Data Analysis With Small Samples and Non-Normal Data - Nonparametrics and Other Strategies

Data Analysis with
Small Samples
and Non-Normal Data
POCKET GUIDES TO
SOCIAL WORK RESEARCH METHODS
Series Editor
Tony Tripodi, DSW
Professor Emeritus, Ohio State University
Determining Sample Size Finding and Evaluating Evidence

Balancing Power, Precision, and Practicality Systematic Reviews and Evidence-Based Practice
Patrick Dattalo Denise E. Bronson and Tamara S. Davis
Preparing Research Articles Policy Creation and Evaluation
Bruce A. Thyer Understanding Welfare Reform in the United States
Systematic Reviews and Meta-Analysis Richard Hoefer
Julia H. Littell, Jacqueline Corcoran, Grounded Theory
and Vijayan Pillai Julianne S. Oktay
Historical Research Systematic Synthesis of Qualitative Research
Elizabeth Ann Danto Michael Saini and Aron Shlonsky
Confirmatory Factor Analysis Quasi-Experimental Research Designs
Donna Harrington Bruce A. Thyer
Randomized Controlled Trials Conducting Research in Juvenile and Criminal
Design and Implementation for Justice Settings
Community-Based Psychosocial Michael G. Vaughn, Carrie Pettus-Davis, and
Interventions Jeffrey J. Shook
Phyllis Solomon, Mary M. Cavanaugh, Qualitative Methods for Practice Research
and Jeffrey Draine Jeffrey Longhofer, Jerry Floersch, and Janet Hoy
Needs Assessment Analysis of Multiple Dependent Variables
David Royse, Michele Staton-Tindall, Patrick Dattalo
Karen Badger, and J. Matthew Webster
Culturally Competent Research
Multiple Regression with Discrete Using Ethnography as a Meta-Framework
Dependent Variables Mo Yee Lee and Amy Zaharlick
John G. Orme and Terri Combs-Orme
Using Complexity Theory for Research
Developing Cross-Cultural Measurement and Program Evaluation
Thanh V. Tran Michael Wolf-Branigin
Intervention Research Basic Statistics in Multivariate Analysis
Developing Social Programs Karen A. Randolph and Laura L. Myers
Mark W. Fraser, Jack M. Richman,
Research with Diverse Groups:
Maeda J. Galinsky, and Steven H. Day
Diversity and Research-Design and Measurement
Developing and Validating Rapid Assessment Equivalence
Instruments Antoinette Y. Farmer and G. Lawrence Farmer
Neil Abell, David W. Springer, and
Conducting Substance Use Research
Akihito Kamata
Audrey L. Begun and Thomas K. Gregoire
Clinical Data-Mining
A Social Justice
Integrating Practice and Research
Approach to Survey Design and Analysis
Irwin Epstein
Llewellyn J. Cornelius and Donna Harrington
Strategies to Approximate Random Sampling
Participatory Action Research
and Assignment
Hal A. Lawson, James Caringi, Loretta Pyles,
Patrick Dattalo
Janine Jurkowski, and Christine Bolzak
Analyzing Single System Design Data
Developing, Selecting, and Using Measures
William R. Nugent
David F. Gillespie and Brian E. Perron
Survival Analysis
Mixed Methods Research
Shenyang Guo
Daphne C. Watkins and Deborah Gioia
The Dissertation
Content Analysis
From Beginning to End
James W. Drisko and Tina Maschi
Peter Lyons and Howard J. Doueck
Group Work Research
Cross-Cultural Research
Charles D. Garvin, Richard M. Tolman, and
Jorge Delva, Paula Allen-Meares,
Mark J. Macgowan
and Sandra L. Momper
Developing Cross-Cultural Measurement in Social
Secondary Data Analysis
Work Research and Evaluation, Second Edition
Thomas P. Vartanian
Thanh V. Tran, Tam H. Nguyen, and Keith T. Chan
Narrative Inquiry
Data Analysis With Small Samples and
Kathleen Wells
Non-Normal Data
Structural Equation Modeling Nonparametrics and Other Strategies
Natasha K. Bowen and Shenyang Guo Carl F. Siebert and Darcy Clay Siebert
C ARL F. SIEBERT
DARC Y CL AY SIEBERT
Data Analysis
with Small Samples and
Non-Normal Data
Nonparametrics and
Other Strategies
1
3
Oxford University Press is a department of the University of Oxford. It furthers
the University’s objective of excellence in research, scholarship, and education
by publishing worldwide. Oxford is a registered trade mark of Oxford University
Press in the UK and certain other countries.
Published in the United States of America by Oxford University Press

198 Madison Avenue, New York, NY 10016, United States of America.
© Oxford University Press 2018
All rights reserved. No part of this publication may be reproduced, stored in

a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by license, or under terms agreed with the appropriate reproduction
rights organization. Inquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above.
You must not circulate this work in any other form

and you must impose this same condition on any acquirer.
Library of Congress Cataloging-in-Publication Data

Names: Siebert, Carl F., author. | Siebert, Darcy Clay, author.
Title: Data analysis with small samples and non-normal data :
nonparametrics and other strategies / Carl F. Siebert, Darcy Clay Siebert.
Description: New York : Oxford University Press, [2018] |
Series: Pocket guides to social work research methods |
Includes bibliographical references and index.
Identifiers: LCCN 2017010962 (print) | LCCN 2017014700 (ebook) |
ISBN 9780199391509 (updf) | ISBN 9780199391516 (epub) |
ISBN 9780199391493 (alk. paper)
Subjects: LCSH: Nonparametric statistics.
Classification: LCC QA278.8 (ebook) | LCC QA278.8 .S52 2017 (print) |
DDC 519.5/4—dc23
LC record available at https://lccn.loc.gov/2017010962
1 3 5 7 9 8 6 4 2
Printed by WebCom, Inc., Canada
Reprint Courtesy of International Business Machines Corporation,

© International Business Machines Corporation.
IBM® SPSS® Statistics software (“SPSS®”). SPSS Inc. was acquired
by IBM in October, 2009.
We dedicate this book to our children and grandchildren,
who are constant sources of joy and inspiration.
Contents
Acknowledgments ix
1 Introduction to Nonparametrics 1
2 Analyzing Single Variables and Single Groups 27
3 Comparing Two or More Independent Groups 69
4 Comparing Two or More Related Groups 107
5 Predicting with Multiple Independent Variables 149
Appendix A: SPSS Syntax 199

Appendix B: Missing Data 209
Appendix C: Other Resources 211
References 213
Index 219
vii
Acknowledgments
We both thank Dr. Tony Tripodi for his encouragement and patience.

Thanks, too, to Dr. Myles Hollander for sparking Carl’s interest in non-
parametric analysis and for all the lessons learned in his courses. We
especially wish to acknowledge Dr. Carl Henley for his early support of
Darcy’s abilities, and for teaching her that statistics are her friends, not
her enemies.
I, Carl, also want to thank my co-author, wife, and partner for
allowing me the privilege of spending my life with her, experiencing
unending love, and being constantly inspired by her intelligence and
dedication to helping others.
ix
Data Analysis with Small Samples
and Non-Normal Data
1
Introduction
to Nonparametrics
OVERVIEW
Research training in the social sciences must cover a wide range of
methodological and statistical coursework in a few years of doctoral
study. Frequently, doctoral students take two or three statistics courses
to learn basic analytic techniques, and students who take additional
courses often do so to learn advanced inferential techniques, such
as Hierarchical Linear Modeling (HLM) and Structural Equation
Modeling (SEM). These techniques can be very useful when one is work-
ing with large data sets and normally distributed data.
However, researchers regularly find themselves working with small
data sets, and this is especially commonplace for intervention research
studies, which are the hallmark of social science research. Similarly,
researchers may find themselves with data sets that simply do not meet
the assumptions required for the parametric techniques they were
trained to use. Yet very little, if any, time in their research training is
devoted to the analytic techniques that would allow them to properly
analyze these kinds of data—nonparametric statistics. This gap in edu-
cation leaves individuals without the necessary tools to conduct suitable
1
2 Data Analysis with Small Samples and Non-Normal Data
analyses of their data, or more importantly, the lack of exposure may

prevent researchers from recognizing that nonparametric tools are
available to them. The researchers’ primary option is to hope for robust-
ness when they utilize parametric methods, and this has the potential
to guide the person to the wrong conclusions based on the inappropriate
use of these statistical procedures.
This book is written in clear, understandable language for non-
statisticians who want to conduct statistical analyses on a small
sample and/or variables that are not normally distributed. The non-
parametric procedures covered herein describe how researchers can
analyze data that otherwise might fall by the wayside. For example,
successful grant proposals are frequently supported by believable
pilot data, and these pilots often consist of small samples. Funding
may hinge on researchers’ ability to show significant findings, and
using parametric methods for data that do not meet the assumptions
or have insufficient power can result in nonsignificant findings when
significant findings are actually present. Conversely, the researcher
may find significant results that are an artifact of using the inappro-
priate analytic technique.
In this book, we assume that you, the reader, want to understand
nonparametric statistics conceptually, and then learn the steps to
complete a nonparametric statistical test. We intend the book to be
a reference source more than a textbook, so we have formatted the
information in ways to make it easy to understand. For example, you
will find sections in this book where we have used bullet points rather
than standard text in paragraphs. This is to make the information easy
to find and simple to read.
We also do not cover the mathematical formulas, nor do we walk
you through the actual derivation of how to obtain the critical values for
hypothesis testing. In addition, we do not cover how to conduct a non-
parametric test by hand, as we assume you have access to a PC and have
a basic understanding of IBM® SPSS® Statistics software (SPSS) soft-
ware. We have chosen SPSS software because it offers a point-and-click
process for nonparametric procedures that is reasonably easy to follow.
Thus, this book does not offer R or SAS users the code to conduct non-
parametric tests; however, we include the SPSS syntax in Appendix A.
Accordingly, this book is written for nonstatisticians, yet it might
prove useful for applied statisticians as well, because we supply links to
Introduction to Nonparametrics 3
additional resources that will meet the needs of those interested in the
mathematical foundations of nonparametrics.
Another important topic that is not integrated into the chapters is
the difference between statistical significance (i.e., results from a statis-
tical test) and practical importance. Statistical tests are mathematical
procedures that provide a mathematical test of a hypothesis. Finding
significance from a statistical test is not the same as determining how
important the statistical finding might be for a particular research ques-
tion. Your data analyses may indicate statistical significance, when in
the real world, the difference is practically inconsequential. We strongly
urge anyone conducting statistical analyses, especially those making
decisions from statistical findings, to review fully the topic of practi-
cal importance related to statistics. (For more information on statistical
significance versus practical importance, see Hojat & Xu, 2004.)
Finally, we do not introduce bootstrapping as an option in our chapters.
Bootstrapping is another option for researchers dealing with small data
sets, and in simple terms, it draws inferences about the population under
investigation by resampling the data set from which a sample of the pop-
ulation came. Since the introduction of the bootstrap (Efron, 1979), proce-
dures have been improved and its use has been demonstrated to be a viable
alternative when dealing with challenging data (Davidson & Hinkley, 1998;
Janssen & Pauls, 2003; Wu, 1986). Because bootstrapping is an advanced
technique, we leave the discussion of it and how to use it to others.
ORGANIZATION
This book is organized so you can quickly locate the nonparametric
technique that is optimal for your data, and then follow the clear, step-
by-step instructions for using SPSS to conduct your analysis. Each
chapter includes a scenario typically found in social science research
to contextualize the analytic techniques, including details about the
research design, population, sample, groups within the sample, num-
ber of cases/observations/data points, time (if relevant), variables, level
of measurement, description of the variables’ distribution, and a series
of research questions and hypotheses. Next, using the scenario as the
example, relevant research questions and hypotheses guide the discus-
sion of the nonparametric method(s) that can be used to answer them.
In this way, you can easily match your particular questions to the
appropriate technique, and then follow the example to its conclusion
and interpretation.
Beginning with Chapter 2, each chapter starts with a list of chap-
ter objectives, a scenario that serves as the backdrop for the chapter’s
research questions, and a table of the nonparametric methods covered
in the chapter. Chapter 2 covers nonparametric methods using single
variables or groups. Chapter 3 presents procedures for comparing two or
more independent groups on a particular characteristic, and Chapter 4
describes techniques for comparing two or more related groups. The
fifth chapter describes methods for predicting a particular value or level
based on a set of independent variables, when the data do not meet the
assumptions for least-squares regressions.
The primary software used throughout the book is SPSS, but because
SPSS currently does not include nonparametric regression techniques
in its software, we utilize an add-in to Excel—X LSTAT—for the steps to
conduct the predictive analyses.
Additional useful information is supplied in the Appendices (e.g.,
SPSS syntax, missingness) and from SPSS Help (e.g., how to prepare
your data for analysis). We also suggest additional resources (i.e., web
links, articles, books) that will provide detailed foundational informa-
tion and formulas. On this book’s companion website you will find all
the data sets used in the book’s analyses. After downloading the data
sets, you can follow the steps for each nonparametric method using
the example data to reinforce your learning and understanding of the
process.
You may be tempted to go straight to the data analysis chapters at
this point, but if you do, you will not necessarily understand the foun-
dation for what you are doing. Consequently, you may err in describ-
ing your procedures to others, or in presenting the conclusions and
interpretations of your analyses. We urge you to review carefully all the
information in this chapter before proceeding.
WHAT ARE NONPARAMETRIC STATISTICS?

For some, nonparametric statistics or tests can be difficult to understand
at first, because most “Introduction to Statistics” courses are based on
parametric procedures. Nonparametric statistics are a set of tests that are

based on rankings, and they are often called distribution-free statistics,
meaning that nonparametric techniques “rely on no or few assumptions
about the shape or parameters of the population distribution from which
the sample was drawn” (Hoskin, n.d., p. 2). In addition, nonparametric
procedures can be used with nominal and ordinal data, as well as the
interval and ratio variables required for many parametric procedures.
KEY TERMS
Several key terms used are used throughout the book. These terms are
as follows:
• Hypothesis —a testable proposition that guides your study. The

null hypothesis indicates no statistically significant difference
between two groups or measured phenomena; that is, they
are drawn from the same sample. Thus, any difference in the
outcome variable is due to sampling error, and the independent
variable won’t make a difference. The alternative (research)
hypothesis suggests that the null hypothesis is incorrect, and
that a relationship does exist. The alternative hypothesis can
be either nondirectional or directional. In a nondirectional
hypothesis, a difference exists, but the direction of the
relationship (e.g., higher or lower) is unknown. The directional
alternative hypothesis indicates the direction of the hypothesis
(e.g., mean scores in X are greater than mean scores in Y), and
it has more power to detect the direction of the relationship in
which you are interested.
• Levels of measurement —a classifying scheme describing the
kind of information ascribed to the values representing a variable.
• Parametric assumptions —guidelines that the data representing
a variable(s) must meet before being used in a parametric
statistical procedure.
• Robustness —a property of a statistical technique in which
assumption violation will not adversely influence the test
results (i.e., the test’s actual Type I error rate is close to the .05
guideline).
• P-value —the probability of getting results at least as extreme

as the ones you observed, given that the null hypothesis (the
hypothesis of no effect/no difference) is correct.
• Independence —when one respondent does not influence how
another person responds to a question.
• Sample size —the number of subjects or records in a data set.
• Effect size —a quantitative measure of the magnitude of a
difference or relationship (e.g., the correlation between two
variables, a mean difference, or the influence of an intervention)
that is not influenced by sample size. Typical effect sizes for
social sciences (see Cohen, 1988, 1992) are: .1 = small effect,
.3 = moderate effect, .5 = large effect.
• Significance —in this book, the term for statistical significance,
meaning that the p-value is less than the a priori alpha level
chosen for the study. Practical significance, or the importance of
a finding in the real world, is labeled as such.
• Response coverage —the degree to which respondents
provide data for all possible response options for an item or
variable.
• Outlier —a response that is extreme in value, either low or high,
from the majority of other responses.
• Ties —when two responders or observations provide the same
value.
WHEN SHOULD NONPARAMETRIC PROCEDURES BE UTILIZED?

Nonparametric methods are the methods of choice when samples are
small and/or do not meet the assumptions for parametric methods.
These parametric assumptions are exceptionally important, as they are
the glue holding the mathematical processes together. If your data do
not meet one or more of the assumptions and you use the parametric
procedure anyway (yes, the statistical software will provide results for
a parametric procedure even if the assumptions are not met), you may
very well come to incorrect conclusions. In addition, when samples are
small, the ability to verify that your data meet parametric assumptions
becomes difficult and sometimes impossible.
Although not all parametric procedures have the same set of

assumptions, the primary parametric assumptions are as follows:
1. The population data are approximately normally distributed.

2. The population variances are equal (i.e., homogeneity of
variance among groups is present).
3. The sample is adequately large.
4. A linear relationship exists between or among the variables.
5. Homoscedasticity is present in regression models.
6. Residuals (error) follow a normal distribution.
Other methodological assumptions are fundamental when performing

any kind of statistical analysis. These assumptions are as follows:
7. The data are sampled randomly from a defined population.

8. The data consist of independent observations (apart from
paired data).
9. The levels of measurement match the variable design.
Let us examine these assumptions, one at a time, to understand,

conceptually, the circumstances in which a violation in an assumption
might require the use of a nonparametric strategy.
PARAMETRIC ASSUMPTIONS
Assumption 1. The Population Data Are Approximately Normally Distributed
That the data are approximately normally distributed may be the most
common parametric assumption (Hoskins, n.d., p. 4). The normal-
ity assumption is important whenever a scale variable (i.e., continu-
ous variable) is used in a parametric statistics test. Most parametric
tests involve a scale variable that is assumed to represent data from
a symmetric bell-shaped curve distribution. Small departures from
the symmetrical bell curve often have limited negative consequences
for an analysis, but larger departures from normality can make
the results wholly incorrect and/or uninterpretable. Furthermore,
although some parametric tests like ANOVA purport to be robust
against mild violations of normality, this does not mean that you can
ignore the normality assumption, particularly with small samples. In
addition, you are left to answer the question as to what, exactly, con-
stitutes a mild departure from normality.
All scale variables must be tested for normality prior to includ-
ing them in a parametric test. This is especially true when conducting
multivariable analyses in which multiple scale variables are included
in a single statistical model (e.g., multivariate regression). Each varia-
ble must demonstrate its association with a smooth, continuous, and
bell-shaped distribution. However, when the sample size is small, the
gaps between values (i.e., the data points on the curve) make the test
of normality more likely to fail. This is particularly true when testing a
Likert-t ype response variable that is being used as a scale variable rather
than an ordinal variable, and even more so when the response options
number five or fewer. Thus, although many social scientists tend to use
Likert-scaled response options as if they represent a continuous varia-
ble, one must be sure to take the size of the sample into consideration.
For example, it is easier to justify the use of a Likert-t ype variable if the
sample size is 1,000 and the response options range from 1 to 7 or 1 to 9
than if the sample size is 50 and the Likert response options are 1 to 4.
Testing to determine if a variable can be treated as a scale variable is
included in upcoming chapters.
You have two primary options when your variable fails the normal-
ity test. Your first option is to use a transformation strategy to normal-
ize your variable so that it will pass a normality test (e.g., log or square
transformation). Not all variables can be transformed into a normal
distribution (e.g., data with zeroes and negative values cannot be log
transformed), and when your transformation attempts fail, you cannot
simply use them anyway, hoping that your test will be robust against the
normality violation. When a non-normal variable does respond well to
transformation, you may use it in a parametric procedure, but then you
cannot interpret your findings as if the variable were not transformed.
You are faced with making sense of the transformed variable in the
analysis and in your findings, and this can be complicated, if not impos-
sible. (For more information on variable transformation, see McDonald,
2014, and Stat Trek, n.d.) Your second choice, when faced with a non-
normal variable, is simply to use a nonparametric test or the aforemen-
tioned bootstrapping.
Another challenge to verifying normality is when your data have

outliers. Because a scale variable is assumed to come from a bell-
shaped normal curve, this includes the assumption that every value
along the curve is a possible value for your variable (i.e., it is a reason-
able representative of your sample). Outliers are data points that fall
away from the mean value and that are multiple standard deviations
away from the center of the normal curve. Therefore, variables with
data that include outliers have a more difficult time passing normal-
ity tests. When your variable fails a normality test due to potential
outliers, you must investigate the validity of including the case with
the outlier. If you have irrefutable evidence that an outlier is not a
legitimate member of your sample (e.g., it was a data entry error),
you can choose to remove it or correct the error. If it does represent a
legitimate data point, although extreme, it must be included in your
analysis, even though it contributes to non-normality. In this situa-
tion, you should not run a parametric test anyway, hoping that the
outlier is not influential. Instead, use a nonparametric procedure.
(For more information on outliers, see Aguinis, Gottfredson, & Joo,
2013; Barnett & Lewis, 1978; Taylor, n.d.)
Finally, we must mention the use of Monte Carlo simulations. You
may find in the literature a reference suggesting that someone has
proved that it is reasonable to ignore assumption violations because
they (or someone else, often referencing Glass, 1972) conducted a
Monte Carlo study —basically research conducted with many dummy
data sets. The dummy data sets are created by drawing values for
variables that represent hypothetical distributions. With hundreds of
data sets that are slightly different from one another, researchers can
explore how sensitive a procedure is to a violation in an assumption.
The challenge is to fully understand how close your data situation
is to the dummy data sets studied and therefore the importance of
the assumption to your analysis. In other words, a Monte Carlo sim-
ulation is basically the creation of many data sets using automated
procedures to represent a particular data model. The data sets allow
a researcher to conduct a risk or other similar analysis to identify
generalizable results related to the data model. However, with small
data sets, the limited information a small data set provides restricts
the ability to identify the actual model the data represent. Because
parametric statistics are constructed with assumptions and because
each data set is unique, assuming your data set is one that conforms
to these earlier simulation studies may place your findings at risk if
you choose to ignore parametric assumptions.
Assumption 2. The Population Variances Are Equal (i.e., Homogeneity of Variance

Among Groups is Present), in that the Scores in Different Groups Have Similar Variances
Variance is how far a group of numbers are spread. No variance means
the data points are equal, a small variance means the data points are close
together around the mean, and a large variance means they are spread
out a good deal from one another. The equal variances assumption is
important for parametric tests that investigate differences between or
among groups within your targeted population. A commonly used
example to explain equal variances is the testing of height differences
between men and women. When examining differences between men
and women, think of the distribution of height for men and women as
separate groups. The variance of men’s height is expected to be similar
to the variance of women’s height, thereby meeting the assumption of
equal variances among groups.
A few parametric tests are appropriate for situations when the equal
variances assumption is not met, but the equal variance assumption is
not the only assumption for these tests (e.g., normality, sample size).
Therefore, when conducting parametric tests to look for group differ-
ences, it is important to know if variances are different among groups.
When your group variances are different, you cannot simply move ahead
as if they meet the assumptions. Instead, you have two options. First, use
a parametric test that allows for unequal variances while meeting all the
other assumptions, or second, use a nonparametric test that does not
assume equal or unequal variances. (For more information on testing
for equal variances, see Elliott, 2015.)
Assumption 3. The Sample Is Adequately Large

Terms in the statistical literature like “adequately large” or “small” are
essentially relative terms. For example, researchers working with large,
secondary data sets may believe that 500 cases constitute a small sam-
ple, but many intervention researchers would find 500 cases to be an
extraordinarily large sample. In practical applications, however, the
conceptualization of sample size is often simply related to the statisti-

cal method being utilized. If the sample in use does not meet the rules
for the statistical method a researcher wishes to use —for example, the
accepted practice of 10 to 15 cases per variable in the case of an ordinary
least squares (OLS) regression —then the sample becomes “too small.”
In addition, as the sample size decreases, the likelihood of verifying
that the assumptions for conducting parametric procedures are met also
decreases. Remember that parametric procedures require data that are
approximately normally distributed, and most texts recommend a min-
imum sample of 100 cases to investigate normality. Thus, you might
infer that samples of fewer than 100 cases cannot adequately verify the
normal distribution assumption, and nonparametric procedures must
be used. Because most nonparametric texts suggest that samples of 30
cases or fewer must always utilize nonparametric procedures (Hoskins,
n.d.; Mathews, 2010; Ryan, 2013), samples between 30 and 100 cases are
unclear, and you must examine the variables in these samples carefully
before proceeding.
Small samples are also problematic for parametric analyses because
sample size is linked to the statistical power needed to make inferences
about a population from your sample. Simply said, a small sample makes
it more likely that you will not find a significant effect in your analysis,
particularly if the effect size (e.g., of an intervention) is small. When
faced with a small sample, you have two options. First, you can always
use nonparametric procedures with small samples. Second, you can
utilize a parametric test along with its parallel nonparametric test and
include both in any report and in publication, explaining the strengths
and limitations of the findings for each.
Assumption 4. A Linear Relationship Exists Between or Among the Variables

Linearity is an important assumption for parametric tests like correlation
and regression. For example, obtaining a value for a correlation between
two variables is accurate only if their relationship can be represented by a
linear combination of parameters. In simple regression, two parameters
called regression coefficients explain the relationship between a depend-
ent variable (i.e., the outcome variable) and an independent variable
(i.e., predictor variable). The two parameters (i.e., intercept and slope)
can be used to graph a straight line, with the slope of the line being the
correlation value representing the relationship between the dependent

and independent variables. Of course, placing a line that best fits a scatter
plot for two variables is always possible, but as the scatter plot reveals a
departure from a straight line, the correlation value may no longer rep-
resent the relationship between the variables. The assumption of a linear
relationship dictates that a straight line can represent the relationship.
Simple and multiple regression are parametric models that assume
a linear relationship, and the models are used to predict values of a
dependent variable based on values from a single independent variable
or multiple independent variables. To make a prediction, the relation-
ship between the independent and dependent variables must be clear. In
other words, the independent variables in a statistical model like mul-
tiple regression must suggest only one value for a dependent variable.
Having the linear relationship between the independent and dependent
variables makes prediction possible.
In simple regression (i.e., regression with only one independent var-
iable), the relationship between the dependent and independent vari-
ables can be seen by graphing a scatter plot of the values. However, in
multiple regression (i.e., regression with multiple independent vari-
ables), verifying the linear relationship is a bit more complicated. In
addition, adding higher power regression coefficients can model curvi-
linear relationships between the dependent and independent variables.
(For more information on testing the linear relationship assumption,
see StatSoft, 2000.)
Conducting a simple or multiple regression when the assumptions
are not met has the potential for a misinterpretation of the relation-
ship between the independent and dependent variable(s), or worse, you
could identify a predicted value for the dependent variable that is truly
unfounded. You have several options in this situation. One option is to
attempt to transform one or a few of your variables to more closely resem-
ble a normal distribution, even though the transformation does make
explaining the variable relationships much more difficult. Another option
is to use nonparametric or nonlinear regression, discussed in Chapter 5.
Assumption 5. Homoscedasticity is Present in Regression Models

Equivalent to the assumption about equal variances, an assumption of
homoscedasticity is required when conducting parametric regression.
In simple terms, you are assuming that the data points for the relation-
ship between the dependent and independent variable have similar var-
iability along the entire range of the prediction line. For example, if you
were to examine a scatter plot of the prediction line and data points,
homoscedasticity is present if the points along the prediction line show
a similar pattern along the entire range of the prediction line. A viola-
tion of this assumption (i.e., heteroscedasticity) results in some cases
having more pull than others, as demonstrated by scatter plots that
can sometimes look like a cone or a fan. Heteroscedasticity can result
in invalid prediction values or inaccurate estimates of variable rela-
tionships. When faced with heteroscedasticity, you have two options.
First, you can look for variables that are causing the heteroscedastic-
ity and examine the cases that appear to be problematic to see if you
can legitimately replace them. However, it is not appropriate to go “case
shopping,” because any case that represents a real respondent cannot
be excluded without cause. The other option is to use a nonparametric
approach, such as nonparametric or nonlinear regression, as discussed
in Chapter 5. (For more information on testing the homoscedasticity
assumption, see Gaskin, 2015.)
Assumption 6. Residuals (Errors) Follow a Normal Distribution

When conducting regression, there is rarely, if ever, a time when the
dependent and independent variables suggest a prediction line that
matches perfectly with the data. The differences between the estimated
prediction line and the data used to establish the line are called residu-
als or errors. These residual values help explain how well your predic-
tion line matches the data. Larger residual values suggest less support
for prediction values, while smaller values suggest more support for
prediction values (i.e., poor prediction model versus better prediction
model, respectively). Given this, the distribution of the residual values
is assumed to follow a normal distribution. If they do not, the potential
to make incorrect predictions or to draw incorrect conclusions about
variable relationships is high. The non-normal residuals indicate that
your model is not constructed appropriately for predicting the depend-
ent variable. When faced with a non-normal residuals distribution, your
options are to find the variable(s) that are causing the issue and deter-
mine if you can legitimately exclude them from the model, or to use a
nonparametric process, such as nonparametric or nonlinear regression.

(For more information on testing the normal residuals distribution
assumption, see NIST/SEMATECH, 2012.)
OTHER ASSUMPTIONS
The assumptions considered next apply to both parametric and nonpa-
rametric statistical techniques, and are the product of a well-designed
study. Violating any of the assumptions below weakens or possibly
invalidates any conclusions you draw from your statistical analyses.
Assumption 7. The Data Are Sampled Randomly from a Defined

Population —A Probability Sample
If you are planning a research study that will select respondents ran-
domly from a population, you should spend a good deal of time iden-
tifying your target population and learning about all the potential
differences and similarities within the population. You must establish
clear inclusion and exclusion criteria so that your population is clearly
defined. Next, you will probably decide to collect data from a sample of
the population rather than the entire population, for one of any number
of reasons —e.g., you do not have access or enough time or money to
collect data from all the population members. Thus, the believability of
your study and the conclusions you draw about the population will be
based on your sample. Importantly, your sample must reflect the same
differences and similarities you identified in the population. To accom-
plish this, you must utilize a probability sampling method that gives
every person in your target population an equal opportunity to be in
your sample. This reassures the consumers of your research that your
sample is representative of the population.
You won’t always be able to utilize a probability sampling strategy
for your research, and sometimes your secondary data will not come
from a random sample. In these situations, you must clearly understand
and then explain that the analysis and findings are from a convenience
sample (i.e., a nonrandom sample from a population). Convenience
samples are commonplace in social science research, and especially
when the population of interest is a clinical population. Nevertheless,
incorrectly treating a convenience sample as a probability sample is

extremely risky, as it undermines your ability to draw accurate conclu-
sions from your data. Using a nonprobability sample must be justified
by the research question and has to be explained carefully in any reports
or publications as a potential limitation.
Assumption 8. The Data Consist of Independent Observations (Apart from Paired

Data, but for Which the Pairs Must Be Independent)
Independent observations involve the assumption that the informa-
tion that one respondent provides does not influence the information
provided by any other respondent. In other words, one response from
an individual does not have anything to do with responses from other
individuals, except for what is being measured. Unfortunately, trying
to anticipate all the possible threats to collecting independent observa-
tions is very challenging, but if this challenge is ignored, it can add
damaging bias to your study. If unrecognized, it can have a potentially
devastating effect on the accuracy of your findings. The independent
observation assumption is related to your ability to understand fully
the potential influences on the data you collect. The examples are end-
less and range from not being able to conduct an unbiased probabil-
ity sample to the natural occurrence of subjects’ falling into previously
unrecognized groups. The most important step in verifying independ-
ent observations is to examine closely how the data were collected, both
procedurally and theoretically.
If your observations are not independent and you can identify the
dependence issue, one option is to use multilevel modeling, which pro-
vides a way to model a dependence among your observations (e.g., stu-
dents within classrooms within schools). Unfortunately, most of these
models require a large data set. Ignoring a recognized dependence prob-
lem is not an option in either parametric or nonparametric strategies.
Assumption 9. The Levels of Measurement Match Variable Design

Measurement level is an important issue for all statistical procedures
(see Box 1.1, “Levels of Measurement”). This assumption requires you
to identify the appropriate level of measurement accurately for each
variable used in a statistical analysis. In social sciences, ratio data are
Box 1.1 Levels of Measurement
Nominal —Qualitative information (i.e., names or labels) that is often
classified into dichotomous typologies (e.g., heads or tails, yes or no) or
more than two (e.g., race, religion). The categories do not have an implied
order. We can determine if cases are equal or different on the variable, but
nothing else. Mode is the only central tendency measure.
Ordinal — Qualitative information that is rank ordered, indicating the

direction of difference, but that does not have equal distances between the
data points (e.g., the distance between strongly disagree, disagree, agree,
and strongly agree). We can determine if cases are equal or not equal, and
greater or less than on the variable only. Median and mode are measures of
central tendency.
Interval —Information that encompasses data points with equal distances
between them, so we can determine amount of difference, but without
an arbitrary zero (e.g., people’s scores on a summary measure). We can
determine if cases are equal or not, greater or less than, and we can add
or subtract. Mode, median, and arithmetic mean are measures of central
tendency.
Ratio —Information that includes a meaningful, nonarbitrary zero and
that is often used in age or counting (e.g., 2 is twice as many as 4; 50 years is
five times older than 10 years). We can determine if cases are equal or not,
greater or less than, and we can add, subtract, multiply, and divide. Mode,
median, arithmetic mean, and geometric mean are measures of central
tendency.
Measurement Levels in SPSS

Categorical
• Nominal —see under “Measurement Levels from Textbooks and
Research Courses”
• Ordinal —see under “Measurement Levels from Textbooks and
Research Courses”
• Single items with Likert-t ype response option scales with fewer than
6 points
Scale —Any variables that can be treated as continuous, such as

• Single items with Likert-t ype response option scales with 6 points
or longer
• Multiple items in a measure, summed for the scale score
• Interval or ratio —see under “Measurement Levels from Textbooks
and Research Courses”
uncommon, because they require a meaningful, nonarbitrary possible

value of zero. However, even interval data are infrequent, as the data
points must have equal distances between them. For scale variables,
the equal-interval assumption is extremely important. (The term equal
interval refers to the implied space between whole-number values of
your variable.) For example, if the theoretical space between a value
of 8 and 10 is not the same as the space between 20 and 22, you do not
have a variable with equal intervals. Not having equal intervals means
that your variable is not really a scale variable; rather, it is actually an
ordinal variable.
Nevertheless, social science researchers often treat variables with
Likert-like measures as if they were interval or scale measures, assum-
ing that the ordered responses (e.g., never, sometimes, frequently,
always) are equally separated. In other words, the space between
“never” and “sometimes” is the same as the space between “sometimes”
and “frequently.” This can be problematic, because the researcher who
designed the question may believe that the spaces between the response
options are the same, but the respondents may believe otherwise. For a
question to be reliable, each respondent must see the response options
in the same way. In other words, if some of the respondents perceive
equal spaces between the response options and other respondents do
not, the study may not be capturing the information in the way that the
researcher intended. The researcher may not recognize that the respon-
dents are approaching the questions differently, leaving the potential for
the researcher to report invalid findings.
Response coverage is another conceptual issue of importance that is
related to level of measurement and to the normality assumption. Response
coverage refers to whether you have enough responses for each response
option in a question. For example, assume that 100 people respond to a
4-point Likert-type variable with response options of strongly disagree,
disagree, agree, and strongly agree. If 37 people choose strongly disagree
and 57 people choose strongly agree, you are left with only six responses
total in the disagree and agree options. The information you have for agree
and disagree does not support the use of the 4-point Likert-type response
options scale, and you should not use the responses as a continuous var-
iable. Instead, you may need to transform the responses into a dichoto-
mous variable (i.e., variable with only two categories). Although you may
need to choose a different statistical test, the validity of the transformed
variable’s contribution is greatly increased. These decisions can be chal-

lenging and you should not base them exclusively on mathematical guide-
lines. A solid conceptual understanding of what the variable represents is
an important consideration in these situations.
In summary, researchers faced with measures that do not meet
the requirements for parametric techniques should not use them any-
way and simply hope that their results are meaningful. It is easy to do,
because statistical software packages like SPSS assume that the vari-
ables are measured and defined accurately. Frankly, some journals
will publish articles based on parametric analyses using inappropriate
measurement levels for the variables, as the practice has become com-
monplace. Yet this practice diminishes the believability of the findings
and weakens the reputation of social science research. Fortunately, non-
parametric techniques offer a viable alternative analytic strategy. When
researchers find themselves with nominal or ordinal data, including
Likert-like measures with few response options, and they are inclined to
utilize parametric techniques, a solid strategy would be to follow up the
parametric analysis with the parallel nonparametric analysis to deter-
mine if meaningful differences exist between the two. Including both
findings will serve to strengthen the science in a meaningful way.
ADVANTAGES AND DISADVANTAGES OF NONPARAMETRIC PROCEDURES

Nonparametric analytic techniques offer a number of important
advantages.
• They have few, if any, assumptions.

