ebook

THE GUILFORD PRESS

Data Analysis with Mplus
Methodology in the Social Sciences
David A. Kenny, Founding Editor
Todd D. Little, Series Editor
www.guilford.com/MSS

This series provides applied researchers and students with analysis and research
design books that emphasize the use of methods to answer research questions.
Rather than emphasizing statistical theory, each volume in the series illustrates when
a technique should (and should not) be used and how the output from available
software programs should (and should not) be interpreted. Common pitfalls as well
as areas of further development are clearly articulated.

R e c en t vol u m e s

ADVANCES IN CONFIGURAL FREQUENCY ANALYSIS

Alexander A. von Eye, Patrick Mair, and Eun-Young Mun

APPLIED MISSING DATA ANALYSIS

Craig K. Enders

DIAGNOSTIC MEASUREMENT: Theory, Methods, and Applications

André A. Rupp, Jonathan Templin, and Robert A. Henson

APPLIED META-ANALYSIS FOR SOCIAL SCIENCE RESEARCH

Noel A. Card

PRINCIPLES AND PRACTICE OF STRUCTURAL EQUATION MODELING,

Third Edition
Rex B. Kline

DATA ANALYSIS WITH Mplus

Christian Geiser

INTENSIVE LONGITUDINAL METHODS: An Introduction to Diary

and Experience Sampling Research
Niall Bolger and Jean-Philippe Laurenceau

DOING STATISTICAL MEDIATION AND MODERATION

Paul E. Jose
Data Analysis
with Mplus

Christian Geiser

Series Editor’s Note by Todd D. Little

THE GUILFORD PRESS

New York  London
Originally published in the German language by VS Verlag
für Sozialwissenschaften, 65189 Wiesbaden, Germany,
as “Geiser—Datenanalyse mit Mplus—978-3-531-16393-2”.

© VS Verlag für Sozialwissenschaften | GWV Fachverlage GmbH,

Wiesbaden 2010

English edition © 2013 The Guilford Press

A Division of Guilford Publications, Inc.
72 Spring Street, New York, NY 11012
www.guilford.com

All rights reserved

No part of this book may be reproduced, translated, stored in

a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, microfilming, recording,
or otherwise, without written permission from the publisher.

Printed in the United States of America

This book is printed on acid-free paper.

Last digit is print number: 9 8 7 6 5 4 3 2 1

Library of Congress Cataloging-in-Publication Data

Geiser, Christian.
[Datenanalyse mit Mplus. English]
Data analysis with Mplus / Christian Geiser.
   pages cm — (Methodology in the social sciences)
Includes bibliographical references and index.
ISBN 978-1-4625-0245-5 (pbk.) — ISBN 978-1-4625-0782-5
(cloth)
1. Multivariate analysis—Data processing. 2. Mplus. I. Title.
QA278.G4513 2013
519.5′35028553—dc23
2012033585
Series Editor’s Note

I’m very pleased to introduce Christian Geiser’s book to you. This book
was originally published in German by Springer Verlag. I felt it was so
important to bring the book to a broader readership that we secured the
rights to produce it in English as part of The Guilford Press series Meth-
odology in the Social Sciences. Given that Christian is a very talented
multilingual scholar, I knew that rendering this excellent work in English
would be both easy for him and an important contribution. Special thanks
go to the legal and contracts team at The Guilford Press who worked dili-
gently to secure the English-­language rights to Christian’s book and to
Whitney Moore, whose assistance in bringing this book to fruition was
above the call of duty.
Why am I so pleased to bring this book to you? Simply, it is the most
thorough and accessible book available on how to use Mplus for all man-
ner of statistical analysis procedures. Mplus is a popular software package
that is capable of estimating a wide variety of useful analyses. Figuring
out how to get Mplus to do most of the analyses, however, is both chal-
lenging and frustrating if you rely on the Mplus User’s Guide. Christian’s
book is clearly organized, covers a wide variety of analyses that Mplus can
estimate, and is easy to follow.
In other words, Data Analysis with Mplus fills an important gap. It is
not focused simply on a cookbook model of how to run the various analy-
ses in Mplus. Because it is written for the applied researcher and not the
quantitative analyst, the models are accessible. Christian explains each
v
vi   Series Editor’s Note

analysis model step-by-step. He does not presume that you will have used
Mplus before. In fact, he actually begins with how to prepare and bring in
data. This feature is a rarity and an important one because it is an essen-
tial step to master before you can do any of the fun analyses!
A wide variety of different analyses are described in this book. In a
clear and methodical manner, Christian explains what the syntax means,
teaches you very useful Mplus tricks, and actually alerts you to the defaults
that exist (which can be a mystery sometimes). Because of his skillful syn-
tax pedagogy, new users will feel comfortable knowing what the syntax
means and not be intimidated about using it, which is important for new
researchers who do not have much exposure to the world of writing syn-
tax for data analyses.
Christian’s focus on different types of analyses will help build your
knowledge and understanding of Mplus syntax. For example, some analy-
ses are presented first as manifest variable techniques before the complex-
ity of a measurement model is added. This piecewise exposure helps avoid
the overwhelming effect of presenting both manifest and latent variables
in the same model and allows for highlighting important differences in
the required syntax. Christian does a terrific job of explaining Mplus
syntax in bite-size pieces so that a reader without prior knowledge can
understand it. This aspect of his book is essential, I think, because often it
is the failure of nonquantitative users to grasp a simple, or basic, concept
that results in their not understanding the following steps. The chapter-­
specific recommendations focus on the applied researcher.
Christian’s book is an excellent resource manual for anyone wishing
to use Mplus for advanced statistical data analyses. It is a terrific comple-
ment to Rex B. Kline’s Principles and Practices of Structural Equation Model-
ing, Timothy A. Brown’s Confirmatory Factor Analysis for Applied Research,
as well as my forthcoming book, Longitudinal Structural Equation Modeling.

Todd D. Little
Berlin, Germany
Preface

Multivariate statistical methods with or without latent variables are more

and more frequently used in virtually all areas of social science research.
One reason is that many studies examine research problems that involve
complex relationships among a variety of manifest and/or latent variables.
Multivariate statistical approaches with latent variables are particularly
useful—­ and are becoming increasingly popular—­ because they allow
researchers to explicitly take errors of measurement into account, thereby
avoiding bias in parameter estimates and standard errors (Bollen, 1989).
The computer program Mplus allows the analysis of a wide variety of
multivariate statistical models with or without latent variables in a single
comprehensive modeling framework (Muthén, 2002; Muthén & Muthén,
1998–2012). A special feature of Mplus is that it allows for both the esti-
mation of models with continuous latent variables (e.g., confirmatory
factor analysis, structural equation modeling, item ­response theory) and
models with categorical latent variables (so-­called latent class variables).
Latent class models (see Chapter 6) can also be combined with models for
continuous latent variables (see Chapter 3), for example, in the analysis of
factor or growth mixture models.
In addition, Mplus is flexible in handling data situations that violate
assumptions of common methods of parameter estimation, such as maxi-
mum likelihood (ML). For example, ML estimation requires multivariate
normality of the variables. The assumption of multivariate normality is
often violated in practice, for instance, when items with ordinal-­scale level
vii
viii  Preface

are used as indicators of latent variables (Finney & DiStefano, 2006). In

addition, many statistical procedures require independent observations
drawn from random samples. Many empirical studies, however, make use
of cluster samples in which individuals are nested within groups (e.g.,
children nested within school classes). Such cluster sampling violates
the assumption of independent observations and requires the use of the
multi­level modeling techniques implemented in Mplus (see Chapter 5).
Mplus offers a variety of robust methods of parameter estimation so
that ordinal, otherwise non-­normal, and clustered data can be properly
analyzed. In addition, Mplus can be seen as a relatively user-­friendly pro-
gram. Even though models have to be specified via syntax commands
in Mplus, the Mplus syntax is fairly straightforward and easy to learn.
Complex models can be specified with relatively few lines of code, and the
user does not necessarily have to be an expert in programming or matrix
algebra to be able to specify a complex latent variable model.
The present book provides an introduction to multivariate data anal-
ysis with Mplus as well as to the theory and practical application of differ-
ent types of multivariate statistical methods that are particularly relevant
to social science research: cross-­ sectional structural equation models
(SEMs; including regression, confirmatory factor, and path analysis;
Chapter 3), longitudinal SEMs (including latent state, latent state–trait,
autoregressive, latent change, and latent growth curve models; Chapter
4), multilevel regression analysis (Chapter 5), and classical latent class
analysis (LCA; Chapter 6).
Because the goal of the book is to introduce readers to statistical com-
puting in Mplus in an application-­oriented and user-­friendly way, empiri-
cal data examples are used throughout the book. Many of these examples
reflect typical analyses carried out in the social sciences and can thus be
directly transferred to a user’s own research questions. Numerous screen
shots and output excerpts guide the reader through the analyses and the
interpretation of the Mplus output in a step-by-step fashion. Furthermore,
all data examples and input, output, and other files are available on the
companion website (http://crmda.ku.edu/guilford/geiser) to further facili-
tate the use of the techniques discussed. All syntax files are annotated,
making it easy for the user to connect the statistical model to the syntax
specification in Mplus. Many of the examples are limited to six or fewer
variables so that they can be reproduced using the free Mplus demo ver-
sion that is available for downloading at www.statmodel.com/demo.shtml.
This has the advantage of making most examples available for teaching
without requiring students to purchase the full Mplus program version.
 Preface  ix

The book starts with the question of how to prepare a data set for use
in Mplus. Given the great popularity of SPSS in the social sciences, I chose
to demonstrate the key steps in data preparation and data export using
this general statistics program. The basic principles can be easily trans-
ferred to other statistics programs, such as R, SAS, STATA, STATISTICA,
and Systat.
The book is meant to be more than a pure Mplus syntax guide, and
it is not a replacement for the Mplus User’s Guide (Muthén & Muthén,
1998–2012; www.statmodel.com/ugexcerpts.shtml). My goal is not only to
show the reader how to specify different types of models using Mplus
syntax, but also to provide guidelines for useful modeling strategies and
the interpretation of the results. Furthermore, the book also points out
some typical caveats in the analysis of complex multivariate data (e.g., the
question of how to assess measurement invariance in longitudinal SEMs
or how to avoid local likelihood maxima in LCA) and how to properly
address them. Specific issues and solution strategies are discussed in vari-
ous text boxes throughout.
Despite the focus on application, the theory of each of the statisti-
cal approaches is briefly reviewed at the beginning of each chapter. It is
important to note that the book is not a replacement for relevant statis-
tics textbooks or for specialized literature. Therefore, key references to
relevant literature are provided at the beginning of each chapter and for
the discussion of specific problems, such as the question of how to most
appropriately test the mediated effect in path analysis (see Section 3.5 in
Chapter 3).
The choice of statistical methods was made on the basis of my experi-
ences as a lecturer for numerous statistical workshops, focusing on what
seemed to be the needs of a large number of empirical scientists. I hope
that the selection of methods and models presented in this book will be
useful for researchers who want to use these techniques in their research.
The use of more complex statistical methods is discussed in Geiser,
Crayen, and Enders (2012).
The present book evolved from my seminar and workshop teaching
experiences. I am grateful to my students and to the participants in my
workshops for the feedback provided in these courses. In addition, I would
like to thank Martin Corth, David Cole, Mark Roosa, and the Deutsches
Zentrum für Altersfragen (German Center for Gerontology) for letting me
use data sets as examples in the book. I also want to thank Christopher
Marx for his help in preparing and analyzing the data set used to illus-
trate path modeling and mediation analysis in Section 3.5 as well as for
x  Preface

his helpful comments on the draft of Section 3.5. Furthermore, I thank

­Claudia Crayen for creating some of the figures and for helpful correc-
tions, as well as Henriette Hunold, Anne Janssen, Tanja Kutscher, Maike
Luhmann, Natalie Mallach, and Tabea Reuter for their careful proofread-
ing of the German draft and for their useful comments. I am grateful to
C. Deborah Laughton and Todd D. Little for their faith in this project and
for their great support, and Whitney G. Moore at the University of Kansas
for her substantive and extremely helpful review of the manuscript. All of
these individuals have contributed significantly to making this book more
comprehensible and user-­friendly and to eliminating errors. All remain-
ing errors, omissions, and weaknesses are my sole responsibility. Please
contact me at [email protected] if you find errors or have sug-
gestions for improvement. I wish all readers the best of luck with their
own analyses using Mplus and for successful publication of the results.

Christian Geiser
Logan, Utah
Contents

1 • Data Management in SPSS 1

1.1 Coding Missing Values   2
1.2 Exporting an ASCII Data File for Mplus   7

2 • Reading Data into Mplus 9

2.1 Importing and Analyzing Individual Data
(Raw Data)  10
2.1.1 Basic Structure of the Mplus Syntax
and BASIC Analysis  10
2.1.2 Mplus Output for BASIC Analysis  14
2.2 Importing and Analyzing Summary Data
(Covariance or Correlation Matrices)   21

3 • Linear Structural Equation Models 24

3.1 What Are Linear SEMs?   24
3.2 Simple Linear Regression Analysis
with Manifest Variables   28
3.3 Latent Regression Analysis   39
3.4 Confirmatory Factor Analysis   51
3.4.1 First‑Order CFA  51
3.4.2 Second‑Order CFA  58
3.5 Path Models and Mediator Analysis   62
3.5.1 Introduction and Manifest Path Analysis   62
3.5.2 Manifest Path Analysis in Mplus   65
3.5.3 Latent Path Analysis   73
3.5.4 Latent Path Analysis in Mplus   74

xi
xii  Contents

4 • Structural Equation Models 81

for Measuring Variability and Change
4.1 LS Analysis  82
4.1.1 LS versus LST Models   85
4.1.2 Analysis of LS Models in Mplus   85
4.1.3 Modeling Indicator‑Specific Effects   88
4.1.4 Testing for Measurement Invariance across Time   99
4.2 LST Analysis  116
4.3 Autoregressive Models  126
4.3.1 Manifest Autoregressive Models   127
4.3.2 Latent Autoregressive Models   133
4.4 Latent Change Models   145
4.5 Latent Growth Curve Models   163
4.5.1 First‑Order LGCMs  164
4.5.2 Second‑Order LGCMs  183

5 • Multilevel Regression Analysis 195

5.1 Introduction to Multilevel Analysis   195
5.2 Specification of Multilevel Models in Mplus   197
5.3 Option TwoLevel BASIC  198
5.4 Random Intercept Models   205
5.4.1 Null Model (Intercept‑Only Model)   205
5.4.2 One‑Way Random Effects ANCOVA   209
5.4.3 Means‑as‑Outcomes Model  214
5.5 Random Intercept and Slope Models   220
5.5.1 Random Coefficient Regression Analysis   221
5.5.2 Intercepts‑and‑Slopes‑as‑Outcomes Model  224

6 • Latent Class Analysis 232

6.1 Introduction to Latent Class Analysis   232
6.2 Specification of LCA Models in Mplus   235
6.3 Model Fit Assessment and Model Comparisons   257
6.3.1 Absolute Model Fit   258
6.3.2 Relative Model Fit   263
6.3.3 Interpretability  268

Appendix A. Summary of Key Mplus Commands 271

Discussed in This Book

Appendix B. Common Mistakes in the Mplus 279

Input Setup and Troubleshooting

Appendix C. Further Readings 283

 Contents  xiii

References 285

Author Index 293

Subject Index 296

About the Author 305

The companion website http://crmda.ku.edu/guilford/geiser provides

all data sets and Mplus input and output files, as well as additional
materials discussed in this book.
1

Data Management in SPSS

In this chapter, I explain how to prepare an SPSS raw data set for use in
Mplus. The procedure is demonstrated in the program SPSS, because it
is one of the most widely used general statistics software packages in the
social and behavioral sciences. The essential steps shown here can be
transferred easily to other common programs, such as R, SAS, STATA,
STATISTICA, SYSTAT, etc.
Mplus cannot directly read data sets in the common SPSS data for-
mat (*.sav). However, it is easy to export a data set as a simple text file (a
file with the typical file extensions *.txt or *.dat) from SPSS or another
general data management program. Such a text file can then be used for
statistical analyses in Mplus. In my experience, the text file format that
works best is a simple tab-­delimited format. Tab-­delimited means that each
column (typically representing a variable) in the text file is separated by
a tab symbol. Reading data in other formats is possible (see Mplus User’s
Guide; Muthén & Muthén, 1998–2012), and one may also use summary
data (e.g., covariance matrices, correlation matrices, means, and standard
deviations) instead of raw data, as described in Section 2.2.
In the following sections, a convenient procedure for preparing and
exporting a raw data set from SPSS is demonstrated, step by step, using
the sample data file KFT.sav. It should be noted that the method described
here is not the only possible way; other procedures yield the same results.
The sample data file as well as an SPSS syntax file with the relevant com-
mands can be found on the companion website.
1
2   Data Analysis with Mplus

BOX 1.1. Sample Data Set KFT.sav

The data set KFT.sav contains data from N = 455 German high school
students who were tested with a German intelligence test battery, Kogni-
tiver Fähigkeitstest (Cognitive Abilities Test, KFT; Heller, Gaedicke, & Wein-
läder, 1976). The KFT is frequently used to measure school achievement in
German-­speaking countries. Subtests in the present data set included verbal
material (variables kft_v1 and kft_v2), numerical (“quantitative”) material
(variables kft_q1 and kft_q3), and figural (“nonverbal”) material (variables
kft_n1 and kft_n3). The data set contains the sum scores of the students for
each of the six KFT subtests.

1.1 Coding Missing Values

It is recommended to assign a numeric missing value code for all miss-

ing values prior to importing data into Mplus, including SPSS “system-­
defined” missing values (i.e., empty cells with a dot in the SPSS data view).
Saving SPSS system-­defined missing values into the *.dat file can cause
errors when reading the data file in Mplus. Therefore, system-­defined
missing values should be recoded into user-­defined missing values using
a numeric missing value code. It is recommended that a numeric missing
value code be used—that is, a number such as “–9,” “–99,” or “999”—to
clearly identify any missing values in the data set. Of course, the miss-
ing value code needs to be chosen in such a way that the number cannot
occur as a valid score on a variable. In our example, the number –99 is
used as missing value code for all six variables. A score of –99 cannot
occur as a real value for any of the six KFT subscales.
It is important to note that one should refrain from deleting missing
values in the raw data set. Mplus has excellent capabilities for dealing
with missing values (e.g., full information maximum likelihood [FIML]
and multiple imputation; e.g., see Enders, 2010; Geiser et al., 2012). In
addition, listwise deletion is also an option available in Mplus; hence,
there is no need to delete missing values before analyzing the data in
Mplus. (Listwise deletion means that only cases with complete data on all
variables are included in the analysis.)
System- and user-­defined missing values in SPSS can be recoded
easily into a unitary (user-­defined) missing value code using the point-
and-click option Transform → Recode into same variables. This is
shown for the KFT data example in Figures 1.1 through 1.4. The use
FIGURE 1.1. SPSS option Transform → Recode into Same Variables.

FIGURE 1.2. To assign (or change) a numeric missing value code, all variables are
transferred to the window on the right-hand side.
4   Data Analysis with Mplus

of the corresponding SPSS syntax command for coding missing values

(RECODE) is shown on the companion website.
Caution: After assigning a single user-­defined missing value code to
both system- and user-­defined missing values, SPSS needs to be informed
that the new code indicates missing values. (Unfortunately, SPSS does
not recognize this indication by default.) Otherwise, SPSS would treat the
newly coded values as if they were actual scores (i.e., –99 as a valid value)
leading to incorrect statistical calculations. The procedure for defining
missing values explicitly as such is illustrated in Figures 1.5 through 1.8.
Use of the corresponding syntax command (MISSING VALUES) for the
same goal is illustrated on the companion website.
In order to tell SPSS that –99 now indicates a missing value for all
variables, one can either use the syntax command MISSING VALUES, as
illustrated on the companion website, or use the following point-and-click
options in SPSS: In the variable view, go to the first variable, then Missing
Values, and click on the three little dots on the right of the cell, as shown
in Figure 1.6.

FIGURE 1.3. The option Old and New Values is used to recode system- and user-­
defined missing values into the same numeric code, here –99.
 Missing values now
coded as –99

FIGURE 1.4. Modified data file with new missing value code –99.

FIGURE 1.5. SPSS does not automatically recognize the new missing value code
as “missing.” This code has to be manually assigned under “Missing” in the SPSS
variable view.
FIGURE 1.6. Entering the value –99 as a user-­defined missing value in the SPSS
variable view.

FIGURE 1.7. Using the copy-and-paste function to assign the value –99 as missing
value to each of the six variables in the data set.

FIGURE 1.8. The SPSS variable view with correctly assigned missing value code
for all six variables.
 Data Management in SPSS   7

1.2 Exporting an ASCII Data File for Mplus

After defining missing values appropriately, one can export the data file
into an ASCII format that is acceptable for Mplus. When using the point-
and-click interface in SPSS, one can simply use the option File → Save as
(as shown in Figures 1.9–1.10). Alternatively, the syntax command SAVE
TRANSLATE (see companion website) can be used for the same purpose.
The resulting tab-­delimited text data file that can be processed by Mplus
is shown in Figure 1.11.

FIGURE 1.9. Exporting an SPSS data file to ASCII format (*.dat).

 Uncheck this box

FIGURE 1.10. Deactivating the SPSS option Write Variable Names to Spread-
sheet.

FIGURE 1.11. Excerpt from the newly created text file that contains the variables
in tab-­delimited format. This file can be processed by Mplus.
2

Reading Data into Mplus

Mplus can process data in different formats. The two most relevant for-
mats are discussed here: individual data and summary data. The indi-
vidual data format is probably the most commonly used format in prac-
tice. Individual data are raw data in which the scores of all individuals
on all variables are preserved. This is the case, for example, in our KFT
data file (see Figure 1.11). The summary data format is used when one
wants to analyze data in summary format, for example, a covariance or a
correlation matrix (and potentially the means and standard deviations of
the variables). The possibility to analyze summary data can be useful, for
example, when a researcher wants to reanalyze data reported in a research
report (e.g., a journal article, grant proposal). Many scholarly publications
provide the covariance or correlation matrices that served as the basis for
the analyses (e.g., path or structural equation models [SEMs]), whereas
raw data are only rarely made available in publications. For many types of
structural equation analyses, it is not necessary to analyze individual data
(although using individual data may often be advantageous, e.g., when
the user wants to take missing data into account or use specific estima-
tion procedures, e.g., for clustered data). For many models, the covariance
matrix (sometimes in combination with the mean vector) of the variables
is sufficient as input for estimating the model. Let us first consider the
case of reading individual data, using the data example from Chapter 1.
9
10   Data Analysis with Mplus

2.1 Importing and Analyzing Individual Data

(Raw Data)

Importing individual data into Mplus is usually unproblematic if the

procedure described in Chapter 1 is followed. Nevertheless, it is recom-
mended to check the proper import of the data into Mplus by first per-
forming only basic statistical analyses (e.g., running descriptive statistics
for all variables in the data set) and making sure that these statistics match
the corresponding statistics calculated in at least one other program (e.g.,
SPSS). This approach can be used to check whether Mplus is reading the
data correctly. This point cannot be emphasized enough. In practice,
users often skip this step and then realize too late that Mplus has not pro-
cessed the data correctly—­leading to incorrect model estimates. For this
reason users should take the time to first run a so-­called basic analysis
in Mplus to ensure the correct processing of the data before carrying out
actual analyses in Mplus.

2.1.1 Basic Structure of the Mplus Syntax

and BASIC Analysis
Mplus is an almost entirely syntax-­based program. This means that the
estimation of statistical models and many of the other functions in Mplus
are executed via syntax commands and are not available through a point-
and-click interface as, for example, in SPSS. Nonetheless, for the construc-
tion of the basic Mplus syntax, a point-and-click interface (the so-­called
Mplus language generator) is available. This option allows users to generate
the most important basic syntax commands. Here I do not discuss the use
of the Mplus language generator in detail, because users usually do not
refer to this option any more after having learned the basic syntax rules in
Mplus—which, fortunately, are not very difficult.
In the following material I demonstrate a useful strategy for reading
data into Mplus and to check the correct processing of the data using the
Mplus basic option. For this purpose we again refer to the sample data
set KFT.dat. In the first step, the Mplus editor has to be opened (in MS
Windows: Start → Programs → Mplus → Mplus Editor). An empty
window becomes visible, into which one can either type the required
syntax commands manually or use the Mplus language generator to get
started. The syntax commands needed to run a basic analysis in Mplus
on the KFT variables are shown in Figure 2.1.
 Reading Data into Mplus   11

FIGURE 2.1. Mplus syntax file for checking the correct data import using the
option type = basic. The menu option File → Save as is used here to save the
input as a *.inp file. It is most convenient to save the input file to the same directory
that also contains the data file to be analyzed. That way, no specific path has to be
included under data: file =.

A new syntax file should immediately be saved as an Mplus input file

(see Figure 2.1). In addition, when writing or editing a syntax file, one
should not forget to save changes regularly, by either using the save sym-
bol in the menu or hitting the combination CTRL + S so that work is not
lost if the computer crashes, etc. Mplus input files contain the file ending
*.inp. Unlike, for example, SPSS syntax, Mplus requires that a separate
input file be specified for each statistical model.
The title command is used to label the analysis that is performed
and to provide explanations on what statistical analysis is carried out (and
12   Data Analysis with Mplus

potentially in which way the specific model or analysis is different from

others in a series of analyses). Although this command is not required, it
is recommended to include a meaningful and informative title section so
as to clearly document what has been done. An additional useful option
is the use of comments. Each comment has to begin with an exclamation
mark (!). No specific ending of the line is required, but every new line of
comment has to begin with an exclamation mark again. Comments will
appear in green font. Lines of comment that exceed 80 characters (in older
versions of Mplus) or 90 characters (in newer versions) will be truncated
at the end, so that comments may be incompletely shown in the output.
For comments (unlike actual commands) this may not be critical; how-
ever, comments would also be incomplete on printouts, etc.
The data command is used to inform Mplus about the name, type,
and location of the data file to be used. If the data set is saved in the same
folder as the Mplus input file, no specific path to the data set needs to be
specified in the Mplus input. The variable command is used to define
the names of the variables in the data set as well as to define the missing
value code (see also Section 1.1). Using the subcommand names =, the
variable names are defined. It is important to list the variable names in
the correct order in which they appear in the data set. Furthermore, it is
important to know that in Mplus, variable names cannot be longer than
eight characters. It is convenient to import variable names directly from
SPSS using the SPSS option Utilities → Variables. This option is illus-
trated in Figures 2.2 and 2.3.
Another advantage of this procedure is that the variable names will
be identical in both programs. Of course, this requires that variable names
be defined that are fewer or equal to eight characters in SPSS as well. It is
also important to check whether SPSS provides the variable names in the
correct order when using the Utilities → Variables option. Furthermore,
it is important to know that each subcommand in Mplus has to end with
a semicolon (;). Missing semicolons are probably the most common cause
of error messages in Mplus. Using the missing subcommand, we tell
Mplus how missing values are coded. In our example, we coded missing
values as –99 (cf. Section 1.1). We add the following additional subcom-
mand under variable: missing = all (–99);. Using the com-
mand analyses: type = basic; we request descriptive statistics
that we can then compare to the descriptive statistics in SPSS. By click-
ing on run (or hitting the combination CTRL + R), the basic analysis is
executed. Box 2.1 gives an overview of some basic rules of Mplus syntax.
Subsequently, we discuss the results of the BASIC analysis.
 Reading Data into Mplus   13

FIGURE 2.2. Exporting variable names from SPSS via the menu option Utilities
→ Variables in SPSS. On the left-hand side, all variable names are highlighted. By
clicking on Paste, the variable names are added to an SPSS syntax window (see
Figure 2.3).

FIGURE 2.3. Variable names in the SPSS syntax window. The names can now be
copied and pasted into an Mplus syntax under variable: names =. Note that
the correct order of the variable names in the SPSS syntax file should be checked
before transferring the names to Mplus.
14   Data Analysis with Mplus

BOX 2.1. Some Basic Mplus Syntax Rules

• In the Mplus syntax no differences are made between upper- and lower-
case letters. (In other words, Mplus is not case sensitive.)
• In addition, the order in which different commands are listed is largely
arbitrary.
• Every command line has to end with a semicolon (;).
• A single command line should not exceed 90 characters (in older Mplus
versions, only 80 characters were allowed per line). Longer lines can
occur, for example, when there is a long list of variables to read into Mplus
or when a long file path has to be specified. Lines that are too long will be
cut after 80/90 characters (depending on the program version), and Mplus
will ignore all following specifications in that line. Clearly, this action can
lead to very significant errors. Mplus will inform users about this problem
in the output by means of an error message. One should not overlook
or ignore this error message. A simple way to deal with the problem of
overly long lines is to simply break the lines using the ENTER key on the
keyboard.
• Variable names cannot be longer than eight characters.
• Each line of comment has to start with an exclamation mark. Comments
appear in green font in the input and output. Comments do not have to
end with a specific symbol; however, every new line of comment has to
start with an exclamation mark again.

