps make accurate classification decisions, this score should be one of many pieces of information used when deciding how to classify an inmate. 382 Do Good Recruits Make Good Cops? The purpose of police training academies is to prepare recruits for work on the street; however, it is not known to what extent academy training relates to on-the-job performance. Henson, Reyns, Klahm, and Frank (2010) attempted to determine whether, and to what extent, recruits’ academy performance is associated with their later effectiveness as police officers. They used three DVs: the evaluations new officers received from their supervisors, the number of complaints that were lodged against those new officers, and the number of commendations the officers received for exemplary actions. The IVs consisted of various measurements of academy performance. Henson et al. obtained the correlations shown in the table. Note that the authors did not report correlation coefficients that were not statistically significant; the nonsignificant coefficients are replaced with ns. p <.01; p <.001. Source: Adapted from Table 3 in Henson et al. (2010). Were the authors’ predictions supported? Is academy performance related to the quality of the job those recruits do once they are on the street? The results were mixed. You can see in the table that many of the correlations were statistically significant, but nearly as many were not. Even those results that were statistically significant were weak in magnitude. It would appear from this analysis that academy performance and on-the-job performance are not strongly related. An officer’s academy performance is only marginally related to how well he or she carries out the police job. For a third example, we will draw from the data set Juvenile Defendants in Criminal Courts (JDCC; Data Sources 11.1). This data set contains information on the dollar amount of penalties imposed on juveniles convicted of crimes whose sentence includes paying a fine. A question we might ask is whether there is a correlation between the number of charges a juvenile is prosecuted for and the size of the fine imposed on him or her. It seems logical to expect these variables to correlate positively, such that more charges result in heavier fines. Therefore, this will be a directional test and we will be working on the right-hand (positive) side of the t distribution. Table 13.4 shows the data for a sample of eight juveniles. Alpha will be .05. Step 1. State the null (H0) and alternative (H1) hypotheses. H0: ρ = 0 H1: ρ > 0

383

Step 2. Identify the distribution and compute the degrees of freedom.

The distribution is t , and the df are computed using Formula 13(1):

df = 8 − 2 = 6

Step 3. Identify the critical value and state the decision rule.

For a one-tailed test, an alpha of .05, and df = 6, the critical value of t is 1.943. The decision rule is that if tobt

is greater than 1.943 , H0 will be rejected.

Step 4. Compute the obtained value of the test statistic.

Using Formula 13(2) and the sums from Table 13.4,

= .74

384

The r value is large, indicating a strong positive relationship between the variables. We still need to carry out

the calculations for tobt, though, because we have not yet ruled out the possibility that this r is a fluke finding.

The t test will tell us that

= .74 (3.65)

= 2.70

Step 5. Make a decision about the null and state the substantive conclusion.

The decision rule stated that the null would be rejected if tobt ended up being greater than 1.943. The value of

tobt greatly exceeds 1.943, so we will reject the null. Among juveniles tried in adult courts and sentenced to

fines, there is a statistically significant correlation between the number of criminal charges filed against a

person and the dollar amount of the fine imposed.

385

Beyond Statistical Significance: Sign, Magnitude, and Coefficient of

Determination

When the null hypothesis is rejected in a correlation hypothesis test, the correlation coefficient r can be

examined with respect to its substantive meaning. We have touched on the topic of magnitude versus

statistical significance already; note that in all three examples of correlation tests, we made a preliminary

assessment of the magnitude of r before moving on to the tobt calculations. In each one, though, we had to do

that last step to check for statistical significance before formally interpreting the strength or weakness of r.

The reverse of this is also true: Statistical significance is not proof that the variables are strongly correlated.

The null can be rejected even when a correlation is of little practical importance. The biggest culprit of

misleading significance is sample size: Correlations that are substantively weak can result in rejected nulls

simply because the sample is large enough to drive up the value of tobt. When the null is rejected, criminal

justice and criminology researchers turn to three interpretive measures to assess the substantive importance of

a statistically significant r : sign , magnitude , and coefficient of determination.

The sign of the correlation coefficient indicates whether the correlation between the IV and the DV is

negative or positive. Take another look at Figure 13.1 to refresh your memory as to what negative and positive

correlations look like. A positive correlation means that a unit increase in the IV is associated with an increase

in the DVs, and a negative correlation indicates that as the IV increases, the DV declines.

The magnitude is an evaluation of the strength of the relationship based on the value of r. As noted in the

outset of this chapter, there are no rules set in stone for determining whether a given r value is strong,

moderate, or weak in magnitude; this judgment is based on a researcher’s knowledge of the subject matter. As

described previously, a general guideline is that values between 0 and ±.29 are weak, from about ±.30 to ±.49

are moderate, ±.50 to ±.69 are strong, and those beyond ±.70 are very strong.

Third, the coefficient of determination is calculated as the obtained value of r , squared (i.e., r2). The result is

a proportion that can be converted to a percentage and interpreted as the percentage of the variance in the DV

that is attributable to the IV. As a percentage, the coefficient ranges from 0 to 100, with higher numbers

signifying stronger relationships and numbers closer to zero representing weaker associations.

Let us interpret the sign, magnitude, and coefficient of determination for the correlation coefficient computed

in the third example. Since we retained the null in the first and second examples, we cannot apply the three

interpretive measures to these r values. In the examination of charges and fines, we found that r = .74.

First, the sign of the correlation coefficient is positive, meaning that a greater number of charges is associated

with a higher fine amount. Second, the magnitude is quite strong, as .74 exceeds the .70 threshold. Third, the

coefficient of determination is r2 =.742 = .55. This means that 55% of the variance in the DV (fine amount)

can be explained by the IV (number of charges). This is a decent amount of shared variance! Of course, we

cannot draw causal conclusions—there are numerous factors that enter into judges’ sentencing decisions.

Anytime you interpret the outcome of a correlation hypothesis test, keep in mind that statistical significance is

386

not, by itself, enough to demonstrate a practically significant or substantively meaningful relationship between

two variables; moreover, even a strong association does not mean that one variable truly causes the other.

387

SPSS

Correlations are run in SPSS using the Analyze → Correlation → Bivariate sequence. Once the dialog box

shown in Figure 13.7 appears, select the variables of interest and move them into the analysis box as shown. In

this example, the JDCC data on charges and fines is used. You can see in Figure 13.7 that both of these

variables have been moved from the list on the left into the box on the right. After selecting your variables,

click OK. Figure 13.8 shows the output.

The output in Figure 13.8 is called a correlation matrix, meaning that it is split by what is called a diagonal

(here, the cells in the upper-left and lower-right corners of the matrix) and is symmetric on both of the off-

diagonal sides. The numbers in the diagonal are always 1.00 because they represent each variable’s correlation

with itself. The numbers in the off-diagonals are the ones to look at. The number associated with Pearson’s

correlation is the value of r. In Figure 13.8, you can see that r = .239, which is much smaller than what we

arrived at by hand, but since the SPSS example is employing the entire data set, variation is expected.

The Sig. value is, as always, the obtained significance level or p value. This number is compared to alpha to

determine whether the null will be rejected. If p < ⍺, the null is rejected; if p > ⍺, the null is retained.

Typically, any p value less than .05 is considered statistically significant. In Figure 13.8, the p value is .000,

which is very small and indicates that we would reject the null even if we set ⍺ at .001, a very stringent test for

statistical significance.

Figure 13.7 Running a Correlation Analysis in SPSS

Correlation matrices can be expanded to include multiple variables. When you do this, SPSS runs a separate

analysis for each pair. Let us add juvenile defendants’ ages to the matrix. Figure 13.9 shows the output

containing all three variables.

Figure 13.8 SPSS Output

388

** Correlation is significant at the 0.01 level (2-tailed).

389

Learning Check 13.2

In a correlation matrix, the numbers in the diagonal will all be 1.00 because the diagonal represents each variable’s correlation with itself.

Why is this? Explain why any given variable’s correlation with itself is always 1.00.

Research Example 13.1, Continued

390

Part 2: Is Perceived Risk of Internet Fraud Victimization Related to

Online Purchases?

Recall that Reisig et al. (2009) predicted that people’s perceived likelihood of falling victim to Internet theft would lead to less-

frequent engagement in the risky practice of purchasing items online using credit cards. They also thought that financial impulsivity,

as an indicator of low self-control, would affect people’s perceptions of risk. The following table is an adaptation of the correlation

matrix obtained by these researchers.

Were the researchers’ hypotheses correct? The results were mixed. On the one hand, they were correct in that the correlations

between the IVs and the DV were statistically significant. You can see the significance of these relationships indicated by the

asterisks that flag both of these correlations as being statistically significant at an alpha level of .05. Since the null was rejected, it is

appropriate to interpret the coefficients. Regarding sign, perceived risk was negatively related to online purchases (i.e., greater

perceived risk meant less online purchasing activity), and financial impulsivity was positively related to perceived risk (i.e., financially

impulsive people were likely to see themselves as facing an elevated risk of victimization).

The reason the results were mixed, however, is that the correlations—though statistically significant—were not strong. Using the

magnitude guidelines provided in this chapter, it can be seen that −.12 and .11 are very weak. The coefficient of determination for

each one is (−.12)2 = .01 and .112 = .01, so only 1% of the variance in online purchases was attributable to perceived risk, and only

1% was due to financial impulsivity, respectively. This illustrates the potential discrepancy between statistical significance and

substantive importance: Both of these correlations were statistically significant, but neither meant much in terms of substantive or

practical implications. As always, though, it must be remembered that these analyses were bivariate and that the addition of more

IVs might alter the IV–DV relationships observed here.

Source: Adapted from Table 1 in Reisig et al. (2009).

p <.05. Figure 13.9 SPSS Output 391 It turns out that age bears no relationship with fine severity (or with the number of charges filed, which was not the original purpose of our analysis, but something we can see in the correlation matrix). The correlation between age and fines is .033, and p = .390, which is nowhere close to .05. Chapter Summary This chapter introduced Pearson’s correlation, the hypothesis-testing technique that is appropriate in studies employing two continuous variables. The correlation coefficient is symbolized r and ranges from −1.00 to +1.00. The value of r represents the amount of linear change in the DV that accompanies a one-unit increase in the IV. Values of r approaching ±1.00 are indicative of strong relationships, whereas values close to zero signify weak correlations. The statistical significance of r is tested using the t distribution because the sampling distribution of r is normal in shape. When the null hypothesis is rejected in a correlation test, the sign, magnitude, and coefficient of determination should be examined in order to determine the substantive, practical meaning of r. The sign indicates whether the correlation is positive or negative. The magnitude can be assessed using general guidelines and subject-matter expertise. The coefficient of determination is computed by squaring r to produce a proportion (or percentage, when multiplied by 100) that represents the amount of variance in the DV that is attributable to the IV. Correlation analyses can be run in SPSS. The program provides the calculated value of the correlation coefficient r and its associated significance value (its p value). When p is less than alpha (typically set at .05), the null is rejected and the correlation coefficient can be interpreted for substantive meaning; when p is greater than alpha, the null is retained and r is not interpreted because the conclusion is that the two variables are not correlated. Thinking Critically 1. At the beginning of the chapter, it was explained that correlations are tests for linear relationships between two variables. Figures 13.4 through 13.6 depicted the straight lines that represent linear relationships. Compare these straight lines to the curvilinear lines pictured in Figure 13.2. Why would it be a problem to run a correlation analysis on two variables whose relationship looked like one of those in Figure 13.2? (Hint: Try drawing a curvilinear line and then a straight line over the top of it.) Do you think the r coefficient would be an accurate reflection of the relationship between these two variables? Explain your answer. 2. In one of the examples, we found a statistically significant, positive relationship between the number of charges filed against juveniles in criminal courts and the severity of the fine imposed (among those sentenced to pay fines). Throughout this book, you have seen examples of the need to consider additional variables before making a decision about the trustworthiness of a bivariate relationship. What more would you want to know about juvenile defendants’ cases in order to feel that you have a clear picture of the charge−fine relationship? Identify the variables that you would add to this analysis. Review Problems 1. A researcher wishes to test the hypothesis that parental incarceration is a risk factor for lifetime incarceration of the adult 392 children of incarcerated parents. She measures parental incarceration as had a parent in prison; did not have a parent in prison and adult children’s incarceration as incarcerated; not incarcerated . 1. Identify the IV. 2. Identify the level of measurement of the IV. 3. Identify the DV. 4. Identify the level of measurement of the DV. 5. What type of inferential analysis should the researcher run to test for a relationship between these two variables? 2. A researcher wishes to test the hypothesis that coercive control can actually increase, rather than reduce, subsequent criminal offending among persons who have been imprisoned. He gathers a sample of people who have spent time in prison and measures coercive control as number of years spent in prison and recidivism as number of times rearrested after release from prison . 1. Identify the IV. 2. Identify the level of measurement of the IV. 3. Identify the DV. 4. Identify the level of measurement of the DV. 5. What type of inferential analysis should the researcher run to test for a relationship between these two variables? 3. A researcher hypothesizes that a new policing strategy involving community meetings designed to educate residents on the importance of self-protective measures will increase the use of such measures. He gathers a sample of local residents and finds out whether they have participated in an educational session (have participated; have not participated) . Among each group, he computes the proportion of residents who now take self-protective measures . 1. Identify the IV. 2. Identify the level of measurement of the IV. 3. Identify the DV. 4. Identify the level of measurement of the DV. 5. What type of inferential analysis should the researcher run to test for a relationship between these two variables? 4. A researcher thinks that there might be a relationship between neighborhood levels of socioeconomic disadvantage and the amount of violent crime that occurs in those neighborhoods. She measures socioeconomic disadvantage as the percentage of neighborhood residents that live below the poverty line , and she measures violent crime as the number of violent offenses reported to police per 1,000 neighborhood residents . 1. Identify the IV. 2. Identify the level of measurement of the IV. 3. Identify the DV. 4. Identify the level of measurement of the DV. 5. What type of inferential analysis should the researcher run to test for a relationship between these two variables? 5. Explain what it means for a relationship to be linear. 6. Explain what it means for a relationship to be nonlinear. 7. Explain the concept of the line of best fit. 8. When a correlation is ____, increases in the IV are associated with increases in the DV. 1. positive 2. negative 3. zero 9. When a correlation is ____, increases in the IV are not associated with any predictable or consistent change in the DV. 1. positive 2. negative 3. zero 10. When a correlation is ____, increases in the IV are associated with decreases in the DV. 1. positive 2. negative 3. zero 11. Is there a correlation between the amount of money states spend on prisons and those states’ violent crime rates? There could be a negative correlation insofar as prisons might suppress crime, in which case money invested in incarceration produces a reduction in violence. There could also be a positive correlation, however, if prison expenditures do not reduce crime but, rather, merely reflect the amount of violence present in a state. Morgan, Morgan, and Boba (2010) offer information on the 393 dollars spent per capita on prison expenditures in 2009, and the UCR provide 2009 violent crime data. The following table contains dollars spent per capita and violent crime rates per 1,000 citizens for a random sample of five states. Using an alpha level of .05, test for a correlation between the variables (note that no direction is being specified). Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 12. One aspect of deterrence theory predicts that punishment is most effective at preventing crime when would-be criminals know that they stand a high likelihood of being caught and penalized for breaking the law. When certainty breaks down, by contrast, crime rates might increase because offenders feel more confident that they will not be apprehended. The following table contains UCR data on the clearance rate for property crimes and the property crime rate per 100. The prediction is that higher clearance rates are associated with lower property crime rates. Using an alpha of .01, test for a negative correlation between the variables. Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 13. Are there more police agencies in areas with higher crime rates? We will find out using data from Morgan et al. (2010). The IV is crimes per square mile , and the DV is police agencies per 1,000 square miles. Using an alpha level of .05, test for a positive correlation between the variables. Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 394 14. In police agencies—much like in other types of public and private organizations—there are concerns over disparities in pay; in particular, lower-ranking officers might feel undercompensated relative to higher-ranking command staff and administrators. Let us investigate whether there is a correlation between the pay of those at the top of the police hierarchy (chiefs) and those at the bottom (officers). We will use a random sample of six agencies from the LEMAS data set. The following table shows the minimum annual salaries of chiefs and officers among the agencies in this sample. (The numbers represent thousands; for instance, 57.5 means $57,500.) Using an alpha level of .05, test for a correlation between the two variables (no direction specified). Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 15. It is well known that handguns account for a substantial portion of murders. This has led some people to claim that stricter handgun regulations would help curb the murder rate in the United States. Others, though, say that tougher gun laws would not work because people who are motivated to kill but who cannot obtain a handgun will simply find a different weapon instead. This is called a substitution effect. If the substitution effect is operative, then there should be a negative correlation between handgun and knife murders. The following table contains data from a random sample of states. For each state, the handgun and knife murder rates (per 100,000 state residents) are shown. Using an alpha level of .01, test for a negative correlation between the variables. Use all five steps. If appropriate, interpret the sign, magnitude, and coefficient of determination. 395 16. Does it take defendants who face multiple charges longer to get through the adjudication process? The following table contains data on a random sample of juveniles from the JDCC data set. The variables are number of charges and months to adjudication; the latter variable measures the total amount of time that it took for these juveniles to have their cases disposed of. Using an alpha level of .05, test for a positive correlation between the variables. Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 17. Does a juvenile’s age correlate with how long it takes for the case to reach disposition? The following table contains a sample of female juveniles from the JDCC. Using an alpha level of .05, test for a correlation between the variables (no direction specified). Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 396 18. Is there a relationship between age and attitudes about crime and punishment? The General Social Survey (GSS) asks respondents whether they oppose or favor the death penalty and whether they think courts are too harsh, about right, or not harsh enough on offenders. These two variables can be summed to form a scale measuring people’s general preference for harsher or more-lenient penalties (higher values represent harsher attitudes). The data are displayed in the table. Using an alpha level of .01, test for a correlation between the variables (no direction specified). Use all five steps. If appropriate, interpret the sign, the magnitude, and the coefficient of determination. 19. Is there a relationship between age and the types of experiences people have with police? The file PPCS for Chapter 13.sav contains data from the PPCS. The sample is narrowed to male respondents and contains the variables age (respondents’ ages in years) and length (the number of minutes in a traffic stop, for those respondents who had been stopped in the past year). Run a correlation analysis. 1. Identify the obtained value of the correlation coefficient r. 2. State whether you would reject the null hypothesis at an alpha of .05 and how you reached that decision. 3. State the substantive conclusion. 4. As appropriate, interpret the sign, the magnitude, and the coefficient of determination. 20. Let’s revisit the variable measuring attitudes toward criminal punishment. The file GSS for Chapter 13.sav (www.sagepub.com/gau) contains data from the GSS, narrowed down to female respondents. The variable severity measures how harshly respondents feel toward persons convicted of crimes (higher values indicate greater preferences for severe punishment). The other variables are children, age , and education. We want to find out whether there are any statistically significant correlations among these variables. Run a correlation analysis using all four variables. 1. Identify the obtained values of the correlation coefficient r for each test. 2. For each of the r values, state whether you would reject the null hypothesis at an alpha of .05 and how you reached 397 http://www.sagepub.com/gau that decision. 3. State the substantive conclusion for each test. 4. As appropriate, interpret the sign, the magnitude, and the coefficient of determination for each test. 398 Key Terms Pearson’s correlation 312 Positive correlation 312 Negative correlation 312 Linear relationship 312 r coefficient 312 Glossary of Symbols and Abbreviations Introduced in This Chapter 399 Chapter 14 Introduction to Regression Analysis 400 Learning Objectives Explain the benefits of regression’s ability to predict values on the dependent variable (y) and explain how regression differs from the tests covered in previous chapters. Explain the difference between bivariate and multiple regression, including the importance of controlling for additional independent variables. Calculate each element of an ordinary least squares regression model, and compile the elements into an equation predicting y. Use an ordinary least squares regression equation to find the predicted value of y given a certain value of x. Read and interpret SPSS regression output, including identifying each key element of the equation and forming a conclusion about the statistical significance and substantive strength of relationships. In Chapter 13, you learned about correlation analysis. Correlation is a statistical procedure for finding two pieces of information: (1) whether two continuous variables are statistically related in a linear fashion and (2) the strength of that relationship. In most criminal justice and criminology research, however, merely finding out whether two variables are correlated is not sufficient. Researchers want to find out whether one of the variables (the independent variable [IV]) can be used to predict the other one (the dependent variable [DV]). Regression analysis does this. Regression goes a step beyond correlation by allowing researchers to determine how well the IV predicts the DV. Regression analysis: A technique for modeling linear relationships between one or more independent variables and one dependent variable wherein each independent variable is evaluated on the basis of its ability to accurately predict the values of the dependent variable. This chapter will discuss two types of regression analyses. Bivariate regression employs one IV and one DV. It is similar to bivariate correlation in many respects. Multiple regression uses several IVs to predict the DV. The specific type of regression modeling discussed here is ordinary least squares (OLS) regression. This is a fundamental form of regression modeling that is used frequently in criminology and criminal justice research. This type of regression produces a line of best fit that is as close to all the data points as it can possibly get. In other words, it minimizes the errors in the prediction of the DV. There are other types of regression, but OLS is the default procedure that is generally employed unless there is good reason to depart from it and use a different technique instead. In OLS, the DV must be continuous and normally distributed. The IVs can be of any level of measurement. Bivariate regression: A regression analysis that uses one independent and one dependent variable. Multiple regression: A regression analysis that uses two or more independent variables and one dependent variable. Ordinary least squares (OLS) regression: A common procedure for estimating regression equations that minimizes the errors in predicting the dependent variable. 401 One Independent Variable and One Dependent Variable: Bivariate Regression Bivariate regression is an extension of bivariate correlations, so we will begin the discussion of OLS using some of the data from Chapter 13. These data were from a sample drawn from the Juvenile Defendants in Criminal Courts (JDCC; see Data Sources 11.1). They measure the number of charges filed against each defendant, and the dollar amount of the fine imposed. (This sample has been narrowed to defendants who were convicted and penalized with fines.) The data from Table 13.4 are reproduced in Table 14.1. They are sorted in ascending order first according to the charges variable (the IV). Sorting them this way prepares the data for easy plotting on a graph. Figure 14.1 Scatterplot of Charges and Fine Amount 402 Learning Check 14.1 Can you create a scatterplot using a set of data points? Try plotting the data on age and police contacts from Table 13.1 in the previous chapter. Every x score in Table 14.1 has a corresponding y score. These scores form the graphing coordinates (x, y) . These coordinates can be graphed on a scatterplot like that in Figure 14.1. You can find each juvenile on the graph according to that person’s x and y scores. We already know that these two variables are correlated. This means we know they relate to one another insofar as they increase in tandem. Now we want to find out how well charges predict fines. In other words, we are looking to estimate what someone’s score on the DV would be, given that person’s score on the IV. In fact, we want to be able to predict someone’s score on the DV even for values not in the data set. For example, what fine would we expect a juvenile defendant with six charges to receive? The value “6” does not occur in the data in Table 14.1, but we can predict it using regression. Correlation does not allow us to do this. Think about drawing a line in the scatterplot in Figure 14.1. You want the line to come as close as possible to as many of data points as possible. What would this line look like? Would it have a positive or a negative slope? Where would it cross the y (vertical) axis? How steep would it be? Answering these questions will inform us as to how well x predicts y. The line of best fit is the one that comes closer to each of the data points than would any other line that could be drawn. The following formula is that for a line. You might recognize it from prior algebra classes. where ŷ = the predicted value of y at x , a = the y intercept, b = the slope coefficient, and x = the raw values of the IV. The predicted values of y —symbolized ŷ (pronounced “y hat”)—are calculated based on the intercept (a) and a certain slope coefficient (b) that is constant across all values of x. The intercept is the point at which the line of best fit crosses the y axis. The intercept is the value of y when x is zero. The slope coefficient conveys information about the steepness of that line. For instance, in a random sample of men, suppose the mean number of arrests is 1.2. If you randomly selected one of the men and wanted to make your best guess as to how many times he had been arrested, your guess would be 1.2. This is the estimate that will most minimize the number of times you are wrong. Now suppose you know each man’s age. In other words, referencing Formula 14(1), you now have more than a : you have x , too. The question becomes whether knowing x significantly improves your ability to predict any given person’s number of arrests, above and beyond the basic 403 predictive power of the mean. This is what the slope coefficient represents. Intercept: The point at which the regression line crosses the y axis; also the value of y when x = 0. Slope: The steepness of the regression line and a measure of the change in the dependent variable produced by a one-unit increase in an independent variable. Once the line given by Formula 14(1) is estimated, you have two different sets of DV scores: the empirical or observed scores (these are the y values), and the predicted values (the ŷ scores). If the IV is a good predictor of the DV, these two sets of numbers will be very similar to one another. If the IV does not predict the DV very well, the predicted values will differ substantially from the empirical values. The difference between a predicted value and an empirical value is called a residual (also called an error term). Small residuals indicate that x is a useful predictor of y , whereas large residuals suggest that x does not add much to our understanding of y. Knowing men’s ages, for example, will improve your ability to predict the number of times they have been arrested, because age is closely linked with criminal offending. Knowing their hair color, on the other hand, will not help predict arrests. The variable age will shrink the residuals, but hair color will not. Residual: The difference between a predicted value and an empirical value on a dependent variable. Ordinary least squares regression provides the formula that minimizes the residuals. If you were to add up all the error terms in a data set after running an OLS model, the total amount of error will be smaller than that you could get using any other regression approach. For this reason, OLS is described with the acronym BLUE: It is the best linear unbiased estimator. To construct the regression line, b and a must be calculated. In OLS, the slope coefficient b is calculated as 404 Learning Check 14.2 If you feel a sense of déjà vu when you look at Formula 14(2), that is good! This formula is very similar to the formula used to calculate the correlation coefficient r. Go back briefly and revisit Formula 13(2) to compare. What elements in the two formulas are different? Once b is known, a can be calculated from the following formula: where ȳ = the mean of the DV, b = the slope coefficient, and = the mean of the IV. Now we can construct the regression equation for our data. The first step is to calculate the slope coefficient b. In Table 13.4 in Chapter 13, we computed the sums required by Formula 14(2), so these numbers can be pulled from that table and entered into the formula: = 134.88 The slope b is 134.88, meaning that for every one-unit increase in the number of charges against a defendant, the expected fine rises by $134.88. This is helpful information, but we cannot yet generate predicted values because we do not have a starting point to build from. Finding a will do that. We first need the DV and IV means. These are found using the mean formula with which you are familiar: Plugging the means and b into Formula 14(3) yields a = 497.50 −134.88(2.75) = 497.50 – 370.92 405 = 126.58 Now the entire regression equation can be constructed using the pieces we just computed: Ŷ = 126.58 + 134.88x This equation tells us that the regression line crosses the y -axis at 126.58 and that every one-unit increase in x produces a 134.88-unit increase in y. The equation can be used to predict people’s fines on the basis of the number of charges against them. This is accomplished by entering a given value for x and solving the equation. For a person with two charges, the predicted fine is ŷ = 126.58 + 134.88x = 126.58 + 134.88(2) = 396.34 Looking back at Table 14.1, the person with two charges received a fine of $500. Our predicted value ( is roughly $100 off. The residual is 500 – 396.34 = 103.66. Even though we found a strong correlation between this IV and DV in Chapter 13 (recall that r = .74), it is clear from the regression residual that there is room for improvement. 406 Learning Check 14.3 Try calculating the predicted values of y for the remaining values in the data set. Compare each predicted value to its corresponding empirical value by calculating residuals. How close are they? What additional variables do you think would reduce the residuals? Let’s try using a value that is not in the data set. Nobody in Table 14.1 had six charges, but we can still use the regression equation to estimate what fine a person with six charges would receive. Plugging in this value of x yields ŷ = 126.58 + 134.88(6) = 935.86 This regression equation predicts that a juvenile defendant facing six charges will receive a fine of $935.86 (if he or she is penalized with a fine). We cannot compute the residual for this value because we do not have an observed value of y against which to compare the predicted value. 407 Inferential Regression Analysis: Testing for the Significance of b The most common use of regression analysis in criminal justice and criminology research is in the context of hypothesis testing. Just like the correlation coefficient r , the slope coefficient b does not itself determine whether the null hypothesis should be rejected. The slope coefficient b is also unstandardized. This means that it is presented in the DVs’ original units. This makes it impossible to figure out whether b is “large” or “small.” The measurement of the DV can have a dramatic impact on the slope coefficient. For instance, if a researcher trying to predict the length of prison sentences imposed on people convicted of burglary measures the DV (“sentence length”) in days and finds that gender (male or female) has a slope of b = 10, then the average difference between male and female offenders is only 10 days. By comparison, if sentence length is measured in months, and the researcher finds b = 5, then even though this slope is half the size of the previous one, the real-world meaning is substantial. A slope of b = 5 means there is a 5-month difference between men and women. This is a big discrepancy! We will first discuss the procedure for determining whether b is statistically significant. If it is not, there is no point converting it to a standardized metric. If it is, then a researcher would take the next step of standardizing it. To figure out whether b is significant, a five-step hypothesis test must be conducted. We will conduct this test on the b from above using ⍺ = .05. Step 1. State the null (H0) and alternative (H1) hypotheses. The null hypothesis in regression is generally that there is no relationship between the IV and the DV and, therefore, that the slope coefficient is zero. The alternative hypothesis is usually two-tailed; one-tailed tests are used only when there is a compelling reason to do so. Here, we will use a two-tailed test because this is the more customary course of action. Two-tailed tests are more conservative than one-tailed tests because they make it harder to reject the null, thereby reducing the risk of a Type I error (refer back to Table 9.1 and the accompanying discussion if you need a refresher). The null and alternative hypotheses are phrased in terms of the population parameters. The default assumption (i.e., the null) is always that B = 0. What we are looking for in an inferential test is evidence to lead us to believe that b is insignificantly different from zero. In regression, B symbolizes the population slope coefficient. The hypotheses are H0: B = 0 H1: B ≠ 0 Step 2. Identify the distribution and compute the degrees of freedom. The t distribution is the one typically used in regression. When the sample size is large, z can be used instead; however, since t can accommodate any sample size, it is more efficient to simply use that distribution in all circumstances. In bivariate regression, the degrees of freedom (df) are calculated as 408 Here, df = 8 – 2 = 6 Step 3. Identify the critical value and state the decision rule. With a two-tailed test, ⍺ = .05, and df = 6, tcrit = ±2.447. The decision rule states that if tobt is either < –2.447 or > 2.447 , H0 will be rejected.