• These techniques are useful for dealing with samples containing
outliers, especially extreme outliers that cannot be logically
removed, because nonparametric techniques reduce excessive
influence of outliers.
• They reduce the effect of heterogeneity of variance.
• Some of the nonparametric tests provide an exact p-value, which
is the probability a null hypothesis is true given the data used in
the test.
• They can be used when (a) variable(s) violate the normality
assumption, (b) transforming the variable(s) does not solve the
normality requirement, or (c) transforming the variable makes

its interpretation unintelligible.
• They are useful for analysis of ordinal data (e.g., 5-point Likert-
type responses).
So why should a researcher not always use nonparametrics tech-

niques? Nonparametric analytic techniques are accompanied by several
disadvantages.
• Most importantly, when data are normally distributed, the

parallel parametric technique is more powerful. The parametric
technique has a higher probability that the procedure will reveal
significant findings —e.g., that two variables are associated with
one another when they are actually associated. However, having
more power justifies the use of parametric statistics only when
all the assumptions for the parametric technique are met.
• Nonparametric procedures provide no estimates of variance and
many times do not provide confidence intervals.
• Because nonparametric procedures frequently use rankings
rather than the actual data, statistical results can sometimes be
more difficult to interpret.
• Tied values can be problematic if too many of them exist in the
data set, and software packages are not always straightforward
in describing how they deal with ties. An example of a tie is
when two respondents provide the same value or response to a
scale-t ype question.
• Calculating power is less straightforward, requiring Monte
Carlo simulation methods (Mumby, 2002). Monte Carlo
simulations will allow the testing of different combinations of
sample sizes, group member comparisons (if applicable), effect
size, etc., all a priori to estimate the sensitivity of a future test to
reject a null hypothesis given different data scenarios.
MISCONCEPTIONS ABOUT NONPARAMETRIC TESTS

Misconceptions are commonplace about the conditions in which non-
parametric statistics should be applied and about the characteristics
that set them apart from parametric procedures. Over the years, many
people have disseminated the misconceptions, while others have refuted
them (e.g., Hunter, 1993; Sawilowsky, 2005). However, the repudiations
do not seem to have made much of an impact in correcting people’s
beliefs. We attribute this, in large part, to the lack of inclusion of nonpa-
rametric training in statistical courses. The confusion about these mis-
conceptions makes it difficult to avoid making errors when choosing
your statistical strategies. The list of misconceptions is quite long, but
a few that seem to come up more frequently than others are as follows:
• Parametric procedures are always more powerful than

nonparametric procedures. This is simply not true, especially
when dealing with non-normal data.
• Nonparametric statistics should be avoided because article
reviewers do not understand these procedures. Not true.
Reviewers are selected by their expertise and ability to review
article content. If you want your article to be accepted, using
an appropriate procedure for your data will improve your odds
greatly. Just be sure to justify why you chose the nonparametric
test for your particular research question or hypothesis.
• Nonparametric statistics should not be used on large data sets.
Untrue. The mathematical procedures for nonparametric
statistics are just as valid for large data sets as for small data sets,
but they are used more frequently with small data sets and data
with non-normal distributions.
• Unlike nonparametric tests, parametric tests are robust. Again,
not true. Robustness implies that violation of statistical
assumptions is not problematic. In fact, neither nonparametric
nor parametric tests are immune to violations of assumptions,
which is why it is important to test assumptions before
accepting any test’s findings.
A FEW WORDS OF CAUTION

1. You should not get the mistaken impression that because
nonparametric analytic techniques are available, it is always
OK to conduct research with small samples. We urge you to
conduct an a priori power analysis when planning a study, so

that all statistical techniques can be available for analyzing the
data, and so you will be able to detect small or moderate effect
sizes. Remember, parametric procedures used on large samples
with normally distributed data that meet all the statistical
assumptions are more powerful than nonparametric strategies.
Nevertheless, nonparametric strategies are extremely important
if data from your large sample do not meet the assumptions of
parametric tests, particularly the normality assumption. Studies
of human behavior, for example, often result in highly skewed
or kurtotic data (e.g., your outcome variable has clear limits of
detection or is censored, or your variables are ranked or ordinal
level). In addition, you may run into situations you do not
anticipate that result in small data sets (e.g., attrition from your
study sample), so it is useful to have nonparametric strategies to
use in these situations.
2. It can be easy to interpret p-values incorrectly, so it is important
to cover this topic a bit more extensively here rather than to
integrate it into the chapters. As defined, p-values are the
probability of getting results at least as extreme as the results you
observed, assuming that the null hypothesis (the hypothesis of no
effect/no difference) is correct. In other words, p-values evaluate
how well the sample data support the notion that the null
hypothesis is true. Thus, a small p-value (typically less than .05)
indicates that the sample provides sufficient evidence to reject the
null hypothesis for the whole population. For example, assume
that you conduct an intervention that found a p-value of .03. The
proper interpretation would be that if you assume the intervention
had no effect, you’d obtain the observed difference (or more) in 3%
of the cases, due to random sampling error.
Nevertheless, the literature contains many misinterpretations of

the p-value. It is not the probability of making an error by rejecting a
true null hypothesis —a Type I error in traditional statistical language.
Furthermore, p-values cannot indicate the likelihood that the null
hypothesis is true or false. A small p-value can specify that your data
are improbable if, indeed, the null were true, but it cannot indicate if the
null is false, or if the null is true but the sample was unusual. Thus, you
cannot say if the null hypothesis is rejected, you have a 3% likelihood of

making an error.
The American Statistical Association arrived at six statements about
p-values, as follows:
1. The p-value can indicate how incompatible the data are with a
specified statistical model.
2. The p-values do not measure the probability that the studied
hypothesis is true, or the probability that the data were produced
by random chance alone.
3. Scientific conclusions and business or policy decisions should
not be based only on whether a p-value passes a specific
threshold.
4. Proper inference requires full reporting and transparency.
5. A p-value, or statistical significance, does not measure the size of
an effect or the importance of an effect.
6. By itself, a p-value does not provide a good measure of evidence
regarding a model or hypothesis. (Wasserman & Lazar, 2016)
It is also important that you choose a level of significance (alpha)

for your threshold before you begin your analyses. Social scientists
typically choose .05 as their threshold, but in nonparametric analyses
it is sometimes reasonable to choose .10 instead (M. Hollander, per-
sonal communication, 2003), despite its tendency to provide an inflated
experimentwise Type I error rate. This information is integrated into
the chapters of this book.
CONVENTIONS UTILIZED IN THIS BOOK

• Each chapter begins with a research scenario, followed
by research questions and hypotheses to illustrate each
nonparametric technique.
• Levels of measurement are described in SPSS terms. Thus,
categorical includes nominal and ordinal measurement, which
includes dichotomous variables. Another term used to describe
categorical variables is discrete, which represents nominal,
ordinal, and count data. In addition, by our recommendation
and for use in this book, categorical also includes Likert-

type response option variables with fewer than 6 points.
Scale includes any variable that can be treated as continuous.
A continuous variable is one that can take on any value within a
particular range of values related to what the variable represents.
In this book, scale variables include multiple items in a measure
that are summed for the scale score, interval and ratio levels of
measurement, and single items with Likert-t ype response option
scales of 6 or more points.
• Notes are strategically placed for clarity and additional
explanation.
• Variable names are capitalized (e.g., Stress, Burnout).
• Response options for an item/question are italicized (low,
medium, high).
• SPSS labels (i.e., words or phrases in the SPSS window) are
boldface (Choose Tests, Fields). When you see boldface words
in SPSS, you can click on the words to make something happen
(e.g., if you click on Variable View, you will find the list of
variables).
• Window names are underlined (e.g., One-Sample
Nonparametric Tests).
• “Select =>” indicates that you should click to change the
appearance of a screen or open/close a window.
• “Click” indicates that you are choosing an option on the screen
or highlighting an item on the screen.
• is the icon representing a move arrow. When in SPSS, clicked
variables can be moved in and out of a variable list.
• In SPSS, the icon for nominal level measures is
• In SPSS, the icon for ordinal level measures is
• In SPSS, the icon for scale level measures is
• We use the language and symbols that SPSS uses.
• Sig. = significance level.
• Adj. Sig. = adjusted significance.
• Success = SPSS’s term for differentiating the values of a
dichotomous variable (e.g., if you have a 1 and a 0 as the values
of a dichotomous variable, you must choose one of them to
represent success). Note: The default for SPSS is to choose
whatever value it encounters first as the success value. If you run
a success-t ype test on a categorical variable with three or more

values, SPSS will choose the first value as success and the rest as
failure. This becomes clearer as you work your way through the
relevant chapters.
• We report the number of decimal places that are included in the
SPSS output (typically three), rather than rounding values.
FIRST STEPS
If you plan to follow the steps in the chapters to analyze your own data,
you will need to get your data into SPSS. Procedures for getting your
data into SPSS are found in SPSS’s help feature (click on the Help menu
option). Once your data are in SPSS, the next step is to verify that all
your variables are assigned to the appropriate level of measurement.
Figure 1.1 displays the data view for the example data used in Chapter 2.
Note that each variable is assigned one of the three levels of measure-
ment values (e.g., nominal, ordinal, or scale).
Once your data are in SPSS, you recognize SPSS notation, and you
are familiar with the key terms, you are ready to move on to Chapter 2.
Figure 1.1. Data view within SPSS.
SUGGESTIONS FOR FURTHER READING

Manfred te Grotenhuis, Manfred, & Matthijssen, Anneke (2015). Basic
SPSS Tutorial. Thousand Oaks, CA: Sage Publications.
Huff, D., & Geis, I. (1993). How to Lie with Statistics. NY: W. W. Norton
& Company.
Pett, M. A. (2015). Nonparametric Statistics for Health Care

Research: Statistics for Small Samples and Unusual Distributions.
Thousand Oaks, Ca: Sage Publications.
Salkind, N. J. (2015). Excel Statistics: A Quick Guide. Thousand Oaks,
CA: Sage Publications.
2
Analyzing Single
Variables and
Single Groups
CHAPTER OBJECTIVES
This chapter focuses on techniques for assessing single variables and
single groups, organized by the variable’s level of measurement. The fol-
lowing topics are covered:
• Examining variables for the quality of information they provide

• Dichotomous nominal variables —binomial test
• A nominal variable with three or more categories —
chi-square test
• An ordinal variable with three or more categories —
chi-square test
• A Likert-t ype variable treated as continuous —
Kolmogorov-Smirnov test
• A summed scale variable —Wilcoxon signed-rank test
27
• A scale variable needing testing for randomness —runs test

• A single group, analyzing the association between two or
more scale variables —Spearman’s rho and Kendall’s tau-b
coefficients
INTRODUCTION
Chapter 1 describes the importance of understanding the data and vari-
ables prior to and while using them in statistical tests. This begins by first
identifying the measurement level of the variables in the data set. The
measurement level dictates which nonparametric procedures can be used
to answer research questions. The second step is to review the quality of
the information provided by each variable. For scale-type variables, the
quality of information is related to capturing the true variability or distri-
bution of the data. In addition, variables that represent concepts or ideas
that have very little variability are not as informative and sometimes con-
tribute less to a statistical model. For example, a variable that has very
little variability (i.e., small variance) does not offer as much information
about the characteristic the variable is attempting to capture as a vari-
able that has more variability. In other words, in most cases, the more
recorded variability, the more information available to answer research
questions. This is not to say that variables with large variances are always
preferred over variables with smaller variances, just that scale (i.e., con-
tinuous) variables with minimal variances do not contribute as much to
the analysis. Therefore, before using any scale variable, you should review
the underlying concept you are measuring to investigate if the observed
variance has theoretical support, and if there is enough variance to war-
rant its inclusion in an analysis.
For categorical and dichotomous variables, the quality of informa-
tion is related to coverage in the variable categories. The more cover-
age, the more information you have for the analyses. For example, if
using a dichotomous variable for ethnicity (i.e., a nominal categori-
cal variable) and the data set contains 30 cases (i.e., a small data set)
with 27 Hispanic/L atino(a) people and three non-Hispanic/L atino(a)
people, you have much more information for Hispanic people than
for non-Hispanic people. In this situation, you must decide if you
have enough information to consider differences between the two
Analyzing Single Variables and Single Groups 29
ethnicities. Similarly, consider a categorical variable that represents

a 4-point Likert-t ype scale (i.e., an ordinal categorical variable) with
response options strongly disagree, disagree, agree, and strongly agree.
In this situation, if, for example, only two respondents selected agree
and the other respondents were equally distributed among the three
other response options, the variable offers very little to no informa-
tion on people who agree. Therefore, one option is to review carefully
the construct of the variable and to decide if collapsing the responses
into disagree and agree has theoretical support. After creating the
new collapsed variable, you should run the analysis twice —first on
the noncollapsed variable and then on the collapsed variable —to
identify how the results differ between the two versions of the varia-
ble. The difference between the two analyses will provide insight into
how to proceed.
Software packages that facilitate the analysis of variables do not
take into consideration the quality of information captured by a vari-
able. Therefore, the responsibility falls to you, the researcher, to deter-
mine when the information provided by a variable is too limited to
use in a statistical test. When using categorical variables (e.g., ana-
lyzing a 2 × 2 table that represents two dichotomous variables), one
frequently used guideline (dating from 1952) dictates that a catego-
rical variable should have a minimum of five respondents represent-
ing each category before using the variable in an analysis (Cochran,
1952). On face value, this makes sense because any category that has
fewer than five responses does not have much representation within
the variable. In the example above involving ethnicity, three non-
Hispanic people do not provide enough information to effectively use
the variable Ethnicity in a statistical test. Categorical variables with
a small number of respondents for a category are especially problem-
atic for parametric procedures involving group comparisons (e.g., t-
tests, ANOVA). The lack of information limits the ability to verify that
the data meet the assumptions for parametric procedures. However,
the lack of abundant information for non-Hispanic people is not as
problematic for nonparametric procedures, because nonparametric
procedures use ranks and order to test hypotheses. Therefore, the
likelihood of a Type I error (reject a true null hypothesis) is much
lower when conducting nonparametric tests. This is not to say that
using nonparametric procedures releases a researcher from fully
understanding the information quality within each variable. A thor-

ough review of the variable characteristics is still the first and most
important step in analyzing data.
You have two major reasons for looking at variables individually.
The first reason is understanding the quality of information captured
by variables in a data set. The second reason relates to investigating
research questions that can be answered with a single variable. Multiple
nonparametric procedures provide the opportunity to run statistical
tests on a single variable.
For most people learning how to use statistics, it is often helpful to
provide a framework or context for the discussion. This is especially true
when the discussion is your initial introduction to new statistical proce-
dures. Therefore, each chapter begins with a scenario describing a hypo-
thetical research line of study that can be explored with research questions.
Box 2.1 Research Scenario for Chapter 2
Dr. Chonody has just completed a study investigating the coming out process
for people who are gender and sexually diverse. She conducted extensive,
semi-structured interviews with 29 LGBTQ youth who were recruited through
websites and newspaper advertisements. Dr. Chonody used grounded theory
and narrative approaches, audio taping the interviews. The tapes were
transcribed and then analyzed using a combination of grounded theory
and cross-case procedures. She ensured the quality of the data through peer
debriefing, independent coding into key categories, and member checking. After
analyzing the qualitative data, Dr. Chonody asked the respondents to fill out
a paper-and-pencil questionnaire that contained several validated measures
that could confirm or extend her findings. In combination, the qualitative and
quantitative data provide improved validity and a more complete picture of
the respondents. Dr. Chonody has hired you as her statistical consultant. Her
research questions follow.
Six research questions guide the presentation of nonparametric pro-

cedures related to the above scenario. Each research question describes
the reasoning for using nonparametric procedures. Table 2.1 lists the
nonparametric procedures presented in Chapter 2, the types of vari-
ables for each procedure, and the names of the variables found in the
database for Chapter 2, which is available online.
Table 2.1. Summary of Analyses and Variables in Chapter 2

Research Nonparametric Test Variable Measure Variable Name
Question # in DB
1 Binomial Dichotomous or Confidante

scale
2 Chi-square Nominal or Coming_out
ordinal
3 Kolmogorov-Smirnov Scale Loneliness
4 Wilcoxon signed-rank Scale Coping_level
5 Runs Dichotomous, Confidante and
nominal, Rater_score
ordinal, or scale
6 Spearman’s rho and Scale Age, Self_esteem,
Kendall’s tau-b and Coming_
coefficients Out
Box 2.2 Research Question #1

After analyzing her qualitative data, Dr. Chonody believed that only 60% of
the youths had a confidante with whom to discuss the coming out process.
However, she did not ask that question explicitly in her interviews — she
simply inferred this from her qualitative data. Dr. Chonody examined this
question explicitly by asking it on her follow-up questionnaire. After she
collects the data from the follow-up questionnaire, Dr. Chonody asks you to
test her inferred percentage.
Research Question #1
What proportion of youths have a confidante?
BINOMIAL TEST
The binomial test is used for testing the statistical significance of possible
differences between an expected proportion and an observed proportion.
The binomial test is called an exact test because it provides the exact prob-
ability that the data represent the stated expected proportion. A compa-
rable parametric test is a t-test that uses z scores to test a hypothesis of an
expected proportion, but it is used rarely for proportion testing outside of
introductory statistics courses. The assumptions associated with the bino-
mial test are that the variable used in the test is dichotomous —two cat-
egories representing success and failure —or can represent a dichotomous
situation by identifying what values represent success and what values

represent failure with at least five values for each of the two categories.
The strengths of the binomial test are that it provides the exact probabil-
ity and, unlike the t-test, it does not assume any associated distribution.
However, like many nonparametric procedures, the binomial test has less
power to find statistical significance compared to the other tests with an
assumed distribution (if the data meet the distribution assumption).
Other examples of binomial tests are as follows:
1. You recently opened a new health clinic, and you know the
percentage of people with STDs in your city, so you want to test
if the proportion of people coming to your new clinic is similar
to the city percentage or significantly different.
2. A treatment center’s administrators believe that 90%
of their patients are satisfied with their care, and you
suspect that the administrators are overestimating the
satisfaction level.
Prior to conducting any hypothesis test, you must examine the informa-
tion provided by the variables. For Research Question #1 (see Box 2.2),
the variable is dichotomous, representing two nominal categories —
having a confidante or not having a confidante. For dichotomous vari-
ables, the easiest way to examine the variable information is to run a
frequency of the responses.
SPSS Process
To begin your analysis
□ Select Analyze => Descriptive Statistics => Frequencies to

open the Frequencies window.
□ Click to select the variable Confidante in the list of variables on
the right (i.e., Do you have a confidante?).

□ Click the move arrow to move the variable name over to
Variable(s): area.
Because the interest is only in frequencies, leave all the options for the
buttons on the right in their default position.
□ Make sure the Display frequency tables option at the bottom of

the Frequencies window is selected.
□ Select => OK at the bottom of the window.
SPSS Output
Two tables will appear in the IBM SPSS Statistical Viewer window
(i.e., analysis output window).
• The first table shows the number of valid and missing cases
found in the data set. Use this table to verify that the table
matches the number of cases intended to be part of the analysis.
• The second table shows the frequency at which the 28 youths
report having a confidante (see Table 2.2 —Frequency Table for
Research Question #1).
Table 2.2. Frequency Table for Research Question #1, Do you have a confidante?

Frequency Percent Valid Percent Cumulative Percent
Valid No 8 27.6 27.6 27.6

Yes 21 72.4 72.4 100.0
Total 29 100.0 100.0
Table 2.2 shows that 21 out of the 29, or 72.4% of the youths, have
a confidante. Therefore, do the data collected in the follow-up support
Dr. Chonody’s inference that 60% of participants had a confidante? At
first glance, it appears that they do not (60% is too low), but a rigorous
researcher will test further to be sure.
You might be tempted to utilize a parametric test like a t-test using
z scores, despite a sample size of fewer than 30. If you run a one-sample
test in SPSS (i.e., Analyze => Compare Means => One Sample
t-
T-test…), the 95% confidence interval for the percentage is .55 to .90
(55% to 90%). This would support Dr. Chonody’s inference that 60% of
the youths have confidantes. However, because the sample has fewer
than the sample-size guideline, you should use a nonparametric proce-
dure called a binomial test, which is well suited for this analysis because
of the ranking approach used in testing hypotheses.
SPSS Process
To begin this analysis,
□ Select => Analyze => Nonparametric Tests => One Sample.
Once the One-Sample Nonparametric Tests window opens, SPSS needs

to know which variable to analyze.
□ Select => Fields tab.

□ Click to select the Use custom field assignments option.
□ Click the variable you want to examine by highlighting Do you
have a confidante?
□ Click the move arrow to move the variable over to the Test
Fields: area.
Note: Depending on how you selected the last nonparametric test

options, some or all the available variables may show up in the Test
Fields: area on the right side of the window. If this is the case, simply
select all the variables except the Confidante variable in the Test Fields:
area and then use the move arrow to move them to the Fields: area
on the left side of the window.
□ Select => Settings tab to select test options and which

nonparametric test to run.
The Settings tab in the One-Sample Nonparametric Tests window

is where you can select which nonparametric test to run, where you can
set different test options, and where you set how you want to handle
missingness. Selecting the different items listed in the Select an item:
area gives you access to these choices.
□ Select => Test Options in the Select an item: area if you wish
to change the significance level from the default (0.05) for a test,
or if you wish to change the confidence interval range from
the default (95.0%). Here you can also change the missing data
default (exclude cases test-by-test).
□ Select => Choose Tests from the Select an item: area to see two
options for choosing tests.

The first option (Automatically choose the test based on the data)
allows SPSS to decide automatically which tests are appropriate based
on the recorded SPSS variable definition (i.e., defined level of meas-
urement for each variable). Because it is important to follow the anal-
ysis process closely to understand fully how the analysis is performed,
click to select the Customize tests option to gain access to the test
selection boxes.
□ Click the first option in the list (i.e., binomial test). Figure 2.1
shows the One-Sample Nonparametric Tests window with the
test selected that matches Research Question #1.
Figure 2.1. One-Sample Nonparametric Tests window.
Shown in Figure 2.1 is the list of different tests available for analyz-
ing a single variable. The icons shown on the right side of the screen
represent the tests that are appropriate for each measurement level. For
Research Question #1, the binomial test is appropriate because it com-
pares a hypothesized proportion of responses to the observed propor-
tion of responses in the variable categories. For a dichotomous variable,
the default is that half of the respondents selected one category while
the other half selected the other category (i.e., successes are as equally
likely to occur as fails, or 50% success and 50% fails). However, Research
Question #1 proposes a 60% rate for yes responses, so
□ To change the default, select => Options… to open the

Binomial Options window.
The Binomial Options window provides an opportunity to change the

expected or hypothesized proportion, request confidence intervals, iden-
tify which category value to use for the hypothesized proportion, and
define how to group low vs. high values if analyzing a scale variable.
Binomial Options Decisions

1. Hypothesized Proportion —You can identify any hypothesized
proportion value between 0 and 1 to test against your data. The
SPSS default hypothesized proportion value is 0.5.
2. Confidence Interval —SPSS provides the option of three
different calculation methods to estimate a confidence interval
for likely proportions based on the data. You may have an
occasion in which obtaining a confidence interval would be
helpful (e.g., wanting to identify a plausible percentage range of
youths with confidantes). When you do, select all three options
and compare their results.
3. Define Success for Categorical Fields —SPSS uses the terms
success and failure to identify values in binomial tests. If a value
for success is not provided, SPSS will use the first value it finds in
the variable as success. Managing what values are designated as
success and failure will help in explaining your results.
4. Define Success for Continuous Fields —The binomial test is
very flexible, and SPSS provides a way to conduct the test using a
scale variable. This option decision allows you to identify a range
of values that represent success and values that represent failure.
For Research Question #1, 0.60 is used as the hypothesized propor-

tion and 1 as the value for success (i.e., Yes —youth has a confidante).
□ Click in the Hypothesized proportion: area and change the

value from .5 to .6.
□ Click all three options under Confidence Interval (i.e., Clopper-
Pearson, Jeffreys, and Likelihood ratio) so you can compare the

results from the three different ways of calculating confidence

intervals.
□ Click the Specify success values option for Define Success for
Categorical Fields.
□ Click in the Success Values: white area and enter a 1 (to indicate
1 as the success value).

□ After you review your choices, select => OK to close the
window.
□ Select => Run to conduct the analysis.
See Appendix A for complete SPSS syntax.
SPSS Output
A variety of information shows up on the output window, including:
• Syntax used by SPSS (i.e., log)

• Location of the data set used (i.e., active data set)
• Hypothesis Test Summary (i.e., link to Model Viewer)
• List of items in the output index on the left side of the
output window
Clicking on an item in the list initiates a scroll to that information

on the right side of the output window. To see the details of the test,
double-click the Hypothesis Test Summary information (i.e., point the
mouse within the Hypothesis Test Summary table and double-click).
The Model Viewer window will open to display the detailed informa-
tion (see Figure 2.2).
The Model Viewer window has access to more information than
what is shown in Figure 2.2.
• At the bottom of the window are dropdown lists on the left

for View and Field Filter, and on the right for Test, Field(s),
and View.
• Because Research Question #1 involves only one variable
(i.e., you moved only one variable to Test Fields:), some of
the dropdowns do not offer multiple options. Changing the
dropdown selection for View: on the left from Hypothesis
Figure 2.2. Model Viewer window with detailed test results.
Summary View to Confidence Interval Summary View

displays the lower and upper values for the three selected
confidence interval calculation methods.
Findings
When you set the View in the Model Viewer window to Hypothesis
Summary View, it displays the results of the hypothesized 60%. The Sig.
value is .120, which fails to reject the hypothesis that the proportions of
with and without confidantes are .60 and .40, respectively. Whenever
you run a statistical test, you want to verify that your settings in SPSS
were as you intended. You can verify your settings by closely reviewing
the wording in the tables, values of the confidence intervals, the look of
the histograms, and the information below the histograms, etc. You can
find all this information by selecting the different dropdown options in
the Model Viewer.
Next, select the dropdown for View: on the left side of the window
and select Confidence Interval Summary View. The values under the
Lower and Upper columns show the boundary values for the 95% con-
fidence interval for the percent success identified in this data set. The
calculation methods are slightly different among the three, which is why
the values are slightly different.
In statistical terms, you can say a binomial test of the null hypoth-
esis that the percent of youths with confidantes is equal to 60% fails to
be rejected at an alpha level of .05 with a p-value of .120. Another expla-
nation from the results is that if the youths responding to the follow-up
survey are typical for gay youths, then there is a 12% chance that the pro-
portion of youths who have confidantes is 60%. A Clopper-Pearson cal-
culation provides a 95% confidence interval for proportion rate of 52.8%
to 87.3%. In this analysis, selecting 95% for the confidence interval pro-
vides a range of possible percentages of youths with a confidante in which
we are 95% confident that the true percentage falls. If 90% was selected,
the range would be a bit wider, but the confidence that the true percent-
age falls within the range would be smaller. Note that this finding differs
from that of the (inappropriate) parametric t-test mentioned earlier that
found lower and upper bounds of 55% to 90%.
Practically, the findings suggest that, for this study, the number of
gay youths with a confidante is a bit higher than the number without a
confidante. For example, if you had a group of 100 gay youths who were
similar to the youths in Dr. Chonody’s study, the number of youths with
a confidante would be somewhere between 53 and 87.

In her qualitative data, Dr. Chonody did not see a general pattern in the
degree to which her participants had come out, but she wanted to test this with
a survey question asking participants the number of people to whom they had
disclosed their sexual orientation. After reviewing this variable’s distribution,
Dr. Chonody found that the responses fell into four main categories — coming
out to no one, to one person, to a few people, and to everyone. However, she
did not have enough responses in the two middle categories (i.e., fewer than
five responses in “to one person” and “to a few people” categories), so she
collapsed those responses into one category. She now has a variable with three
categories of responses — coming out to no one, to one or a few people, and to
everyone. Dr. Chonody expects no differences in these three categories, but she
wants you to test this statistically.
Are the proportions equivalent among youths coming out to a single
confidante, to multiple confidantes, or not coming out at all?
CHI-SQUARE TEST
Different kinds of chi-square tests address different data analysis sit-
uations (e.g., 2 × 2 contingency table analysis, Pearson chi-square,
likelihood-ratio chi-square, etc.). Basically, the chi-square test is suited
for testing possible proportion differences between an expected propor-
tion and an observed proportion when more than two combinations of
categories are possible. When proportions among categories are exam-
ined, this is similar to investigating whether the numbers within each
category are similar. The kind of variable used in a chi-square test in
this situation is either a nominal or an ordinal variable (i.e., categorical
variable representing two or more responses —e.g., none, sometimes,
or often for each respondent). The assumptions associated with the chi-
square test are that the variables used in the test are categorical and that
responses are independent. In addition, each category should be repre-
sented with at least five responses to allow for appropriate information
coverage for the respondent categories. The strengths of the chi-square
test are that the results are easy to interpret and you can apply the test to
all kinds of categorical variables. One limitation is that the chi-square
test is much more sensitive to sample sizes and often requires a larger
sample to find significant proportional differences; however, if the sam-
ple is too large, the test finds significance too easily. Lastly, no alter-
native parametric procedure is possible without manipulating the data
into something testable with a t-test.
Other uses for the chi-square test are as follows:
1. You are examining the data from a pilot study, and you want
to investigate the proportions of people subscribing to different
religions (i.e., the expected vs. observed).
2. In your small sample of gay youths, you wish to explore your
suspicion that only 10% have come out to no one and the other
90% are evenly distributed between coming out to one or a few
and coming out to everyone.
The binomial test investigates dichotomous variables, but the chi-

square test examines if the response proportions are equal across three
or more categories, which you need to answer Research Question #2.
The SPSS steps, however, are similar. For Research Question #2, the
variable Coming_Out is operationalized with three categories (1 = no

one, 2 = one or a few, and 3 = everyone).
SPSS Process
To begin your analysis,
□ Select => Analyze => Nonparametric Tests => One Sample to

open the One-Sample Nonparametric Tests window.
When the One-Sample Nonparametric Tests window opens, SPSS needs

to know which variables to use in the analysis and which analysis rou-
tines to run.
□ Select => Fields.

□ Click to select Use custom field assignments option.
□ Click to select Coming_Out from the list of variables in the
Fields: area.
□ Click the move arrow to move Coming_Out over to Test
Fields: area.
Note: If other variables are listed in the Test Fields: area, select them and
move them back to the Fields: area on the left side of the window.
Now that SPSS knows what variables to use,
□ Select => Settings tab to see the list of available

nonparametric tests.
□ Select => Choose Tests in the Select an item: area.
□ Click to select Customize tests and then click to select the
checkbox for the Chi-Square test.
In Figure 2.1, the variable measure icons indicate the variables with
the appropriate level of measurement. For the chi-cquare test, the levels
of measurement are nominal and ordinal, and the icons change to color
when you check the Chi-Square Test checkbox.
The options for the chi-square test are different from the binomial
test. The only time a change to the default options is necessary is when
you are testing the category proportions for something other than being
equal across all categories. Because Research Question #2 is testing for

equal proportions, the chi-square test options can remain as the default
(i.e., all categories have equal probability).
□ After carefully reviewing your choices, select => Run to conduct

the analysis
SPSS Output
To see the details of the test, double-click the Hypothesis Test Summary
information (i.e., point the mouse within the Hypothesis Test Summary
table and double-click). The Model Viewer window will open to display
the detailed information.
Findings
The null hypothesis for the chi-square test for Coming_Out is that equal
proportions of responses exist for the three categories. When the Model
Viewer window first opens, it presents a histogram on the right and the test
results on the left. The Sig. value for Coming_Out indicates that the test fails
to reject the null hypothesis (i.e., Sig. value = .185 is larger than the default
alpha level of .05), thereby suggesting that there is not enough information
to identify a proportional difference among youths’ Coming_Out catego-
ries. Another explanation is that the proportions are similar enough to not
be identified statistically as significantly different. In statistical terms, you
may say that the chi-square test for a null hypothesis of equal proportion of
youth Coming_Out categories fails to be rejected at α = .05, with p = .185.
However, when you examine the histograms for Coming_ Out
among gay youth categories (i.e., no one, one or a few, and everyone),
they seem to show possible differences despite not being significantly
different, statistically. More than likely, if the sample size were larger
and followed similar proportions, there would be enough power to
identify differences. This is one important reason always to include as
many respondents in your study as possible, to increase the likelihood of
finding statistical differences. In this situation, you do not have enough
information from the variables to identify statistical significance; that is,
the statistics are keeping you from drawing a potentially wrong conclu-
sion based on viewing the histograms only.

Dr. Chonody believed, after reviewing the qualitative data, that her respondents’
feelings of loneliness may have important relationships with other issues in
their lives. To investigate this possibility, one of the questions on her follow-up
survey asked respondents how frequently they feel lonely. The response options
were never, sometimes, often, and all the time. Dr. Chonody wants to use the
responses to this question as a continuous variable for further analysis, but
she isn’t sure if this ordinal-level variable sufficiently approximates a normal
distribution, so she asks you to test this for her.
Should a categorical Likert- type variable with four response options
representing levels of loneliness be treated as a continuous variable
representing a normal distribution in a statistical analysis?
KOLMOGOROV-SMIRNOV TEST
The Kolmogorov-Smirnov test examines how well a distribution from a
scale variable matches a probability distribution. In SPSS’s menu struc-
ture for this nonparametric procedure, you can test the variable against
four different types of distribution —normal, uniform, exponential,
and Poisson. The normal distribution is most often selected because
of the applications of the normal distribution to parametric assump-
tions. On the parametric side of SPSS’s menu structure, the option to
use the Kolmogorov-Smirnov test exists only for the normal distribution
(Normality plots with tests option within the Explore: Plots window).
You can consider the Kolmogorov-Smirnov test a type of goodness-of-fit
test, because it examines how well the data conform to the selected dis-
tribution. For information on distributions, see http://pages.stern.nyu.
edu/~adamodar/New_Home_Page/StatFile/statdistns.htm. (Be sure not
to include the concluding period in the web link.) The kind of variable
used in a Kolmogorov-Smirnov test is a scale variable, a continuous vari-
able representing responses along a range of potential values —e.g., Age,
Burnout level, etc. If you want to examine a Likert-like variable to see if it
follows a normal distribution, you will need to redefine its level of meas-
urement to scale before attempting to perform the test.
Other applications for the Kolmogorov-

Smirnov test are as
follows:
1. From your pilot study, you wish to analyze a scale/continuous

variable to determine if it is represented by any of four
distributions (i.e., normal, uniform, exponential, or
Poisson).
2. You have conducted a pilot for a larger survey, and you wish to
test if the ages of respondents follow a normal distribution with
a mean of x and a standard deviation of y.
Use great caution when deciding to treat a Likert-like categori-

cal variable as continuous. A continuous variable has the charac-
teristic that all values are possible between each response point and
within a specific range supported by the variable’s conceptual design.
Altering the theoretical construct of a categorical variable should not
be taken lightly and must be preceded by a close review to verify that
the change does not disrupt how the variable represents the initial
concept.
One important component of a close review includes the exam-
ination of the potential distributions of the categorical variable. The
Kolmogorov-Smirnov test offers a convenient way to look at a vari-
able’s distribution. However, the Kolmogorov-Smirnov test will work
only with scale variables. Therefore, for Research Question #3, you
must change the definition for the variable How frequently do you
feel lonely? temporarily from ordinal to scale within SPSS’s variable
view area.
SPSS Process
To change the level of measurement for a variable
□ Select => Variable View tab at the bottom of the IBM SPSS
Statistics Data Editor window.
□ Find the Measure column and the row for Loneliness, and then
click on the Ordinal icon to show the dropdown list.

□ Select Scale. (Note: If Scale is already selected, then you have
no need to make a measure change.)
To begin the analysis,
□ Select=> Analyze => Nonparametric Tests => One Sample.

□ Click to highlight the Loneliness variable and then click the
move arrow to move it to Test Fields: area.
Note: Another method for moving variables from one side of the win-
dow to the other side is to point your mouse at the variable name and
double-click. SPSS interprets the double-click as a message to move the
variable to opposite area.
Now that SPSS knows what variable to use,

□ Select => Choose Tests to highlight it in the Select an item: area.
□ Click to select the Customize tests option and then click to
select the checkbox for the Kolmogorov-Smirnov test (uncheck

other Customize tests options if checked).
□ Select => Options… to open the Kolmogorov-Smirnov Test
Options window. The options allow you to test your variable

against four different hypothesized distributions.
□ Click to select all four distributions (i.e., Normal, Uniform,
Exponential, and Poisson) to compare all four potential

distributions rather than just testing for normality.
□ For Research Question #3, use the default (i.e., the Use sample
data/Sample mean) for each of the four distribution mean and

parameters options.
Note: The decision to use the default is based on the details in Research
Question #3, but in other situations the default may not be appropriate. For
example, if the desire was to investigate if a Likert-like variable can be from
a normal distribution with a mean of 5 and a standard deviation of 1, then
changing the Distribution Parameters for the Normal option is accom-
plished by clicking on Custom and entering the Mean and Std Dev. values.
□ Select => OK at the bottom to close the Kolmogorov-Smirnov

Test Options window.
□ After reviewing your selections, select => Run at the bottom of

the window to begin the analysis.
SPSS Output
Shortly after selecting Run, SPSS’s output window will display and
show the Hypothesis Test Summary results for the analysis. Because all
four distributions were selected to test against the distribution for the
Loneliness variable, four results are presented in the summary.
• To see the details of the test, double-click the Hypothesis

Test Summary information (i.e., point the mouse within the
Hypothesis Test Summary table and double-click).
• The Model Viewer window will open to display the detailed
information.
• When the Model Viewer window opens, the first test for the
normal distribution is highlighted in the list of tests shown
on the left side of the window, and the corresponding test
information is displayed on the right. Some of the corresponding
test information is a histogram with an overlay of a normal
distribution using the sample data to establish parameter values.
Findings
According to the Kolmogorov-Smirnov test, there is sufficient informa-
tion to conclude that the data for the Loneliness variable do not come from
a normal distribution. A review of the graph clearly shows a highly right-
skewed distribution for Loneliness. Recall, one of the challenges of using
nonparametric procedures is lower power in rejecting the null hypothesis,
but in this case, it does not seem to be even close with a Sig. value less than
.001. Even so, the decision to use an ordinal variable as continuous (i.e.,
scale) should not be based on a single statistical test alone.
Clicking on the other test summaries on the left side of the window (i.e.,
clicking within the Hypothesis Test Summary table) shows the correspond-
ing information for each test. A review of the histograms and the hypoth-
esized distribution shows the level of conformity of the Loneliness data
to the uniform, Poisson, and exponential distributions. The Kolmogorov-

Smirnov test found that the Loneliness data did not reflect uniform or expo-
nential data (i.e., Sig. values less than or equal to .001), but did suggest a
distribution that followed a Poisson distribution (i.e., Sig. = .836). In more
statistical terms, the test found that there is less than a 5% chance that the
Loneliness data follow a normal distribution if the null hypothesis is true.
For Research Question #3, although the Kolmogorov- Smirnov
test did reject the normal distribution, a review of the histograms for
Loneliness provides enough information that you should not use it as a
continuous variable with an assumption of normality. However, if you
had an opportunity to use Loneliness in a statistical model in which it
would represent a Poisson distribution, it would be acceptable.
Practically, you should take great care when choosing to use an ordi-
nal variable with Likert-like response options as continuous. If you have
any doubts about the ordinal variable’s ability to represent a normally
distributed set of values, then do not use the ordinal variable in a test
that assumes normality.