2.1.2 Mplus Output for BASIC Analysis

After running the input file for the basic analysis, a new window auto-
matically pops up containing the output for the analysis. The correspond-
ing output file is automatically saved to the same folder that contains the
input file. The output file has the same file name as the input file but can
be distinguished from the input file through the different file ending (*.out
instead of *.inp). The output file for our BASIC analysis is shown in the
following. First of all, the commands used to specify the analysis are repro-
duced in the output. This is useful because it allows us to review the input
specifications to check if everything is set up correctly. In addition, that
way, the input specifications will be included in any printout of the results.

Mplus VERSION 6.1

MUTHEN & MUTHEN
11/10/2011 4:36 PM
 Reading Data into Mplus   15

INPUT INSTRUCTIONS

title: Read data set “KFT.dat” in Mplus

Check that the data set is read correctly in Mplus
using the TYPE = BASIC option
Here: Using the Mplus default with FIML estimation
including missing data

data: file = KFT.dat;

! This is a comment
variable: names = kft_v1 kft_v3 kft_q1 kft_q3 kft_n1 kft_n3;
missing = all(-99);

analysis: type = basic;

Next, we receive the following warning message, which is caused by

the fact that in our data set, 131 students had missing values on all six
variables:

*** WARNING
Data set contains cases with missing on all variables.
These cases were not included in the analysis.
Number of cases with missing on all variables: 131
1 WARNING(S) FOUND IN THE INPUT INSTRUCTIONS

This message appears because Mplus, by default, uses FIML estima-

tion with missing data (e.g., see Enders, 2010). This procedure cannot be
used for participants who have no valid scores on any of the variables,
because these individuals do not provide any information about the vari-
ables.
Following is the title and some technical information about the anal-
ysis with regard to, for example, the sample size, the variables used in the
analysis, and the data structure in general:

Read data set “KFT.dat” in Mplus

Check that the data set is read correctly in Mplus
using the TYPE = BASIC option
Here: Using the Mplus default with FIML estimation
including missing data

SUMMARY OF ANALYSIS

Number of groups 1
Number of observations 456

Number of dependent variables 6

Number of independent variables 0
Number of continuous latent variables 0
16   Data Analysis with Mplus

Observed dependent variables

Continuous
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1 KFT_N3

Estimator ML
Information matrix OBSERVED
Maximum number of iterations 1000
Convergence criterion 0.500D-04
Maximum number of steepest descent iterations 20
Maximum number of iterations for H1 2000
Convergence criterion for H1 0.100D-03

Input data file(s)

KFT.dat

Input data format FREE

This information is useful to check, among other things, that the cor-
rect data set was used as well as the intended variables. For Number of
observations we can see that 456 students contributed one or more
values (here, those cases that have missing data on all six variables are
already excluded). Next is a summary of the missing data patterns that
occurred in the present application:

SUMMARY OF DATA

Number of missing data patterns 2

SUMMARY OF MISSING DATA PATTERNS

MISSING DATA PATTERNS (x = not missing)

1 2
KFT_V1 x x
KFT_V3 x x
KFT_Q1 x x
KFT_Q3 x x
KFT_N1 x x
KFT_N3 x

MISSING DATA PATTERN FREQUENCIES

Pattern Frequency Pattern Frequency

1 455 2 1

Under SUMMARY OF DATA we see that two distinct missing data

patterns occurred in this example. Under SUMMARY OF MISSING
DATA PATTERNS Mplus shows what those patterns were. Each missing
data pattern is represented by a separate column (with the exception of
 Reading Data into Mplus   17

the pattern in which all values are missing, which is not shown). Column
1 (missing data pattern 1) contains only “x’s,” indicating that this is the
pattern in which no missing data occurred. The second column contains
“x’s” except for the last variable KFT_N3. This means that individuals
with pattern 2 have valid scores on all variables except for the variable
KFT_N3.
The section on MISSING DATA PATTERN FREQUENCIES shows
that of the 456 students, all but one student contributed complete data
(values on all six variables): 455 students showed missing data pattern 1
(frequency = 455), whereas only one student showed missing data pattern
2 (frequency = 1), that is, a missing value on KFT_N3.
The covariance coverage shows us the proportion of cases that con-
tributes values for the calculation of each variance or covariance. In our
case, for the variances and covariances, those are 100% of cases, with the
exception of the variance and covariances that are associated with vari-
able KFT_N3 (only 99.8% of cases contribute to this variable because
of the one student who has a missing value on this variable). The mini-
mum acceptable value for the covariance coverage is 10% according to
the Mplus default (“Minimum covariance coverage value 0.100”). If the
covariance coverage fell below this value, Mplus would no longer estimate
a model by default because the coverage would be seen as too weak.

COVARIANCE COVERAGE OF DATA

Minimum covariance coverage value 0.100

PROPORTION OF DATA PRESENT

Covariance Coverage
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1
________ ________ ________ ________ ________
KFT_V1 1.000
KFT_V3 1.000 1.000
KFT_Q1 1.000 1.000 1.000
KFT_Q3 1.000 1.000 1.000 1.000
KFT_N1 1.000 1.000 1.000 1.000 1.000
KFT_N3 0.998 0.998 0.998 0.998 0.998

Covariance Coverage
KFT_N3
________
KFT_N3 0.998

Note that the covariance coverage does not refer to the actual vari-
ances and covariances of the variables, but instead simply informs us about
the “completeness” of the data (the amount of data available to calculate
those statistics). The actual estimated sample variances and covariances
18   Data Analysis with Mplus

are obtained next, under the header RESULTS FOR BASIC ANALY-
SIS/ESTIMATED SAMPLE STATISTICS:

RESULTS FOR BASIC ANALYSIS

ESTIMATED SAMPLE STATISTICS

Means
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1 KFT_N3
________ ________ ________ ________ ________ ________
11.904 8.978 12.377 7.730 11.088 8.277

Covariances
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1
________ ________ ________ ________ ________
KFT_V1 21.592
KFT_V3 10.761 17.938
KFT_Q1 6.010 5.267 11.235
KFT_Q3 4.645 4.406 4.027 6.754
KFT_N1 11.072 11.644 6.870 6.111 29.826
KFT_N3 6.347 7.190 4.415 3.986 9.298

Covariances
KFT_N3
________
KFT_N3 11.782

Correlations
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1
________ ________ ________ ________ ________
KFT_V1 1.000
KFT_V3 0.547 1.000
KFT_Q1 0.386 0.371 1.000
KFT_Q3 0.385 0.400 0.462 1.000
KFT_N1 0.436 0.503 0.375 0.431 1.000
KFT_N3 0.398 0.495 0.384 0.447 0.496

Correlations
KFT_N3
________
KFT_N3 1.000

MAXIMUM LOG-LIKELIHOOD VALUE FOR THE UNRESTRICTED (H1) MODEL IS -7152.289

Beginning Time: 16:36:14

Ending Time: 16:36:15
Elapsed Time: 00:00:01

MUTHEN & MUTHEN

3463 Stoner Ave.
Los Angeles, CA 90066

Tel: (310) 391-9971

Fax: (310) 391-8971
Web: www.StatModel.com
Support: [email protected]

Copyright (c) 1998-2010 Muthen & Muthen

 Reading Data into Mplus   19

The descriptive statistics obtained for each of the six observed KFT
variables (i.e., the estimated means, variances, covariances, and product–­
moment correlations of the variables) were estimated using the FIML pro-
cedure. Therefore, the statistics in this form are not directly comparable
to the corresponding statistics calculated in SPSS. The reason is that SPSS
statistics are based either on pairwise or listwise deletion of cases rather
than on FIML estimation.
In order to directly compare Mplus and SPSS descriptive statistics,
one can request the use of listwise deletion in Mplus rather than the
default (FIML). Note that we discourage the use of listwise deletion in
later analysis steps that involve actual model fitting. The procedure is
used here for technical reasons only, in order to facilitate the process
of checking the proper data entry into Mplus. In later analysis steps,
researchers should consider using more advanced missing data analytic
techniques, such as FIML, as described by Enders (2010) or by Schafer
and Graham (2002), because these techniques are usually more reason-
able than listwise deletion.
Listwise deletion of cases is obtained in Mplus by adding the subcom-
mand listwise = on; under the data command. The listwise
= on; subcommand deactivates the FIML procedure and excludes all
cases that have missing data on at least one of the variables included in
the analysis. Note that FIML has been the default since Mplus version 5.
In previous versions of the program through Mplus version 4, listwise
deletion was the default.
The extended Mplus syntax using listwise deletion is shown in Fig-
ure 2.4. The resulting descriptive statistics are shown below. For compari-
son, the corresponding SPSS statistics (with listwise deletion) are shown
in Figure 2.5. The SPSS statistics were generated using the SPSS option
Analyze → Scale → Reliability Analysis → Statistics → Descriptives
for Item/Inter-Item Covariances/Correlations. The results are identical
to the Mplus results, rounded to three decimals, which shows us that the
data set KFT.dat was apparently correctly read by Mplus. In addition,
SPSS returns the same sample size (N = 455 listwise cases). After this
check, we can start the first actual model in Mplus (see Chapter 3).

SAMPLE STATISTICS [Now based on listwise deletion of cases]

Means
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1 KFT_N3
________ ________ ________ ________ ________ ________
11.899 8.974 12.387 7.732 11.112 8.281
20   Data Analysis with Mplus

Covariances
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1
________ ________ ________ ________ ________
KFT_V1 21.677
KFT_V3 10.799 18.008
KFT_Q1 6.057 5.310 11.242
KFT_Q3 4.669 4.429 4.038 6.783
KFT_N1 11.172 11.745 6.794 6.120 29.686
KFT_N3 6.383 7.230 4.417 4.001 9.294

Covariances
KFT_N3
________
KFT_N3 11.811

Correlations
KFT_V1 KFT_V3 KFT_Q1 KFT_Q3 KFT_N1
________ ________ ________ ________ ________
KFT_V1 1.000
KFT_V3 0.547 1.000
KFT_Q1 0.388 0.373 1.000
KFT_Q3 0.385 0.401 0.462 1.000
KFT_N1 0.440 0.508 0.372 0.431 1.000
KFT_N3 0.399 0.496 0.383 0.447 0.496

Correlations
KFT_N3
________
KFT_N3 1.000

FIGURE 2.4. Modified Mplus input file for the basic analysis, now with FIML
estimation turned off. This file generates descriptive statistics based on a listwise
deletion of cases that can be more easily compared to results in SPSS.
 Reading Data into Mplus   21

FIGURE 2.5. Descriptive statistics for the six KFT variables produced in SPSS
through the option Analyze → Scale → Reliability → Statistics → Descriptive
Statistics for Items/Inter-Items → Covariances/Correlations. The values match
the Mplus estimates obtained under the listwise = on; option.

2.2 Importing and Analyzing Summary Data

(Covariance or Correlation Matrices)

As already mentioned at the beginning of this chapter, it is sometimes

practical to use summary data rather than individual data for an analysis.
Summary data are, for example, covariance or correlation matrices, some-
times supplemented by the means and/or standard deviations of all vari-
ables. We now show the procedure of reading summary data into Mplus
22   Data Analysis with Mplus

using the six KFT variables as an example. The simplest way to enter
summary data is to copy and paste the covariance or correlation matrix
into a simple text file. For this purpose one can use, for example, the
simple Editor program in Windows (Start → Programs → Accessories)
or WordPad.
Figure 2.6 shows an example of a text file that contains the means
(first row), standard deviations (second row), and product moment cor-
relation matrix of the six observed KFT variables (the file is located on
the companion website and is named KFT_summary-­data.txt). The
data shown in this text file can easily be used by Mplus to estimate, for
example, an SEM. The Mplus syntax to read the data from the file KFT_
summary-­data.txt is shown in Figure 2.7. The data command is again
used to define the name of the data set. In addition, it has to be specified
what type of summary statistics are being read from this file. In our case
the subcommand type = means std corr; means that the data
set KFT_summary-­data.txt contains the means (means) followed by the
standard deviations (std) and correlations (corr) of the variables. Note
that the order of the subcommands is important. (They have to be in line
with the order in which the summary statistics appear in the data file.) To
read a covariance matrix, one would use the subcommand cova instead
of corr.
The additional subcommand nobservations = 455; is used
to specify the sample size (number of individuals on which the summary
data set is based). In this case, the data are based on N = 455 listwise

FIGURE 2.6. Text file with summary data for the six KFT variables. The first line
contains the means of the variables. The second line contains their standard devia-
tions. The matrix contains the product–­moment correlations of the variables. Figure
2.7 shows how these data can be properly processed in Mplus.
 Reading Data into Mplus   23

FIGURE 2.7. Mplus input file for reading the summary data shown in Figure 2.6.
The basic option is not available for summary data. Therefore, descriptive statistics
for data checking are requested via the output: sampstat; option. Given that
we do not explicitly specify a model for the data, by default Mplus estimates a so-­
called null model for all variables listed under variable: names =.

cases. This information has to be provided, because it is not possible for

Mplus to infer the number of observations from the summary data set
(whereas this is possible for individual data). The option variable:
names = is again used to define variable names—which also are not
given in the data file.
The BASIC option is not available for summary data in Mplus. For
this reason, descriptive statistics for data checking are now requested
using the option output: sampstat;. The abbreviation sampstat
stands for sample statistics and provides us with descriptive statistics for
the six variables (in this case, the observed means and the covariance
matrix will be provided). Given that no model is explicitly specified, by
default Mplus estimates a so-­called null model (sometimes referred to as
an independence model). The null model assumes that there are no rela-
tionships among any of the six variables. The only model parameters to
be estimated in this model are the means and variances of the observed
variables. The output for this analysis is not shown here but can be found
on the companion website.
The next chapter provides an introduction to the basics of model
specification in Mplus using linear SEMs as an example. I first consider
simple linear regression models and subsequently more complex SEMs
such as confirmatory factor analysis and latent path analysis.
3

Linear Structural Equation Models

3.1 What Are Linear SEMs?

Linear SEMs with latent variables (often called covariance structure models)
are used to model complex relationships between continuous variables at
the latent level, that is, at the levels of variables that are corrected for mea-
surement error. Linear SEMs can also be viewed as multivariate regres-
sion models. In these models, structural relationships between multiple
dependent variables and multiple independent variables can be analyzed
simultaneously. Variables that serve exclusively as independent variables
are referred to as exogenous variables, because variability in these variables
is not explained by other variables in the model. For exogenous variables,
only nondirectional relationships (e.g., covariances or correlations) are
analyzed.
Variables that are regressed on one or more other variables in the
model are called endogenous variables. Endogenous variables that serve as
independent and dependent variables at the same time are called inter-
vening variables or mediator variables, because those variables mediate the
relationship between other variables. Mediated effects are frequently of
interest and often examined using manifest or latent path analysis, as
discussed in detail in Section 3.5.
Linear SEMs with latent variables can be described as consisting of
two parts (see Figure 3.1): a measurement model and a structural (or latent
variable) model. The parameters pertaining to both parts are estimated

24
 Linear Structural Equation Models   25

ζ2 Structural model

β21
η1 η2

λ11 λ21 λ32 λ42

Y1 Y2 Y3 Y4
ε1 ε2 ε3 ε4

Measurement
model

FIGURE 3.1. Path diagram of a simple structural equation model with two latent
variables (h1 and h2), each of which is measured by two indicators (Y1 and Y2 are
indicators of h1, whereas Y3 and Y4 are indicators of h2). The parameters l11–l42
denote factor loadings. The variables e1–e4 are residual (error) variables. In the struc-
tural model, the latent regression of the factor h2 on h1 is estimated. In this latent
regression, b21 denotes the regression slope coefficient and z2 denotes a latent resid-
ual variable.

simultaneously. In the measurement model, the researcher specifies how

the latent variables (or factors; here indicated as hj) are measured using
observed (manifest) variables (Yi) as indicators. It is often assumed that
the latent variables are responsible for the covariance between different
indicators of the same construct (so-­called reflective measurement model).
This means that variation in the latent variable is assumed to cause varia-
tion in the indicators and to explain relationships between different indi-
cators.
The manifest and latent variables are connected via linear regres-
sions in the measurement model. In reflective measurement models the
indicators serve as dependent variables, whereas the latent variables are
independent variables (in so-­called formative measurement models this is
reversed; however, formative measurement models are not discussed in
this book). The regression coefficients lij of these regressions are often
referred to as factor loadings. In the structural (latent variable) model,
relationships between the latent variables are specified, either through
covariances, latent regressions, or path analyses. If the structural model
consists only of covariances between latent variables (i.e., nondirectional
relationships), the analysis is often referred to as a confirmatory factor anal-
ysis (CFA; see Section 3.4). If the structural model includes directional
26   Data Analysis with Mplus

paths (regressions) between latent variables, the model could be referred

to as a latent regression analysis (see Figure 3.1 as well as Section 3.3 for
an example). If multiple latent endogenous variables are considered, the
model is often referred to as a latent path analysis (see Section 3.5).
An important advantage of linear SEMs with latent variables is that
the use of latent variables allows researchers to explicitly account for
measurement error in the analysis. As a consequence, the relationships
between variables in the structural model can be more accurately esti-
mated compared to conventional correlation, regression, or path analyses
at the level of manifest variables. In addition, SEMs allow researchers to
test complex relationships between variables through formal model tests.
It is also possible to compare competing models statistically against each
other.
SEMs are very flexible and can be used to analyze a variety of com-
plex research questions. For example, SEMs can be used to analyze vari-
ability and change in longitudinal studies, to separate stable from time-­
specific influences on psychological measurements (see Section 4.2 as
well as Steyer, Ferring, & Schmitt, 1992; Steyer, Schmitt, & Eid, 1999),
to test for measurement variance (Section 4.1.4), and to analyze latent
change (Sections 4.3–4.5) and latent growth curves (Section 4.5; see also
Bollen & Curran, 2006; Duncan, Duncan, & Strycker, 2006; Steyer, Eid,
& Schwenkmezger, 1997). In addition, SEMs are often used to analyze
complex data observed from multiple sources of information (so-­called
multimethod data; see Eid et al., 2008; Eid, Lischetzke, & Nussbeck, 2006;
Geiser, 2009).
In order to graphically illustrate SEMs, path diagrams like the one
shown in Figure 3.1 are often used. In this book, I frequently use path
diagrams (instead of mathematical equations) to illustrate the models ana-
lyzed in Mplus. In path diagrams, observed variables (manifest variables,
indicators) are typically shown in boxes, whereas not directly observed
variables (latent variables, latent factors) are typically shown in circles or
ellipses. Directional relationships (regressions) are symbolized through
single-­headed arrows (→). The variable that emits the arrow is the inde-
pendent variable (predictor, regressor) with respect to the variable to
which the arrow points. All variables that “receive” an arrow are consid-
ered endogenous (dependent) variables. Variables that only emit paths but
do not receive unidirectional paths are referred to as exogenous variables.
Straight or curved double-­headed arrows (↔) indicate nondirectional
relationships (i.e., covariances or correlations). In this book, manifest
 Linear Structural Equation Models   27

variables are indicated by the letter Y or with the actual variable name.
Latent variables are indicated by h or receive a substantive label that char-
acterizes their meaning. Error variables (residual variables) associated
with manifest variables are indicated by e, and latent variable residuals
are indicated by z.
Useful introductory textbooks on the theory of linear SEMs are, for
example, books by Bollen (1989), Kaplan (2009), Kline (2011), Loeh-
lin (1998), Raykov and Marcoulides (2006), as well as Schumacker and
Lomax (1996).

BOX 3.1. Overview of the Most Important Parameters That Are

Estimated in SEMs with Latent Variables and Common Notation

Parameters of the measurement model:

• Factor loading (regression slope coefficients) lij of the ith observed vari-
able Yi on the jth latent variables hj
• Intercept ai of the observed variable Yi
• Residual (error) variance Var(ei) of the observed variable Yi
• Residual covariance Cov(ei, ei′) or correlation Corr(ei, ei′)

Parameters of the structural model (model at the level of the hj variables/

factors):

• Factor variance Var(hj) (for exogenous factors)

• Factor mean E(hj) (for exogenous factors)
• Factor covariances Cov(hj, hj′) or correlations Corr(hj, hj′) (for exogenous
factors)
• Latent path coefficients/regression coefficients bjj′ for the regression of an
endogenous factor hj on another factor hj′. In a higher-­order factor model,
we denote the factor loading of a lower-order factor hj on a higher-­order
factor with gj.
• Latent intercepts for endogenous factors hj are denoted as b0j in regression
or path analyses and as g0j in higher-­order factor analyses.
• Latent residual variances Var(zj) (for endogenous factors)
• Latent residual covariances Cov(zj, zj′) or correlations Corr(zj, zj′) (for
endogenous factors)

Correlations are also denoted by r in the present book.

28   Data Analysis with Mplus

3.2 Simple Linear Regression Analysis

with Manifest Variables

We introduce the basic principles of the Mplus model specification based

on linear regression analysis with manifest variables, which can be seen as
a special case of more complex SEMs. Our first model is a simple (bivari-
ate) linear regression model with one manifest independent variable (IV)
and one manifest dependent variable (DV).
Formally, the simple linear regression model can be written as fol-
lows (e.g., see Cohen, Cohen, West, & Aiken, 2003):

Y = b0 + b1X + e

where b0 indicates the intercept (the expected value of Y when X is zero),

b1 indicates the slope (the expected increase in Y for a 1-unit increase in X)
for the regression of the dependent variable (criterion Y) on the indepen-
dent variable (predictor variable X), and e indicates a residual variable [the
difference between Y and the expected value of Y given X: e = Y – E(Y|X)].
A simple way to specify and run the first model on a newly imported
data set in Mplus is to use the Mplus input file for the BASIC analysis as
a template by first saving it under a new name (so as not to overwrite the
original BASIC analysis). The new name could be, for example, 1_sim-
ple_regression.inp (see Figure 3.3). In the next step, one can change the
syntax commands in the newly saved file in order to estimate the model
shown in Figure 3.2.
For simplicity, we use listwise deletion in the following when speci-
fying models in Mplus rather than the default FIML procedure. As noted
previously, in actual analyses, users should consider using FIML when
missing data are present (e.g., see Enders, 2010; Geiser et al., 2012; as well
as Schafer & Graham, 2002).

BOX 3.2. Example of Simple Linear Regression

As a concrete example imagine that a researcher is interested in predicting

numerical ability from figural ability. For simplicity we use only the variables
kft_q1 (as a measure of quantitative ability; this represents the dependent
variable = Y in our model) and kft_n1 (as a measure of figural ability; this is
the independent variable = X). The corresponding simple linear regression
model is depicted as a path diagram in Figure 3.2. This path diagram already
contains the estimated Mplus model parameters in unstandardized form.
 Linear Structural Equation Models   29

(9.65)

KFT_N1 0.23 KFT_Q1

(29.83) (11.24)

FIGURE 3.2. Path diagram of the estimated bivariate regression model for pre-
dicting the observed variable kft_q1 (quantitative ability) from kft_n1 (nonverbal/
figural ability). The numbers indicate the estimated unstandardized regression coef-
ficient (0.23) as well as the estimated variances of the variables (29.83 and 11.24) and
the estimated residual variance for kft_q1 (9.65).

FIGURE 3.3. Saving the Mplus input file from Figure 2.1 under a new name to
generate a new input file for the bivariate regression model.
30   Data Analysis with Mplus

BOX 3.3. Tips for the First Model Specification

After checking the correct data import into Mplus using the BASIC option
(see Chapter 2), one can begin with the specification of the first actual model.
In general, it is useful to start with a simple model that may represent one
part of a more complex model with many variables. This approach has the
advantage that possible errors or technical problems that may arise in a com-
plex model are easier to detect in a small submodel than in a larger model.
For example, if one wants to analyze a complex model with many variables,
it is useful to build up the model in a step-by-step fashion (Mulaik & Millsap,
2000). This means starting with only some of the manifest variables and/or
latent factors instead of including all of the variables at once. Then one can
slowly increase the complexity of the model. In this way, possible errors and
problems can be more easily traced back to a specific part of the model or
constellation of variables. As a consequence, troubleshooting will be easier
than for a complex model that already contains many variables and param-
eters. It can also be useful to first estimate the measurement model before
placing restrictions on the structural part of the model (Mulaik & Millsap,
2000).

The Mplus specification for the bivariate linear regression model

is shown in Figure 3.4. We can see that a new subcommand has been
added to the variable command. The relevant command is the so-­called
­usevar subcommand, which allows us to specify which variables in the
data set will actually be used in the model. In our simple manifest regres-
sion model we only use the variables kft_q1 and kft_n1. Therefore,
the usevar command is specified as follows in our example:

usevar = kft_q1 kft_n1;

It is important to note that only variables that are actually used in the
model should be listed under usevar, because Mplus assumes that all
variables listed under usevar are part of the model. If variables listed
under usevar are not included in the actual model specification, Mplus
will automatically assume that those variables are uncorrelated with all
other variables in the model (see discussion below as well as Appendix B),
resulting in the following error message:

*** WARNING in Model command

Variable is uncorrelated with all other variables: KFT_N3
 Linear Structural Equation Models   31

*** WARNING in Model command

All least one variable is uncorrelated with all other
variables in the model.
Check that this is what is intended.
2 WARNING(S) FOUND IN THE INPUT INSTRUCTIONS

In this example, the variable kft_n3 is erroneously listed under

usevar, but is not used in the actual model specification. Similar prob-
lems occur when a usevar command is entirely omitted, in which case
Mplus assumes that all variables listed under names = are part of the
model. Whenever a message like the above one appears in the output, the
user should go back to the input file and add or correct the usevar sub-
command and then reestimate the model based on the corrected input.
Failure to do so can result in serious model misspecification and biased
parameter estimates.
Note that in the analysis command the type of the analysis changes
from type = basic to a type = general because Mplus does not
estimate an actual model under the BASIC option. The BASIC option
provides only descriptive statistics, but no actual model parameters can

FIGURE 3.4. Mplus input file for the bivariate linear regression model from Figure
3.2.
32   Data Analysis with Mplus

be obtained. Two new commands have been added to this input file. The
first one is the model command, which is used to specify the desired
model. The second one is the output command that is used to request
additional output that is not part of the Mplus default output:

model: kft_q1 on kft_n1;

output: sampstat stdyx;

The model command specifies the simple regression analysis in

which kft_q1 is regressed on kft_n1. The key word on thus stands
for “regressed on.” Note that different types of regression analyses (e.g.,
logistic regression) could be requested by specifying a dependent variable
as categorical (using the subcommand categorical =) and requesting
a specific method of estimation (see Geiser et al., 2012, as well as Muthén
& Muthén, 1998–2012).
In the output command, we request the output of descriptive sta-
tistics (the means, covariances, and correlations) for the manifest vari-
ables (keyword sampstat for “sample statistics”) as well as the com-
pletely standardized solution (keyword stdyx), in addition to the Mplus
default output. In the following section, the most important part of the
output for the simple regression model is shown. First of all, it is impor-
tant for the user to check whether the following message appears in the
output:

INPUT READING TERMINATED NORMALLY

This message means that Mplus has not found any kind of fatal syn-
tax errors that prevented the program from running the input file. For
example, this message indicates that Mplus had no difficulties finding
and reading the data set for the analysis. It is important to note that this
message is not a guarantee that Mplus read the data correctly or that the
model was correctly specified. It is still important to check the accuracy
of the calculated sample size and the descriptive statistics, as well as the
plausibility of the estimated model parameter estimates, even when the
user has already carried out a BASIC analysis (see Section 2.1.1). In addi-
tion, this message is not a guarantee that the actual estimated statistical
model converged or that all parameter estimates are proper and mean-
ingful. Problems related to the actual estimation of the specific statistical
model would appear further below in the Mplus output (below the sample
statistics).
 Linear Structural Equation Models   33

If the above message (INPUT READING TERMINATED NOR-

MALLY) does not appear in the output, error messages will appear in its
place. These messages will point to errors in the input (e.g., model) speci-
fication and/or to errors that occurred during attempts by the program to
locate and read the data set. Appendix B provides an overview of typical
problems with the input specification and related troubleshooting.
Under SUMMARY OF ANALYSIS Mplus reports technical details
about the analysis (number of groups—­here just one because it is not a
multigroup analysis, number of observations, number and name of the
dependent and independent manifest and latent variables in the model,
name and format of the data file).

SUMMARY OF ANALYSIS

Number of groups 1
Number of observations 456

Number of dependent variables 1

Number of independent variables 1
Number of continuous latent variables 0

Observed dependent variables

Continuous
KFT_Q1

Observed independent variables

KFT_N1

Estimator ML
Information matrix OBSERVED
Maximum number of iterations 1000
Convergence criterion 0.500D-04
Maximum number of steepest descent iterations 20

Input data file(s)

KFT.dat

Input data format FREE

This information is useful to check whether the analysis has been

done correctly, as intended by the user. In the following box, descriptive
statistics (called sample statistics in Mplus) for the variables used in the
model are shown. These values should once again be thoroughly checked
for accuracy, in case anything went wrong while reading the data into
34   Data Analysis with Mplus

Mplus. In addition, these statistics are often useful for the interpretation
of the model results. In our case, we can see that the correlation between
kft_q1 and kft_n1 is moderately large (r = .375) as shown in the sub-
section Correlations.