Step 4. Compute the obtained value of the test statistic.

The first portion of this step entails calculating b and a in order to construct the regression equation in

Formula 14(1). We have already done this; recall that the regression equation is ŷ = 126.58 + 134.88x. Just

like all other statistics, b has a sampling distribution. See Figure 14.2. The distribution centers on zero because

the null predicts that the variables are not related. We need to find out whether b is either large enough or

small enough to lead us to believe that B is actually greater than or less than zero, respectively.

Finding out whether b is statistically significant is a two-step process. First, we compute this coefficient’s

standard error, symbolized SEb. The standard error is the standard deviation of the sampling distribution

depicted in Figure 14.2. The standard error is important because, all else being equal, slope coefficients with

larger standard errors are less trustworthy than those with smaller standard errors. A large standard error

means that there is substantial uncertainty as to the accuracy of the sample slope coefficient b as an estimate of

the population slope B.

Figure 14.2 The Sampling Distribution of Slope Coefficients

The standard error of the sampling distribution for a given regression coefficient (SEb) is computed as

409

where

sy = the standard deviation of y

sx = the standard deviation of x

r = the correlation between x and y

Recall that standard deviations are the mean deviation scores for a particular variable. The standard error of

the sampling distribution for a given slope coefficient is a function of the standard deviation of the DV, the

standard deviation of the IV in question, and the correlation between the two. This provides a measure of the

strength of the association between the two variables that simultaneously accounts for the amount of variance

in each. All else being equal, more variance (i.e., larger standard deviations) suggests less confidence in an

estimate. Large standard deviations produce a larger standard error, which in turn reduces the chances that a

slope coefficient will be found to be statistically significant.

Here, the standard deviation of x (the IV) is 1.75 and the standard deviation of y (the DV) is 317.89. The

correlation between these two variables is .74. Plugging these numbers into Formula 14(5) and solving

produces

= 181.65(.28)

= 50.86

This is the standard error of the slope coefficient’s sampling distribution. Now SEb can be entered into the tobt

formula, which is

The obtained value of t is the ratio between the slope coefficient and its standard error. Entering our numbers

into the equation results in

Step 4 is complete! The obtained value of t is 2.65. We can now make a decision about the statistical

410

significance of b.

Step 5. Make a decision about the null and state the substantive conclusion.

The decision rule stated that the null would be rejected if tobt turned out to be either <–2.447 or >2.447. Since

2.65 is greater than 2.447, the null is rejected. The slope is statistically significant at an alpha of .05. There is a

positive relationship between the extent to which a prison is over capacity and the number of major

disturbances that occur in that institution. In other words, knowing how many inmates a prison has beyond its

rated capacity helps predict the number of major disturbances that a prison will experience in a year.

As with correlation, rejecting the null requires further examination of the IV–DV relationship to determine

the strength and quality of that connection. In the context of regression, a rejected null indicates that the IV

exerts some level of predictive power over the DV; however, it is desirable to know the magnitude of this

predictive capability. The following section describes two techniques for making this assessment.

411

Beyond Statistical Significance: How Well Does the Independent Variable

Perform as a Predictor of the Dependent Variable?

There are two ways to assess model quality. The first is to create a standardized slope coefficient or beta

weight (symbolized β , the Greek letter beta) so the slope coefficient’s magnitude can be gauged. The second

is to examine the coefficient of determination. Each will be discussed in turn. Remember that these

techniques should be used only when the null hypothesis has been rejected: If the null is retained, the analysis

stops because the conclusion is that there is no relationship between the IVs and the DVs.

Beta weight: A standardized slope coefficient that ranges from –1.00 to +1.00 and can be interpreted similarly to a correlation so that

the magnitude of an IV–DV relationship can be assessed.

412

Standardized Slope Coefficients: Beta Weights

As noted earlier, the slope coefficient b is unstandardized, which means that it is specific to the units in which

the IV and DV are measured. There is no way to “eyeball” an unstandardized slope coefficient and assess its

strength because there are no boundaries or benchmarks that can be used with unstandardized statistics—they

are specific to whatever metric the DV is measured in. The way to solve this is to standardize b. Beta weights

range between 0.00 and ±1.00 and, like correlation coefficients, rely more on guidelines than rules for

interpretation of their strength. Generally speaking, betas between 0 and ±.19 are considered weak, from

about ±.20 to ±.29 are moderate, ±.30 to ±.39 are strong, and anything beyond ±.40 is very strong. These

ranges can vary by topic, though; subject-matter experts must decide whether a beta weight is weak or strong

within the customs of their fields of study.

Standardization is accomplished as follows:

We saw in the calculation of SEb that the The standard deviation of x is 1.75 and the standard deviation of y

is 317.89. We already computed b and know that it is 134.88. Plugging these numbers into Formula 14(7), we

get

In this calculation, rounding would have thrown the final answer off the mark, so the division and

multiplication were completed in a single step. The beta weight is .74. If this number seems familiar, it is! The

correlation between these two variables is .74. Beta weights will equal correlations (within rounding error) in

the bivariate context and can be interpreted the same way. A beta of .74 is very strong.

413

Learning Check 14.4

You just learned that standardized beta weights are equal to regression coefficients in bivariate regression models. As we will see soon,

however, this does not hold true when there is more than one IV. Why do you think this is? If you are not sure of the answer now,

continue reading and then come back to this question.

414

The Quality of Prediction: The Coefficient of Determination

Beta weights help assess the magnitude of the relationship between an IV and a DV, but they do not provide

information about how well the IV performs at predicting the DV. This is a substantial limitation because

prediction is the heart of regression—it is the reason researchers use this technique. The coefficient of

determination addresses the issue of the quality of prediction. It does this by comparing the actual, empirical

values of y to the predicted values (ŷi ). A close match between these two sets of scores indicates that x does a

good job predicting y , whereas a poor correspondence signals that x is not a useful predictor. The coefficient

of determination is given by

where = the correlation between the actual and predicted values of y.

The correlation between the y and ŷ and values is computed the same way that correlations between IVs and

DVs are and so will not be shown here. In real life, SPSS generates this value for you. The correlation in this

example is .74. This makes the coefficient of determination

.742 = .55

This means that 55% of the variance in y can be attributed to the influence of x. In the context of the present

example, 55% of the variance in fine amounts is attributable to the number of charges. Again, this value looks

familiar—it is the same as the coefficient of determination in Chapter 13! This illustrates the close connection

between correlation and regression at the bivariate level. Things get more complicated when additional IVs

are added to the model, as we will see next.

415

Adding More Independent Variables: Multiple Regression

The problem with bivariate regression—indeed, with all bivariate hypothesis tests—is that social phenomena

are usually the product of many factors, not just one. There is not just one single reason why a person commits

a crime, a police officer uses excessive force, or a prison experiences a riot or other major disturbance. Bivariate

analyses risk overlooking variables that might be important predictors of the DV. For instance, in the bivariate

context, we could test for whether having a parent incarcerated increases an individual’s propensity for crime

commission. This is probably a significant factor, but it is certainly not the only one. We can add other

factors, such as having experienced violence as a child, suffering from a substance-abuse disorder, and being

unemployed, too. Each of these IVs might help improve our ability to understand (i.e., predict) a person’s

involvement in crime. The use of only one IV virtually guarantees that important predictors have been

erroneously excluded and that the results of the analysis are therefore suspect, and it prevents us from

conducting comprehensive, in-depth examinations of social phenomena.

Multiple regression is the answer to this problem. Multiple regression is an extension of bivariate regression

and takes the form

Revisit Formula 14(1) and compare it to Formula 14(9) to see how 14(9) expands on the original equation by

including multiple IVs instead of just one. The subscripts show that each IV has its own slope coefficient.

With k IVs in a given study, ŷ is the sum of each bkxk term and the intercept.

In multiple regression, the relationship between each IV and the DV is assessed while controlling for the

effect of the other IV or IVs. The slope coefficients in multiple regression are called partial slope coefficients

because, for each one, the relationship between the other IVs and the DV has been removed so that each

partial slope represents the “pure” relationship between an IV and the DV. Each partial slope coefficient is

calculated while holding all other variables in the model at their means, so researchers can see how the DV

would change with a one-unit increase in the IV of interest, while holding all other variables constant. The

ability to incorporate multiple predictors and to assess each one’s unique contribution to is what makes

multiple regression so useful.

Partial slope coefficient: A slope coefficient that measures the individual impact of an independent variable on a dependent variable

while holding other independent variables constant.

It is interesting and surprising that intelligence outweighed parenting in predicting children’s self-control.

Intelligence was, in fact, by far the strongest predictor of low self-control: More-intelligent children had more

self-control relative to their peers who scored lower on intelligence tests. Paternal low self-control significantly

predicted children’s low self-control, but the beta was very small. The only other significant variable is sex,

with boys displaying higher levels of low self-control compared to girls. The model R2 = .225, meaning that

the entire set of predictors explained 22.5% of the variance in children’s self-control. Clearly, childhood

intelligence is integral in the development of self-control and, ultimately, in the prevention of delinquency and

416

crime.

Research Example 14.1 Does Childhood Intelligence Predict the Emergence of Self-Control?

Theory suggests—and research has confirmed—that low self-control is significantly related to delinquency and crime. People with

low self-control tend to be impulsive and to have trouble delaying gratification and considering possible long-term consequences of

their behavior. Self-control is said to be learned during the formative years of a child’s life. Parenting is critical to the development of

self-control; parents who provide clear rules and consistent, fair punishment help instill self-discipline in their children. But what

about children’s innate characteristics, such as their intelligence level? Petkovsek and Boutwell (2014) set out to test whether

children’s intelligence significantly affected their development of self-control, net of parenting, and other environmental factors.

They ran OLS regression models and found the following results (note that SE = standard error).

Source: Adapted from Table 2 in Petkovsek and Boutwell (2014).

p < .01;
p < .001.
Before getting into more-complex examples, let us work briefly with a hypothetical regression equation
containing two IVs, x1 and x2. Suppose the line is
ŷ = 1.00 + .80x1 + 1.50x2
We can substitute various values for x1 and x2 to find . Let’s find the predicted value of the DV when x1 = 4
and x2 = 2:
ŷ =1.00 + .80(4) +1.50(2) =1.00 + 3.20 + 3.00 = 7.20.
There it is! If x1 = 4 and x2 = 2, the DV is predicted to be 7.20.
417
418
Learning Check 14.5
Use the equation ŷ = 1.00 + .80x1 + 1.50x2 to find the predicted value of y when . ..
1. x1 = 2 and x2 = 3.
2. x1 = 1.50 and x2 = 3.
3. x1 = .86 and x2 = –.67.
4. x1 = 12 and x2 = 20.
The formulas involved in multiple regression are complex and are rarely used in the typical criminal justice
and criminology research setting because of the prevalence of statistical software. We now turn to a discussion
of the use of SPSS to obtain and interpret OLS regression output.
419
Ordinary Least Squares Regression in SPSS
As described earlier, researchers rarely fit regression models by hand. Data sets are typically far too large for
this, and the prevalence of user-friendly software programs like SPSS put impressive computing power right
at researchers’ fingertips. Of course, the flipside of this wide availability of user-friendly interfaces is the
potential for them to be used carelessly or incorrectly. People producing research must possess a solid
comprehension of the theory and math underlying statistical techniques before they attempt to run analyses in
SPSS or other programs. Consumers of research (e.g., police and corrections officials) need to have enough
knowledge about statistics to be able to evaluate results, including spotting mistakes when they occur.
Consumers can be led astray if they fail to critically examine statistics and if they do not know when to trust
empirical findings and when not to. As with the techniques discussed in previous chapters, GIGO applies to
regression modeling. Statistical programs will frequently run and produce results even when errors have been
made. For instance, SPSS will run an OLS model when the dependent variable is nominal. The results of this
test are meaningless and useless, so it is up to producers and consumers to be smart and avoid making these
mistakes and being deceived by them if they do occur.
Before discussing the analytical element of running OLS models in SPSS, we should revisit the null and
alternative hypotheses. In multiple regression, the null and alternative each apply to every IV. For each IV, the
null predicts that the population slope coefficient Bk is zero, and the alternative predicts that it is significantly
different from zero. Since the analysis in the current example has two IVs, the null and alternative are
H0: B1 = 0 and B2 = 0
H1: B1 and/or B2 ≠ 0
Since each IV has its own null, it is possible for the null to be rejected for one of the variables and not for the
other.
To run a regression analysis in SPSS, go to Analyze → Regression → Linear. This will produce the dialog box
shown in Figure 14.3. Here, the Police–Public Contact Survey (PPCS; see Data Sources 2.1) is being used.
The DV is the length of time a vehicle or pedestrian stop lasted. The IVs are characteristics of the
respondents. Respondents’ sex, age, and race are included. Age is a continuous variable (measured in years),
and sex and race are nominal-level variables each coded as a dummy variable such that one category is 0 and
the other is 1. In this example, 1 = male and 0 = female for the gender variable, and 1 = white and 0 =
nonwhite for the race variable. Move the DV and IVs into their proper locations in the right-hand spaces, and
then press OK. This will produce an output window containing the elements displayed in the following
figures.
Dummy variable: A two-category, nominal variable with one class coded as 0 and the other coded as 1.
The first portion of regression output you should look at is the analysis of variance (ANOVA) box. This
420
might sound odd since we are running a multiple regression analysis, not an ANOVA, but what this box tells
you is whether the set of IVs included in the model explains a statistically significant amount of the variance
in the DV. If F is not significant (meaning if p > .05), then the model is no good. In the event of a

nonsignificant F , you should not go on to interpret and assess the remainder of the model. Your analysis is

over at that point and what you must do is revisit your data, your hypothesis, or both to find out what went

wrong. The problem might be conceptual rather than statistical—the IVs you predicted would impact the DV

might not actually do so. There could be an error in your choice of variables to represent theoretical concepts,

or there could be a deeper flaw affecting the theory itself. Before you consider possible conceptual issues,

check the data to make sure the problem is not caused by a simple coding error.