Based on her interviews, Dr. Chonody believes that her respondents have
an average ability to cope with stressful events (e.g., coming out). To test this
hypothesis, she investigates her respondents’ scores on a validated measure of
coping with stressful events. An average score would be 15 out of 30 possible,
and Dr. Chonody would like you to test the likelihood that a score of 15
represents her respondents’ median score.
Is the median score 15 for these respondents’ ability to cope with stressful events?
WILCOXON SIGNED-R ANK TEST

The Wilcoxon signed-rank test is used for many different tests, but in
SPSS under the One-Sample Nonparametric Tests, the Wilcoxon signed-
rank test examines the probability that a specified median value is
possible for a target population given the available data from a single
continuous variable. Therefore, the Wilcoxon signed-rank test examines
the likelihood that a specific value represents a median value for the data.
When examining small samples, the use of a median value is often pre-
ferred over the use of a mean value because a single response can change
the mean score dramatically in a small sample (i.e., an extreme score can
overly influence a mean calculation). When outliers are present, using
a median value is preferred because this removes the outliers’ excessive
influence on the analysis. However, if no outliers are present, the mean is
the preferred choice.
The parametric t-test is similar to the Wilcoxon signed-rank test,
except that the t-test uses a specified mean value rather than a median
value. The kind of variable used in a Wilcoxon signed-rank test is a scale/
continuous variable representing responses along a range of potential val-
ues —e.g., Age, Burnout level —based on a summed scale score from
multiple items in the measure, etc. The Wilcoxon signed-rank test is
the test to replace the parametric t-test when you cannot verify the nor-
mality assumption. However, like many nonparametric procedures, the
Wilcoxon signed-rank test has less power to find statistical significance
than the t-test when the data meet the normality assumption for the t-test.
Other uses for the Wilcoxon signed-rank test are as follows:
1. You are conducting a small study of a new weight-loss

intervention, and you want to determine if the average weight
loss is equal to or greater than 5 pounds.
2. Before starting a new therapy group, you wish to learn if the
median level of self-esteem for your future clients is similar to,
or different from, that of people in general.
For Research Question #4, the continuous variable from the data set is
Coping_Level. The Coping_Level variable in the data set represents a
scale score from a validated measure that examines coping levels with
stressful events. Twenty-nine youths completed the validated measure,
and their responses were used to set the values for Coping_Level.
SPSS Process
To begin your analysis
□ Select => Analyze => Nonparametric Tests => One Sample.


□ Click to highlight the Coping_Level variable in the list of
variables in the Field: area.
□ Click the move arrow to move it over to Test Fields: area.

□ Select => Choose Tests in the Select an item: area.
□ Click to select the Customize tests option to gain access to the
list of tests.
□ Click the checkbox for the Wilcoxon signed-rank test.
□ Click in the Hypothesized median: area to place focus on the
white space, and enter the number 15.
Unlike the other test examples, the Wilcoxon signed-rank test does not
have any options —it requires only the identification of the hypothesized
median value. The level of measurement icon to the right indicates that
this test is only for interval or scale variables. Therefore, if the variable
is nominal, testing for a hypothesized median value is not appropriate.
□ After carefully reviewing your choices, select => Run to begin

the analysis
SPSS Output
To see the details of the test, double-click the Hypothesis Test Summary
information (i.e., point the mouse within the Hypothesis Test Summary
table and double-click). The Model Viewer window will open to display
the detailed information.
Findings
The histogram displayed on the right side of the Model Viewer
window shows the relative distribution of the Coping_L evel variable,
the hypothetical median used in the Wilcoxon signed-rank test, and

the observed median. A review of this histogram does not show any
significant difference between the hypothetical and observed median,
which indicates that no information supports the belief that the youths
as a group have a different expected level of coping with stressful situa-
tions (i.e., Sig. = .537, which is larger than the α = .05 threshold).
In statistical terms, the Wilcoxon signed-rank test fails to reject the
null hypothesis that the median value for Coping_Level is 15 (p = .537)
at α = .05.
Practically, the only way to know for sure that a set of values does
not have a specific median value is to have enough evidence to be cer-
tain to think otherwise. A statistical test is one method that contributes
to identifying what “enough evidence” means. In this situation, the data
show that a median value of 15 is possible.

Dr. Chonody had four experts in this substantive area review her semi-
structured interview questions before she began her study, asking them to rate
each question on a 1 to 10 scale for how well it addressed an important issue.
When she collated the experts’ ratings, she noticed what she thought was a
pattern in one of the rater’s scores. To test this possibility, she wants you to
conduct a runs test on the rater’s sum scale scores. Of course, Dr. Chonody
placed the questions randomly in a list prior to expert rating to minimize the
possibility of question quality order.
Are the values in the experts’ ratings showing a pattern (i.e., not random)?
RUNS TEST
The runs test, also called the Wald-Wolfowitz test, provides an opportu-
nity to see if the values in a variable are random. The runs test in SPSS
is constructed around a dichotomous variable, but SPSS can examine
other levels of measurement variables for randomness by identifying two
groups within the data. Randomness associated with testing for runs is
the hypothesis that the order in which the two values in a dichotomous
variable appear does not show any kind of pattern (i.e., the two values in
the dichotomous variable appear at random). In statistics, many different
procedures can test for randomness in all different kinds of variables,

because knowing when a sequence of numbers is random or dependent
is extremely important when analyzing data. However, in SPSS the runs
test is the only procedure that directly tests for randomness in a variable.
Other examples of a Runs test are as follows:
1. A supervisor hopes that her caseworkers’ client satisfaction

surveys vary in a random way. However, she suspects that they
are more negative at the end of the day, or that they are more
positive based on the specific work group.
2. The administrator of a small organization wants to know if the
number of people who do not show up for their appointments is
greater toward the end of the month.
For Research Question #5, the variable from the dataset is Rater_
score.
Part 1 of 2
Because an investigation for randomness may be necessary with any
level of variable (i.e., nominal, ordinal, or scale), a few additional steps
are sometimes required before actually running the tests. For exam-
ple, because SPSS defines Rater_score as a scale variable, you must
first identify a cut point —the value used to demarcate the two sides
on which the variable values fall. The cut point, sometimes called a
threshold, will be used to identify how many times the variable values
switch from above to below (or vice versa) the cut point value, which
is the method used by the Runs test to determine if the values are not
necessarily random.
One way to identify the cut point is to run a descriptive analysis on
Rater_score.
□ Select => Analyze => Descriptive Statistics => Descriptive to

open the Descriptives window.
□ Click to highlight the Rater_score variable in the list of variables

□ Click the move arrow to move it over to the Variable(s): area.
□ Select => OK button to run the descriptive analysis.
The analysis reveals that the mean = 16.45 and the standard devia-
tion = 9.482, so you must next decide if 16.45 is a practically reason-
able threshold for this measure before using it for your analysis. Using
the mean value, as you did in this procedure, is a statistical option for
determining a cut point value. However, other options are available to
you. For example, you can utilize a cut point commonly found in the
literature (e.g., a cut point of 16 for the CESD depression scale). Another
option is to review the variable’s distribution to see if you can identify
visually a cut point value. When you have identified your cut point/
threshold, you may now conduct the Runs test.
SPSS Process
After you identify your cut point, continue your analysis:

□ Select => Fields tab to display the list of available variables.
□ Click on Use custom field assignments option.
□ Click to highlight Rater_score in the Fields: area.
□ Click the move arrow to move Rater_score to the Test
Fields: area.

□ Click to select Customize tests option and then click the
checkbox for the Runs test (uncheck other Customize tests

options if checked).
□ Select => Options… to open the Runs Test Options window
(see Figure 2.3 —Runs Test Window).
When the Runs Test Options window opens, you will notice that the
default for scale variables is to use the Sample median option. Because
the descriptive information was obtained for Rater_score and the mean
value was identified so it could be used in the test, you will need to make
a change.
□ Click to select Sample mean option in the Define Cut Pont for
Continuous Fields area.
Figure 2.3. Runs Test Options window.
Note: If a cut point other than the mean value is more appropriate for
the situation being tested, selecting the Custom option provides access
to the Cut point area, where a specific value can be entered.
□ Select => OK button at the bottom to close the Runs Test

Options window.
□ After carefully reviewing your choices, select => Run to begin
the analysis.

SPSS Output
Shortly after you select => Run, SPSS’s output window will display and
show the Hypothesis Test Summary results for the analysis. To see the
details of the test, double-click the Hypothesis Test Summary informa-
tion (i.e., point the mouse within the Hypothesis Test Summary table and
double-click). The Model Viewer window will open to display the detailed
information.
Findings for Part 1 of 2
The graph displayed on the right side of the Model Viewer window shows
a normal distribution. The distribution represents the probability of
observing the calculated number of runs based on the expected number
of runs given the selected cut point. The red line on the graph shows the
calculated number of runs and the vertical dashed line shows the expected
number of runs. Therefore, the farther away the red line is from the dashed
line, the less likely the observations for Rater_score are random.
The Runs test for Rater_score, using the sample mean as the cut point,
suggests that you do not have enough information to reject the possibility
that Rater_score values are random. In statistical terms, the Runs test of the
null hypothesis that the Rater_score values are random, based on the sample
mean of 16.448, fails to reject the null hypothesis at α = .05 (p = .577 in this
example).
Practically, patterns in values are sometimes difficult to recognize,
especially when the sample size is relatively small. In this situation, the
Runs test did not have enough evidence to suggest a pattern. If a pattern
does not exist, the natural conclusion is that the values are possibly ran-
dom, which is the case for Part 1. However, do keep in mind that another
explanation is that the small sample size simply did not have enough
information to recognize a pattern. Remember, use all the information
available (e.g., data distribution, graphs, categorical coverage, etc.) to
draw a conclusion for your analysis.
Box 2.7 Research Question #5, Part 2

After reviewing the results from Part 1 above, Dr. Chonody wants you to do one
additional test to check her assumption that the order of the youths reporting to
have or not have a confidante was random. If Dr. Chonody discovers that the
youths reporting to have a confidante is not random — for example, the first 15
report “no” and the next 15 report “yes” — then she will need to review other
aspects of her study (e.g., how she selected respondents, when the respondents
filled out their questionnaires) to learn the potential influences causing the
nonrandom respondent order.
Part 2 of 2
Unlike investigating randomness with a scale variable, the runs test on a
dichotomous variable does not require the selection of a cut point. With
the dichotomous variables’ two response options (e.g., yes or no, positive or
negative), the runs test will examine the number of times the values switch
from one to another. For Part 2 of Research Question #5, the dichotomous
variable to investigate is Do you have a confidante? (1-yes, 0-no).
To begin this part of your analysis,

□ Select => Fields tab to display the list of available variables.
□ Click to highlight Do you have a confidante? in the Fields: area.
□ Click the move arrow to move Do you have a confidante? to
the Test Fields: area.

□ Click on Customize tests option and then click the checkbox
for the Runs test (uncheck other Customize tests options if

checked).
□ Select => Options… to open the Runs Test Options window.
When the Runs Test Options window opens,
□ Verify that the option selected for Define Groups for

Categorical Fields is There are only 2 categories in the sample.
□ Once verified, select => OK button at the bottom to close the
Runs Test Options window.

□ When you have carefully reviewed your choices, select => Run
to begin the analysis
SPSS Output
Shortly after you select => Run, SPSS’s output window will display and
show the Hypothesis Test Summary results for the analysis. To see the
details of the test, double-click the Hypothesis Test Summary informa-
tion (i.e., point the mouse within the Hypothesis Test Summary table
and double-click). The Model Viewer window will open to display the
detailed information.
Note: If you left the Rater_score variable from Part 1 in the Fields:
area when Do you have a confidante? was added, the Hypothesis Test
Summary will present information on two analyses, one for Rater_score
and one for Do you have a confidante? Simply click on the second test
(i.e., the second line in the Hypothesis Test Summary area) to display
the results of the runs test for the dichotomous variable.
Findings for Part 2 of 2
For Confidante, the red line in the graph on the right is farther away
from the dashed line than it was for Rater_score, but the distance
is still not far enough to suggest that Do you have a confidante? is
not random. Clicking between the two tests in the Hypothesis Test
Summary area, if you tested both Rater_score and Do you have a
confidante? at the same time, will show the differences in distance
between the two lines, which is reflected in the Sig. values. For Rater_
score, Sig. = .577, and for Do you have a confidante? Sig. = .164.
Therefore, once again, you do not have enough information to reject
the possibility that youths who have a confidante is random (i.e., you
have no information to question Dr. Chonody’s assumption of youth
randomness).
Practically, it is not a good idea to assume something like ran-
domness and then not check to see if it is true. Part 1 was clear in that
there was not enough information to suggest a pattern in rater scores.
In Part 2, a variable somewhat related to the concern of patterns in
rater scores was tested for randomness. Neither variable indicated a

Based on the data from her interviews, Dr. Chonody suspected that, for her
participants who had not come out to anyone, the older they became, the
lower they would score on a self-esteem measure. However, for those who had
come out to at least one person, she hypothesized that the older they became,
the higher they would score on self-esteem. Dr. Chonody included a validated
measure of self-esteem on the follow-up questionnaire to test her hypotheses.
She grouped the responses into two groups — those who had come out to at
least one person, and those who had not come out to anyone. Because of the
small number of participants and the possibility of non-normal data, she
should not conduct a Pearson correlation to analyze the data. Dr. Chonody
asks you to conduct the analysis.
Does a relationship (i.e., correlation) exist between age and the youth’s level
of self-esteem, and does the relationship differ between youths who have not
come out and youths who have come out to at least one person?
possible pattern, so the findings indicate that there is no pattern in

question quality or youth order. Of course, the runs test works only on
unsorted data.
SPEARMAN’S RHO AND KENDALL’S TAU-B COEFFICIENTS

Spearman’s rho, also called the Spearman’s rank correlation coefficient,
and Kendall’s tau-b, also called the Kendall rank correlation coefficient,
are measures of the relationship between two scale variables. However,
unlike the commonly used Pearson correlation, these two coefficients
use ranking in the calculation of the correlation value, rather than
the assumed linear relationship used in Pearson correlation calcula-
tions. The reason for using Spearman’s rho and Kendall’s tau-b is to
avoid the restriction of the assumed linear association and because the
Pearson correlation does not effectively deal with outliers, departure
from normality, unequal variances, and nonlinearity. In addition, the
interpretation of the coefficient values for these two procedures is the
same as for the Pearson correlation —e.g., if y increases as z increases,
it is a positive relationship; if y increases as z decreases, it is a negative
relationship. Another strength of the two coefficients is that you can
include ordinal variables in the procedure. Note that SPSS will provide a
Pearson correlation value that involves an ordinal variable, but if either

of the more appropriate coefficients is requested, the Pearson correla-
tion is not included in the results.
The assumptions for Spearman’s rho and Kendall’s tau-b are that
the variables in the analyses are not nominal and that the participants’
responses are independent. In addition, if you are including an ordinal
variable, you should determine if you have enough categories to pro-
vide conceptual support for calculating a correlation value —e.g., does
a correlation with an ordinal variable with three categories make sense?
The assumptions for the Pearson correlation are the absence of outli-
ers, equal variances among variables, linear association, and bivariate
normality. If you are dealing with a smaller data set, verifying the mul-
tiple assumptions may make the decision to use the Pearson correlation
risky. When comparing the results of Spearman’s rho and Kendall’s tau-
b, keep in mind that Kendall’s tau-b is more efficient at adjusting for
matching values (i.e., ties in rank-ordered values) than Spearman’s rho.
Other examples are as follows:
1. In your small inpatient unit, you wish to learn if patients who

are more depressed have less self-esteem than those who are not
depressed, so you can improve treatment protocols.
2. You are analyzing data from a small pilot study, and you wish
to explore if general anxiety is associated with different levels of
socioeconomic status.
Note from SPSS: “Before calculating a correlation coefficient, screen

your data for outliers (which can cause misleading results) and evidence
of a linear relationship. Pearson’s correlation coefficient is a measure of
linear association. Two variables can be perfectly related, but if the rela-
tionship is not linear, Pearson’s correlation coefficient is not an appropri-
ate statistic for measuring their association.”
Note from authors: Research Question #6 illustrates a situation in
which the desire is to understand the relationship between two con-
cepts. However, it is always good practice to run Pearson correlations,
Spearman’s rho, or Kendall’s tau-b among all the continuous variables
used in a study. The primary reason is that these values will identify if
two variables are mimicking each other and thus duplicating the infor-
mation for a respondent. If you don’t fully understand how similar
two variables are, using two almost identical variables will bias the findings
and invalidate analysis results. Generally, examining bivariate and univari-
ate scatter plots for outliers and linearity is always a good practice when first
examining the data. Another reason to run correlations on all the variables
is to help validate the data, which provides another opportunity to find data-
entry mistakes or to catch an error in how a variable was operationalized.
For Research Question #6, the variables from the data set are Age
and Self_esteem. Because Dr. Chonody is interested in learning more
about the relationship between Age and Self_esteem, and whether this
relationship differs for youth who have come out, a good way to begin is
to examine the information contained in the two variables. You can do
this by having SPSS create histograms of their distributions.
SPSS Process
To request histograms,
□ Select => Analyze => Descriptive Statistics => Frequencies to

□ Click to select Age in the list of available variables on the left
side of the window, then click the move arrow to move the
variable to the Variable(s): area.
□ Click to select Self_esteem in the list of available variables on
the left side of the window, then click the move arrow to
move the variable to the Variable(s): area.
□ Select => Charts… to open the Frequencies: Charts window.
□ Click to select the Histograms: option and then click to select
the Show normal curve on histogram checkbox.

□ Select => Continue button to close the Frequencies: Charts
window.
□ Select => OK button at the bottom to close the Frequencies
window and generate the histograms.
Shortly after you select OK, SPSS’s output window will display a few
tables, and if you scroll down the window you will see the two histo-
grams for Age and Self_esteem. These two histograms show the variable
distributions for all the youths who completed the follow-up question-
naire. However, along with these two graphs, you want histograms for
the two groups of youths. To do this, you need to set a data filter in SPSS.
First filter to get only youths who have not come out to anyone, request
histograms, and then filter to get youths who have come out to one or
more people.
To filter respondents to select only the youths who have not come out
to anyone, and ignoring youths who have come out to one or more people,
□ Select => Data => Select cases… to open the Select Cases
window.
□ Click to select the If condition is satisfied option in the
Select area.
□ Select => If… button to open the Select Cases: If window.
□ Click to select the Coming_Out variable on the left side of the
window, then click the move arrow to move the variable to

the white space to the right of the move arrow.
□ Right after the variable name Coming_Out in the white space,
click to move the focus to right after the variable name.

□ Type =1 (i.e., an equals sign and a 1). This tells SPSS to use
cases that have a value of 1 for the variable Coming_Out.

Recall, a 1 represents youths who have not come out to
anyone.
□ Select => Continue button to close the Select Cases: If window.
□ Select => OK button to close the Select Cases window and use
only the cases that have a 1 for Coming_Out.
You can verify that the filter is working by looking at SPSS’s cases view
datasheet. You can identify the cases that have been filtered by the
diagonal line that appears across some of the line numbers on the left
side of the datasheet. Another way to verify that the filter is working as
expected is to review the value for N (i.e., number of cases used in the
analysis) on the notes in the histogram.
Now that the youths who have come out to one or more people have
not been included, you can rerun the steps for requesting histograms
(see “To request histograms”). After you have obtained the histograms
for your first group of youths, you need to change the SPSS data filter to
select the other group of youths.
To filter respondents to select youths who have come out to one or
more people, ignoring youths who have not come out to anyone,
window.
Select area.
The white area should still have the previous filter command (i.e.,
Coming_Out = 1), but, if not,
□ Click to select the Coming_Out variable on the left side of the

□ Right after the variable name Coming_Out in the white space,
Next, to filter youths who have come out to one or a few people (Coming_
Out values of 2) and youths who have come out to everyone (Coming_
Out values of 3), you need to use a “not equal to” symbol instead of the
equal sign. Therefore, you want the filter command in the white area
that will tell SPSS to use all the cases that have something other than a
1 for Coming_Out (i.e., Coming_Out ~=1, where ~= is the same as ≠).
□ Once you set the filter command to use all the cases that do not
have a 1 for Coming_Out, select => Continue button to close
the Select Cases: If window.
□ Select => OK button to close the Select Cases window and filter
out all the cases that do not have a 1 for Coming_Out.
Now that your filter is set, you can again rerun the request to get his-
tograms for the second group of youths by following the same steps as
before (i.e., steps under “To request histograms”).
SPSS Output
Figures 2.4, 2.5, and 2.6, show the histograms produced by completing
the steps above (i.e., Age and Self_esteem histograms for all the youths,
youths who have not come out to anyone, and youths who have come
out to one or more people). A review of the values for N in each of the
Age Self_esteem
10 Mean = 13.96 6 Mean = 14.25
Std. Dev. = 2.487 Std. Dev. = 3.855
N = 28 5 N = 28
8
4
Frequency
Frequency
6
3
4
2
2
1
0 0
10 12 14 16 18 20 5 10 15 20 25 30
Age Self_esteem
Figure 2.4. Age and Self_esteem histograms for all youths.
figures shows the number of youths represented by the histograms (i.e.,

28, 9, and 19, respectively).
The reason for reviewing all the histograms is to fully understand
the information captured by the two variables, Age and Self_esteem,
by the different categories for youths coming out. A review of the histo-
grams for Age does not show signs of a normal distribution. In addition,
the histogram for Self_esteem for all the youths appears to follow closely
a normal distribution, but when only youths who have not come out to
Age Self_esteem
4 Mean = 13.67 2.0 Mean = 15.67
Std. Dev. = 2.598 Std. Dev. = 3
N=9 N=9
3 1.5
Frequency
Frequency
2 1.0
1 0.5
0 0.0
10 12 14 16 18 20 10 12 14 16 18 20 22
Age Self_esteem
Figure 2.5. Age and Self_esteem histograms for youths who have not come out
to anyone.
Age Self_esteem
5 Mean = 14.11 4 Mean = 13.58
Std. Dev. = 2.492 Std. Dev. = 4.1
N = 19 N = 19
4
3
3
Frequency
Frequency
2
2
1
1
0 0
10 12 14 16 18 20 5 10 15 20 25 30
Age Self_esteem
Figure 2.6. Age and Self_esteem histograms for youths who have come out to
one or more people.
anyone are reviewed (i.e., Figure 2.5), the distribution appears to follow
a uniform distribution.
Another reason a close review of the data using histograms is
important is that a single statistical test alone is not enough to sup-
port a strong conclusion. For example, if the above distributions for
all respondents were tested for normality (don’t forget to remove the
filter) using the Kolmogorov-Smirnov test (see Research Question #3),
both normal and Poisson distributions are possible for Self_esteem.
The reality is that, when dealing with small samples, the amount of
available information to reject a hypothesis is small, making it more
of a challenge to find significance (i.e., more difficult to reject a null
hypothesis). Therefore, a researcher must use different approaches
to draw conclusions about analysis findings as well as information
to validate assumptions. Because the histograms for Age are non-
normal, the procedures to answer Research Question #6 must be
nonparametric.
Note: When using filters, you must pay extra attention to make
sure your filters are working properly before conducting any analysis
or requesting graphs. SPSS has no way to warn you if your filter is not
working as you intended, so you must look at what cases are and are not
filtered out. In addition, make sure you remove any filters when you are
finished using subsets of your dataset. Forgetting to remove a filter can
be very confusing if you move forward with other analyses and cannot
understand the unexpected analysis results!
SPSS Process
If you still have an active filter from requesting the histograms above,
window.
□ Click to select All cases option in the Select area.
□ Select => OK button to close the Select Cases window and
remove any filter on the dataset.
Now that there is no active filter,
□ Select => Analyze => Correlate => Bivariate to open the

Bivariate Correlations window
Note that Spearman’s rho and Kendall’s tau-b are not found in the
Nonparametric Tests menu structure.
After the Bivariate Correlations window opens,
□ Click to select Age in the list of variables on the left side of the
the Variables: area.
□ Click to select Self_esteem in the list of variables on the left
side of the window, then click the move arrow to move

the variable to the Variables: area (both Age and Self_esteem
appear in the Variables: area).
□ Click the check boxes next to both Kendall’s tau-b and
Spearman tests. (Note: Even if Pearson is checked, it is ignored

if either of the other two is selected.)
□ Click to select the Flag significant correlations checkbox.
□ After reviewed your choices carefully, select => OK button at
the bottom of the Bivariate Correlations window to run the

correlation analyses

SPSS Output
Shortly after you click OK, SPSS’s output window will display and show
the Correlations results for the analyses (see Table 2.3).
Findings
Table 2.3 shows that both Spearman’s rho and Kendall’s tau-b meas-
ure a negative relationship between Age and Self_esteem. Note that the
Kendall’s tau-b finds a significant correlation (p = .042) between the two
variables, while the Spearman’s rho fails to find significance (p = .061)
at α = .05. You can attribute the difference between the two tests to their
differences in dealing with ties within the ranking process. To verify
that you have ties in the data, you can sort the data by Age and look for
multiple occurrences of Age and Self_esteem combinations.
Table 2.3. Correlation Results for Spearman’s rho and Kendall’s tau-b

Correlations
Age Self_ esteem
Kendall’s Age Correlation Coefficient 1.000 -.296*

tau-b
Sig. (2-tailed) . .042
N 28 28
Self_esteem Correlation Coefficient -.296* 1.000
Sig. (2-tailed) .042 .
N 28 28
Spearman’s Age Correlation Coefficient 1.000 -.359
rho
Sig. (2-tailed) . .061
N 28 28
Self_esteem Correlation Coefficient -.359 1.000
Sig. (2-tailed) .061
N 28 28
* Correlation is significant at the 0.05 level (2-tailed).
The dot in SPSS represents a non-reported value.
Note: A highly respected statistician whose specialty is nonpara-

metric statistics, Dr. Myles Hollander, believed that when using non-
parametric statistics, the strict adherence to α = .05 was too aggressive.
He believed that with a thorough understanding of the data, the use of

α = .10 was just as informative (M. Hollander, personal communica-
tion, 2003). Best practice, of course, is to identify the alpha you will use
prior to beginning the analyses. In the example above, we believe that
concluding that you have a difference between a p = .042 and .061 would
be a bit arbitrary and that both are sufficient to conclude a significant
correlation finding.
To sort data in SPSS,
□ Select => Data => Sort cases… to open the Sort Cases window.
□ Click to select Age in the list of variables on the left side of the
the Sort by: area.
□ Select => OK button to close the Sort Cases window.
Note: When planning to sort cases, it is a good practice to have one var-
iable that will allow you to return the cases to their original order. You
can accomplish this by adding a variable to the dataset that assigns an
ID to the cases, counting from 1 through k. The variable in the data set
for Chapter 2 that identifies the original order is ID.
After you sort the data by Age, a review of the values for Age and
Self_esteem shows multiple occasions in which cases have the same
Age and Self_esteem values —suggesting that the tests managed mul-
tiple ties. Because Kendall’s tau-b handles ties more efficiently than
Spearman’s rho, it makes logical sense that Kendall’s tau-b calculated a
smaller p-value than Spearman’s rho. Again, you should always pursue
multiple ways to verify your findings and the performance of your test
so that you can have confidence in your conclusions.
However, in Research Question #6, Dr. Chonody is interested in
more than just the relationship between Age and Self_esteem for all the
youths. She is interested in the possible differences in the relationships
between youths who have not come out and youths who have. Therefore,
you must run the correlation analysis two more times, once after filter-
ing the cases so that youths who have come out are excluded, and then
again after the youths who have not come out are excluded (i.e., filter on
the variable Coming_Out).
Using the steps discussed earlier to filter out certain cases and the
steps to conduct the Spearman’s rho and Kendall’s tau-b tests, you
collect the Sig. values for the multiple tests into a single table for review
(see Table 2.4).
Although the negative correlations still exist when you divide the
youths into two groups, youths who have come out and youths who
have not, the Sig. values in Table 2.4 show that the negative relation-
ship between Age and Self_esteem is being heavily influenced by the
youths who have come out (i.e., Sig. values are significant at .032 and
.043). In addition, Table 2.4 shows that the relationship between Age
and Self_esteem is not significant for youths who have not come out,
and the negative relationship between Age and Self_esteem is moderate
(i.e., correlation value is between -.3 and -.6) for youths who have come
out. Therefore, the Sig. values for youths who have come out to anyone
suggest that there is only a 3% to 4% chance that there is no relationship
between Age and Self_esteem.
A more statistical explanation of the results of the correlation anal-
ysis is that a test of the relationship between Age and Self_esteem using
Kendall’s tau-b is -.384 and found to be significant (p = .032) for youths
who have come out to one or more people (n = 19).
Practically, if the gay youths in Dr. Chonody’s study are typical of
all gay youths, the findings suggest that self-esteem is not necessarily
linked to age for youths who have not come out. However, for youth
who have come out to anyone, the findings suggest that their self-
esteem will decrease as they get older. This finding should concern you
and push you to conduct a closer examination of, among other things,
your data, your sample, how the data were collected, and the wording
of the questions provided to the respondents. The reason for the closer
examination is that these findings seem to be in conflict of what you
Table 2.4. Correlation for Age and Self_esteem for Two Youth Groups

Youths who have not come Youths who have come out to
out to anyone one or more people
Correlation Correlation Sig. N Correlation Sig. N

test (2-tailed) (2-tailed)
Kendall’s .031 .912 9 .384 .032 19
tau-b
Spearman’s .057 .885 9 -.469 .043 19
rho
might expect —that individuals who have come out would experience

improved self-esteem as they grow older. Remember, your responsibility
as a researcher goes beyond just following the statistical steps to test a
hypothesis. Your responsibility includes understanding all the potential
threats, limitations, and possible biases that are present in your study.
Examining all possible issues is the only way you can be sure that your
findings are ready to present to others.
Research Question #6 explored one way to compare two groups
of youths. To explore further the possible ways to compare groups,
Chapter 3 introduces nonparametric procedures that involve two or
more independent groups, and Chapter 4 introduces nonparametric
procedures that involve related or dependent groups. Knowing these
nonparametric group-related procedures will greatly expand your abil-
ity to analyze data, especially when your data do not meet the assump-
tions for parametric statistics.
3
Comparing Two or More

Independent Groups
CHAPTER OBJECTIVES
This chapter covers analyses that compare two or more independent
groups, and the topics presented are as follows:
• Determining dependent (related) and independent (unrelated)

groupings
• Selecting tests for two independent groups (similar to t-test for
independent samples)
• Mann-W hitney U test
• Kolmogorov-Smirnov test
• Wald-Wolfowitz runs
• Moses test of extreme reactions
• Hodges-Lehmann estimate
• Selecting tests for more than two independent groups (similar to
ANOVA and MANOVA)
• Kruskal-Wallis one-way ANOVA
• Jonckheere-Terpstra for k samples
• Median test
69
INTRODUCTION
Recall that Chapter 1 describes the importance of understanding the
data and individual variables used in statistical tests, and Chapter 2
explains how to examine closely the individual variables and how to
conduct nonparametric tests that involve variables for a single group.
This includes discussion of how to inspect individual variables to gain
a deeper understanding of exactly how much information the variable
contributes to its conceptual construct and characteristics. Most impor-
tantly, a deeper understanding of the individual variable will help you
select variables for other purposes, such as analyses that involve two or
more groups of respondents.
Chapter 3 explores independent groups and how to use nonpara-
metric statistics that are specific to group comparisons. The term inde-
pendent groups refers to the situations in which a response from an
individual in one group is not influenced by a response from an indi-
vidual in another group. In comparison, related groups, discussed in
Chapter 4, involve responses that have a connection (e.g., examining the
difference between two responses from the same individual, so that the
individual is the connection).
Many of the data concerns from previous chapters remain when
moving from single to multiple groups, but the concerns are more pro-
nounced in this chapter because the information available from the
data is now being shared among a number of groups. This becomes
more of a challenge when dealing with the limited information avail-
able in a small sample. The limited information shared among the
groups makes meeting parametric analysis assumptions much more
difficult, if not impossible. Thus, knowing how to use nonparamet-
ric procedures becomes critical to good research. Becoming familiar
with nonparametric procedures for groups will help avoid the poten-
tial mistake of reporting invalid results due to running a paramet-
ric procedure on data that do not meet the necessary assumptions. In
addition, verifying that the assumptions are met is more challenging
when the data are broken up into even more pieces when involving
even more groups.
Research Questions 7 to 11 illustrate situations that involve com-
paring independent groups and the manner in which data are analyzed.
Comparing Two or More Independent Groups 71
After each research question, detailed steps are provided for how to
perform an analysis, issues to be aware of when conducting the anal-
ysis, and how to interpret test results. Simply stated, Chapter 3 covers
nonparametric statistics for independent group comparison, includ-
ing the researchers’ responsibility to understand the variables individ-
ually as well as how variables are used to investigate potential group
differences.
Box 3.1 Scenario for Chapter 3

Dr. Rutledge conducted a small feasibility study of an anticipated clinical trial
for healthy, sexually active HIV + men and women. The intervention, a single
motivational interviewing session, targets the reduction of HIV transmission by
addressing participants’ ambivalence about safer sexual practices. Dr. Rutledge
recruited potential participants through advertisements in newspapers and flyers
placed in physicians’ offices. Participants attended an initial screening interview,
during which they answered questions to determine the appropriateness of their
inclusion for the study. At this time, they also responded to a computer-based
pretest questionnaire that included behavioral questions and validated measures
of knowledge, attitudes, and beliefs. Dr. Rutledge randomly assigned a total of
60 men and women into one of two groups —a treatment group that would
receive the intervention immediately, and a control group that would receive the
intervention after the completion of the feasibility study. The 30-member treatment
group participated in the motivational interview and completed the posttest, a
computer-based questionnaire. The control group completed the posttest as well.
You will act as Dr. Rutledge’s statistical consultant. His research questions follow.
Like Chapter 2, Chapter 3 begins with a scenario describing a

hypothetical line of study that can be explored with specific research
questions.
Chapter 3 contains five research questions that guide the presenta-
tion of the nonparametric procedures related to the research scenario
above. Each research question illustrates the reasoning for using non-
parametric procedures. Table 3.1 lists the nonparametric procedures
presented in Chapter 3, the kinds of variables for each procedure, and
the names of the variables found in the database for Chapter 3 that is
available online.
Research Nonparametric Variable Measure Variable Name
Question # Test in DB
7 Moses extreme One dichotomous Received_MI

reaction, and and Percent_Pre
Hodges-Lehmann one scale
confidence
intervals
8 Kolmogorov- One dichotomous Received_ MI
Smirnov, and and Percent_DIF
Mann-W hitney U, one scale
Wald-Wolfowitz
9 Kruskal-Wallis One categorical Received_ MI,
with 3 or more Percent_Post,
groups, and Partner
either nominal or
ordinal,
and one scale
10 Jonckheere-Terpstra One categorical Stage_of_Change
with 3 or more and Percent_DIF
groups,
either nominal or
ordinal,
and one scale
11 Median One categorical Condom_Request
with 3 or more and Self_efficacy
groups,
either nominal or
ordinal,
and one scale

Dr. Rutledge’s main research question is whether the Motivational Interviewing
(MI) intervention reduced the percentage of times the participants had
unprotected sex in the two months following the intervention. The first step was
to place the participants randomly into two groups (control and treatment) in
which the treatment group will receive the intervention. Next, to be certain that
he has comparable groups prior to conducting the intervention, Dr. Rutledge
wants you to compare the two groups on their pretest scores for their report
of the approximate percentage of time they use condoms when having sex.
He wants to be sure that the median value of each group is similar so he can
properly assess the influence of the MI intervention.
Are the median and distribution of percentage values for the treatment and
control groups similar?
MOSES EXTREME REACTION TEST

The Moses test of extreme reactions focuses on the tail ends of the range
of scores to analyze whether two distributions are similar. As in other
nonparametric tests, a ranking strategy is used while focusing on the
lower and upper values to test the hypothesis that extreme values are
equally likely in both distributions. The test is restricted to two groups,
with the assumption that the group values captured by a single scale
variable are independent. A common way to compare distributions for
extreme values is to graph the distributions using either histograms or
box-a nd-whisker plots, and if you assume the distributions are normal,
you have the option of testing for equal means. However, other than a
visual examination of the data, SPSS does not offer a parametric test
for equal probability of extreme values. Thus, the Moses test of extreme
reactions is conducted in concert with the Hodges-Lehmann estimate,
as follows.
HODGES-LEHMANN ESTIMATE (CONFIDENCE INTERVALS)

Multiple Hodges-Lehmann estimates exist, but described here is the
estimate of a confidence interval for the median difference between two
distributions. Thus, the Hodges-Lehmann estimate returns a calcu-
lated confidence interval for median difference values, which provides
an opportunity to examine a potential range of median differences.
However, the estimate assumes that the distributions are similar (i.e.,
kurtosis and skewness values are similar), so the estimate is valid only
after it is verified that the two distributions are similar in shape, which
is especially challenging when sample sizes are small.
Note: The Hodges-Lehmann assumes that individual group distri-
butions are similar and provides a confidence interval on the plausible
difference in median values for a continuous variable (e.g., percentage of
times using a condom) across two groups (e.g., treatment and control).
Other examples for use of the Moses test are as follows:
1. You believe the arrival times of visitors to a small agency are

similar on Monday and Thursday, and you need the information
for scheduling.
2. You believe that depression levels are similar in your small

sample of men and women, and you need to know this before
developing an intervention plan.
Running the Moses extreme reaction test and requesting a confi-

dence interval from the Hodges-L ehmann procedure together pro-
vide a statistical test of equal medians and give you information on
the possible range of the median difference. Of course, if the Moses
extreme reaction test rejects the null hypothesis (i.e., finds that the
two groups have different medians), the assumption for the Hodges-
Lehmann procedure is not met. However, if the Moses extreme reac-
tion test does not reject the null hypothesis, the Hodges-L ehmann
procedure provides information on a plausible range of median dif-
ference values that gives you more insight into how similar the two
groups are.
The variables needed for the Moses extreme reaction test and the
Hodges-Lehmann confidence interval are a dichotomous variable rep-
resenting two groups and a scale/continuous variable. For Research
Question #7 (see Box 3.2), the dichotomous variable is Received_MI,
a grouping variable to distinguish who did and did not receive the
Motivational Interviewing (MI) intervention. The scale variable is
Percent_Pre, the percentage of unsafe sex episodes measured prior to
giving one group the intervention.
SPSS Process
□ Select => Analyze => Nonparametric Tests => Independent

Samples.
Once the Nonparametric Tests: Two or More Independent Samples

window opens, SPSS needs to know which variables to use in the analy-
sis and which analysis routines to run.
□ Select => Fields tab to see the list of available variables for the
analyses.

□ Click to select the grouping variable Received_MI.
□ Click the move arrow closest to the Groups: area to move the
variable to that area.
□ Click to select the scale variable Percent_Pre that contains
percent values for respondents’ condom use.

□ Click the move arrow closest to the Test Fields: area to move
the variable to that area.