SAMPLE STATISTICS

SAMPLE STATISTICS

Means
KFT_Q1 KFT_N1
________ ________
1 12.377 11.088

Covariances
KFT_Q1 KFT_N1
________ ________
KFT_Q1 11.235
KFT_N1 6.870 29.826

Correlations
KFT_Q1 KFT_N1
________ ________
KFT_Q1 1.000
KFT_N1 0.375 1.000

Following the descriptive statistics, the user should check once again
for possible error messages. As mentioned above, Mplus would output
additional error messages at this point if there were any kinds of specific
problems related to the actual estimation of the model. Possible reasons
for such types of error messages include (1) the model is not identified, (2)
inadmissible estimation results are encountered (e.g., negative variance or
residual variance estimates or correlations > | 1 |), or (3) when the estima-
tion procedure failed to converge on a final optimal set of parameter esti-
mates. In our example, no error messages appear. The model estimation
terminated without problems, as is expected for a simple linear regression
model with manifest variables. Therefore, we just obtain the message

THE MODEL ESTIMATION TERMINATED NORMALLY

It is important to note than the user should be cautious whenever

additional messages appear at this point in the output. In this case, all fol-
lowing results (including the fit indices and model parameter estimates)
 Linear Structural Equation Models   35

should be interpreted only with great caution or not at all, because severe
estimation problems may have occurred.
Subsequently, a number of statistical measures are reported that can
be used to evaluate the overall goodness of fit of the model under MODEL
FIT INFORMATION (formerly referred to as TESTS OF MODEL FIT
in older versions of Mplus). A description of these indices can be found
in Box 3.7 on page 45 (for more detailed discussions, see Bollen & Long,
1993, or Schermelleh-­Engel, Moosbrugger, & Müller, 2003). In our case
we are analyzing just a simple linear regression model with only mani-
fest variables. The simple regression model uses the entire information
in the data to estimate model parameters. Such a model is saturated so
that most of the global indices of model fit are not of great interest to us
here. Saturated models always show a perfect fit with the observed data.
Only the chi-­square test for the baseline model (the so-­called independence
model) is relevant. The independence model assumes that all variables
used in the model are uncorrelated. In our case, this means that kft_q1
and kft_n1 are assumed to be uncorrelated. The chi-­square test for the
independence model is significant in our example, c2 = 69.235, df = 1, p
< .0001, which indicates that the independence model has to be rejected.
This shows us that the correlation between kft_q1 and kft_n1 is
significantly different from zero and that the estimation of a regression
model is this meaningful.

MODEL FIT INFORMATION

Number of Free Parameters 3

Loglikelihood

H0 Value -1163.957
H1 Value -1163.957

Information Criteria

Akaike (AIC) 2333.913

Bayesian (BIC) 2346.281
Sample-Size Adjusted BIC 2336.760
(n* = (n + 2) / 24)

Chi-Square Test of Model Fit

Value 0.000
Degrees of Freedom 0
P-Value 0.0000
36   Data Analysis with Mplus

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.000
90 Percent C.I. 0.000 0.000
Probability RMSEA <= .05 0.000

CFI/TLI

CFI 1.000
TLI 1.000

Chi-Square Test of Model Fit for the Baseline Model

Value 69.235
Degrees of Freedom 1
P-Value 0.0000

SRMR (Standardized Root Mean Square Residual)

Value 0.000

Following are the parameter estimates for the model, first of all in
unstandardized form (under MODEL RESULTS).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_Q1 ON
KFT_N1 0.230 0.027 8.647 0.000

Intercepts
KFT_Q1 9.823 0.329 29.833 0.000

Residual Variances
KFT_Q1 9.652 0.639 15.100 0.000

The first number in the column Estimate (0.23) is the estimated

unstandardized regression slope coefficient bˆ 1 for the regression of kft_
q1 on kft_n1. In the second column, the estimated standard error (SE)
for bˆ 1 is given, which is estimated to be σˆ b1 = 0.027. The ratio estimate/
SE (given in the third column) yields a test statistic that is equivalent to a
z-score in large samples and that can be used for significance testing. In
our example, the resulting z-score is 8.647. The fourth column (two-­tailed
p-value) gives the probability of the observed estimate of bˆ 1 under the null
hypothesis that the coefficient is zero in the population for a two-­tailed
 Linear Structural Equation Models   37

test. Here the p-value is smaller than .0001, which tells us that the slope
coefficient is significantly different from zero at an a-level of .0001. This
means that kft_n1 significantly predicts kft_q1 (or equivalently, that
the two variables are significantly correlated).
The next row gives the estimated intercept parameter, which is bˆ 0 =
9.823. The intercept is also statistically significant (z = 29.833). Using the
estimates for bˆ 1 and bˆ 0 we can formulate the regression equation for our
example:

kft_q1 = 9.823 + 0.23 kft_n1 + e

This means that for individuals with a kft_n1 score of zero (none
of the items solved correctly), a score of 9.823 on kft_q1 is predicted.
Furthermore, according to this equation, the expected increase in the
kft_q1 score for a one-unit increase in the kft_n1 score is 0.23 points.
The third row of the MODEL RESULTS table (p. 36) provides us with
the unstandardized residual variance (estimated variance of the error
variable e). Here the error of variance equals σˆ 2e = 9.652. The value of the
error variance becomes meaningful when we relate it to the total variance
of the dependent variable (σˆ 2Y ). The total variance of the dependent vari-
able can be found under SAMPLE STATISTICS in the upper part of the
output (in the main diagonal of the covariance matrix). The estimated
variance of kft_q1 is 11.235 in our example. This shows us that the
residual variance (i.e., that part of the variance that is not explained in the
regression analysis) is relatively large compared to the total variance of the
dependent variable. In other words, only a small portion of individual dif-
ferences in kft_q1 scores can be explained by individual differences in
the kft_n1 scores. By calculating the so-­called coefficient of determination
(R 2), we can see that only about 14.1% of the variability in kft_q1 scores
can be explained by this simple regression model:

ˆ2
ˆR 2 = 1 − σ e = 1 − 9.652 = .141
σˆ Y2 11.235

Below the unstandardized model parameter estimates, we obtain

the so-­called STANDARDIZED MODEL RESULTS in Mplus (here: the
fully standardized or STDYX solution). This solution contains the fully
standardized parameter estimates as well as the estimated R 2 value that
we have just calculated by hand, using the ratio of the unstandardized
residual variance to the total variance.
38   Data Analysis with Mplus

BOX 3.4. Centering Independent Variables

The intercept term b0 in the linear regression analysis can be meaning-

fully interpreted only when all independent variables in the equation have
a meaningful zero point. Many social science variables, however, do not
have a meaningful zero point, because, for instance, zero age, zero intel-
ligence, zero extraversion, or zero success in school are often not expected
for any individual, and/or psychological scales to measure these constructs
are designed in a way that scores of zero are not possible or meaningful. One
way to make zero a meaningful score for such variables is to center them.
Centering refers to the subtraction of some type of meaningful constant from
each raw score. Often the arithmetic average (mean) is used for this purpose.
Centering scores at the mean causes the zero point of a variable to shift to
the observed mean of the variable. This results in a meaningful zero point
(even if the original zero point was not meaningful), because the value of
zero after centering refers to the typical (average) case in the sample (Cohen
et al., 2003).
The intercept term for centered variables therefore represents the
expected value on the criterion variable for individuals with an average value
on the predictor variable, which is typically a meaningful score. Centering
plays an especially important role in moderated regression (i.e., regression
analysis with interaction terms; see Aiken & West, 1991) as well as in multi-
level regression analysis (see Chapter 5).
In Mplus, centering is easily done through adding the subcommand
centering to the variable command. In our case we add the follow-
ing line to the Mplus input to center the predictor variable kft_n1 prior to
running the regression analysis (the full input/­output files with the centered
independent variable can be found on the companion website):

variable: centering = grandmean(kft_n1);

The centering subcommand causes Mplus to center the predictor vari-

able kft_n1 before running the regression analysis. In this case Mplus uses
so-­called grand-mean centering, which means that the overall mean of kft_
n1 is subtracted from each individual kft_n1 score. After centering, our
regression equation reads as

kft_q1 = 12.377 + 0.23 kft_n1 + e

It can be seen that in a simple linear regression analysis, centering only

changes the estimated intercept bˆ 0 (which is now 12.377 instead of 9.823 for
uncentered scores), but not the estimated regression bˆ 1. The new intercept
of 12.377 represents the expected kft_q1 score for individuals with an
average kft_n1 score (this score is exactly identical to the kft_q1 mean).
 Linear Structural Equation Models   39

The estimated standardized regression slope coefficient equals 0.375

and hence is exactly identical to the bivariate correlation between kft_
n1 and kft_q1 (in a bivariate regression analysis the standardized
regression slope coefficient is always identical to the bivariate correlation
between the predictor and the criterion variable).

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_Q1 ON
KFT_N1 0.375 0.040 9.329 0.000

Intercepts
KFT_Q1 2.931 0.166 17.678 0.000

Residual Variances
KFT_Q1 0.859 0.030 28.448 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

KFT_Q1 0.141 0.030 4.664 0.000

The standardized intercept term (here, the solution with uncentered

independent variable) is calculated by dividing the unstandardized inter-
cept estimate (9.823) by the model-­implied standard deviation of the
dependent variable. This standard deviation is not reported directly by
Mplus; however, it can be calculated as the square root of the variance
σˆ 2Y of kft_q1 (σˆ Y = σˆ 2Y = 11.235 = 3.352 ), which can be found in
the sample statistics output (see above). The standardized intercept term
is therefore given by 9.823/3.352 = 2.931. The standardized residual vari-
ance of 0.859 equals 1 – R̂ 2 (1 – .141 = .859). This value indicates the
proportion of variability in kft_q1 scores that cannot be explained by
kft_n1 (here, approximately 85.9%).

3.3 Latent Regression Analysis

The simple linear regression model described in Section 3.2 considered

the relationship between quantitative and figural intelligence scores at
40   Data Analysis with Mplus

the level of manifest variables. One important disadvantage of regression

analyses with manifest variables is that these analyses make the (unre-
alistic) assumption that the manifest independent variables are perfectly
reliable (i.e., free of measurement error). However, manifest scores typi-
cally contain random measurement error, and the results of the regres-
sion analyses (e.g., the estimated regression coefficients) may therefore be
biased. (Although measurement error variance in the dependent variable
is accounted for in these analyses through the residual variance, it is con-
founded with systematic unexplained variance, which is typically also
undesirable.)
SEMs with latent variables allow us to specify regression analyses
at the latent level, that is, at the level of latent variables (factors) that are
corrected for measurement error. This procedure has the advantage that
errors of measurement can be taken into account explicitly for both the
independent and the dependent variable(s). Furthermore, SEMs with
latent variables allow us to obtain estimates of the reliabilities of the man-
ifest variables. The explicit consideration of measurement error leads to
a more precise estimation of the parameters of the regression model (as
parameters of a latent structural model) compared to manifest regression
analyses with observed variables that are not adjusted for measurement
error.
For each latent variable in a latent regression model, multiple indica-
tors are required (at least two manifest variables for each latent variable;
for reasons of identification and estimation it is better, however, to use at
least three indicators per factor if possible). In this section, we show how
the relationship between quantitative and figural KFT scores can be mea-
sured at the latent rather than the observed level.

BOX 3.5. Example of Latent Regression Analysis

In the data set KFT.dat there are two indicators for each ability facet (i.e., ver-
bal, quantitative, and nonverbal). Therefore, we can model both quantitative
and nonverbal ability through a latent factor. The manifest variables kft_q1
and kft_q3 are used as indicators of a quantitative ability factor (KFT_Q).
The variables kft_n1 and kft_n3 are used as indicators of a latent vari-
able representing nonverbal abilities (KFT_N). The regression of quantitative
abilities on nonverbal abilities is modeled at the level of latent variables (i.e.,
in the structural model). The resulting model can be referred to as a latent
regression analysis. It is shown as a path diagram in Figure 3.5.
 Linear Structural Equation Models   41

(1.24)

KFT_N 0.48
KFT_Q
(14.22)

1.00f 0.65 1.00f 0.90

KFT_N1 KFT_N3 KFT_Q1 KFT_Q3

(29.62) (11.79) (11.22) (6.77)

(15.41) (5.73) (6.76) (3.13)

FIGURE 3.5. Latent regression model for predicting quantitative ability (fac-
tor KFT_Q) versus nonverbal ability (factor KFT_N). Each factor is measured by
two indicators (manifest variables = KFT subscale scores). The values indicate the
unstandardized Mplus parameter estimates (the loadings, path coefficients, and vari-
ances/residual variances). The superscript f indicates factor loadings that were fixed
to 1 to identify the metric of the latent variables.

The Mplus syntax command for specifying the model shown in Fig-
ure 3.5 is depicted in Figure 3.6. First of all, note that all manifest vari-
ables that are used in the model (and only those) have to be listed under
usevariables. The measurement models are specified in the model
command using by statements. These by statements serve to indicate
which manifest variables load onto which latent factor. The word by here
stands for “measured by.” In our example, the latent factor KFT_Q is mea-
sured by the manifest variables kft_q1 and kft_q3, whereas the factor
KFT_N is measured by the variables kft_n1 and kft_n3.
The factor loading of the first indicator listed in each by statement
will be fixed to 1 by default in Mplus (here the loadings of kft_q1 and
kft_n1). Fixing one loading per factor is necessary to assign a metric to
each factor (see also Box 3.6). The loadings of all remaining indicators are
freely estimated as the default.
The command KFT_Q on KFT_N; is used in the same way as the
on command is used in the manifest regression analysis in Section 3.2.
Here this command refers to the regression of the latent factor KFT_Q on
the factor KFT_N. It therefore defines the structural part of this simple
structural equation model. Residual variances for the manifest variables
42   Data Analysis with Mplus

BOX 3.6. Alternative Strategies for Assigning a Metric

to a Latent Factor

The metric of a latent variable is arbitrary and therefore has to be defined by

the user (for details see Bollen, 1989). The default setting in Mplus is that the
factor loading of the first indicator listed in the by statement is fixed to 1 so
that the metric of the factor is defined through the metric of this reference
(or marker) variable. As an alternative to the fixation of one loading per fac-
tor, one can freely estimate all loadings and define the factor metric by fixing
the factor variance (or, for endogenous variables, the residual variance) to a
positive value (usually to 1). Fixing the factor variances to 1 (instead of fixing
factor loadings) is sometimes useful for technical reasons in practical appli-
cations. This is the case, for example, when indicators of the same factor dif-
fer strongly in their metrics (e.g., when they have very different variances) so
that the specification with one fixed loading leads to convergence problems.
Another case in which one may prefer fixing the factor variances rather than
the loadings is when certain parameters of the structural model (e.g., factor
correlations) are supposed to be fixed to a priori hypothesized values. For
example, researchers are sometimes interested in fixing factor correlations
to certain values based on theory. If the factors are assigned variances of 1,
fixing their correlations to a certain hypothesized value is much easier than
when their variances are freely estimated. Fixing the factor variances to 1
puts the factors into a standardized metric, in which the covariance is identi-
cal to the correlation.
On the other hand, fixations of factor variances can sometimes be unde-
sirable. This is the case, for example, in longitudinal SEMs when changes
in the factor variances across time are expected or subject to a researcher’s
hypotheses.
As an example of how to override the Mplus default, we consider the
factor KFT_N. To set the loading of the first indicator free, we add an asterisk
(*) after the first indicator:

KFT_N by kft_n1* kft_n3;

This causes Mplus to set the first loading free. The variance of factor KFT_N
can be fixed to 1 by adding an additional command that lists the name of the
factor and adds “@1” after the name of the factor:

KFT_N@1;

Neither the model fit nor the standardized parameter estimates are affected
by fixing the factor variance rather than the loading in this example (see
output on the companion website).
 Linear Structural Equation Models   43

FIGURE 3.6. Mplus input file for the latent regression model from Figure 3.5.

and the latent dependent variable KFT_Q are estimated by Mplus as the
default. All other commands shown in this input have already been dis-
cussed. By clicking on run, we obtain the model results (maximum likeli-
hood estimates are the default), which are shown, in part, below:

MODEL FIT INFORMATION

Number of Free Parameters 13

Loglikelihood

H0 Value -4696.080
H1 Value -4696.079

Information Criteria

Akaike (AIC) 9418.160

Bayesian (BIC) 9471.724
Sample-Size Adjusted BIC 9430.466
(n* = (n + 2) / 24)

Chi-Square Test of Model Fit

Value 0.002
Degrees of Freedom 1
P-Value 0.9653
44   Data Analysis with Mplus

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.000
90 Percent C.I. 0.000 0.000
Probability RMSEA <= .05 0.980

CFI/TLI

CFI 1.000
TLI 1.015

Chi-Square Test of Model Fit for the Baseline Model

Value 406.899
Degrees of Freedom 6
P-Value 0.0000

SRMR (Standardized Root Mean Square Residual)

Value 0.000

The model fit information shows that the model fits the data very
well. The chi-­square value for the target model shows that the model-­
implied covariance matrix and mean vector do not differ significantly
from the corresponding observed values, c2 = 0.002, df = 1, p = .9635.
The chi-­square value for the baseline model is very large and highly sta-
tistically significant, c2 = 406.899, df = 6, p < .0001, which shows us that
the assumption that all four observed KFT variables are unrelated is not
tenable. The remaining fit statistics also indicate a good model fit, given
that they all show values better than the recommended cutoff scores (see
Box 3.7 for details). According to these cutoff scores, CFI and TLI should
be larger than 0.95 or 0.97, and RMSEA and SRMR should both be smaller
than 0.05 (Schermelleh-­Engel et al., 2003). In the present example, CFI
and TLI are both larger than 0.97, and RMSEA and SRMR are both smaller
than 0.05. The test of close fit also leads to a positive outcome for the
model. This test can be used to test the hypothesis that RMSEA is smaller
than or equal to .05 in the population. This hypothesis is not rejected at
the .05 level in the present case given that p = .980.
In the following (beginning on page 47), the unstandardized Mplus
parameter estimates for the latent regression model from Figure 3.5 are
shown. The unstandardized parameter estimates appear under the header
MODEL RESULTS. The most important estimates are also shown in Fig-
ure 3.5.
 Linear Structural Equation Models   45

BOX 3.7. Criteria for Assessing Goodness of Fit

in Structural Equation Modeling

This box provides an overview of the most important tests and indices avail-
able in Mplus to assess model fit of SEM as well as common cutoff values for
acceptable fit (Hu & Bentler, 1999; Schermelleh-­Engel et al., 2003).

Chi-­Square Test. This tests the null hypothesis that the covariance
matrix and mean vector in the population are equal to the model-­implied
covariance matrix and mean vector (test of exact model fit). A significant chi-­
square value leads to the rejection of the null hypothesis that the model fits
exactly in the population. The degrees of freedom of this test are calculated
as the difference between the number of pieces of available information (vari-
ances, covariances, and means of the manifest variables) minus the number
of estimated model parameters. In Mplus, the chi-­square value, the corre-
sponding degrees of freedom, and the p-value are reported as the default.

Chi-­Square Difference Test. This test can be used to statistically com-

pare two nested SEMs against each other. Nested means that one model is
a special case of another, more general model (i.e., that it can be derived
directly from this model by means of parameter constraints). Important
assumptions of the test are that (1) the more general model fits the data, and
(2) the additional restrictions that lead to the more restricted model do not
cause parameters to be fixed to boundary values of the admissible parameter
space (e.g., a variance fixed to 0 or a correlation set to 1). To conduct a chi-­
square difference test the difference between the chi-­square values of two
nested models as well as the difference in the degrees of freedom are calcu-
lated. The chi-­square difference value is then tested for significance based
on the difference in degrees of freedom. If the chi-­square difference value is
significant, this means that the more restricted model fits the data signifi-
cantly worse than the more general model. Chi-­square difference tests are
frequently used in longitudinal modeling when the assessment of measure-
ment invariance is of interest. Researchers can use Crayen’s (2010) software,
which allows calculating chi-­square difference tests on the basis of Mplus
outputs. This program is presented in Section 4.1.4 (Box 4.4) and is included
on the companion website.

Comparative Fit Index (CFI). The CFI belongs to the family of so-­
called incremental fit indices that compare the fit of the target model to the fit
of a baseline model. In Mplus, the baseline model currently is the so-­called
independence model, which assumes that the population covariance matrix of
the observed variables is a diagonal matrix. This means that the observed
46   Data Analysis with Mplus

variables are allowed to have different variances, but are assumed to have
zero covariances. In other words, it is assumed that there are no relation-
ships between any of the variables. The CFI indicates by how much the target
model fits better than the more parsimonious independence model. For a
good model, the CFI should be larger than 0.95 (better: larger than 0.97).
The CFI is reported as part of the Mplus standard output.

Tucker–Lewis Index (TLI). The TLI also belongs to the family of

incremental fit indices and compares the fit of the target model to the fit of
the independence model. The same cutoff values as for the CFI apply. The
TLI is also part of the Mplus standard output.

Root Mean Square Error of Approximation (RMSEA). The RMSEA

coefficient is a measure of approximate model fit. A good model should have
an RMSEA value smaller than 0.05. Mplus reports the point estimate of
RMSEA as well as a 90% confidence interval around this estimate. Further-
more, a p-value is provided that can be used to test the null hypothesis that
the population RMSEA is ≤ .05.

Standardized Root Mean Square Residual (SRMR). The SRMR coef-

ficient is a standardized measure for the evaluation of the model residuals
(sample minus model-­implied covariances and means). Small SRMR values
indicate that the observed variances, covariances, and means are well repro-
duced by the model on average. Values below .05 are usually seen as indica-
tive of a good fit. SRMR is output by default in Mplus.

Information Criteria (AIC, BIC, Sample-Size Adjusted BIC). Infor-

mation criteria can be used to compare different models on a descriptive
basis. In contrast to the chi-­square difference test, these indices can also be
used to compare models that are not nested. As a rule, the model with the
smallest value of the relevant information criterion is preferred. Information
criteria are part of the Mplus standard output.

Residuals. Covariance and mean residuals can be used as measures

of local (as opposed to global or omnibus) model fit. Residuals represent
the difference between the observed (sample) variances, covariances, and
means and the corresponding model-­implied values. For each observed vari-
ance, covariance, and mean, residuals indicate how well these statistics were
reproduced by the model. Residuals are therefore an important diagnostic
tool to detect possible causes of global model misfit. Unstandardized covari-
ance residuals are often hard to interpret, because they depend on the metric
of the observed variables. Researchers, therefore, often study standardized
residuals (provided as z-scores in Mplus) to evaluate model fit. Standardized
residuals can be more easily compared across different variables, and the
 Linear Structural Equation Models   47

z-score metric allows us to test whether a residual is significantly different

from zero. Both unstandardized and standardized residuals are available in
Mplus as of version 5. They are not part of the standard output, but can be
requested by the user by adding the command

output: residual;

Model Modification Indices. Using model modification indices, a

researcher can examine which restrictions may be relaxed to obtain a sig-
nificant improvement of the global model fit. A modification index for the
single modification (a single additional parameter) is approximately equiva-
lent to a chi-­square value with 1 degree of freedom. It is therefore roughly
equivalent to a chi-­square difference test between the estimated model and
the less restrictive model in which the corresponding restriction is relaxed.
Large modification indices may indicate that a specific restriction is related
to global misfit. In Mplus, modification indices can be obtained using the
command

output: modindices;

Mplus also outputs the expected parameter change, that is, the actual value
that is expected for the additional parameter to be estimated in a less restric-
tive version of the model.
Similar to residuals, modification indices often provide useful hints as
to the cause of an insufficient global model fit. However, they can also some-
times lead to confusing results and often are not directly related to the true
cause of a bad fit. Therefore, model modification indices should never be
used in a purely data-­driven way. In addition, users should always imple-
ment just one (theoretically meaningful) modification at a time. Model modi-
fications should be theoretically justified rather than purely data-­driven, and
they should be validated with independent data.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_Q BY
KFT_Q1 1.000 0.000 999.000 999.000
KFT_Q3 0.903 0.094 9.588 0.000

KFT_N BY
KFT_N1 1.000 0.000 999.000 999.000
KFT_N3 0.652 0.064 10.222 0.000
48   Data Analysis with Mplus

KFT_Q ON
KFT_N 0.476 0.056 8.467 0.000

Intercepts
KFT_Q1 12.387 0.157 78.890 0.000
KFT_Q3 7.732 0.122 63.398 0.000
KFT_N1 11.112 0.255 43.552 0.000
KFT_N3 8.281 0.161 51.457 0.000

Variances
KFT_N 14.215 2.074 6.854 0.000

Residual Variances
KFT_Q1 6.758 0.608 11.107 0.000
KFT_Q3 3.128 0.395 7.921 0.000
KFT_N1 15.406 1.591 9.683 0.000
KFT_N3 5.734 0.644 8.911 0.000
KFT_Q 1.238 0.425 2.916 0.004

Under KFT_Q BY KFT_Q1 KFT_Q3 and KFT_N BY KFT_N1

KFT_N3 we obtain estimates of the unstandardized factor loadings (col-
umn Estimate), as well as their standard errors (S.E.), test statistics
(Est./S.E.), and two-­tailed p-values. As explained above, the first factor
loading has been fixed to 1 by Mplus as the default so that the metric of
each of the two latent factors is identified. For this reason, the loadings
of the indicators KFT_Q1 and KFT_N1 do not have a standard error.
The value of 999.000 in the columns Est./S.E. and Two-­Tailed
P-Value indicates that a significance test is not available for these
parameters because they are fixed.
The loadings of the two remaining indicators (KFT_Q3 and KFT_N3)
were estimated to 0.903 and 0.652, respectively, and are statistically signif-
icantly different from zero (two-­tailed p-values < .0001). The unstandard-
ized regression coefficient for the regression of the latent factor KFT_Q on
the factor KFT_N equals 0.476 and is also statistically significantly differ-
ent from zero (z = 8.467, p < .0001). This shows that the factor KFT_N is
a significant predictor of the factor KFT_Q.
In addition, Mplus outputs the unstandardized intercepts and resid-
ual variances of the observed variables as well as the latent variance of the
exogenous factor KFT_N and the residual variances of the endogenous
factor KFT_Q. (For all latent exogenous variables, Mplus estimates the
variance of these variables as a model parameter. For all latent endog-
enous variables, the corresponding residual variance is estimated.)
The standardized solution STANDARDIZED MODEL RESULTS
(STDYX Standardization) contains fully standardized versions
 Linear Structural Equation Models   49

of the model parameters. The fully standardized factor loadings in this

model can be interpreted as correlations between the observed variables
and the factor on which they load (this is possible here because each
variable loads only onto one factor). For example, 0.631 is the estimated
correlation between the observed variable KFT_Q1 and the latent factor
KFT_Q.
Under KFT_Q ON KFT_N the standardized regression coefficient
for the latent regression of KFT_Q on KFT_N is 0.85. In the present
example, this value can be interpreted as the correlation between the two
latent factors, because KFT_N is the only predictor in the latent regres-
sion (this would not be true for multiple regression analyses with several
correlated predictor variables). Hence, the estimated latent correlation
between KFT_Q and KFT_N equals .85, indicating a very strong relation-
ship between numeric and nonverbal abilities as measured by the KFT at
the latent level.

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_Q BY
KFT_Q1 0.631 0.041 15.548 0.000
KFT_Q3 0.733 0.040 18.247 0.000

KFT_N BY
KFT_N1 0.693 0.038 18.019 0.000
KFT_N3 0.717 0.038 18.717 0.000

KFT_Q ON
KFT_N 0.850 0.049 17.245 0.000

Intercepts
KFT_Q1 3.698 0.131 28.177 0.000
KFT_Q3 2.972 0.109 27.240 0.000
KFT_N1 2.042 0.082 24.798 0.000
KFT_N3 2.412 0.093 26.024 0.000

Variances
KFT_N 1.000 0.000 999.000 999.000

Residual Variances
KFT_Q1 0.602 0.051 11.781 0.000
KFT_Q3 0.462 0.059 7.840 0.000
KFT_N1 0.520 0.053 9.765 0.000
KFT_N3 0.487 0.055 8.869 0.000
KFT_Q 0.278 0.084 3.315 0.001
50   Data Analysis with Mplus

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value
KFT_Q1 0.398 0.051 7.774 0.000
KFT_Q3 0.538 0.059 9.124 0.000
KFT_N1 0.480 0.053 9.010 0.000
KFT_N3 0.513 0.055 9.359 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

KFT_Q 0.722 0.084 8.622 0.000

The standardized residual variances for the observed variables give

the proportion of variability in each manifest variable that is not accounted
for by the corresponding latent factor. Likewise, the latent standardized
residual variance of KFT_Q (0.278) indicates that 27.8% of the variability
in latent (error-free) KFT_Q scores are not accounted for by KFT_N.
In contrast, the estimated R 2 values indicate the proportion of vari-
ability in each endogenous variable that is explained in the model. In this
example, the R 2 values equal the squared standardized factor loading of
each indicator because each variable loads only onto one factor. The esti-
mated R 2 values for the observed variables can be seen as a lower bound
estimate of the reliabilities of these indicators. The values represent a lower
bound because indicator- or scale-­specific variance cannot be separated
from random measurement error in this model. Here, the values indicate
rather low reliabilities of the observed variables. Another interpretation of
the rather small R 2 values is that the scales are rather heterogeneous. This
implies that a relatively large part of their true score (i.e., reliable) variance
is scale-­specific and thus becomes confounded with measurement error in
this model—­leading to an underestimation of the reliabilities.
Under R-SQUARE—Latent Variable we can see that about
72.2% of the variability in the KFT_Q factor can be explained by the
KFT_N factor, which is a pretty substantial amount. In comparison to
the manifest regression analysis (see Section 3.2), in which only the two
observed variables, KFT_Q1 and KFT_N1, were used and no correction for
measurement error was implemented, the amount of explained variance
is much higher in the latent model. This can partly be explained by the
fact that the manifest model did not take measurement error into account,
so that the actual strength of the relationship between the two variables
was underestimated owing to the influences of random measurement
error. In addition, the latent regression model used multiple indicators so
 Linear Structural Equation Models   51

that indicator-­specific variance was separated out and became part of the
error variance. This likely contributed to the fairly large difference in the
amount of explained variance between the manifest and the latent model.