Figure 14.3 Running a Multiple Regression Analysis in SPSS

Figure 14.4 SPSS Regression Output

421

In Figure 14.4, you can see that F = 11.654 and p = .000, so the amount of the variance in the DV variance

that is explained by the IVs is significantly greater than zero. Note that a significant F is not by itself proof

that the model is good—this is a necessary but insufficient condition for a high-quality regression model.

Second, look at the “R Square column” in the “Model Summary” box. This is the multiple coefficient of

determination and indicates the proportion of the variance in the DV that is explained by all the IVs

combined. It is an indication of the overall explanatory power of the model. There are no specific rules for

evaluating R square. Generally, values up to .20 are considered fairly low, .21 to .30 are moderate, .31 to .40

are good, and anything beyond .41 is very good. Here, R square is .008, meaning the IVs that we selected

explain a trivial .8% of the variance in stop length. There are definitely important variables omitted from this

model.

Third, go to the “Coefficients” box at the bottom of the output to see the unstandardized b values,

standardized beta weights, and significance test results. The “Unstandardized Coefficients: B” column

contains the slope for each variable (the constant is the intercept). The IVs age and sex are statistically

significant, but race is not. We know this because the “Sig.” values (i.e., p values) for sex and age are both less

than .05, but the p value for race is much greater than .05.

Since age and sex are statistically significant, we can interpret their meaning to the model. The

unstandardized slope for age is b = –.045. This means that each one-year increase in a person’s age is

associated with a reduction of .045 minutes in the total length of the stop. Older people’s stops are shorter, on

average, than younger people’s. This makes sense, because younger people are more active in deviant behavior

than older people are. Police officers probably take more time with younger drivers and pedestrians they stop

to make sure they are not engaged in illegal activity.

Since sex is a dummy variable, the interpretation of b is a bit different from the interpretation of the slope of a

continuous predictor. Dummy variables’ slopes are comparisons between the two categories. Here, since

female is coded 0 and male is coded 1, the slope coefficient b = .948 indicates that males’ stops last an average

of .948 minutes longer than females’ stops. This finding, like that for age, makes sense in light of the fact that

males commit more crime than females do. Police might subject males to more scrutiny, which extends the

length of the stop.

The Beta column shows the standardized values. The utility of beta weights over b values is the ability to

compare the relative strength of each IV. Using the unstandardized slopes results in an apples-to-oranges type

of comparison. We are left not knowing whether age or gender is a stronger predictor of stop length. Beta

weights answer this question. You can see that β = .049 for sex and β = –.075 for age. Since –.075 represents a

stronger relationship than .049 does (it is the absolute value we are examining here), we can conclude that age

is the more impactful of the two. Still, –.075 is very small.

Research Example 14.2 Does Having a Close Black Friend Reduce Whites’ Concerns About Crime?

Mears, Mancini, and Stewart (2009) sought to uncover whether whites’ concerns about crime as a local and as a national problem

were affected by whether or not those whites had at least one close friend who was black. Concern about crime was the DV in this

422

study. White respondents expressed their attitudes about crime on a 4-point scale where higher values indicated greater concern.

The researchers ran an OLS regression model and arrived at the following results with respect to whites’ concerns about local crime.

The authors found, contrary to what they had hypothesized, that having a close friend who was black actually increased whites’