□ Select => Choose Tests from the Select an item: area.
□ Click to select Customize tests to enable the ability to select
specific tests.
□ Click the check boxes next to Moses extreme reaction.
□ Click the check box next to Hodges-L ehmann estimate.
For Research Question #7, the only test option to consider is the Moses
extreme reaction for identifying outliers. Use the default option (i.e.,
Compute outliers from sample) to use the observed data.
□ When you have reviewed your choices carefully, select => Run
to conduct the analysis.
SPSS Output
After you click Run, a variety of information shows up on SPSS’s output
window, including:

• Location of the data set used (i.e., Active Dataset)
• Hypothesis Test Summary (i.e., Model Viewer)
• List of items in the output index on the left side of the output
window
Clicking on an item in the list initiates a scroll to that information on

the right side of the output window. To see the details of the test, double-
click within the area outlined by the Model Viewer (i.e., point the mouse
at the Hypothesis Test Summary information and double-click). The
Model Viewer window will open to display the detailed information for
the analysis.
Findings
The information for the Moses extreme reaction test shows that the
two groups have similar values for percentage of time using a con-
dom (i.e., Sig. = .665 is larger than the α = .05 significance level). The
Moses extreme reaction test provides an exact significance (Sig. is the
notation SPSS uses to represent statistical significance), indicating
that there is a 66.5% chance that the ranges are the same. By tradi-
tion, statistical significance must be small, with only a 5% or smaller
chance that the medians are similar, before concluding that the groups
are different.
Note: Not all statistical tests provide exact significance values.
Therefore, be sure to translate a Sig. value to “percent chance” interpre-
tation only when the test allows this.
To see the Hodges-Lehmann results, you must change the view. At
the bottom left of the Model Viewer window, one of the dropdown lists
is named View:.
□ Click on the View: dropdown list and click to select Confidence

Interval Summary View.
After you select the Confidence Interval Summary View, the informa-
tion for the Confidence Interval Summary is displayed. It reveals a
95% confidence interval for the difference between median values range
from -.20 to .10. This range includes a plausible difference of zero and,
therefore, is in line with the conclusion drawn by the Moses extreme
reaction test.
In addition, a box-a nd-whiskers plot associated with the test is
included with the other analysis results. Box-a nd-whiskers plots are
a graphical representation of a distribution of values. The box part
of the graph represents the first and third quartiles, called the inter-
quartile range or IQR, and the thicker dark line represents the
median value. The lines extending from the box are called whiskers
and are calculated differently by different applications. SPSS draws
the end of each whisker as the minimum and maximum values in
the data that are not identified as statistical outliers. SPSS identifies
outliers as values that fall farther away from the first and third quar-
tiles than 1.5 times the size of the IQR. SPSS identifies outliers with
small circles and with asterisks when outliers are three times beyond
the IQR. Figure 3.1 does not identify the presence of any outliers, but
Figure 3.5 does. The similarity of the vertical positions of these areas
for the two groups illustrates and supports the values for the confi-
dence interval and the hypothesis test finding of similar range values
(see Figure 3.1).
In summary, the Moses extreme reaction test did not find a signif-
icant difference in the range of percentage of condom use between the
case and control groups. In addition, the Hodges-Lehmann confidence
interval for plausible median difference identified a lower and upper
bound of −.20 and .10, respectively. Therefore, the two analyses support
the conclusion that there is not enough information to suggest that the
median value or range of values for condom use is different.
Independent-Samples Moses Test of Extreme Reaction
1.00
0.80
Percent_Pre
0.60
0.40
0.20
0.00
Received MI Did not Receive MI
Received_MI
Control Group N = 30 Experimental Group N = 30
Figure 3.1. Box-a nd-whiskers plot for Moses test of extreme reactions.

In statistical terms, the Moses extreme reaction test (p = .648) did

not reject the null hypothesis at α = .05 of similar range values for per-
cent condom use between case and control groups. In addition, the
Hodges-Lehmann confidence interval found a 95% confidence interval
of –.20 to .10. These two values indicate the range in which we are 95%
confident that the true percent difference falls. It is not a coincidence
that the null hypothesis was not rejected and the confidence interval
includes a zero percent difference.
Practically, the above tests verify that the randomization process
did not create dissimilar groups that would bias the examination of
the intervention. If the tests had suggested a possible difference, then
Dr. Rutledge would need to conduct a new randomization of the respon-
dents until the two groups were similar in their condom use. Without
a comparison of the two groups prior to conducting the intervention,
the results could have been inaccurate without any evidence to inform
Dr. Rutledge otherwise. Of course, a statistical test using α = .05 sug-
gests a 5% probability that a difference between the case and control
groups is identified when the truth is that there isn’t a difference (i.e., .05
probability of incorrectly rejecting the null).

Now that Dr. Rutledge has verified that the two groups are similar in their
condom use, he can proceed to conducting the MI intervention on the treatment
group members and then test to find out if the intervention makes a difference.
To examine this, he created change scores for each group. The change scores
represent the difference in the percentage of time having unprotected sex from
pre-intervention condom use to post-intervention condom use. Dr. Rutledge
asks you to compare the changes scores for the treatment group with those of
the control group to answer his question.
Are the changes in percentage of time having unprotected sex different between
those who have received the intervention and those who have not?
Note: A nonparametric ANCOVA procedure does exist (Tsangari & Akritas,

2004), and you could use it to answer this research question. However,
it is not a procedure available in SPSS, so it is not discussed herein.
KOLMOGOROV-SMIRNOV TEST
The Kolmogorov-Smirnov test calculates the amount of separation of
the empirical distributions for the two groups, and it uses the amount
of separation to decide if the two distributions are significantly dif-
ferent. Said differently, it tests to find out if it is possible that the val-
ues for the two groups come from the same distribution. Because the
focus is on the separation of the two distributions, no prior assump-
tion about the distribution exists, unlike the t-test, in which the data
are assumed to come from a normal distribution. Recall that when the
data clearly meet all parametric assumptions, the parametric proce-
dure does have more power to find significance than its nonparamet-
ric counterpart (emphasis on the word clearly). Importantly for the
Kolmogorov-Smirnov test, not having a normality assumption does
make the test less able to find significant differences between distribu-
tions than the parametric t-test when the data clearly meet the t-test’s
normality assumption. In addition, SPSS restricts the Kolmogorov-
Smirnov test to two groups.
Other examples for uses of the Kolmogorov-Smirnov test are as
follows:
1. You suspect a difference in PTSD between men and women

clients at your small treatment center; therefore, you have
everyone complete a standardized measure of PTSD so you can
examine the differences by gender.
2. You hypothesize that those in your youth group who identify as
being spiritual are happier than those who do not, so you ask
them to complete a small survey to investigate this.
Remember, prior to conducting any hypothesis test, you must

examine the information provided by the variables. For Research
Question #8 (see Box 3.3), one variable represents group member-
ship (i.e., participated or not in MI) and the other represents the
percent change in having unsafe sex. We assume that you have
already used the steps discussed in Chapter 2 to examine the vari-
ables individually.
SPSS Process

Samples.
Once the Nonparametric Tests: Two or More Independent Samples win-

dow opens, SPSS needs to know which variables to use in the analysis
and which analysis routines to run.
analyses.
□ Click to select the scale variable (Percent_DIF) that contains
change in percentage values.

□ Select => Settings tab to see the list of available nonparametric

tests.
specific tests.
□ Click the check box to the left side of Kolmogorov-Smirnov
test in the Compare Distribution across Groups list of

tests area.
You may run into situations in which you want to change the signifi-
cance level for your test. In addition, some of the tests within this list of
nonparametric procedures can estimate confidence intervals. The steps
below show how to change the options associated with significance level
and confidence intervals.
□ Select => Test Options from the Select an item: area.

□ Decide on the significance level (default is 0.05) for the test.
□ Decide on the confidence interval range (default is 95.0%).
□ Decide how to handle missing data (default is Exclude cases
test-by-test).
□ After reviewing your choices carefully, select => Run to conduct
the analysis.
SPSS Output
window, including:

output window
Clicking on an item in the list initiates a scroll to that information on the

right side of the output window. To see the details of the test, double-click
within the area outlined by the Model Viewer (i.e., point the mouse at the
Hypothesis Test Summary information and double- click). The Model
Viewer window will open to display the detailed information for the analysis.
Findings
The Kolmogorov-Smirnov test shows Sig. = .056, which is very close to
the Significance level: set in the Test Options from the Select an item:
area. In the Hypothesis Test Summary information, SPSS is not indi-
cating that this test rejects the null hypothesis (i.e., the null hypothesis is
that distributions between the two groups are the same). We discuss in
Chapter 2 how nonparametric procedures are more conservative when
finding test significance. Therefore, a case can be made that this test
is close enough to rejecting the null hypothesis to report a difference
in group distributions, although some statistical purists will definitely

disagree. Ideally, you would make the decision to select a more nontra-
ditional significance level (e.g., α = .10) before conducting the analysis.
We have learned that clearly describing your reasoning for determining
significance is sufficient information to have a manuscript published or
a grant proposal funded.
Another approach is to gather more information before drawing a
conclusion about the findings. To do this, simply conduct the other two
Compare Distribution across Groups tests to learn how all three tests
report their findings. The two others are the Mann-W hitney U test and
the Wald-Wolfowitz test, as follow.
MANN-WHITNEY U TEST
The Mann-Whitney U test, also called the Wilcoxon rank sum test, uses
ranking of the observed values to determine if the two groups come from
the same distribution. Unlike the parametric equivalent independent t-test
that assumes a normal distribution, the Mann-Whitney has no distribu-
tion assumption. The only restrictions in SPSS are that the test is for only
two groups and the comparison variable is on a scale level of measurement.
The lack of any distribution assumption does make the Mann-Whitney
U appropriate for any type of distribution, but, as with the Kolmogorov-
Smirnov and Wald-Wolfowitz tests, the flexibility does make finding sig-
nificant differences between the distributions a bit more difficult.
WALD-WOLFOWITZ TEST
The Wald-Wolfowitz test analyzes runs for the values while focusing
on group membership. The test’s structure provides an opportunity to
test the hypothesis that values for a scale variable from two groups have
similar value distributions. The test is restricted to two groups for com-
paring the distribution of a scale variable. Because the Wald-Wolfowitz
uses runs, no assumption about the distribution is required. However,
when a sample size is greater than 30, SPSS uses a normal approxima-
tion for a test statistic because as the number of possible runs reaches
30+, the distribution for the number of possible runs is similar to a
normal distribution. Another use of this test is to examine if responses

from two groups captured by a single scale variable are independent.
The execution of the test is the same, only the interpretation of the find-
ings changes (i.e., distributions of a scale variable found to be differ-
ent for two groups imply that responses between the two groups are
independent).
The summary of tests for Research Question #8 is:
• Kolmogorov-Smirnov —uses distribution separation to draw

conclusion about similarity between distributions.
• Mann-W hitney U —uses ranking to draw conclusion about
similarity between distributions.
• Wald-Wolfowitz —uses runs to draw conclusion about
similarity between distributions.
SPSS Process
To begin your analysis using all three tests,

Samples.

dow opens, SPSS needs to verify that the variables to use in the analysis
are still selected.
analyses.
□ Click to select the Use custom fields assignments option.
If Percent_ MI are not in the Test Fields: and

DIF and Received_
Groups: areas,

□ Click to select the scale variable (Percent_DIF).

specific tests.
Figure 3.2 shows many options for nonparametric tests, some for two
groups and some for more than two groups. Looking closely at Figure
3.2 will show five tests specific to having only two groups.
□ Click to select the three check boxes on the left side of the
Compare Distribution across Groups list of tests —Mann-
Whitney U, Kolmogorov-Smirnov, and Wald-Wolfowitz.
Figure 3.2. Nonparametric Tests: Two or More Independent Samples window.

All three “for 2 samples” tests are selected because, as discussed

above, each one approaches the test of similar groups differently.
Therefore, to fully understand if the two groups are actually different,
the results from each of the three tests need to be reviewed.
Before continuing, review the test options to make sure they are set
appropriately.
□ Select => Test Options in the Select an item: area to see the
place where to change default settings.
□ Decide on the significance level (default is 0.05) for the test.
□ Decide on the confidence interval range (default is 95.0%).
□ Decide how to handle missing data (default is Exclude cases
test-by-test).
the analysis.
SPSS Output
window, including:

window

Model Viewer window will open to display the detailed information (see
Figure 3.2).
The Model Viewer window has access to more than what is shown
in Figure 3.2. By clicking on one of the three test results listed in the
Hypothesis Test Summary table on the left side of the Model Viewer
window, you change the information on the right side to that particular
test information.
Another way to switch between test information is to change the
selection from the dropdown lists on the left for View and Field Filter,
and on the right for Test, Field(s), and View. Figure 3.2 shows the
information for the Kolmogorov-Smirnov test because that test is the
one highlighted in the list of tests on the right.
Findings
A review of the Sig. values finds that the Wald-Wolfowitz runs test is
clearly not rejecting the null hypothesis (p = .994). The Mann-W hitley
U test (p = .059) and the Kolmogorov-Smirnov test (p = .056) are similar
in their findings. In other words, the two are close to rejecting the null
hypothesis at α = .05 that the groups are similar in their Percent_DIF
values. Differing values for these three tests is common, and this situa-
tion demonstrates how important it is to review all three. To complicate
matters even further, if you conduct an independent-samples t-test on
the data, the parametric test will return a Sig. value of .012 —leading
you to conclude that the two groups are significantly different for values
of Percent_DIF.
When faced with conflicting information, begin with what you know
about the tests. You know that the nonparametric tests are less likely
to find significance than parametric tests when parametric assump-
tions are met. You also know that one guideline for deciding if you
have enough information about a group is 30 per group, although small
group sizes can typically detect only large effect sizes. An a priori power
analysis is another method for understanding the strengths of your
data, but the discussion on how to conduct a power analysis is left for
other resources. For Research Question #8, you again find yourself with
a marginal group size (i.e., two groups of 30). You know also that when
the data for each group resembles normality, the independent-sample
t-test may be appropriate (i.e., after looking for outliers, missingness,
data value gaps, etc.). Figure 3.4 shows histograms of the distributions
for the two groups by filtering on the variable Received_MI. Recall that
the process on how to filter data is discussed in Chapter 2, and it is dis-
cussed later in this chapter.
Neither of the two histograms shows a large departure from normal-
ity, providing additional support for the findings of the independent-
samples t-test. In addition, the two nonparametric tests related to
distribution comparisons were close to being significant, with Sig.
values of .059 and .056. Thus, Dr. Rutledge has enough information to
cautiously reject the null hypothesis and conclude that the MI inter-
vention significantly changed the percentage of unsafe sexual behav-
ior over a two-month period. However, when cautiously rejecting a null
Received MI Intervention Did Not Received MI Intervention

Mean = .10 4 Mean = –.05
Std. Dev. = .209 Std. Dev. = .30
6 N = 30 N = 29
3
Frequency
Frequency
4
2
2 1
0 0
–0.40 –0.20 0.00 0.20 0.40 0.60 –.50 –.25 .00 .25 .50
Percent_DIF Percent_DIF
Figure 3.4. Histograms for Percent_DIF values for MI and non-MI groups.

hypothesis, it is very important that he explain to his readers that addi-

tional studies with larger samples need to be conducted before making a
final determination about the intervention’s effectiveness.
In statistical terms, both the Mann-W hitney U test (p = .059) and
the Kolmogorov-Smirnov test (p = .056) for distribution comparison
reject the null hypothesis of equal distributions at α < .10. Additionally,
results from conducting a parametric independent-sample t-test also
reject the null hypothesis (p = .012)
Practically, the findings suggest that Dr. Rutledge’s MI interven-
tion influenced the respondents’ unsafe sex behavior over a two-month
period. After he divided his respondents into two groups, one group
received a motivational intervention (MI) while the other group did not.
He measured both groups on their unsafe sexual behavior using pre and
post surveys. The analysis provided enough evidence to suggest that the
MI intervention was effective in increasing the percent of safe sex activ-
ity for the group who received the intervention.

Based on the literature, Dr. Rutledge suspects that frequency of condom use
is similar among three groups of participants —those who have exclusively
same-sex sexual partners, those who have opposite-sex sexual partners,
and those who have both same- sex and opposite- sex sexual partners.
Dr. Rutledge asks you to test this for the participants in his treatment group,
as the findings can be useful when analyzing the post-intervention data,
and in determining if this will be a reasonable research question for his
clinical trial.
After participants in the treatment group receive the MI intervention, is
condom use different among these participants based on who has exclusively
same-sex, exclusively opposite-
sex, and both same- sex and opposite-sex
sexual partners?
KRUSKAL-WALLIS ONE-WAY ANOVA TEST

The parametric ANOVA analysis is frequently used to investigate group
differences, but the independent groups, homogeneity of variances, and
normality assumptions are sometimes very difficult to meet, especially
when sample sizes are small. The Kruskal-Wallis one-way ANOVA test
does assume independent groups but has no normality assumption.
This distribution-free characteristic makes the test applicable to many
data analyses and it is best suited for testing data that represent groups
that are nominal, conceptually (i.e., the compared groups do not have a
natural ordering). As in other nonparametric tests, the distribution-free
characteristic comes with the cost of being more conservative, or less
likely to find significant group differences. The Kruskal-Wallis one-way
ANOVA test uses runs of values from two, three, or more independ-
ent groups to determine if the groups come from the same distribu-
tion. Of course, as the number of groups increases, so does the required
size of the sample. The test pools the values from the groups to see how
often the values switch from one group to the other to determine group
differences.
Other examples for use of the Kruskal-Wallis one-way ANOVA test
are as follows:
1. You conducted a pilot study for a larger research study about

college drinking. You suspect that risky drinking is different
among students in fraternities and sororities than it is among
students who are not members, so you test this hypothesis with
your small pilot sample before including those variables in your
larger study.
2. You plan to conduct a comprehensive analysis of a secondary
data set, but you are unsure about some of the relationships
among your variables. You use a small subset of the data and
place people in groups according to their socioeconomic status
(SES) so that you can test if the level of resilience, as indicated by
a standardized measure, is different among SES categories.
Box 3.5 Research Question #9, Continued

From the findings of the analysis conducted in Research Question #8,
Dr. Rutledge has evidence to suggest that receiving an MI intervention changes
the amount of condom use. Now he wants you to verify that the change from
the intervention is not related to sexual partner choice. To do this, the first
SPSS process to perform is to filter out the control group and then conduct the
Kruskal Wallis one-way ANOVA analysis.
SPSS Process
To filter the data so that only respondents who received the MI interven-
tion are analyzed,
window.
Select area.
□ Click to select the Received_MI variable on the left side of the

□ Right after the variable name Received_MI in the white space,

□ Type =1 (i.e., equals sign and the number 1). This tells SPSS to
use cases that have a value of 1 for the variable Received_MI.

Recall, a 1 represents respondents in the treatment group.
□ Select => Continue button to close the Select Cases: If window.
□ Select => OK button to close the Select Cases window and use
only the cases that have a 1 for Received_MI.
To begin your analysis of the treatment group,

Samples.

analyses.
Note: If the Received_MI variable is in the Groups: area, move it back

to the list of variables on the left. In addition, if the Percent_Pre variable
is in the Test Fields: area, click to select the variable and then click the
move arrow to move it back to the Fields: area.
□ Click to select Partner variable in the list of variables on the left

of the window.
□ Click to select the scale variable (Percent_Post) that contains
percentage condom use post intervention.


specific tests.
Figure 3.2 shows many options of nonparametric tests, some for two
groups, and some for three or more groups, frequently referred to as k
samples in published literature. A close review of Figure 3.2 shows three
tests specific to having three or more groups.
□ Click to select the Kruskal-Wallis one-way ANOVA check box on

the right side of the Compare Distribution across Groups area.
Make sure that none of the other test check boxes are checked.
Before running the test, you should consider the test option. For the
Kruskal-Wallis test, the dropdown list for Multiple comparisons: should
be set to All pairwise to compare the distributions of all the possible
pairing of groups in the data. When the number of groups is too large,
the comparison of all possible pairwise combinations becomes cumber-
some, and the Stepwise step-down option is helpful. This option sorts
the groups by their calculated group median value and compares only
adjacent groups, making the comparison of groups possible. However,
with only three groups, the comparison of all pairwise combinations is

completely manageable.

the analysis.
SPSS Output
window, including:

output window

the analysis (see Figure 3.5). As in the process discussed earlier, change
the settings in the dropdown lists to find detailed information about the
data and the analyses.
Findings
Figure 3.5 shows the results of the analyses for Research Question #9. The
box-and-whiskers plot in Figure 3.5 shows that the median values among
the three groups are very similar, but the group variability appears differ-
ent. Specifically, the respondents who are exclusively with opposite-sex
partners fall much closely around the median value, while the other two
groups spread out more (i.e., whiskers indicate a wider range of values for
exclusively same-sex and both same-sex and opposite-sex partner groups
than for the exclusively opposite-sex partner group).
Even though the box-a nd-whiskers plot in Figure 3.5 shows only
a slight difference in Percent_Post median values along the horizon-
tal axis for the three Partner groups, you still need to verify observed
difference. Another way to examine the difference in median values is
to review the calculated mean ranks for the three groups. To view the
mean rank values, hover the mouse pointer over the black line in the
box-a nd-whiskers plots. Hovering will cause a Mean Rank value to
appear (14.25 for exclusively same-sex, 15.27 for exclusively opposite
sex, and 17.43 for both).
The Kruskal-Wallis test did not find significant distribution differ-
ences among the three groups. The Kruskal-Wallis Sig. value = .775,
larger than the traditionally used threshold value of α = .05. This non-
significant finding is graphically supported by the amount of similar
vertical placement in the box-and-whiskers plots.
In statistical terms, the Kruskal-Wallis test for distribution compar-
ison (p = .775) failed to reject the null hypothesis of equal distributions
at α = .05.
Practically, this means that the evidence suggests that the influence
of an MI intervention is similar among groups of people who are exclu-
sively with same-sex partners, exclusively with opposite-sex partners,
and those who have both same and opposite-sex partners. Of course,
another explanation is that the small sample simply does not pro-
vide enough information to identify differences. Therefore, adding
the information from Research Question #8, we can infer that there is
not enough information to conclude that the intervention benefits one
group over the other.

Dr. Rutledge utilized Motivational Interviewing (MI) as the intervention in his
feasibility study to reduce unprotected sexual behavior by moving participants
through a variety of stages of change. The goal of MI is to facilitate the
movement from the lowest stage, precontemplation, through the subsequent
stages of contemplation, preparation, action, and maintenance. During the
initial screening, Dr. Rutledge assessed his participants’ stage of change, and
found that all the participants in the treatment group fall into one of three
stages of change —precontemplation, contemplation, and preparation for
change. Dr. Rutledge suspects that the amount of change in condom use after the
intervention differs among the groups based on the participants’ stage of change,
so he divided his sample into these three groups. He hypothesizes that he will
find the lowest percent change in condom use among those in precontemplation,
and the highest change in condom use among those in preparation. He asks you
to examine this hypothesis, because if it is true, the use of MI to improve the
participants’ stage of change will be a valuable goal for the clinical trial.
Is the percentage of change in condom use after receiving the MI intervention
different for respondents who are in different stages of change (i.e.,
precontemplation, contemplation, and preparation)?
Because the grouping of respondents suggests an order, the test to

use in this kind of situation is the Jonckheere-Terpstra test. Unlike the
other nonparametric independent sample tests, the Jonckheere-Terpstra
is better equipped to look for group differences when the groups have an
order like the one in Research Question #10.
JONCKHEERE-TERPSTRA TEST
Like the Kruskal-Wallis test, the Jonckheere-Terpstra test uses runs to
determine if the values from the groups come from the same distribution.
However, the alternative hypothesis is that the distributions are somehow
ordered, suggesting that the distributions are naturally decreasing or

increasing among the groups. This difference from the Kruskal-Wallis
test makes the Jonckheere-Terpstra more powerful in detecting distri-
bution differences when the ordering characteristic is present. Similar
to the Kruskal-Wallis test, the Jonckheere-Terpstra test assumes only
independent groups and does not assume normality, which makes this
test applicable whenever groups have some kind of order. Although the
ordering of the groups does increase its strength when seeking group dif-
ferences, it will still be more conservative than a one-way ANOVA when
the data meet the normality assumption.
Other examples for use of the Jonckheere-Terpstra test are as follows:
1. As a supervisor in a small organization, you wonder if your

professionals’ negative behavior with clients is related to their
caffeine intake. You know that measuring the number of cups of
coffee is one way to examine this question, but you imagine that
the size of their coffee cups is important to measure as well. You
divide your workers into groups based on whether people drink
from small, medium, and large cups of coffee, and test this with
a standardized measure of behavior.
2. A personal trainer suspects that people who exercise a lot
may express more narcissistic symptoms than those who don’t
exercise much or at all. To test this, she asks all her clients to
fill out a questionnaire that includes a standardized measure
of narcissism along with a measure of how frequently they
exercise. She divides her trainees into three groups –those who
exercise only occasionally, those who exercise one to three times
a week, and those who exercise four or more time a week, to test
her hypothesis.
The variables needed for the Jonckheere-Terpstra tests are a catego-

rical ordinal variable representing three or more ordered groups, and a
scale or continuous variable. For Research Question #10, the categorical
variable is Stage_of_Change and the scale variable is Percent_DIF.
SPSS Process
Similar to the steps for Research Question #9, Research Question #10
involves analyzing only the data that represent those who were part
of the treatment group. Therefore, the data filter used earlier on the
Received_MI variable needs to be active before beginning the analysis.
If you do not recall how to set the filter, see the first few steps under the
SPSS Process for Research Question #9.

Samples.

analyses.
□ Click to select Stage_of_Change variable in the list of variables
in the Fields: area.
□ Click to select Percent_DIF scale variable that contains percent
change difference in condom use.


specific tests.
□ Click the check box for Jonckheere-Terpstra for k samples
on the right side of the Compare Distribution across

Groups area.
□ Make sure that none of the other test check boxes are checked.
Before running the test, you should consider a few options. First, set the
dropdown list for Multiple comparisons: to All pairwise to compare
the distributions of all the possible pairings of groups in the data. To

take advantage of the strengths of the Jonckheere-Terpstra test and
increase the probability to finding significance among the groups when
group differences exist, you will likely want to select the best option
for hypothesis order:. The ordering refers to the theoretical or natu-
ral order of the groups identified by the categorical variable. Because
the Jonckheere-Terpstra test uses an ordinal variable to identify groups,
the ordinal variable places the groups in a sequence or specific order.
While examining the groups in this specific order, you can compare
the median values among the groups by running a Kruskal-Wallis test.
Note: If the medians are from smaller to larger values, set the Hypothesis
order: dropdown option to Smallest to largest. If the medians are from
larger to smaller values, set the dropdown option to Largest to smallest.
□ For Research Question #10, click on Hypothesis order:

dropdown and select Smallest to largest.
the analysis.
SPSS Output
window, including:

output window

the analysis. As in the process discussed earlier, change the settings in
the dropdown lists to find detailed information about the data and the
analysis.
Findings
The results displayed in the Model Viewer window for the Jonckheere-
Terpstra test did not indicate that the distributions were significantly
different among the three groups (Sig. = .055). Because SPSS did not
find significant differences, the multiple comparisons All pairwise
option was ignored. However, a Sig. value of .055 is very close to the
α = .05 threshold. Considering the closeness of the Sig. value to .05
and your understanding that nonparametric tests tend to be more
conservative in finding significance, a .055 may be enough to proceed
further. But how?
One option is to adjust the significance level slightly so that SPSS
will provide the pairwise comparison output, which in turn will give
you more information for you to make a final determination about the
test (despite inflating the experiment-w ise Type I error rate). To adjust
the significance level value,

Samples to reopen the Nonparametric Tests: Two or More
Independent Samples window.
□ Select => Settings tab to see the options in the Select an
item: area.
□ Change the value displayed for the Significance level: area from
0.05 to 0.10.
□ Select => Run button to rerun the analysis.
Now that SPSS recognizes that the Sig. value is smaller than the
Significance level option value, the pairwise comparisons are con-
ducted. Notice that although the significance level value has changed, the
Sig. value remains the same. The reason is that the decision to set the sig-
nificance level at a specific value is theoretical, and that decision affects
only the choice of rejecting or failing to reject a test’s null hypothesis. The
chance that the three groups have similar means is still the same.
After double-clicking within the Hypothesis Test Summary area

in the output window, and then changing the value of the dropdown
for View: on the right side of the window to Pairwise Comparisons,
you can see which comparisons are more different than others (see
Figure 3.6).
Figure 3.6 shows that the group that is dissimilar to the other two
is Precontemplation. In addition, the pairwise comparison that is most
dissimilar is between Precontemplation and Preparation, with Adj.
Sig. = .082. Like the Sig. value, the Adj. Sig. is used to decide when to
Pairwise Comparisons ...
Precontemplation
11.14
Contemplation
17.60
Preparation
18.50
Each node shows the sample average rank of Stage_of_Change.
Sample1-Sample2 Test Std. Std. Test Sig. Adj.Sig.

Statistic Error Statistic
Precontemplation-Contemplation 78.000 14.090 1.632 .051 .154
Precontemplation-Preparation 74.500 13.018 1.920 .027 .082
Contemplation-Preparation 47.000 12.199 .164 .435 1.000
Each row tests the null hypothesis that the Sample 1 and Sample 2 distributions are the same.
Asymptotic significances (1-sided tests) are displayed. The significance level is .10.
Figure 3.6. Pairwise comparison graph of Stage of Change groups.

reject the null hypothesis based on the established significance level. The
difference is that there is an adjustment to the value due to the increased
likelihood of finding significance just because of conducting multiple
tests. Recall, in this figure the reason that the pairwise is showing signif-
icance is that earlier you changed the significance level to .10, allowing
SPSS to conduct the pairwise comparisons.
Because the initial test was very close to rejecting the null hypothe-
sis and the review of the pairwise comparisons shows a clear difference
between Precontemplation and Preparation, there is enough evidence
to claim that respondents within this sample that are in different stages
of change do, indeed, respond differently to the MI. The stage difference
is seen in the percent change in condom use prior to and then after the
intervention, with the largest difference being between those in precon-
templation and preparation.
In statistical terms, the Jonckheere-Terpstra test (p = .055) rejects
the null hypothesis of equal distributions at α = .10. A pairwise compari-
son found adjusted p = .082 between Precontemplation and Preparation,
adjusted p = .154 between Precontemplation and Contemplation, and
adjusted p = 1.0 between Contemplation and Preparation.
Practically, this means that the evidence suggests that individuals
in the precontemplation stage of change decrease their level of unsafe
sexual behavior after experiencing an MI intervention, but less of a
decrease than that of individuals in the preparation stage. This informa-
tion makes theoretical sense, and it will greatly benefit the effort to
improve the effectiveness of an MI intervention that targets decreasing
the amount of unsafe sexual activity.

Improving participants’ self- efficacy can be a goal for Motivational
Interviewing, and Dr. Rutledge suspects that self-efficacy is related to the
likelihood that participants will ask their sexual partners to use condoms.
Based on their reports in the initial screening, the treatment group
participants were divided into three groups —those unlikely to ask a partner
to use a condom, those somewhat likely, and those very likely to request
condom use. Participants also completed a validated measure of self-efficacy,
so Dr. Rutledge requests that you assess if those unlikely to ask a partner to use
a condom have lowest self-efficacy, and if those very likely to request condom
use have highest self-efficacy scores. This finding can be useful when analyzing
the post-intervention treatment group, and it will support the use of these
measures in the clinical trial.
Do the median values for self-efficacy differ among respondents who are
unlikely to ask their partner to use condoms, those somewhat likely, and those
very likely to request condom use?
MEDIAN TEST
The median test uses the frequency in which the values for each group
are above and below a median value and then employs these counts
to decide if the group medians are significantly different. In SPSS, the
median value can be either a pooled value, in which SPSS calculates a
median value using all the values from the groups, or a specified value
identified by the person conducting the test. Unlike the parametric
t-test that assumes a normal distribution for the groups, the median test
has no assumption about the group’s distribution. The median test does
calculate a chi-square statistic based on expected vs. observed frequen-
cies, so the test will identify significant median differences when the
sample size is large. Note that “large” for three groups is a sample size
of 300+ per group. The other median test assumption is that the groups
are independent, so the values in one group do not have influence on the
values in another group.
Other examples of uses for the median test are as follows:
1. You wish to know if the scale values from a depressive symptoms

measure completed by the participants in your pilot study
identify median differences for men and women.
2. You wish to verify that median values are not significantly
different between subgroups within a sample in your
feasibility study.
The variables needed for the median test are a categorical variable (i.e.,
nominal or ordinal) representing three or more groups of an inde-
pendent variable, and a scale or continuous variable representing the
dependent variable. For Research Question #11, the categorical variable
is Condom_Request and the scale variable is Self_efficacy.
SPSS Process
Similar to Research Questions #9 and #10, Research Question #11 is ana-
lyzing only the data that represent those who were part of the treatment
group. Therefore, the data filter for the Received_MI variable needs to be
active before beginning the analysis. If you do not recall how to set the fil-
ter, see the first few steps under the SPSS Process for Research Question #9.

Samples.
After the Nonparametric Tests: Two or More Independent Samples win-

analyses.
□ Click to select the grouping variable Request_Condom.
□ Click to select the scale variable (Self_efficacy) that contains
values representing levels of self-efficacy for the respondents.

□ Click to select the Customize tests option to enable the ability
to select specific tests.
□ Click to select the check box next to the Median test on the
lower right side of the window in the Compare Medians across

Groups area.
□ Make sure that none of the other test check boxes are checked.
Before running the test, you must consider a few options. You must
decide what median value to use when the median test is counting the
larger and smaller values. If you do not have a specific value you iden-
tified from the literature or another source to use as a cut point, then
it is best to allow SPSS to use a pooled sample median, which is the
median value calculated from all the values in the groups. Therefore,
for Research Question #11, you click to select Pooled sample median.
To test a specific median value, you would select the Custom
option that enables you to enter a specific value in the Median area.
To compare the medians of all the possible pairings of groups in the
data, set the dropdown list for Multiple comparisons: to All pairwise.
When the number of groups is too large, making the comparison of all
possible pairwise combinations cumbersome, the Stepwise step-down
option is helpful. This option sorts the groups by their calculated
group median value and compares only adjacent groups, making the
comparison of groups possible. However, with only three groups, the
comparison of all pairwise combinations is completely manageable.

the analysis.
SPSS Output
window, including:

output window
Clicking on an item in the list initiates a scroll to that informa-

tion on the right side of the output window. To see the details of the
test, double-click within the area outlined by the Model Viewer (i.e.,
point the mouse at the Hypothesis Test Summary information and

double-click). The Model Viewer window will open to display the
detailed information for the analysis. Information for the Hypothesis
Test Summary is displayed on the left, and box-a nd-whiskers plots
on the right.
This information is for the median test and indicates enough
information to conclude the median values for Self_efficacy are differ-
ent across Condom_Request categories, because the Sig. value of .035 is
smaller than the α = .05 significance level. To see the pairwise compar-
ison results,
□ At the bottom right side of the Model Viewer window, click

on the dropdown list named View: and select Pairwise
Comparisons.
The information for the pairwise comparisons is displayed and shows

Sig. and Adj. Sig. values for the three-category pairing (see Figure 3.7).
The difference between Sig. and Adj. Sig. values is an important dis-
tinction. The Sig. values represent a straight comparison test of mean
values without an adjustment for the multiple statistical tests that were
conducted. However, the family-w ise error rate when conducting mul-
tiple statistical tests within an analysis inflates the possibility of finding
significant differences in median values for a category pair. The Adj.
Sig values include an adjustment for this family-w ise error rate and
provide a more accurate view of testing median differences among the
category pairs.
Findings
The overall median test indicates sufficient evidence to conclude that
there is a difference in median self-efficacy across likelihood of request-
ing condom use categories (Sig. value = .035). The results of the pair-
wise tests show that median self-efficacy between respondents reporting
they are unlikely to request a partner to use a condom and very likely
to request a partner to use a condom is significantly different, with
median values of 14.50 for unlikely and 18.50 for very likely categories.
somewhat likely
16.00
unlikely
14.50
very likely
18.50
Each node shows the sample median of Condom_Request.

Test
Sample1-Sample2 Sig. Adj.Sig.
Statistic
unlikely-somewhat likely .486 .486 1.000
unlikely-very likely 6.600 .010 .031
somewhat likely-very likely 2.423 .120 .359
Each row tests the null hypothesis that the Sample 1 and
Sample 2 distributions are the same.
Asymptotic significances (2-sided tests) are displayed. The
significance level is .05.
Figure 3.7. Pairwise comparison of Condom_Request.