3.4 Confirmatory Factor Analysis

In this section, we discuss confirmatory factor analysis (CFA) in Mplus.

An SEM is commonly referred to as a CFA when only nondirectional rela-
tionships (covariances, correlations) are assumed between the latent vari-
ables in the structural model (rather than directional paths such as latent
regressions) or when the factors are assumed to be orthogonal (uncorre-
lated). CFA is often used to test whether a set of measures has a specific
hypothesized dimensionality, for example, one-­factorial, two-­factorial,
three-­factorial, and so forth.

3.4.1 First‑Order CFA

The first model tested here is a one-­factor model (so-­called general factor or
g-factor model; see Figure 3.7), in which all six KFT scale scores load onto
a single common factor. This model suggests that all six KFT scales assess
a single latent ability dimension and, in addition, include only random
measurement error and indicator-­specific variance. If the g-factor model
fits the data well, one would have to assume that it may be unnecessary
to distinguish between verbal, numerical, and figural abilities or at least
that such a distinction is not possible based on the measures used here.
In more technical terms, the scales would show a lack of discriminant
validity.

BOX 3.8. Example of CFA

As an example, consider the six KFT variables in the data set KFT.dat. The
following research question may be of interest: Do the six KFT scales kft_v1,
kft_v2, kft_q1, kft_q3, kft_n1, and kft_n3 measure three distinct dimen-
sions (verbal, quantitative, and nonverbal/figural abilities) or are they merely
indicators of the single latent ability dimension, for example, a general factor
of intelligence? We examine this question by means of two different first-
order CFA models in Mplus. Subsequently, we also discuss the specification
of a second-­order CFA model.
52   Data Analysis with Mplus

KFT-g
(9.41)

1.00f 1.00 1.23 0.76 0.63 0.53

KFT_V1 KFT_V3 KFT_N1 KFT_N3 KFT_Q1 KFT_Q3

(21.63) (17.97) (29.62) (11.79) (11.22) (6.77)

(12.22) (8.53) (15.44) (6.35) (7.49) (4.12)

FIGURE 3.7. Single-factor (g-factor) model for the six KFT variables with unstan-
dardized Mplus parameter estimates. The superscript f indicates that the factor load-
ing of the first indicator (KFT_V1) was fixed to 1 to identify the metric of the latent
variable.

The second model (see Figure 3.8) contains three latent factors sug-
gesting that the six KFT scales measure three distinct dimensions of men-
tal ability (kft_v1 and kft_v3 as indicators of a verbal ability factor, kft_q1
and kft_q3 as indicators of a quantitative/numerical ability factor, and
kft_n1 and kft_n3 as indicators of a nonverbal/figural ability factor). A
statistical comparison of the two models allows us to find out whether a
one- or a three-­factor model is more appropriate to represent what is mea-
sured by the KFT scales.
The Mplus input files for the one- and three-­factor models are shown
in Figure 3.9 and Figure 3.10, respectively. Note that as an exception, the
usevariables command is not needed in this example, because all
variables in the data set (all variables that are listed under names =) are
actually used in the model statement. In all other cases, the usevari-
ables command needs to be included (compare Section 3.2).
Further note that the covariances between the three latent variables
in the three-­factor model do not have to be explicitly specified in the
Mplus input file. By default Mplus estimates all covariances between exog-
enous factors automatically (because none of the three factors is predicted
by another variable in the model, they are all considered exogenous vari-
ables).
Sometimes a user may be interested in explicitly setting the covari-
ance/correlation between two factors to zero (e.g., see Section 4.1.3). This
 Linear Structural Equation Models   53

5.25 (r = .75)

11.09 (r = .89) 7.12 (r = .85)

KFT_V KFT_N KFT_Q
(10.42) (14.95) (4.66)

1.00f 1.03 1.00f 0.62 1.00f 0.87

KFT_V1 KFT_V3 KFT_N1 KFT_N3 KFT_Q1 KFT_Q3

(21.63) (17.97) (29.62) (11.79) (11.22) (6.77)

(11.21) (6.83) (14.67) (6.03) (6.56) (3.28)

FIGURE 3.8. Three-­factor model for the six KFT variables with unstandardized
Mplus parameter estimates (latent correlations are given in parentheses). The super-
script f indicates factor loadings that were fixed to 1 to identify the metric of the
latent variables.

FIGURE 3.9. Mplus input file for the single-­factor model from Figure 3.7.
54   Data Analysis with Mplus

FIGURE 3.10. Mplus input file for the three-­factor model from Figure 3.8.

can be done by specifying, for example, F1 with F2@0; in the model

command. In this case, the correlation between the two factors F1 and
F2 would be fixed to zero, which could be of interest, for example, if one
wanted to test the hypothesis that the correlation between these factors
is zero in the population. (In the present example, we are expecting sig-
nificant positive correlations between all factors, and thus we will not use
this command here.)
Using the additional tech4 option in the output command allows us
to output the estimated covariance and correlation matrices for the latent
factors in matrix form (in addition to the output of these parameters in the
regular Mplus model results table). Obtaining these estimates in matrix
form can be practical and is sometimes clearer, especially when there are
many latent factors.
Fit statistics estimated for both the one- and the three-­factor model
are shown in Table 3.1. It can be seen that the one-­factor model shows a
significant chi-­square value, whereas the chi-­square value for the three-­
factor model is not significant. According to this criterion, the one-­factor
model would have to be rejected, whereas the three-­factor model fits the
data well. In addition, the RMSEA value for the one-­factor model is larger
than the commonly recommended cutoff value of 0.05. According to the
 Linear Structural Equation Models   55

TABLE 3.1. Model Fit Statistics for the One- and Three-­Factor

Models for the Six KFT Variables

One factor Three factors

c2 (df) 33.939 (9) 7.093 (6)

p(c2) 0.0001 0.3123
CFI/TLI 0.969/0.948 0.999/0.997
RMSEA (90% confidence interval) 0.078 (0.051, 0.107) 0.020 (0.000, 0.066)
p(RMSEA ≤ 0.05) 0.043 0.820
SRMR 0.031 0.014
AIC 14343.362 14322.516

Note. A detailed description of these fit indices is provided in Box 3.7.

CFI, TLI, and SRMR values, both the one- and the three-­factor model
would be acceptable. The AIC value for the one-­factor model is somewhat
higher than the AIC value for the three-­factor model, so that according to
this criterion, the three-­factor model would again be preferred. In sum-
mary, the fit statistics seem to point more to the three-­factor solution than
to the one-­factor solution.
The Mplus parameter estimates for the three-­factor model are shown
in the following. We again begin with the MODEL RESULTS table, which
provides the unstandardized solution. It turns out that all loadings (listed
under BY) that were freely estimated are significantly different from zero
(the first loadings of the first indicator is again set to 1 for each factor
to identify the metric of each factor). All three estimated factor covari-
ances (shown under WITH) are also statistically significant. The size of
the relationship between the three factors is difficult to evaluate based on
the covariances, however, because covariances are unstandardized mea-
sures of association. Standardized covariances (correlations r) are easier to
interpret because they are restricted to a standardized range (–1 ≤ r ≤ 1).
The correlations of the latent factors are provided in the standardized
solution and are discussed below.
The unstandardized solution furthermore contains the unstandard-
ized intercepts of the manifest variables. In this case, the intercepts are
identical to the observed means of the manifest variables because there
are no restrictions on the mean structure in this model. Also given are
the estimated factor variances and the residual variances for the manifest
variables.
56   Data Analysis with Mplus

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_V BY
KFT_V1 1.000 0.000 999.000 999.000
KFT_V3 1.034 0.083 12.434 0.000

KFT_Q BY
KFT_Q1 1.000 0.000 999.000 999.000
KFT_Q3 0.865 0.084 10.337 0.000

KFT_N BY
KFT_N1 1.000 0.000 999.000 999.000
KFT_N3 0.620 0.050 12.394 0.000

KFT_Q WITH
KFT_V 5.247 0.697 7.533 0.000

KFT_N WITH
KFT_V 11.088 1.222 9.074 0.000
KFT_Q 7.123 0.839 8.488 0.000

Intercepts
KFT_V1 11.899 0.218 54.574 0.000
KFT_V3 8.974 0.199 45.155 0.000
KFT_Q1 12.387 0.157 78.888 0.000
KFT_Q3 7.732 0.122 63.396 0.000
KFT_N1 11.112 0.255 43.552 0.000
KFT_N3 8.281 0.161 51.457 0.000

Variances
KFT_V 10.424 1.401 7.442 0.000
KFT_Q 4.659 0.720 6.471 0.000
KFT_N 14.953 1.968 7.596 0.000

Residual Variances
KFT_V1 11.205 1.004 11.156 0.000
KFT_V3 6.830 0.853 8.011 0.000
KFT_Q1 6.559 0.586 11.192 0.000
KFT_Q3 3.283 0.366 8.980 0.000
KFT_N1 14.667 1.382 10.615 0.000
KFT_N3 6.033 0.550 10.967 0.000

Under STANDARDIZED MODEL RESULTS (STDYX Stan-

dardization) Mplus provides the fully standardized solution. The
standardized factor loadings again appear under BY, the latent factor
 Linear Structural Equation Models   57

correlations under WITH. It can be seen that all three KFT factors are
strongly positively correlated (.75 ≤ r ≤ .89), which shows that there is
rather little discriminant validity among the three KFT subscales.
Despite the rather large correlations among the three factors, the
fit statistics indicated that more than one factor is needed to sufficiently
account for the observed variances and covariances of the six KFT sub-
scales and that a g-factor solution (which implies perfectly correlated
factors) had to be rejected for the present data. Therefore, we can con-
clude that even though the three subscales appear to share a substantial
amount of common variance, there also is a nontrivial portion of system-
atic domain-­specific (verbal, numerical, and figural) ability variance that
needs to be taken into account and that explains why the correlations
among the three factors differ from 1.

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

KFT_V BY
KFT_V1 0.694 0.033 20.963 0.000
KFT_V3 0.787 0.031 25.290 0.000

KFT_Q BY
KFT_Q1 0.644 0.039 16.698 0.000
KFT_Q3 0.718 0.038 18.959 0.000

KFT_N BY
KFT_N1 0.711 0.033 21.591 0.000
KFT_N3 0.699 0.033 21.046 0.000

KFT_Q WITH
KFT_V 0.753 0.051 14.764 0.000

KFT_N WITH
KFT_V 0.888 0.041 21.589 0.000
KFT_Q 0.853 0.049 17.458 0.000

Intercepts
KFT_V1 2.558 0.097 26.401 0.000
KFT_V3 2.117 0.084 25.083 0.000
KFT_Q1 3.698 0.131 28.176 0.000
KFT_Q3 2.972 0.109 27.239 0.000
KFT_N1 2.042 0.082 24.798 0.000
KFT_N3 2.412 0.093 26.024 0.000
58   Data Analysis with Mplus

Variances
KFT_V 1.000 0.000 999.000 999.000
KFT_Q 1.000 0.000 999.000 999.000
KFT_N 1.000 0.000 999.000 999.000

Residual Variances
KFT_V1 0.518 0.046 11.267 0.000
KFT_V3 0.380 0.049 7.753 0.000
KFT_Q1 0.585 0.050 11.753 0.000
KFT_Q3 0.485 0.054 8.933 0.000
KFT_N1 0.495 0.047 10.589 0.000
KFT_N3 0.512 0.046 11.037 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

KFT_V1 0.482 0.046 10.482 0.000

KFT_V3 0.620 0.049 12.645 0.000
KFT_Q1 0.415 0.050 8.349 0.000
KFT_Q3 0.515 0.054 9.479 0.000
KFT_N1 0.505 0.047 10.796 0.000
KFT_N3 0.488 0.046 10.523 0.000

3.4.2 Second‑Order CFA

The high (albeit imperfect) correlations among the three KFT factors in the
three-­factor CFA model made us speculate about whether there might be
a common ability factor that underlies all three ability domains (a g-factor
of intelligence?). Given the high correlations, there appeared to be a large
amount of shared variance and only a relatively minor amount of specific
verbal, numerical, and figural ability variance. A model that is in line with
the idea of a g-factor, while still allowing for domain-­specific (residual)
variability, is the second-­order factor model shown in Figure 3.11. In this
model, the three first-order KFT factors (KFT_V, KFT_Q, and KFT_N)
themselves load onto a factor, a so-­called second-­order factor (KFT_g).
The Mplus model specification for the second-­order CFA model shown
in Figure 3.11 is provided in Figure 3.12. Note that in this example, even
though a second-­order factor model can be specified, the model is statisti-
cally equivalent to the correlated three-­factor model discussed previously.
At least four first-order factors are required for the second-­order factor
structure to be overidentified and more restrictive than a model with cor-
related first-order factors. This means that the second-­order factor model
 Linear Structural Equation Models   59

KFT_G
(8.17)

1.00f 0.64
1.36

ζ 1(2.26) ζ 2(–0.10) ζ 3 (1.29)

KFT_V KFT_N KFT_Q

1.00f 1.03 1.00f 0.62 1.00f 0.87

KFT_V1 KFT_V3 KFT_N1 KFT_N3 KFT_Q1 KFT_Q3

(21.63) (17.97) (29.62) (11.79) (11.22) (6.77)

(11.21) (6.83) (14.67) (6.03) (6.56) (3.28)

FIGURE 3.11. Second-order factor model for the six KFT variables with unstan-
dardized Mplus parameter estimates. In this model, the three first-order KFT factors
KFT_V, KFT_N, and KFT_Q load onto a second-­order factor (KFT_G). The super-
script f indicates factor loadings that were fixed to 1 to identify the metric of the
latent variables. Variances and residual variances are given in parentheses. Note the
estimated negative residual variance (“Heywood case”) for the factor KFT_N (–0.10).

cannot be statistically distinguished from the correlated three-­ factor

model based on model fit in this example. Another possibility to obtain
an overidentified structural model would be to impose restrictions on
the loadings of the first-order factors on the second-­order factor (e.g., the
restriction of equal loadings).
In this example, even though the fit of the second-­order model is
identical to the fit of the correlated three-­factor model, there is still some
doubt as to whether the second-­order factor model should be used. The
reason is that this model results in an inadmissible parameter estimate
(so-­called Heywood case; Chen, Bollen, Paxton, Curran, & Kirby, 2001;
see also Box 3.9). In this case, the residual variance of the factor KFT_N is
estimated to a negative value (–0.099), which represents a nonadmissible
60   Data Analysis with Mplus

FIGURE 3.12. Mplus input file for the second-­order factor model from Figure
3.11.

parameter estimate because variances, by definition, cannot be negative.

Mplus indicates this problem with the following warning message:

WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE

DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE
FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN
TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT
VARIABLES. CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM
INVOLVING VARIABLE KFT_N.

The estimated negative residual variance could be an indication that

the first-order factors show a heterogeneous correlation pattern that is not
in line with the idea of a common second-­order factor. In this case, it
seems best to work with the less problematic model of three correlated
first-order factors instead of the second-­order factor model and to inter-
pret the parameters of the second-­order model only with great caution.
In the following, the Mplus estimates for the residual variances of the
model from Figure 3.11 are shown. Note the negative residual variance for
the factor KFT_N (–0.099).
 Linear Structural Equation Models   61

BOX 3.9. Inadmissible Parameter Estimates (Heywood Cases)

in Structural Equation Modeling

A solution is commonly referred to as a Heywood case if it contains inadmis-

sible (out-of-range) parameter estimates such as negative (residual) variances
or correlations larger than 1. These kinds of inadmissible estimates can have
multiple causes, which can also occur in combination with each other. The
most typical causes are:

• Model misspecification, for example, because of heterogeneous indicators

in the measurement model or the specification of too many or too few
factors
• Use of too small a sample size
• Use of too few indicators per factor
• Sampling error
• Outliers or extreme cases

In my experience, the most common cause of inadmissible parameter

estimates are model misspecifications. For this reason, inadmissible esti-
mates should not simply be eliminated through parameter fixations (e.g., set
a negative variance to zero) or other cosmetic actions. Instead, the possible
causes of these problems should be examined carefully. Often, a model can
be respecified in a more meaningful way that does not produce offending
parameter estimates. A more detailed discussion of the issue of Heywood
cases can be found in Chen et al. (2001).

MODEL RESULTS
Residual Variances
KFT_V1 11.206 1.004 11.156 0.000
KFT_V3 6.829 0.853 8.010 0.000
KFT_Q1 6.559 0.586 11.192 0.000
KFT_Q3 3.283 0.366 8.979 0.000
KFT_N1 14.667 1.382 10.614 0.000
KFT_N3 6.033 0.550 10.967 0.000
KFT_V 2.256 0.754 2.992 0.003
KFT_Q 1.289 0.402 3.209 0.001
KFT_N -0.099 1.206 -0.082 0.935

Negative residual variance

(“Heywood case”)!
62   Data Analysis with Mplus

3.5 Path Models and Mediator Analysis

3.5.1 Introduction and Manifest Path Analysis
Path analyses, like CFAs, are among the most frequently used models
that are estimated in the SEM framework. In general, path analysis can
be seen as a multivariate regression model—that is, a regression analysis
that simultaneously considers multiple dependent (and often also multiple
independent) variables, in contrast to conventional regression analyses
that are restricted to a single dependent variable (see Section 3.2). In a
path analysis, exogenous variables are typically allowed to be correlated.
Path analyses can be carried out at either the level of manifest vari-
ables (an example is shown in Figure 3.13) or at the level of latent vari-
ables (latent path analysis; see Section 3.5.3). Here, we first discuss the
simpler case of a manifest path analysis.
Figure 3.13 shows the path model proposed by Whitelaw and Liang
(1991) using three manifest variables. In this model, it is assumed that the
exogenous variable, physical health (X; here measured as the number of
diseases in the last 12 months; variable SICK), predicts the endogenous
variable, functional health (Y1; the sum score of the SF-36 questionnaire
for measuring functional health; variable FH), which itself predicts the
second endogenous variable, subjective health (Y2; measured through the
self-­report of subjective health in the present; variable SHP). In addition,
it is assumed that the construct physical health also has a direct effect on
subjective health. The coefficients b1, b2, and b3 denote regression or path
coefficients that characterize the strength of the influence of one variable
on another. These coefficients are equivalent to slope coefficients in linear
regression analysis.

β3
ε1 ε2
X β1 Y1 β2 Y2
Physical Functional Subjective
Health Health Health
(SICK) (FH) (SHP)

FIGURE 3.13. Manifest path model for the relationship between physical health
(X), functional health (Y1), and subjective health (Y2) proposed by Whitelaw and
Liang (1991). The constants b1, b2, and b3 indicate regression (path) coefficients.
The variables e1 and e2 are residual variables. The model contains one indirect effect,
which can be expressed as the product b1 · b2. The variable names used in the Mplus
analyses are given in parentheses.
 Linear Structural Equation Models   63

BOX 3.10. Example Path Analysis

As an example, we consider the theoretical model proposed by Whitelaw

and Liang (1991) that is used in health psychology. This model deals with
the relationship between three constructs: physical health, functional health,
and subjective health. Physical health refers to the actual objective physiologi-
cal health status (e.g., presence or absence of trouble or disease). Functional
health refers to the ability to carry out everyday activities and actions (e.g.,
walking, climbing stairs). Subjective health refers to the individual, global sub-
jective judgment of one’s own health (a more detailed description of these
three constructs can be found in Pinquart, 2001).
The data considered here are taken from the age survey sample of the
German center of gerontology (Deutsches Zentrum für Altersfragen, DZA)*
and are located in the file health.dat. This data set contains individual data
for six variables: the number of diseases or health problems (variable SICK)
as well as the number of consultations of medical doctors in the last 12
months (variable CONSULT) as indicators of physical health, the subscale
for the assessment of functional health taken from the SF-36 Health Survey
questionnaire (Ware & Sherborne, 1992). This scale is represented in terms
of both a global sum score (variable FH) and also in terms of two scale scores
(so-­called item parcels) as represented by the variables FH1 and FH2. Finally,
two items are included to assess subjective health: (1) subjective health in the
present (variable SHP) and (2) changes in subjective health across the last 6
years (variable SHC).

*I would like to thank Susanne Wurm from the DZA for providing the data for this
example as well as Christopher Marx for running some of the analyses.

One reason why path analysis is very popular in the social sciences is
that many social science theories include hypotheses with regard to direct
and indirect effects of variables on each other (e.g., see MacKinnon, 2008).
Indirect effects are effects that are mediated through other variables and are
therefore also referred to as mediated effects (Baron & Kenny, 1986). Medi-
ated effects are present in a path model when the model contains one or
more variables that are dependent and independent variables at the same
time. Those types of variables are often referred to as intermittent or media-
tor variables. Mediated effects can be analyzed by means of either manifest
or latent path analyses.
In the model in Figure 3.13, the variable Y1 (functional health, FH)
is a mediator variable, because this variable is a dependent variable with
regard to the variable X (physical health, SICK) but an independent
variable with regard to the variable Y2 (subjective health, SHP). In other
64   Data Analysis with Mplus

words, Y1 receives an effect from X (path b1) while at the same time send-
ing an effect out to Y2 (path b2). Hence, Y1 partly mediates the effect of X
on Y2, which explains the term mediator variable. Formally, the indirect
(or mediated) effect of X on Y2 can be quantified as the product of b1 times
b2 (MacKinnon, 2008). Variable X, in this example, also has a direct effect
on Y2, which is expressed through the path coefficient b3.
In substantive terms, all this means that, in this model, it is assumed
that the objective health status (physical health) of a person has both
a direct (nonmediated) and an indirect (mediated) effect (via functional
health) on subjective health. The mediated effect is caused by the fact that
objective physical problems often result in a loss of functional competen-
cies that, in turn, cause a decrease in subjective health. By fitting the path
model to our data, we can test whether physical health indeed has both a
direct and an indirect effect on subjective health, or whether the effect of
physical health on subjective health is fully mediated through functional
health.
In path analyses that contain one or more indirect effects, an effect
decomposition is often of interest. In this decomposition, the total effect is
split up into the sum of all indirect effects plus the direct effect (if a direct
effect is assumed). In other words, the total effect is equal to the sum of
the direct effect plus all indirect effects. In our example, the direct effect
of X on Y2 is measured by the coefficient b3, whereas the indirect effect is
measured by the product b1 · b2. Therefore, the total effect can be calcu-
lated as the sum b3 + (b1 · b2).

BOX 3.11. Mediated versus Moderated Effects

Mediator effects should not be confused with moderator effects (see also
Baron & Kenny, 1986). In contrast to a mediator effect, a moderator effect
is present when there is an interaction between two variables. A variable is
referred to as a moderator variable when the strength of the direct influence
of one variable (say, X) on another variable (say, Y) depends on the values of
a third variable (say, Z). If, for example, the regression coefficient b3 (which
characterizes the direct effect of X on Y2 in Figure 3.13) varied across dif-
ferent levels of Y1 (i.e., if the direct influence of X on Y2 was moderated by
Y1), then Y1 could be seen as a moderator variable. The analysis of moderator
(interaction) effects is not discussed in this book. Aiken and West (1991;
as well as Cohen et al., 2003, their Chapters 7–9) discuss the analysis of
moderated effects in detail in the context of manifest regression models (so-­
called moderated regression). Latent moderator models (SEMs with interaction
effects) are discussed, for example, in Marsh, Wen, and Hau (2006).
 Linear Structural Equation Models   65

BOX 3.12. Significance Testing of Indirect Effects

Often, the statistical significance of indirect effects is of interest when medi-

ated effects are the subject of the study. One difficulty for statistical inference
with mediated effects is that the indirect effect is the product of two or more
regression coefficients (in our example, the product b1 · b2). The assumption
of a normal distribution of this product term in the population is often vio-
lated. For this reason, conventional methods of significance testing for indi-
rect effects may lead to biased results (MacKinnon, Lockwood, & Williams,
2004). MacKinnon and his colleagues have studied different methods for
testing indirect effect in detail (MacKinnon, 2008; MacKinnon et al., 2004;
MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002). Among other
methods, these authors recommend testing mediated effects via asymmet-
ric confidence intervals derived from bootstrapping. The theoretical back-
ground of these methods is discussed in detail in MacKinnon (2008). An
indirect effect is significant at the .05 level, according to this method, if the
value of 0 is not part of the 95% bootstrap confidence interval around the
indirect effect. MacKinnon et al. (2004) recommend, among others, the so-­
called bias-­corrected bootstrap as a method for statistical inference in statisti-
cal mediation analysis using confidence intervals. Bootstrap methods as well
as confidence intervals are implemented in Mplus and are easy to obtain. An
application of these methods to testing mediated effects is shown in Section
3.5.2.

3.5.2 Manifest Path Analysis in Mplus

Figure 3.14 shows the Mplus input file for the specification of manifest
path model in Figure 3.13. In this model, the variable SICK is used for
the operationalization of the construct physical health. Functional health
is operationalized by the total score of the SF-36 subscale for measuring
functional health (variable FH). Subjective health is represented by the
item referring to subjective health in the present (variable SHP).
In contrast to the regression analysis discussed in Section 3.2, we
now use more than one on statement in the model command, because
we simultaneously consider two linear regression analyses. In our model,
there are two dependent variables (FH and SHP). Therefore, two on state-
ments are required: one for the (multiple) regression of SHP on FH and
SICK, and one for the (simple) regression of FH on SICK.

   FH on SICK;
   SHP on SICK FH;
66   Data Analysis with Mplus

FIGURE 3.14. Mplus input file for the manifest path model from Figure 3.13.

By using the additional model indirect subcommand, we request

that in addition to the standard output, we obtain a separate listing of all
direct, indirect, and total effects on the outcome variable SHP (note that
the model indirect subcommand can be used in both manifest and
latent path analyses):

   model indirect: SHP ind SICK;

The keyword ind between the two variable names causes Mplus to out-
put all indirect effects that the variable SICK has on the variable SHP (in
this case, there is only one indirect effect via FH), along with the relevant
standard errors and tests of significance. The model indirect sub-
command is particularly useful when a researcher wants to test specific
indirect effects for significance (see below).
In the Mplus output, we obtain sample statistics with the means,
covariances, and correlations of the three manifest variables (only the
correlation table is shown here). As expected, subjective health corre-
lates positively with functional health and negatively with the number
of health problems in the last 12 months. In addition, the frequency of
health problems is negatively correlated with functional health.
 Linear Structural Equation Models   67

Correlations
SHP FH SICK
________ ________ ________
SHP 1.000
FH 0.594 1.000
SICK -0.458 -0.447 1.000

Under the rubric MODEL FIT INFORMATION we can see that the
estimated manifest path model in our example has zero degrees of free-
dom. The reason is that all means, variances, and covariances of the three
manifest variables are used up in the model to estimate model parameters.
One therefore speaks of a so-­called saturated or just identified model in this
case (this was the same issue for the manifest regression model discussed
in Section 3.2). One consequence is that the model does not contain any
testable restrictions with respect to the mean and covariance structure of
the observed variables and therefore fits perfectly (for details see Bollen,
1989). This can be seen from the chi-­square model test:

MODEL FIT INFORMATION

Chi-Square Test of Model Fit

Value 0.000
Degrees of Freedom 0
P-Value 0.0000

The chi-­square value is zero with zero degrees of freedom. The test
is therefore not useful to evaluate the model fit in this case (the p-value is
erroneously reported to be 0 by Mplus; it should in fact be printed as p =
1 in this case, given that the model has a perfect fit to the data). In gen-
eral, the judgment of the model fit of a saturated model focuses more on
the estimated model parameters (in particular, the estimated path coef-
ficients) and the obtained proportion of the explained variability in the
endogenous variables (as measured by the R 2 value) rather than the global
fit statistics output by Mplus.
The Mplus MODEL RESULTS table provides us with the estimates
of the unstandardized path coefficients (here for the saturated path model
that includes the direct effect from SICK to SHP). Fully standardized
path coefficients can again be found under STANDARDIZED MODEL
RESULTS (STDYX Standardization). It can be seen that there is
a significant negative direct effect of the frequency of health issues on sub-
jective health (direct effect from SICK to SHP, bˆ 3 = –0.094, z = –11.258,
p < .001, bˆ standardized
3 = –0.24). For every additional health issue encoun-
tered in the last 12 months, the expected decrease in subjective health is
68   Data Analysis with Mplus

BOX 3.13. Saturated versus Nonsaturated Path Models

Note that not every manifest path model is necessarily saturated. If, for
example, in the present model one deletes the direct path from SICK to SHP
(which is equivalent to fixing the corresponding path coefficient b3 to zero),
the model would no longer be saturated. Instead, it would have one degree
of freedom, implying one testable restriction. Using tests or indices of model
fit, the assumption that the effect of physical health on subjective health is
fully mediated by functional health could then be tested. In other words,
this restricted model would be in line with the idea that physical health only
indirectly affects subjective health.
The Mplus model specification would have to be changed, as follows,
to obtain a nonsaturated model as described above (the full input for this
alternative model can be found on the companion website):

   FH on SICK;
   SHP on FH;

The resulting more restricted model, without the direct effect from SICK to
SHP, shows a highly significant chi-­square value of 122.11 with one degree
of freedom (p < .0001). The remaining fit indices (especially the RMSEA
index) also point to a rather bad fit of this model (RMSEA = .27, CFI = .90,
SRMR = .06). Therefore, the hypothesis of complete mediation has to be
rejected (there is also a significant direct effect from physical health to sub-
jective health).