concerns about crime. You can see this in the fact that the slope coefficient for have a black friend is statistically significant (p < .05)
and positive. Age was also related to concern, with older respondents expressing more worry about local crime. Income had a
negative slope coefficient such that a one-unit increase in annual income was associated with a .06 reduction in concern. Finally,
living in an urban area substantially heightened whites’ worry about crime—looking at the beta weights, you can see that urban is the
strongest IV in the model. The researchers concluded that for whites living in urban areas, having black friends might make the
crime problem seem more real and immediate because they are exposed to it vicariously through these friends.
Source: Adapted from Table 2 in Mears et al. (2009).
p < .05;
p < .01;
p < .001.
For another example, we will turn to the JDCC analysis predicting the dollar amount of fine imposed on
juveniles adjudicated delinquent and ordered to pay a fine. We calculated the bivariate equation using
“number of charges” as the IV and a small sample of eight juveniles. What happens when we use the full data
set and add relevant case characteristics and information about the juvenile defendants? Figure 14.5 shows the
OLS output. The variables measuring the number of charges and juveniles’ ages at the time of arrest are
continuous, and the rest are dummy coded (i.e., 0 and 1). The output shows the coding, so you can see which
categories were coded as 0 and which as 1.
The model F is 9.126 and statistically significant (p < .001). The R square is .101, suggesting the six IVs
collectively explain 10% of the variance in fine amount. This is a small percentage and indicates that
important variables have been omitted from the model. Turning to the slopes, the variable measuring the total
number of charges remains statistically significant (p < .001) and is the strongest predictor in the model (β =
.239). More charges mean a higher fine. The second-largest predictor is a juvenile’s history of prior arrests or
convictions (measured as 0 = no priors and 1 = priors). This slope is statistically significant (p < .01) and
moderate in strength (β = –.120). Interestingly, this slope is negative, meaning that juveniles with prior
423
records receive smaller fines. This seems backward at first glance but makes sense if we consider that repeat
offenders are sentenced to harsher penalties. Juveniles with priors might get smaller fines because they receive
other sentences along with those fines (such as probation), whereas first-time offenders are sentenced to only
fines and therefore those fines are larger. The final IV that reaches statistical significance at .05 is whether the
charge for which a juvenile was adjudicated delinquent was violent. The β of .092 suggests this is a small
effect; fines are probably not used as the primary punishment for juveniles adjudicated delinquent on violent
charges, so this small effect is not surprising. The race variable fell just shy of the customary .05 alpha level. At
p = .054, though, most researchers would probably consider it statistically significant since it missed the cut off
by only .004. It appears that white juveniles receive larger fines than nonwhite juveniles, but at β = .084, this
effect is small.
Figure 14.5 SPSS Output
The regression equation can be constructed from the output in Figure 14.5 and used to calculate predicted
values. The equation is
ŷ = −1101.82 + 155.17xcharges −448.26xpriors + 187.74xviolent + 85.47xage – 757.08xsex + 188.58xrace
424
Let’s compute the predicted value of y (i.e., the predicted fine amount) for a white male juvenile with three
charges, no prior record, who was not adjudicated of a violent crime, and was 15 at the time of arrest.
Plugging these numbers in,
ŷ = −1,101.82 +155.17(3) − 448.26(0) +187.74(0) + 85.47(15) −757.08(0) +188.58(1)
= –1,101.82 + 465.51 + 1,282.05 + 188.58
= 834.32
We would expect a juvenile with these characteristics to receive a fine of $834.32. Try replicating this analysis,
then visit Learning Check 14.6 for additional examples you can use for practice.
425
Learning Check 14.6
Use the equation ŷ = −1101.82 + 155.17xcharges −448.26xpriors + 187.74xviolent + 85.47xage – 757.08xsex + 188.58xrace to calculate the
predicted value of y for each of the following juveniles:
1. Nonwhite female age 16 with no priors who had three charges and was adjudicated guilty of a violent offense.
2. White male age 17 with priors who had two charges and was adjudicated guilty of a nonviolent offense.
3. Nonwwhite male age 15 with no priors who had one charge and was adjudicated guilty of a nonviolent offense.
4. White male age 15 with priors who had one charge and was adjudicated guilty of a violent offense.
Research Example 14.3 Do Multiple Homicide Offenders Specialize in Killing?
In Chapter 11, you read about a study by Wright, Pratt, and DeLisi (2008) wherein the researchers examined whether multiple
homicide offenders (MHOs) differed significantly from single homicide offenders (SHOs) in terms of the diversity of offending.
Diversity was measured as a continuous variable with higher values indicating a greater spectrum of offending. We saw in Chapter
11 that Wright et al. first ran a t test to check for differences between the group means for MHOs and SHOs; that test showed that
the difference was not statistically significant. Given that bivariate results are untrustworthy for the reasons discussed in this chapter,
Wright et al. ran a multiple OLS regression model. They found the following results.
Source: Adapted from Table 2 in Wright et al. (2008).
p < .05; ** p < .01.
Offenders’ current age, the age at which they started offending, and their race were statistically significant predictors of offending
diversity. As foreshadowed by the nonsignificant t test, the coefficient for Offender type: SHO was not statistically significant. A
dichotomous IV like that used here, where people in the sample were divided into two groups classified as either MHOs or SHOs , is
a dummy variable. The slope is interpreted just as it is with a continuous IV: It is the amount of predicted change in the DV that
occurs with a one-unit increase in the IV. Here, you can see that being an SHO increased offending diversity by only .004 of a unit.
This is a very trivial change and was not statistically significant. Race is also a dummy variable, because offenders in the sample were
classified as either nonwhite or white. Can you interpret this slope coefficient with respect to what it means about race as a predictor
of the diversity index? If you said that white offenders score significantly lower on the diversity index, you are correct.
Wright et al.’s (2008) multiple regression model confirmed that MHOs and SHOs do not differ in terms of offending diversity.
This suggests that MHOs do not specialize in killing; to the contrary, they display as much diversity as other types of homicide
offenders. The theoretical implication of this finding is that the theories that have been developed to help explain violent offending
might be applicable to MHOs because these people are similar to other offenders.
426
427
Alternatives to Ordinary Least Squares Regression
This chapter has focused on OLS regression because it is the most basic regression modeling strategy and is
generally the starting point for the study of regression. Recall that OLS can be used only when a DV is
continuous and (reasonably) normally distributed. As you might have already guessed, researchers often
confront DVs that do not meet one or both of these criteria. A DV might be nominal or ordinal, or it could
be heavily skewed. It might contain a substantial number of zeros, which throws the OLS calculations off.
Various regression techniques are available for situations when the DV violates the OLS assumptions. We will
not go into them in detail, but it is worth knowing the names of at least some of the most common models.
Each one is specific to a certain kind of DV.
When a DV is dichotomous (i.e., categorical with two classes), binary logistic regression is used. Binary
logistic regression calculates the probability of any given case in the data set falling into the category of
interest. The DV might be, for instance, whether a defendant was sentenced to prison. Binary logistic
regression would tell the user whether and to what extent each of the IVs (offense type, prior record, and so
on) increases the probability of a defendant being sent to prison. Additionally, a severely skewed DV can
sometimes be dichotomized (i.e., split in half) and put through binary logistic rather than OLS. Binary
logistic regression is very popular. There are two other types of logistic regression that are used for categorical
DVs. For a nominal DV with three or more classes, multinomial logistic is available. For an ordinal DV with
three or more classes, ordered logistic can be employed. Each of these types of models functions by
sequentially comparing pairs of classes.
A technique called Poisson regression is used when a DV is made of count data (sometimes also called event
data). Count data are phenomena such as the number of times the people in a sample have been arrested or
the number of homicides that occur in a city in a year. Throughout this book, we have treated count data as
being part of the continuous level of measurement, and this is often perfectly fine to do, but it can cause
problems if done so in the regression realm. Part of the problem is that count data are usually highly skewed,
because most of the DVs that are studied are relatively rare events (violent crimes, criminal convictions, and
the like). In a data set containing the number of homicides experienced by a sample of cities, the vast majority
of cities will cluster at the low end of the distribution, a few will be in the middle, and a small minority will
extend out in the tail. Additionally, OLS assumes that the DV can take on any value, including negatives and
decimals, but count data can only be positive, whole numbers (integers). For these reasons, OLS is
inappropriate for count data. The Poisson distribution is a probability distribution that works well for count
data. Poisson regression is, not surprisingly, based on the Poisson probability distribution. This type of
regression is common in criminal justice and criminology research, although most data violate the assumptions
for using the Poisson distribution and so researchers typically turn to a technique called negative binomial
regression. This type of regression works well with count data containing a lot of zeros.
There are also modeling strategies available for times when none of the previously discussed techniques is
adequate given the type of hypothesis being tested or the type of data a researcher is working with. A
428
technique called structural equation modeling (SEM) can be applied when a researcher has multiple DVs
rather than just one. A researcher might posit that one variable causes another and that this second variable, in
turn, causes a third one. Standard regression can accommodate only one DV, but SEM can handle a more
complex causal structure. Multilevel modeling (also called hierarchical linear modeling) is appropriate when
data are measured at two units of analysis (such as people nested within neighborhoods or prisoners nested
within institutions). Basic regression assumes that all data are of the same unit of analysis, and when that is
not the case, the standard-error estimates can be inaccurate, and the significance tests thrown off as a result.
Both structural equation modeling and multilevel modeling are based in regression.
Research Example 14.4 Is Police Academy Performance a Predictor of Effectiveness on the Job?
In Chapter 13, you encountered a study by Henson, Reyns, Klahm, and Frank (2010). The researchers sought to determine whether
recruits’ performance while at the academy significantly influenced their later success as police officers. Henson et al. measured
success in three ways: the scores new officers received on the annual evaluations conducted by those officers’ supervisors, the number
of complaints lodged against these new officers, and the number of commendations they earned. These three variables are the DVs
in this study.
You saw in Chapter 13 that the bivariate correlations indicated mixed support for the prediction that academy performance was
related to on-the-job performance; however, to fully assess this possible link, the researchers ran an OLS regression model. They
obtained the following results.
The results from the three OLS models showed that recruits’ civil service exam scores, physical agility exam scores, and overall
academy ratings were—with only one exception—unrelated to on-the-job performance. The exception was the positive slope
coefficient between overall academy ratings and evaluation scores (b = .06; p < .01). The demographic variables gender, age, race, and
education also bore limited and inconsistent relationships with the three performance measures. These results seem to indicate that
the types of information and training that recruits receive are not as clearly and directly related to on-the-job performance as would
be ideal. There might be a need for police agencies to revisit their academy procedures to ensure that recruits are receiving training
that is current, realistic, and practical in the context in which these recruits will be working once they are out on the street.
429
Source: Adapted from Table 5 in Henson et al. (2010).
p < .01.
This has been a brief overview of regression and regression-based alternatives to OLS. There are many
options available to accommodate hypotheses and data of all types. More information on these and other
techniques is available in advanced statistics textbooks and online.
Chapter Summary
This chapter introduced you to the basics of bivariate and multiple regression analysis. Bivariate regression is an extension of
bivariate correlation and is useful because correlation allows for only a determination of the association between two variables, but
bivariate regression permits an examination of how well the IV acts as a predictor of the DV. When an IV emerges as a statistically
significant predictor of the DV, it can then be assessed for magnitude using the standardized beta weight, β , and the coefficient of
determination, r2.
The problem with bivariate analyses of all types, though, is that every phenomenon that is studied in criminal justice and
criminology is the result of a combined influence of multiple factors. In bivariate analyses, it is almost certain that one or more
important IVs have been omitted. Multiple regression addresses this by allowing for the introduction of several IVs so that each one
can be examined while controlling for the others’ impacts. In the bivariate regression example conducted in this chapter, age of onset
significantly predicted lifetime arrests; however, when education was entered into the regression model, age lost its significance and
education emerged as a significant and strong predictor of lifetime arrests. This exemplifies the omitted variable bias: Failing to
consider the full gamut of relevant IVs can lead to erroneous results and conclusions.
This also brings us to the end of the book. You made it! You struggled at times, but you stuck with it, and now you have a solid
grasp on the fundamentals of the use of statistics in criminal justice and criminology research. You know how to calculate univariate
and bivariate statistics and how to conduct hypothesis tests. Just as important, you know how to evaluate the statistics and tests
conducted by other people. You know to be critical, ask questions, and always be humble in drawing conclusions because every
statistical test contains some level of error, be it the Type I or Type II error rate, omitted variables, or some other source of
430
uncertainty. Proceed with caution and a skeptical mind when approaching statistics as either a producer or a consumer. Make GIGO
a part of your life—when the information being input into the system is deficient, the conclusions are meaningless or possibly even
harmful. The bottom line: Question everything!
Thinking Critically
1. People commonly make the mistake of assuming that since multiple regression analyses control for many factors and help
rule out spuriousness, they are proof of causation (i.e., proof that the IV of interest causes the DV). Critically examine this
misunderstanding about regression. Use the criteria for establishing causation to explain the flaw in this assumption. Refer
back to Chapter 2 if you need to.
2. In criminal justice and criminology research (as in most other social sciences), regression models usually produce fairly
modest R2 values. Even explaining a relatively small fraction of the variance in a DV is frequently considered good, even
though an R2 value considered robust within social sciences (such as .40, for instance) leaves the majority of the variance
unaccounted for (in this case, 60%). Why do you think this is? That is, why do regression models commonly explain less than
half of the variance? (Hint: Think about the DVs that criminology and criminal justice researchers work with.)
Review Problems
1. You learned in this chapter that the key advantage of bivariate regression over correlation is that regression can be used for
prediction. Explain this. How is it that regression can be used to predict values not in the data set, but correlation cannot?
2. Identify the two criteria that a DV must meet for OLS regression to be used.
3. Does OLS regression place restrictions on the levels of measurement of the IVs?
4. Explain the advantage of multiple regression over bivariate regression. What does multiple regression do that bivariate
regression does not? Why is this important?
5. In a hypothetical example of five prison inmates, let us find out whether prior incarcerations influence in-prison behavior.
The following table contains data on the prior incarcerations and the number of disciplinary reports filed against each
inmate. Use the data to do the following:
1. Calculate the slope coefficient b.
2. Calculate the intercept a.
3. Write out the full regression equation.
4. Calculate the number of disciplinary reports you would expect to be received by a person with
1. 3 prior incarcerations.
2. 15 prior incarcerations.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether b is statistically significant. (sy = 2.59; and sx = 5.36; r = −.09)
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
6. Does the amount of crime in an area predict the level of police presence? The following table displays data on the number of
crimes per square mile and the number of police agencies per 1,000 square miles in a sample of states. Use the data to do the
following:
431
1. Calculate the slope coefficient b .
2. Calculate the intercept a .
3. Write out the full regression equation.
4. Calculate how many police agencies per 1,000 square miles you would expect in a state with
1. 5 crimes per square mile.
2. 10 crimes per square mile.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether the IV is a significant predictor of the DV. sy = 6.39; and sx = 2.76; r = −.71
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
7. Research has found that socioeconomic disadvantage is one of the strongest and most consistent predictors of crime.
Negative socioeconomic factors such as poverty and unemployment have been shown to profoundly impact crime rates. The
following table contains a random sample of states. The IV consists of data from the U.S. Census Bureau on the percentage
of adults in the civilian labor force that was unemployed. The DV is UCR-derived violent crime rates per 1,000 persons. Use
the data to do the following:
432
1. Calculate the slope coefficient b.
2. Calculate the intercept a.
3. Write out the full regression equation.
4. Calculate how many violent crimes per 1,000 citizens you would expect in a state with
1. a 4% unemployment rate.
2. an 8% unemployment rate.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether the IV is a significant predictor of the DV. sy = 1.76; and sx = 1.29; r = −.79
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
8. Deterrence theory suggests that as the number of crimes that police solve goes up, crime should decrease because would-be
offenders are scared by the belief that there is a good chance that they would be caught and punished if they commit an
offense. The following table contains regional data from the UCR. The IV is clearance and is the percentage of violent crimes
that were cleared by arrest or exceptional means in one year. The DV is violent crime rate and is the number of violent crimes
that occurred in the following year (rate per 1,000 residents). Use the data to do the following.
1. Calculate the slope coefficient b .
2. Calculate the intercept a.
3. Write out the full regression equation.
4. Calculate the rate of violent crimes per 1,000 citizens you would expect in a region where
1. 30% of violent crimes were cleared.
433
2. 50% of violent crimes were cleared.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether the IV is a significant predictor of the DV. sy = .88; and sx = 4.41; r = −.25
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
9. Let us now consider the possible relationship between poverty and crime. The following table contains a random sample of
states and the violent crime rate (the DV). The IV is the percentage of families living below the poverty line. Use the table to
do the following.
1. Calculate the slope coefficient b.
2. Calculate the intercept a .
3. Write out the full regression equation.
4. Calculate the rate of violent crimes per 1,000 citizens you would expect in a state with
1. a 5% poverty rate.
2. a 10% poverty rate.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether the IV is a significant predictor of the DV. sy = 1.87; and sx = 2.29; r = .78
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
10. Is there a relationship between the handgun murder rate and the rate of murders committed with knives? The table shows
data from a sample of states. Consider the knife murder rate to be the DV. Use the table to do the following.
1. Calculate the slope coefficient b.
2. Calculate the intercept a.
434
3. Write out the full regression equation.
4. Calculate the rate of knife murders per 1,000 citizens you would expect in a state with
1. a handgun murder rate of 3.00.
2. a handgun murder rate of 1.75.
5. Using an alpha level of .05 and two-tailed alternative hypothesis, conduct a five-step hypothesis test to determine
whether the IV is a significant predictor of the DV. sy = .15; sx = .70; r = −.68
6. If appropriate (i.e., if you rejected the null in Part e), calculate the beta weight.
11. The data set JDCC for Chapter 14.sav at www.sagepub.com/gau contains data from the JDCC survey. The sample has been
narrowed to those facing drug charges who were convicted and sentenced to probation. The DV (probation) is the number of
months to which convicted juveniles were sentenced. The IVs are the number of charges filed against each defendant
(charges) and whether the defendant had a prior record of juvenile arrests or convictions (priors) . Run a bivariate regression
using the DV probation and the IV charges to determine whether the number of charges significantly predicts the severity of
the probation sentence at the bivariate level. Do not include the other variable. Then do the following.
1. Report the value of the ANOVA F and determine whether you would reject the null at an alpha of .05. What does
the rejection or retention of F mean?
2. Report the R -square value. Using the guidelines provided in the text, interpret the strength of the explanatory power
of the IV.
3. Using the numbers in the output, write out the bivariate regression equation for ŷ.
4. Determine whether b is statistically significant at ⍺ = .05, and explain how you arrived at this decision.
5. If appropriate, identify the value of the beta weight.
6. What is your substantive conclusion? Do the number of charges predict probation sentences? Explain your answer.
12. Keeping the same data set and model from Question 11, add priors to the model to account for whether or not each
defendant had a prior record of arrests or convictions as a juvenile. Recall from the chapter that this is a binary (dummy)
variable with “yes” coded as 1 and “no” as zero, so what you will be looking for in the output is a comparison between those
with and without priors (i.e., those who had and had not been arrested or convicted before). The slope coefficient will tell
you the impact of having a prior record as opposed to not having one (the larger the slope, the stronger the impact).
1. Report the value of the ANOVA F and determine whether you would reject the null at an alpha of .05. What does
the rejection or retention of F mean?
2. Report the R -square value. Using the guidelines provided in the text, interpret the strength of the explanatory power
of the IVs.
3. Using the numbers in the output, write out the multiple regression equation for ŷ.
4. Determine whether each b is statistically significant at ⍺ = .05, and explain how you arrived at these decisions.
5. If appropriate, identify the value of the beta weight for each significant IV and compare them. Which one is
stronger? Weaker?
6. What is your substantive conclusion? Do age and/or number of charges seem to be strong predictors of probation
sentences? Explain your answer.
13. Using the multiple regression equation you wrote in Question 12, do the following:
1. Setting charges at its mean (2.09), calculate the predicted probation sentence for a juvenile who does not have a prior
record (i.e., xpriors = 0 in the equation).
2. Holding charges constant at its mean, calculate the predicted probation sentence for a juvenile who does have a prior
record (i.e., xpriors = 1).
3. By how much did the predicted sentence change? Was this change an increase or decrease?
14. Using the multiple regression equation you wrote in Question 12, do the following:
1. Setting priors at its mean (.79), calculate the predicted probation sentence for a juvenile facing two charges.
2. Holding age constant at its mean, calculate the predicted probation sentence for a juvenile facing five charges.
3. By how much did the predicted sentence change? Was this change an increase or a decrease?
15. The data file Socioeconomics and Violence for Chapter 14.sav at www.sagepub.com/gau contains a sample of states. Violent
crime rates (the variable violentrate) is the DV. The IVs are the percentage of households receiving food stamps
(Supplemental Nutrition Assistance Program, or SNAP; snap) ; the percentage of the adult civilian workforce that is
unemployed (unemployed) ; and the percentage of families that are below the poverty line (poverty) . Run a bivariate
regression using the DV violentrate and the IV unemployed to test for a relationship between unemployment and violent
crime. Do not include the other variables. Then do the following:
435
http://www.sagepub.com/gau
http://www.sagepub.com/gau
1. Report the value of the ANOVA F and determine whether you would reject the null at an alpha of .05. What does
the rejection or retention of F mean?
2. Report the R -square value. Using the guidelines provided in the text, interpret the strength of the explanatory power
of the IV.
3. Using the numbers in the output, write out the bivariate regression equation for ŷ.
4. Determine whether b is statistically significant at ⍺ = .05, and explain how you arrived at this decision.
5. If appropriate, identify the value of the beta weight.
6. What is your substantive conclusion? Does the unemployment rate seem to be a strong predictor of violent crime?
Explain your answer.
16. Using the data file Socioeconomics and Violence for Chapter 14. sav , run a multiple regression using all three IVs (snap,
unemployed , and poverty) . Then do the following:
1. Report the value of the ANOVA F and determine whether you would reject the null at an alpha of .05. What does
the rejection or retention of F mean?
2. Report the R -square value. Using the guidelines provided in the text, interpret the strength of the explanatory power
of the IVs.
3. Using the numbers in the output, write out the multiple regression equation for ŷ.
4. Determine whether each b is statistically significant at ⍺ = .05, and explain how you arrived at these decisions.
5. Identify the beta weights for each significant IV and compare them. Which one is strongest? Weakest?
6. What is your substantive conclusion? Do these variables seem to be strong predictors of violent crime rates? Explain
your answer.
17. Using the multiple regression equation you wrote in Question 16, find the predicted violent crime rate for a state where 15%
of households receive benefits, 10% of people are unemployed, and 16% of families live below the poverty line.
18. Using the multiple regression equation you wrote in Question 16, do the following:
1. Setting snap at its mean (8.25) and poverty at its mean (8.94), calculate the predicted violent crime rate for a state
with a 4% unemployment rate.
2. Holding snap and poverty constant at their means, calculate the predicted violent crime rate for a state with a 10%
unemployment rate.
3. By how much did the predicted rate change? Was this change an increase or decrease?
19. Using the multiple regression equation you wrote in Question 16, do the following:
1. Setting unemployment at its mean (6.77) and poverty at its mean (8.94), calculate the predicted violent crime rate for a
state where 11% of households receive benefits.
2. Holding unemployment and poverty constant at their means, calculate the predicted violent crime rate for a state where
6% of households receive benefits.
3. By how much did the predicted rate change? Was this change an increase or a decrease?
20. Using the multiple regression equation you wrote in Question 16, do the following:
1. Setting unemployment at its mean (6.77) and snap at its mean (8.25), calculate the predicted violent-crime rate for a
state where 15% of families live below the poverty line.
2. Holding unemployment and snap constant at their means, calculate the predicted violent-crime rate for a state where
20% of families live below the poverty line.
3. By how much did the predicted rate change? Was this change an increase or decrease?
436
Key Terms
Regression analysis 336
Bivariate regression 336
Multiple regression 336
Ordinary least squares (OLS) regression 336
Intercept 337
Slope 337
Residual 338
Beta weight 345
Partial slope coefficient 347
Dummy variable 350
Glossary of Symbols and Abbreviations Introduced in This Chapter
437
Appendix A. Review of Basic Mathematical Techniques
In order to succeed in this class, you must have a solid understanding of basic arithmetic and algebra. This
appendix is designed to help you review and brush up on your math skills.
438
Section 1: Division
You will be doing a lot of dividing throughout this book. The common division sign “÷” will not be used.
Division will always be presented in fraction format. Instead of “6 ÷ 3,” for example, you will see “ .” For
example,
Try the following as practice.
1.
2.
3.
4.
439
Section 2: Multiplication
Multiplication is another oft-used technique in this book. The common multiplication sign “×” will not be
used, because in statistics, the symbol “x ” represents a raw score in a data set. Employing the “×”
multiplication sign would therefore introduce confusion. Parentheses will be the most commonly used
multiplication sign. Also, when operands or variables are right next to one another, this is an indication that
you should use multiplication. Sometimes, too, multiplication will be indicated by a dot between two
numbers. For example,
7(3) = 21 7⋅4 = 28 10(3)(4) = 120
Try the following as practice.
1. 3(4) =
2. 9 . 8 =
3. 12(2) =
4. 4 . 5 ⋅ 3 =
440
Section 3: Order of Operations
Solving equations correctly requires the use of proper order of operations. The correct order is parentheses;
exponents; multiplication; division; addition; subtraction. Using any other order could result in erroneous final
answers. For example,
3(5) + 2 = 15 + 2 = 17
(4 + 7) − 6 = 11 − 6 = 5
= (4)2 = 16
Try the following as practice.
1. 3 + 2 − 4 =
2. 4(5) + 7 =
3.
4.
5.
6. 22 + 33 =
7. (3 + 2)2 =
441
Section 4: Variables
The formulas in this book require you to plug numbers into equations and solve those equations. You must,
therefore, understand the basic tenets of algebra, wherein a formula contains variables, you are told the values
of those variables, and you plug the values into the formula. For example,
If x = 9 and y = 7, then x + y = 9 + 7 = 16
If x = 10 and y = 7, then xy = 10(7) = 70
If x = 2, y = 5, and z = 8, then = 4 ⋅ 5 = 20
Try the following as practice.
1. , where x = 12 and y = 3
2. xy , where x = 1 and y = 1
3. x + y + z , where x = 1, y = 19, and z = 4
4. + 2, where x = 6 and y = 3
5. , where x = 36 and y = 11
442
Section 5: Negatives
There are several rules with respect to negatives. Negative numbers and positive numbers act differently when
they are added, subtracted, multiplied, and divided. Positive numbers get larger as the number line is traced
away from zero and toward positive infinity. Negative numbers, by contrast, get smaller as the number line is
traced toward negative infinity. Adding a negative number is equivalent to subtracting a positive number.
When a positive number is multiplied or divided by a negative number, the final answer is negative. When
two negative numbers are multiplied or divided, the answer is positive. For example,
5 + (−2) = 5 − 2 = 3
−5 + (−2) = −5 − 2 = −7
−10(9) = − 90
Try the following as practice.
1. −3 + (−2) =
2. −3 − 4 =
3. −5 + 3 =
4. 3 − 8 =
5. (−2)2 =
6. −22 =
7. (−4)(−5) =
8.
443
Section 6: Decimals and Rounding
This book requires you to round. Two decimal places are used here; however, your instructor might require
more or fewer, so pay attention to directions. When rounding to two decimal places, you should look at the
number in the third decimal place to decide whether you will round up or whether you will truncate. When
the number in the third (thousandths) position is 5 or greater, you should round the number in the second
(hundredths) position up. When the number in the thousandths position is 4 or less, you should truncate. The
diagram below shows these positions pictorially.
Examples:
.506 rounded to two decimal places = .51.632 rounded to two decimal places = .63
.50 + .70 = 1.20 (.42)(.80) = .336 ≈.34 = 3.742 ≈ 3.74 = (1.33)5 = 6.65
Try the following as practice.
1. .50 + .55 =
2.
3. 2.23 − .34 =
4.
5. 1 − .66 =
6. (.20)(.80) =
7.
444
8. =
9. Round this number to two decimal places: .605
10. Round this number to two decimal places: .098
If these operations all looked familiar to you and you were able to do them with little or no difficulty, then you
are ready for the course! If you struggled with them, you should speak with your course instructor regarding
recommendations and options.
445
Answers to Appendix A Problems
446
Section 1
1. 3
2. 7
3. 5
4. 3
447
Section 2
1. 12
2. 72
3. 24
4. 60
448
Section 3
1. 1
2. 27
3. 3
4. 26
5. 14
6. 13
7. 25
449
Section 4
1. 4
2. 1
3. 24
4. 4
5. 71
450
Section 5
1. −5
2. −7
3. −2
4. −5
5. 4
6. −4
7. 20
8. −3
451
Section 6
1. 1.05
2. 7.80
3. 1.89
4. 3.46
5. .34
6. .16
7. 1.58
8. 2.88 or 2.86
9. .61
10. .10
452
Appendix B. Standard Normal (z) Distribution
453
Area Between the Mean and z
454
455
456
Appendix C. t Distribution
457
458
Source: Abridged from R. A. Fisher and F. Yates, Statistical Tables for Biological, Agricultural and Medical Research , 6th ed. Copyright ©
R. A. Fisher and F. Yates 1963. Reprinted by permission of Pearson Education Limited.
459
Appendix D. Chi-Square (χ2) Distribution
460
461
Source: R. A. Fisher & F. Yates, Statistical Tables for Biological, Agricultural and Medical Research , 6th ed. Copyright © R. A. Fisher and
F. Yates 1963. Reprinted by permission of Pearson Education Limited.
462
Appendix E. F Distribution
463
464
Source: R. A. Fisher and F. Yates, Statistical Tables for Biological, Agricultural and Medical Research , 6th ed. Copyright © R. A. Fisher and
F. Yates 1963. Reprinted by permission of Pearson Education Limited.
465
466
467
Glossary
Alpha level:
The opposite of the confidence level; that is, the probability that a confidence interval does not contain
the true population parameter. Symbolized α.
Alternative hypothesis:
In an inferential test, the hypothesis predicting that there is a relationship between the independent and
dependent variables. Symbolized H1. Sometimes referred to as a research hypothesis.
Analysis of variance (ANOVA):
The analytic technique appropriate when an independent variable is categorical with three or more
classes and a dependent variable is continuous.
Beta weight:
A standardized slope coefficient that ranges from –1.00 to +1.00 and can be interpreted similarly to a
correlation so that the magnitude of an independent variable–dependent variable relationship can be
assessed.
Between-group variance:
The extent to which groups or classes in a sample are similar to or different from one another. This is a
measure of true group effect, or a relationship between the independent and dependent variables.
Binomial coefficient:
The formula used to calculate the probability for each possible outcome of a trial and to create the
binomial probability distribution.
Binomial probability distribution:
A numerical or graphical display showing the probability associated with each possible outcome of a
trial.
Binomial:
A trial with exactly two possible outcomes. Also called a dichotomous or binary variable empirical
outcome.
Bivariate regression:
A regression analysis that uses one independent variable and one dependent variable.
Bivariate:
Analysis involving two variables. Usually, one is designated the independent variable and the other the
dependent variable.
Bonferroni:
468
A widely used and relatively conservative post hoc test used in ANOVA when the null is rejected as a
means of determining the number and location(s) of differences between groups.
Bounding rule:
The rule stating that all proportions range from 0.00 to 1.00.
Categorical variable:
A variable that classifies people or objects into groups. Two types: nominal and ordinal.
Cell:
The place in a table where a row and a column meet.
Central limit theorem:
The property of the sampling distribution that guarantees that this curve will be normally distributed
when infinite samples of large size have been drawn.
χ² distribution:
The sampling or probability distribution for chi-square tests. This curve is nonnormal and contains only
positive values. Its shape depends on the size of the crosstabs table.]
Chi-square test of independence:
The hypothesis-testing procedure appropriate when both the independent variable and the dependent
variables are categorical.
Classes:
The categories or groups within a nominal or ordinal variable.
Combination:
The total number of ways that a success r can occur over N trials.
Confidence interval:
A range of values spanning a point estimate that is calculated so as to have a certain probability of
containing the population parameter.
Constant:
A characteristic that describes people, objects, or places and takes on only one value in a sample or
population.
Contingency table:
A table showing the overlap between two variables.
Continuous variable:
A variable that numerically measures the presence of a particular characteristic. Two types: interval and
ratio.
469
Cramer’s V:
A symmetric measure of association for χ² when the variables are nominal or when one is ordinal and
the other is nominal. V ranges from 0.00 to 1.00 and indicates the strength of the relationship. Higher
values represent stronger relationships. Identical to phi in 2 × 2 tables.
Critical value:
The value of z or t associated with a given alpha level. Symbolized z⍺ or t⍺.
Cumulative:
A frequency, proportion, or percentage obtained by adding a given number to all numbers below it.
Dependent samples:
Pairs of samples in which the selection of people or objects into one sample directly affected, or was
directly affected by, the selection of people or objects into the other sample. The most common types
are matched pairs and repeated measures.
Dependent variable (DV):
The phenomenon that a researcher wishes to study, explain, or predict.
Descriptive research:
Studies done solely for the purpose of describing a particular phenomenon as it occurs in a sample.
Deviation score:
The distance between the mean of a data set and any given raw score in that set.
Dispersion:
The amount of spread or variability among the scores in a distribution.
Dummy variable:
A two-category variable with one class coded as 0 and the other coded as 1.
Empirical outcome:
A numerical result from a sample, such as a mean or frequency. Also called an observed outcome.
Empirical relationship:
The causality requirement holding that the independent and dependent variables possess an observed
relationship with one another.
Empirical:
Having the qualities of being measurable, observable, or tangible. Empirical phenomena are detectable
with senses such as sight, hearing, or touch.
Evaluation research:
Studies intended to assess the results of programs or interventions for purposes of discovering whether
those programs or interventions appear to be effective.
470
Exhaustive:
A property of all levels of measurement whereby the categories or range within a variable capture all
possible values.
Expected frequencies:
The theoretical results that would be seen if the null were true—that is, if the two variables were, in fact,
unrelated. Symbolized fe.
Exploratory research:
Studies that address issues that have not been examined much or at all in prior research and that
therefore might lack firm theoretical and empirical grounding.
F distribution:
The sampling distribution for ANOVA. The distribution is bounded at zero on the left and extends to
positive infinity; all values in the F distribution are positive.
F statistic:
The statistic used in ANOVA; a ratio of the amount of between-group variance present in a sample
relative to the amount of within-group variance.
Factorial:
Symbolized !, the mathematical function whereby the first number in a sequence is multiplied
successively by all numbers below it down to 1.00.
Failure:
Any outcome other than success or the event of interest.
Familywise error:
The increase in the likelihood of a Type I error (i.e., erroneous rejection of a true null hypothesis) that
results from running repeated statistical tests on a single sample.
Frequency:
A raw count of the number of times a particular characteristic appears in a data set.
Goodman and Kruskal’s gamma:
A symmetric measure of association used when both variables are ordinal or one is ordinal and the other
is dichotomous. Ranges from –1.00 to +1.00.
Hypothesis:
A single proposition, deduced from a theory, that must hold true in order for the theory itself to be
considered valid.
Independent samples:
Pairs of samples in which the selection of people or objects into one sample in no way affected, or was
471
affected by, the selection of people or objects into the other sample.
Independent variable (IV):
A factor or characteristic that is used to try to explain or predict a dependent variable.
Inferential analysis:
The process of generalizing from a sample to a population; the use of a sample statistic to estimate a
population parameter. Also called hypothesis testing.
Inferential statistics:
The field of statistics in which a descriptive statistic derived from a sample is employed probabilistically
to make a generalization or inference about the population from which the sample was drawn.
Intercept:
The point at which the regression line crosses the y-axis; also the value of y when x = 0.
Interval variable:
A quantitative variable that numerically measures the extent to which a particular characteristic is
present or absent and does not have a true zero point.
Kendall’s taub:
A symmetric measure of association for two ordinal variables when the number of rows and columns in
the crosstabs table are equal. Ranges from –1.00 to +1.00.
Kendall’s tauc:
A symmetric measure of association for two ordinal variables when the number of rows and columns in
the crosstabs table are unequal. Ranges from –1.00 to +1.00.
Kurtosis:
A measure of how much a distribution curve’s width departs from normality.
Lambda:
An asymmetric measure of association for χ² when the variables are nominal. Lambda ranges from 0.00
to 1.00 and is a proportionate reduction in error measure.
Leptokurtosis:
A measure of how peaked or clustered a distribution is.
Level of confidence:
The probability that a confidence interval contains the population parameter. Commonly set at 95% or
99%.
Level of measurement:
A variable’s specific type or classification. Four types: nominal, ordinal, interval, and ratio.
472
Linear relationship:
A relationship wherein the change in the dependent variable associated with a one-unit increase in the
independent variable remains static or constant at all levels of the independent variable.
Longitudinal variables:
Variables measured repeatedly over time.
Matched-pairs design:
A research strategy where a second sample is created on the basis of each case’s similarity to a case in an
existing sample.
Mean:
The arithmetic average of a set of data.
Measures of association:
Procedures for determining the strength or magnitude of a relationship after a chi-square test has
revealed a statistically significant association between two variables.
Measures of central tendency:
Descriptive statistics that offer information about where the scores in a particular data set tend to
cluster. Examples include the mode, the median, and the mean.
Median:
The score that cuts a distribution exactly in half such that 50% of the scores are above that value and
50% are below it.
Methods:
The procedures used to gather and analyze scientific data.
Midpoint of the magnitudes:
The property of the mean that causes all deviation scores based on the mean to sum to zero.
Mode:
The most frequently occurring category or value in a set of scores.
Multiple regression:
A regression analysis that uses two or more independent variables and one dependent variable.
Mutually exclusive:
A property of all levels of measurement whereby there is no overlap between the categories within a
variable.
Negative skew:
A clustering of scores in the right-hand side of a distribution with some relatively small scores that pull
the tail toward the negative side of the number line.
473
Negative correlation:
When a one-unit increase in the independent variable is associated with a decrease in the dependent
variable.
Nominal variable:
A classification that places people or objects into different groups according to a particular characteristic
that cannot be ranked in terms of quantity.
Nonparametric statistics:
The class of statistical tests used when dependent variables are categorical and the sampling distribution
cannot be assumed to approximate normality.
Nonspuriousness:
The causality requirement holding that the relationship between the independent variable and
dependent variable not be the product of a third variable that has been erroneously omitted from the
analysis.
Normal curve:
A distribution of raw scores from a sample or population that is symmetric and unimodal and has an
area of 1.00. Normal curves are expressed in raw units and differ from one another in metrics, means,
and standard deviations.
Normal distribution:
A set of scores that clusters in the center and tapers off to the left (negative) and right (positive) sides of
the number line.
Null hypothesis:
In an inferential test, the hypothesis predicting that there is no relationship between the independent
and dependent variables. Symbolized H0.
Observed frequencies:
The empirical results seen in a contingency table derived from sample data. Symbolized fo.
Obtained value:
The value of the test statistic arrived at using the mathematical formulas specific to a particular test. The
obtained value is the final product of Step 4 of a hypothesis test.
Omega squared:
A measure of association used in ANOVA when the null has been rejected in order to assess the
magnitude of the relationship between the independent and dependent variables. This measure shows
the proportion of the total variability in the sample that is attributable to between-group differences.
Omitted variable bias:
An error that occurs as a result of unrecognized spuriousness and a failure to include important third
474
variables in an analysis, leading to incorrect conclusions about the relationship between the independent
and dependent variables.
One-tailed tests:
Hypothesis tests in which the entire alpha is placed in either the upper (positive) or lower (negative) tail
such that there is only one critical value of the test statistic. Also called directional tests.
Ordinal variable:
A classification that places people or objects into different groups according to a particular characteristic
that can be ranked in terms of quantity.
Ordinary least squares (OLS) regression:
A common procedure for estimating regression equations that minimizes the errors in predicting the
dependent variable.
p value:
In SPSS output, the probability associated with the obtained value of the test statistic. When p < α, the
null hypothesis is rejected.
Parameter:
A number that describes a population from which samples might be drawn.
Parametric statistics:
The class of statistical tests used when dependent variables are continuous and normally distributed and
the sampling distribution can be assumed to approximate normality.
Partial slope coefficient:
A slope coefficient that measures the individual impact of an independent variable on a dependent
variable while holding other independent variables constant.
Pearson’s correlation:
The bivariate statistical analysis used when both independent and dependent variables are continuous.
Percentage:
A standardized form of a frequency that ranges from 0.00 to 100.00.
Phi:
A symmetric measure of association for chi-square with nominal variables and a 2 × 2 table. Identical to
Cramer’s V.
Platykurtosis:
A measure of how flat or spread out a distribution is.
Point estimate:
A sample statistic, such as a mean or proportion.
475
Pooled variances:
The type of t test appropriate when the samples are independent and the population variances are equal.
Population distribution:
An empirical distribution made of raw scores from a population.
Population:
The universe of people, objects, or locations that researchers wish to study. These groups are often very
large.
Positive correlation:
When a one-unit increase in the independent variable produces an increase in the dependent variable.
Positive skew:
A clustering of scores in the left-hand side of a distribution with some relatively large scores that pull
the tail toward the positive side of the number line.
Post hoc tests:
Analyses conducted when the null is rejected in ANOVA in order to determine the number and
location of differences between groups.
Probability distribution:
A table or graph showing the entire set of probabilities associated with every possible empirical
outcome.
Probability sampling:
A sampling technique in which all people, objects, or areas in a population have a known chance of
being selected into the sample.
Probability theory:
Logical premises that form a set of predictions about the likelihood of certain events or the empirical
results that one would expect to see in an infinite set of trials.
Probability:
The likelihood that a certain event will occur.
Proportion:
A standardized form of a frequency that ranges from 0.00 to 1.00.
r coefficient:
The test statistic in a correlation analysis.
Range:
A measure of dispersion for continuous variables that is calculated by subtracting the smallest score from
the largest. Symbolized as R.
476
Ratio variable:
A quantitative variable that numerically measures the extent to which a particular characteristic is
present or absent and has a true zero point.
Regression analysis:
A technique for modeling linear relationships between one or more independent variables and one
dependent variable wherein each independent variable is evaluated on the basis of its ability to accurately
predict the values of the dependent variable.
Repeated-measures design:
A research strategy used to measure the effectiveness of an intervention by comparing two sets of scores
(pre and post) from the same sample.
Replication:
The repetition of a particular study that is conducted for purposes of determining whether the original
study’s results hold when new samples or measures are employed.
Representativeness:
How closely the characteristics of a sample match those of the population from which the sample was
drawn.
Residual:
The difference between a predicted value and an empirical value on a dependent variable.
Restricted multiplication rule for independent events:
A rule of multiplication that allows the probability that two events will both occur to be calculated as the
product of each event’s probability of occurrence: That is, p(A and B) = p(A) · p(B).
Rule of the complement:
Based on the bounding rule, the rule stating that the proportion of cases that are not in a certain
category can be found by subtracting the proportion that are in that category from 1.00.
Sample distribution:
An empirical distribution made of raw scores from a sample.
Sample:
A subset pulled from a population with the goal of ultimately using the people, objects, or places in the
sample as a way to generalize to the population.
Sampling distribution:
A theoretical distribution made out of an infinite number of sample statistics.
Sampling error:
The uncertainty introduced into a sample statistic by the fact that any given sample is only one of many
477
samples that could have been drawn from that population.
Science:
The process of gathering and analyzing data in a systematic and controlled way using procedures that
are generally accepted by others in the discipline.
Separate variances:
The type of t test appropriate when the samples are independent and the population variances are
unequal.
Slope:
The steepness of the regression line and a measure of the change in the dependent variable produced by
a one-unit increase in an independent variable.
Somers’ d:
An asymmetric measure of association for χ² when the variables are nominal. Somers’ d ranges from –
1.00 to +1.00.
Standard deviation:
Computed as the square root of the variance, a measure of dispersion that is the mean of the deviation
scores. Notated as s or sd.
Standard error:
The standard deviation of the sampling distribution.
Standard normal curve:
A distribution of z scores. The curve is symmetric and unimodal and has a mean of zero, a standard
deviation of 1.00, and an area of 1.00.
Statistic:
A number that describes a sample that has been drawn from a larger population.
Statistical dependence:
The condition in which two variables are related to one another; that is, knowing what class
persons/objects fall into on the independent variable helps predict which class they will fall into on the
dependent variable.
Statistical independence:
The condition in which two variables are not related to one another; that is, knowing what class persons
or objects fall into on the independent variable does not help predict which class they will fall into on
the dependent variable.
Statistical significance:
When the obtained value of a test statistic exceeds the critical value and the null is rejected.
478
Success:
The outcome of interest in a trial.
t distribution:
A family of curves whose shapes are determined by the size of the sample. All t curves are unimodal and
symmetric and have an area of 1.00.
t test:
The test used with a two-class, categorical independent variable and a continuous dependent variable.
Temporal ordering:
The causality requirement holding that an independent variable must precede a dependent variable.
Theoretical prediction:
A prediction, grounded in logic, about whether or not a certain event will occur.
Theory:
A set of proposed and testable explanations about reality that are bound together by logic and evidence.
Trends:
Patterns that indicate whether something is increasing, decreasing, or staying the same over time.
Trial:
An act that has several different possible outcomes.
Tukey’s honest significant difference:
A widely used post hoc test used in ANOVA when the null is rejected as a means of determining the
number and location(s) of differences between groups.
Two-tailed test:
A statistical test in which alpha is split in half and placed into both tails of the z or t distribution.
Type I error:
The erroneous rejection of a true null hypothesis. α
Type II error:
The erroneous retention of a false null hypothesis. β
Unit of analysis:
The object or target of a research study.
Univariate:
Involving one variable
Variable:
479
A characteristic that describes people, objects, or places and takes on multiple values in a sample or
population.
Variance:
A measure of dispersion calculated as the mean of the squared deviation scores. Notated as s2.
Variation ratio:
A measure of dispersion for variables of any level of measurement that is calculated as the proportion of
cases located outside the modal category. Symbolized as VR.
Within-group variance:
The amount of diversity that exists among the people or objects in a single group or class. This is a
measure of random fluctuation, or error.
z score:
A standardized version of a raw score that shows that offers two pieces of information about the raw
score: (1) how close it is to the distribution mean and (2) whether it is greater than or less than the
mean.
z table:
A table containing a list of z scores and the area of the curve that is between the distribution mean and
each individual z score.
480
Answers to Learning Checks
481
Chapter 1
1.1.
1. sample
2. sample
3. population
4. sample
1.2.
1. descriptive
2. evaluation
3. theory testing
4. exploratory
482
Chapter 2
2.1. The variable offense type could be coded as either a nominal or an ordinal variable because it could be
either purely descriptive or it could represent a ranking of severity. An example of a nominal coding
scheme is “violent offense, weapons offense, sex offense, property offense.” This coding approach does
not lend itself to clear ranking in terms of the severity of the crime. An example of ordinal coding is
“property offense, violent offense (non-lethal), homicide.” This approach allows for a moderate
comparison of crime severity across the different categories.
2.2. Zip codes are nominal variables. They are made up of numbers, but these numbers are not
meaningful in a statistical sense. Zip codes cannot be added, subtracted, multiplied, or divided. They are
merely placeholders designating particular locations. They offer no information about ranking or
quantification.
483
Chapter 3
3.1. A sum greater than 1.00 (for proportions) or 100.00 (for percentages) would suggest that one or
more cases in the table or data set got counted twice or that there is an error in the count of the total
cases. If the result is less than 1.00 or 100.00, then the opposite has occurred; that is, one or more cases
in the sample have been undercounted or the total has been overestimated.
3.2. State is a nominal variable, because it comprises categories with no rank order. Property crime rate is
ratio level, because it is continuous and has a true zero point.
3.3. Rates measure the prevalence of a certain event or characteristic in a population or sample.
Percentages break samples or populations down into elements that either do or do not possess a certain
characteristic. The calculation of rates requires information about the total population, something not
needed for percentages. Rates do not sum to 100 because they are a ratio-level variable that can range
from 0 to infinity (theoretically); that is, there is no maximum value that a rate cannot exceed.
Percentages are confined to the range 0 to 100.
3.4. In Table 3.8, column percentages would show the percentage of jails that do (and do not) offer
GED courses that are small, medium, and large. The column percentages for the jails offering courses
would be small = (155/988)100 = 15.69; medium = (315/988)100 = 31.88; and large = (518/988)100 =
52.43. This column sums to 100.00. The column percentages for prisons not providing courses is small
= (623/1,383)100 = 45.05; medium = (471/1,383)100 = 34.06; and large = (289/1,383)100 = 20.90.
With rounding, this column sums to 100.01.
3.5. Pie charts can be composed only of percentages because the idea behind them is that the entire “pie”
represents 100% of the total available options, categories, or cases. The total is then partitioned into its
constituent parts, with the “slices” representing percentages. This would not work with rates, because
rates are independent numbers that have no maximum values and do not sum to 100% when combined
with other rates.
3.6. The histogram in Figure 3.8 and the frequency polygon in Figure 3.10 probably would not benefit
substantially from grouping. Their ungrouped distributions display clear shapes; they are distinctly
different from the flat, shapeless histogram in Figure 3.14. Grouping might help if a researcher was
short on space and wanted to reduce the size of the chart, but it would not meaningfully enhance the
interpretability of the data.
3.7. Rates are advantageous compared to raw counts because rates are standardized according to the size
of the population. For instance, a police department with 25 officers serving a town of 1,000 residents is
much different than a department of 25 officers in a city of 100,000 residents. Without population
information, it is impossible to tell whether a 25-officer department is an appropriate size or is far too
small. Rates are more informative and useful than raw counts are in this type of situation.
484
Chapter 4
4.1. The positively skewed distribution should look similar to Figure 4.2, and the negatively skewed one
should look like Figure 4.3.
4.2. The modal category is “large” because there are more jails in this group (807) than in either of the
other two.
4.3. The modal number of assaults is 1, since the bar associated with this number is higher than any
other. The frequency is 136.
4.4. Adding Maryland at 8.59 makes the rank order 1.331 1.882 1.933 4.404 5.635 5.796 6.087 7.168
8.599. The MP = . The median is therefore 5.63, since this is the number in position 5 in
the ordered list. Adding Maryland moved the median up slightly (from 5.02).
4.5. First, rearrange the categories so that they are in order from Never to Every day or Almost every day.
Next, sum the frequencies until the cumulative sum is equal to or greater than 772.5. Since 457 + 37 +
82 + 217 = 793, we reach the same conclusion as before: A few days a week is the modal driving
frequency.
4.6. Formula 4(2) would be used to calculate the mean for Table 4.3. The mean property crime rate
among this sample of cities is
4.7. Formula 4(3) would be used to calculate the mean for Table 2.9. This is because the numbers are
arranged so that the left-hand column (Number Executed) contains the values in the data set and the
right-hand column (Frequency) displays the number of times each value occurs. The mean formula with
the f in the numerator (i.e., Formula 4[3]) is therefore the correct one. Using this formula, the fx
column and sum would be: 0 + 3 + 4 + 3 + 6 + 7 + 16 = 39. Divided by the sample size (N = 50), the
mean is .78. Across all 50 states, the mean number of executions per state in 2013 was .78. Since Table
2.9 includes all 50 states, including the 14 that do not authorize the death penalty, there are several
more zeroes (14, to be exact!) than there would be if only those 36 states that authorize capital
punishment were included in the calculations. These zeros pull the mean down. Similarly, the mean
calculated on the data in Table 4.8, which contains only the 36 authorizing states, would be lower if the
14 non-authorizing states were added because there would be more zeros in the numerator and a larger
denominator. In Table 2.9, dropping the 14 no-death penalty states from the analysis causes the mean
to increase to 39/36 = 1.08. In Table 4.8, adding 14 states with zero executions changes the calculation
to 81/50 = 1.62. These numerical outcomes confirm the conceptual predictions.
4.8. The mean cannot be calculated on ordinal data because ordinal variables are made of categories
rather than numbers. There is no way to add A few days a week to A few days a month. The arithmetic
required to calculate the mean cannot be done on a categorical variable.
4.9. Subtracting 196.65 from each of the raw scores in Table 4.3 produces the deviation scores –66.28,
485
100.59, –158.15, 128.38, 42.44, and –46.98. The deviation scores sum to 0.00.
4.10. The histogram in Figure 4.1 is fairly normal, so the mean and the median will be close to one
another. There is a slight positive skew, so the mean is probably a bit higher than the median.
486
Chapter 5
5.1. 45% is male; 4% were not convicted; and 10% received fines.
5.2. This distribution would be leptokurtic because the majority of the values cluster within a small
range and only a few values are outside of that range. Thus, the distribution’s shape or curve would be
tall and thin.
5.3. The x -axis will contain the numbers 0, 0, 1, 1, 2, 3, 10, 19 in that order (or 0 to 19), with a vertical
bar at 4.50 representing the mean. Points will be plotted above each number representing its distance
from the mean and whether it is above the mean or below it.
5.4. The sample with 10 cases would have a mean of 100/10 = 10, whereas the sample with 50 cases
would have a mean of 100/50 = 2. The first sample’s mean is much larger than the second sample’s
mean because in it, the sum of 100 is spread across just 10 cases; in the second sample, by contrast, the
100 is spread across 50 cases. Thus, two samples can have the same sum but very different means. This
is why sample size matters.
5.5. The reason is mathematical: Since variances and squared deviation scores are calculated by squaring
numbers, and since any number squared is positive (even if the original number was negative), variances
and squared deviations are always positive.
5.6. The mean salary is 732,644.99/8 = 91,580.62. The deviation scores are –4,007.62; 47.98; –
25,069.56; 5,919.38; 6,419.38; 5,279.71; 5,991.38; and 5,419.38. Squaring and summing the deviation
scores produces 813,935,309.50, and dividing by n – 1 (i.e., 8 – 1 = 7) gives us the variance
813,935,309.50/7 = 116,276,472.78. The standard deviation is the square root of the variance and is
10,783.16.
5.7. The upper and lower limits, in the order they appear in the table, are.85 – 1.23; 58.03 – 83.73; 2.15
– 9.05; and 1.38 – 3.62.
487
Chapter 6
6.1.
1. The probability of the die landing on 3 is 1/6 = .17. This probability remains the same no matter
what the original prediction was.
2. The probability of drawing the ace of hearts is 1/52 = .02. Again, the probability is not dependent
on the original prediction.
3. The probability of the die landing on any value except 1 is 5/6 = .83. This calculation would apply
to any initially predicted outcome.
4. The probability that the drawn card is anything except the nine of spades is 51/52 = .98. Again,
this would apply to any predicted outcome.
6.2. Homicide = 100.00 – 61.50 = 38.50, so there is a .62 probability of clearance and a .39 probability
that no arrest will be made. For rape, 62.20% (or .62) are not cleared. For robbery, the rate of non-
clearance is 70.70% (.71). Non-clearance is 46.00% (.46) for aggravated assault, 78.1% (.78) for larceny-
theft, and 86.9% (.87) for motor vehicle theft.
6.3. Answers will vary.
6.4. It is in Hospital B that the sex ratio of babies born in a single day would be expected to roughly
mirror that in the population. This is because 20 is a larger sample size than 6 and, therefore, is less
likely to produce anomalous or atypical results. In Hospital A, it would be relatively easy to have a day
wherein 4 (67%), 5 (83%), or even all 6 (100%) babies were of the same gender. In Hospital B, however,
it would be highly unlikely to see a result such as 16 of 20 babies being the same sex. Over the course of
a year, though, it would be expected that births in both hospitals would be approximately 50% female
and 50% male. This is because the sample size in both hospitals would be large enough that variations
seen on a day-to-day basis (even fairly large variations) would even out over time.
6.5. N/A
6.6. N/A
6.7. With a clearance probability of p = .62, q = 1.00 – .62 = .38, and N = 6, the resulting probabilities
are: p (0) = .003; p (1) = .04; p (2) = .11; p (3) = .24; p (4) = .32; p (5) = .21; and p (6) = .06.
6.8. The four types of deviation from normality include two types of skew (positive and negative) and
two types of kurtosis (leptokurtosis and platykurtosis).
6.9. Conceptually, a standard deviation is the mean of the deviation scores; in other words, it is the
average distance between the raw data points and the distribution mean. The mean offers information
about the location of the center of the distribution, and the standard deviation describes the amount of
spread or variability in the scores. Together, the mean and standard deviation offer a more
comprehensive picture of the shape of the distribution than either one of these numbers would provide
by itself.
6.10.
1. area between = .4162; area beyond = .50 – .4162 = .0838
2. area between = .2422; area beyond = .2578
488
3. area between = .4931; area beyond = .0069
4. area between = .4990; area beyond = .001
6.11. In the chapter, it was discovered that z scores greater than 2.33 are in the upper 1% of the
distribution. The corresponding z score for the bottom 1% is –2.33, and we can say that scores less than
–2.33 are in the bottom 1% of the distribution. We know this because the z distribution is symmetric.
Whatever numbers and areas are on one side will appear identically (with the opposite sign) on the other
side.
6.12.
489
Chapter 7
7.1.
1. The sample mean is noticeably far from the population mean (15-unit difference), so the sample is
not representative.
2. The sample proportion is fairly close to the population proportion (.05 units away), suggesting
reasonable representativeness.
3. Both sample means are less than their respective population means. While the difference for
violent crime is not huge (13 units), that for property crime is substantial (587 units). Therefore,
this sample appears to not be representative of the population.
7.2. The sampling distribution takes on a normal shape because most sample statistics cluster around the
population parameter. For example, if a mean in a population is 14, and someone pulls sample after
sample after sample, computing and plotting each one’s mean, then the majority (2/3, in fact) of the
sample means will be 14 or close to it. Values significantly far from 14 will occur but infrequently. This
means the distribution will peak at 14 and curve downward in the positive and negative directions.
490
Chapter 8
8.1. The 99% confidence interval (with z⍺ = ±2.58) will produce a wider confidence interval than the
95% level (z⍺ = ±1.96). Mathematically, this is because the 99% level’s z score is larger and will therefore
create a larger interval. Conceptually, this is an illustration of the trade-off between confidence and
precision: A higher confidence level means a wider (i.e., less precise) interval.
8.2.
1. t = ±2.228
2. t = ±1.725
3. t = ±2.660
8.3. The reason that critical values of t change depending on sample size is that t is not a single, static
curve like z is; instead, t is a family of curves that are taller and more normal shaped at larger sample
sizes and smaller and flatter at smaller sample sizes.
8.4. Two options to shrink the size of a confidence interval in order to increase the precision of the
estimate are to (1) reduce the confidence level or alpha and (2) increase the sample size.
8.5. The final result of the calculation is .91 ±.06, so the 95% CI : .85 ≤ P ≤ .97. The interval shrank
(became more precise) when the confidence level was reduced. This exemplifies the trade-off between
confidence and precision. The probability that the interval is correct fell (from 99% to 95%), but the
estimate is more precise.
491
Part III: Hypothesis Testing
1.
1. ratio
2. ordinal
3. nominal
4. ordinal
5. nominal
6. interval
7. Answers will vary
492
Chapter 9
9.1. The probability of a coin flip resulting in heads is .50. Results of flips will vary. The distribution of
flips should be closer to a 50/50 split in the 10-flip exercise than in the 6-flip exercise.
9.2. Agency type is the IV and officer-to-resident ratio is the DV. This is because we are predicting that
an agency’s jurisdiction characteristics (population and land area) will impact the size of its workforce.
493
Chapter 10
10.1. Expected-frequency final answers should match Table 10.8.
10.2. The expected frequencies, from cell A to cell D, are 31.18, 17.82, 24.82, and 14.18. The obtained
value of chi-square is.04 + .08 + .06 + .10 = .28. The obtained value is smaller than the critical value
(3.841), so the null is retained. This is the opposite of the conclusion reached in the chapter, where the
obtained value exceeded the critical value and the null was rejected. This illustrates the sensitivity of chi-
square tests to sample size.
494
Chapter 11
11.1. Using the sequence 3 + 4/2 results in 5 as an answer, while (3 + 4)/2 is 3.5. This is because the
calculator performed the division first when parentheses were not used, but did the addition first and the
division second when they were used. Similarly, –32 = –9, while (–3)2 = 9. Without parentheses, the
negative number remains negative; with parentheses, it becomes positive.
495
Chapter 12
12.1. The final answers to all the elements of the Fobt calculations can only be positive because SST and
SSB formulas require squaring (hence their names, sums of squares) . This eliminates negative signs and
ensures that the final answers will be positive.
12.2. In the first example, it was determined that juvenile defendants’ attorney types (public, private, or
assigned) did not significantly influence the jail sentences that these defendants received on conviction.
Since the null was retained, omega squared would not be computed, and post hoc tests would not be
examined; all these posttest procedures are appropriate only when the null is rejected.
496
Chapter 13
13.1. The reason why there will always be correspondence between the sign of the r value and the sign
of the t statistic lies in the sampling distribution for r. The negative values of r are located on the left
side of the sampling distribution (i.e., the negative side), and the positive values are on the right side
(i.e., the positive side). Therefore, r and t will always have the same sign.
13.2. Variables always correlate perfectly with themselves. You can test this by running a correlation
between two variables with the same coding. For instance, if each variable is coded 1, 2, 3, 4, 5, the
correlation between them will be 1.00. The same holds true if the coding is reversed to be 5, 4, 3, 2, 1.
There is a one-to-one correspondence between a variable and itself.
497
Chapter 14
14.1. Scatterplot:
14.2. The numerators are identical. The difference is in the denominators. There are more terms in
Formula 13(2) than in Formula 14(2), and information about y appears on the correlation formula but
not the formula for the slope coefficient.
14.3. Using the formula ŷ = 126.58 + 134.88x , the predicted fines are as follows. For x = 1, ŷ = 261.46;
for x = 3, ŷ = 531.22; for x = 4, ŷ = 666.10; and for x = 5, ŷ = 800.98. Residuals vary depending on each
person’s observed y value. For instance, the residual for the person with x = 1 and y = 120 is –141.46
(120 – 261.46) and that for the person who also had x = 1 but whose y value was 305 is 43.54 (305 –
261.46).The residuals could be reduced by adding variables measuring relevant legal characteristics of
cases, such as the severity of the charge, whether a juvenile has a prior criminal record, and whether she
or he was on probation at the time of the offense.
14.4. The reason is that adding more variables to the model changes the estimation of the relationship
between each IV and the DV. For example, a relationship that appears strong at the bivariate level can
become weaker when more IVs are added. Multiple regression, unlike bivariate regression, estimates
each IV’s influence on the DV while controlling for the impact of other variables, so each IV’s slope
coefficient changes when additional IVs are added to the model.
14.5.
1. ŷ = 7.10
2. ŷ = 6.7
3. ŷ = 2.69
4. ŷ = 40.60
14.6.
1. ŷ = –1,101.82 + 155.17(3) – 448.26(0) + 187.74(1) + 85.47(16) – 757.08(1) + 188.58(0) = 161.87
2. ŷ = –1,101.82 + 155.17(2) – 448.26(1) + 187.74(0) + 85.47(17) – 757.08(0) + 188.58(1) = 401.83
3. ŷ = –1,101.82 + 155.17(1) – 448.26(0) + 187.74(0) + 85.47(15) – 757.08(0) + 188.58(0) = 335.40
4. ŷ = –1,101.82 + 155.17(1) – 448.26(1) + 187.74(1) + 85.47(15) – 757.08(0) + 188.58(1) = 263.46
498
499
Answers to Review Problems
500
Chapter 1
1. Science is a systematic and controlled way of gathering information about the world. Methods are
integral to science because scientific results are trustworthy only when the procedures used to reach
them are considered correct by others in the scientific community. Scientific methods are logical,
transparent, and replicatable.
3. Samples are subsets of populations. Researchers draw samples because populations are too large to be
studied directly. Samples are smaller and therefore more feasible to work with.
5. Hypothesis testing is used to test individual components of theories as a means of determining the
validity of those theories. In this type of research, researchers use samples to draw inferences about
populations. Evaluation research assesses the effectiveness of a program or intervention. People, places,
or objects are measured before and after an intervention, or a control group is used for comparison to the
treatment group. Exploratory research delves into new areas of study about which little is known.
Hypotheses are generally not possible, because researchers usually do not have theory or prior evidence
to guide them. Descriptive research analyzes a sample and provides basic information about those
people, places, events, or objects. Usually, inferences are not made to the population and there is no
effort to discern the possible impact one variable might have on another.
7. Any three programs or policies; the correctness of students’ responses is up to the instructor’s
judgment. Correct answers will contain elements of random assignment to treatment and control
groups, repeated measures, or matched pairs.
9. This would be descriptive research, because that type of research focuses solely on a sample rather
than using a sample to draw conclusions about a population.
501
Chapter 2
1.
1. education
2. crime
3. people
3.
1. poverty (or income)
2. violent crime
3. neighborhoods
5.
1. money spent on education, health, and welfare
2. violent crime
3. countries
7.
1. police department location (urban or rural)
2. entry-level pay
3. police departments
9. The researcher has failed to consider additional variables. A statistical relationship between ice cream
and crime is not proof that one causes the other; one or more variables have been erroneously omitted
from the analysis. In the present case, the missing variable is probably ambient temperature—both ice
cream sales and crime might be higher in warmer months, so the empirical association between them is
in fact spurious.
11.
1. nominal
2. ratio
3. nominal
4. ratio
5. ratio
6. ordinal
7. nominal
13.
1. The first would produce a ratio variable, the second would create an ordinal variable, and the third
would make a nominal variable.
2. The phrasing that yields a ratio-level variable is best. Researchers who collect data should always
use the highest level of measurement possible. Continuous variables can be made into categorical
ones later on, but categorical data can never be made continuous.
15.
1. nominal
502
2. ratio
17.
1. victim advocacy (presence or absence of witness office)
2. nominal
3. sentencing (months of incarceration imposed)
4. ratio (the sample includes all offenders in each court, not just those sentenced to prison, so there is
a zero point)
5. courts
19.
1. homicide rate (homicides per population)
2. ratio
3. handgun ownership (own or not)
4. nominal
5. cities for homicide rates and people for gun ownership
503
Chapter 3
1.
1.
2. Since this variable is nominal, a pie chart or a bar graph would be appropriate. The pie chart
requires percentages be used, whereas the bar graph can be percentages or frequencies.
(Frequencies shown here.)
3.
3.
504
1. The data are ratio, so either a histogram or a frequency polygon would be the correct chart type.
2. The range is 42 − 0 = 42. With 10 intervals, the width of each would be
. Rounding to the nearest whole number, the width is 4.00.
505
3. Grouped data are technically ordinal, but the underlying scale is continuous, and so the chart type
would be a histogram or frequency polygon.
506
5. The line chart displays an upward trend, meaning that support has been increasing over time.
7.
1.
2. The variable city is nominal, narrowing the available options down to pie charts and bar graphs.
Since rates cannot be used for pie charts, a bar graph is correct.
507
9.
11. SPSS exercise. Using Transform → Compute and the command sworn/population*1000 will produce
the correct variable.
13. The variable is categorical, so a pie chart (percentages) or bar chart (percentages or frequencies) can
be used.
508
509
510
Chapter 4
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
1.
1. ratio
2. mode, median, mean
3.
1. nominal
2. mode
5. The mean is the midpoint of the magnitudes because it is the value that perfectly balances the
deviation scores. Deviation scores are produced by subtracting the mean from each raw score; deviation
scores measure the distance between raw scores and the mean. They always sum to zero.
7. c
9. a
7. c
11.
1. nominal; mode
2. mode = acquaintance
13.
a.
1. rank order: 170, 176, 211, 219, 220, 258, 317, 345
2. MP = (8 + 1)/2 = 4.5
3. Md = (219 + 220)/2 = 439/2 = 219.50
b. mean = (317 + 170 + 211 + 258 + 219 + 345 + 176 + 220)/8 = 1916/8 = 239.50
15.
1. mean = ((1 · 15) + (2 · 4) + (3 · 2) + (4 · 2) + (5 · 15) + (7 · 1) + (10 · 11) + (15 · 7) + (20 · 1) + (24 ·
1) + (30 · 1) + (45 · 2) + (60 · 2) + (120 · 2) + (180 · 1))/70 = (15 + 8 + 6 + 8 + 75 + 7 + 8 + 110 +
105 + 20 + 24 + 90 + 90 + 120 + 240 + 180)/70 = 1,106/70 = 15.80.
2. The mean (15.80) is substantially greater than the median (5.00), so this distribution is positively
skewed.
17.
1. MP = (7 + 1)/2 = 4, and Md = 26
2. mean = (31 + 26 + 31 + 15 + 13 + 27 + 4)/7 = 147/7 = 21.00
3. Alaska = 31 − 21.00 = 10.00; Arkansas = 26 − 21.00 = 5.00; Connecticut = 31 − 21.00 = 10.00;
Kansas = 15 − 21.00 = −6.00; Montana = 13 − 21.00 = −8.00; South Carolina = 27 − 21.00 = 6.00;
Vermont = 4 − 21.00 = −17.00; sum = 0
19. The mean = 19.87, the median = 10.00, and the mode = 10.
511
512
Chapter 5
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
For numbers close to zero, decimals are extended to the first nonzero number.
1. Measures of central tendency offer information about the middle of the distribution (i.e., where scores
tend to cluster), but they do not provide a picture of the amount of variability present in the data.
Measures of dispersion show whether the data cluster around the mean or whether they are very spread
out.
3. two-thirds
5. VR = 1 − (807/2,371) = 1 − .34 = .66
7. VR = 1 − (6,036/6,956) = 1 − .87 = .13
9.
1. R = 190 − 48 = 142
2. mean = (84 + 50 + 81 + 122 + 190 + 48)/6 = 575/6 = 95.83
3. variance = 14,300.84/(6 − 1) = 2,860.17
4. sd = = 53.48
11.
1. R = 18.70 − 9.90 = 8.80
2. mean = (13.70 + 18.70 + 9.90 + 12.40 + 16.20 + 14.00 + 10.50 + 10.40 + 11.00)/9 = 116.80/9 =
12.98
3. variance = 71.21/(9 − 1) = 8.90
4. sd = = 2.98
13.
1. 63.10 − 18.97 = 44.13 and 63.10 + 18.97 = 82.07
2. 1.75 − .35 = 1.40 and 1.75 + .35 = 2.10
3. 450.62 − 36.48 = 414.14 and 450.62 + 36.48 = 487.10
15. The standard deviation is the mean of the deviation scores. It represents the average (i.e., mean)
distance between the mean and the individual raw scores in the data set.
17. range = 42; variance = 67.15; standard deviation = 8.19
19. range = 64585; variance = 12294723.87; standard deviation = 3506.38
513
Chapter 6
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
For numbers close to zero, decimals are extended to the first nonzero number. Areas from the z table are
reported using all four decimal places.
1.
1. N = 8
2. r = 2
3. 28
3.
1. N = 7
2. r = 3
3. 35
5.
1. N = 8
2. r = 4
3. 70
7.
1. standard normal
2. standard normal
3. binomial
9.
1. With p = .62, q = .38, and N = 4: p(0) = .02; p(1) = .12; p(2) = .132; p(3) = .36; p(4) = .15
2. r = 3, or 3 of the 4 charges would be for aggravated assault
3. r = 0, or none of the 4 charges would be for aggravated assault
4. p(2) + p(1) + p(0) = .32 +.12 +.02 = .46
5. p(3) + p(4) = .36 +.15 = .51
11.
1. With p = .61, q = .39, and N = 5: p(0) = .01; p(1) = .07; p(2) = .22; p(3) = .35; p(4) = .27; p(5) = .08
2. r = 3, or 3 of the 5 murders committed with firearms
3. r = 0, or none of the 5 murders committed with firearms
4. p(1) + p(0) = .07 +.01 = .08
5. p(4) + p(5) = .27 +.08 = .35
13.
1. z4.28 = (4.28 − 1.99)/.84 = 2.29/.84 = 2.73
2. area between the mean and z = 2.73 is .4968
3. area in the tail beyond the mean and z = 2.73 is .50 − .4968 = .0032
15.
1. z1.29= (1.29 − 1.99)/.84 = − .70/.84 = −.83
514
2. area between the mean and z = −.83 is .2967
3. area in the tail beyond the mean and z = −.83 is .50 − .2967 = .2033
17. .50 − .03 = .47. The closest area on the table is .4699, which corresponds to a z score of 1.88. This is
the upper tail, so z = 1.88.
19. .50 − .10 = .40. The closest area on the table is .3997, which corresponds to a z score of 1.28. This is
the lower tail, so z = −1.28.
515
Chapter 7
1. a
3. b
5. sampling
7. b
9. z
11. a
13. c
15. b
516
Chapter 8
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
Calculation steps are identical to those in the text. For numbers close to zero, decimals are extended to the
first nonzero number.
1. At least 100.
3. The z distribution.
5. The z distribution (or standard normal curve) is fixed; it cannot change shape to accommodate small
samples. Small samples violate the assumption that the scores are perfectly normally distributed, so z
cannot be used in these instances.
7. df = 17, t = ±2.110.
95% CI: 6.80 ≤ μ ≤ 19.20
There is a 95% chance that the interval 6.80 to 19.20, inclusive, contains the population mean.
9. z = 1.96.
95% CI: 1.12 ≤ μ ≤ 1.16
There is a 95% chance that the interval 1.11 to 1.17, inclusive, contains the population mean.
517
11. z = 2.58
99% CI: 2.80 ≤ μ ≤ 3.32
There is a 99% chance that the interval 2.80 to 3.32, inclusive, contains the population mean.
13. df = 29, t = ±2.045
99% CI: 2.85 ≤ μ ≤ 5.83
There is a 95% chance that the interval 2.85 to 5.83, inclusive, contains the population mean.
15. z = 1.96
95% CI: .32 ≤ P ≤ .36
There is a 95% chance that the interval .32 to .36, inclusive, contains the population proportion.
518
17. z = 1.96
95% CI: .29 ≤ P ≤ .33
There is a 95% chance that the interval .29 to .33, inclusive, contains the population proportion.
19. z = 1.96
=.28 ± 1.96(.03)
= .28 ± .06
95% CI: .22 ≤ P ≤ .34
There is a 95% chance that the interval .22 to .34, inclusive, contains the population proportion.
519
Chapter 9
1. One possible explanation for the difference is sampling error. There might have been males with
unusually long sentences or females with atypically short sentences; these extreme values might have
pulled males’ mean upward, females’ mean downward, or both. The other possible explanation is that
men truly are given longer sentences than women, on average. It could be that there is a real difference
between the two population means.
3. The symbol is H1, and it predicts that there is a difference between populations; in other words, it
predicts that the observed differences between two or more samples’ statistics reflects a genuine
difference in the population.
5. A Type I error is the erroneous rejection of a true null (also called false positive). This occurs when a
researcher concludes that two or more variables are related when, in fact, they are not.
7. Reducing the likelihood of one type of error increases the probability that the other one will occur.
Preventing a Type I error requires increasing the amount of evidence needed to reject the null, which
raises the chances that a false null will not be rejected as it should be (Type II error). Preventing a Type
II error requires a reduction in the amount of evidence needed to reject the null, thus increasing the
probability that a true null will be wrongly rejected (Type I error).
9.
Step 1: State the null (N0) and alternative (N1) hypotheses. These are the competing predictions
about whether or not there is a true difference between population values.
Step 2: Identify the distribution and calculate the degrees of freedom. Each type of statistical test
uses a particular sampling distribution, so the correct one (and correct table) must be identified at
the outset. With the exception of the z distribution, sampling distributions are families of curves
and the degrees of freedom determine the shape of the curve that will be used.
Step 3: Identify the critical value of the test statistic and state the decision rule. The critical value
is located by using the table associated with the selected distribution. The decision rule states what
the obtained value (calculated in Step 4) must be in order for the null to be rejected.
Step 4: Calculate the obtained value of the test statistic. This is the mathematical part of the test.
Sample means, proportions, standard deviations, and sizes are entered into formulas, which are
then solved to produce the obtained value.
Step 5: Make a decision about the null and state the substantive conclusion. In this final step, the
obtained value is evaluated according to the decision rule (Step 3). If the criteria are met, the null
is rejected; if they are not, it is retained. The substantive conclusion is the interpretation of the
statistical outcome in the context of the specific variables and samples being analyzed.
11.
1. True effect. Events with low probabilities are atypical. They are unlikely to occur by chance alone.
2. Reject
13. b (categorical IV with two classes and continuous DV)
15. d (continuous IV and continuous DV)
520
17. a (categorical IV and categorical DV)
521
Chapter 10
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
For numbers close to zero, decimals are extended to the first nonzero number. Calculation steps are identical
to those in the text; using alternative sequences of steps might result in answers different from those presented
here. These differences might or might not alter the final decision regarding the null.
1. Yes. The IV and the DV are both categorical, so chi-square can be used.
3. No. The IV is categorical, but the DV is continuous, so chi-square cannot be used.
5.
1. The IV is gender and the DV is sentence received.
2. Both are nominal.
3. Two rows and three columns
7.
1. The IV is crime type and the DV is sentence length.
2. The IV is nominal and the DV is ordinal.
3. Three rows and three columns
9.
Step 1: H0: χ2 = 0; H1: χ2 > 0