In addition, the pairwise tests do not show a significant difference

between unlikely and somewhat likely, and somewhat likely and very
likely categories.
A review of Figure 3.7 shows that the unlikely-very likely pairwise
comparison is significant with an adjusted p = .031, which is smaller
than the α = .05 threshold. The other comparisons have adjusted signifi-
cance values of 1.0 and .359, both larger than the significance threshold.
Therefore, the median levels of self-efficacy are significantly different for
only one of the three pairwise comparisons.
In statistical terms, the median test (p = .035) rejects the null
hypothesis of equal median values at α = .05 for self-efficacy between
three respondent categories on likelihood of requesting a partner to use
a condom (unlikely, somewhat likely, and very likely). A pairwise test
comparison identified one category pair (i.e., unlikely and very likely)
that are significantly different (adjusted p = .031).
Practically, this means that levels of self-efficacy are different in
Dr. Rutledge’s sample for people grouped by their willingness to ask
their partners to use a condom, and the findings provide support for
further research into the association of self-efficacy with risky sex
behavior.
4
Comparing Two or More

Related Groups
CHAPTER OBJECTIVES
Chapter 4 presents information about comparing two or more groups
that relate to one another. The chapter covers the following topics:
• Determining dependent (related) and independent (unrelated)

groupings
• Selecting tests for two dependent groups (similar to the paired
t-test for dependent samples)
• Comparing scale scores from pretests and posttests —Wilcoxon
matched pair signed rank and sign tests
• Comparing dichotomous scores/responses for two related
groups — McNemar’s test
• Comparing ordinal (Likert-t ype) scores on two related
groups —marginal homogeneity test
• Selecting tests for more than two dependent groups (similar to
ANOVA and MANOVA)
107
• Comparing ordinal (dichotomous) scores on three or more

related groups —Cochran’s Q test
• Comparing scale scores on three or more related groups —
Friedman’s 2-way ANOVA by ranks test
INTRODUCTION
Chapter 4 covers information about comparing related samples and the
way you analyze these data. Before beginning, if you are new to research,
it might be helpful to clarify the use of the word sample. In this chap-
ter, when comparing samples that are related, the term samples refers to
collecting data multiple times from the same group of people, not to be
confused with the use of the word sample to mean a subset of a target
population (i.e., a sample of the population). It is very important to rec-
ognize when the term sample relates to a single data collection effort from
one group of respondents, as described in Chapter 3, a sample from a
population, and when the term samples relates to multiple data-collection
activities from one group of respondents.
Prior to analyzing data for related samples, it is very important to
make sure your data are identified with the appropriate level of measure-
ment in SPSS and are in the appropriate format for related-samples kinds
of tests. For example, in Research Question #12, the data format requires
that responses for the two nominal variables for each respondent fall on
the same row on the data spreadsheet, as they might for pre and post
scores for the same respondent. This lets SPSS know that the pre and post
scores are related to a single person or a single respondent.
As a reminder, to avoid any unexpected surprises or worse, report a
finding not supported by the data, be thorough in your approach, and
examine the information provided by each variable individually before
moving forward with an analysis that involves multiple variables. This
includes examining categorical variables to verify that the appropriate
response coverage is present (i.e., a minimum of 5 responses for each
category), and reviewing a graphical depiction of the distribution of
scale variables to examine the theoretical/conceptual validity of the data
in the variable. Chapter 1 describes a few things to consider when exam-
ining variables individually, and Chapter 2 provides multiple analytical
procedures to assist the examination.
Comparing Two or More Related Groups 109
Like Chapters 2 and 3, Chapter 4 starts with a scenario describing a

hypothetical line of research that we can explore with research questions.
Box 4.1 presents the hypothetical research study for Chapter 4.
Box 4.1 Research Scenario for Chapter 4

Dr. Fleming’s study is a secondary data analysis of a longitudinal study
of youth in a major metropolitan area, and you are his statistician on the
project. The original researchers utilized a systematic probability sampling
strategy to collect data from public school students in four waves, beginning
with children in third grade and surveying the students every three years.
Dr. Fleming’s interest was in the influence of gang membership on several
outcomes, so he determined that data from all four waves —waves 1 (grade
three), 2 (grade six), 3 (grade nine), and 4 (grade twelve) —would be suited
for his needs. Because of attrition, the original sample size of 1000 decreased
by 10% with each data collection, so wave 2 had 900 participants, wave 3
had 810 participants, and wave 4 had 729 participants. Of the 729 youth
who responded to all four data collections, 3.4% (N = 25) reported that they
eventually had joined a gang in one of the four waves of data collection and
remained in the gang through wave 4.
Chapter 4 contains seven research questions to guide the presen-

tation of the nonparametric procedures related to the above scenario.
Each research question section describes the reasoning for using non-
parametric procedures. Table 4.1 lists the nonparametric procedures
presented in Chapter 4, the kinds of variables for each procedure, and
available online.
MCNEMAR’S TEST
McNemar’s test uses combination frequencies to determine if yes/
no kinds of responses at two different times are different. The com-
binations are constructed from responses captured by two dichot-
omous variables (i.e., yes-yes, yes-no, no-yes, and no-no). The 2 × 2
contingency table structure of McNemar’s test is similar to Pearson’s
chi-square test; however, the chi-square test is for independent sam-
ples (i.e., unpaired data) and not very effective when dealing with
Question Test in DB
12 McNemar’s Two dichotomous Joined_gang_

variables w2 and
Violence_
crime_w2
13 Marginal Two categorical variables Friends_
homogeneity (i.e., ordinal variables w2 and
representing two related Friends_w4
responses, such as pre
and post values)
14 Sign test Two continuous variables Anxiety_
(i.e., paired scale w2 and
variables representing Anxiety_w4
two related responses,
such as pre and post
values for respondents)
15 Wilcoxon Two continuous variables Anxiety_
matched-pair (i.e., paired scale w2 and
signed-rank variables representing Anxiety_w4
Hodges-Lehmann two related responses,
confidence such as pre and post
interval values for respondents)
16 Cochran’s Q Three or more Assault_w2,
dichotomous variables Assault_w3,
and Assault_
w4
17 Kendall’s Three or more continuous Support_w1,
coefficient of variables (i.e., scale Support_w2,
concordance variables representing Support_
three or more related w3, and
responses, such as Support_w4
pre-observation, second
observation, third
observation, etc., for
respondents)
18 Friedman’s Three or more continuous Support_w1,
2-way variable (i.e., scale Support_w2,
ANOVA by variables representing Support_w3,
ranks three or more related and Support_
responses, such as w4
pre-observation, second
observation, third
observation, etc., for
respondents)

Dr. Fleming is curious about the reasons youth join gangs, and previous
studies suggest that youth who are victims of a violent crime may be more
likely to join a gang for protection. To test this in his sample, Dr. Fleming
wants you to compare responses to two questions in wave 2 —“Have you ever
been the victim of a violent crime?” and “Have you joined a gang?” None of
the respondents at wave 1 selected for this analysis was in a gang or had been
a victim of a violent crime, thereby providing an opportunity to explore a
relationship between violence and joining a gang.
Does being a victim of a violent crime increase the probability that you will
join a gang?
small sample sizes (i.e., less than 100 records). Another related test
is the paired-sample t-test, but the paired-sample t-test is only for
continuous variables that meet the normality assumption. Because
McNemar’s test uses two dichotomous variables, the null hypothesis
is that the marginal probabilities for each of the two data point out-
comes are equal. This means, for example, that the probability that a
Hispanic/L atino person will answer yes to a specific question is equal
to the probability that a non-Hispanic/L atino person will answer yes
to the same question.
Other examples of uses for McNemar’s test are as follows:
1. A researcher wants to know if a certain drug can keep people

from getting the flu, so she conducts a small pilot study.
2. In a feasibility study for a larger research agenda, a researcher
wants to test if those who regularly participate in risky drinking
(also known as binge drinking) also regularly participate in
risky sexual behaviors so he can decide if these variables are
appropriate for the larger study.
The variables needed for the McNemar’s tests for Question #12 are two
dichotomous variables representing two related responses —e.g., pre
and post values for respondents. For this question, the dichotomous
variables are victim of a violent crime and joined a gang in wave 2 (i.e.,
Violent_crime_w2 and Joined_gang_w2).
SPSS Process
□ Select => Analyze => Nonparametric Tests => Related

Samples….
After the Nonparametric Tests: Two or More Related Samples win-

dow opens, SPSS needs to know which variables to use in the analysis.

□ Click to select the first variable, Violent_crime_w2, found in the
Fields: area.
□ Click the move arrow to move the variable to the Test
Fields: area.
□ Click to select the second variable, Joined_gang_w2, found in
the Fields: area.
Fields: area.
Now that SPSS knows what variables to use, you need to choose the
kind of nonparametric test to conduct.
□ Select => Settings tab.

□ Select => Choose Tests found in the Select an item: area.
□ Click to select Customize tests to show all the available tests for
related samples (see Figure 4.1).
Figure 4.1 shows many different tests, some for two related observa-
tions, some for three or more related observations, some for categorical
variables, and some for scale variables. A close examination of Figure
4.1 reveals four tests that are specific to having only two related obser-
vations. Because Research Question #12 involves binary data (i.e., two
dichotomous variables), you should select McNemar’s test.
□ Click to select the checkbox next to the McNemar’s test in the

Test for Change in Binary Data area.
Figure 4.1. Nonparametric Tests: Two or More Related Samples Window.
Note: When you select Choose Tests in the Select an item: area, you will
see one option that allows SPSS to decide automatically which tests are
appropriate based on the variable definitions (i.e., identified by the level
of measurement). It is prudent to monitor the analysis process closely
by always choosing Customize tests and then selecting the specific test
you want to conduct.
Before running McNemar’s test, SPSS needs to know which value in
the dichotomous variables is associated with success. SPSS uses the ter-
minology of success and failure to identify the two options in dichoto-
mous variables. You should keep track of the values identified as success
to verify the use of your data and to understand how best to interpret
the test results.
□ Select => Define Success… to open the McNemar’s Test: Define

Success window.
Once the window opens, two options are displayed. The default option
is to let SPSS use the First value found in data. To be fully aware of how
the data are defined and used,
□ Click to select the Combine values into success category option.

After making this selection, you gain access to the Success: area.
□ Click on the white space below the word Value and enter a 1.
Entering a 1 tells SPSS to associate the number 1 with a success.
□ Select => OK to close the window.
Note: If the two variables had more than two outcomes that represent
success, multiple lines in the white space identify the success outcomes.
Recall, McNemar’s test is based on dichotomous events. Having the
capability to combine values expands the usability of McNemar’s test to
categorical variables with more than two values.
To continue your analysis,
□ Select => Test Options in the Select an item: area to see

where to change the default significance level of 0.05 for
tests, the default 95.0% confidence interval range, and the
default on how to handle missing data, which is Exclude cases
test-by-test.
the analysis.
SPSS Output
window, including:

window

Figure 4.2. Model Viewer window for McNemar’s test.
the analysis (see Figure 4.2).
The Model Viewer window has access to more than the default view
shown in Figure 4.2. At the bottom of the window are dropdown lists on
the left for View: and Field Filter: and dropdown lists on the right for
Test:, Field(s):, and View:.
□ Changing the default View: on the right from Related Samples

Test to Categorical Field Information is necessary so you can
verify that you constructed the data in the analysis as intended —
an important part of any analysis before providing any findings.
A review of the Model Viewer information provides a few interest-

ing observations.
• Total N has a value of 23 even though the number of respondents

for Research Question #12 was 25. This is due to the influence of
missing data and the fact that Exclude cases test-by-test option
was selected. The option for how to handle missing data is set
under Test Options in the Select an item: area.
• The footnote on the right side of the Model Viewer window

explains that a different statistical distribution testing for
significant differences was used (i.e., Binomial) because there are
fewer than 25 records. The footnote explains the use of counts in
observed frequency rather than the number of respondents. Due
to the use of the Binomial distribution, the analysis provides an
exact p-value, giving the probability that the responses for the
two variables are random.
• In addition to switching views by changing dropdown values,
hovering the mouse over the bar charts shows additional
information related to the test. Figure 4.2 displays information
for combination of victim of a violent crime and did not join
a gang if you hover the mouse over the area. The difference
between the observed and expected values is used to determine
if the frequencies for the combination of responses are not
random, suggesting a relationship between the two variables.
Findings
The Hypothesis Test Summary information on the left side of the
Model Viewer window reveals the exact significance of the test,
p = .031. The exact significance values suggest only a 3.1% chance that
being a victim of a violent crime is unrelated to the probability of join-
ing a gang. The values in the bar charts in Figure 4.2 reveal that every
respondent who joined a gang in wave 2 was a victim of a violent crime.
Therefore, the data suggest that joining a gang is related to being a vic-
tim of a violent crime.
In statistical terms, McNemar’s test rejects the null hypothesis of
equal marginal probabilities at α = .05 (p = .031). The small sample
and use of the Binomial distribution in the statistical analysis provide
an exact significance value (p = .031). The p-value indicates that if the
null hypothesis is true, there is only a 3.1% chance of equal success
probabilities.
Practically, this means that you found, in Dr. Fleming’s sample of
adolescents, a significant relationship between adolescents’ joining a
gang and having been a victim of a violent crime. Not explained in this
analysis is if their reason for joining the gang is for protection.

Dr. Fleming has noticed that other researchers use a friendship measure as
a proxy for the degree to which a youth is integrated in a gang. His data set
includes an item that asks respondents, “How many of your friends are in
a gang?” with response options of a few, many, or most. He expects that, in
his sample, the numbers of participants reporting many and most should
increase over time, but he isn’t certain. To test this, he asks you to compare the
data from wave 2 and wave 4. If the numbers of gang friends increase as he
suspects, this may be a valuable variable to use in subsequent analyses (e.g.,
investigating social support, peer influence).
Are the number of adolescents reporting that their friends are in a gang
increasing over time?
MARGINAL HOMOGENEITY TEST
The marginal homogeneity test is an extension of McNemar’s test in that it
addresses two ordinal categorical variables rather than two dichotomous
variables. The number of categories can be three or more, but, as described
in Chapter 2, the more categories you have, the more responders you need
to be sure each category has enough coverage to be used in an analysis.
The marginal homogeneity test uses response combination frequencies to
determine if the number of positive outcomes has changed, and because
the test uses two categorical variables, the null hypothesis is that the mar-
ginal probabilities for each of the two data point outcomes are equal.
Other examples of uses for the marginal homogeneity test are as
follows:
1. A clinician believes that an event in his therapy group increased

the participants’ level of anxiety. Prior to the group’s first
session, he had participants respond to an anxiety measure with
4-point Likert-like response options. He asked them to respond
to the same measure after the event to learn if his participants’
level of anxiety increased.
2. A social services administrator believes that she has improved
her organization’s services, and that clients’ satisfaction should
reflect this. At regular intervals, clients fill out satisfaction surveys
based on the following response options: not satisfied, somewhat

satisfied, mostly satisfied, and very satisfied. She wants to compare
earlier responses with the most recent responses to find out if the
clients are more satisfied now.
For Research Question #13 (see Box 4.3), the categorical variables are
Friends in wave 2 and wave 4 (Friends_w2 and Friends_w4). These vari-
ables identify the number of the respondent’s friends who are in a gang,
with categorical response options of few, many, and most. A review of the
two categorical variables shows that coverage is just above the five responses
per category guideline. However, although the data meet the coverage
guideline, the frequency for each category is small, which does not provide
much information for identifying differences. Nevertheless, when analyz-
ing small datasets, your only option is to use the best test that matches your
situation while making sure the data meet all the related assumptions.
SPSS Process
□ Select => Analyze => Nonparametric Tests => Related Samples.
Once the Nonparametric Tests: Two or More Related Samples window

opens, SPSS needs to know which variables to use in the analysis.

□ Click to select the first variable, Friends_w2, found in the
Fields: area.
Fields: area.
□ Click to select the second variable, Friends_w4, found in the
Fields: area.
Fields: area.
Now that SPSS knows what variables to use, the type of nonparametric
test to conduct needs to be chosen.


□ Because Research Question #13 involves categorical data (i.e.,
two ordinal variables), click to select the checkbox next to

the Marginal Homogeneity test in the Test for Change in
Multinomial Data area.
Unlike the important step in Research Question #12 to identify values

related to success, the marginal homogeneity test does not use a success
distinction.

the analysis.
SPSS Output
window, including:

output window

the right side of the output window.
To see the details of the test, double-click within the area outlined
by the Model Viewer (i.e., point the mouse at the Hypothesis Test
Summary information and double-click). The Model Viewer window
will open to display the detailed information for the analysis. Use the
dropdown lists at the bottom of the Model Viewer window to verify that
the data used in the analysis are what you expected and supported the-
oretically. Hovering the mouse over the bar charts shows observed fre-
quencies related to the test.
Findings
The Hypothesis Test Summary information on the left side of the Model
Viewer window shows a Sig. value of .414. This Sig. value is larger than the
α = .05 threshold for significance. Therefore, the data suggest that there is not
enough information to determine that the number of friends in a gang has
changed from wave 2 to wave 4. However, it is important to note that, with so
little information for the combinations of categories, you would most likely
need a sizable difference (effect size) for any statistical test to find significant
differences. With so few respondents, you simply do not have enough power.
In statistical terms, the marginal homogeneity test fails to reject the
null hypothesis of equal marginal probabilities at α = .05 (p = .414).
Practically, this means that there is not enough information in the
responses from the few adolescents to conclude that there is an increase
in the number of reported friends in gangs over time.

Dr. Fleming has reviewed studies investigating anxiety among gang members.
Some studies show that anxiety increases with the length of gang membership,
while other analyses demonstrate that anxiety decreases over time. To
investigate this question with his sample, Dr. Fleming asks you to compare
participants’ scores on a validated anxiety measure at wave 2 (i.e., early gang
membership) with their scores on the anxiety measure at wave 4 (i.e., later
gang membership).
Does the amount of reported anxiety among gang members change over time?
SIGN TEST
The sign test uses the positive and negative differences for the matched
pairs of values from two continuous variables to determine if the two
sample medians have equal value. The null hypothesis is that the median
of the calculated differences for the two samples is zero (i.e., the initial
premise is that the median of one group is equal to the median of the sec-
ond group). The equivalent parametric test is the related-sample t-test,
but this test assumes that the difference scores are normally distributed in
the population, whereas the sign test has no distribution assumption. The
lack of a distribution assumption does make the sign test more applicable
to a variety of data situations, especially with smaller samples for which

you cannot verify the normality assumption. However, this also makes
the test more conservative in finding significant differences between the
two samples, meaning that the test is less likely to detect a statistically
significant difference.
Other examples of uses for the sign test are as follows:
1. A researcher testing a new depression intervention on a small

sample wishes to learn if his participants’ depressive symptoms
scale scores have changed from six months earlier.
2. A hospital administrator wishes to find out if the frequency of
different diagnoses among his inpatient mental health unit has
changed from the previous year.
The variables needed for the sign test are two paired scale or continuous
variables representing two related responses, such as pre and post values
for respondents. For Research Question #14 (see Box 4.4), the scale vari-
ables are Anxiety in waves 2 and 4 (i.e., Anxiety_w2 and Anxiety_w4).
SPSS Process


□ Click to select the first variable, Anxiety_w2, found in the
Fields: area.
Fields: area.
□ Click to select the second variable, Anxiety_w4, found in the
Fields: area.
Fields: area.

□ Click to select Choose tests from the Select an item: area.

□ Because Research Question #14 involves scale data (i.e., two
continuous variables) click to select the checkbox related

to the Sign test in the Compare Median Difference to
Hypothesized area.
Unlike the step in Research Question #12 in which you identify val-
ues related to success, the sign test does not use a success distinction.
Therefore, no options button is available other than the Test Options
found in the Select an item: area.

the analysis.
SPSS Output
window, including:

output window

the analysis.
□ Use the dropdown lists at the bottom of the Model Viewer

window to verify that the data used in the analysis are what you
expected and are supported theoretically.
Findings
The histogram shows that nine of the paired values had a positive dif-
ference and 10 had a negative difference. In the total of 25 records, one
pair had missing data and five pairs were equal. Note that the statistical
calculations do not include pairs with zero differences. The number of
positive-difference pairs to negative-difference pairs suggests that there
is no information to indicate an increase in anxiety levels. In addition,
the sign test shows a Sig. value (1.000) that is the largest significance
value a test can return. Therefore, the sample of adolescents does not
indicate any increase in anxiety from wave 2 to wave 4.
However, if you look at the histogram on the right side of the Model
Viewer window (see Figure 4.3), the bars for the negative differences are
in a very different pattern from the bars for the positive differences. This
does suggest that something is happening related to anxiety levels.
Figure 4.3 reveals that the distribution of negative differences is
much wider than the distribution of positive differences. Knowing that
the sign test uses rankings and not difference values to test for equal
medians, you might consider that another test may offer more insight
into anxiety values over time (see Research Question #15).
10.0
8.0
Frequency
6.0
4.0
2.0
0.0
–6.00 –4.00 –2.00 0.00 2.00
Anxiety_w4 - Anxiety_w2
Positive Differences (N = 9) Negative Differences (N = 10)

(Number of Ties = 5)
Figure 4.3. Histogram for anxiety levels for wave 2 and wave 4.

For Research Question #14, using statistical terms, the sign test fails
to reject the hypothesis of equal medians at α = .05 (p = 1.00), indicat-
ing that there is not enough information to indicate that the median
anxiety levels for wave 2 are different from the median anxiety levels
for wave 4.
Practically, this means that no information suggests that adoles-
cent anxiety levels are changing. However, a review of the histogram
in Figure 4.3 reveals that adolescents who report drops in their anxiety
level experience a larger drop, on average, than the increase for adoles-
cents reporting an increase in anxiety. Further analysis of these data is
warranted to fully understand what the data suggest.

After reviewing the findings from Research Question #14 about anxiety
in waves 2 and 4, Dr. Fleming wants to take the additional step of using a
different statistical procedure more suited to variations in difference values
between related observations. Thus, he requests that you return to SPSS to
rerun the analysis using a different statistical procedure.
Is the median difference in anxiety level between waves 2 and 4 equal to zero?
WILCOXON MATCHED-PAIR SIGNED-R ANK TEST

The Wilcoxon matched- pair signed- rank test, also called Wilcoxon
signed-rank test, is different from the Wilcoxon rank-sum test. The
Wilcoxon matched-pair signed-rank test determines if median values
from two samples are equal. The null hypothesis is that the median differ-
ence between related pairs is zero. Like the sign test in Research Question
#14, the Wilcoxon matched-pair signed-rank test uses the frequency of
positive and negative matched pairs value differences, but includes the
ordering of these differences to test if median values are equal. This addi-
tional step makes the Wilcoxon test more powerful in detecting differ-
ences, but it does have an added assumption that you need to test. The
assumption is that the median differences values are symmetrical about
the median of the difference values.
Other examples of uses for the Wilcoxon matched-pair signed-rank

test are as follows:
1. A researcher conducts a pilot study of a self-esteem intervention,

and she wishes to test if the self-esteem scores immediately after
the intervention are the same as when measured three months
after the intervention ends.
2. A new manager believes he is doing well in his new role, but
he wonders if the clients have any new concerns about his
organization. He decides to examine the differences between
the clients’ satisfaction levels while his predecessor was in the
managing role and their satisfaction levels now that he has been
manager for six months.
HODGES-LEHMANN CONFIDENCE INTERVAL

The Hodges-Lehmann confidence interval is based on the Wilcoxon
matched-pair signed-rank procedure that calculates an upper and lower
limit of a confidence interval for median difference values. The confi-
dence interval range is based on an alpha, typically of 0.05, representing
a 95% confidence interval.
The variables needed for the Wilcoxon matched-pair signed-rank test
and the Hodges-Lehmann confidence interval estimate are two continu-
ous variables —paired scale variables representing two related responses,
such as pre and post values for respondents. For Research Question #15
(see Box 4.5), the scale variables are Anxiety_w2 and Anxiety_w4.
Along with rerunning the analysis using the Wilcoxon matched-pair
signed-rank test, Dr. Fleming wants a confidence interval. The confidence
interval will provide a potential range in median differences between the
two waves of data. In addition, knowing that the sign test Sig. value was
1.00, Dr. Fleming requests that you adjust the Significance level option
from .05 to .10, because SPSS’s Hodges-Lehmann procedure produces a
confidence interval only if the test finds significance.
SPSS Process
To adjust the significance level value,
□ Select => Settings tab to see the options in the Select an

item: area.
□ Change the value displayed for the Significance level: area from
0.05 to 0.10.
Next, SPSS needs to know which variables to use in the analysis.

If the anxiety variables are no longer in the Test Fields: area,
□ Click to select the first variable, Anxiety_w2, found in the

Fields: area.
Fields: area.
□ Click to select the second variable, Anxiety_w4, found in the
Fields: area.
Fields: area.

□ Select => Choose tests from the Select an item: area.
Because Research Question #15 involves two scale/continuous variables,
□ Click the checkbox next to the Wilcoxon matched-pair

signed-rank test in the Compare Median Difference to
Hypothesized area.
□ Click the checkbox next to Hodges-L ehmann to request a
confidence interval of the median difference between the two

variables.
Unlike the step in Research Question #12 in which you identify values
related to success, the Wilcoxon matched-pair signed-rank test does not
use a success distinction.
□ After carefully reviewing your choices, select => Run to

conduct the analysis.
SPSS Output
window, including:

output window

the analysis (see Figure 4.4).
□ Use the dropdown lists at the bottom of the Model Viewer

window to verify that the data used in the analysis are as you
expected and are conceptualized correctly for the analysis.
Findings
The histogram in Figure 4.4 is the same as the histogram in Figure 4.3.
As before, the number of positive difference pairs to negative differ-
ence pairs suggests that there is no increase in median anxiety levels.
However, instead of looking at the pairs, a review of the value differ-
ences suggests a different finding. The Wilcoxon signed-rank test finds
Figure 4.4. Model Viewer window for the Wilcoxon matched-pair test.
that median of the differences does not equal zero, indicating that there
is a difference between anxiety levels for wave 2 and wave 4 (Sig. value
of .062, which is smaller than α = .10).
To view the upper and lower bounds of a 95% confidence interval for
median difference values,
□ Click on the dropdown list next to the View: area on the bottom
left of the Model Viewer window.
□ Click to change Hypothesis Summary View to the Confidence
Interval Summary View.
After changing to the Confidence Interval Summary View, you will

find the lower and upper values for a 95% confidence interval for the dif-
ference between the median values. The Confidence Interval Summary
indicates that the Hodges-Lehmann procedure calculated a median dif-
ference estimate of –1.00 and a 95% confidence interval of –2.00 to 0.50.
These results suggest that the median drop in reported adolescent anxi-
ety levels from wave 2 to wave 4 is –1.00, and that the plausible range of
median changes in anxiety is from –2.00 to 0.50.
Before reporting on the test results, you must verify that the data
meet the median difference assumption for the Wilcoxon matched-pair
signed-rank test. Recall, this assumption is that the distribution for the
median difference values is symmetrical about the median.
To verify that the data meet the assumption, you need to generate a
new variable from the two samples and review its distribution.
SPSS Process
□ Select => Transform from the main SPSS menu.
□ Select Compute Variable… to open the Compute Variable
window.
□ Enter Median_DIFF in the white space below the Target
Variable: area.
□ Click to select the variable Anxiety_w2 from the list of variables

□ Click the move arrow to move the variable to the Numeric
Expression: area.
□ After Anxiety_w2 in the Numeric Expression: area, enter a
space, minus sign, and a space.

□ Click to select the variable Anxiety_w4 from the list of variables

□ Click the move arrow to move the variable to the Numeric
Expression: area.
The above steps tell SPSS to create a new variable called Median_DIFF
by calculating the anxiety difference value for each respondent (i.e.,
Anxiety_w2 – Anxiety_w4)
□ Select => OK to compute the new variable.
After the new variable is computed, you need to ask SPSS to graph the
new computed values.
□ Select => Analyze => Descriptive Statistics => Frequencies to

□ Click to select the new computed variable Median_DIFF from
the list of variables on the left side of the window.

□ Click the move arrow to move the variable to the
Variable(s): area.
□ Select => Statistics… to open the Frequencies: Statistics window.
□ Click to select the Median option in the Central

Tendency area.
□ Select => Continue to close the window.
□ Click to select the Histograms: option in the Chart Type area.
□ Click to select the Show normal curve on histogram option.
□ Select => OK to ask SPSS to create the graph of the distribution.
Note: The steps above include a request to draw a normal curve on the
graph. The assumption you are attempting to verify refers to the sym-
metry about the median and NOT an assumption of normality. The
reason for requesting SPSS to draw a normal curve is only to assist in
visually interpreting the symmetry and NOT to compare the distribu-
tion against a normal curve.
SPSS Output
Figure 4.5 shows the graph of the computed variable Median_DIFF.
A visual examination of the graph reveals a departure from symmetry
Histogram
10 Mean = 1.13
Std. Dev. = 2.383
N = 24
8
6
Frequency
0
–2.00 0.00 2.00 4.00 6.00 8.00
Median_DIFF
Figure 4.5. Distribution of value differences for two related samples.
about the median value of 0.00. Is the departure enough to determine

that the data do not meet the assumption? This kind of question can be
challenging, and this example is no exception. If you are certain that the
data do not meet the assumption, then a different analytic process must
be used to answer the research question. However, if you are uncertain,
then use additional procedures to find out if they agree. In the case of
the Wilcoxon matched-pair test, the sign test is a close relative and can
give you additional insight into the data. The bottom line is that it is your
responsibility as a researcher to gather enough statistical information to
be comfortable with your results. Most of the time, this will include mul-
tiple statistical procedures that examine your data from different angles,
because one statistical test rarely provides enough evidence for a conclu-
sion. Even running the same statistical procedures on a subset of your
data is more informative than a single stand-alone test.
In this situation, the distribution for Median_DIFF is not symmet-
rical about the median. Therefore, the relatively balanced number of
values on each side of the median, the upper and lower values of the
Hodges-Lehmann confidence interval, and the size of the p-value from
the sign and Wilcoxon signed-rank test indicate that there is not enough
evidence to conclude the two samples are different. If the distribution
for Median_DIFF were closer to being symmetrical about the median,
this decision would have been much more difficult to make. However,
the skewness shown in Figure 4.5 and the significance level of the sign
test at 1.00 are clear indications that you cannot reject the hypothesis of
similar median values.
In statistical terms, the data do not meet the required assumption of
a symmetrical distribution of difference values for the Wilcoxon signed-
rank test. Therefore, the findings of significance at α = .10 (p = .062) may
not be valid. The Hodges-Lehmann confidence interval procedures
provided an estimated mean difference of –1.00 and upper and lower
bound values for a 95% CI for median difference of –2.00 to 0.50, which
includes the possibility of a median difference value of 0, thereby fail-
ing to reject a hypothesis of equal median values. Recall, the confidence
interval values provide a range in which we are 95% confident that the
true median value falls between –2.00 and 0.50.
Practically, this means that you do not have enough information to
suggest that adolescent anxiety has changed from wave 2 to wave 4.
Note: Do not forget to reset your Significance level: value from .10 to .05
if you prefer a more restrictive alpha level.

Dr. Fleming wonders if personal assaults committed by youth in his sample of
gang members increase, decrease, or show no pattern across data collection
waves. To investigate this, he asks you to analyze the participants’ responses
to the question, “Have you been involved in the commission of an assault since
the last time we interviewed you?” for waves 2, 3, and 4.
Is the frequency of committed assaults changing over time in this sample of
gang members?
COCHRAN’S Q TEST
Cochran’s Q test uses frequency of success and failure for k-related sam-
ples/observations to determine if the number of successes/failures is dif-
ferent among the samples. For Conchran’s Q, success could represent
testing positive on a medical test, answering yes to a yes/no question, or
endorsing, rather than not endorsing, a candidate. Comparing samples
to find potential differences is similar to repeated measures ANOVA, but
with dichotomous variables. Remember that repeated measures ANOVA
compares mean values and assumes normality for the sample values
within groups. Before conducting Cochran’s Q test, you must verify that
you have enough data for all the responses to have enough information
for a statistical test, and that the values are independent within groups
(i.e., the relationships among the values exist only across groups/waves of
data). The variables needed for Cochran’s Q test are three or more dichot-
omous variables representing three or more related response values for
respondents (e.g., observation 1, observation 2, observation 3). Because
Cochran’s Q test uses dichotomous variables, the null hypothesis is that
the distribution of frequencies for each of the k data point outcomes is
similar.
Other examples of uses for Cochran’s Q test are as follows:
1. A researcher conducts a pilot study of a brief alcohol-related

intervention for college students. To test the intervention’s
effectiveness after the conclusion of the intervention, he collects
data on their drinking at three, six, and twelve months.
2. A researcher is training four research assistants to score a
recorded interview, and she wishes to test if the raters
agree with each other or if they are more often not in

agreement.
For Research Question #16 (see Box 4.6), the dichotomous variables are
Committed an Assault for waves 2, 3, and 4 (i.e., Assault_w2, Assault_w3,
and Assault_w4). These three variables represent three time points in which
data were collected from adolescents, asking them if they had participated
in an assault on another individual since the last time interviewed.
SPSS Process
After the Nonparametric Tests: Two or More Related Samples window


□ Click to select the first variable, Assault_w2, found in the
Fields: area.
Fields: area.
□ Click to select the second variable, Assault_w3, found in the
Fields: area.
Fields: area.
□ Click to select the third variable, Assault_w4, found in the
Fields: area.
Fields: area.
Note: Another method to move variables from one side of the window to
the other side is to point your mouse at the variable name and double-
click. SPSS interprets the double-click as a message to move the variable
to the opposite area.
Now that SPSS knows what variables to use, you must choose the
kind of nonparametric test to conduct.

Because Research Question #16 involves binary data (i.e., three

dichotomous variables),
□ Click to select the checkbox next to Cochran’s Q.
Before running the Cochran’s Q test, SPSS needs to know which value in
the dichotomous variables are associated with a success.
□ Select => Define Success… to gain access to the Cochran’s Q

Test: Define Success window.
After the window opens, two options are displayed. The default option
is to let SPSS use the First value found in data.
□ Click to select the Combine values into success category option

to be fully aware of how the data are defined and used.
After making this selection, you gain access to the Success: area.
□ Click on the white space below the word Value and enter a
1. Entering a 1 tells SPSS to associate the number 1 with a
success.
□ Select => OK to close the window.
Note: If the variables have more than two outcomes that represent
success, then the success outcomes are identified on multiple lines
in the white space. Recall, Cochran’s Q test is based on dichotomous
events. Having the capability to combine values expands the usabil-
ity of Cochran’s Q test to categorical variables with more than two
values.
You must set one more option for the Cochran’s Q test, the Multiple
comparisons: option. The default is for All pairwise and is shown in
the white area for the dropdown list under Multiple comparisons:. The
other options are Stepwise step-down or None; click on the dropdown

to show all three options. To compare the medians of all the possible
waves in the data, set the dropdown list to All pairwise.
Note: When the number of comparisons is too large, requesting all pos-
sible pairwise combinations may be problematic due to the number of
comparisons to review. When this is the case, the Stepwise step-down
option may be a better choice. This option sorts the groups (e.g., waves
of data) by their calculated group median value and compares only adja-
cent groups, making the comparison of groups manageable. However,
with only three groups, the comparison of all pairwise combinations is

the analysis.
SPSS Output
window, including:

output window

the analysis.
□ Using the View: dropdown list at the bottom right of the Model
Viewer window, select Pairwise Comparisons to display the
information shown in Figure 4.6.
Pairwise Comparisons
Assault_w2
7.00
Assault_w3
12.00
Assault_w4
15.00
Each node shows the sample number of successes.

Test Std. Std. Test
Sample1-Sample2 Sig. Adj.Sig.
Statistic Error Statistic
Assault_w2_Assault_w3 –.208 .136 –1.531 .126 .377
Assault_w2_Assault_w4 –.333 .136 –2.449 .014 .043
Assault_w3_Assault_w4 –.125 .136 –.919 .358 1.000
Each row tests the null hypothesis that the Sample 1 and Sample 2 distributions
are the same.
Figure 4.6. Pairwise Comparisons for Cochran’s Q.
Findings
The Adj. Sig. values shown in Figure 4.6 indicate that the only signif-
icant pairwise difference for assault frequencies is from wave 2 (i.e.,
Assault_w2) to wave 4 (i.e., Assault_w4). The other pairwise differences
are not significant (Adj. Sig. values of .377 and 1.00, both greater than
the α = .05 threshold).
Hovering the mouse over the lines in the Model Viewer’s Pairwise
Comparisons graph will display the related Adj. Sig. values. Therefore,
what you can draw from the Adj. Sig. values is that the number of respon-
dents committing assaults increased at each wave and by wave 4 adoles-
cents committing assaults had significantly increased from wave 2.
In statistical terms, Cochran’s Q test (p = .047) rejects the null hypoth-
esis that the number of respondents committing assaults has not changed
at α = .05 for the three waves of collecting data from the sample of gang
members. In addition, a test of equal probabilities was conducted for all
pairwise comparisons. The adjusted significance for multiple tests was
calculated, and it revealed that a single pair —wave 2 and wave 4 —was
significantly different (p = .043). The remaining pairs, wave 2 and wave 3
(p = .377), and wave 3 and wave 4 (p = 1.00) were not significant at α = .05.
Practically, this means that as a group in the sample, the number par-
ticipating in assaults significantly increased from wave 2 to wave 4. The
frequency changes from wave 2 to 3 and wave 3 to 4 are not large enough
to claim a significant difference. In addition, viewing the frequencies at
each wave indicates the change is from fewer assaults in wave 2 to signif-
icantly more in wave 4. Another interesting relationship to investigate
would be to look at joining a gang and committing assaults.

Dr. Fleming learned from the literature that gang members report staying
in gangs because of the social support they offer. He hypothesized that this
feeling of support would increase over time, as the youth became more
integrated with their gang. To investigate this hypothesis, Dr. Fleming asks
that you compare the participants’ responses to a validated measure of social
support at baseline (i.e., in grade three) with the responses at the subsequent
three time points.
Is the level of social support reported by the adolescents changing across the
four waves of data collection?
KENDALL’S COEFFICIENT OF CONCORDANCE

Kendall’s coefficient of concordance, also called Kendall’s W, is a calculated
value or coefficient that represents the amount of similarity among samples.
The value range for the coefficient is from 0 for no similarity to 1 for com-
plete similarity. Kendall’s coefficient requires three or more continuous/
scale variables that represent three or more samples, in which the value (i.e.,
row) in the sample is somehow related to the same row in the other sam-
ples. The commonly used parametric procedures associated with Kendall’s
coefficient is the Pearson correlation coefficient, but the Pearson correlation
coefficient assumes that the variables come from a normally distributed
population of values, and it can compare only two variables in a test. As
mentioned, Kendall’s coefficient handles three or more variables and its only
assumption is that the samples are dependent. Kendall’s coefficient uses
rankings and the number of ties to look for trends among the observations
to determine if the trends in the ranks are different among the samples. The
null hypothesis is that the distribution of the samples is the same.
Other examples of uses for Kendall’s coefficient of concordance are
as follows:
1. A researcher asks three or more raters to score a small group

of research participants. Each rater (i.e., sample) scores each
individual in a group (i.e., row in sample data). The researcher
knows that the scores for each rater are related because they
are assigning scores to the same group of participants. The
researcher uses Kendall’s coefficient of concordance because
it provides a process to determine how much agreement exists
among the raters.
2. An instructor conducts satisfaction surveys for three different
classes that she is teaching. She wishes to know if the satisfaction
scores for her three classes are similar or different.
For Research Question #17 (see Box 4.7), the scale variables are Levels of
Support in waves 1, 2, 3, and 4 (i.e., Support_w1, Support_w2, Support_
w3, and Support_w4), which represent the values recorded for social
support for the four waves of data collections.
SPSS Process
□ Select => Analyze => Nonparametric Tests => Related

Samples.


□ Click to select the first variable, Support_w1, found in the
Fields: area.
Fields: area.
□ Click to select the second variable, Support_w2, found in the
Fields: area.
Fields: area.
□ Click to select the third variable, Support_w3, found in the
Fields: area.
Fields: area.
□ Click to select the fourth variable, Support_w4, found in the
Fields: area.
Fields: area.
to the opposite area.
type of nonparametric test to conduct.

□ Select => Choose Tests from the Select an items: area.

□ Because Research Question #17 involves scale data (i.e.,
continuous variables), click to select the checkbox next to the

Kendall’s coefficient of concordance procedure.
Unlike the step in Research Question #12 in which you identified

values related to success, Kendall’s coefficient does not use a success
distinction. Therefore, no buttons to select options are available.

However, you can change the Multiple comparisons option. The
default is All pairwise, shown in the white area for the dropdown
list under Multiple comparisons: for the test. The other options for
Multiple comparisons: are Stepwise step-down or None. When the
number of groups is too large, requesting the comparison of all pos-
sible pairwise combinations can be cumbersome. The Stepwise step-
down option is helpful in these situations. This option sorts the groups
by their calculated group median value and compares only adjacent
groups, making the comparison of groups possible. However, with
only three groups, the comparison of all pairwise combinations is

the analysis.
SPSS Output
window, including:

output window

the analysis.
To change the detailed information displayed in the Model Viewer
window, you change the selected values in the dropdown lists at the bot-
tom of the window.
□ Click the dropdown list for the View: dropdown list at the
bottom right of the Model Viewer window and select Pairwise
Comparisons to display the information shown in Figure 4.7.
Findings
The Adj. Sig. values shown in Figure 4.7 indicate that one of the six pair-
wise comparisons is significant —Support_w1 vs. Support_w4 (p =.031).
Pairwise Comparisons
Support_w2 Support_w1
2.58 1.86
Support_w3
Support_w4
2.68
2.88
Each node shows the sample average rank.