–0.094, controlling for functional health. Functional health is positively

related to subjective health (direct effect bˆ 2 = 0.885, z = 22.765, p < .001,
bˆ 2standardized = 0.486): For every one-unit increase in functional health as
measured by the SF-36 questionnaire, the expected increase in subjective
health is 0.885, controlling for the number of health problems reported
for the last 12 months. Physical health problems are significantly nega-
tively related to functional health (bˆ 1 = –0.096, z = –20.273, p < .001,
bˆ 1standardized = –0.447): For every additional health issue in the last 12
months, the expected decrease in the functional health score is –0.096
units. The indirect effect from physical health to subjective health is dis-
cussed in detail below.
The estimated R 2 values for the two endogenous variables appear
under the standardized parameter estimates. With the present model,
approximately 39.9% of the variability in subjective health (R̂ 2 = .399)
and approximately 20% of the variability in functional health (R̂ 2 = .20)
can be explained.
 Linear Structural Equation Models   69

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

FH ON
SICK -0.096 0.005 -20.273 0.000

SHP ON
SICK -0.094 0.008 -11.258 0.000
FH 0.885 0.039 22.765 0.000

Intercepts
SHP 1.408 0.117 12.034 0.000
FH 2.936 0.016 179.357 0.000

Residual Variances
SHP 0.432 0.015 28.705 0.000
FH 0.174 0.006 28.705 0.000

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed
Estimate S.E. Est./S.E. P-Value

FH ON
SICK -0.447 0.020 -22.661 0.000

SHP ON
SICK -0.240 0.021 -11.441 0.000
FH 0.486 0.019 25.057 0.000

Intercepts
SHP 1.661 0.151 10.973 0.000
FH 6.304 0.099 63.378 0.000

Residual Variances
SHP 0.601 0.019 32.153 0.000
FH 0.800 0.018 45.431 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

SHP 0.399 0.019 21.304 0.000

FH 0.200 0.018 11.330 0.000
70   Data Analysis with Mplus

The additional output that we requested through the model indi-

rect command provides information about the size and statistical signif-
icance of the indirect effect from physical health to subjective health that
is estimated in the model (reminder: formally, the indirect effect is the
product of the path coefficients b1 · b2). In the first part of the additional
output (labeled TOTAL, TOTAL INDIRECT, SPECIFIC INDI-
RECT, AND DIRECT EFFECTS), the effects are reported in unstan-
dardized form together with their respective standard errors and tests of
significance:

TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Effects from SICK to SHP

Total -0.178 0.009 -20.892 0.000

Total indirect -0.085 0.006 -15.140 0.000

Specific indirect

SHP
FH
SICK -0.085 0.006 -15.140 0.000

Direct
SHP
SICK -0.094 0.008 -11.258 0.000

The estimated total effect is the sum of the estimated direct effect bˆ 3
and the estimated indirect effect bˆ 1 · bˆ 2 :

bˆ 3 + (bˆ 1 · bˆ 2 ) = –0.094 + (–0.096 · 0.885) = –0.178

The total indirect effect is given by

bˆ 1 · bˆ 2 = –0.096 · 0.885 = –0.085

In this example, we have to deal with only one indirect effect from
physical health to subjective health (mediated by functional health), such
that the total indirect effect is equal to the specific indirect effect (given
under Specific indirect). The default test used by Mplus to test
the indirect effect for significance is the so-­called Sobel test (Sobel, 1982).
 Linear Structural Equation Models   71

The standard error for the indirect effect provided by Mplus is 0.006
(z = –15.14). Therefore, the indirect effect is significant (p < .001) accord-
ing to this test.
Mplus also outputs the estimated direct effect bˆ 3 again, which equals
–0.094. The standardized solution lists the same effects again now in fully
standardized form:

STANDARDIZED TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND

DIRECT EFFECTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Effects from SICK to SHP

Total -0.458 0.019 -23.496 0.000

Total indirect -0.217 0.013 -16.600 0.000

Specific indirect

SHP
FH
SICK -0.217 0.013 -16.600 0.000

Direct
SHP
SICK -0.240 0.021 -11.441 0.000

The use of conventional tests of significance for testing indirect effects

can be problematic in practice, because the product of two or more regres-
sion coefficients is tested. This product often is not normally distributed,
which can make conventional tests of significance unreliable (MacKinnon,
2008; see Box 3.12). MacKinnon and his colleagues (MacKinnon, 2008;
MacKinnon et al., 2004) therefore recommend the use of asymmetric con-
fidence intervals based on bootstrap methods for significance testing of
indirect effects as a more appropriate alternative. In Mplus, confidence
intervals around model parameter estimates can generally (i.e., with or
without bootstrap methods) be obtained via the following command:

output: cinterval;

If one wants to estimate confidence intervals based on bias-­corrected

bootstrap methods, the following additional specification is required in
the analysis statement:
72   Data Analysis with Mplus

analysis: bootstrap = <number of bootstrap samples>;

output: cinterval (bcbootstrap);

The subcommand bootstrap = is used to choose the number of

bootstrap samples. For a high precision, a large number of bootstrap sam-
ples should be chosen (e.g., 10,000). The type of bootstrap method used is
chosen in the output command (in parentheses behind cinterval).
The keyword bcbootstrap refers to the bias-­corrected version of the
bootstrap recommended by MacKinnon et al. (2004). Conventional boot-
strap confidence intervals can be obtained using the command output:
cinterval (bootstrap);
The complete input and output files for the estimation of the path
model with bias-­corrected bootstrap confidence intervals can be found
on the companion website. In the following we show only the part of
the Mplus output that contains the bias-­corrected bootstrap confidence
intervals estimated based on 10,000 bootstrap draws for both the unstan-
dardized and the standardized parameter estimates. Under the columns
Lower .5% and Upper .5% we find the lower and upper limits of 99%
confidence intervals around each parameter estimate, respectively. The
columns labeled Lower 2.5% and Upper 2.5% contain the lower and
upper bounds of 95% confidence intervals, respectively.

CONFIDENCE INTERVALS OF TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT,

AND DIRECT EFFECTS

Lower .5% Lower 2.5% Estimate Upper 2.5% Upper .5%

Effects from SICK to SHP

Total -0.203 -0.197 -0.178 -0.159 -0.153
Total indirect -0.102 -0.098 -0.085 -0.072 -0.068

Specific indirect
SHP
FH
SICK -0.102 -0.098 -0.085 -0.072 -0.068
Direct
SHP
SICK -0.117 -0.111 -0.094 -0.076 -0.071

It turns out that also according to the bias-­corrected bootstrap con-

fidence interval, the indirect effect is significant at the 1% level. This can
be seen from the fact that neither the 95% confidence interval (–0.098,
–0.072) nor the 99% confidence interval (–0.102, –0.068) around the indi-
rect effect includes the value of zero. The same is true for the standardized
 Linear Structural Equation Models   73

indirect effect. The relevant output lines are shown in boldface, respec-
tively:

CONFIDENCE INTERVALS OF STANDARDIZED TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT,

AND DIRECT EFFECTS (STDYX Standardization)

Lower .5% Lower 2.5% Estimate Upper 2.5% Upper .5%

Effects from SICK to SHP

Total -0.513 -0.500 -0.458 -0.415 -0.402
Total indirect -0.257 -0.247 -0.217 -0.187 -0.178

Specific indirect
SHP
FH
SICK -0.257 -0.247 -0.217 -0.187 -0.178

Direct
SHP
SICK -0.299 -0.285 -0.240 -0.196 -0.182

The results therefore indicate that physical health has a significant

direct, as well as a significant indirect effect on subjective health.

3.5.3 Latent Path Analysis

In latent path analysis, the idea of CFA is combined with the idea of a
manifest path analysis. This means that, as in CFA, a measurement model
with multiple indicators for each construct of interest is added to the
model. Each construct of interest is represented by a latent factor, and the
latent factors are connected via directional regression paths in the same
way as in a manifest path analysis. (This is in contrast to CFA in which
only nondirectional relationships between factors are assumed, with the
exception of higher-­order factor models; see Section 3.4.2.)
Figure 3.15 shows a path diagram of a latent path analysis with three
latent variables (h1, h2, and h3), each of which is measured with two indi-
cators (Y11–Y23), respectively. Substantively, the path model again corre-
sponds to the theoretical model of Whitelaw and Liang (1991); however,
this time, the mediator model is specified at the level of latent variables
(i.e., at the level of the h-variables). A latent path analysis has the advan-
tage over manifest path analysis in that it takes random measurement
error in the observed variables into account when estimating direct, indi-
rect, and total effects between the constructs of interest.
 74   Data Analysis with Mplus

β 31
ζ2 ζ3
η1 β 21 η2 β 32 η3
Phys. Health Funct. Health Subj. Health
(PH) (FH) (SH)
λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

Health problems Consultations Present Change
(FH1) (FH2)
(SICK) (CONSULT) (SHP) (SHC)

ε11 ε21 ε12 ε22 ε13 ε23

FIGURE 3.15. Latent path model with three latent variables h1, h2, and h3, each
of which is measured with two manifest variables Yij. The variable names used in
the Mplus analyses appear in parentheses. The constants lij indicate factor loadings.
The variables eij are measurement error variables. The constants bjj′ indicate latent
regression (path) coefficients. The zj variables are latent residual variables. At the
level of the latent variables, the model contains the indirect effect b21 · b32.

3.5.4 Latent Path Analysis in Mplus

Each of the three constructs in our example is measured by two indica-
tors (manifest variables): physical health is measured by the number of
diseases (variable SICK) as well as the number of consultations of doctors
in the last 12 months (variable CONSULT); functional health is measured
by the two test halves created for the SF-36 questionnaire (variables FH1
and FH2); and subjective health is measured by two single items, one of
which refers to the subjective health in the present (SHP), whereas the
second item refers to changes in subjective health across the last 6 years
(variables SHC).
The three h-variables represent the latent scores with regard to the
three constructs of physical health, functional health, and subjective
health, which are corrected for random measurement error as well as for
item- and scale-­specific influences. The theoretical model by Whitelaw
and Liang (1991) can therefore now be examined at the level of latent vari-
ables, which are corrected for measurement error and indicator-­specific
influences. Figure 3.16 shows the Mplus input file for the estimation of the
parameters of the latent path analysis represented in Figure 3.15.
 Linear Structural Equation Models   75

FIGURE 3.16. Mplus input file for the latent path model from Figure 3.15.

In the latent path model, we first specify the measurement model for
the latent factors as in a CFA:

SH by SHP SHC;
PH by SICK CONSULT;
FH by FH1 FH2;

The latent factor SH represents subjective health, the factor PH refers

to physical health, and the factor FH represents functional health. In the
next step, the structural model is specified—­which, in our case, is the
path analysis at the latent level:

FH on PH;
SH on PH FH;

Note that SH, PH, and FH are now latent variables that are not part of
our data set and that therefore do not have to be listed under variable:
names. The subcommand for obtaining a summary of all direct, indirect,
76   Data Analysis with Mplus

and total effect remains the same as in the manifest path model (but now
refers to the latent variables):

model indirect: SH ind PH;

In the output, we see under MODEL FIT INFORMATION that the

latent path model—in contrast to the manifest model—is not saturated,
but has 6 degrees of freedom. In this case, all 6 degrees of freedom are
coming out of the measurement model (due to using more than one indi-
cator for each construct). The structural model (the path analysis at the
latent level) is saturated in this model as well. That is, no testable restric-
tions arise from the structural model. The model fit indices indicate that
the latent path model fits the data well.

MODEL FIT INFORMATION

Chi-Square Test of Model Fit

Value 5.509
Degrees of Freedom 6
P-Value 0.4804

CFI/TLI

CFI 1.000
TLI 1.000

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.000
90 Percent C.I. 0.000 0.031
Probability RMSEA <= .05 1.000

SRMR (Standardized Root Mean Square Residual)

Value 0.007

With regard to the parameter estimates (see table MODEL RESULTS),

we can first of all see that all freely estimated factor loadings are sig-
nificantly different from zero (the first unstandardized loading is fixed
to one for each factor, respectively, to identify the metric of the latent
variables as the default in Mplus). In addition, all standardized loadings
(see STANDARDIZED MODEL RESULTS [STDYX Standardiza-
tion]) are substantial (range: 0.494–0.972). The indicators of the FH
(functional health) factor show particularly high standardized loadings
 Linear Structural Equation Models   77

(lˆ 12
standardized
= 0.972; lˆ 22
standardized
= 0.955). This finding shows that
the indicators of this factor are very homogeneous (the two test halves
seem to measure very similar aspects of functional health) and are highly
reliable.
The factor for physical health (PH) has the smallest standardized
loadings (lˆ 11standardized
= 0.623, lˆ 21
standardized
= 0.494). This finding can be
explained by the fact that even though the frequency of health problems
is positively correlated with the number of physician consultations, this
correlation is rather moderate in size in the present data set (r = .308,
p < .001; see full output on the companion website). Therefore, the two
indicators of physical health can be seen as rather heterogeneous (they
measure different facets of the construct physical health). The variable
SICK shows a somewhat larger standardized loading on the factor than
the number of physician consultations and might therefore be seen as the
marker variable for this factor. (Marker variables are indicators that show
high loadings on the intended factor and are therefore seen as most rel-
evant for the interpretation of this factor.)
With regard to the structural model, the relationships between the
three constructs go in the same direction: physical health problems neg-
atively impact both functional health ( bˆ 21 = –0.246, z = –13.189, p <
.001, bˆ 21
standardized
= –0.722) and subjective health ( bˆ 31 = –0.379, z =
–6.944, p < .001, bˆ standardized
31 = –0.709), whereas functional health is
positively related to subjective health ( bˆ 32 = 0.314, z = 2.631, p = .009,
bˆ standardized
32 = –0.200).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

SH BY
SHP 1.000 0.000 999.000 999.000
SHC 0.710 0.031 23.145 0.000

PH BY
SICK 1.000 0.000 999.000 999.000
CONSULT 5.547 0.363 15.265 0.000

FH BY
FH1 1.000 0.000 999.000 999.000
FH2 0.981 0.014 68.107 0.000

FH ON
PH -0.246 0.019 -13.190 0.000
78   Data Analysis with Mplus

SH ON
PH -0.379 0.055 -6.945 0.000
FH 0.314 0.119 2.632 0.008

Intercepts
SHP 3.525 0.021 168.751 0.000
SHC 2.521 0.020 128.763 0.000
SICK 2.706 0.054 50.435 0.000
CONSULT 13.082 0.375 34.892 0.000
FH1 2.646 0.012 226.225 0.000
FH2 2.709 0.012 232.059 0.000

Variances
PH 1.839 0.179 10.277 0.000

Residual Variances
SHP 0.194 0.019 10.359 0.000
SHC 0.367 0.016 23.627 0.000
SICK 2.906 0.159 18.304 0.000
CONSULT 175.061 7.166 24.429 0.000
FH1 0.012 0.002 5.135 0.000
FH2 0.020 0.002 8.231 0.000
SH 0.132 0.026 5.115 0.000
FH 0.102 0.009 11.015 0.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

SH BY
SHP 0.854 0.016 54.333 0.000
SHC 0.647 0.018 36.103 0.000

PH BY
SICK 0.623 0.026 24.309 0.000
CONSULT 0.494 0.025 19.907 0.000

FH BY
FH1 0.972 0.006 173.398 0.000
FH2 0.955 0.006 165.700 0.000

FH ON
PH -0.722 0.030 -24.379 0.000

SH ON
PH -0.709 0.080 -8.836 0.000
FH 0.200 0.076 2.631 0.009

Intercepts
SHP 4.157 0.076 54.352 0.000
 Linear Structural Equation Models   79

SHC 3.172 0.060 52.435 0.000

SICK 1.242 0.033 37.891 0.000
CONSULT 0.860 0.029 29.817 0.000
FH1 5.573 0.100 55.647 0.000
FH2 5.716 0.103 55.731 0.000

Variances
PH 1.000 0.000 999.000 999.000

Residual Variances
SHP 0.270 0.027 10.062 0.000
SHC 0.581 0.023 25.043 0.000
SICK 0.612 0.032 19.203 0.000
CONSULT 0.756 0.025 30.786 0.000
FH1 0.055 0.011 5.063 0.000
FH2 0.088 0.011 7.966 0.000
SH 0.251 0.045 5.549 0.000
FH 0.479 0.043 11.219 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

SHP 0.730 0.027 27.166 0.000

SHC 0.419 0.023 18.052 0.000
SICK 0.388 0.032 12.155 0.000
CONSULT 0.244 0.025 9.953 0.000
FH1 0.945 0.011 86.699 0.000
FH2 0.912 0.011 82.850 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

SH 0.749 0.045 16.517 0.000

FH 0.521 0.043 12.189 0.000

It is noteworthy that, in the latent model, the relationships between

the constructs are much stronger than in the manifest model. This can be
seen, in particular, from the latent variable R 2 values. According to these
values, the latent model allows us to account for approximately 74.9% of
the variability in subjective health and about 52.1% of the variability in
functional health. In the manifest model, the corresponding proportions
of variability accounted for were only 39.9% and 20%, respectively. There
are two explanations for this rather large discrepancy. First, the latent
analysis takes measurement error into account, which is not the case in
the manifest model. The correction for measurement error leads to a disat-
tenuation of the relationships at the latent level. Second, specific variance
80   Data Analysis with Mplus

in the indicators (due to indicator heterogeneity) also contributes to the

different estimates in the manifest and latent models. As seen above, the
indicators of physical health are rather heterogeneous, as shown by the
rather modest correlation between those indicators. In addition, the items
used as indicators of subjective health are only moderately correlated
(r = .553). These indicators therefore also show relatively large specific
variance components that are treated as part of the error variables in the
latent model.
In the manifest path model we did not use the variable SHC as indi-
cator of subjective health, and we did not use the variable CONSULT as
indicator of physical health. Therefore, the constructs of subjective health
and physical health were defined in a narrower way in the manifest than
in the latent model, and specific variance not shared with the other indi-
cators used in the latent model could not be separated from variance that
is shared with these indicators. This issue is probably not relevant for the
indicators of functional health, because FH1 and FH2 are test halves cre-
ated from the same subscale and are strongly correlated (r = .928), indicat-
ing that there is little specific variance in each test half.
In summary, in the present example there is a substantial amount
of indicator specificity related to the indicators of subjective health and
physical health. Removal of specific variance is partly responsible for the
stronger relationships seen in the latent model compared to the mani-
fest model because indicator-­specific variance is separated from com-
mon variance and treated as part of the error. This means that the results
should be interpreted with great caution in this case because the large
amounts of specific variance may have inflated the relationships between
the constructs. Hence, one should think carefully about whether the pres-
ent operationalization of the latent variables for subjective and physical
health is appropriate. This is a theoretical rather than a statistical problem
that the user needs to solve based on substantive considerations.
The additional output obtained based on the model indirect
command is not shown here in detail, but is available on the compan-
ion website. This output again provides us with detailed information on
the size and statistical significance of the indirect effect in our model.
In addition—­and analogous to the manifest path model—we may again
request bootstrap confidence intervals for testing the indirect effect. The
principle is the same as for the manifest path model and is therefore not
shown here in detail. The relevant input and output files can be found on
the companion website.
4

Structural Equation Models

for Measuring Variability
and Change

Because of their high flexibility in modeling complex relationships

between a large number of variables, and because of their ability to
explicitly account for measurement error, linear SEMs are also very useful
for analyzing longitudinal data (data obtained from repeated-­measures
designs). In the following sections we discuss the specification and analy-
sis of five different types of longitudinal SEMs in Mplus. We distinguish
between models for analyzing variability (variability models) and mod-
els for analyzing change (change models; Eid, Courvoisier, & Lischetzke,
2011). Variability models are used to model situation-­specific fluctuations
in scores around a stable set point or trait value. In contrast, change models
are often used when the goal is to measure more long-­lasting and poten-
tially irreversible changes of psychological constructs over time.
With regard to variability models, we discuss latent state (LS) and
latent state–trait (LST) models (Steyer et al., 1992, 1999) in Sections 4.1
and 4.2. In Section 4.1, we also discuss general issues relevant to longitu-
dinal data analysis using SEMs; for example, the modeling of indicator-­
specific effects in multiple-­indicator models and the issue of testing mea-
surement invariance across time. Subsequently, we discuss three types of
change models commonly used in the literature: (1) manifest and latent
81
82   Data Analysis with Mplus

autoregressive models (Hertzog & Nesselroade, 1987; Jöreskog, 1979a,

1979b; see Section 4.3), (2) latent change (LC) models (McArdle, 1988;
Raykov, 1993; Steyer et al., 1997; Steyer, Partchev, & Shanahan, 2000;
see Section 4.4), and (3) latent growth curve models (LGCMs; Bollen &
­Curran, 2006; Duncan et al., 2006; see Section 4.5).
The main goal of the analysis of variability models is to determine
the degree to which individual differences on an attribute or construct
depend on situation- or occasion-­specific effects versus stable person-­
specific effects. Following Steyer and colleagues (Steyer et al., 1992, 1999),
we distinguish between a trait component (a component that character-
izes stable person-­specific effects), a state residual component (a compo-
nent that characterizes effects of the situation and/or person × situation
interactions), and a random measurement error component. We therefore
speak of LST models (Steyer et al., 1992, 1999; see Section 4.2). For exam-
ple, it is often of interest to which degree measurements (e.g., question-
naire or test scores) measure aspects of the situation or occasion of mea-
surement (e.g., anxiety in a specific situation) versus stable personality
dispositions (traits; e.g., trait anxiety = general anxiety of a person across
situations) and random measurement error. These questions concern the
temporal stability, reliability, and validity of the measurement of individ-
ual differences. LST models allow us to determine the degree of occasion
specificity, consistency, and reliability through the calculation of variance
components (Steyer et al., 1992, 1999).

4.1 LS Analysis

In our introduction to longitudinal SEMs, we first consider different ver-

sions of the less complex LS models (see Figure 4.1 as well as Steyer et al.,
1992). LS models can be seen as “precursors” of LST models and represent
general CFA models for the analysis of longitudinal data (CFA models
for cross-­sectional data were discussed in Section 3.4). In LS models, no
separation of stable from occasion-­specific variance components is pos-
sible. Nonetheless, LS models are frequently used in practice for analyzing
longitudinal data, partly because they are rather simple and unrestrictive.
The analysis of an LS model requires that multiple (and preferably the
same) indicators are available at each time point to measure the construct
of interest (e.g., depression). In an LS model, each observed variable Yik (i
= indicator, k = occasion of measurement) is decomposed into a constant
 Structural Equation Models for Measuring Variability and Change   83

State 1: State 2: State 3:

Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

FIGURE 4.1. Latent state (LS) model for depression measured on three occasions
of measurement (T1–T3). Yik = observed variable i measured at time point k; eik =
measurement error variable; lik = factor loading. The variable names of the indicators
used in Mplus appear in parentheses.

BOX 4.1. Example of LS Analysis

We use data from Proyecto: La Familia (The Family Project; Roosa et al., 2008),
which assessed the health and adjustment of Mexican American children
and their parents every 2 years, beginning in the fifth grade for the target
children.* In the applications presented in this chapter, we fit longitudinal
SEMs to three waves of fathers’ (N = 323 complete cases) ratings of their own
depressive symptoms, using the 20-item Center for Epidemiological Stud-
ies Depression scale (CES-D; Radloff, 1977). Sample items include “You felt
depressed” and “You felt that everything you did was an effort.” Items were
answered on a 4-point scale ranging from “Rarely or none of the time” to
“Most or all of the time.” For the purpose of the present illustration, the items
of this questionnaire were aggregated to two test halves, so that there would
be two indicators of depression at each time point. The data set depression.
dat contains the summary data (means, standard deviations, and correla-
tions) for the two test halves measured on three occasions of measurement
(variables d11, d21, d12, d22, d13, and d23).

*I would like to thank Mark Roosa for providing the data set for the analyses.
84   Data Analysis with Mplus

intercept term aik, an occasion-­specific latent state variable (State k), and
an occasion-­specific measurement error (residual) variable eik:

Yik = aik + lik · (State k) + eik

where lik denotes a factor loading. Figure 4.1 shows an LS model for
depression, measured with two variables Y1k and Y2k on three occasions
of measurement (k = 1, 2, 3).
Figure 4.1 shows that all latent state variables can be correlated. These
correlations reflect the “true” stability of individual differences across time
(i.e., stability corrected for measurement error). The correlations between
the state variables allow us to examine whether a construct is better
described as a temporally stable disposition (“trait”) or as an attribute that
is prone to substantial occasion-­specific fluctuations (a more state-like
construct). Correlations between state factors that approach 1 indicate
high stability of individual differences and therefore the construct can be
seen as rather trait-like (e.g., intelligence). In contrast, moderate to small
correlations between LS factors indicate that a construct depends strongly
on the situation (i.e., is more state-like; e.g., anger, mood states, cortisol
levels) or that some individuals changed more than others over time.
In addition, the correlations between the state factors also reveal
whether an autoregressive process is present in the data. This is the case
if the state factor correlations decline with increasing temporal distance
between the measurements. Latent autoregressive models, which can be
seen as special cases of LS models, are discussed in Section 4.3.
LS analyses typically do not involve any specific assumptions as to
the structure of the correlations between the state factors (as is the case,
e.g., in LST and latent autoregressive models; see below). This means that
all possible correlations between LS variables are usually estimated as
free, unconstrained parameters. This implies that the structural model of
LS models is typically saturated, meaning that no testable restrictions can
be derived from the structural model and therefore no degrees of freedom
are obtained from this part of the model. This aspect makes LS models
rather unrestrictive, which is one reason why they are often useful as a
starting point for a longitudinal statistical analysis. Causes of potential
misfit that may occur in more complex longitudinal SEMs are more easily
identified in the LS model, because in an LS model misfit can be caused
only by a misspecified measurement model (the structural model is satu-
rated and hence does not contribute to any global misfit of the model).
 Structural Equation Models for Measuring Variability and Change   85

4.1.1 LS versus LST Models

In LS models—­in contrast to LST models, which are discussed in Section
4.2—only LS and measurement error variables are modeled. The LS vari-
ables in LS models are not further decomposed into trait and state residual
components as in LST models (Steyer et al., 1992). Therefore, LS models
do not allow for a separation of stable from occasion-­specific components
of variance.
Despite this limitation, LS models are useful for longitudinal data
analysis. They serve as good baseline models not only in an LST analysis,
but also in other types of longitudinal analyses. The reason is that many
other types of longitudinal models are either reformulations (e.g., as LC
models; see Section 4.4) or special cases of the LS model (e.g., as latent
autoregressive models and second-­order LGCMs; see Sections 4.3 and 4.5,
respectively), some of which make more restrictive assumptions than the
LS model (e.g., autoregressive models and LGCMs).
LS models allow the testing of basic aspects of longitudinal measure-
ment models (e.g., the question of whether indicators used to measure
a construct are homogeneous, whether the factor structure is invariant
across time, and whether measurement invariance holds across time)
before additional restrictive assumptions with regard to the LS factors are
introduced (e.g., regarding a state–trait, autoregressive, or specific growth
structure). Different options for modeling indicator-­specific effects (situ-
ations where indicators are not perfectly homogeneous) are presented in
Section 4.1.3. The issue of testing measurement invariance across time is
discussed in Section 4.1.4. In Section 4.2, we discuss the extension of an
LS model to an LST model.