Step 2: χ2 distribution with df = (2 − 1)(2 − 1) = 1

Step 3: χ2

crit = 6.635. Decision rule: If χ2

obt > 6.635, the null will be rejected.

Step 4: Expected frequencies are 28.14 for cell A, 38.86 for B, 34.86 for C, and 48.14 for D. χ2

obt

= 23.76 + 17.21 + 19.18 + 13.89 = 74.04

Step 5: The obtained value is greater than 6.635, so the null is rejected. There is a relationship

between whether a jail offers alcohol treatment and whether it offers psychiatric counseling. Row

percentages can be used to show that 80.6% of jails that offer alcohol treatment provide

counseling, compared to only 10.8% of those that do not offer alcohol treatment. It appears that

most jails supply either both of these services or neither of them; relatively few provide only one.

11.

Step 1: H0: χ2 = 0; H1: χ2 > 0

Step 2: χ2 distribution with df = (2 − 1)(2 − 1) = 1

Step 3: χ2

crit= 3.841. Decision rule: If χ2

obt > 3.841, the null will be rejected.

Step 4: Expected frequencies are 314.73 for cell A, 1023.27 for B, 67.27 for C, and 218.73 for D.

χ2

obt = .09 + .03 + .41 + .13 = .66

Step 5: The obtained value is less than 3.841, so the null is retained. There is no relationship

between victims’ gender and the likelihood that their injuries resulted from fights. Row

percentages show that 23.9% of males’ injuries occurred during fights, compared to 21.7% of

females’ injuries. The two percentages are substantively similar to one another, and the small

difference between them appears to be a chance finding.

522

13.

Step 1: H0: χ2 = 0; H1: χ2 > 0

Step 2: χ2 distribution with df = (2 − 1)(2 − 1) = 1

Step 3: χ2

crit = 3.841. Decision rule: If χ2

obt > 3.841, the null will be rejected.

Step 4: Expected frequencies are 46.51 for cell A, 27.49 for B, 85.49 for C, and 50.51 for D. χ2

obt

= .26 + .44 + .14 + .24 = 1.08.

Step 5: The obtained value is less than 3.841, so the null is retained. Gender and support for

marijuana legalization are statistically independent among black Americans. Looking at row

percentages, 67.6% of men and 60.3% of women believe that marijuana should be made legal.

There appears to be more support for legalization by men than by women, but this difference is

not statistically significant (i.e., appears to be a chance finding).

15.

Step 1: H0: χ2 = 0; H1: χ2 > 0.

Step 2: χ2 distribution with df = (3 − 1)(3 − 1) = 4

Step 3: χ2

crit = 13.277. Decision rule: If χ2

obt > 13.277, the null will be rejected.

Step 4: Expected frequencies are 315.90 for cell A, 42.69 for B, 13.42 for C, 200.41 for D, 27.08

for E, 8.51 for F, 260.70 for G, 35.23 for H, and 11.07 for I. χ2

obt = .003 + .01 + .19 +.10 + .57 +

.03 + .11 + .30 + .39 = 1.70

Step 5: The obtained value is less than 13.277, so the null is retained. There is no relationship

between annual income and the frequency of contact with police. Row percentages show that 85%

of people in the lowest-income category, 83% of those in the middle-income category, and 87% of

those in the highest-income group had between zero and two recent contacts. The vast majority of

people have very few annual contacts with officers, irrespective of their income.

17.

1. The SPSS output shows χ2

obt = 16.125.

2. The null is rejected at an alpha of .05 because p = .003, and .003 is less than .05.

3. Race and attitudes about courts’ harshness are statistically dependent.

4. Asking SPSS for row percentages shows the majority of people in all racial groups think courts are

not harsh enough, but this percentage is higher among whites (63.0%) than blacks (55.6%) or

members of other racial groups (61.8%). Likewise, blacks are more likely than the other two

groups to say that the courts are overly harsh on offenders (25.9%). The applicable measures of

association are Cramer’s V, lambda, gamma, and tauc. All of these values show a fairly weak

relationship between these two variables. This makes sense, because people’s attitudes about

courts’ approach to crime control are too complex to be determined by race alone.

19.

1. The SPSS output shows χ2

obt = 25.759.

2. The null is rejected at an alpha of .05 because p = .003, and .003 < .05. 3. There is a relationship between race and perceived stop legitimacy. Asking SPSS for row percentages shows that 84.4% of white drivers, 71.3% of black drivers, and 85.5% of drivers of 523 other races thought the stop was for a legitimate reason. Black drivers appear to stand out from nonblack drivers in that they are less likely to believe their stop was legitimate. 4. The null was rejected, so measures of association can be examined. Since both variables are nominal and there is a clear independent and dependent designation, Cramer’s V and lambda are both available. The SPSS output shows that lambda = .000, meaning that the relationship between race and stop legitimacy, while statistically significant, is substantively meaningless. Cramer’s V = .107, also signaling a very weak relationship. This makes sense looking at the percentages from Part C. A clear majority of all drivers believed their stop was for a legitimate reason. Black drivers deviated somewhat, but a large majority still endorsed stop legitimacy. 524 Chapter 11 Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer. For numbers close to zero, decimals are extended to the first nonzero number. Calculation steps are identical to those in the text; using alternative sequences of steps might result in answers different from those presented here. These differences might or might not alter the final decision regarding the null. 1. 1. whether the defendant plead guilty or went to trial 2. nominal 3. sentence 4. ratio (there is no indication that the sample was narrowed only to those who were incarcerated, so theoretically, zeroes are possible) 3. 1. judge gender 2. nominal 3. sentence severity 4. ratio 5. a 7. a 8. t; z 11. b 13. Step 1: H0: μ1= μ2; H0: μ1≠ μ2. (Note: No direction of the difference was specified, so the alternative is ≠.) Step 2: t distribution with df = 155 + 463 − 2 = 616 Step 3: tcrit = ±1.960 (±1.980 would also be acceptable). The decision rule is that if tobt is greater than 1.960 or less than −1.960, the null will be rejected. Step 4: Step 5: Since tobt is not greater than 1.960 or less than −1.960, the null is retained. There is no difference between MHOs and SHOs in the diversity of the crimes they commit. In other words, there does not appear to be a relationship between offenders’ status as MHOs or SHOs and the variability in their criminal activity. 525 15. Step 1: H0: μ1= μ2; H1: μ1 < μ2. Step 2: t distribution with Step 3: tcrit = −1.658 (−1.645 would also be acceptable). The decision rule is that if tobt is less than −1.658, the null will be rejected. Step 4: and tobt = −1.79. Step 5: Since tobt is less than −1.658, the null is rejected. Juveniles younger than 16 at the time of arrest received significantly shorter mean jail sentences relative to juveniles who were older than 16 at arrest. In other words, there appears to be a relationship between age at arrest and sentence severity for juveniles transferred to adult court. 17. Step 1: H0: μ1= μ2; H1: μ1 > μ2.