Test Std. Std. Test
Sample1-Sample2 Statistic Error Statistic Sig. Adj.Sig.
Support_w1_Support_w2 –.720 .365 –1.972 .049 .292
Support_w1_Support_w3 –.820 .365 –2.246 .025 .148
Support_w1_Support_w4 –1.020 .365 –2.793 .005 .031
Support_w2_Support_w3 –.100 .365 –.274 .784 1.000
Each row tests the null hypothesis that the Sample 1 and Sample 2 distributions are the same.
Figure 4.7. Pairwise Comparisons for Kendall’s coefficient of concordance.

The other five pairwise comparisons have Adj. Sig. values larger than the
Significance level: set in Test Options. Hovering the mouse over the lines
in the Pairwise Comparisons graph shown in the Model Viewer window
will cause the related Adj. Sig. values to display. Therefore, from the Adj.
Sig. values, you can conclude that the level of support reported by the ado-
lescents is significantly different from wave 1 to wave 4. If you review the
descriptive information for the four variables, you will see that reported
support levels are larger in wave 4 than in wave 1, indicating that the dif-
ference is an increase in reported support. In addition, you can see this
increase in the Pairwise Comparisons graph under each variable name.
In statistical terms, Kendall’s coefficient of concordance (p < .016)
at α = .05 rejected the null hypothesis of similar observed trends for the
four samples. This analysis says that if the null hypothesis is true, there
is less than a 1.6% chance that all four trends are similar. The test of
equal responses was conducted for all pairwise comparisons. Results of
the statistical analysis identify only one pair that is significantly differ-
ent (Support_w1 and Support_w4), with adjusted p = .031.
Practically, this means that reported support levels are changing
over time for the adolescents, and that a significant increase in the sup-
port levels exists for the adolescents in the sample between wave 1 and
wave 4. Understanding why the support levels increase over time will
require additional data collection and analyses.

Now that Dr. Fleming has learned from his previous analysis that support
levels are not remaining the same over time for the adolescents in the study,
he wonders if, at each wave, the reported levels are similarly distributed. You
recognize that, to investigate this, you can utilize another nonparametric
procedure to examine social support in the same four data collection waves,
this time comparing distributions rather than changes over time.
Are the reported support levels for each wave similarly distributed?
FRIEDMAN’S TWO-WAY ANOVA BY RANKS TEST

Friedman’s two-way ANOVA is a statistical procedure that tests if
related samples come from the same population. A ranking strategy,
along with a calculated mean rank for each sample, is used to draw a
conclusion about the null hypothesis that the sample distributions are
similar. Because the comparable related-samples ANOVA has a nor-
mality assumption that is difficult to meet and sometimes to verify
when samples are small, you can use the Friedman’s test in a wider
range of data situations. Both Kendall’s coefficient of concordance
and Friedman’s two-way ANOVA test examine sample similarities.
However, Friedman’s test is calculating the possibility that the samples
all come from one population, and Kendall’s coefficient is looking only
for similarities among the samples, a slight but recognizable difference
when selecting which procedures to use when analyzing your data.
The variables needed for Friedman’s two-way ANOVA test are three
or more continuous/scale variables representing three or more related
responses (e.g., pre-observation, second observation, third observa-
tion, fourth observation for respondents).
Other examples of uses for the Friedman’s two-way ANOVA test are
as follows:
1. A college professor teaches a statistics course, and he wishes to

determine if his students’ test scores on three different quizzes
are comparable.
2. A researcher preparing a large-scale research proposal wonders
if the time of day will have an influence on how participants
report their depressive symptoms. She measures depressive
symptoms in the participants in her feasibility study in
the early morning, late morning, early afternoon, and late
afternoon.
For Research Question #18 (see Box 4.8), the scale variables are reported
Support Levels at waves 1, 2, 3, and 4 (i.e., Support_w1, Support_w2,
Support_w3, and Support_w4).
SPSS Process

If you just completed the analysis for Research Question #17, the vari-
ables may still be in the Test Fields: area. If they are, skip down to the
Note: below. If they are not, SPSS needs to know which variables to use
in the analysis.

□ Click to select the first variable, Support_w1, found in the
Fields: area.
Fields: area.
□ Click to select the second variable, Support_w2, found in the
Fields: area.
Fields: area.
□ Click to select the third variable, Support_w3, found in the
Fields: area.
Fields: area.
□ Click to select the fourth variable, Support_w4, found in the
Fields: area.
Fields: area.
to opposite area.
type of nonparametric test to conduct.

□ Select => Choose Tests from the Select an items: area.
Because Research Question #18 involves scale data (i.e., continuous

variables),
□ Click the checkbox next to the Friedman’s 2-way ANOVA by

ranks procedure.
Unlike the step in Research Question #12 in which you identified val-
ues related to success, Friedman’s test does not use a success distinction.
Therefore, no buttons to select options are available. However, you can
change the Multiple comparisons option. The default is All pair-
wise, shown in the white area for the dropdown list under Multiple
comparisons: for the test. The other options for Multiple compari-
sons: are Stepwise step-down or None. When the number of groups
is too large, requesting the comparison of all possible pairwise com-
binations can be cumbersome. The Stepwise step-down option is
helpful in these situations. This option sorts the groups by their
calculated group median value and compares only adjacent groups,
making the comparison of groups possible. However, with only four
groups, the comparison of all pairwise combinations is completely
manageable.

the analysis.
SPSS Output
window, including:

output window

at the Hypothesis Test Summary information and double-click).
Within the Model Viewer window, on the right side, you will see the
distributions for each sample and their individual Mean Rank score.
Also on the right side, you will see the Total N count for the number
of responders, the Test Statistic, the Degrees of Freedom, and the
Asymptotic Sig. (2-sided test) value. You can use this information in
the test to decide if the distributions are dissimilar enough to reject the
possibility that they come from the same population.
Findings
The Asymptotic Sig. value (.016) tells us to reject the possibility that they
come from the same population. A review of the horizontal histograms
in the Model Viewer window clearly reveals different distributions for
the four variables because the shapes of the histograms are different.
However, there is more analysis information to review than just the sig-
nificance test. To see the descriptive information and the distribution
for the individual waves, change the selected option in the dropdown
list for View: at the bottom of the window.
□ Click the View: dropdown list at the bottom of the Model

Viewer window.
□ Select Continuous Field Information to display the overall
distribution for the samples.
□ Click the dropdown list of variables Field(s): to change the
selected variable and the display from one variable to another.
To see the results of the pairwise comparisons, change the selected

option in the dropdown lists for View: at the bottom of the window.
□ Click the View: dropdown list at the bottom right of the Model
Viewer window.
□ Select Pairwise Comparisons to display the pairwise
information.
In statistical terms, Friedman’s Two-way ANOVA test (p < .016) rejected

the null hypothesis at α = .05 of similar distributions. The test of equal
distributions was conducted for all pairwise comparisons. Results of
the statistical analysis identify only one pair that is significantly differ-
ent (Support_w1 and Support_w4), with adjusted p = .031.
Practically, this means not only that the support levels from one
wave to another increased over time, but also the reported levels clus-
tered differently. More adolescents reported lower levels of support in
wave 1 and more reported higher levels in wave 4.
5
Predicting with Multiple
Independent Variables
CHAPTER OBJECTIVES
This chapter presents nonparametric models that allow for prediction of
a dependent variable based on values for a set of independent variables.
The chapter covers the following:
• Nonparametric regression (LOWESS)

• Nonlinear regression
• Factorial design regression
INTRODUCTION
Chapter 5 covers prediction using both dependent and independent
variables in a nonparametric model. Similar to parametric models like
linear regression, nonparametric models are available that allow for pre-
diction of a dependent variable based on values for a set of independent
variables. As in the earlier chapters, the reason for selecting a nonpa-
rametric procedure is because the data do not meet the assumptions
149
of the traditional parametric procedures. For linear regression models,

the assumptions involve linearity, independence, homoscedasticity, and
normality.
You may recall that, in your beginning statistics course, the
instructor showed you two scatter plots on an X-Y axis that included
an estimated regression line. One of the scatter plots showed dots fol-
lowing relatively close to a straight line, while the other plot showed
dots that followed a curved line. The curved line was an example of
a nonlinear relationship between two variables (X and Y). If you are
in this kind of situation with your own data, you can consider trans-
forming the X variable to see if linearity is possible (see Chapter 6
in Chatterjee & Hadi, 2012, for information on transformations).
Alternatively, you can consider using nonparametric regression that
does not have the linearity assumption (i.e., the assumption that the
relationship among the variables can be represented by a linear com-
bination of parameters).
The independence assumption, when violated, means that the model
is more than likely not fitting the data or that the p-values are smaller.
In other words, the errors in the model have a pattern or trend (e.g.,
serial correlation). One way to test the independence assumption (i.e.,
check for serial correlation) is to conduct the Durbin-Watson test, which
is a test available in the SPSS list of statistics options on the Linear
Regression: Statistics window. A YouTube video that offers additional
explanation can be found at https://w ww.youtube.com/watch?v=Lyk_
S9T-oCw (Grande, 2015). Homoscedasticity is another assumption for
linear regression, and it requires no relationship between the residu-
als and predicted values. To check for homoscedasticity, you can plot
standardized residuals against standardized predicted values to look for
patterns. SPSS provides options for requesting different combinations
of prediction and residuals plots to test for linearity, homoscedasticity,
and normality.
Normality is another necessary assumption when performing para-
metric regression. Violations of normality can occur due to the presence
of a few extreme outliers that cause the distribution of the error terms
to be skewed, not representing a normal distribution. Variables in the
model that do not represent normal distributions also have the potential
to cause a violation of normality. Thus, it is important to examine the
distribution of residuals for your regression model.
Predicting with Multiple Independent Variables 151
You must explore all potential violations of assumptions before

running a parametric test. This is not to say that nonparametric regres-
sions are exempt from variable examination. On the contrary, you must
closely examine every variable so you can make an informed selection
of the appropriate statistical test. When you are faced with the need to
conduct a regression and the parametric assumptions cannot be met by
transformation, by dropping cases with inappropriate outliers, etc., the
following nonparametric regression procedures are your next option.
Note: Over time, many people have asked IBM-SPSS to add a nonpara-
metric regression test. Unfortunately, at the time we wrote this chapter,
SPSS did not offer a menu-driven option for conducting a nonparamet-
ric regression analysis. Therefore, we use a second software package to
conduct some of the statistical tests in Chapter 5. The software pack-
age, XLSTAT, and is an add-in to Excel (see www.xlstat.com for more
information on XLSTAT).
Another note: R 2 is the amount of variance explained in your regression
model. For social science research, in general, R 2 values are smaller than
in other disciplines due to the complexity of topics/issues under inves-
tigation. For example, consider the complexity and number of different
issues that may be involved in understanding burnout or self-efficacy. R 2
values in the range of 20% to 30% are common, while values above 60%
are extremely rare. However, if find that your analysis does not explain
much of the variation, then there is something your model does not cap-
ture. Departures from what you hypothesize are always a good indica-
tion to verify the statistical assumptions.
Chapter 5 describes situations in which a researcher wishes to pre-
dict a value or level based on a set of independent variables, but the data
do not meet the assumptions for least-squares regressions. Following
the same format used in the previous chapters, a representative scenario
related to prediction is presented, along with research questions that can
be answered by using nonparametric methods for regression.
Box 5.1 Scenario for Chapter 5

Dr. Northcott’s area of study is adoption and foster care, and she also provides
training for practitioners working in this field. Recently, in response to comments
she received in several of her training sessions, Dr. Northcott decided to conduct
a survey about the personal and occupational stressors that may influence
the quality of these social workers’ practice. After assembling her questions,
receiving comments from expert content and measurement reviewers, and
revising her items, she was ready to pilot test her questionnaire. She sent a
questionnaire to every kth person on the membership list of a professional
organization for adoption and foster care workers, following up with reminder
postcards and duplicate questionnaires at regular intervals, as recommended in
the survey research literature (e.g., Dillman, 2014). She mailed questionnaires
to 50 professionals and achieved a 70% response rate (N = 35). However, eight
professionals did not answer all the questions, reducing her usable sample to 27.
Dr. Northcott included many variables for the study, knowing from the literature
that both personal and occupational issues might be relevant. A sample size of
27 does not provide sufficient power to conduct the planned regression analyses,
nor does it meet the guidelines that states 10 to 15 cases are needed for each
independent variable in the regression model. She hires you to assist her in
answering her questions.
Chapter 5 contains three research questions that guide the presen-

tation of the nonparametric procedures related to the above scenario.
Each research question provides the details for using the nonpara-
metric procedures. Table 5.1 lists the nonparametric procedures pre-
sented in Chapter 5, the kinds of variables for each procedure, and
available online.
Table 5.1. Summary of Analysis and Variables in Chapter 5

Question Test in DB
19 Nonparametric IV —categorical or Impairment,

regression continuous Stressors,
DV — continuous Burnout, Pred_
Stressors, and
Pred_Burnout
20 Nonlinear IV & DV — Edu_Level and
regression categorical Depression
(ordinal) or
continuous
21 Factorial design IV — categorical LifeSatisfaction,
DV — continuous Gender, and
Employment

Dr. Northcott’s training participants reported that the reason the quality of
their work was suffering was because they were dealing with too many stressors
at work and that they were experiencing burnout. Dr. Northcott created a
professional impairment instrument that encompassed several measures
of impaired practice (e.g., missing appointments, missing days from work,
employer disciplinary action). She asks you to find a regression model that
could predict level of professional impairment from standardized measures of
burnout and work stressors.
Can levels of burnout and work stressors predict levels of professional
impairment?
ROBUST LOWESS REGRESSION

Robust LOWESS regression is a nonparametric regression modeling
process available in XLSTAT that you can use when your data do not
meet the assumptions for linearity, independence, homoscedasticity,
and normality for the comparable parametric modeling process, mul-
tiple regression. Robust LOWESS regression, like all nonparametric
regression models, does not involve any assumptions about normality.
However, unlike multiple regression, which can examine individual
variable effects on a dependent variable and make predictions based
on a set of independent variables, the robust LOWESS can be used for
prediction only. LOWESS stands for LOcally WEighted regression and
Smoothing Scatter plots (Cleveland, 1979; Cleveland & Devlin, 1988).
Robust LOWESS is an updated version. Simply stated, the process uses
a proportion of the available data to begin estimating prediction values
for the dependent variable, and through many computations identifies
a prediction model. Calculations for smoothing are related to soften-
ing any sharp changes in the prediction model due to gaps between
data points that naturally occur in scale variables, thereby creating a
smoother prediction line that can be viewed in graphical form.
Other examples of nonparametric regression are as follows:
1. A researcher wishes to know if measures of self-efficacy and gender

predict levels of life satisfaction in his small clinical sample.
2. The researcher above also wants to know if demographics

are related to depressive symptoms in his clinical sample.
Specifically, he wants to know if current employment, age, and
socioeconomic status predict any part of depressive symptoms.
For Research Question #19 (see Box 5.2), three variables are used in
the analysis —two continuous independent variables, Burnout and
Stressors, and one continuous dependent variable, Impairment.
Before beginning a nonparametric procedure, you want to check
if the data do, indeed, violate parametric assumptions, which would
prevent you from using the customary linear regression. This is espe-
cially true when choosing between linear regression and nonparametric
regression. The major objective when building a nonparametric regres-
sion statistical model is to find the model that best fits your data. Thus,
unlike linear regression, nonparametric regression does not provide
beta values that you can use to interpret variables’ contribution to the
model. The reason lies in the technique used to estimate the requested
model for prediction. Therefore, viewing the results from a nonpara-
metric regression analysis is very different, and the process of identify-
ing the best model with the available variables is a bit more repetitive,
as it requires manually moving variables in and out of the model while
monitoring R 2 and prediction charts.
The first step is to examine the variables individually to fully under-
stand the quality and type of information they contain. Begin by run-
ning frequency distributions on each so you can review, among other
things, levels of skewness and kurtosis. For Research Question #19, the
distributions of both Burnout and Stressors are left-skewed and the dis-
tribution of Impairment is somewhat bimodal (see Figure 5.1).
SPSS Process
To obtain frequency distributions,
□ Select => Analyze => Descriptive Statistics => Frequencies….

□ Click to select Impairment in the list of variables on the left side
of the window.
Variable(s): area.
□ Click to select Burnout in the list of variables on the left side of

the window.
Variable(s): area.
□ Click to select Stressors in the list of variables on the left side of
the window.
Variable(s): area.
to opposite area.
Now that SPSS knows what variables to use in the descriptive
analysis,

□ Click to select the Histograms: option.
□ Click to select the checkbox next to Show normal curve on
histogram.
□ Select => OK to run the analysis, which will open the Model
Viewer window.
See Appendix A for SPSS syntax.
SPSS Output
The histograms in Figure 5.1 are displayed in the Model Viewer
window.
If these variables were normally distributed, you would continue
assessing the assumptions. However, given the variables’ departure
from normality and the small data set (N = 27), using parametric
regression is not an option. In addition, one of the cases has a miss-
ing value (ID 11), and XLSTAT requires you either to drop the case
from the analysis or to select a missing value replacement option
(e.g., mean, nearest neighbor replacement). Because of the repetitive
nature of conducting nonparametric regression, it is probably best to
Impairment
10 Mean = 2.93
Std. Dev. = 1.269
N = 27
8
6
Frequency
0
0 1 2 3 4 5 6
Impairment
Burnout
6 Mean = 16.15
Std. Dev. = 9.134
5 N = 27
4
Frequency
0
0 10 20 30 40
Burnout
Stressors
6 Mean = 7.04
Std. Dev. = 3.117
5 N = 26
4
Frequency
0
2 4 6 8 10 12 14
Stressors
Figure 5.1. Histograms of continuous variables.

examine model fit first by dropping case(s) with missingness. Next,

compare the model fit when using the different replacement options.
If the model fit is appropriate, the best option is to choose the model
that dropped case(s), as this model will avoid adding bias into your
prediction values. See Appendix B for more information on handling
missing values.
ANALYZING THE DATA IN XLSTAT

Prior to running an analysis, you must arrange the data to meet the
requirements of the software you are using. XLSTAT expects that you
have arranged the variables in columns, and that each row represents
a unique case or respondent. Therefore, for this example, you should
arrange values for Burnout, Stressors, and Impairment in three separate
columns. SPSS uses the same layout for the data as XLSTAT, with one
important difference —how to identify missing values. SPSS provides a
method in which to name multiple values as identifiers for missingness,
but XLSTAT identifies missing values only wherever a blank or empty
cell exists.
After installing XLSTAT on your PC, you need to make XLSTAT
active. To activate XLSTAT within Excel,
□ Select => ADD-INS tab at the top of the Excel window.

□ Click to select the ToolBar Commands button to open the
XLSTAT toolbar.
After you click on the button, Excel will take a few minutes to bring
up the toolbar and add a new tab named XLSTAT to the list of tabs at
the top of the Excel window.
XLSTAT Process
The nonparametric regression analysis to run is found under the
Modeling Data button .
□ Select => XLSTAT tab in the Excel menu options at the top of
the Excel window.
□ Select => the Modeling Data button to display the list of

analyses associated with modeling data.
□ From the list, select => Nonparametric regression, which opens
the Nonparametric regression window.
At the top of the Nonparametric regression window are seven tabs (General,
Options, Validation, Prediction, Missing data, Outputs, and Charts).
The information and selections you provide within each of these tabs tell
XLSTAT how to conduct the nonparametric regression (see Figure 5.2).
For robust LOWESS, XLSTAT dims some of the options to indicate that
this specific nonparametric regression model does not use them.
□ Select => General tab.

□ Be sure to check the option Variable labels on the right side of
the window if variable names appear in the first row of the Excel
spreadsheet above the data.
□ Click on the white space for the field name Y/Dependent
variables:, right below the word Quantitative:.
By clicking in the white space, you draw focus to that area, which allows
for the identification of the dependent variables. The dependent variable
in the spreadsheet for Impairment is in column B, so find that column
on the spreadsheet and
□ Click on the B at the top of the column.
Figure 5.2. Nonparametric regression window.

When you click on the B, you tell XLSTAT where to find the values
for the dependent variable, Impairment. Your click will both high-
light the column in the traditional way Excel selects a column and add
text to the Y/Dependent variables: field (i.e., Data!$B:$B). Data!$B:$B
tells XLSTAT to use all the data in that column for your dependent
variable.
□ Click in the white space for the X/Explanatory variables: field

just below Quantitative:, drawing focus to that area.
You can find the independent or explanatory variables in our exam-

ple in columns C and D.
□ While holding down the “Ctrl” key on the keyboard, click and
drag the mouse from C to D at the top of the two columns.
Doing this tells XLSTAT which data to use for the independent vari-
ables and will place text in the field to identify the two columns (i.e.,
Data!$C:$D).
Note: Using the mouse to select the data for the analysis is only one
method of telling XLSTAT what to use. If you want to avoid using the
mouse, you can type the information directly into the fields that repre-
sent the dependent and independent data.
Because this research question does not include categorical data, the
Qualitative checkbox should be unchecked. Next,
□ Click to select the Sheet option to tell XLSTAT to place the

analysis results in its own Excel sheet.
□ Click to select the dropdown list for Method: to select Robust
LOWESS from the dropdown list.
Robust LOWESS uses a form of curve smoothing in a scatter plot to

identify the relationship between the dependent and independent
variables.
The other options within this dropdown are for other nonparamet-
ric regression model procedures. Selection of a model option is based on
the kind of data you are modeling and can become quite complicated.
For Research Question #19, and for many data situations, the robust
LOWESS will meet the prediction requirements.
Note: For more information on the above options available, including
more information on model procedures, select => Help button at the
lower right side of the Nonparametric regression window. Selecting
Help will open a separate window named HTML Help that provides
access to help information for XLSTAT and available procedures.
To proceed with the analysis,
□ The default for the Polynomial degree: value is 1 (one). If it is

not 1, click to change the Polynomial degree: value back to 1.
The reason for this is that the first time you run your model, you start with
a linear polynomial (i.e., value of 1). Starting with 1 in this example tells
XLSTAT not to investigate the squared values of Burnout or Stressors for
predicting Impairment. Once you review the fit indices, you can return to
this option and try higher levels of polynomials to see if model fit can be
improved, as long as increasing the polynomial makes theoretical sense.
For example, does adding the squared values of Stressor (polynomial = 2)
improve model fit? However, keep in mind that variables with influen-
tial outliers can suggest a curvilinear relationship, especially with small
samples. This is another example of how important it is to understand the
data thoroughly before building regression models or testing hypotheses.
As you explore all the options provided by XLSTAT for nonpara-
metric regression, you will recognize many different analytic possi-
bilities. However, for this analysis most of the default options fit with
Research Question #19.
□ Select => Options tab to see where in the Nonparametric

regression window you can change the proportion of the data
used to estimate prediction values (i.e., proportion of the sample
for “learning”).
“Learning” is how XLSTAT describes the process of building a model

that fits the data. XLSTAT will use some of the data to build the model
and other data to predict model fit. The default is 50% and is a good
place to start, but if you are experiencing poor model fit, you may adjust
this option up or down to see if you can improve the fit.
Another option is Tolerance:, which has as a default to check and set

to .01. This option tells XLSTAT how much a variable can be out of con-
formity with the estimated model before it is ignored altogether. If you
want the variables to be more in line with the model, you can decrease
the default of .01. In other words, if a variable is not contributing enough
to the model, XLSTAT will decide to move on without it. For Research
Question #19, we accept the default.
The final option is the Interactions/Level:, which is not checked as a
default. If you want to include variable interactions in your model, you
can check this box. For Research Question #19, we did not include vari-
able interactions in the model.
Note: Dimmed options on the window indicate that they are not used by
the robust LOWESS procedure.
□ Select => Validation tab if a subset of your data set is to be used

to validate the model.
With larger data sets, validation is an option to further validate the fit of
a model. This is like splitting a data set so you can conduct an explora-
tory factor analysis (EFA) on one half and then validate the model with a
confirmatory factor analysis (CFA) using the other half. Unfortunately,
with smaller data sets, taking some of the records to validate the esti-
mated model is impractical. This is the case for Research Question #19;
therefore, due to the limited number of records, you should leave the
Validation option unchecked.
□ Select => Missing data tab to verify that cases with missing data
will be dropped.
Dropping the cases with missing data is the safest way to avoid adding
unrecognized bias. If you replace the missing values, you introduce the
possibility of unnaturally increasing the power of the model to predict
the independent variable, and it may not be a true reflection of the con-
cepts you are analyzing.
□ Select => Outputs tab to verify that all four options are
selected.
For Research Question #19, we want to request all the information avail-
able to help understand the data and the model.
□ Select => Charts tab to request that the analysis output includes
charts.
Be sure that the Data and predictions option is checked and that
the As a function of X1 option is selected. In addition, be sure to
select the Residuals option to add residual information to the anal-
ysis output.
Note: It is important that you set these options as described above for
both the Outputs and Charts tabs, because these options provide a clear
graphical view into the quality of the data in the model as well as model
fit (e.g., predicted vs. observed).
Now that the options are set,
□ Select => OK button at the bottom of the window to execute the

analysis.
Once OK is selected, a final verification window (i.e., XLSTAT —

Selections) appears. This gives you one last look at the data and options
selected for the analysis. After reviewing your choices carefully,
□ Select => Continue button and a new Excel sheet will appear
with all the requested analysis output.
Note: If some analysis information is missing, or if XLSTAT recognizes

a conflict in your analysis request, an XLSTAT — Message window
will appear with an explanation of the issue instead of the XLSTAT —
Selections window.
XLSTAT OUTPUT —VERIFYING MODEL FIT

Remember that the major objective when building a nonparametric
regression statistical model is to find the model that best fits your data.
The search for the best-fitting model involves trying different model set-
tings (i.e., options) while monitoring the model fit information. Before
using any results from a statistical model, you must assess model fit to
make sure your model represents your data. If you are familiar with
factor analysis or latent class analysis (LCA), this exploratory approach
will sound familiar.
The first step in an examination of nonparametric regression model
fit in XLSTAT is to consider the amount of variance for the depend-
ent variables explained by the independent variables, identified as the
determination coefficient R 2. R 2 is part of the Goodness of fit statistics:
information found on the new Excel tab at the bottom of the Excel win-
dow (see Figure 5.3). R 2 is a calculated value that falls between 0 and
1. Zero represents the worst possible fit, in which the data do not follow
the model at all, and 1 represents a model that supports every data point
exactly.
Figure 5.3. Fit indices.

R 2 for Impairment in Research Question #19 is 0.618, which indi-

cates that 62% of the observed variability in Impairment is explained
by Burnout and Stressors. Figure 5.3 shows that the robust LOWESS
results provide three other fit statistics —SSE (sum of squares of the
errors), MSE (means of the squares of the errors), and RMSE (root mean
squares of the errors). All four of the fit indices relate in some way to
one other. For example, RMSE is the square root of MSE, but because
of their interpretation, the two most often reported are R 2 and RMSE.
RMSE is the standard deviation of the value differences between the
data and the model, thereby providing information on the variations of
differences between the two. Because the best-fitting models are those
with the smallest errors and that explain the most observed variability,
we can use R 2 and RMSE values to identify which model has the best fit.
Note: Because R 2 will, in almost all cases, increase as you add more
variables to the model, whether or not the added variables are signif-
icant predictors of the dependent variable, R 2 alone is insufficient in
determining fit. In addition to following RMSE, a close review of vari-
able information is important. Following the variables individually, as
discussed in Chapter 1, is just as important when building prediction
models.
So just how much variation (i.e., R 2) needs to be explained before the
model can be used? Unfortunately, the response to this question is that
it depends on the research question and the specific analysis. For non-
linear regression in which you have the flexibility to try different models
(i.e., functions), it comes down to the variables in the study, to how much
variation other similar studies have identified, and what would be con-
sidered enough based on theoretical understanding of the issues being
measured. No matter how you make the decision, the responsibility falls
on you, the researcher, to justify that the variance amount explained is
enough before sharing your findings with other researchers.
Another important responsibility is to make sure you have elimi-
nated other model options and to verify that you have found the appro-
priate model for your data. This is an iterative process, so, as mentioned
earlier, this includes trying a few other XLSTAT options prior to run-
ning the analysis. Another way to eliminate other models is to try other
combinations of independent variables that have theoretical support
for predicting the dependent variable —in this case, Impairment. In
addition, exploring other possible models will help you understand
your data at a deeper level, give you an opportunity to build confidence

that you have selected the right variables for your final model, and help
you draw a conclusion that you feel comfortable sharing with other
researchers.
In addition to presenting fit indices, XLSTAT for LOWESS regression
provides graphs to examine model fit (see Figures 5.4, 5.5, and 5.6). The
first graph, Figure 5.4, is a scatter plot of the model prediction for the rela-
tionship between Burnout and Impairment as a function of X1. Burnout
was used because it was the first listed independent variable (i.e., X1).
If you want to see a graph for Stressors and Impairment, you must
switch the order of the independent variables and rerun the analysis.
The second graph, Figure 5.5, is a scatter plot of predicted vs. observed
values for Impairment. Because LOWESS assumes that the errors for a
model are random, an appropriate fitting model should show random
error values on the predicted vs. observed plot. Nonrandom patterns
along the diagonal line may suggest poor model fit to the data. The third
graph, Figure 5.6, is the residuals bar chart showing the amount of error/
residual for each case. Residuals are the differences between observed
Nonparametric regression (Impairment)

6
4
Impairment
0
0 5 10 15 20 25 30 35 40
Burnout
Impairment Pred(Impairment)
Figure 5.4. Fit assessment graph as a function of X1.

Pred(Impairment)/Impairment
6
4
Impairment
0
0 1 2 3 4 5 6
Pred(Impairment)
Figure 5.5. Fit assessment graph for predicted vs. observed.
and model prediction values. Like the scatter plot, histograms that show
a pattern as you review the bars from left to right suggest the possibility
of a poor-fitting model.
In addition, the bar chart provides the opportunity to identify cases
that have unusually large error (i.e., a bar on the chart that is much
larger than other bars). In these cases, you need to verify that you do not
a have a data-entry problem and that the case does follow all the original
guidelines for inclusion in the target population. In extreme situations,
you may remove the case if the case clearly does not belong theoretically
or conceptually, but we do not recommend this practice unless you have
irrefutable information to support the fact that the respondent is outside
of the parameters of the study.
Fortunately, for Research Question #19, the scatter plot in Figure 5.4
does not show a pattern along an envisioned diagonal line. In addition,
the bars in Figure 5.6 appear relatively random and do not indicate a
pattern or problem with cases’ being dependent. Therefore, you can be
Residuals
1.5
0.5
0
Residual
Obs1
Obs2
Obs3
Obs8
Obs9
Obs13
Obs14
Obs19
Obs20
Obs21
Obs22
Obs23
Obs26
Obs27
Obs10
Obs7
–0.5
Obs5
Obs17
Obs4
Obs15
Obs16
Obs18
–1
Obs6
Obs24
Obs12
–1.5
Obs25
–2
Observations
Figure 5.6. Fit assessment graph using residuals.
comfortable using the selected model to predict levels of Impairment

with values of Burnout and Stressors.
Once you are comfortable that you have identified the appropri-
ate model for the data, the next step is to add the predictor values for
Burnout and Stressors into the nonparametric regression model. To add
these values,
□ Select => Modeling data => Nonparametric regression to open

the Nonparametric regression window.
Verify that the information for each tab is correct in that it matches
what you intended for the final model, and that options are set the same
as when you found the best-fitting model.
□ Select => Prediction tab and click to select the checkbox next to
the Prediction option.
□ Click on the white space right below Quantitative: to move

focus to the Quantitative: area.
By clicking in the white space, you draw focus to the area that allows
the identification of the predictor values. The predictor values represent
values for Burnout and Stressors variables to predict impairment. In the
data set, the values are stored in two columns named Pred_Burnout and
Pred_ Stressors.
□ Find these columns on the spreadsheet. Next, left-click and hold

while dragging the mouse from the first value for Pred_Burnout
to the last value for Pred_Stressors.
Doing this will place text in the field to identify the data to use for the
prediction (i.e., Data!$J$2:$K$3). Unlike the process used to select a col-
umn for the independent and dependent variables, the selection of the
predictor values cannot include the column headings.
The use of the mouse to select the data for the predictors is only one
method of telling XLSTAT what to use. If you want to avoid using the
mouse, you can type the information directly into the field.
Important: The order from left to right of the independent variables
in the spreadsheet must match the order of the predictor values. For
example, the Burnout variable column in the spreadsheet is to the left
of the Stressors column, which means that the values for Burnout and
Stressors in a prediction model must be Burnout values on the left and
Stressors values on the right.
Now that XLSTAT knows what information to use for the prediction,
□ Select => OK button at the bottom of the window.
The XLSTAT —Selections window appears, giving you one last chance
to see what data you have selected for the prediction.
□ After reviewing your selections carefully, select => Continue

button at the bottom of the window to conduct the analysis and
open a new Excel sheet with the results.
XLSTAT Output
A new Excel sheet appears that contains all the LOWESS results
requested and a scatter plot graph that shows the two prediction val-
ues along with the other data points. Figure 5.7 shows the two pre-
dicted Impairment values (5.038 and 4.466) for the two sets of values
for Burnout (27.000 and 15.000) and Stressors (5.000 and 15.000).
A review of the predicted values shown in the scatter plot reveals that
two are on the fringe of other predicted values, but not so far out that
they appear to be outliers. When predicted values are not within the
proximity of other data values, this suggests the predictor values may
be outside of the model’s prediction capability. When this happens,
you need to use caution in using these predictions for decision-making
purposes.
Nonparametric regression (Impairment)

6
4
Impairment
0
0 5 10 15 20 25 30 35 40
Burnout
Impairment Pred(Impairment) Predictions
Figure 5.7. Two predicted values.

Findings
In statistical terms, a nonparametric regression model using the robust
LOWESS estimation method achieved model fit with R 2 = .618 explain-
ing 62% of the variance in Impairment using two independent variables,
Burnout and Stressors. Two predictions for Impairment with values for
Burnout (27 and 15) and Stressors (5 and 15) were 5.038 and 4.466 and
were found to be within the model’s range for effective prediction.
Practically, this means that the new Impairment scale is showing
signs of reliability for use with two previously validated scales of Burnout
and Stressors. This means that you can estimate level of Impairment
with values for Burnout and Stressors.
Note: For comparison, we ignored the issue around meeting assump-
tions and conducted a parametric linear regression on the Research
Question #19 data. The results established an R 2 of 0.604, indicating that
the parametric model explained 60% of the variance in Impairment, in
comparison to the 62% we found for the nonparametric model. In addi-
tion, the RMSE for the linear regression was 0.809, while the RMSE for
the nonparametric regression as 0.747, meaning that there is a bit more
variation in error values for the linear regression model than the error
in the LOWESS model. The difference may be due to a variety of things,
including how close the data are to violating assumptions. However,
this finding does not mean that nonparametric models will always show
improved model fit when the data violate the assumptions. Rather, it just
means that you must consider many important factors when selecting
a statistical approach so that you are not presenting results that are not
supported statistically.

Dr. Northcott found evidence in the literature of a relationship between
depressive symptoms and education level in social service professionals, which
is important because rates of depressive symptoms are quite high among
this professional group and can be a barrier to good practice. Soon to teach
a master’s level social work course, Dr. Northcott is interested in exploring
the predictability of education level using scores on a standardized depressive
symptoms measure. Therefore, her dependent variable is an ordinal measure of
education and training (bachelor’s degree only, master’s degree with no license,
and master’s degree with clinical license) and depressive symptom scores are
her independent variable. Dr. Northcutt asks if you can find a regression model
that could verify this relationship in her pilot sample. If she finds a relationship,
she will use this information in her upcoming class.
Are depressive symptoms significantly related to an ordinal measure of
education and training?
NONLINEAR REGRESSION
Nonlinear regression is a method that models a nonlinear relationship
between a dependent variable and a single variable or multiple inde-
pendent variables. Unlike the frequently used multiple regression, the
relationship is not assumed to be linear —the relationship between the
dependent and the independent variables is not assumed to represent
a straight line. The dependent and independent variables can be either
categorical/ordinal or continuous, and it is assumed that the responses
captured by the dependent variable are independent and that the inde-
pendent variables are not overly correlated (i.e., multicollinearity). As in
nonparametric regression, the goal in conducting a nonlinear regres-
sion is to find a well-fitting model.
Other examples of uses for nonlinear regression are as follows:
1. A researcher conducting a feasibility study for a clinical trial

wants to estimate the number of participants she can recruit
related to the number of advertising spots she runs that are
aimed at recruiting viable participants meeting the inclusion
criteria for her trial.
2. An administrator is considering opening her agency one hour
later during the weekend. She wants to predict the number of
clients who will arrive in that 60-minute window based on data
collected on number of arrivals throughout the day for 365 days.
Note: Another frequently used procedure is the multinomial regres-

sion procedure, which requires a categorical dependent variable with
no implied order to the categories. If you have non-normal contin-
uous variables, highly correlated continuous variables (i.e., ρ > .8), or
implied order in your dependent categorical variable, then multinomial
regression is not an option. Both multinomial regression and nonlinear

regression are procedures available in SPSS.
So, what is the difference between nonparametric and nonlinear
regression? Because there are many variations of both, answering this
question is very complex. An oversimplification is that nonparamet-
ric regression builds a model based on some of the observed data, and
nonlinear regression uses numerical optimization and a predetermined
function (i.e., the formula that describes the probability of a relationship
between the dependent and independent variables) to best fit param-
eter estimates to the data. Identifying the predetermined function for
a specific data set can be challenging and may require more statistical
background or assistance from a statistician, but if a function can be
identified, your estimate errors more than likely will be smaller than
when using a nonparametric approach. However, not all dependent to
independent variable relationships will effectively fit, even a nonlin-
ear relationship, which leaves the nonparametric regression procedure
important to have as an option.
XLSTAT provides preprogrammed functions to help you begin
identifying possible functions. However, to conduct a nonlinear regres-
sion in SPSS, you will need additional knowledge on function devel-
opment that reaches beyond the scope of this book, therefore we use
XLSTAT for Research Question #20 (see Box 5.3).
For Research Question #20, the dependent variable, Edu_Level, is
ordinal (1 = bachelor’s degree, 2 = master’s degree with no license, and
3 = master’s degree with clinical license) and the independent variable,
Depression (representing depressive symptoms) is continuous.
Before performing the nonlinear analysis, you must verify that the
data support a nonlinear approach. The assumptions of constant error
variance and independence are the same as they are for linear regres-
sion. However, you now replace the linear relationship assumption
between the dependent and independent variable(s) with an assump-
tion that an underlying probability model (i.e., function) exists for the
observed data. Identifying the kind of underlying probability model is
the most important task in developing a nonlinear regression model —
a task demonstrated in Research Question #20.
Just as in Research Question #19, first you must carefully exam-
ine the variables individually to fully understand the quality and type
of information they contain. Because this example uses a categorical
variable, Edu_Level, and one continuous variable, Depression, you will

use box-and-whiskers and Q-Q plots to examine the data. A review of
the two sets of plots, along with other things, will help identify if the
distributions among the three Edu_Level groups (i.e., bachelor’s degree,
master’s degree with no license, and master’s degree with clinical license)
are possibly different from each other. If they are, you cannot use Edu_
Level as your dependent variable in nonlinear regression. The nonlinear
regression process requires that the distributions among the categories
for the independent variables be similar.
You can use XLSTAT to review your data. If the XLSTAT menu bar
is not visible,

□ Click on the ToolBar Commands button to open the
XLSTAT toolbar.
up the toolbar and add a new tab named XLSTAT to the list of tabs at
the top of the Excel window.
To request the plots,
□ Select => Visualizing Data => Univariate plots next to the

box-and-whiskers icon to open the Descriptive statistics
window.
After the Descriptive statistics window opens,

□ Click to select Quantitative data: (if not selected) and then click
in the white space representing the Quantitative data: area.
Now that the quantitative data field has focus,
□ Click on the variable column for the Depression values (i.e.,

click on the letter “E” at the top of the column). Once you click
on the “E,” Data!$E$E appears in the field.
□ Click to select Subsamples: (if not selected) and then click in the
white space below Subsamples: to direct focus to that area.