4.1.2 Analysis of LS Models in Mplus

In the first step of an LS analysis, it is often a good idea to start with a rather
unrestrictive model, which, for example, does not make any assumptions
about time-­invariant factor loadings, etc. An important question that
should be clarified at the beginning is whether the basic hypothesized fac-
tor structure holds across time and whether indicators are homogeneous
(or whether so-­called indicator-­specific effects are present).
Indicator specificity refers to method effects of indicators that become
apparent in longitudinal analyses because the same indicators are repeat-
edly measured (Raffalovich & Bohrnstedt, 1987). An indicator that shares
specific variance (i.e., variance that is not shared with the other indicators)
86   Data Analysis with Mplus

with itself over time violates the assumption of uncorrelated residual

variables made in the standard LS model. The reason is that, due to the
presence of indicator-­specific effects, the indicator would be more highly
correlated with itself over time than with other indicators—­v iolating the
assumption of indicator homogeneity made in an LS model with uncor-
related residual variables. In other words, the standard LS model with
uncorrelated residual variables would, in this case, be too restrictive, and
the researcher would have to think of possibilities to adequately account
for indicator-­specific effects.
In the following, I first show the specification of a simple LS model in
Mplus that does not account for indicator-­specific effects. As we will see,
this model shows a somewhat unsatisfactory fit in the present example
due to the presence of indicator-­specific effects. Subsequently, I demon-
strate how the simple LS model can be extended to take indicator-­specific
effects into account. Figure 4.2 shows the Mplus input file for the estima-
tion of the simple LS model without indicator-­specific effects from Figure
4.1, based on the two repeatedly measured depression test halves. Using
the subcommands

FIGURE 4.2. Mplus input file for the specification of a simple LS model for three
measurement occasions without measurement invariance restrictions.
 Structural Equation Models for Measuring Variability and Change   87

state1 by d11
d21;
state2 by d12
d22;
state3 by d13
d23;

the indicators are assigned to the relevant LS factors. This command is

analogous to the by statements in a conventional cross-­sectional CFA
model (see Section 3.4).
In the following, the model fit criteria reported in the Mplus output
are shown for the simple LS model:

MODEL FIT INFORMATION

Chi-Square Test of Model Fit

Value 18.474
Degrees of Freedom 6
P-Value 0.0052

BOX 4.2. Mplus Default for the Mean Structure

In our basic LS model, we do not include latent means for the state factors
and instead make use of the Mplus default settings, according to which all
latent variable means are set to zero and all observed variable intercepts
(aik) are freely estimated. To the degree that no restrictions are imposed on
the intercepts of the observed variables, the intercepts will be identical to
the observed variable means [expected values E(Yik)]. This follows from the
decomposition of the observed variable means in the LS model, because the
LS means are set to zero by default in Mplus and the residual variables eik
have means of zero by definition (see Steyer et al., 1992):

E(Yik) = E(aik) + E(lik · State k) + E(eik)

= aik

In Section 4.1.4, I demonstrate how the means of the LS variables (that are
often of interest in longitudinal studies) can be identified and estimated in
Mplus.
88   Data Analysis with Mplus

Information Criteria

Number of Free Parameters 21

Akaike (AIC) 1088.313
Bayesian (BIC) 1167.643
Sample-Size Adjusted BIC 1101.034
(n* = (n + 2) / 24)

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.080
90 Percent C.I. 0.040 0.123
Probability RMSEA <= .05 0.098

CFI/TLI

CFI 0.990
TLI 0.976

SRMR (Standardized Root Mean Square Residual)

Value 0.012

It can be seen that the model shows a somewhat unsatisfactory fit, which
is indicated, in particular, by the significant chi-­square test (c2 = 18.474,
df = 6, p = .0052) and the high RMSEA coefficient (.08), whereas the CFI
(0.99) and SRMR (0.012) coefficients show acceptable values, according
to commonly used criteria (see Box 3.7). A bad model fit in longitudinal
SEMs with multiple indicators can often be explained by the presence of
indicator-­specific effects, which are not sufficiently accounted for in the
simple LS model.

4.1.3 Modeling Indicator‑Specific Effects

One possibility for modeling indicator-­specific effects is to allow for cor-
relations between the measurement error (residual) variables that pertain
to the same indicator over time (e.g., Sörbom, 1975; see Figure 4.3). This
is often referred to as an LS model with autocorrelated error variables (or—in
the context of multitrait–­multimethod analysis—­as a correlated uniqueness
model; e.g., see Cole & Maxwell, 2003; Lance, Noble, & Scullen, 2002).
Given that the estimation of error correlations is relatively straight-
forward, this approach to accounting for indicator-­specific effects over
time is popular among longitudinal researchers. A disadvantage of this
method, however, is that it leaves indicator-­specific effects confounded
 Structural Equation Models for Measuring Variability and Change   89

State 1: State 2: State 3:

Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

FIGURE 4.3. LS model with autocorrelated error variables for modeling indicator-­
specific effects. This model is underidentified in the present application because
there are only two indicators per time point.

with random measurement error. Therefore, indicator-­specific effects as

a reliable source of individual differences are confounded with random
error variance. As a consequence, the reliabilities of the indicators tend
to be underestimated in models with correlated residuals (e.g., Geiser
& Lockhart, 2012). An alternative way to account for indicator-­specific
effects is to include additional latent variables instead of correlated resid-
ual variables in the model—so-­called indicator-­specific (or method) factors.
I first show the specification of correlated residuals in Mplus and then an
alternative model version with an indicator-­specific factor.
In our example with only two indicators per LS factor and only three
time points, an LS model that allows all possible autocorrelations between
residual variables is not identified (due to a lack of information) and there-
fore cannot be estimated (for a more detailed discussion of the issue of
model identification in SEMs, see Bollen, 1989). In order to estimate a
model with all possible autocorrelations between error variables, more
indicators per time point and/or more than three time points are required.
In order to illustrate the principle of specifying correlated error
variables in Mplus, we choose a slightly more restrictive version of the
LS model, with autocorrelated error variables in which only the error
90   Data Analysis with Mplus

correlations between adjacent time points are included (see Figure 4.4).
This model version is identified for the present data and the model param-
eters can thus be estimated. The relevant Mplus syntax to specify this
model is shown in Figure 4.5.
The model fit statistics for this model are slightly better compared
to the LS model without correlated residual variables, c2 = 8.536, df =
2, p = .014; RMSEA = .101, CFI = 0.995, SRMR = 0.008, except for the
higher RMSEA value. Nonetheless, the model with autocorrelated residu-
als between adjacent time points also shows a significant chi-­square value,
indicating a significant discrepancy between the model and the data. In
addition, the RMSEA value is even higher than for the original model
(presumably due to the loss in degrees of freedom through the estimation
of four additional parameters). These results might suggest that indicator-­
specific effects generalize beyond adjacent time points, so that the current
model (which only allows for correlated errors of adjacent time points)
does not sufficiently account for these effects.
Eid, Schneider, and Schwenkmezger (1999; see also Eid, 2000)
proposed an approach for modeling indicator-­specific effects that does
not require the estimation of correlated error variables and thus allows

State 1: State 2: State 3:

Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

FIGURE 4.4. LS model with autocorrelated error variables for modeling indicator-­
specific effects. In contrast to the model in Figure 4.3, error correlations are admit-
ted only between adjacent measurement occasions. This model is identified in the
present application.
 Structural Equation Models for Measuring Variability and Change   91

FIGURE 4.5. Mplus model command for the specification of the LS model with
autocorrelated error variables between adjacent measurement occasions shown in
Figure 4.4.

researchers to avoid the problem of underestimated reliabilities of indica-

tors. In this approach, one indicator is chosen as comparison standard
(i.e., a reference indicator). The reference indicator could, for example, be
a marker variable that best represents the construct—­either theoretically
or empirically. For simplicity, we assume in the following that the first
indicator (Y1k) is selected as the reference indicator. However, in principle,
any other indicator could serve as reference indicator as well.
The reference indicators Y1k load only on the LS factors, whereas the
remaining (non-­reference) indicators Yik, i ≠ 1, have loadings on both the
state factors and a so-­called indicator-­specific (or method) factor (IS2; see
Figure 4.6). The indicator-­specific factor reflects the specificity of the non-­
reference indicators Yik, i ≠ 1, relative to the reference indicators Y1k. In
the present case, we therefore need only one indicator-­specific factor (i.e.,
for the second indicator Y2k). In general, if I indicates the total number of
indicators, then I – 1 indicator-­specific factors are needed to account for
indicator-­specific effects in LS models.
The indicator-­specific factors are defined as residual factors with
regard to the state variables pertaining to the reference indicator (Eid,
2000; Eid et al., 1999; Geiser, 2009). The indicator-­specific factors there-
fore represent that part of the reliable variance of the non-­reference indi-
cators that is not shared with the reference indicators. It follows from this
definition that the mean of each indicator-­specific factor is zero and that
the means are uncorrelated with all LS factors that pertain to the same
construct. The latter aspect is particularly relevant for the specification
92   Data Analysis with Mplus

State 1: State 2: State 3:

Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

λIS22 λIS23
λIS21

IS2

FIGURE 4.6. LS model with an indicator-­specific (method) factor (IS2) for the sec-
ond indicator. The parameters lIS2k indicate factor loadings on the indicator-­specific
factor. The indicator-­specific factor is defined as a residual factor with respect to the
LS factors and is therefore by definition uncorrelated with all state factors that per-
tain to the same construct.

in Mplus, because we need to tell Mplus explicitly which factors cannot

be correlated. (The mean of all factors is assumed to be zero by default in
Mplus [see Box 4.2] so that the zero mean of the indicator-­specific factor
is not a specific issue in Mplus.)
Note that the meaning of the LS factors in Figure 4.6 is now deter-
mined by the meaning of the reference indicator (Y1k). In our case, for
example, the meaning of State 1 depends on the meaning of indicator
d11. For this reason, as mentioned above, researchers should choose a
reference variable that best represents the construct (a so-­called marker
variable). This step is particularly important when a researcher deals with
rather heterogeneous indicators that represent different facets of a con-
struct. We return to this issue below when we discuss the model results
in detail. Figure 4.7 shows the Mplus syntax for the specification of the LS
model with an indicator-­specific factor for the second indicator.
 Structural Equation Models for Measuring Variability and Change   93

The indicator-­specific factor (here named is2) is introduced analo-

gously to the state factors by using a by statement. Of importance, the
correlations of the factor is2 and the state factors state1–state3 (that
Mplus would try to estimate as the default) have to be set to zero explicitly,
because is2 is defined as a residual factor with respect to all three state
factors. The following command is used to set these correlations to zero:

is2 with state1@0 state2@0 state3@0;

In the output, we can see that the LS model with indicator-­specific

factor shows a better fit than either the simple LS model or the LS model
with autocorrelated residual variables between adjacent time points. The
chi-­square value for the LS model with indicator-­specific factors is smaller
and statistically nonsignificant, c2 = 3.569, df = 3, p = .3119. Furthermore,
RMSEA (.024) and SRMR (0.008) also indicate an improved fit. In addi-
tion, the AIC value for the model with indicator-­specific factor (1079.408)
is lower than the AIC values for both the simple LS model (1088.313) and
the LS model with autocorrelated residual variables (1086.375) between
adjacent time points. We therefore only show the results for the well-­
fitting model with an indicator-­specific factor for the second indicator.
The complete outputs for the two previous LS models can be found on the
companion website.
As usual, Mplus first outputs the unstandardized model parameter
estimates. It can be seen that the unstandardized state factor loadings
of the second indicator are slightly lower than the loadings of the first

FIGURE 4.7. Mplus model command for the specification of the LS model with
one indicator-­specific factor (is2) for the second indicator shown in Figure 4.6.
94   Data Analysis with Mplus

indicator (the unstandardized loadings of the first indicator have been

fixed to 1 for identification as the default). This may indicate that the sec-
ond indicator does not discriminate as well as the first indicator between
individuals with lower versus higher scores on the LS. In other words,
a one-unit change in each LS factor score is associated with a slightly
smaller expected increase in the score on the second indicator than in the
score on the first one.
In addition, we see that the state loadings of the second indicator are
estimated to very similar values at each time point. This already indicates
that the assumption of time-­invariant loadings is probably tenable in this
case (a formal test of this assumption is discussed in Section 4.1.4). Fur-
thermore, we can see that the two freely estimated factor loadings on the
indicator-­specific factor is2 are not significantly different from zero. This
indicates that indicator-­specific effects are rather weak in this example,
showing that the indicators are rather homogeneous. The standardized
loadings (see output STANDARDIZED MODEL RESULTS [STDYX
Standardization]), however, are significantly different from zero
and range between .157 and .278.
In practice, it is very important to closely examine the degree of het-
erogeneity of the indicators (i.e., strength of indicator-­specific effects).
If the standardized loadings on the indicator-­specific factors are large
(e.g., say, .5 or larger), this finding would indicate that 25% or more of
observed individual differences are due to indicator specificity (i.e., the
squared standardized loadings indicate the proportion of variability in
the observed variables that is accounted for by the relevant factor). This
would mean that the indicators are rather heterogeneous and mirror at
least somewhat different facets of a construct.
In such cases, one should carefully think about which indicator best
serves as the reference indicator, because the substantive meaning and
interpretation of the LS factors depend on the choice of the reference
indicator in the model with I – 1 indicator-­specific factors. For example,
if indicators are rather heterogeneous, correlations of state factors with
external variables might vary substantially for different reference indica-
tors. In addition, different indicators may show differential mean changes
over time.
If strong indicator-­specific effects are present in a study, primarily
theoretical/substantive criteria should be used to decide which indicator
best represents the construct of interest, and the corresponding indicator
should be selected as the reference indicator to ensure that the state fac-
tors represent the intended construct (for a more detailed discussion of
 Structural Equation Models for Measuring Variability and Change   95

this issue, see Geiser, Eid, & Nussbeck, 2008). In addition, researchers
should carry out careful sensitivity analyses to examine in which way the
choice of different reference indicators does, or does not, affect the results
obtained for the structural model (e.g., intercorrelation of the state factors,
relationships of state factors to external variables, LS factor means and
variances).
In our example, indicator-­specific effects are minor, as indicated by
the small standardized loadings on the is2 factor, which, as mentioned
above, range from .157 to .278. Therefore, at most .2782 = 7.7% of the
observed variability is due to indicator-­specific effects, which can be seen
as unproblematic. From a substantive point of view this suggests that the
indicators are rather homogeneous (i.e., measure, to a large extent, the
same true score) and have only small indicator-­specific variance compo-
nents.
In addition, we see that all covariances between the indicator-­specific
factor and the three state factors were set to zero as requested. Therefore,
the is2 factor can be interpreted as a residual factor. In contrast, the
covariances among the state factors were freely estimated. All state covari-
ances are significantly different from zero (8.354 ≤ z ≤ 7.586). Hence,
a significant portion of individual differences with regard to depression
remained stable across time. The exact amount of stability, however, is dif-
ficult to assess based on the covariances, because covariances are unstan-
dardized measures of association. For this reason, researchers typically
report the correlations between the state factors as measures of stability
because correlations are standardized measures of association that do not
depend on the metric of the variables and are therefore easier to interpret
(see below).
In the unstandardized solution, we furthermore obtain the estimated
intercepts, factor variances, and residual variances. The unstandardized
intercepts are, in this case, identical to the observed variable means,
because the present model does not impose any restrictions on the mean
structure of the variables (cf. Box 4.2).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 1.000 0.000 999.000 999.000
D21 0.700 0.049 14.221 0.000
96   Data Analysis with Mplus

STATE2 BY
D12 1.000 0.000 999.000 999.000
D22 0.749 0.043 17.342 0.000

STATE3 BY
D13 1.000 0.000 999.000 999.000
D23 0.755 0.047 16.192 0.000

IS2 BY
D21 1.000 0.000 999.000 999.000
D22 0.623 0.378 1.650 0.099
D23 0.793 0.507 1.564 0.118

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.120 0.015 7.917 0.000

STATE3 WITH
STATE1 0.110 0.014 7.586 0.000
STATE2 0.128 0.015 8.354 0.000

Intercepts
D11 1.584 0.027 58.307 0.000
D21 1.354 0.021 65.148 0.000
D12 1.642 0.028 58.527 0.000
D22 1.389 0.023 60.772 0.000
D13 1.631 0.027 60.284 0.000
D23 1.402 0.023 62.159 0.000

Variances
STATE1 0.205 0.022 9.287 0.000
STATE2 0.230 0.023 10.185 0.000
STATE3 0.214 0.021 10.012 0.000
IS2 0.011 0.008 1.402 0.161

Residual Variances
D11 0.033 0.012 2.717 0.007
D21 0.028 0.009 3.134 0.002
D12 0.025 0.011 2.284 0.022
D22 0.036 0.007 5.326 0.000
D13 0.022 0.011 2.028 0.043
D23 0.035 0.008 4.692 0.000

Based on the completely standardized solution (STANDARDIZED

MODEL RESULTS [STDYX Standardization]), in which all
 Structural Equation Models for Measuring Variability and Change   97

observed and all latent variables are standardized, we can see that the
observed variables have rather high loadings on the state factors (≥ .849)
and rather small loadings on the indicator-­specific factor (largest loading
= .278; see discussion above). The standardized loadings can again be
interpreted as the estimated correlations between each indicator and the
corresponding latent variable, because the state factors are uncorrelated
with the indicator-­specific factor. Based on the indicators’ high standard-
ized loadings on the state factors and the low loadings on the is2 factor,
we can conclude that the indicators are highly reliable and rather homo-
geneous.
The standardized solution also provides us with the estimated latent
correlations between the state factors. In this example, the states show
moderately high correlations across time (.523 ≤ r ≤ .578), which indicates
a moderate degree of stability of individual differences over time. Note
that adjacent state factors (State 1 and State 2 as well as State 2 and State
3) are slightly more highly correlated than the more “distant” state factors
State 1 and State 3. This finding indicates the presence of a small autore-
gressive effect (cf. Section 4.3).
Of interest are also the estimated R 2 values for the observed variables,
which are reported at the end of the standardized solution. These values
can be used as estimates of the reliabilities of the indicators. As we already
saw based on the high standardized loadings, all six variables show high
reliabilities (Rel), which range between .785 and .907. This shows that the
six indicators of depression (here: questionnaire scales) allow for a high
precision of measurement (small error).

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 0.928 0.028 32.992 0.000
D21 0.849 0.029 28.934 0.000

STATE2 BY
D12 0.950 0.023 42.147 0.000
D22 0.874 0.024 36.260 0.000

STATE3 BY
D13 0.952 0.024 39.012 0.000
D23 0.862 0.026 33.195 0.000
98   Data Analysis with Mplus

IS2 BY
D21 0.278 0.099 2.810 0.005
D22 0.157 0.062 2.551 0.011
D23 0.203 0.076 2.687 0.007

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.552 0.045 12.293 0.000

STATE3 WITH
STATE1 0.523 0.047 11.194 0.000
STATE2 0.578 0.043 13.524 0.000

Intercepts
D11 3.244 0.139 23.299 0.000
D21 3.625 0.153 23.727 0.000
D12 3.257 0.140 23.313 0.000
D22 3.381 0.144 23.463 0.000
D13 3.354 0.143 23.420 0.000
D23 3.459 0.147 23.556 0.000

Variances
STATE1 1.000 0.000 999.000 999.000
STATE2 1.000 0.000 999.000 999.000
STATE3 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000

Residual Variances
D11 0.139 0.052 2.673 0.008
D21 0.202 0.066 3.063 0.002
D12 0.097 0.043 2.255 0.024
D22 0.212 0.042 5.063 0.000
D13 0.093 0.046 2.007 0.045
D23 0.215 0.048 4.501 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.861 0.052 16.496 0.000

D21 0.798 0.066 12.080 0.000
D12 0.903 0.043 21.073 0.000
D22 0.788 0.042 18.865 0.000
D13 0.907 0.046 19.506 0.000
D23 0.785 0.048 16.401 0.000
 Structural Equation Models for Measuring Variability and Change   99

The LS model with an indicator-­specific factor for the second indica-

tor showed a very good fit to the data and fit better than both a model that
did not account for indicator-­specific effects and a model with correlated
residual variables between adjacent time points. It therefore makes sense
to use the model with the indicator-­specific factor as the basis for tests of
measurement invariance across time (see next section) and other more
advanced longitudinal analyses. Had there not been any indicator-­specific
effects, we would have preferred the simpler model without the indicator-­
specific factor and would have used this simpler model as the basis for
further analysis.

4.1.4 Testing for Measurement Invariance across Time

In most longitudinal investigations the question of measurement invari-
ance or measurement equivalence across time is an important concern. Usu-
ally, researchers want to make sure that the indicators (e.g., items, scale
scores) that they used did not change their relationship to the latent vari-
ables in any critical way over the course of the study. The reason is that
changes in, for example, the factor loadings of the same indicator over
time may result in a different interpretation of an LS factor on different
occasions of measurement. This is typically not desirable, because it can
make across-­time comparisons of latent scores difficult. In particular, we
want to make sure that key parameters of our LS measurement model for
depression do not change over time so that the latent states are compa-
rable across time. If a sufficient degree of measurement invariance does
not hold, comparisons of, for example, LS factor means across time may
be uninterpretable. In this case, the LS variables may not represent the
same constructs, or the constructs may be measured on different scales
with different origins and/or different units of measurement, making it
hard to compare them across time.
SEMs for longitudinal data that are based on the same multiple indi-
cators at each time point (such as the LS model) allow us to formally
test whether, and to what degree, measurement invariance holds over
time, using model tests and fit indices. Specifically, the question of mea-
surement invariance refers to the time invariance of the measurement
model parameters—­that is, the factor loadings, intercepts, and residual
variances. The level of measurement invariance present in the data can
be determined by imposing equality constraints on these parameters
and checking whether these additional restrictions lead to a significant
decrease in model fit. An equality constraint is a constraint that forces a
100   Data Analysis with Mplus

parameter to be estimated to an equivalent value as another parameter

and is easily implemented in Mplus.
A minimum level of measurement invariance is typically desired
in longitudinal studies, because certain across-­time comparisons (e.g.,
examination of latent mean differences) depend on a certain level of mea-
surement invariance. The absence of measurement invariance can render
these comparisons impossible or at least difficult to interpret.
The degree of measurement invariance (see Box 4.3) that can be
assumed for the data at hand is often determined through statistical
model comparisons. In these comparisons, models with different levels of

BOX 4.3. Different Levels of Measurement Invariance

According to Widaman and Reise (1997)

Widaman and Reise (1997; see also Meredith, 1993; Meredith & Horn, 2001)
distinguish between the following levels of measurement invariance:

• Configural invariance is the weakest form of measurement invariance. Con-

figural invariance requires only the factor structure (the number of factors
and the loading pattern) to be constant across time. No specific parameters
have to be equal across time to satisfy this criterion.
• The next highest level, weak factorial invariance, requires the factor load-
ings lik to be the same over time for all indicators.
• The condition of strong factorial invariance is satisfied when, in addition to
the loadings, the intercepts aik of the indicators are also time-­invariant.
• Strict factorial invariance means that, in addition to loadings and intercepts,
indicator residual variances Var(eik) are constant across time.

Additional, even stricter, forms of parameter invariance such as, for example,
the assumption of equal LS factor variances and/or covariances across time,
can also be tested. In addition to the stability of the latent variance/covari-
ance structure, latent mean stability can be tested, which is often of inter-
est in longitudinal studies. A prerequisite for a meaningful interpretation
of latent mean change over time is that at least strong factorial invariance
can be assumed for the measurement model. Sometimes the assumption of
invariance in parameters across time can be made for only a subset of the
indicators, whereas other indicators show noninvariant parameters. In this
case, one speaks of partial measurement invariance. Byrne, Shavelson, and
Muthén (1989) discuss issues related to partial measurement invariance in
detail.
 Structural Equation Models for Measuring Variability and Change   101

measurement invariance are estimated and evaluated in terms of their rel-

ative fit compared to other more and less restrictive models. For example,
a configural invariance model (model without equality restrictions) can
be tested against a model in which the loadings are set equal across time
(weak factorial invariance model) and/or against a model with equal load-
ings and equal intercepts (strong factorial invariance model). The most
restrictive model that still shows an acceptable fit is selected, because it is
most parsimonious while still representing the data adequately.
One strategy to determine the degree of measurement invariance is
to specify more and more restrictive versions of the LS model until model
fit criteria (e.g., the chi-­square test of model fit) indicate that the latest set
of restrictions is no longer tenable for the data. The LS model version with
the most comprehensive set of restrictions that still shows an acceptable fit
is then selected. The most restricted model is chosen, because it typically
allows for an easier interpretation and comparability of certain parameters
and because it is more parsimonious (uses fewer parameters) than less
constrained model versions.
In our application in Section 4.1.3, we already estimated an LS model
with configural invariance. We therefore start our analysis of measure-
ment invariance with the next highest level of measurement invariance,
which is weak factorial invariance. In this model, only the factor loadings,
but not the intercepts and residual variances, are set equal across time.
In general, parameters can be set equal in Mplus by providing the
same number or parameter label in parentheses on the same input line
after parameters that should be held equal. Which particular number (or
label) is chosen is arbitrary. The user only has to make sure that the same
number is chosen for parameters to be set equal and that different numbers
are chosen for different groups of parameters that should be set equal,
respectively. (This point will become clearer below when we specify a
model with time-­invariant loadings and intercepts.) In our case, for the
weak factorial invariance model, only the loadings of the second indicator
have to be set equal, because the loadings of the first indicator are fixed
to 1 at each time point to set the scale of the LS factors and are therefore
already implicitly set equal across time.
We chose the number (1) to set the state factor loadings of the sec-
ond indicator equal across time:

state1 by d11
d21 (1);
102   Data Analysis with Mplus

state2 by d12
d22 (1);

state3 by d13
d23 (1);

It is important to note that the specification (1) does not mean that
the parameter is supposed to be estimated to a value that is equal (or even
close) to 1. The number (1) merely causes Mplus to estimate the same
value for each of the loadings l21, l22, and l23. This specification does not
tell Mplus which specific value it is. The optimal value in terms of the ML
criterion is determined by Mplus when the parameters of the model are
estimated. (In order to fix parameters to specific values, the @ command
is used in Mplus.)
For the indicator-­specific factor (is2), we also specify equal loadings
over time. Although equality of indicator-­specific factor loadings is not
strictly necessary for most purposes, this restriction is often plausible and
meaningful (because the indicator remains the same). In addition, it con-
tributes to the parsimony of the model. The equality restrictions for this
factor are slightly different, because the parameters that are subject to the
equality constraint here are loadings on a time-­unspecific factor. Given
that the first loading on this factor is again fixed to 1 by default in Mplus
(i.e., lIS21 = 1), we have to fix the remaining two indicator-­specific factor
loadings (lIS22 and lIS23) to 1 as well to obtain equal loadings on this
factor. (This was not necessary for the three state factor loadings of the
second indicator because none of the loadings was fixed to any specific
value as the default.) The loadings lIS22 and lIS23 are therefore fixed to 1
in the model command using the @ symbol:

is2 by d21 d22@1 d23@1;

The complete Mplus model command for the model with weak
factorial invariance is shown in Figure 4.8. The fit indices reported by
Mplus show that—­similar to the configural invariance model (see Sec-
tion 4.1.3)—the weak factorial invariance model shows a good fit to the
data, c2 = 5.094, df = 7, p = .6484, RMSEA = .00, p(RMSEA ≤ .05) = .925,
SRMR = .012, CFI = 1.00. To test whether the more restrictive model with
weak factorial invariance fits the data significantly worse than the con-
figural model with unrestricted factor loadings for the second indicator,
 Structural Equation Models for Measuring Variability and Change   103

FIGURE 4.8. Mplus model command for the specification of the LS model with
one indicator-­specific factor (is2) for the second indicator and time-­invariant load-
ings (model of weak factorial invariance).

a chi-­square difference test can be used, because the models are hierar-
chically nested (cf. Box 3.7). Nested means that the model with invariant
loadings can be seen as a special case of the baseline model with unre-
stricted loadings: The model with invariant loadings directly follows from
the configural model by constraining certain parameters.
Chi-­square difference tests can be easily calculated by hand or using
specialized software. In Box 4.4, I provide a description of how chi-­square
difference tests can be conducted automatically, based on Mplus outputs,
using Crayen’s (2010) software Chi-­Square Difference Calculator (CDC). The
CDC program (version 3) as well as a manual can be found on the com-
panion website.
In our example, the comparison of the weak factorial invariance
model with the configural invariance model yields a chi-­square differ-
ence value of c2∆ = 5.094 – 3.569 = 1.525. A chi-­square value of 1.525
with df∆ = 7 – 3 = 4 is not significant at the .05 level (p = .82; see also
Box 4.4). This means that the more restrictive model with equal loadings
across time does not fit significantly worse than the configural model with
unrestricted loadings. Hence, the assumption of weak factorial invari-
ance is not rejected for these data. The AIC also favors the more restricted
model (AIC = 1072.933) over the configural model (AIC = 1079.408), even
though the difference in the values is not very large.
By looking at the unstandardized parameter estimates (rubric MODEL
RESULTS; see below), we can see that the desired invariance restric-
tions on the factor loadings were correctly implemented by Mplus. The
104   Data Analysis with Mplus

BOX 4.4. Automated Chi-­Square Difference Test Using Crayen’s

(2010) CDC Software

The Chi-­Square Difference Calculator (CDC) program allows for an auto-

mated comparison of model fit between two nested models based on Mplus
output files. In the first step, the two nested models have to be estimated in
Mplus. Subsequently, the file CDC.exe (see companion website) is executed.