Step 2: t distribution with df = 160 + 181− 2 = 339

Step 3: tcrit = ±1.960 (±1.980 would also be acceptable). The decision rule is that if tobt is greater

than 1.960 or less than −1.960, the null will be rejected.

Step 4:

Step 5: Since tobt is greater than 1.960, the null is rejected. There is a statistically significant

difference between property and drug offenders’ mean fines. In other words, there does not appear

to be a relationship between crime type and fine amount.

19.

Step 1: H0: μ1= μ2; H1: μ1 ≠ μ2

Step 2: t distribution with df = 5 − 1 = 4

Step 3: tcrit = ±2.776. The decision rule is that if tobt is greater than 2.776 or less than −2.776, the

526

null will be rejected.

Step 4:

Step 5: Since tobt is not greater than 2.776 or less than −2.776, the null is retained. There is no

difference between states with high and low arrest rates in terms of officer assault. In other words,

there does not appear to be a relationship between arrest rates and officer assaults.

21.

Step 1: H0: P1= P2; H1: P1 ≠ P2.

Step 2: z distribution.

Step 3: zcrit = ±1.96 (recall that .50 − .025 = .475). The decision rule is that if zobt is less than −1.96

or greater than 1.96, the null will be rejected.

Step 4:

Step 5: Since zobt is greater than 1.96, the null is rejected. There is a significant difference between

juveniles represented by public attorneys and those represented by private counsel in terms of the

time it takes for their cases to reach disposition. In other words, there appears to be a relationship

between attorney type and time-to-disposition among juvenile drug defendants.

23.

1. Equal/pooled variances. Levene’s F = 2.810 with a p value of .095. Since .095 > .05, the F statistic

is not significant at alpha = .05 (i.e., the null of equal variances is retained).

2. tobt = .977

3. No. The p value for tobt is .330, which well exceeds .01; therefore, the null is retained.

527

4. There is no statistically significant difference between daytime and nighttime stops in terms of

duration. That is, there seems to be no relationship between whether a stop takes place at day or

night and the length of time the stop lasts.

25.

1. Unequal/separate variances. Levene’s F = 36.062 with a p value of .000. Since .000 < .05, the F
statistic is significant at alpha = .05 (i.e., the null of equal variances is rejected).
2. tobt = 8.095.
3. Yes. The p value for tobt is .000, which is less than .01; therefore, the null is retained.
4. There is a statistically significant difference between prosecutors’ offices that do and do not use
DNA in plea negotiations and trials in the total number of felony convictions obtained each year.
In other words, there is a relationship between DNA usage and total felony convictions. (Though
one would suspect, of course, that this relationship is spurious and attributable to the fact that
larger prosecutors’ offices process more cases and are more likely to use DNA as compared to
smaller offices.)
528
Chapter 12
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
For numbers close to zero, decimals are extended to the first nonzero number. Calculation steps are identical
to those in the text; using alternative sequences of steps might result in answers different from those presented
here. These differences might or might not alter the final decision regarding the null.
1.
1. judges’ gender
2. nominal
3. sentence severity
4. ratio
5. independent-samples t test
3.
1. arrest
2. nominal
3. recidivism
4. ratio
5. ANOVA
5.
1. poverty
2. ordinal
3. crime rate
4. ratio
5. ANOVA
7. Within-groups variance measures the amount of variability present among different members of the
same group. This type of variance is akin to white noise: it is the random fluctuations inevitably present
in any group of people, objects, or places. Between-groups variance measures the extent to which groups
differ from one another. This type of variance conveys information about whether or not there are actual
differences between groups.
9. The F statistic can never be negative because it is a measure of variance and variance cannot be
negative. Mathematically, variance is a squared measure; all negative numbers are squared during the
course of calculations. The final result, then, is always positive.
11.
Step 1: H0: μ1 = μ2 = μ3 and H1: some μi ≠ some μj.
Step 2: F distribution with dfB = 3 − 1 = 2 and dfW = 21 − 3 = 18
Step 3: Fcrit = 3.55 and the decision rule is that if Fobt is greater than 3.55, the null will be rejected.
529
Step 4:
SSB = 7(2.86 − 4.43)2 + 7(9.71 − 4.43)2 + 7(.71 − 4.43)2 = 7(2.46) + 7(27.88) + 7(13.84) = 309.26;
SSw = 1,045.14 − 309.26 = 735.88;
Step 5: Since Fobt is greater than 3.55, the null is rejected. There is a statistically significant
difference in the number of wiretaps authorized per crime type. In other words, wiretaps vary
significantly across crime types. Since the null was rejected, it is appropriate to examine
omega squared:
This means that 21% of the variance in wiretap authorizations is attributable to crime type.
13.
Step 1: H0: μ1 = μ2 = μ3 = μ4 and H1: some μi ≠ some μj.
Step 2: F distribution with dfB = 4 − 1 = 3 and dfW = 23 − 4 = 19
Step 3: Fcrit = 5.01 and the decision rule is that if Fobt is greater than 5.01, the null will be rejected.
Step 4:
SSB = 5(1.03 − 2.52)2 + 6(2.52 − 2.52)2 + 5(2.85 − 2.52)2 + 7(3.35 − 2.52)2 = 5(2.22) + 6(0) +
5(.11) + 7(.69) = 16.48;
530
SSw = 55.17 − 16.48 = 38.69
Step 5: Since Fobt is less than 5.01, the null is retained. There are no statistically significant
differences between regions in terms of the percentage of officer assaults committed with firearms.
In other words, there is no apparent relationship between region and firearm involvement in
officer assaults. Since the null was retained, it is not appropriate to calculate omega squared.
15.
Step 1: H0: μ1 = μ2 = μ3 = μ4 and H0: some μi ≠ some μj.
Step 2: F distribution with dfB = 4 − 1 = 3 and dfW = 20 − 4 = 16.
Step 3: Fcrit = 5.29 and the decision rule is that if Fobt is greater than 5.29, the null will be rejected.
Step 4:
SSB = 6(11.67 − 13.15)2 + 5(20.40 − 13.15)2 + 5(10.60 − 13.15)2 + 4(9.50 − 13.15)2 = 6(2.19) +
5(52.56) + 5(6.50) + 4(13.32) = 361.72;
SSw = 4,536.55 − 361.72 = 4,174.83
Step 5: Since Fobt is less than 5.29, the null is retained. There are no statistically significant
differences between juveniles of different races in the length of probation sentences they receive.
531
In other words, there is no apparent relationship between race and probation sentences among
juvenile property offenders. Since the null was retained, it is not appropriate to calculate omega
squared.
17.
Step 1: H0: μ1 = μ2 = μ3 = μ4 and H1: some μi ≠ some μj.
Step 2: F distribution with dfB = 4 − 1 = 3 and dfW = 20 − 4 = 16.
Step 3: Fcrit = 3.24 and the decision rule is that if Fobt is greater than 3.24, the null will be rejected.
Step 4:
=37,062 – 34,777.80
= 2,284.20
SSB = 5(44.20 – 41.70)2 + 5(36.20 – 41.70)2 + 5(29.00 – 41.70)2 + 5(57.40 – 41.70)2
= 5(2.50)2 + 5(–5.50)2 + 5(–12.70)2 + 5(15.70)2
= 5(6.25) + 5(30.25) + 5(161.29) + 5(246.49)
= 31.25 + 151.25 + 806.45+ 1,232.45
= 2,221.40
SSW = 2,284.20 − 2,221.40 = 62.80
Step 5: Since Fobt is greater than 3.24, the null is rejected. There is a statistically significant
difference in the percentage of patrol personnel police managers tasked with responsibility for
engaging in problem solving, depending on agency type. Since the null was rejected, it is
appropriate to examine omega squared:
532
This means that 97% of the variance in patrol assignment to problem-oriented tasks exists
between agencies (i.e., across groups). Agency type is an important predictor of the extent to
which top managers allocate patrol personnel to problem solving.
19.
1. Fobt = 9.631.
2. Yes. The p value is .000, which is less than .01, so the null is rejected.
3. Among juvenile property defendants, there are significant differences between different racial
groups in the amount of time it takes to acquire pretrial release. In other words, there is a
relationship between race and time-to-release.
4. Since the null was rejected, post hoc tests can be examined. Tukey and Bonferroni tests show that
there is one difference, and it lies between black and white youth. Group means reveal that black
youths’ mean time-to-release is 40.36 and white youths’ is 20.80. Hispanics, with a mean of 30.30,
appear to fall in the middle and are not significantly different from either of the other groups.
5. Since the null was rejected, it is correct to calculate omega squared:
Only about 2.7% of the variance in time-to-release is attributable to race. (This means that
important variables are missing! Knowing, for instance, juveniles’ offense types and prior records
would likely improve our understanding of the timing of their release.)
21.
1. Fobt = 3.496.
2. The null would be rejected because p = .000, which is less than .05.
3. There is a relationship between shooters’ intentions and victims’ ages. In other words, victim age
varies across assaults, accidents, and officer-involved shootings.
4. Since the null was rejected, post hoc tests can be examined. Tukey and Bonferroni both show that
the mean age of people shot by police is significantly different from the mean ages of victims of
assaults and accidental shootings. The group means are 28.99, 29.04, and 33.55 for people shot
unintentionally, in assaults, and by police, respectively. This shows that people shot by police are
significantly older than those shot in other circumstances.
5. It is appropriate to calculate and interpret omega squared. Using the SPSS output,
This value of omega squared is nearly zero (.01%) and suggests that shooter intent explains a
miniscule amount of the variance in victim age.
533
534
Chapter 13
Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer.
For numbers close to zero, decimals are extended to the first nonzero number. Calculation steps are identical
to those in the text; using alternative sequences of steps might result in answers different from those presented
here. These differences might or might not alter the final decision regarding the null.
1.
1. parental incarceration
2. nominal
3. lifetime incarceration
4. nominal
5. chi-square
3.
1. participation in community meetings
2. nominal
3. self-protective measures
4. ratio
5. two-population z test for a difference between proportions
5. A linear relationship is one in which a single-unit increase in the independent variable is associated
with a constant change in the dependent variable. In other words, the magnitude and the direction of
the relationship remain constant across all levels of the independent variable. When graphed, the IV
−DV overlap appears as a straight line.
7. The line of best fit is the line that minimizes the distance from that line to each of the raw values in
the data set. That is, it is the line that produces the smallest deviation scores (or error). No other line
would come closer to all of the data points in the sample.
9. c
11.
Step 1: H0: ρ = 0 and H1: ρ ≠ 0.
Step 2: t distribution with df = 5 − 2 = 3.
535
Step 4:
Step 5: Since tobt is not greater than 3.182, the null is retained. There is no correlation between
prison expenditures and violent crime rates. In other words, prison expenditures do not appear to
impact violent crime rates. Because the null was retained, it is not appropriate to examine the sign,
the magnitude, or the coefficient of determination.
13.
Step 1: H0: ρ = 0 and H1: ρ > 0.

Step 2: t distribution with df = 9 − 2 = 7.

Step 3: tcrit = 1.895 and the decision rule is that if tcrit is greater than 1.895, the null will be

rejected.

Step 4:

Step 5: Since tobt is greater than 1.895, the null is rejected. There is a positive correlation between

536

crime concentration and concentration of police agencies. In other words, where there is more

crime, there are apparently more police agencies. Since the null was rejected, the sign, the

magnitude, and the coefficient of determination can be examined. The sign is positive, meaning

that a one-unit increase in the IV is associated with an increase in the DV. The magnitude is very

strong, judging by the guidelines offered in the text (where values between 0 and ±.29 are weak,

from about ±.30 to ±.49 are moderate, ±.50 to ±.69 are strong, and those beyond ±.70 are very

strong). The coefficient of determination is .752 = .56. This means that 56% of the variance in

police agencies can be attributed to crime rates.

15.

Step 1: H0: ρ = 0 and H1: ρ < 0.
Step 2: t distribution with df = 7 − 2 = 5.
Step 3: tcrit = −3.365 and the decision rule is that if tcrit is less than −3.365, the null will be rejected.
Step 4:
Step 5: Since tobt is not less than −3.365, the null is retained. There is no correlation between
handgun and knife murder rates. In other words, murder handgun rates do not appear to affect
knife murder rates. Because the null was retained, it is not appropriate to examine the sign, the
magnitude, or the coefficient of determination.
17.
Step 1: H0: ρ = 0 and H1: ρ ≠ 0.
Step 2: t distribution with df = 8 − 2 = 6.
Step 3: tcrit = ±2.447 and the decision rule is that if tcrit is less than −2.447 or greater than 2.447,
the null will be rejected.
537
Step 4:
Step 5: Since tobt is not greater than 2.447, the null is retained. There is no correlation between
age and time-to-disposition among female juveniles. In other words, girls’ ages do not appear to
affect the time it takes for their cases to reach adjudication. Because the null was retained, it is not
appropriate to examine the sign, the magnitude, or the coefficient of determination.
19.
1. For age and contacts, r = .044; for age and length, r = − .113; for contacts and length, r = .312.
2. For age and contacts, the null is retained because .5556 > .05; for age and length, the null is