Once the focus is on the Subsamples: area,
□ Click on the column in Excel that contains the categorical

values named Edu_Level (i.e., click on the letter “F”).
Before leaving the General tab, verify that Variable-Category labels

and Sample labels are both selected. Next,
□ Select => Options tab.

□ Make sure Descriptive statistics, Charts, Compare to the total
sample, and Sort the categories alphabetically are selected.
□ Select => Chart(1) tab.
□ Make sure only Box plots and Normal Q-Q plots are selected.
The information and selections provided within each of these tabs tells
XLSTAT how to produce the two plots.
□ Select => OK button at the bottom of the window to produce

the plots.
Once OK is clicked, a final verification window appears (i.e., XLSTAT —

Selections) giving one last look at the data selected for creating the
plots.
Note: Only a few procedures in XLSTAT do not have ways to handle
missing values. Unfortunately, Visualizing Data happens to be one
of them. If you have missing data, you will need to sort and reselect
your data to exclude cases with missingness prior to asking for Box or
Normal Q-Q plots. After carefully reviewing the choices,
□ Select => Continue button.
A new Excel sheet will appear with all the requested descriptive output.
EXCEL Output
The Q-Q plots help identify when distributions among the groups are
too dissimilar to use in a nonlinear regression. Fortunately, the Q-Q
plots in Figure 5.8 show only minor differences among the three groups,
which is important because the distribution must be similar for the
three groups if you are going to use nonlinear regression with a cate-
gorical dependent variable. In addition, the Q-Q plots show a specific
Q-Q plot (Depression | Edu_Level-1)

5
Quantile - Normal (2.13, 1.36)
0
0 1 2 3 4 5
Depression | Edu_Level-1

16
14
12
10
2
2 4 6 8 10 12 14 16

25
20
15
10
0
0 5 10 15 20 25
Figure 5.8. Q-Q plots.
pattern along the diagonal line (i.e., first above and then below), which
does support the possibility of a nonlinear model.
The requested box-and-whiskers plots show the data distributions
for each group, which are the categories for the dependent variable. Like
the Q-Q plots, the differences among the box-and-whiskers plots are not
enough to indicate that we cannot continue.
As stated for other statistical procedures, another important step
before moving on is to make sure you have enough information for
each group, which in this situation translates into enough represen-
tation in all three groups (i.e., n = 8, 12, and 7 for bachelor’s degree,
master’s degree with no license, and master’s degree with clinical license,
respectively). The minimum guideline we recommend is to have at least
5 responses for each category in a statistical procedure, but remember
that you should use any parameter estimates with care due to the lack of
precision from using a very small sample.
Note: The Q-Q plots presented in Figure 5.8 are those that should be
produced prior to conducting a linear regression to verify that the nor-
mality assumption is not violated.
Now that you have examined the data closely and have veri-
fied similarity among the categories, you can move on to finding a
model that fits your data. So how many different models are possi-
ble? Unfortunately, the list is extremely long. One of the strengths of
nonlinear regression is that a researcher can explore many different
models with many different levels of complexity. Up to now, we have
avoided discussing formulas, and we will continue to leave formula
development to other discussions outside of this book. However, to
accomplish the nonlinear regression, you will need to use three dif-
ferent formulas we provide for this procedure. The three formulas
represent three possible probability function models that fit nonlin-
ear relationships. Many different probability functions are built into
XLSTAT for conducting nonlinear regression. In addition to the built-
in functions, XLSTAT offers a process that allows for the development
of other function models.
While answering Research Question #20, both the built-in and the
user-defined methods are described. You may ask, “With all this flexi-
bility, how do I know where to start?” Fortunately, the three function
options cover the use of two-, three-, and four-parameter models (i.e.,
probability models predicting a value for the dependent variable) that
will meet the needs of most people who are not statisticians but have
the occasion to use nonlinear regression. These three examples will get
you well on your way to conducting a nonlinear regression analysis. We
present the functions related to the examples later when discussing the
steps for the XLSTAT procedure.
Exploring different models is the main process for identifying an
effective nonlinear relationship (unlike nonparametric regression, in
which the data identify the relationship). The first model relates to one
that uses only two parameters to identify the relationship between a
dependent and an independent variable. If you took a basic statistics
class, a simple linear regression formula included an intercept and
a slope (i.e., β0 and β1, respectively). The intercept and slope are two
parameters that help explain the relationship between a dependent and
independent variable. The two-parameter nonlinear model attempts to
do the same thing, which is to find estimates for the two parameters
that appropriately explain the relationship. You can add parameters
when the two parameters are not enough to explain a more complex
situation. Each model has a unique way of using the parameters in a
formula and explaining the relationship. As you would expect, using
more parameters than are necessary adds complexity and possibly
unwanted estimate error to your prediction. Therefore, the goal is to
find an effective model by finding the formula that has the fewest num-
ber of parameters while still effectively representing the relationship
between the variables.
A nice feature of XLSTAT is that two of the three functions that
include two and three parameters are already in a list of available func-
tions; however, SPSS offers no list of functions to get you started. In
addition, XLSTAT provides a very friendly way to include user-defined
functions —which is demonstrated as part of answering Research
Question #20.
ANALYZING THE DATA IN XLSTAT AND EXCEL —TWO PARAMETERS

Prior to running the analysis, you must arrange the data to meet the
requirements of the software you are using. XLSTAT expects that you
arrange the variables in columns, and that each row represents a unique
case or respondent. Check that you arrange the values for Depression
and Edu_Level in two adjacent columns, respectively. If the XLSTAT

menu bar is not available,

□ Click on the ToolBar Commands button to open the
XLSTAT toolbar.
up the toolbar and add a new tab named XLSTAT to the list of tabs at the
top of the Excel window. You will find the nonlinear regression analysis
to run under the Modeling Data button.
□ Select => Modeling Data => Nonlinear regression, which

opens the Nonlinear regression window.
At the top of the Nonlinear regression window are seven tabs. The
information and selections provided within each of these tabs tell
XLSTAT how to conduct the nonlinear regression. For Research
Question #20,

□ Click on the white space below Quantitative: to move focus to
the Y/Dependent variables: area.
By clicking in the white space, you shift your focus to the area that
allows for the identification of the dependent variable. The dependent
variable in the spreadsheet for Edu_Level is in column F, so find that
column on the spreadsheet and
□ Click on the “F” at the top of the column.
Doing this will both highlight the column in the traditional way Excel
selects a column and add text to the Y/Dependent variables: area (i.e.,
Data!$F:$F). Data!$F:$F indicates that you want to use all the data in
that column for the dependent variable.
□ Make sure you check the Variable labels option on the right side
of the window.
□ Click in the white space below Quantitative: for the X/

Explanatory variables: area, drawing focus to that area.
You can find the independent or explanatory variable in the example

in column E.
□ Find the E column on the spreadsheet and click on the “E” at the
top of the column.
Doing this will place text in the field to identify the column (i.e.,
Data!$E:$E).
Note: The use of the mouse to select the data for the analysis is only one
method for telling XLSTAT what to use. If you want to avoid using the
mouse, you can type the information directly into the fields.
The next step is to select the probability model to use.
□ Click on the Functions tab to display the options for function

selection (see Figure 5.9 —Nonlinear regression window).
Remember that Function is the formula that describes the probability of

a relationship between the dependent and independent variables.
When the information for the Functions tab appears, you will
notice two sections. The section on the left (Built-in functions) lists the
Figure 5.9. Nonlinear regression window.

probability models already defined in XLSTAT. The function high-

lighted on the left in Figure 5.9 is the first one you should examine to
review for model fit.
□ Click to select Built-in functions.

□ Scroll down the Built-in functions list and click to select the
one that is defined as 1/(1 + Exp(–pr1–pr2*X1)).
This function represents a nonlinear regression model that uses only

two parameters (i.e., pr1 and pr2). Once the function has been selected,
you have chosen to examine the relationship between the variables
using a model with only two parameters.
Note: XLSTAT offers many functions that includes the one used first for
this example. If you are not comfortable with formulas, then this exam-
ple may be a bit unsettling. We recommend that, as you experience this
nonlinear procedure for the first time, you don’t spend time trying to
interpret the formulas right away. Simply use this opportunity to work
through the entire process for nonlinear regression and see how to per-
form a nonlinear regression analysis. Later, you can request assistance
from a statistician for a deeper understanding.
□ Select => Options tab and verify that none of options on the left
side is selected.
□ Select => Validation and Prediction tabs and verify that none
of the options under these two tabs is selected.

□ Select => Missing data tab and verify that Remove the
observations option is selected.

□ Select => Outputs tab and verify that all the options are selected.

analysis.
Once OK has been clicked, a final verification window appears (i.e.,
XLSTAT —Selections) giving one last view of the data selected
for the analysis.
□ After carefully reviewing the choices, select => Continue button
and a new Excel sheet will appear with all the requested analysis
output.
Excel Output —Two Parameters

At the completion of the analysis, a new Excel tab will appear and will
contain the goodness-of-fit statistics as part of the analysis output.
For the function representing two parameters, the R 2 (amount of
variation explained) is very small (0.065). A nonlinear regression model
that explains only 6% is not a useful model for a study. Therefore, you
need to select a different probability model or function in XLSTAT to
determine if one of the other models provides an adequate fit. The fol-
lowing information in this section of the chapter refers to three-and
four-parameter models, which are models that can add different curves
or shapes, along with other planes, to the slope and intercept param-
eters. Don’t worry. We provide the formulas for them.
ANALYZING THE DATA IN XLSTAT AND EXCEL —THREE PARAMETERS

To investigate a different model, you will need to reopen the Nonlinear
regression window.
□ Select => Modeling data => Nonlinear regression to reopen the

Nonlinear regression window.
□ Select => Functions tab.
Fortunately, XLSTAT also includes the second model. The second

function is in the list just below the first one you tried, and is the func-
tion pr3/(1 + Exp(-pr1-pr2*X1)). This function attempts to model the
data using three parameters (p1, p2, and p3).
□ Click to select the function pr3/(1 + Exp(-pr1-pr2*X1)) from

the list.
analysis.
Once OK has been clicked, a final verification window appears (i.e.,

XLSTAT — Selections) giving one last view of the data selected for the
analysis.
□ After carefully reviewing the choices, select => Continue button

and a new Excel sheet will appear with all the requested analysis
output.
Note: Excel will not overwrite the previous output appearing in sepa-
rate Excel sheets. Each run of an analysis will create a new Excel sheet.
Keeping track of the sheets by renaming them will save you time later
when you are writing up your findings.
Excel Output —Three Parameters

With the newly selected function that involves three parameters, the R 2
improves to 0.571 (i.e., 57% of the variance in Depression is explained
by Edu_Level). In social science settings, 57% is sufficient variance
explained to feel positive about the model in a study. However, with-
out further investigation of other possible models, you will not know
if a three-parameter model is the best option to explain the variable
relationship.
The three graphs in Figure 5.10 show some of the analysis output
from nonlinear regression. The new Excel sheet containing the analysis
output includes these graphs. You can see that
• The relationship between Depression and Edu_Level is

nonlinear —the line showing the relationship is curved.
• In the graph of Edu_Level predictions vs. Edu_Level, the model
does better at predicting bachelor’s degree (category 1) and
master’s degree with no license (category 2) than at predicting the
master’s with clinical license (category 3).
• The residuals graph does not show signs of a pattern in the
errors, though one case does have a larger error than the
other cases.
It is fair to say that the results from the nonlinear regression model
containing three parameters show that you have sufficient evidence to
conclude that the levels of depressive symptoms are different for people
with different education and training levels. You can draw this conclu-
sion primarily because you could find a model that showed appropriate
fit and prediction capabilities. If you were not able to find an effective
model, you could not draw this conclusion, and you would need to test
the probability model with four parameters. To be thorough in present-
ing the process and to complete the introduction to nonlinear regres-
sion, we next describe the four-parameter model option.
Nonlinear regression (Edu_Level)

3.5
Edu_Level
2.5
1.5
1
0 5 10 15 20 25
Depression
Active Model
3.5 Pred(Edu_Level)/Edu_Level
2.5
Edu_Level
1.5
0.5
0.5 1 1.5 2 2.5 3 3.5
Pred(Edu_Level)
Residuals
2
1.5
1
Residual
0.5
0
Obs1
Obs2
Obs5
Obs7
Obs8
Obs10
Obs13
Obs16
Obs17
Obs18
Obs22
Obs23
Obs24
Obs27
Obs6
Obs11
Obs25
–0.5
Obs21
Obs4
Obs26
Obs19
Obs3
Obs9
Obs15
Obs12
Obs14
Obs20
–1
Observations
Figure 5.10. Analysis output graphs from nonlinear regression.
ANALYZING THE DATA IN XLSTAT AND EXCEL —FOUR PARAMETERS

To conduct a nonlinear regression with a model with four parameters,
you must add a few additional bits of information into Excel, because
the probability model with four parameters is not part of the predefined
functions within XLSTAT. The additional bits of information are four

formulas that represent the derivatives of the four-parameter function.
If you do not know what a derivative is, don’t worry —it involves calcu-
lus and is beyond the scope of this book. You can simply follow the steps,
which are very easy to perform, to get XLSTAT ready to use this new
model, and learn more about derivatives after you have mastered the
steps. One nice feature is that once you add the four-parameter model to
XLSTAT, you will not have to do it again when you conduct a nonlinear
regression analysis in the future.
By adding the four formulas below, you are providing XLSTAT
with the next model option in the set of options for this example. These
four model options (i.e., functions with one, two, three, and now four
parameters) are part of the exponential family of functions for mod-
eling nonlinear relationships between a dependent variable and inde-
pendent variable(s). In the future, if you have a prediction model that
involves a nonlinear relationship, these four formulas are a great place
to start. Simply follow the steps presented in this example for answering
Research Question #20 with your own data.
Before returning to the nonlinear regression window in XLSTAT,
you must create a new Excel sheet by naming it derivatives and enter-
ing the following four formulas into column A. However, if you have
downloaded the data set for Chapter 5 from the companion website,
you will find the sheet named derivatives ready for you to use. The four
formulas are:
(pr3 / pr 4) * exp(− pr1 − pr2 * X1) / (1 + exp(− pr1 − pr2 * X1))(1+1/ pr 4 )

(pr3 * X1 / pr 4) * exp(− pr1 − pr2 * X1) / (1 + exp(− pr1 − pr2 * X1))(1+1/ pr 4 )
1 / (1 + exp(− pr1 − pr2 * X1))(1/ pr 4 )
(pr1 / pr 42 ) * ln(1 + exp(− pr1 − pr2 * X1)) / (1 + exp(− pr1 − pr2 * X1))(1/ pr 4 )
Said differently, if you have not yet downloaded the Chapter 5 data
set, open an Excel sheet that contains no information, type the four
formulas exactly as written above into the first four lines in column
A in the empty sheet and label it derivatives in the label space at the
bottom of the sheet. These four formulas are the same formulas you
can use with other data when examining a model with four param-
eters. Again, you do not need to understand the workings of these
formulas for now, just know that XLSTAT needs these four to sup-
port your effort to examine a four-parameter model that XLSTAT
did not predefine. Once you enter the four, and after verifying that
you typed them correctly, return to XLSTAT and set up the new
probability model.
To continue your analysis in XLSTAT, you’ll need to reopen the
□ Select => Modeling data => Nonlinear regression to open the

□ Select => Functions tab.
□ This time, click to select the User defined functions option on
the right side of the Nonlinear regression window.
After the User defined functions option is selected,
□ Click on the Add button toward the bottom right of the window.
When you click the Add button, the two white areas for Function: Y =
and Derivatives: become active.
□ Be sure the Derivatives: option is selected.

□ Click to move focus to the Derivatives: white space area.
Next, we want to let XLSTAT know where to find the four formu-
las entered earlier into that blank Excel sheet that you have named
derivatives.
□ Navigate to where the four formulas are, and click and drag to
highlight all four of them (i.e., click and drag adds the location
information into the area —derivatives!$A$1:$A$4).
□ Click to move the focus to the white space for the Function:
Y = area.
□ Type the formula pr3/(1 + exp(-pr1-pr2*X1))^(1/pr4)
Note: The formula starts with “p” and ends with “)” —a closed paren-
thesis symbol. Also, once you add this as a user-defined function,
you will not have to add it again for future analyses unless you update
XLSTAT with a new version.
□ After you verify the formula is correct, click the Save button to
add the formula to the User defined functions list.
The new four-parameter function should be highlighted, telling

XLSTAT to use it for the analysis. Therefore, adding the four formu-
las into the blank Excel sheet and adding the new function into the
Function: Y = area has enabled XLSTAT to use this new probability
model with four parameters in an analysis.
After carefully reviewing your work,

analysis.
Once OK is clicked, a final verification window appears (i.e., XLSTAT

— Selections), giving you one last view of the data selected for the
analysis.
□ Select => Continue button and a new Excel sheet will appear
with all the requested analysis output.
Excel Output —Four Parameters

A review of the output from the four-parameter model shows an insig-
nificant increase in R 2 from the three-parameter model (from 0.571 to
0.575). Recall that when a less complicated model fits relatively the same
as a more complicated model, the less complicated model is the better
selection, so you should clearly use the results from the three-parameter
model. A review of the three-parameter model fit statistics shows that
the RMSE is relatively small, and a review of the estimates of the three
parameters seems reasonable. Therefore, our final nonlinear model is
within the Excel sheet under Equation of the model:. After rounding of
a few of the parameter estimates, the final model is:
(
Edu_Level = 2.80 / 1+ Exp ( +0.61 - 0.24* Depression ) )
The above formula can be found in the Excel sheet that was generated
from the analysis. The formula is located on, or close to, line 41 on the
Excel sheet. Because the model fit is adequate, you can enter a value
for Depression into the above equation to predict what category is most
likely —the education level category in which the respondent will most
likely be. One way to accomplish this is to use Excel.
□ Open a new Excel sheet.

□ Enter a value for Depression (e.g., 10) into cell A1.
□ Enter the following into cell B1:
(
= 2.80 / 1 + exp (.61 − .24 * A1) )
□ After typing in the formula above, press the Enter key on your
keyboard to calculate a value for Edu_Level that will appear in
cell B1. The value that appears in cell B1 is most likely a number
that you can round to the nearest integer. The integer number
is the predicted Edu_Level category for the value of Depression
you entered in A1.
Findings
A three-parameter nonlinear model with an ordinal dependent vari-
able Edu_Level and a continuous independent variable Depression
was identified for estimating education and training level (Edu_Level)
using scores on a depressive symptoms measure (R 2 = .571). The three-
parameter fit indices for RMSE and R 2 were much improved over a
two-parameter model. In addition, R 2 for a four-parameter model was
the same as the three-parameter model, and the RMSE showed only a
slight increase, results that support the three-parameter model as most
efficient.
Practically, this means that the nonlinear model has identified a
relationship between Depression and Edu_Level (level of education
and training). These findings from this sample suggest that levels of
depressive symptoms do predict variations in education and training,
but more data collection and analyses are necessary to conclude a causal
relationship. A different and perhaps more intuitive regression model
could reverse the roles of the variables (i.e., education and training as
a predictor of depressive symptoms level). Knowing which to choose as
the dependent variable centers on the research question you are attempt-
ing to answer and the type of data available. This alternative regression
model would explore the possibility that education and training lower
depressive symptoms levels.
Note: Another SPSS analysis method for models with ordered catego-
rical dependent variables is PLUM (Polytomous Universal Model).
For more information on this nonparametric strategy, see the website
https://s tatistics.laerd.com/spss-t utorials/ordinal-regression-u sing-
spss-statistics-2.php

Dr. Northcott is interested in examining professional life satisfaction, as the
literature suggests that this is related to turnover, and turnover is a serious
problem in adoption and foster care agencies. She suspects that those working
in the public sector may experience lower professional life satisfaction than
those working in the private sector. Based on studies of other professions, she
also suspects that men have higher professional life satisfaction than women.
To investigate these relationships in her pilot study, Dr. Northcott wants
you to examine the possible relationships between life satisfaction and two
factors —working in the private or public sector and if the person is male or
female. Note that we are assuming a dichotomous variable for gender for this
example only, but that gender, along with biological sex, is more complicated
than a dichotomy.
Are employment setting and gender significant for understanding levels of
professional life satisfaction?
FACTORIAL ANALYSIS
Factorial analysis, also called general linear model (GLM) analysis, is
a method for conducting a regression analysis in which the dependent
variable is continuous (in this example) and the independent variables
are categorical (i.e., the factors in factorial analysis). This analysis exam-
ines the effects factors have on the dependent variable. In other words,
a factorial analysis explores the potential effects on a dependent variable

due to being in one group rather than another. The test similar to a fac-
torial analysis is the traditional regression with a continuous dependent
variable and with multiple dummy variables as independent variables.
The dummy variables represent whether the respondent is or is not in
each group for each variable. In some situations, adding the necessary
number of dummy variables is problematic and meeting the assump-
tions of parametric regression is sometimes challenging, especially
when dealing with small samples.
Unfortunately, no nonparametric approach is easily available to
conduct a factorial analysis that matches Research Question #21 (see
Box 5.4). Yes, a few methods that will work have been published, but
they require programming/coding experience in SAS or R. Those who
do not have that kind of experience need an alternative. Most impor-
tantly, the alternative must not be to run an analysis just because
you found software that will run an analysis. Without the statisti-
cal resources and experience, Research Question #21 becomes much
more challenging because you need to lean on the available parametric
approach to proceed. Although a nonparametric approach is not easily
available, we present a suggested SPSS parametric procedure because of
its many applications to small data sets, and knowing what to look for
when building factorial models is a very important skill to have when
analyzing different kinds of data.
Note: SPSS will be happy to provide you with analysis results, even if
the results are fundamentally flawed from running an analysis on data
that do not meet the required assumptions, misinterpreting a variable,
having the wrong level of measurement, experiencing unrecognized
bias, etc. It is the responsibility of every researcher to examine fully all
the possible threats and biases to their findings before presenting their
results to others.
Factorial models have many assumptions and requirements, most of
which are discussed earlier in this book. The assumptions are normal-
ity, homoscedasticity, independence, and a continuous variable for the
dependent variable. In addition, you need to make sure that you have
enough coverage in your categorical variables. Coverage of the variable
categories is related to having enough representation in the options to
support the use of the variable in the analysis. For example, if a varia-
ble for gender has data for 20 females and three males, according to an
accepted guideline of five per category, you do not have enough males
representing the male category —the variable does not have enough
information to compare males and females.
The five-count guideline becomes more difficult to meet as you
add more categorical variables into a factorial model. For example, if
Research Question #21 aimed at knowing if gender and education level
were predictors of depressive symptoms, the data do not have enough
coverage to pursue this analysis (see Table 5.2). Remember that Edu_
Level is an ordinal measure of education and training (1 = bachelor’s
degree only, 2 = master’s degree with no license, and 3 = master’s degree
with clinical license).
Review of Table 5.2 reveals that cross-tabulation of males with edu-
cation level does not meet the five-count guideline. In other words,
you do not have enough males in Education Levels 1 and 2 to represent
properly all the potential males who are at those education levels, and
the same is true for both males and females in the Education Level 3
category.
Sometimes the coverage for the various categories is such that col-
lapsing a category is necessary. For example, one option for Research
Question #21 would be to merge education categories 1 and 2 to com-
pensate for the low number of men, but the few numbers and therefore
the lack of information for category 3 would still be unresolved, and
just managed for analysis purposes. It is important to note that SPSS
will perform a factorial analysis on a model with Gender and Edu_Level
unchanged, but the potential of coming to a flawed conclusion based on
these variables is high. SPSS does not always warn you when cell cov-
erage is too low. Therefore, with small data sets, at times you may not
Table 5.2. Frequency for Gender and Edu_Level —Edu_Level * Gender

Crosstabulation
Count Gender Total
Female Male
Edu_Level 1 6 2 8
2 6 4 10
3 3 4 7
Total 15 10 25
be able to use specific combinations of variables, depending upon the

coverage in each cell.
SPSS Process (Using SPSS to Build a Factorial Model)

We now return to using SPSS, because SPSS offers the process needed to
analyze Research Question #21. To begin your analysis,
□ Select => Analyze => General Linear Model => Univariate.
You select the Univariate option because you have only one dependent
variable, named LifeSatisfaction. Once the Univariate window opens,
you need to tell SPSS which variable to analyze. The list on the left
includes the variables in the data set.
□ Click to select LifeSatisfaction from the list of variables on the left.

□ Click the move arrow closest to Dependent Variable: to
move the variable to the Dependent Variable: area.
□ While holding down the Ctrl key, click to select both the Gender
and Employment variables so they are highlighted.

□ Click the move arrow closest to Fixed Factor(s): to move the
variables to the Fixed Factor(s): area.
Why use Fixed Factor(s): instead of random (i.e., Random Factor(s))?

Variables considered fixed factors are those that contain information
on all the possible categories represented by the variable. (For more
information on fixed vs. random factors, see Raudenbush, 1993.)
Before you select the OK button at the bottom of the window, you
must select a few analysis options.
□ Select => Model… button on the left to open the

Univariate: Model window.
Verify that the Full factorial option is selected because this will test
potential differences for Gender, Employment, and their interaction
(Gender * Employment) for LifeSatisfaction.

Next, we want to request plots to help verify the results, so,
□ Select => Plots… button on the right to open the

Univariate: Profile Plots window.
After the window opens, you will see the two factors Gender and
Employment in the Factors: area. With only two factor variables, we
suggest that you request two plots —one with Gender on separate lines
and one with Employment on separate lines.
□ Click to select Gender in the Factors: area.

□ Click the move arrow closest to Separate Lines: to copy
Gender to the Separate Lines: area.
□ Click to select Employment in the Factors: area
□ Click the move arrow closest to Horizontal Axis: to copy
Employment to the Horizontal Axis: area.
□ Request the plot by clicking on the Add button, which will add
Employment*Gender to the Plots: area. Follow these same steps

to add Gender*Employment to the Plots: area to see the same
information, just a different view.
□ After carefully reviewing your choices, select => Continue
button to close the window.
Because the factors Gender and Employment are dichotomous, running

a post hoc test to examine differences between variable categories is not
necessary. However, we do need to select some options to verify a few
assumptions about the data.
□ Select => Options… button on the right to open the

Univariate: Options window.
□ Click on the (OVERALL) entry in the Factor(s) and Factor
interactions: area.
□ Click the move arrow to copy (OVERALL) to the Display
Means for: area.
□ Select four options in the Display area (Descriptive
statistics, Parameter estimates, Homogeneity tests,

and Residual plot).
Note: You can click the Help button on this window to get additional
information on all the available options.

□ Select => OK to conduct the analysis.
See Appendix A for complete syntax.
SPSS Output
After you select OK, a variety of information shows up on the SPSS out-
put window. On the left side of the output window, you will see an index
showing the names of the information sections created from running the
analysis. The names are arranged in an index, and clicking on the name
displays the details of that section on the right. On the right side of the out-
put window, you will see all the detailed information within each section.
If you have been running other analyses before your factorial model,
you may have other information in your output window. If this is the
case, simply click on Univariate Analysis of Variance in the index on
the left side of the output window. This will cause the left side of the win-
dow to scroll to the detailed information for the analysis output.
Findings
In the Descriptive Statistics table in the IBM SPSS Statistics Viewer win-
dow, you can verify that you selected the correct variables for the anal-
ysis and that the data provide sufficient coverage for the combination of
categories (i.e., values for n meet the guideline of five or more per cell).
The next table, Levene’s Test of Equality of Error Variance, sug-
gests (p = .213) that the statistical assumption about homogeneity of
variances among the groups has not been violated. A Sig. value smaller
than α =.05 would indicate the variances are potentially different among
the factors and therefore may require a different analysis approach,
because an assumption has been violated.
When you scroll down to the next section of output information,
the Test of Between-Subjects Effects table shows the results of your
general linear model (GLM) analysis —Factorial Models are part of the

GLM family. The values in the Sig. column show that LifeSatisfaction
between private and public employment is not significantly different,
and neither is the interaction of male and female with employment type
(Employment, p = .951, and Gender*Employment, p = .167). This indi-
cates that there is not enough information to assume that life satisfaction
levels for males with public employment are different from life satisfac-
tion levels for females with private employment. However, the Sig. value
for Gender (p = .043) falls below the commonly used threshold of α =.05.
This indicates that information from the data suggests that life satisfac-
tion levels, in general, are different for males and females.
Another important value to review in the Test of Between-Subjects
Effects table is the Adjusted R Squared at the bottom of the table. R2
is the same as R Squared on SPSS’s note at the bottom of the table, and
recall that R2 represents the amount of variability in the dependent var-
iable explained by the independent variables. In this analysis, Gender
and Employment explain about 15% of the variability in LifeSatisfaction.
Many authors elect to report the R2 value, but the Adjusted R2 calcula-
tion is more informative about the performance of the model because it
takes into consideration the number of independent variables used. As
more variables are added to a model, the R2 value naturally increases,
and the Adjusted R2 corrects for this general increase and takes into
account the sample size.
The remaining three graphs in the analysis output provide an
opportunity to verify that another assumption has not been violated
and to see a graphical representation of the Sig. values found in Test
of Between-Subjects Effects table. The first graph, with Dependent
Variable: LifeSatisfaction at the top, illustrates the observed vs. pre-
dicted values for the model, which reveals the amount of difference
between the two. A review of this graph should include a search for
possible patterns that may indicate a deviation from equal variance
for LifeSatisfaction among the groups. For Research Question #21,
the graph of predicted vs. observed does not show any patterns, which
is what you would expect due to the findings of the Levene’s Test of
Equality of Error Variance discussed earlier.
The last two plots in the Profile Plots section show estimated means
for the factors, one by Employment and the other by Gender. Recall,
we asked for two plots that contained the same information under the
Plots button prior to running the analysis. The differences in how the
lines are drawn using the Estimated Marginal Means values is larger for
Gender than it is for the Employment Settings. Using the vertical case for
Estimated Marginal Means will help in comparing the two graphs. The
observed differences in the two graphs support the p-values for Gender
(p = .043) and Employment Setting (p = .951).
Another important observation is the intersection of the two lines
for Gender (see Figure 5.11). Intersecting lines can indicate a potential
significant factor interaction. However, in our analysis, the p-value for
Gender*Employment was not significant (p = .167). Because the sam-
ple is very small, there may not be enough information to identify a
significant interaction of Gender and Employment, especially with
Employment categories involved (p = .951).
A good way to test the effects of the nonsignificant interaction on the
model is to remove the interaction term altogether and rerun the anal-
ysis. To accomplish this,
□ Select => Analyze => General Linear Model => Univariate to

reopen the Univariate window.
Estimated Marginal Means of Life Estimated Marginal Means of Life

Satisfaction Satisfaction
17 17
Estimated Marginal Means
Estimated Marginal Means
16 16
15 15
14 14
13 13
12 12
f m Private Public
Gender Employment
Employment Gender
Private f
Public m
Figure 5.11. Estimated Marginal Means of Life_Satisfaction.

□ Select => Model… button to open the Univariate: Model

window.
□ Click to select the Custom option to gain access to the
Model: area.
□ Click to select Gender in the Factors & Covariates: area.
□ Click the move arrow to move Horizontal Axis: to the
Model: area.
In other words, because the interaction term Gender*Employment and

the factor Employment were found not to be significant, we are going to
rerun the analysis with only the factor Gender.