We then have to specify the location of the output file for (1) the more restric-
tive (H0) model and (2) the less restrictive (H1) model. Either the corre-
sponding paths can be typed directly into the respective fields or the Select
H0/H1 File buttons can be used to browse the computer for these files. By
clicking on Calculate, the program determines the chi-­square difference
value, the differences in df, and the p-value for the chi-­square difference
under the null hypothesis that the more restrictive model fits the data. In
addition, the program reports which model would be favored based on AIC,
 Structural Equation Models for Measuring Variability and Change   105

BIC, and sample-­size-­adjusted BIC (see Box 3.7). In this example, the more
restrictive model with equal loadings is preferred according to all criteria
(see discussion in the text).
It is important to note that the CDC program does not check whether
the assumptions required for proper chi-­square difference testing (see Box
3.7) are met. Specifically, the program does not check whether the two mod-
els chosen by the user are, in fact, nested. This has to be determined by the
user prior to the analysis. For certain non-­nested models, statistical compari-
sons via the AIC, BIC, and/or sample-­size-­adjusted BIC may still be mean-
ingful, but chi-­square difference testing would not be appropriate for these
models.

unstandardized state factor loadings of the second indicator were all esti-
mated to the same value ( lˆ 21 = lˆ 22 = lˆ 23 = 0.737 as intended). In addi-
tion, all three loadings show the same standard error (SE = 0.03), test
statistic (z = 24.857), and p-value (p < .001). All loadings on the indicator-­
specific factor is2 were fixed to 1, as desired. The fact that no standard
errors (and consequently no test statistic) were estimated for any of the
is2 factor loadings shows that they are fixed parameters in this model—
as was intended.

BOX 4.5. Checking the Correct Implementation of Constraints

The Mplus output for all models that include specific equality or other
constraints should be carefully screened for correct implementation of the
desired restrictions. The reason is that, especially with complex models,
errors can easily be made in the input specification. One should therefore
always check carefully whether the intended parameters were set equal (or
not equal). Sometimes an implausibly bad model fit can be a sign that one
or more parameter constraints were incorrectly specified. For example, it
can occur that erroneously, the same number is used twice for setting differ-
ent parameter groups equal. This may lead to the implementation of overly
restrictive and/or meaningless model constraints (e.g., the constraint that
loadings are set equal to intercepts). Such errors can lead to a significant
misfit of the model and to incorrect conclusions. Therefore, a careful inspec-
tion of the correct implementation of all restrictions should always be car-
ried out before interpreting key model results, conducting statistical model
comparisons, etc.
106   Data Analysis with Mplus

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 1.000 0.000 999.000 999.000
D21 0.737 0.030 24.857 0.000

STATE2 BY
D12 1.000 0.000 999.000 999.000
D22 0.737 0.030 24.857 0.000

STATE3 BY
D13 1.000 0.000 999.000 999.000
D23 0.737 0.030 24.857 0.000

IS2 BY
D21 1.000 0.000 999.000 999.000
D22 1.000 0.000 999.000 999.000
D23 1.000 0.000 999.000 999.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.118 0.015 7.963 0.000

STATE3 WITH
STATE1 0.108 0.014 7.622 0.000
STATE2 0.129 0.015 8.419 0.000

Intercepts
D11 1.584 0.027 58.666 0.000
D21 1.354 0.021 64.730 0.000
D12 1.642 0.028 58.461 0.000
D22 1.389 0.023 60.797 0.000
D13 1.631 0.027 60.172 0.000
D23 1.402 0.022 62.502 0.000

Variances
STATE1 0.195 0.019 10.505 0.000
STATE2 0.233 0.021 11.021 0.000
STATE3 0.218 0.020 10.982 0.000
IS2 0.007 0.002 3.456 0.001

Residual Variances
D11 0.040 0.009 4.462 0.000
D21 0.028 0.005 5.422 0.000
 Structural Equation Models for Measuring Variability and Change   107

D12 0.022 0.009 2.434 0.015

D22 0.035 0.006 6.288 0.000
D13 0.019 0.009 2.173 0.030
D23 0.037 0.006 6.592 0.000

Notice that the standardized loadings (shown in the following) are not
invariant, but can vary across time in this model. The reason is that the
weak factorial invariance model does not involve invariance constraints
on the state factor variances and the residual variances. The standardized
loadings would be the same across time only if the state factor variances
and the residual variances were also held equal across time.

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 0.911 0.021 43.928 0.000
D21 0.867 0.022 39.796 0.000

STATE2 BY
D12 0.957 0.018 52.262 0.000
D22 0.867 0.020 43.413 0.000

STATE3 BY
D13 0.959 0.019 49.979 0.000
D23 0.855 0.021 41.265 0.000

IS2 BY
D21 0.221 0.033 6.773 0.000
D22 0.202 0.030 6.761 0.000
D23 0.206 0.030 6.774 0.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.552 0.045 12.346 0.000

STATE3 WITH
STATE1 0.521 0.047 11.179 0.000
STATE2 0.574 0.043 13.455 0.000

Intercepts
D11 3.264 0.138 23.632 0.000
D21 3.602 0.149 24.153 0.000
108   Data Analysis with Mplus

D12 3.253 0.139 23.399 0.000

D22 3.383 0.141 23.959 0.000
D13 3.348 0.143 23.490 0.000
D23 3.478 0.144 24.194 0.000

Variances
STATE1 1.000 0.000 999.000 999.000
STATE2 1.000 0.000 999.000 999.000
STATE3 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000

Residual Variances
D11 0.171 0.038 4.517 0.000
D21 0.200 0.036 5.529 0.000
D12 0.085 0.035 2.429 0.015
D22 0.207 0.033 6.316 0.000
D13 0.080 0.037 2.172 0.030
D23 0.227 0.034 6.744 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.829 0.038 21.964 0.000

D21 0.800 0.036 22.122 0.000
D12 0.915 0.035 26.131 0.000
D22 0.793 0.033 24.138 0.000
D13 0.920 0.037 24.989 0.000
D23 0.773 0.034 22.926 0.000

In the next step, we test a model in which strong factorial invariance

is assumed. In this model, the intercepts of all indicators are set equal
across time in addition to the factor loadings. Given that we want to allow
for differences in the LS variable means across time, we have to override
the Mplus default according to which all latent variables have means of
0 (cf. Box 4.2). Otherwise, our specification with equal intercepts in con-
junction with zero latent means would be very restrictive because it would
reflect the assumption of zero mean change across time. (In the weak fac-
torial invariance model, this was not an issue, because all intercepts were
freely estimated and hence there were no constraints whatsoever on the
observed or latent mean structure.)
We first have to figure out how we can identify the LS factor means.
For this purpose, we once again consider the decomposition of the means
of the observed variables, E(Yik), in the LS model (cf. Box 4.2):
 Structural Equation Models for Measuring Variability and Change   109

E(Yik) = E(aik) + E(lik · State k) + E(eik)

= aik+ lik · E(State k)

This formula shows that the means of the LS variables, E(State k), can
be identified in a relatively straightforward way by setting the intercepts,
a1k, pertaining to the marker indicators, Y1k, to zero. If we set a1k = 0 for
all k, then all LS factor means become identified through the means of the
marker indicators:

E(Y1k) = a1k+ l1k · E(State k)

= 0 + 1 · E(State k)
= E(State k)

In this way, the latent variable means are (1) identified and esti-
mable, and (2) they can potentially take on different values at each time
point, reflecting latent mean change over time. Furthermore, this speci-
fication ensures that the intercepts of the marker indicators are time-­
invariant (each one takes on the same value of 0), reflecting measurement
invariance across time for these indicators. The intercepts of the remain-
ing (non-­marker) indicators are set equal across time, following the same
principle as shown for the factor loadings in the weak factorial invariance
model.
The full Mplus model specification for the strong factorial invariance
model is shown in Figure 4.9. Note that commands related to the mean
structure (observed and latent variable means and intercept parameters)
involve variable names in brackets []. Depending on whether a variable
is exogenous or endogenous in the model (as determined automatically
by Mplus, based on the model specification), a variable name in brackets
refers to either the mean (for exogenous variables) or the intercept (for
endogenous variables) of that variable.
Figure 4.9 shows that the intercepts of the marker indicators (d11,
d12, and d13) are set to zero to identify the LS factor means:

[d11@0 d12@0 d13@0];

Furthermore, the estimation of the LS factor means is explicitly requested

using the command [state1 state2 state3];. Otherwise, the LS
factor means would remain fixed at zero, which is no longer adequate.
110   Data Analysis with Mplus

FIGURE 4.9. Mplus model command for the specification of the LS model with
one indicator-­specific factor (is2) for the second indicator, time-­invariant loadings,
and time-­invariant intercepts (model of strong factorial invariance).

Finally, the intercepts of the indicators d21, d22, and d23 are set
equal across time. Note that we use the number (2) to set these parameters
equal, which is a different number than the one used to set the loadings
equal. [The number (1) is still used to set the loadings of the indicators
d21, d22, and d23 equal across time.] The mean of the indicator-­specific
factor does not have to be specified, because it is set to zero by default (as
is desired in this model).
The Mplus output shows that the model with strong factorial invari-
ance also fits the data well, c2 = 6.656, df = 9, p = .6729, RMSEA = .00,
p(RMSEA) ≤ .05 = .95, SRMR = .014, CFI = 1.00. In addition, the chi-­
square difference test against the model with weak factorial invariance
is not significant, c2∆ = 1.562, df∆ = 2, p = .46, indicating that the more
restrictive, strong factorial invariance model is preferred. A comparison of
the AIC values also indicates that the more constrained model should be
chosen (strong factorial invariance model: AIC = 1070.494; weak factorial
invariance model: AIC = 1072.933). We can conclude that the assumption
of strong factorial invariance is not rejected in this example.
Based on the unstandardized parameter estimates (MODEL RESULTS,
see below), we can see that our equality constraints were correctly
 Structural Equation Models for Measuring Variability and Change   111

implemented by Mplus ( lˆ 21 = lˆ 22 = lˆ 23 = 0.738; aˆ 21 = aˆ 22 = aˆ 23 =

0.187). Mplus also correctly fixed the intercepts pertaining to the first indi-
cator to zero (i.e., a11 = a12 = a13 = 0), as can be seen from the fact that no
standard error was estimated any more for these intercepts. We now also
obtain estimates of the LS variable means in the output (under Means).
We can see that the latent depression means have slightly increased from
time 1 to time 2 (M1 = 1.583 vs. M2 = 1.639), whereas there is almost no
change from time 2 to time 3 (M3 = 1.634). A formal test of equality of
means is demonstrated in Box 4.6.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 1.000 0.000 999.000 999.000
D21 0.738 0.030 24.997 0.000

STATE2 BY
D12 1.000 0.000 999.000 999.000
D22 0.738 0.030 24.997 0.000

STATE3 BY
D13 1.000 0.000 999.000 999.000
D23 0.738 0.030 24.997 0.000

IS2 BY
D21 1.000 0.000 999.000 999.000
D22 1.000 0.000 999.000 999.000
D23 1.000 0.000 999.000 999.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.118 0.015 7.964 0.000

STATE3 WITH
STATE1 0.108 0.014 7.623 0.000
STATE2 0.129 0.015 8.419 0.000

Means
STATE1 1.583 0.026 60.157 0.000
STATE2 1.639 0.028 58.762 0.000
STATE3 1.634 0.027 60.652 0.000
112   Data Analysis with Mplus

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.187 0.049 3.861 0.000
D12 0.000 0.000 999.000 999.000
D22 0.187 0.049 3.861 0.000
D13 0.000 0.000 999.000 999.000
D23 0.187 0.049 3.861 0.000

Variances
STATE1 0.195 0.019 10.510 0.000
STATE2 0.233 0.021 11.023 0.000
STATE3 0.218 0.020 10.983 0.000
IS2 0.007 0.002 3.433 0.001

Residual Variances
D11 0.040 0.009 4.477 0.000
D21 0.028 0.005 5.425 0.000
D12 0.022 0.009 2.453 0.014
D22 0.035 0.006 6.299 0.000
D13 0.019 0.009 2.192 0.028
D23 0.037 0.006 6.608 0.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 0.911 0.021 43.988 0.000
D21 0.867 0.022 39.895 0.000

STATE2 BY
D12 0.956 0.018 52.311 0.000
D22 0.867 0.020 43.500 0.000

STATE3 BY
D13 0.959 0.019 50.014 0.000
D23 0.854 0.021 41.331 0.000

IS2 BY
D21 0.220 0.033 6.732 0.000
D22 0.202 0.030 6.721 0.000
D23 0.205 0.031 6.733 0.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STATE2 WITH
STATE1 0.553 0.045 12.350 0.000
 Structural Equation Models for Measuring Variability and Change   113

STATE3 WITH
STATE1 0.521 0.047 11.184 0.000
STATE2 0.574 0.043 13.463 0.000

Means
STATE1 3.583 0.181 19.792 0.000
STATE2 3.395 0.164 20.655 0.000
STATE3 3.499 0.169 20.664 0.000

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.498 0.134 3.707 0.000
D12 0.000 0.000 999.000 999.000
D22 0.456 0.124 3.677 0.000
D13 0.000 0.000 999.000 999.000
D23 0.465 0.126 3.686 0.000

Variances
STATE1 1.000 0.000 999.000 999.000
STATE2 1.000 0.000 999.000 999.000
STATE3 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000

Residual Variances
D11 0.171 0.038 4.532 0.000
D21 0.200 0.036 5.535 0.000
D12 0.086 0.035 2.448 0.014
D22 0.208 0.033 6.332 0.000
D13 0.081 0.037 2.191 0.028
D23 0.228 0.034 6.767 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.829 0.038 21.994 0.000

D21 0.800 0.036 22.155 0.000
D12 0.914 0.035 26.155 0.000
D22 0.792 0.033 24.165 0.000
D13 0.919 0.037 25.007 0.000
D23 0.772 0.034 22.941 0.000

Finally, we test a model of strict factorial invariance. In this model,

we also set the residual variances of the indicators [Var(eik)] equal across
time. (We use the model with unequal latent means as the basis here.)
We use the number (3) to set the error variances of the first indicator
equal across time. The number (4) is used to set the error variances of
the second indicator equal over time. The complete model statement for
114   Data Analysis with Mplus

the specification of the model with strict factorial invariance is shown in

Figure 4.10.
The model with strict factorial invariance and unequal means also
shows a good absolute fit, c2 = 11.61, df = 13, p = .5599, RMSEA = .00,
p(RMSEA) ≤ .05 = .95, SRMR = .024, CFI = 1.00. (The full output for this
model can be found on the companion website.) The chi-­square difference
test against the strong factorial invariance model with unequal means is
again nonsignificant, c2∆ = 4.954, df∆ = 4, p = .29, favoring the strict invari-
ance model. A comparison of the AIC values leads to the same conclu-
sion (strict factorial invariance model: AIC = 1067.448; strong factorial
invariance model: AIC = 1070.494). We can conclude that the assumption

BOX 4.6. Test of Equality of Latent Means across Time

A question of substantive interest is whether the three depression state fac-

tor means are significantly different from each other (which would indicate
significant mean change between at least two measurement occasions). For-
mally, the assumption of equality of means can be tested by fitting an addi-
tional model to the data in which all three latent state means are set equal
across time. If this model fits significantly worse than the strong factorial
invariance model without restrictions on the latent means, we can conclude
that mean change occurred over time between at least two time points. The
following command can be used to set the latent state factor means equal
across time in Mplus (the full Mplus input for this model can be found on the
companion website):

[state1 state2 state3] (mean);

Here the label (mean) is used to set the state means equal across time. (We
could have equally well used a number again.) In our example, the model
with equal means does not fit significantly worse than the strong factorial
invariance model with unequal means, according to the chi-­square differ-
ence test, c2∆ = 5.181, df∆ = 2, p = .07, indicating that the observed mean
differences are not statistically significant at the .05 level. On the other hand,
according to the AIC index, the less constrained model would be preferred
(strong factorial invariance model with equal means: AIC = 1071.675; strong
factorial invariance model with unequal means: AIC = 1070.494), although
the difference in AIC values is rather small. In practice, a researcher would
have to take substantive considerations as well as the actual size of the mean
differences into account before making a final decision as to whether the
model with equal or unequal means should be preferred.
 Structural Equation Models for Measuring Variability and Change   115

FIGURE 4.10. Mplus model command for the specification of the LS model with
one indicator-­specific factor (is2) for the second indicator, time-­invariant loadings,
time-­invariant intercepts, and time-­invariant error variances (model of strict facto-
rial invariance).

of strict factorial invariance is not rejected in this example. Therefore,

the more parsimonious model with equal residual variances should be
selected.
We now terminate our series of invariance tests, noting that—­
depending on the research question—­further invariance tests may be of
interest, for example, with respect to the LS factor variances and covari-
ances. In summary, our analysis of the LS model yielded a number of
interesting findings that can be summarized as follows:

• The indicators used to measure depression show relatively high

reliabilities and low indicator-­specificity.
• The parameters of the measurement model did not change sig-
nificantly across time; that is, strict factorial invariance could be
assumed for these data.
116   Data Analysis with Mplus

• True individual differences in depression were moderately stable

across time, as indicated by the medium-­size correlations among
LS factors; in other words, there was a substantial amount of reli-
able occasion-­specific variance that may be examined in more
detail using LST analyses (see next section).
• The correlation structure of the LS factors indicates a slight autore-
gressive effect across time.
• A global test of latent mean differences across time was not sta-
tistically significant at the .05 level, indicating that there was no
significant mean change over time in depression.

As mentioned above, the LS model is a useful baseline model for the anal-
ysis of more complex longitudinal SEMs. In the following material, we
discuss its extension to an LST model, using the same data example.

4.2 LST Analysis

One limitation of the LS model is that the degree to which measurements

are affected by occasion-­ specific effects versus stable person-­ specific
effects can only be quantified indirectly in this model by means of the
correlations among the state factors. However, a decomposition of the
observed variance into stable, occasion-­specific, and error components
is not available in this model (e.g., Steyer et al., 1992). By extending the
LS model to an LST model, we can obtain more detailed information on
the amount of state- versus trait-­related variance of observed and latent
variables.
In an LST model, the observed variables Yik (i = indicator, k = occasion
of measurement) can again be decomposed into a constant intercept term
aik, an occasion-­specific LS variable (State k), and an occasion-­specific
measurement error (residual) variable eik:

Yik = aik + lik · (State k) + eik

where lik again indicates the factor loading of Yik on the LS factor.
In contrast to the LS model, the LST model further decomposes the
LS factors into a latent trait factor (which reflects that part of the LS fac-
tors that is due to stable individual differences or trait effects) and an LS
residual variable zk that characterizes occasion-­specific deviations of the
latent state from the latent trait scores (cf. Steyer et al., 1992, 1999):
 Structural Equation Models for Measuring Variability and Change   117

State k = g0k + gk · Trait + zk

where g0k indicates a constant latent intercept term and gk a second-­order

factor loading of State k on the factor Trait.
In our analysis of the LS model in Section 4.1.3, we saw that the
covariances (or, in the completely standardized solution, the correlations)
among the state factors were an indicator of the degree of stability of indi-
vidual differences across time. In the LST model, the covariance structure
of the LS factors is explained by a single underlying latent (trait) factor.
The single latent trait factor reflects that part of the variance in the state
factors that is stable across time (i.e., their covariance).
Figure 4.11 shows an LST model for two observed variables Y1k and
Y2k measured on three measurement occasions (k = 1, 2, 3). It can be seen
that the LST model implies a second-­order factor model (cf. Section 3.4.2).

Trait
Depression

γ1 γ3
γ2
ζ2
ζ1 ζ3
State 1: State 2: State 3:
Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

FIGURE 4.11. Latent state–trait (LST) model for depression measured on three
occasions of measurement (T1–T3). Yik = observed variable i measured at time point
k; eik = measurement error variable; lik = first-order factor loading; gk = second-­order
factor loading; zk = LS residual variable. The variable names of the indicators used in
Mplus appear in parentheses.
118   Data Analysis with Mplus

As an example, we specify an LST model for the depression data (data

set depression.dat), which we have already used to illustrate different
analyses based on the LS model in Section 4.1. As a starting point, we
use the LS model with an indicator-­specific factor and without invariance
constraints that showed a good fit (cf. Section 4.1.3). The extension of this
model to an LST model is shown in Figure 4.12. A similar LST model with
an indicator-­specific factor for the second indicator is discussed in detail
in Eid et al. (1999).

Trait
Depression

γ1 γ3
γ2
ζ2
ζ1 ζ3
State 1: State 2: State 3:
Depression T1 Depression T2 Depression T3

λ11 λ21 λ12 λ22 λ13 λ23

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

λIS22 λIS23
λIS21

IS2

FIGURE 4.12. LST model for depression measured on three occasions of mea-
surement (T1–T3). Yik = observed variable i measured at time point k; eik = measure-
ment error variable; lik = first-order factor loading; gk = second-­order factor load-
ing; zk = LS residual variable. The model additionally contains an indicator-­specific
(method) factor (IS2) for the second indicator. The parameters lIS2k indicate factor
loadings on the indicator-­specific factor. The variable names of the indicators used in
Mplus appear in parentheses.
 Structural Equation Models for Measuring Variability and Change   119

FIGURE 4.13. Mplus model command for the specification of the LST model
with one indicator-­specific factor (is2) for the second indicator shown in Figure
4.12.

The Mplus model statement for estimating the LST model in Figure
4.12 is shown in Figure 4.13. Note that the identification of the latent vari-
able mean structure is slightly more complicated in the LST model than
in the LS model. This is the case at least if one is interested in estimating
the mean of the latent trait factor. (A more detailed discussion of differ-
ent options for including a latent mean structure in LST analyses can be
found in Box 4.7.)
One way to identify the latent trait factor mean as well as the latent
intercept parameters g0k is to fix the intercept parameters pertaining to
the marker indicators to zero, that is, a11 = a12 = a13 = 0.
In the next step, the intercept term pertaining to an LS factor cho-
sen as reference indicator of the latent trait is set to zero as well—in our
case, the intercept pertaining to State 1 (i.e., g01 = 0). This state factor is
chosen for simplicity, because Mplus automatically sets the loading of this
state factor on the trait factor to 1 (i.e., g1 = 1). This is because Mplus, by
default, treats the first variable listed after the by statement as the refer-
ence variable and sets its loading to 1. As a consequence, the mean of
the latent trait factor is then directly identified through the mean of the
marker variable Y11. Finally, the intercepts of the remaining state factors
(in our case, g02 and g03) are freely estimated. The mean of the is2 factor
120   Data Analysis with Mplus

BOX 4.7. Latent Mean Structures in the LST Model

Whether or not a latent variable mean structure is of interest in an LST anal-

ysis depends on the specific research question. Many LST analyses focus
exclusively on the covariance structure of the variables and completely
ignore the mean structure. In some situations, however, researchers may be
interested in studying mean structures as well—for example, when mean
differences across time are of interest or are theoretically expected. In the
text, I described a way to identify the latent mean structure in an LST model
that allows estimating the mean of the latent trait factor. In addition to the
model version discussed in the text, there are other, statistically equivalent
options for specifying a mean structure in an LST model. If one is not inter-
ested in estimating the mean of the latent trait factor, one can simply follow
the Mplus default, according to which the trait factor mean is fixed at 0, and
estimate all three latent intercept parameters g01, g02, and g03. The complete
model specification is then simplified, as shown in the following:

state1 by d11 d21;

state2 by d12 d22;
state3 by d13 d23;
trait by state1-state3;
is2 by d21 d22 d23;
is2 with state1-state3@0 trait@0;
[d11@0 d12@0 d13@0];
[state1 state2 state3];

This alternative equivalent parameterization has the advantage that the

intercepts of the three state factors are now identical to the LS factor means
(which is not the case for the parameterization discussed in the text). Latent
mean comparisons could be conducted by comparing the intercepts. Keep in
mind, however, that such a comparison would require at least strong facto-
rial invariance of the measurement model (see Section 4.1.4).
Another (equivalent) possibility is to completely ignore the latent vari-
able mean structure and to leave all latent means and latent intercepts fixed
at zero (as is the default in Mplus). In this case, only the intercepts aik per-
taining to the observed variables would be estimated. This parameterization
would be obtained through the following model specification:

state1 by d11 d21;

state2 by d12 d22;
state3 by d13 d23;
trait by state1-state3;
is2 by d21 d22 d23;
is2 with state1-state3@0 trait@0;
 Structural Equation Models for Measuring Variability and Change   121

Additional alternative mean structures are possible. Furthermore, it can be

useful in practice to place restrictions on the mean structure to test specific
assumptions, as shown in Section 4.1.4. The type of mean structure included
in an LST analysis should depend on the substantive research questions.
For example, if the process studied is hypothesized to be a pure variability
process (with no trait change over time), there should not be any changes in
the LS means over time. Hence, in such a model, strong factorial invariance
would be assumed to hold over time for the Yik variables, all trait loadings
gk would be expected to be 1, and all LS intercept parameters g0k would be
expected to be 0. This restricted model could be tested against a model with
free parameters to find out whether the process under study involves just
variability and no trait changes.

again remains fixed at 0 (following the Mplus default settings) because

this factor is defined as a residual factor.
From the fit statistics (see below) we can see that the LST model, in
this particular case, shows the same fit as the LS model with configural
invariance from Section 4.1.3. This finding can be explained by the fact
that in this case, data from only three time points are analyzed. With only
three time points, the higher-­order covariance structure of the LS factors
is saturated in the LST model because no overidentifying restrictions are
included for the loadings of State 2 and State 3 on the trait factor.
The mean structure is saturated as well because no overidentifying
restrictions were placed on the latent trait mean or the intercept parame-
ters. Therefore, the structural model does not provide any additional df, so
that the fit of the model is identical to the fit of the unrestricted LS model.
This would be different if we were analyzing more than three time points.
With four or more time points, the LST model implies testable restrictions
for the covariance structure of the LS factors, so that the structural model
is no longer saturated in these cases.