retained because .308 > .05; for contacts and length, the null is rejected because .004 < .05. 3. There is no correlation between age and the number of contacts male Asian respondents had with police in the past year; there is no correlation between age and the duration of traffic stops for those who had been stopped while driving; the number of recent contacts with police is significantly correlated with the duration of traffic stops. 4. The only test for which it is appropriate to examine the sign, the magnitude, and the coefficient of determination is the test for contacts and length. The sign is positive, meaning that a one-unit increase in one of these variables is associated with an increase in the other one. This suggests that those who experience more contacts also have longer encounters with officers. The magnitude is moderate. The coefficient of determination is.3122 = .097, meaning that only 9.7% of the variance in stop length is attributable to number of contacts (or 9.7% of these variables’ variance is shared). This suggests that although the variables are statistically related, the relationship is of very modest substantive importance. 538 Chapter 14 Note: Rounding, where applicable, is to two decimal places in each step of calculations and in the final answer. For numbers close to zero, decimals are extended to the first nonzero number. Calculation steps are identical to those in the text; using alternative sequences of steps might result in answers different from those presented here. These differences might or might not alter the final decision regarding the null. Numbers gleaned from SPSS output are presented using three decimal places. 1. Regression’s advantage over correlation is its ability to predict values of the DV (y) using specified values of the IV (x). Because regression fits a line to the data, any value of y can be predicted using any value of x. Values of x can be plugged into the formula for the line of best fit (ŷ = a + bx, at the bivariate level), and the equation can be solved to produce the predicted value of y at the given value of x. 3. No. The OLS model can accommodate IVs of any level of measurement. 5. a. b. a = 3.80 − (−.05)4.80 = 3.80 − (−.24) = 4.04 c. ŷ = 4.04 − .05x d. 1. For x = 3, ŷ = 4.04 − .05(3) = 3.89 2. For x = 15, ŷ = 4.04 − .05(15) = 3.26 e. Step 1: H0: B = 0 and H1: B ≠ 0 Step 2: t distribution with df = 5 − 2 = 3 Step 3: tcrit = ±3.182 and the decision rule is that if tobt is greater than 3.182 or less than −3.182, the null will be rejected. Step 4: Step 5: Since tobt is not less than −3.182, the null is retained. There is not a statistically significant relationship between prisoners’ prior incarcerations and their in-prison behavior. That is, prior incarceration history does not appear to significantly predict in-prison behavior. 5f. The null was not rejected, so the beta weight should not be calculated. 7. 539 a. b. a = 3.56 − (1.08)6.60 = −3.57 c. ŷ = −3.57 + 1.08x d. 1. For x = 4, ŷ = −3.57 + 1.08(4) = .75 2. For x = 8, ŷ = −3.57 + 1.08(8) = 5.07 e. Step 1: H0: B = 0 and H1: B ≠ 0. Step 2: t distribution with df = 10 − 2 = 8. Step 3: tcrit = ±2.306 and the decision rule is that if tobt is greater than 2.306 or less than −2.306, the null will be rejected. Step 4: Step 5: Since tobt is greater than 2.306, the null is rejected. There is a statistically significant relationship between unemployment rates and violent crime. That is, the concentration of unemployment in a state appears to help predict that state’s violent crime rate. f. 9. a. b. a = 4.28 − (.63)9.58 = −1.76 c. ŷ = −1.76 + .63x d. 1. For x = 5, ŷ = −1.76 + .63(5) = 1.39 2. For x = 10, ŷ = −1.76 + .63(10) = 4.54 e. 540 Step 1: H0: B = 0 and H1: B ≠ 0. Step 2: t distribution with df = 10 − 2 = 8. Step 3: tcrit = ±2.306 and the decision rule is that if tobt is greater than 2.306 or less than −2.306, the null will be rejected. Step 4: Step 5: Since tobt is greater than 2.306, the null is rejected. There is a statistically significant relationship between poverty rates and violent crime. That is, the concentration of poverty in a state appears to help predict that state’s violent crime rate. f. 11. 1. F = 19.176. The p value is .000. Since .000 < .05, the null is rejected. The model is statistically significant, meaning that the IV explains a statistically significant proportion of the variance in the DV. This means that it is appropriate to continue on and examine the slope coefficient. (Recall that an F value that is not statistically significant means that you should not move on and examine the specifics of the model because the model is not useful.) 2. R2 = .042. This means that charges explains 4.2% of the variance in probation. This is a very small amount of variance explained. 3. ŷ = 23.947 + 1.170x 4. The slope coefficient’s p value is .000, which is less than .05, so b is statistically significant. 5. Since b is statistically significant, it is appropriate to examine the beta weight. Here, beta = .205. 6. The number of charges juvenile drug defendants face is a statistically significant predictor of the length of their probation sentences; however, this relationship is weak, and the number of charges is not a substantively meaningful predictor. Clearly, important variables have been omitted from the model. 13. 1. ŷ = 27.438 + 1.079(2.09) − 4.322(0) = 29.69 2. ŷ = 27.438 + 1.079(2.09) − 4.322(1) = 25.37 3. The predicted sentence decreased by 4.322 units—exactly the slope coefficient for the priors 541 variable! The decrease means that juveniles with prior records received probation sentences that were, on average, 4.322 months shorter compared to juveniles without records. This seems counterintuitive, but there are possible explanations (such as judges treating first-time offenders more harshly in order to “teach them a lesson”) that we are not able to explore with the present data. 14. 1. ŷ = 27.438 + 1.079(2) − 4.322(.79) = 26.18 2. ŷ = 27.438 + 1.079(5) − 4.322(.79) = 29.42 3. The predicted sentence increased by 3.24 units. This means that a juvenile facing five charges should receive, on average, a sentence that is 3.24 months longer than a juvenile facing two charges. 15. 1. F = 14.270. Since p = .001, which is less than .05, the F statistic is significant. The IV explains a statistically significant amount of variance in the model, so it is appropriate to continue on and examine the slope coefficient. 2. R2 = .302. This means that the IV explains 30.2% of the variance in the DV. This is a strong relationship, suggesting that unemployment is a meaningful predictor of violent crime rates. 3. ŷ = − .410 + .637x 4. The p value for the slope coefficient is .001, which is less than .05, so b is statistically significant. 5. Since b is significant, it is appropriate to examine beta. Beta = .549. 6. Unemployment rates are statistically significant and substantively meaningful predictors of violent crime. Knowing the unemployment rate in an area significantly improves the ability to predict that area’s violent crime rate. 17. ŷ = −1.437 + .648(10) − .231(15) + .320(16) = 6.70 19. 1. ŷ = −1.437 + .648(6.77) − .231(11) + .320(8.94) = 3.27 2. ŷ = −1.437 + .648(6.77) − .231(6) + .320(8.94) = 4.42 3. The rate declined by 1.15 units. This means that states with 11% of the population receiving SNAP benefits should have violent crime rates that are, on average, 1.15 points lower than states where 6% of the population receives SNAP benefits. The negative association between SNAP rates and violent crime might seem backward, but it makes sense if SNAP is viewed as a measure of social support. Research has found that states and countries that provide greater levels of social support to disadvantaged citizens have less violent crime compared to those states or countries that do not provide as much social support. 542 References Chapter 1 Bureau of Justice Statistics. (2006). National Crime Victimization Survey, 2004: Codebook. Washington, DC: U.S. Department of Justice. Corsaro, N., Brunson, R. K., & McGarrell, E. F. (2013). Problem-oriented policing and open-air drug markets: Examining the Rockford pulling levers deterrence strategy. Crime & Delinquency, 59 (7), 1085–1107. Farrell, A. (2014). Environmental and institutional influences on police agency responses to human trafficking. Police Quarterly , 17 (1), 3–29. Hughes, C., Warren, P. Y., Stewart, E. A., Tomaskovic-Devey, D., & Mears, D. P. (2017). Racial threat, intergroup contact, and school punishment. Journal of Research in Crime and Delinquency, 54 (5), 583–616. Kerley, K. R., Hochstetler, A., & Copes, H. (2009). Self-control, prison victimization, and prison infractions. Criminal Justice Review, 34 (4), 553–568. Muftić, L. R., Bouffard, L. A., & Bouffard, J. A. (2007). An exploratory analysis of victim precipitation among men and women arrested for intimate partner violence. Feminist Criminology, 2 (4), 327–346. Paoline, E. A. III, Terrill, W., & Ingram, J. R. (2012). Police use of force and officer injuries: Comparing conducted energy devices (CEDs) to hands- and weapons-based tactics. Police Quarterly, 15 (2), 115–136. White, M. D., Ready, J., Riggs, C., Dawes, D. M., Hinz, A., & Ho, J. D. (2013). An incident-level profile of TASER device deployments in arrest-related deaths. Police Quarterly, 16 (1), 85–112. Chapter 2 Bouffard, L. A., & Piquero, N. L. (2010). Defiance theory and life course explanations of persistent offending. Crime & Delinquency, 56 (2), 227–252. 543 Bureau of Justice Statistics. (2011). The Police–Public Contact Survey, 2011. Ann Arbor, MI: Inter- University Consortium for Political and Social Research. Dass-Brailsford, P., & Myrick, A. C. (2010). Psychological trauma and substance abuse: The need for an integrated approach. Trauma, Violence, & Abuse, 11 (4), 202–213. Davis, J. A., & Smith, T. W. (2009). General social surveys, 1972–2008. Chicago: National Opinion Research Center, producer, 2005; Storrs, CT: Roper Center for Public Opinion Research, University of Connecticut. Gau, J. M., Mosher, C., & Pratt, T. C. (2010). An inquiry into the impact of suspect race on police use of Tasers. Police Quarterly, 13 (1), 27–48. Graham v. Connor. 490 U.S. 386 (1989). Haynes, S. H. (2011). The effects of victim-related contextual factors on the criminal justice system. Crime & Delinquency, 57 (2), 298–328. Kleck, G., & Kovandzic, T. (2009). City-level characteristics and individual handgun ownership: Effects of collective security and homicide. Journal of Contemporary Criminal Justice, 25 (1), 45–66. Lindsey, A. M., Mears, D. P., Cochran, J. C., Bales, W. D., & Stults, B. J. (2017). In prison and far from home: Spatial distance effects on inmate misconduct. Crime & Delinquency, 63 (9), 1043–1065. Social Justice, 26 (1), 115–138. Schmidt, J. D., & Sherman, L. W. (1993). Does arrest deter domestic violence? American Behavioral Scientist, 36 (5), 601–609. Crime & Delinquency, 60 (4), 569–591. Snell, T. L. (2014). Capital punishment in the United States, 2013–Statistical tables (BJS Publication No. 248448). Washington, DC: U.S. Department of Justice. Terrill, W., & Paoline, E. A. III. (2013). Examining less lethal force policy and the force continuum: Results from a national use-of-force study. U.S. Department of Justice. (2012). National Crime Victimization Survey, 2012. Ann Arbor, MI: Inter- University Consortium for Political and Social Research. 544 Chapter 3 Lane, J., & Fox, K. A. (2013). Fear of property, violent, and gang crime: Examining the shadow of sexual assault thesis among male and female offenders. Criminal Justice and Behavior, 40 (5), 472–496. Morgan, K. O., Morgan, S., & Boba, R. (2010). Crime: State rankings 2010. Washington, DC: CQ Press. Nuñez, N., Myers, B., Wilkowski, B. M., & Schweitzer, K. (2017). The impact of angry versus sad victim impact statements on mock jurors’ sentencing decisions in a capital trial. Criminal Justice & Behavior, 44 (6), 862–886. Reaves, B. A. (2013). Felony defendants in large urban counties, 2009—Statistical tables. Washington, DC: U.S. Department of Justice, Bureau of Justice Statistics. Steffensmeier, D., Zhong, H., Ackerman, J., Schwartz, J., & Agha, S. (2006). Gender gap trends for violent crimes, 1980 to 2003: A UCR–NCVS comparison. Feminist Criminology, 1 (1), 72–98. Truman, J. L., & Morgan, R. E. (2016). Criminal victimization, 2015 (Publication No. NCJ 250180). Washington, DC: Bureau of Justice Statistics. Chapter 4 Janssen, H. J., Weerman, F. M., & Eichelsheim, V. I. (2017). Parenting as a protective factor against criminogenic settings? Interaction effects between three aspects of parenting and unstructured socializing in disordered areas. Journal of Research in Crime and Delinquency, 54 (2), 181–207. Logue, M. A. (2017). The interactive effects of national origin and immigration status on Latino drug traffickers in U.S. federal courts. Race and Justice, 7 (3), 276–296. Morgan, K. O., Morgan, S., & Boba, R. (2010). Crime: State rankings 2010. Washington, DC: CQ Press. Piquero, A. R., Sullivan, C. J., & Farrington, D. P. (2010). Assessing differences between short-term, high- rate offenders and long-term, low-rate offenders. Criminal Justice and Behavior, 37 (12), 1309–1329. 545 Snell, T. L. (2014). Capital punishment, 2013–Statistical tables (Publication No. 248448). Washington, DC: U.S. Department of Justice. Sorensen, J., & Cunningham, M. D. (2010). Conviction offense and prison violence: A comparative study of murderers and other offenders. Crime & Delinquency, 56 (1), 103–125. Truman, J. L., & Langton, L. (2015). Criminal victimization, 2014 (BJS Publication No. 248973). Washington, DC: U.S. Department of Justice. Chapter 5 Copes, H., Kovandzic, T. V., Miller, J. M., & Williamson, L. (2014). The lost cause? Examining the Southern culture of honor through defensive gun use. Crime & Delinquency, 60 (3), 356–378. Perry, S. W., & Banks, D. (2011). Prosecutors in state courts, 2007—Statistical tables (Publication No. NCJ 234211). Washington, DC: Bureau of Justice Statistics. Pogarsky, G., & Piquero, A. R. (2003). Can punishment encourage offending? Investigating the “resetting” effect. Journal of Research in Crime and Delinquency, 40 (1), 95–12. Sydes, M. (2017). Revitalized or disorganized? Unpacking the immigration-crime link in a multiethnic setting. Journal of Research in Crime and Delinquency, 54 (5), 680–714. Chapter 6 Klinger, D. A. (1995). Policing spousal assault. Journal of Research in Crime and Delinquency, 32 (2), 308–324. Paoline, E. A. III, Lambert, E. G., & Hogan, N. L. (2006). A calm and happy keeper of the keys: The impact of ACA views, relations with coworkers, and policy views on the job stress and job satisfaction of correctional staff. The Prison Journal, 86 (2), 182–205. Ray, B., Grommon, E., Buchanan, V., Brown, B., & Watson, D. P. (2017). Access to recovery and recidivism among former prison inmates. International Journal of Offender Therapy and Comparative Criminology, 61 (8), 874–893. 546 Reaves, B. A. (2013). Felony defendants in large urban counties, 2009—Statistical tables. Washington, DC: Bureau of Justice Statistics. Chapter 7 Bouffard, L. A. (2010). Period effects in the impact of Vietnam-era military service on crime over the life course. Crime & Delinquency. Advanced online publication. doi:10.1177/0011128710372455 Morris, R. G., & Worrall, J. L. (2010). Prison architecture and inmate misconduct: A multilevel assessment. Crime & Delinquency. Advanced online publication. do:10.1177/0011128709336940 Rosenfeld, R., & Fornango, R. (2014). The impact of police stops on precinct robbery and burglary rates in New York City, 2003 – 2010. Justice Quarterly, 31 (1), 96–122. Chapter 8 Deslauriers-Varin, N., Beauregard, E., & Wong, J. (2011). Changing their mind about confessing to police: The role of contextual factors in crime confession. Police Quarterly, 14 (5), 5–24. Dixon, T. L. (2017). Good guys are still always white? Positive change and continued misrepresentation of race and crime on local television news. Communication Research, 44 (6), 775–792. Martin, K. R., & Garcia, L. (2011). Unintended pregnancy and intimate partner violence before and during pregnancy among Latina women in Los Angeles, California. Journal of Interpersonal Violence, 26 (6), 1157–1175. Orth, U., & Maercker, A. (2004). Do trials of perpetrators retraumatize crime victims? Journal of Interpersonal Violence , 19 (2), 212–227. Saum, C. A., Hiller, M. L., & Nolan, B. A. (2013). Predictors of completion of a driving under the influence (DUI) court for repeat offenders. Criminal Justice Review, 38 (2), 207–225. U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Injury Prevention and Control, & United States Consumer Product Safety Commission. (n.d.). Firearm Injury Surveillance Study, 1993–2010 [computer file]. ICPSR33861-v1. Atlanta, GA: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center 547 for Injury Prevention and Control [producer]; Ann Arbor MI: Inter-University Consortium for Political and Social Research [distributor]. Chapter 10 Griffin, T., Pason, A., Wiecko, F., & Brace, B. (2016). Comparing criminologists’ views on crime and justice issue with those of the general public. Criminal Justice Policy Review. doi:10.1177/0887403416638412 Petersen, N. (2017). Examining the sources of racial bias in potentially capital cases: A case study of police and prosecutorial discretion. Race and Justice, 7 (1), 7–34. Chapter 11 Barnoski, R. (2005). Sex offender sentencing in Washington State: Recidivism rates. Olympia: Washington State Institute for Public Policy. Bureau of Justice Statistics. (1998). Juvenile Defendants in Criminal Court, 1998 [data file]. Washington, DC: U.S. Department of Justice. Corsaro, N., Brunson, R. K., & McGarrell, E. F. (2013). Problem-oriented policing and open-air drug markets: Examining the Rockford pulling levers deterrence strategy. Crime & Delinquency, 59 (7), 1085–1107. Kaskela, T., & Pitkänen, T. (2016). The effects of gender and previous prison sentence on the risk of committing crime among a substance-abusing population. Crime & Delinquency. doi:10.1177/0011128716682229 Ostermann, M., & Matejkowski, J. (2014). Estimating the impact of mental illness on costs of crimes: A matched samples comparison. Criminal Justice and Behavior, 41 (1), 20–40. Wright, K. A., Pratt, T. C., & DeLisi, M. (2008). Examining offending specialization in a sample of male multiple homicide offenders. Homicide Studies, 12 (4), 381–398. Chapter 12 548 De Melo, S. N., Pereira, D. V. S., Andresen, M. A., & Matias, L. F. (2017). Spatial/temporal variations of crime: A routine activity theory perspective. International Journal of Offender Therapy and Comparative Criminology. doi: 0306624X17703654 Duff, J. C. (2010). Report of the director of the Administrative Office of the United States Courts. Retrieved from www.uscourts.gov/uscourts/Statistics/WiretapReports/2009/2009Wiretaptext Franklin, T. W., & Fearn, N. E. (2010). Sentencing Asian offenders in state courts: The influence of a prevalent stereotype. Crime & Delinquency. OnLineFirst. Tang, C. M., Nuñez, N., & Bourgeois, M. (2009). Effects of trial venue and pretrial bias on the evaluation of juvenile defendants. Criminal Justice Review, 34 (2), 210–225. Chapter 13 Henson, B., Reyns, B. W., Klahm, C. F., IV, & Frank, J. (2010). Do good recruits make good cops? Problems predicting and measuring academy and street-level success. Police Quarterly, 13 (1), 5–26. Morgan, K. O., Morgan, S., & Boba, R. (2010). Crime: State rankings 2010. Washington, DC: CQ Press. Reisig, M. D., Pratt, T. C., & Holtfreter, K. (2009). Perceived risk of Internet theft victimization: Examining the effects of social vulnerability and financial impulsivity. Criminal Justice and Behavior, 36 (4), 369–384. Walters, G. D. (2015). Criminal thinking as a predictor of prison misconduct and mediator of the static risk– infractions relationship. The Prison Journal, 95 (3), 353–369. Chapter 14 Henson, B., Reyns, B. W., Klahm, C. F., IV, & Frank, J. (2010). Do good recruits make good cops? Problems predicting and measuring academy and street-level success. Police Quarterly, 13 (1), 5–26. Mears, D. P., Mancini, C., & Stewart, E. A. (2009). Whites’ concern about crime: The effects of interracial contact. Journal of Research in Crime and Delinquency, 46 (4), 524–552. Petkovsek, M. A., & Boutwell, B. B. (2014). Childhood intelligence and the emergence of low self-control. 549 http://www.uscourts.gov/uscourts/Statistics/WiretapReports/2009/2009Wiretaptext Criminal Justice and Behavior. Advanced online publication. doi:10.1177/0093854814537812 Wright, K. A., Pratt, T. C., & DeLisi, M. (2008). Examining offending specialization in a sample of male multiple homicide offenders. Homicide Studies, 12 (4), 381–398. 550 Index alpha level, 180 alternative hypothesis (H1), 210–216 analysis of variance (ANOVA) defined, 281 measure of association and post hoc tests, 297–298 types of variances, 282–297 anecdotes, vs. data, 8 associations empirical vs. causation, 19 measures of, 234–236, 297–298 bar graphs, 49–50 beta weight, 345 between-group variance, 282 binomial coefficients, 140–141 binomial probability distributions, 140–147 binomials, 140 bivariate displays, 43–45 bivariate inferential tests, 214 bivariate regression, 335–346 Bonferroni, 298 bounding rule, 109 Bureau of Justice Statistics (BJS), 3–7, 30 categorical variables, 21 causation, vs. empirical associations, 19 cells, 37 Census of Jails (COJ), 43 Census of State Court Prosecutors (CSCP), 110–111 central limit theorem (CLT), 170 central tendency vs. dispersion, 107–108 measures of, 76 charts line charts, 56–57 pie charts, 47–48 See also graphs chi-square distributions, 222–223, 375–376 551 chi-square test of independence overview, 221–234 defined, 219 statistical dependence, 220 statistical independence, 219–220 use of, 214 classes, 47 coefficients binomial coefficients, 140–141 coefficient of determination, 325 partial slope coefficients, 347 r coefficient, 312 column proportions and percentages, 44 combinations, 143–144 complement, rule of, 109 confidence, 178–179 confidence intervals (CIs) defined, 177 for means with large samples, 179–185 for means with small samples, 185–190 with proportions and percentages, 190–198 constants, 15 contingency tables, 43–44 continuous variables defined, 21 frequency polygons, 54–55 histograms, 50–54 hypothesis testing with, 311–325 normal curve, 148–158 correlation analysis, 311–329 Cramer’s V, 235 critical value, 180 cumulative, defined, 40 data vs. anecdotes, 8 data distributions binomial probability distributions, 140–147 bivariate displays, 43–45 chi-square distributions, 222–223, 375–376 empirical distributions, 163–168 552 F distributions, 37–42, 282, 377–380 normal distributions, 76 population distributions, 163–168 probability distributions, 140 sample distributions, 163–172 sampling distributions, 163, 169–174 shapes, 96 t distributions, 172–174, 373–374 theoretical distributions, 168–172 univariate displays, 37–42 z distributions, 371–372 data sources Bureau of Justice Statistics (BJS), 3–7, 30 Census of Jails (COJ), 43 Census of State Court Prosecutors (CSCP), 110–111 Factfinder Series’ Crime Rankings, 57 Firearm Injury Surveillance Study (FISS), 188 General Social Survey (GSS), 25 Juvenile Defendants in Criminal Courts (JDCC), 252 Law Enforcement Management and Administrative Statistics (LEMAS) survey, 48–49 National Crime Victimization Survey (NCVS), 8 Police–Public Contact Survey (PPCS), 22 Uniform Crime Reports (UCR), 7 decimals, 368–369 dependent samples, 249 dependent variables (DVs), 18 descriptive research, 10 deviation scores, 97–98 dispersion vs. central tendency, 107–108 measures of, 107–108, 126–127 SPSS, 126–127 division, 367 dummy variables, 350 empirical distributions, 163–168 empirical outcomes, 138 empirical phenomena, 18 empirical relationships, vs. causation, 19 errors, Type I and Type II, 213–214 553 evaluation research, 10 exhaustive, defined, 23 expected frequencies, 224 expected outcomes, 208–210 exploratory research, 10 F distribution, 282, 377–380 F statistic, 282 Factfinder Series’ Crime Rankings, 57 factorial symbol, 144 failures, 144 familywise error, 281 Firearm Injury Surveillance Study (FISS), 188 frequencies defined, 37 expected and observed, 224 frequency polygons, 54–55 frequency polygons, 54–55 garbage in, garbage out (GIGO), 9 General Social Survey (GSS), 25 Goodman and Kruskal’s gamma, 237 graphs, 49–50. See also charts grouped data, 57–63 histograms, 50–54 hypotheses defined, 10 testing, 134 independent events, restricted multiplication rule for, 145 independent variables (IVs), 18 inferential analyses, 203 inferential statistics, 134 information, nonscientific, 8 intercept, 337 interval variables, 26–31 IV–DV relationships. see variables Juvenile Defendants in Criminal Courts (JDCC), 252 Kendall’s taub, 237 554 Kendall’s tauc, 237 kurtosis, 108 lambda, 235–236 Law Enforcement Management and Administrative Statistics (LEMAS) survey, 48–49 leptokurtosis, 108 level of confidence, 178 levels of measurement, 21–31, 37–45 line charts, 56–57 linear relationships, 312 longitudinal variables, line charts, 56–57 magnitude, 96, 325 matched-pairs designs, 250 mathematical techniques review, 367–369 mean, 88–92 measurement data displays for, 37–45 levels of, 21–31, 37–45 measures of association, 234–236, 297–298 measures of central tendency, 76, 107–108 measures of dispersion, 107–108, 126–127 median (Md), 83–88 methods, 8 midpoint of the magnitudes, 96 mode, 78–83 multiple regression, 335–336, 347–349 multiplication, 367 mutually exclusive, defined, 23 National Crime Victimization Survey (NCVS), 8 negative correlation, 311–312 negative skew, 77 negatives, 368 nominal variables, 21–23 nonparametric statistics, 219 nonscientific information, 8 nonspuriousness, 19 normal curve continuous variables, 148–158 defined, 148 555 and the standard deviation, 121–126 normal distributions, 76 null hypothesis(H0), 210–216 observations. see empirical phenomena observed frequencies, 224 observed outcomes, 208–210 obtained value, 223 omega squared, 286 omitted variables, 19–20 one-tailed tests, 251 order of operations, 367–368 ordinal variables, 21–25 ordinary least squares (OLS) regression alternatives to, 356–359 defined, 335–336 outcomes dependent variables (DVs) as, 20 empirical outcomes, 138 expected and observed, 208–210 p value, 238 parameters, 165 parametric statistics, 219–220 partial slope coefficients, 347 Pearson’s correlation, 311–312 percentages, 38 phi, 237 pie charts, 47–48 platykurtosis, 108 point estimates, 177 pooled variances, 250 population distributions, 163–168 populations defined, 8 student populations, 5 positive correlation, 311–312 positive skew, 76–77 post hoc tests, 286 precision, 178–178 556 predictors, independent variables (IVs) as, 20 probabilities overview, 135–140 distributions, 140 and proportions, 135–137 sampling, 10–11 theory, 134 proportions defined, 38 and probabilities, 135–137 pulling levers approach, 4 r coefficient, 312 range (R), 112–113 ratio variables, 28–31 regression analysis, 335–336 replication, 9 representativeness, 165 research methods, types of, 10–11 residual, 338 restricted multiplication rule for independent events, 145 rounding, 368–369 row proportions and percentages, 44 rule of the complement, 109 sample distributions, 163–172 samples, 8 sampling distributions, 163, 169–174 sampling error, 168 SAS, 11–12 science, defined, 8 separate variances, 250 sign, 325 slope, 337 software, 11–12. See also SPSS Somers’ d, 237 SPSS overview, 11–12, 63–69 analysis of variance (ANOVA), 299–302 chi-square tests, 238–239 correlation analysis, 326–329 557 dependent-samples t test, 270–273 measures of central tendency, 99–102 ordinary least squares (OLS) regression, 349–355 standard deviation, 118–120 standard error, 170 standard normal curve, 151. See also normal curve Stata, 11–12 statistical analysis, vs. research methods, 10–11 statistical dependence and independence, 219–221 statistical significance, 227 structural equation modeling (SEM), 357 student populations, 5 successes, 142 t distributions, 172–174, 373–374 t test overview, 247–264 for differences between proportions, 265–269 for independent samples, 249 for repeated-measures designs, 250 use of, 214 temporal ordering, 19 theoretical distributions, 168–172 theoretical predictions, 138 theories, defined, 10 trends, 56 trials, 140 Tukey’s honest significant difference (HSD), 298 two-tailed test, 180 Type I and Type II errors, 213–214 Uniform Crime Reports (UCR), 7 units of analysis, 15–18 univariate displays, 37–42 variables overview, 15–18 categorical variables, 21 continuous variables, 21, 50–55, 148–158, 311–325 dependent variables (DVs), 18 dummy variables, 350 558 independent variables (IVs), 18 interval variables, 26–31 longitudinal variables, 56–57 nominal variables, 21–23 omitted variables, 19–20 ordinal variables, 21–25 as outcomes, 20 as predictors, 20 ratio variables, 28–31 relationships between, 19–20 review of, 368 variances overview, 113–118 analysis of variance (ANOVA), 214, 281, 282–298 pooled variances, 250 separate variances, 250 SPSS, 299–302 within-group variance, 282 variation ratio (VR), 109–111 z distribution, 371–372 z scores, 150, 172–174 z tables, 154 559 Half Title Publisher Note Title Page Copyright Page Brief Contents Detailed Contents Preface to the Third Edition Acknowledgments About the Author Part I Descriptive Statistics Chapter 1 Introduction to the Use of Statistics in Criminal Justice and Criminology Chapter 2 Types of Variables and Levels of Measurement Chapter 3 Organizing, Displaying, and Presenting Data Chapter 4 Measures of Central Tendency Chapter 5 Measures of Dispersion Part II Probability and Distributions Chapter 6 Probability Chapter 7 Population, Sample, and Sampling Distributions Chapter 8 Point Estimates and Confidence Intervals Part III Hypothesis Testing Chapter 9 Hypothesis Testing A Conceptual Introduction Chapter 10 Hypothesis Testing With Two Categorical Variables Chi-Square Chapter 11 Hypothesis Testing With Two Population Means or Proportions Chapter 12 Hypothesis Testing With Three or More Population Means Analysis of Variance Chapter 13 Hypothesis Testing With Two Continuous Variables Correlation Chapter 14 Introduction to Regression Analysis Appendix A. Review of Basic Mathematical Techniques Appendix B. Standard Normal (z) Distribution Appendix C. t Distribution Appendix D. Chi-Square (χ2) Distribution Appendix E. F Distribution Glossary Answers to Learning Checks Answers to Review Problems References Index