□ Select => OK button to rerun the analysis. If you get a warning
message that one of your factors (Employment) is not being

used, simply click the OK button to continue.
With the updated model, repeat your review of the Levene’s test and
check the assumptions as previously described. For this analysis, after
repeating the process, the Tests of Between-Subjects Effects table shows
that Gender is significant (p = .031) and that the Adjusted R Squared of
.152 is just .003 smaller than the previous model. Therefore, this simpler
model, with just a trivial drop in Adjusted R Squared explaining 15% of
LifeSatisfaction variance, suggests a more efficient model for analyzing
Research Question #21. The Parameter Estimates table shows a signif-
icant parameter estimate for Gender (β = -3.067, p = .031), providing
evidence that mean differences for LifeSatisfaction exist between males
and females (see Table 5.3).
Findings
In statistical terms, a GLM model examining LifeSatisfaction for factors
Gender and Employment found that Employment and the interaction
of Employment and Gender were not significant. A model with only
Gender was significant (p = .031) at α = .05 with an adjusted R2 = .152.
Table 5.3. Parameter Estimates

Parameter B Std. Error t Sig. 95% Confidence
Interval
Lower Upper
Bound Bound
Intercept 15.800 1.031 15.323 .000 13.667 17.933

[Gender = f] –3.067 1.331 –2.304 .031 –5.820 –.313
[Gender = m] 0a . . . . .
a
This parameter is set to zero because it is redundant.
Dependent variable: LifeSatisfaction.
b
Practically, this means that the GLM analysis shows that you do not
have enough evidence to suggest a difference in life satisfaction between
individuals who have public or private industry employment. However,
there is sufficient evidence to suggest that, in general, males and females
have different life satisfaction levels, and that females, on average, score
three points lower than males when measured by a validated profes-
sional life satisfaction scale. In addition, gender accounts for 15% of the
variance in life satisfaction in the way life satisfaction was measured for
this study.
Appendix A
SPSS Syntax
Appendix A presents the SPSS syntax for each of the procedures pre-
sented in this book, identified by chapter and by research question.
Working with the syntax is more comfortable for some researchers,
as they find it easier to follow the specific characteristics of each ana-
lytic procedure. In addition, the capability and usefulness of SPSS can
greatly increase when a researcher moves beyond simply using menus
and point-and-click procedures.
CHAPTER 2
*Frequency Analysis:
FREQUENCIES VARIABLES=Confidante
/ORDER=ANALYSIS.
*Parametric T-Test
T-TEST
/TESTVAL=0
/MISSING=ANALYSIS
199
200 Appendix A
/VARIABLES=Confidante
/CRITERIA=CI(.95).
*Binomial Analysis:
*Nonparametric Tests: One Sample.
NPTESTS
/
ONESAMPLE TEST (Confidante) BINOMIAL(TESTVALUE=0.6
CLOPPERPEARSON JEFFREYS LIKELIHOOD
SUCCESSCATEGORICAL=LIST(1) SUCCESSCONTINUOUS=CUTPOINT(MIDP
OINT))
/
MISSING SCOPE=ANALYSIS USERMISSING=EXCLUDE
/
CRITERIA ALPHA=0.05 CILEVEL=95.
*Chi-square test:
NPTESTS
/
ONESAMPLE TEST (Coming _
Out) CHISQUARE(EXPECTED=EQUAL)
/
/
*Kolmogorov-Smirnov test:
NPTESTS
/
ONESAMPLE TEST (Loneliness)
KOLMOGOROV _
SMIRNOV(NORMAL=SAMPLE UNIFORM=SAMPLE
EXPONENTIAL=SAMPLE POISSON=SAMPLE)
/
/
*Wilcoxon signed-
rank test:
NPTESTS
/
ONESAMPLE TEST (Coping _
Level) WILCOXON(TESTVALUE=15)
/
/
Appendix A 201
*Descriptive Analysis:
DESCRIPTIVES VARIABLES=Rater _
score
/
STATISTICS=MEAN STDDEV MIN MAX.
*Runs test with mean cutpoint:
NPTESTS
/
ONESAMPLE TEST (Rater _
score) RUNS(GROUPCATEGORICAL=SAMPLE
GROUPCONTINUOUS=CUTPOINT(SAMPLEMEAN))
/
/
*Runs test with dichotomous variable:
NPTESTS
/
ONESAMPLE TEST (Rater _
score Confidante)
RUNS(GROUPCATEGORICAL=SAMPLE
GROUPCONTINUOUS=CUTPOINT(SAMPLEMEAN))
/
/
*Histograms for Age and Self _ esteem
FREQUENCIES VARIABLES=Age Self _ esteem
/HISTOGRAM NORMAL
/ORDER=ANALYSIS.
*Filter cases (use only youths who have not come out to
anyone)
USE ALL.
COMPUTE filter _
$=(Coming _
Out=1).
VARIABLE LABELS filter _ $ ‘Coming _
Out=1 (FILTER)’.
VALUE LABELS filter _
$ 0 ‘Not Selected’ 1 ‘Selected’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
*Filter cases (use only youths who have come out to one or
more people)
USE ALL.
COMPUTE filter _
$=(Coming _
Out<>1).
Out<>1 (FILTER)’.
FORMATS filter _
$ (f1.0).
202 Appendix A
FILTER BY filter _
$.
EXECUTE.
*Remove any active filter
FILTER OFF.
USE ALL.
EXECUTE.
*Correlation with all youths
NONPAR CORR
/
VARIABLES=Age Self _ esteem
/
PRINT=BOTH TWOTAIL NOSIG
/
MISSING=PAIRWISE.
*Sort cases
SORT CASES BY Age(A).
*filter and correlation with only youths who have not come
out to anyone
USE ALL.
COMPUTE filter _
$=(Coming _
Out=1).
Out=1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
NONPAR CORR
/
/
/
MISSING=PAIRWISE.
*filter and correlation with only youths who have come out
to one or more people
USE ALL.
COMPUTE filter _
$=(Coming _
Out<>1).
Out<>1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
NONPAR CORR
/
/
/
MISSING=PAIRWISE.
*Remove any active filter
FILTER OFF.
USE ALL.
EXECUTE.
Appendix A 203
CHAPTER 3
*Independent Group Comparison and CI
*Nonparametric Tests: Independent Samples.
NPTESTS
/
INDEPENDENT TEST (Percent _
Pre) GROUP (Received _
MI)
MOSES(TRIMOUTLIERS=SAMPLE) HODGES _
LEHMANN
/
/
*Independent Group Analysis:
NPTESTS
/INDEPENDENT TEST (Percent _ DIF) GROUP (Received _ MI)
KOLMOGOROV _ SMIRNOV
/MISSING SCOPE=ANALYSIS USERMISSING=EXCLUDE
/
*Independent Group Analysis using three tests:
NPTESTS
/INDEPENDENT TEST (Percent _ DIF) GROUP (Received _ MI)
MANN _
WHITNEY KOLMOGOROV _ SMIRNOV WALD _ WOLFOWITZ
/MISSING SCOPE=ANALYSIS USERMISSING=EXCLUDE
/
*Frequency for MI=1 respondents
USE ALL.
COMPUTE filter _
$=(Received _
MI=1).
VARIABLE LABELS filter _ $ ‘Received _
MI=1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
FREQUENCIES VARIABLES=Percent _ DIF
/HISTOGRAM NORMAL
/ORDER=ANALYSIS.
*Frequency for MI=2 respondents
USE ALL.
COMPUTE filter _
$=(Received _
MI=2).
MI=2 (FILTER)’.
204 Appendix A

FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
FREQUENCIES VARIABLES=Percent _ DIF
/HISTOGRAM NORMAL
/ORDER=ANALYSIS.
*Filter to get only respondents in the treatment group:
USE ALL.
COMPUTE filter _
$=(Received _
MI=1).
MI=1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
NPTESTS
/
INDEPENDENT TEST (Percent _ Post) GROUP (Partner)
KRUSKAL _
WALLIS(COMPARE=PAIRWISE)
/
/
*Filter to get only respondents in the treatment group:
USE ALL.
COMPUTE filter _
$=(Received _
MI=1).
MI=1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
*Jonckheere _ Terpstra test:
NPTESTS
/
INDEPENDENT TEST (Percent _ DIF) GROUP (Stage _
of _
Change)
JONCKHEERE _
TERPSTRA(ORDER=ASCENDING COMPARE=PAIRWISE)
/
/
Appendix A 205
NPTESTS
/
INDEPENDENT TEST (Percent _
DIF) GROUP (Stage _
of _
Change)
JONCKHEERE _
TERPSTRA(ORDER=ASCENDING COMPARE=PAIRWISE)
/
/
* set filter to only use respondents in treatment group
USE ALL.
COMPUTE filter _
$=(Received _ MI=1).
VARIABLE LABELS filter _ $ ‘Received _ MI=1 (FILTER)’.
FORMATS filter _
$ (f1.0).
FILTER BY filter _
$.
EXECUTE.
NPTESTS
/
INDEPENDENT TEST (Self _ efficacy) GROUP (Condom _ Request)
MEDIAN(TESTVALUE=SAMPLE COMPARE=PAIRWISE)
/
/
CHAPTER 4
* Analysis using McNemar test:
*Nonparametric Tests: Related Samples.
NPTESTS
/
R ELATED TEST(Violent _
crime _
w1 Violent _
crime _
w2)
MCNEMAR(SUCCESS=LIST(1))
/
/
*Descriptive Analyses:
FREQUENCIES VARIABLES=Friends _
w2 Friends _
w4
/ORDER=ANALYSIS.
NPTESTS
206 Appendix A
/
R ELATED TEST(Friends _
w2 Friends _
w4)
MARGINAL _
HOMOGENEITY
/
/
* Analysis using Sign test:
NPTESTS
/
R ELATED TEST(Anxiety _
w2 Anxiety _
w4) SIGN
/
/
*Analysis using Wilcoxon Signed- Rank test and CI:
NPTESTS
/
R ELATED TEST(Anxiety _
w2 Anxiety _
w4)
WILCOXON HODGES _
LEHMAN
/
/
COMPUTE Median _
DIFF=Anxiety _
w2 -Anxiety _
w4.
EXECUTE.
*Analysis using Cochran’s Q test:
NPTESTS
/
R ELATED TEST(Assault _
w2 Assault _
w3 Assault _
w4)
COCHRAN(SUCCESS=LIST(1) COMPARE=PAIRWISE)
/
/
*Analysis using Kendall’s Coefficient of Concordance:
NPTESTS
/
RELATED TEST(Support _
w1 Support _
w2 Support _
w3 Support _
w4)
Appendix A 207
KENDALL(COMPARE=PAIRWISE)
/
/
*Analysis using Friedman’s Two- Way ANOVA test:
NPTESTS
/
RELATED TEST(Support _
w1 Support _
w2 Support _
w3 Support _
w4)
FRIEDMAN(COMPARE=PAIRWISE)
/
/
CHAPTER 5
n/a
n/a
*Factor Analysis:
CROSSTABS
/TABLES=Edu _
Level BY Gender
/FORMAT=AVALUE TABLES
/CELLS=COUNT
/COUNT ROUND CELL.
UNIANOVA LifeSatisfaction BY Gender Employment
/M ETHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/POSTHOC=Gender(LSD BONFERRONI)
/PLOT=PROFILE(Gender*Employment Employment*Gender)
/EMMEANS=TABLES(OVERALL)
/PRINT=LOF HOMOGENEITY DESCRIPTIVE PARAMETER
/PLOT=RESIDUALS
/CRITERIA=ALPHA(.05)
208 Appendix A
/DESIGN=Gender Employment Gender*Employment.

UNIANOVA LifeSatisfaction BY Gender Employment
/M ETHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/POSTHOC=Gender(LSD BONFERRONI)
/PLOT=PROFILE(Gender*Employment Employment*Gender)
/EMMEANS=TABLES(OVERALL)
/PRINT=LOF HOMOGENEITY DESCRIPTIVE PARAMETER
/PLOT=RESIDUALS
/CRITERIA=ALPHA(.05)
/DESIGN=Gender.
Appendix B
Missing Data
Most researchers will have to deal with missing data at some point in
their careers. The challenge of missing data is that there are many reasons
for the missingness, and probably even more ways in which to address
missingness in a data set. As you may remember from your introduc-
tion to statistics course, missing data fall into three categories—missing
completely at random (MCAR), missing at random (MAR), and missing
not at random (MNAR). MCAR and MAR are, for the most part, man-
ageable when it comes to dealing with the gaps in the information (i.e.,
the missing data).
However, MNAR can be problematic. For example, the gaps in
information might be due to how a few respondents reacted to the data-
collection procedures—say they interpreted a survey question differ-
ently from the majority of other respondents, or perhaps they elected
to skip a response for personal reasons outside of the study issues. This
systematic missingness undermines the accuracy of your findings, and
in some cases can invalidate your study altogether, without your even
recognizing it. For this reason, whenever missing data are present, you
must conduct an extensive review of the reliability and validity of the
study methods. MNAR can be caused by issues in the data collection
methods, such as bias in sampling methods, uneven interpretation of
209
210 Appendix B
survey questions, and unbalanced response options on a question—just

to name a few. Graphing the variables, comparing groups of respon-
dents, asking others to offer reasons why a respondent would skip ques-
tions, and looking to see if any variables in the data set can predict
which respondents have missingness are just a few ways to examine a
data set for reasons for potential MNAR. This examination is necessary
to identify what type of missingness exists, to detect an explanation for
the missingness, and to help you decide what to do next.
Unfortunately, exploring the reasons for the missing data is the easy
part. Deciding what to do about the missing data can be much more
challenging, in part because there are many ways to address missing-
ness and thoughtful scholars occasionally disagree. The disagreement
is primarily due to the varying levels in which researchers are com-
fortable using information from other respondents to fill in the gaps.
With smaller data sets, a limited amount of information is available,
and using this limited amount to estimate other values that are missing
is ill advised. In a small data set, not enough information is available
to effectively estimate the missing values. This leaves very few options
when dealing with missing data in these small data sets.
When making these decisions, return to the primary responsibil-
ity of any researcher—to search for accurate answers using only the
information (both data and theory) available. Therefore, in these situa-
tions it is recommended to use only listwise (also called casewise) dele-
tion for dealing with missingness. This process deletes the cases with
missing values for variables included in your analyses. By using list-
wise deletion, you are acknowledging that the data are in some fashion
missing at random (i.e., not MNAR) and that replacing the missing data
with values from a type of statistical procedure (e.g., imputation, partial
imputation, mean replacement, etc.) would only create additional bias
in your findings. (For more information on missing data, see Allison,
2001; Enders, 2010; Rubin & Roderick, 2002.)
Appendix C
Other Resources
Appendix C contains a list of resources for additional information about

dealing with small samples, using nonparametric procedures, and sta-
tistics in general. The information below is in no particular order of
importance or quality, and is in no way an exhaustive list of helpful
options. The resources are grouped into three categories—books (and
book chapters), articles, and websites. If you are looking for an extremely
detailed and comprehensive statistics-driven reference book, we highly
recommend Nonparametric Statistical Methods by Myles Hollander and
Douglas A. Wolfe. However, if you are not ready for in-depth statisti-
cal explanations, we suggest Nonparametric Statistics for Health Care
Research, Second Edition, by Marjorie A. Pett. The articles listed below
are just a few resources that support ideas and concepts discussed in this
book, but if you have time to read only one article, we strongly suggest
you read David Kenny’s online article titled Measuring Model Fit. The
listed websites will take you to web locations that are rich with informa-
tion and how-to examples for all different levels of statistical experience.
Our favorite site happens to be the one managed by UCLA, which offers
so much information you can spend years reading their content on sta-
tistical procedures.
211
References
CHAPTER 1
Aguinis, H., Gottfredson, R. K., & Joo, H. (2013). Best-practice recommen-
dations for defining, identifying, and handling outliers. Organizational
Research Methods, 16(2), 270–301. doi:10.1177/1094428112470848
Barnett, V., & Lewis, T. (1978). Outliers in statistical data. Chichester: Wiley.
Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2d ed.).
Hillsdale, NJ: Lawrence Earlbaum Associates.
Cohen, J. (1992). A power primer. Psychological Bulletin, 112, 155–159.
doi:10.1037/0 033-2909.112.1.155
Corder, G. W., & Foreman, D. I. (2009). Nonparametric statistics for non-
statisticians. Hoboken, NJ: Wiley.
Davidson, A. C., & Hinkley, D. V. (1998). Bootstrap methods and their applica-
tion. Cambridge, U.K.: Cambridge University Press.
Efron, B. (1979). Bootstrap methods: Another look at the jackknife. The Annals
of Statistics, 7, 1–26. https://doi.org/10.1214/aos/1176344552
Elliott, A. C. (2015). Determine equality of variance for two groups. http://stat-
tutorials.com/EXCEL/EXCEL-EQUAL-VARIANCE.html
Gaskin, J. (2015). Plotting homoscedasticity in SPSS [Video file]. Available from
https://w ww.youtube.com/watch?v=WrQ1O_He63Q
Hojat, M., & Xu, G. (2004). A visitor’s guide to effect sizes: Statistical sig-
nificance versus practical (clinical) importance of research findings.
Advances in Health Sciences Education, 9, 241–249. https://doi.org/10.1023/
b:ahse.0000038173.00909.f6
213
214 References
Hoskin, T. (n.d.) Parametric and nonparametric: Demystifying the terms. Retrieved

from http://w ww.mayo.edu/mayo-edu-docs/center-for-translational-science-
activities-documents/berd-5-6.pdf
Hunter, M. A., & May, R. B. (1993). Some myths concerning parametric and
nonparametric tests. Canadian Psychology, 34, 384–389. https://doi.org/
10.1037/h0078860
Janssen, A., & Pauls, T. (2003). How do bootstrap and permutation tests
work?. Annals of Statistics, 31(3), 768–8 06. https://doi.org/10.1214/aos/
1056562462
Mathews, P. (2010). Sample size calculations: Practical methods for engineers
and scientists. Fairport Harbor, OH: Mathews Malnar and Bailey, Inc.
McDonald, J. H. (2014). Handbook of biological statistics (3d ed.). Baltimore,
MD: Sparky House Publishing.
Mumby, P. J. (2002). Statistical power of non-parametric tests: A quick guide for
designing sampling strategies. Marine Pollution Bulletin, 44, 85–87. https://
doi.org/10.1016/s0025-326x(01)00097-2
NIST/SEMATECH (2012). e-Handbook of statistical methods. Retrieved from
http://w ww.itl.nist.gov/div898/handbook/pri/section2/pri24.htm
Ryan, T. P. (2013). Sample size determination and power. Hoboken, NJ: Wiley.
Sawilowsky, S. (2005). Misconceptions leading to choosing the t-test over the
Wilcoxon Mann-W hitney test for shift in location parameter. Journal of
Applied Statistical Methods, 4, 598–600. Available from http://digitalcom-
mons.wayne.edu/coe_tbf/12
Stat Trek (2016). Transformations to achieve linearity. Retrieved from http://
stattrek.com/regression/linear-transformation.aspx?Tutorial=AP
StatSoft (2000). Multiple regression. Retrieved from http://w ww.uta.edu/fac-
ulty/sawasthi/Statistics/stmulreg.html#aassumption
Taylor, C. (n.d.). How do we determine what is an outlier? Retrieved from http://
statistics.about.com/od/Descriptive-Statistics/a /How-Do-We-Determine-
What-Is-An-Outlier.htm
Wasserman, R. L., & Lazar, N. A. (2016). The ASA’s statement on p-
values: Context, process, and purpose. The American Statistician, 70, 129–
133. http://d x.doi.org/10.1080/0 0031305.2016.1154108
Wu, C. F. J. (1986). Jackknife, bootstrap and other resampling methods in
regression analysis. The Annals of Statistics, 14, 1261–1295. https://doi.org/
10.1214/aos/1176350142
CHAPTER 2
Cochran, W. G. (1952). The χ2 test of goodness of fit. Annals of Mathematical
Statistics, 25, 315–345. https://doi.org/10.1214/aoms/1177729380
References 215
CHAPTER 3
Tsangari, H., & Akritas, M. G. (2004). Nonparametric ANCOVA with two and
three covariates. Journal of Multivariate Analysis, 88, 298–319. https://doi.
org/10.1016/s0047-259x(03)00098-8
CHAPTER 5
Chatterjee, S., & Hadi, A. S. (2012). Regression analysis by example. Hoboken,
NJ: Wiley.
Cleveland, W. S. (1979). Robust locally weighted regression and smoothing
scatterplots. Journal of the American Statistical Association 74, 368, 829–836.
doi:10.2307/2286407
Cleveland, W. S. (1981). LOWESS: A program for smoothing scatterplots by
robust locally weighted regression. The American Statistician 35, 54. https://
doi.org/10.2307/2683591
Cleveland, W. S., & Devlin, S. J. (1988). Locally- weighted regression: An
approach to regression analysis by local fitting. Journal of the American
Statistical Association, 83, 596–610. doi:10.2307/2289282
Dillman, D. A., Smyth, J. D., & Christian, L. M. (2014). Internet, phone, mail,
and mixed-mode surveys: The Tailored Design Method. Hoboken, NJ: Wiley.
Grande, T. L. (2015). Testing the assumption of independent errors with
ZERESID, ZPRED, and Durbin-Watson using SPSS. Retrieved from https://
www.youtube.com/watch?v=Lyk_S9T-oCw
Raudenbush, S.W. (1993). Hierarchical linear models and experimental design.
In L. K. Edwards (Ed.), Applied analysis of variance in behavioral science.
New York: Marcel Dekker.
APPENDIX B
Allison, P. D. (2001). Missing data. Thousand Oaks, CA: Sage Publications, Inc.
Enders, C. K. (2010). Applied missing data analysis. New York: Guilford Press.
Rubin, D. B., & Roderick J. A. L. (2002). Statistical analysis with missing data.
New York: Wiley.
APPENDIX C
Books and Book Chapters
Corder, G. W., & Foreman, D. I. (2009). Nonparametric statistics for non-
statisticians: A step-by-step approach. Hoboken, N.J: Wiley.
216 References
Deshpande, J. V., Gore, A. P., & Shanubhogue, A. (1995). Statistical analysis of

nonnormal data. Hoboken, NJ: Wiley Interscience.
Gibbons, J. D. (1993). Nonparametric statistics: An introduction. Newbury Park,
Green, S. B., & Salkind, N. J. (2014). Using SPSS for Windows and
Macintosh: Analyzing and understanding data. New York: Pearson.
Grotenhuis, M., & Matthijssen, A. (2016). Basic SPSS tutorial. Thousand Oaks,
CA: Sage.
Hollander, M., & Wolfe, D. A. (1999). Nonparametric statistical methods.
New York: Wiley.
Hoyle, R. H. (1999). Statistical strategies for small sample research. Thousand
Oaks, CA: Sage Publications.
Huff, D. (1993). How to lie with statistics. New York: Norton.
Krauth, J. (1988). Distribution-free statistics: An application-oriented approach.
Amsterdam: Elsevier.
Layder, D. (2013). Doing excellent small- scale research. Thousand Oaks,
Noether, G. E., & Dueker, M. (1991). Introduction to statistics: The nonparamet-
ric way. New York: Springer-Verlag.
Petersen, M. A. (2005). In K. Sijtsma & I. W. Molenaar (Eds.), Introduction to
nonparametric item response theory, Sage Publications Inc., Thousand Oaks,
CA, 2002, 168 pp. ISBN: 0-7619-0812-9.
Pett, M. A. (2016). Nonparametric statistics for health care research: Statistics
for small samples and unusual distributions. Thousand Oaks, CA: Sage
Publications.
Salkind, N. J. (2010). Statistics for people who (think they) hate statistics: Excel
2007 edition. Thousand Oaks, CA: Sage Publications.
Sprent, P., & Smeeton, N. C. (2001). Applied nonparametric statistical methods.
Boca Raton, FL: Chapman & Hall/CRC.
Wilcox, R. R. (2009). Basic statistics: Understanding conventional methods and
modern insights. New York: Oxford.
Wilcox, R. R. (2009). Nonparametric estimation. In D. A. Belsey &
E. Kontoghiorghes (Eds.), Handbook of computational econometrics (pp. 153–
182). West Sussex, U.K.: Wiley
Articles
Carroll, R. J., Ruppert, D., Crainiceanu, C. M., Tosteson, T. D., & Karagas, M. R.
(2004). Nonlinear and nonparametric regression and instrumental variables.
Journal of the American Statistical Association, 99, 736–750. http://d x.doi.
org/10.1198/016214504000001088
References 217
Friedman, M. (1937). The use of ranks to avoid the assumption of normal-
ity implicit in the analysis of variance. Journal of the American Statistical
Association, 32, 675–701. http://d x.doi.org/10.1080/01621459.1937.10503522
Friedman, M. (1939). A correction: The use of ranks to avoid the assumption
of normality implicit in the analysis of variance. Journal of the American
Statistical Association, 34, 109. http://d x.doi.org/10.2307/2279169
Geyer, C. J. (2005). Fuzzy p-values and ties in nonparametric tests. Retrieved
from http://w ww.stat.umn.edu/geyer/f uzz/ties.pdf
Han, L. (2008). Calculating the point estimate and confidence interval of Hodges-
Lehmann’s median using SAS software. Proceedings from the Southeastern
SAS Users’ Group (SESUG). Retrieved November 20, 2016 from http://ana-
lytics.ncsu.edu/sesug/2008/ST-154.pdf
Kenny, D. (November 24, 2015). Measuring model fit. Retrieved from http://
davidakenny.net/cm/fit.htm
Kraemer, H. C. (1976). The small sample nonnull properties of Kendall’s coef-
ficient of concordance for normal populations. Journal of the American
Statistical Association, 71(355), 608–613. http://d x.doi.org/10.1080/
01621459.1976.10481536
Park, H. M. (2003). Estimating regression models for categorical dependent vari-
ables using SAS, Stata, LIMDEP, and SPSS. Retrieved from https://w ww.ikhe-
beenvraag.be/mediastorage/FSDocument/101/LogitProbit.pdf
Sawilowsky, S. S. (1990). Nonparametric tests of interaction in experimental
design. Review of Educational Research, 60(1), 91–126. http://d x.doi.org/
10.3102/0 0346543060001091
Steele, J. M. (2005). Darrell Huff and fifty years of how to lie with sta-
tistics. Statistical Science, 20, 205–209. http://d x.doi.org/10.1214/
088342305000000205
Warner, C. B., & Meehan, A. M. (2001). Microsoft Excel as a tool for teach-
ing basic statistics. Teaching of Psychology, 28, 295–298. http://d x.doi.org/
10.1207/S15328023TOP2804_11
Wilcox, R. (2009). Comparing Pearson correlations: Dealing with heterosce-
dasticity and non-normality. Communications in Statistics—Simulation and
Computation, 38, 2220–2234. http://d x.doi.org/10.1080/03610910903289151
Wilcox, R. R. (2005). Depth and a multivariate generalization of the Wilcoxon-
Mann-W hitney test. American Journal of Mathematical and Management
Sciences, 25, 343–361. http://d x.doi.org/10.1080/01966324.2005.10737655
Websites
Data analysis examples by Institute for Digital Research and Education at
UCLA http://w ww.ats.ucla.edu/stat/dae/
218 References
Downloadable books listed by Institute for Digital Research and Education at

UCLA http://w ww.ats.ucla.edu/stat/books/downloadable_books.htm
Resources to help you learn and use SPSS by Institute for Digital Research and
Education at UCLA http://w ww.ats.ucla.edu/stat/spss/
Nonparametric methods by StatsDirect http://w ww.statsdirect.com/help/
default.htm#nonparametric_methods/nonmarametric_methods.htm%3FT
ocPath%3DNonparametric%2520methods%7C_____0
Nonparametric tests by IBM Knowledge Center http://w ww-01.ibm.com/sup-
port/k nowledgecenter/SSLVMB_20.0.0/com.ibm.spss.statistics.help/nonpa-
rametric_tests_intro_legacy.htm
Index
A C
assumptions, 2, 4, 5, 7, 10–15, 40, 43, 58, CFA, 161
73, 82, 88, 150, 172, 189, 193 coverage, 28, 40, 55, 117–8, 189–90, 193
equal interval, 17 response, 6, 17, 108
equal variances, 10 cut point, 51–3, 55, 103
heteroscedasticity, 13
homogeneity of variance, 7, 10, D
88, 193 dependent and independent variables
homoscedasticity, 7, 11, 150, dependent variable, 11–3, 101, 149, 153,
153, 189 171, 184, 188
independence, 6, 150, 172, 189 independent variable, 5, 11–3, 101, 149,
independent groups, 69–70, 88–9, 95 153–4, 171
independent observations, 7, 15 distributions, 73, 79, 83, 93–5, 108,
linear relationship, 7, 11–12, 116, 131
57–8, 150 bell-shaped curve, 7–8
multicollinearity, 171 binomial, 116
nonlinearity, 57 curvilinear, 12, 160
normality, 7–9, 47–8, 57, 88 distribution-f ree, 5, 89
parametric assumptions, 5, 7, exponential, 43, 45, 47, 184
10, 151 frequency, 154
unequal variances, 10, 57 kurtosis, 73, 154
non-normal, 13, 20, 57
B normal, 7–8, 12–3, 43, 54, 95, 101, 130,
bias, 59, 68, 157, 161, 189, 210 132, 153, 155, 176, 189
book conventions, 22 Poisson, 43, 45, 47, 63
bootstrapping, 3 probability distribution, 43
219
220 Index
distributions (cont.) M
skewness, 73, 131, 154 marginal probabilities, 111, 117
uniform, 43, 45, 47 measures of central tendency
mean, 5–6, 9, 16, 48, 52
E median, 16, 47–50, 73–4, 77, 91–3, 97, 101,
EFA, 161 103–4, 120, 123–5, 128, 131
effect size, 6, 19, 21, 87, 120 mode, 16
expected vs. observed, 40, 101 missing cases/missingness
missing, 4, 33–4, 87, 115, 123, 157, 161
F Monte Carlo simulation, 9, 19
filter, 60–1, 63–4, 87, 90, 96, 102 MSE, 164
frequency, 32–3, 88, 101, 118, 124, 132
N
G Nonlinear relationship, 150, 171
GLM, 188, 194, 196–7 nonparametric statistical procedures
group comparison, 29, 70–1 binomial test, 27, 31–3, 35–6, 39–41
guideline, 5, 18, 29, 33, 87, 118, 176, chi-square test, 27, 31, 40–2 , 109
190, 193 Clopper-Pearson, 36, 39
Cochran’s Q test, 108, 110, 132,
H 134, 137
HLM, 1 Durbin-Watson test, 150
hypothesis Friedman’s 2-way ANOVA, 108, 110,
alternative, 5, 94 142–3, 145–6
null, 5–6, 18, 21 Hodges-Lehmann confidence interval,
test, 2–3, 32, 37 73, 77, 110, 125, 128, 131
Hodges-Lehmann estimate, 69, 73
I Jonckheere-Terpstra for k samples, 69,
independent groups, 4, 68–9 72, 94–8, 100
independent samples, 69, 86–8, 109 Kendall rank correlation coefficient, 57
interventions, 10, 21, 78, 88, 93–4, 100 Kendall’s coefficient of concordance,
research, 1 110, 137–9, 142–3
IQR, 77 Kendall’s tau-b, 28, 31, 57–8, 64–7
Kendall’s W, 137
L Kolmogorov-Smirnov test, 27, 31, 43–7,
LCA, 163 63, 69, 72, 79–84, 86, 88
levels of measurement, 5, 7, 15, 41, 50, Kruskal-Wallis one-way ANOVA, 69,
108, 113, 189 72, 88–9, 91, 93–5, 97
categorical, 16, 22–4, 28–9, 36, 40, 44, Levene’s test, 193–4, 196
101, 117–8, 134, 171, 175, 188–90 Mann-W hitney U test, 69, 82–4, 86, 88
continuous, 7–8, 16–7, 23, 27, 36, Marginal homogeneity test, 107, 109,
43–4, 171 117, 119–20
dichotomous, 16, 28–9, 31–2 , 35, 40, 50, McNemar’s test, 107, 109–17
55, 111, 132 median test, 69, 101–4
interval, 5, 16–17 moses test of extreme reactions,
nominal, 5, 16, 22, 49, 58, 89 69, 73, 77
ordinal, 5, 16–19, 22, 40, 46–7, 57–8, 97 Nonparametric regression (LOWESS),
ratio, 5, 16, 23 149, 153, 165
scale, 7–9, 16, 23, 28–9, 43, 46, 48, Pearson’s chi-square test, 109
52, 107–8 rank order, 16, 58
Index 221
ranking, 138 nonlinear regression, 149, 152, 171–2 ,

Robust LOWESS regression, 153, 174–7, 180, 182–4
158–61, 164 paired-sample t-test, 111
runs test, 28, 31, 50–6 related-samples ANOVA, 143
sign test, 107, 110, 120–4, 131 repeated measures ANOVA, 132
Spearman’s rho, 28, 31, 57–8, 64–6 simple linear regression, 177
Wald-Wolfowitz runs, 50, 69, structural Equation Modeling, 1
82–4, 86 t-test, 31–3, 39–40, 48, 69, 79, 82, 86–8,
Wilcoxon matched pair signed rank 101, 107, 111, 120
test, 107, 110, 124–5, 127–8, 131 plots
Wilcoxon rank sum test, 82 box-a nd-whisker, 73, 76, 92–3, 104,
Wilcoxon signed-rank test, 27, 31, 173, 176
47–50, 124 histogram, 38, 42, 46, 59–63, 73, 87, 166
null hypothesis, 5–6, 18–19, 21–2 , 29, Q-Q plot, 173–6
39, 42, 63, 81, 116–17, 120, 124, scatter plot, 12–13, 59, 150, 159,
132, 138 165–6, 169
PLUM, 188
O power, 2, 5, 11, 19–21, 32, 42, 46, 48, 79,
outliers, 6, 9, 18, 48, 57–9, 75, 77, 87, 87, 120, 161
150–1, 160, 169 practical importance, 3
extreme outliers, 18, 150 predicted vs. observed, 162, 165, 194
extreme score, 48 prediction, 12–3, 149–51, 153–4, 164–6,
extreme values, 73 168–70, 184
probability model, 172, 176,
P 179–83, 185–6
pairwise probability sampling, 14–5, 109
combinations, 91, 103, 135, 140, 145
comparisons, 98–100, 137, 142 R
parameter, 5, 11, 45–6, 150, 166, 172, randomness, 28, 50–1, 55–6
176–7, 180–7 residuals, 7, 13, 150, 162, 165, 182
parametric statistical procedures response options
ANCOVA, 78 Likert-like measures, 17–8, 43–5,
ANOVA, 7, 29, 69, 88, 95, 107 47, 117
confirmatory factor analysis, 161 Likert-t ype, 8, 16–17, 23, 29, 107
correlation, 6, 11–12, 57–8, 138 RMSE, 164, 170, 186–7
exploratory factor analysis, 161 robustness, 2, 5, 20
factor analysis, 163, 189
factorial design regression, 149, S
152, 188 sampling
general linear model, 188, 194 random sample, 14
Hierarchical Linear Modeling, 1 randomization, 78
latent class analysis, 163 sample distribution, 143
least-squares regression, 4, 11, 151 sample mean, 45, 54
linear regression, 149–50, 154, 170, sample size, 6, 8, 11, 33, 40, 42, 54–5,
172, 176 89, 101, 111, 194
MANOVA, 69, 107 sampling error, 5, 21
multilevel modeling, 15 scales
multinomial regression, 171–2 validated measure, 30, 48, 57, 71
multiple regression, 12, 153, 171 SEM, 1
222 Index
small samples, 2, 8, 10–11, 20, 48, 55, 63, V

70, 94, 111, 116, 160, 176, 189 validity, 9, 17, 30, 108, 209
smoothing, 153, 159 variable
software age, 31, 59, 61–7
Excel, 4, 151, 157, 182–3 Anxiety_w2, 110, 121, 125–6, 129
R, 2, 189 Anxiety_w4, 110, 121, 125–6, 129
SAS, 2, 189 Assault_w2, 110, 133, 136
SPSS, 2–4, 151, 157, 177, 188 Assault_w3, 110, 133, 136
XLSTAT, 4, 151, 157, 172, 177, Assault_w4, 110, 133, 136
180, 184 Burnout, 152–5, 157, 160, 164–5, 167–70
SPSS options CESD, 52
all pairwise, 91, 96, 98, 103, 135 Coming_Out, 31, 41–2 , 60–1, 66
help feature, 24 Condom_Request, 72, 101, 104
SSE, 164 Confidante, 31–4, 36, 38–9, 55–6
statistical significance Coping_Level, 31, 48–50
alpha, 22, 39, 66, 125 Depression, 152, 172–3, 177, 182, 186–7
exact probability, 31–2 Depressive symptoms, 170–2 , 182,
exact p-value, 18, 116 187–8, 190
exact test, 31 Do you have a confidante, 32–4, 55–6
level of significance, 22 Edu_Level, 152, 172–5, 178, 182,
p-value, 6, 21–2 , 66, 116, 131, 150, 195 186–7, 190
significance, 3, 6, 22–3, 31, 34, 40, 48, Employment, 152, 188, 191–2 , 194–6
63, 76, 79–82, 87, 97–8, 100, 116, Explanatory variable, 159, 179
123, 137 Friends_w2, 110, 118
success and failure, 31, 36, 113 Friends_w4, 110, 118
failure, 24, 32, 113, 132 Gang membership, 109, 120
success, 23–4, 32, 36, 113, 119, 122, 127, Gender, 152–3, 189–92, 194–7
132, 134, 139, 145 Impairment, 152–4, 157–60,
164–5, 167–70
T Joined_gang_w2, 110–2
test statistic life satisfaction, 153, 188
Adjusted R Squared, 194, 196 LifeSatisfaction, 152, 191, 194, 197
chi-square, 101 Loneliness, 31, 44–7
confidence intervals, 19, 33–4, 36, 38–9, Median_DIFF, 129–31
73–4, 76–8, 80, 85, 125, 128, 131 Partner, 72, 91–3
goodness-of-fit, 43, 163, 181 Percent_DIF, 72, 80, 83, 86, 95–6
model fit, 157, 160, 162–3, 165, 170, 187 Percent_Post, 72, 91, 93
R Squared, 151, 154, 163–4, 170, Percent_Pre, 72, 74–5, 90
181–2 , 186–7 Pred_Burnout, 152, 168
standard deviation, 9, 44, 52, 164 Pred_Stressors, 152, 168
sum of squares of the errors, 164 Professional impairment, 153
variance explained, 151, 182 Rater_score, 31, 51–2 , 54, 56
z score, 31, 33 Received_MI, 72, 74–5, 80, 83, 87, 90,
thresholds, 22, 50–2 , 93, 98, 106, 96, 102
120, 194 Request_Condom, 102
ties, 6, 19, 58, 65–6, 138 Self_efficacy, 72, 101–2 , 104
transformation, 8, 12, 150–1 Self-efficacy, 100, 153
Type I error, 5, 21–2 , 29, 98 Self_esteem, 31, 59, 61–7
Index 223
Socioeconomic status, 58, 89, 154 Support_w4, 110, 138–9,

Stage_of_Change, 72, 95–6, 99, 100 141–4, 147
Stressors, 152–7, 160, 164–5, 167–70 turnover, 188
Support_w1, 110, 138–9, 141–4, 147 unsafe sex, 74, 79, 87–8, 100
Support_w2, 110, 138–9, 141, 143–4 Violent_crime_w2, 111–12
Support_w3, 110, 138–9, 141, 143–4 work stressors, 153
24
25
26
27
28

Data Analysis With Small Samples and Non-Normal Data - Nonparametrics and Other Strategies

Uploaded by

Copyright:

Available Formats

Data Analysis With Small Samples and Non-Normal Data - Nonparametrics and Other Strategies

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Data Analysis With Small Samples and Non-Normal Data - Nonparametrics and Other Strategies

Uploaded by

Copyright:

Available Formats

Data Analysis with

Determining Sample Size Finding and Evaluating Evidence

Published in the United States of America by Oxford University Press

© Oxford University Press 2018

All rights reserved. No part of this publication may be reproduced, stored in

You must not circulate this work in any other form

Library of Congress Cataloging-​in-​Publication Data

Reprint Courtesy of International Business Machines Corporation,

Appendix A: SPSS Syntax 199

We both thank Dr. Tony Tripodi for his encouragement and patience.

analyses of their data, or more importantly, the lack of exposure may

WHAT ARE NONPARAMETRIC STATISTICS?

parametric procedures. Nonparametric statistics are a set of tests that are

• Hypothesis —​a testable proposition that guides your study. The

• P-​value —​the probability of getting results at least as extreme

WHEN SHOULD NONPARAMETRIC PROCEDURES BE UTILIZED?

Although not all parametric procedures have the same set of

1. The population data are approximately normally distributed.

Other methodological assumptions are fundamental when performing

7. The data are sampled randomly from a defined population.

Let us examine these assumptions, one at a time, to understand,

Another challenge to verifying normality is when your data have

Assumption 2. The Population Variances Are Equal (i.e., Homogeneity of Variance

Assumption 3. The Sample Is Adequately Large

conceptualization of sample size is often simply related to the statisti-

Assumption 4. A Linear Relationship Exists Between or Among the Variables

correlation value representing the relationship between the dependent

Assumption 5. Homoscedasticity is Present in Regression Models

Assumption 6. Residuals (Errors) Follow a Normal Distribution

nonparametric process, such as nonparametric or nonlinear regression.

Assumption 7. The Data Are Sampled Randomly from a Defined

incorrectly treating a convenience sample as a probability sample is

Assumption 8. The Data Consist of Independent Observations (Apart from Paired

Assumption 9. The Levels of Measurement Match Variable Design

Ordinal —​ Qualitative information that is rank ordered, indicating the

Measurement Levels in SPSS

Scale —​Any variables that can be treated as continuous, such as

uncommon, because they require a meaningful, nonarbitrary possible

variable’s contribution is greatly increased. These decisions can be chal-

ADVANTAGES AND DISADVANTAGES OF NONPARAMETRIC PROCEDURES

• They have few, if any, assumptions.

normality requirement, or (c) transforming the variable makes

So why should a researcher not always use nonparametrics tech-

• Most importantly, when data are normally distributed, the

MISCONCEPTIONS ABOUT NONPARAMETRIC TESTS

• Parametric procedures are always more powerful than

A FEW WORDS OF CAUTION

conduct an a priori power analysis when planning a study, so

Nevertheless, the literature contains many misinterpretations of

cannot say if the null hypothesis is rejected, you have a 3% likelihood of

It is also important that you choose a level of significance (alpha)

CONVENTIONS UTILIZED IN THIS BOOK

and for use in this book, categorical also includes Likert-​

a success-​t ype test on a categorical variable with three or more

Figure 1.1. Data view within SPSS.

SUGGESTIONS FOR FURTHER READING

Pett, M. A. (2015). Nonparametric Statistics for Health Care

• Examining variables for the quality of information they provide

• A scale variable needing testing for randomness —​runs test

Library of Congress Cataloging-in-Publication Data

• Hypothesis —a testable proposition that guides your study. The

• P-value —the probability of getting results at least as extreme

Ordinal — Qualitative information that is rank ordered, indicating the

Scale —Any variables that can be treated as continuous, such as

and for use in this book, categorical also includes Likert-

a success-t ype test on a categorical variable with three or more

• A scale variable needing testing for randomness —runs test

Once the One-Sample Nonparametric Tests window opens, SPSS needs

The Settings tab in the One-Sample Nonparametric Tests window

Figure 2.1. One-Sample Nonparametric Tests window.

The binomial test investigates dichotomous variables, but the chi-

variable Coming_Out is operationalized with three categories (1 = no

When the One-Sample Nonparametric Tests window opens, SPSS needs

checkbox for the Chi-Square test.

Other applications for the Kolmogorov-

1. From your pilot study, you wish to analyze a scale/continuous

Use great caution when deciding to treat a Likert-like categori-

select the checkbox for the Kolmogorov-Smirnov test (uncheck

data/Sample mean) for each of the four distribution mean and

□ Select => OK at the bottom to close the Kolmogorov-Smirnov

• To see the details of the test, double-click the Hypothesis

to the uniform, Poisson, and exponential distributions. The Kolmogorov-

WILCOXON SIGNED-R ANK TEST

1. You are conducting a small study of a new weight-loss