Chi-Square Test of Model Fit

Value 3.541
Degrees of Freedom 3
P-Value 0.3155

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.024
90 Percent C.I. 0.000 0.100
Probability RMSEA <= .05 0.615
122   Data Analysis with Mplus

CFI/TLI

CFI 1.000
TLI 0.998

SRMR (Standardized Root Mean Square Residual)

Value 0.008

Information Criteria

Akaike (AIC) 1085.741

Bayesian (BIC) 1176.404
Sample-Size Adjusted BIC 1100.279
(n* = (n + 2) / 24)

The Mplus parameter estimates (MODEL RESULTS) show that the

state factors have fairly large and statistically significant loadings on the
trait factor (see under Trait by). The estimated unstandardized load-
ings for State 2 and State 3 are ĝ 2 and ĝ 3 (z = 8.731 and 8.805, respec-
tively; the loading of State 1 was fixed to 1 for identification). The esti-
mated standardized loadings are ĝ 1standardized = .707, ĝ 2standardized = .782,
and ĝ 1standardized = .739. The size of the standardized loadings shows that
a sizable proportion of true individual differences (individual differences
on the state factors) is explained by stable influences (trait effects). This
finding reflects the substantial correlations seen between the LS variables
in the LS model. The amount of state factor variance that is explained by
the trait factor can be found in the output under R-SQUARE—Latent
Variable (under the standardized model results). The R 2 values for the
latent variables (here the LS factors) are identical to the squared stan-
dardized loadings on the trait factor: R12 = .7072 = .500; R22 = .7872 = .611;
R32 = .7392 = .546. Hence, between 50 and 61.1% of true individual dif-
ferences in depression states can be explained by stable personality dis-
positions, whereas the remaining 38.9–50% are due to occasion-­specific
effects.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
DEPDAD11 1.000 0.000 999.000 999.000
DEPDAD21 0.700 0.049 14.248 0.000
 Structural Equation Models for Measuring Variability and Change   123

STATE2 BY
DEPDAD12 1.000 0.000 999.000 999.000
DEPDAD22 0.751 0.043 17.260 0.000

STATE3 BY
DEPDAD13 1.000 0.000 999.000 999.000
DEPDAD23 0.755 0.047 16.087 0.000

IS2 BY
DEPDAD21 1.000 0.000 999.000 999.000
DEPDAD22 0.612 0.375 1.629 0.103
DEPDAD23 0.799 0.518 1.542 0.123

TRAIT BY
STATE1 1.000 0.000 999.000 999.000
STATE2 1.168 0.134 8.731 0.000
STATE3 1.067 0.121 8.805 0.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000
TRAIT 0.000 0.000 999.000 999.000

Means
TRAIT 1.584 0.027 58.184 0.000

Intercepts
DEPDAD11 0.000 0.000 999.000 999.000
DEPDAD21 0.246 0.079 3.112 0.002
DEPDAD12 0.000 0.000 999.000 999.000
DEPDAD22 0.156 0.073 2.147 0.032
DEPDAD13 0.000 0.000 999.000 999.000
DEPDAD23 0.171 0.078 2.197 0.028
STATE1 0.000 0.000 999.000 999.000
STATE2 -0.209 0.214 -0.975 0.330
STATE3 -0.060 0.194 -0.311 0.756

Variances
TRAIT 0.103 0.019 5.553 0.000
IS2 0.011 0.008 1.390 0.164

Residual Variances
DEPDAD11 0.033 0.012 2.711 0.007
DEPDAD21 0.028 0.009 3.091 0.002
DEPDAD12 0.025 0.011 2.297 0.022
DEPDAD22 0.036 0.007 5.351 0.000
DEPDAD13 0.022 0.011 1.986 0.047
DEPDAD23 0.035 0.008 4.592 0.000
STATE1 0.103 0.016 6.308 0.000
STATE2 0.089 0.017 5.146 0.000
STATE3 0.098 0.017 5.901 0.000
124   Data Analysis with Mplus

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
DEPDAD11 0.928 0.028 32.979 0.000
DEPDAD21 0.849 0.029 29.046 0.000

STATE2 BY
DEPDAD12 0.950 0.023 41.661 0.000
DEPDAD22 0.874 0.024 36.051 0.000

STATE3 BY
DEPDAD13 0.953 0.025 38.686 0.000
DEPDAD23 0.862 0.026 32.966 0.000

IS2 BY
DEPDAD21 0.278 0.100 2.785 0.005
DEPDAD22 0.155 0.062 2.502 0.012
DEPDAD23 0.205 0.077 2.659 0.008

TRAIT BY
STATE1 0.707 0.046 15.245 0.000
STATE2 0.782 0.045 17.409 0.000
STATE3 0.739 0.046 16.188 0.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000
TRAIT 0.000 0.000 999.000 999.000

Means
TRAIT 4.937 0.453 10.909 0.000

Intercepts
DEPDAD11 0.000 0.000 999.000 999.000
DEPDAD21 0.656 0.220 2.980 0.003
DEPDAD12 0.000 0.000 999.000 999.000
DEPDAD22 0.378 0.182 2.085 0.037
DEPDAD13 0.000 0.000 999.000 999.000
DEPDAD23 0.420 0.197 2.131 0.033
STATE1 0.000 0.000 999.000 999.000
STATE2 -0.435 0.440 -0.989 0.323
STATE3 -0.130 0.416 -0.313 0.755

Variances
TRAIT 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000
 Structural Equation Models for Measuring Variability and Change   125

Residual Variances
DEPDAD11 0.139 0.052 2.668 0.008
DEPDAD21 0.201 0.067 3.026 0.002
DEPDAD12 0.098 0.043 2.267 0.023
DEPDAD22 0.212 0.042 5.081 0.000
DEPDAD13 0.092 0.047 1.966 0.049
DEPDAD23 0.215 0.049 4.417 0.000
STATE1 0.500 0.066 7.634 0.000
STATE2 0.389 0.070 5.534 0.000
STATE3 0.454 0.067 6.741 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

DEPDAD11 0.861 0.052 16.490 0.000

DEPDAD21 0.799 0.067 11.993 0.000
DEPDAD12 0.902 0.043 20.830 0.000
DEPDAD22 0.788 0.042 18.925 0.000
DEPDAD13 0.908 0.047 19.343 0.000
DEPDAD23 0.785 0.049 16.155 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

STATE1 0.500 0.066 7.622 0.000

STATE2 0.611 0.070 8.705 0.000
STATE3 0.546 0.067 8.094 0.000

The latent mean structure is also shown in the output. In this model,
the unstandardized latent trait mean is estimated to be 1.584. The inter-
cepts of State 2 and State 3 are estimated to be ĝ 02 = –0.209 and ĝ 03 =
–0.060. These values reflect slight mean differences relative to the first
state factor. However, these mean differences are not significant (p ≥ .33).
Further steps of an LST analysis may involve tests of measurement
invariance to obtain more parsimonious and substantively meaningful
models. Given that we have already discussed the basic principles of test-
ing measurement invariance for the LS model in Section 4.1.4, I will not
go into the details here. The procedure shown for the LS model can be
easily transferred and applied to the LST model as well.
LST models were primarily developed as variability models to model
systematic, occasion-­ specific fluctuations of measurements around a
constant trait value. As we saw, these models can be used to estimate
variance components to separate stable influences from occasion-­specific
126   Data Analysis with Mplus

influences on LS variables. In the following sections, we consider lon-

gitudinal models that are less concerned with occasion-­specific fluctua-
tions versus stable components of constructs, but rather focus on actual
changes in constructs over time. Such change models are used particu-
larly when it has to be assumed that an attribute has actually changed
(rather than just showing fluctuations around a fixed set point). We first
describe autoregressive models, which can be seen as “classical” models
for measuring change within the framework of path analysis and SEM.

4.3 Autoregressive Models

For a long time, autoregressive models (e.g., Jöreskog, 1979a, 1979b) were
seen as the “gold standard” of methodology for analyzing longitudinal
data. With the development of latent growth curve models (LGCMs; see
Section 4.5), autoregressive models have become somewhat less popu-
lar (for a critique of autoregressive models, see Hertzog & Nesselroade,
1987). Nonetheless, autoregressive models are useful to answer a variety
of important research questions in longitudinal studies.
Autoregressive models are based on the (fairly plausible) idea that
previous behavior is the best predictor of present behavior. For this
reason, repeatedly measured manifest or latent variables are connected
through so-­called autoregressions in autoregressive models (as reflected in
the path coefficients b1–b4 in the manifest autoregressive model shown
in Figure 4.14). The strengths of autoregressive (or autopredictive) effects
as well as the stability of individual differences across time can be exam-
ined by looking at the size of the autoregressive coefficients as well as
the amount of residual variance (variance of the residual variables ejk in
Figure 4.14). When the ratio of residual variance to total variance is large,
this means that the autoprediction is rather weak—­indicating individual
differences in change over time. (In manifest autoregressive models that
do not account for measurement error, such error can be an additional or
even alternative explanation for a large residual variance.)
However, the goal of an analysis of autoregressive models is usually
not only to determine the degree of stability of (or changes in) a single
construct over time. Often, an additional goal is to predict individual dif-
ferences not explained by the autoregressive effect(s) through external
variables that are also included in the model. The impact of such addi-
tional variables can be examined by means of so-­called cross-­lagged effects
 Structural Equation Models for Measuring Variability and Change   127

ε 12 ε 13

Y11 β1 Y12 β2 Y13

(d11) (d12) (d13)

β5 β7
β6 β8
Y21 β3 Y22 β4 Y23
(c11) (c12) (c13)

ε 22 ε 23

FIGURE 4.14. Manifest autoregressive model for two manifest variables Yjk (j =
construct, k = time point) measuring two constructs (j = 1, 2; depression and com-
petence) on three measurement occasions (k = 1, 2, 3): b1–b4 = path coefficients rep-
resenting autoregressive effects; b5 –b8 = path coefficients representing cross-­lagged
effects; ejk = residual variables. The variable names used in Mplus appear in paren-
theses.

(see path coefficients b5 –b8 in Figure 4.14). Cross-­lagged effects are the
effects of additional, temporally preceding variables that are included in
the model in addition to the autoregressive effects. The goal is to explain
at least part of the “leftover” variability that is not explained by the autore-
gressive effect(s). In other words, one attempts to explain the “unstable”
part of individual differences by external variables that may be possible
causes of change.

4.3.1 Manifest Autoregressive Models

Originally, autoregressive and/or cross-­lagged models were often speci-
fied exclusively at the level of manifest variables (sometimes referred to
as single-­indicator models, as shown in Figure 4.14), and analyses with
observed variables are still common. These models imply simple path
analyses without latent variables (manifest path models for cross-­sectional
data are described in Section 3.5.2). In the following, I first describe the
analysis of manifest autoregressive models in Mplus, and in Section 4.3.2,
I discuss latent autoregressive models that explicitly account for measure-
ment error.
128   Data Analysis with Mplus

BOX 4.8. Example of Manifest Autoregressive Models

Here we analyze a longitudinal data set with two constructs (self-­reported

depression and competencies, N = 586 children) on three measurement
occasions.* There were 6 months between each time point. The data set,
depression-­competence.dat, contains summary data (means, standard devi-
ations, and correlations) for six manifest variables. The first three variables
pertain to a repeatedly administered depression scale (the Child Depression
Inventory, CDI; Kovacs, 1985). The last three variables represent a summary
score of various competencies (for details on this study, see Cole, Martin, &
Powers, 1997, as well as Cole, Martin, Powers, & Truglio, 1996). Our goal
is to estimate a manifest autoregressive model for the six manifest variables.
In this model, we want to examine the stability of both constructs as well as
potential cross-­lagged effects.

*I would like to thank David A. Cole for providing the data used for this example.

Figure 4.15 shows the Mplus input file for the manifest autoregres-
sive model with cross-­lagged effects from Figure 4.14. The autoregressive
and cross-­lagged effects in the model are represented by partial regression
coefficients (also referred to as path coefficients; cf. Section 3.5). Therefore,
on statements are used in the Mplus model command to specify these
effects. As noted above, the model represents a special case of a manifest
path analysis, here specified for longitudinal data.
An important aspect in the model specification is that the residual
variables shown in Figure 4.14 are assumed to be uncorrelated. This has

FIGURE 4.15. Mplus input file for the specification of the simple manifest autore-
gressive model for three measurement occasions shown in Figure 4.14.
 Structural Equation Models for Measuring Variability and Change   129

to be explicitly specified for the residuals on the third time point, because
Mplus otherwise estimates a residual correlation between these residuals
by default. For this reason, we add the following command in the model
specification:

d13 with c13@0;

This command overrides the Mplus default and sets the residual correla-
tion for the final outcome variables to zero. All other residual correlations
are automatically set to zero by Mplus. (As we will see shortly, the original
model with uncorrelated residuals does not fit the data in this example, so
that this restriction will be relaxed again in later analysis steps.)
The model fit indices reported by Mplus indicate a very bad fit of
the model to the data, c2 = 201.90, df = 6, p < .001, CFI = 0.86, RMSEA
= 0.24, SRMR = 0.09 (the full output can be found on the companion
website). The lack of fit can in this case be explained by the fact that the
model does not allow the residuals at the same time point to be corre-
lated. The estimation of these additional parameters is necessary in this
example, because there are shared occasion-­specific effects between the
depression and competencies constructs at the same time point, in line
with LST theory. In other words, children who rated themselves as more
depressed also tended to report a lower level of competencies at the same
time point. These situation-­specific effects are not sufficiently accounted
for in the standard autoregressive model with uncorrelated residual vari-
ables, shown in Figure 4.14. This can be changed by allowing the residual
variables at the same time point to correlate (see Figure 4.16).
In addition to situation-­specific effects, the present example also
requires the inclusion of second-­order autoregressive effects to achieve
an acceptable fit. Second-order autoregressive effects are autoregressive
effects between nonadjacent time points (represented by the path coef-
ficients b9 and b10 in Figure 4.16). The Mplus model statement for the
extended model is shown in Figure 4.17. This model fits the data well,
c2 = 2.74, df = 2, p = .25, CFI = 0.999, RMSEA = 0.03, SRMR = 0.01. In
addition, it fits significantly better than the original model with uncor-
related residual variables and no second-­order autoregressive effects, Dc2
= 199.16, Ddf = 4, p < .001. The AIC value also indicates that the less
restrictive model in Figure 4.16 (AIC = 2493.08) should be preferred (the
AIC value for the model in Figure 4.14 was 2690.01). In the following, the
parameter estimates for the extended model are presented.
130   Data Analysis with Mplus

β9
ε 12 ε 13

Y11 β1 Y12 β2 Y13

(d11) (d12) (d13)

β5 β7
β6 β8
Y21 β3 Y22 β4 Y23
(c11) (c12) (c13)

ε 22 ε 23
β10

FIGURE 4.16. Manifest autoregressive model for two manifest variables Yjk (j =
construct, k = time point) measuring two constructs (j = 1, 2; depression and com-
petence) on three measurement occasions (k = 1, 2, 3). b1–b4 = path coefficients
representing first-order autoregressive effects; b9 and b10 = latent path coefficients
representing second-­order autoregressive effects; b5 –b8 = path coefficients repre-
senting cross-­lagged effects; ejk = residual variables. In contrast to the model shown
in Figure 4.14, this model allows for second-­order autoregressive effects as well as
correlated residual variables at the same measurement occasion. The variable names
used in Mplus appear in parentheses.

FIGURE 4.17. Mplus model command for the specification of the extended
autoregressive model with second-­order autoregressive effects and correlated resid-
ual variables shown in Figure 4.16.
 Structural Equation Models for Measuring Variability and Change   131

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

D12 ON
D11 0.526 0.037 14.225 0.000
C11 -0.085 0.020 -4.184 0.000

D13 ON
D12 0.510 0.040 12.782 0.000
D11 0.156 0.035 4.411 0.000
C12 -0.056 0.020 -2.749 0.006

C12 ON
C11 0.422 0.037 11.352 0.000
D11 -0.392 0.068 -5.786 0.000

C13 ON
C12 0.386 0.044 8.797 0.000
C11 0.192 0.037 5.181 0.000
D12 -0.436 0.072 -6.066 0.000

D12 WITH
C12 -0.040 0.005 -8.067 0.000

D13 WITH
C13 -0.035 0.005 -7.868 0.000

Intercepts
D12 0.288 0.050 5.789 0.000
D13 0.210 0.050 4.198 0.000
C12 1.335 0.091 14.640 0.000
C13 0.995 0.107 9.322 0.000

Residual Variances
D12 0.061 0.004 17.117 0.000
D13 0.051 0.003 17.117 0.000
C12 0.206 0.012 17.117 0.000
C13 0.207 0.012 17.117 0.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

D12 ON
D11 0.531 0.033 15.947 0.000
C11 -0.156 0.037 -4.209 0.000
132   Data Analysis with Mplus

D13 ON
D12 0.517 0.037 13.820 0.000
D11 0.160 0.036 4.426 0.000
C12 -0.100 0.036 -2.754 0.006

C12 ON
C11 0.440 0.036 12.214 0.000
D11 -0.224 0.038 -5.872 0.000

C13 ON
C12 0.361 0.040 9.128 0.000
C11 0.187 0.036 5.214 0.000
D12 -0.231 0.038 -6.150 0.000

D12 WITH
C12 -0.353 0.036 -9.778 0.000

D13 WITH
C13 -0.344 0.036 -9.434 0.000

Intercepts
D12 0.909 0.159 5.718 0.000
D13 0.674 0.163 4.146 0.000
C12 2.386 0.197 12.142 0.000
C13 1.663 0.200 8.315 0.000

Residual Variances
D12 0.610 0.031 19.374 0.000
D13 0.526 0.030 17.590 0.000
C12 0.657 0.032 20.657 0.000
C13 0.579 0.031 18.679 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D12 0.390 0.031 12.403 0.000

D13 0.474 0.030 15.875 0.000
C12 0.343 0.032 10.800 0.000
C13 0.421 0.031 13.602 0.000

It turns out that all autoregressive effects (including both first- and
second-­order effects) are statistically significant (see estimated path coef-
ficients and p-values under D12 ON D11, D13 ON D12 D11, C12 ON
C11, and C13 ON C12 C11). This indicates that a significant portion
of individual differences remained stable over time for both depression
and competencies. Nonetheless, individual differences were not perfectly
stable over time. Part of the “instability” in depression and competencies
 Structural Equation Models for Measuring Variability and Change   133

is accounted for by temporally preceding scores on the other construct,

respectively, as shown by significant cross-­lagged coefficients. The sig-
nificant cross-­lagged effects can be interpreted in terms of less competent
individuals showing higher depression scores at later time points, even
when the prior level of depression is controlled for. Conversely, children
reporting higher depression tend to report a lower level of competencies at
later occasions of measurement, even when the prior level of competence
is taken into account.
In this regard, it is particularly interesting that the estimated cross-­
lagged effects from depression to competence are larger (standardized
path coefficients bˆ standardized = –0.224 and –0.231) than the correspond-
ing cross-­lagged effects from competence to depression (standardized
path coefficients bˆ standardized = –0.156 and –0.100). Overall, the cross-­
lagged effects are relatively small in size compared to the autoregressive
coefficients, indicating that cross-­lagged effects account for only a small
(albeit significant) amount of individual differences in the model.
The correlations between the residual variables are statistically sig-
nificant and of medium size (rs = –.35 and –.34, ps < .001), indicating
that there is a small amount of shared variance in the depression and
competence scores at the same occasion of measurement, over and above
what can be explained by autoregression and cross-­lagged effects. These
correlations are likely due to shared situation-­specific effects that influ-
ence both constructs at the same time point.
The R 2 values indicate the amount of variability in the endogenous
(T2 and T3) variables that can be explained by the sum of autoregressive
and cross-­lagged effects. It turns out that, overall, 39% of individual dif-
ferences in depression and 47.4% of individual differences in competence
scores can be explained at T2. At T3, the amount of explained variability
is 34.3% for depression and 42.1% for competencies. The remaining per-
centages are due to other influences not represented in the model, shared
occasion-­specific effects and random measurement error. All these influ-
ences are represented by the residual variables ejk. Latent autoregressive
models, which separate random measurement error from reliable indi-
vidual differences, are discussed in the following section.

4.3.2 Latent Autoregressive Models

The models discussed in the previous section were based on manifest
variables. As a consequence, these models are plagued by the same prob-
lems as other manifest regression or path models (see Sections 3.2 and
134   Data Analysis with Mplus

3.5.2), namely, that measurement error is not explicitly accounted for and
that the path coefficients and their SEs may therefore be biased. In the
context of autoregressive models this can lead to bias in the estimation of
stability, change, and/or cross-­lagged effects and corresponding tests of
statistical significance.
Autoregressive models with latent variables (latent autoregressive
models, LAMs) explicitly account for random measurement error by
using multiple indicators at each time point. LAMs have the additional
advantages that they allow testing of measurement models, modeling of
indicator-­specific effects, and examining of the degree of measurement
invariance over time (cf. Section 4.1.4). These models were discussed in
detail by Jöreskog (1979a, 1979b) as well as by Hertzog and Nesselroade
(1987).
LAMs are based on the LS model described in Section 4.1.2. LAMs
restrict the covariances of the LS variables by imposing an autoregressive
structure. Cross-­lagged models are based on multiconstruct versions of
the basic LS model and additionally include cross-­lagged effects between
LS variables. Figure 4.18 shows a model for two constructs, each of which
is measured by two manifest variables Yijk at each of three time points (i
= indicator, j = construct, k = time point). Note that the structural model
(containing the paths between the LS variables) is equivalent to the struc-
ture of the manifest first-order autoregressive model with uncorrelated
residual variables in Figure 4.14.
A good starting point for a latent autoregressive analysis is the basic
LS model discussed in Section 4.1. This model can be used to clarify tech-
nical issues such as (1) the question of whether the basic measurement

BOX 4.9. Example of LAMs

In the following, we show how a multiple-­indicator LAM can be specified

for the depression and competence data. In this example, we use two indi-
cators for each LS at each time point, as illustrated in Figure 4.18. In order
to obtain two indicators, the depression and competence scales were split
into two test halves (“parcels”) at each time point by assigning the items to
two separate sum scores instead of a single one. The summary data for N =
569 children are provided in the file depression-­competence-­latent.dat on
the companion website. Among other advantages, the LAM allows us (1) to
clarify whether the effects in the manifest autoregressive model were partly
biased due to random measurement error, (2) to more properly estimate the
autoregressive and cross-­lagged coefficients, and (3) to quantify the amount
of (un)reliability in each scale.
 Structural Equation Models for Measuring Variability and Change   135

ε 111 ε 211 ε 112 ε 212 ε 113 ε 213

Y111 Y211 Y112 Y212 Y113 Y213

(d11) (d21) (d12) (d22) (d13) (d23)

λ111 λ211 λ112 λ212 λ113 λ213

ζ12 ζ13
State 11 β1 State 12 β2 State 13
(Depr1) (Depr2) (Depr3)
β5 β7

β6 β8
State 21 β3 State 22 β4 State 23
(Comp1) (Comp2) (Comp3)
ζ22 ζ23
λ121 λ221 λ122 λ222 λ123 λ223

Y121 Y221 Y122 Y222 Y123 Y223

(c11) (c21) (c12) (c22) (c13) (c23)

ε 121 ε 221 ε 122 ε 222 ε 123 ε 223

FIGURE 4.18. Latent autoregressive model for 12 manifest indicators Yijk (i = indi-
cator, j = construct, k = time point) measuring two constructs (j = 1, 2; depression
and competence) on three measurement occasions (k = 1, 2, 3). lijk = state factor
loadings; b1–b4 = latent path coefficients representing autoregressive effects; b5 –b8
= latent path coefficients representing cross-­lagged effects; eijk = measurement error
variables; zjk = LS residual variable. The variable names used in Mplus appear in
parentheses.

model fits for each construct, (2) whether indicator-­specific factors need
to be included due to indicator heterogeneity, and (3) which level of mea-
surement invariance can be assumed across time for each construct. The
final best-­fitting LS model can then serve as the basis for specifying LAMs
with or without cross-­lagged effects.
Figure 4.19 shows the Mplus input file for the estimation of the LAM
from Figure 4.18. For simplicity, we first show a simple model version
without indicator-­specific factors. Later on, we discuss a more complex
model that accounts for indicator-­specific effects.
136   Data Analysis with Mplus

FIGURE 4.19. Mplus input file for the specification of the simple latent autore-
gressive model for three measurement occasions shown in Figure 4.18.

The input shows that the basic measurement model specification is

identical to the LS model. The only difference is that we now consider two
constructs simultaneously. In the structural model, we include autore-
gressive and cross-­lagged effects using on statements. The command
depr3 with comp3@0; again ensures that the residual correlation for
time point 3 is set to zero in the initial model. (Mplus would otherwise
estimate this correlation as in the manifest model.)
The analysis of the model fit reveals that the model does not fit the
data well: c2 = 408.42, df = 45, p < .001, CFI = 0.93, RMSEA = 0.12, SRMR
= 0.05. This is shown especially by the very large chi-­square and RMSEA
values. For this reason, we do not discuss the parameter estimates for this
 Structural Equation Models for Measuring Variability and Change   137

model here (the full output can be found on the companion website). The
bad model fit can be explained by several factors. First, our analysis of the
manifest autoregressive model already indicated that the inclusion of cor-
related residuals and second-­order autoregressive effects was necessary. It
is likely that these effects are also required in the LAM. Furthermore, in
our multiple-­indicator LAM there may be indicator-­specific effects due to
the use of the same indicators over time (cf. Section 4.1.3).
The LAM can be extended in a similar way as the manifest model.
In addition, we can include indicator-­specific factors as discussed for the
LS model in Section 4.1.3. The extended model is shown in Figure 4.20
(p. 143). Figure 4.21 (p. 144) shows the corresponding Mplus model
command.
Even though the extended model still shows a significant chi-­square
value, the fit of the model is much improved compared to the model in
Figure 4.18, c2 = 46.72, df = 28, p = .02; CFI = 0.996; RMSEA = 0.03; SRMR
= 0.02. A comparison of the AIC values also reveals that the more complex
model (AIC = 2601.53) should be preferred over the initial model (AIC =
2929.24). In the following, we show the parameter estimates obtained for
the more complex model.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

DEPR1 BY
D11 1.000 0.000 999.000 999.000
D21 0.957 0.032 29.972 0.000

DEPR2 BY
D12 1.000 0.000 999.000 999.000
D22 0.909 0.030 30.500 0.000

DEPR3 BY
D13 1.000 0.000 999.000 999.000
D23 0.924 0.028 32.900 0.000

ISD BY
D21 1.000 0.000 999.000 999.000
D22 1.093 0.277 3.941 0.000
D23 1.385 0.425 3.256 0.001

COMP1 BY
C11 1.000 0.000 999.000 999.000
C21 1.031 0.059 17.501 0.000
138   Data Analysis with Mplus

COMP2 BY
C12 1.000 0.000 999.000 999.000
C22 0.913 0.054 16.840 0.000

COMP3 BY
C13 1.000 0.000 999.000 999.000
C23 0.857 0.049 17.346 0.000

ISC BY
C21 1.000 0.000 999.000 999.000
C22 1.167 0.165 7.084 0.000
C23 1.151 0.172 6.685 0.000

DEPR2 ON
DEPR1 0.720 0.055 13.141 0.000
COMP1 -0.030 0.034 -0.878 0.380

DEPR3 ON
DEPR2 0.590 0.063 9.406 0.000
DEPR1 0.153 0.051 3.033 0.002
COMP2 -0.049 0.033 -1.471 0.141

COMP2 ON
COMP1 0.693 0.071 9.715 0.000
DEPR1 -0.208 0.100 -2.080 0.037

COMP3 ON
COMP2 0.597 0.101 5.930 0.000
COMP1 0.195 0.073 2.679 0.007
DEPR2 -0.243 0.104 -2.338 0.019

ISD WITH
DEPR1 0.000 0.000 999.000 999.000
DEPR2 0.000 0.000 999.000 999.000
DEPR3 0.000 0.000 999.000 999.000
COMP1 0.000 0.002 0.151 0.880
COMP2 -0.002 0.002 -1.089 0.276
COMP3 0.000 0.002 -0.242 0.809

ISC WITH
COMP1 0.000 0.000 999.000 999.000
COMP2 0.000 0.000 999.000 999.000
COMP3 0.000 0.000 999.000 999.000
DEPR1 -0.011 0.004 -2.823 0.005
DEPR2 -0.007 0.003 -2.601 0.009
DEPR3 -0.008 0.003 -3.122 0.002
ISD -0.002 0.001 -1.622 0.105

DEPR2 WITH
COMP2 -0.035 0.004 -8.310 0.000

DEPR3 WITH
COMP3 -0.036 0.004 -8.584 0.000
 Structural Equation Models for Measuring Variability and Change   139

COMP1 WITH
DEPR1 -0.093 0.008 -11.127 0.000

Intercepts
D11 0.306 0.013 22.757 0.000
D21 0.274 0.013 21.672 0.000
D12 0.274 0.013 20.497 0.000
D22 0.253 0.013 20.079 0.000
D13 0.283 0.013 21.352 0.000
D23 0.256 0.013 20.152 0.000
C11 2.074 0.024 85.370 0.000
C21 2.121 0.026 82.201 0.000
C12 2.091 0.024 88.301 0.000
C22 2.115 0.025 83.180 0.000
C13 2.082 0.025 82.382 0.000
C23 2.184 0.025 85.819 0.000

Variances
DEPR1 0.085 0.006 13.491 0.000
ISD 0.003 0.001 2.508 0.012
COMP1 0.225 0.021 10.628 0.000
ISC 0.054 0.012 4.578 0.000

Residual Variances
D11 0.018 0.002 8.220 0.000
D21 0.010 0.002 5.108 0.000
D12 0.014 0.002 6.928 0.000
D22 0.014 0.002 7.801 0.000
D13 0.010 0.002 5.251 0.000
D23 0.008 0.002 3.802 0.000
C11 0.111 0.012 9.473 0.000
C21 0.086 0.011 7.648 0.000
C12 0.084 0.011 7.546 0.000
C22 0.094 0.012 8.178 0.000
C13 0.087 0.013 6.982 0.000
C23 0.085 0.011 7.693 0.000
DEPR2 0.039 0.003 11.783 0.000
DEPR3 0.038 0.003 12.569 0.000
COMP2 0.096 0.013 7.650 0.000
COMP3 0.102 0.013 7.901 0.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

DEPR1 BY
D11 0.909 0.013 72.457 0.000
D21 0.925 0.012 75.929 0.000

DEPR2 BY
D12 0.929 0.011 82.978 0.000
D22 0.896 0.013 71.363 0.000
140   Data Analysis with Mplus

DEPR3 BY
D13 0.948 0.011 90.229 0.000
D23 0.914 0.012 76.546 0.000

ISD BY
D21 0.193 0.038 5.013 0.000
D22 0.211 0.041 5.215 0.000
D23 0.266 0.046 5.721 0.000

COMP1 BY
C11 0.818 0.022 36.435 0.000
C21 0.794 0.023 34.255 0.000

COMP2 BY
C12 0.858 0.021 40.148 0.000
C22 0.729 0.026 28.125 0.000

COMP3 BY
C13 0.871 0.021 42.281 0.000
C23 0.741 0.026 28.907 0.000

ISC BY
C21 0.377 0.040 9.400 0.000
C22 0.446 0.043 10.403 0.000
C23 0.439 0.043 10.224 0.000

DEPR2 ON
DEPR1 0.708 0.046 15.362 0.000
COMP1 -0.048 0.055 -0.880 0.379

DEPR3 ON
DEPR2 0.583 0.057 10.175 0.000
DEPR1 0.149 0.049 3.048 0.002
COMP2 -0.078 0.053 -1.475 0.140

COMP2 ON
COMP1 0.678 0.057 11.966 0.000
DEPR1 -0.125 0.060 -2.080 0.038

COMP3 ON
COMP2 0.551 0.084 6.591 0.000
COMP1 0.176 0.065 2.699 0.007
DEPR2 -0.137 0.059 -2.330 0.020

ISD WITH
DEPR1 0.000 0.000 999.000 999.000
DEPR2 0.000 0.000 999.000 999.000
DEPR3 0.000 0.000 999.000 999.000
COMP1 0.010 0.064 0.150 0.880
COMP2 -0.091 0.082 -1.111 0.267
COMP3 -0.020 0.083 -0.241 0.809
 Structural Equation Models for Measuring Variability and Change   141

ISC WITH
COMP1 0.000 0.000 999.000 999.000
COMP2 0.000 0.000 999.000 999.000
COMP3 0.000 0.000 999.000 999.000
DEPR1 -0.169 0.051 -3.295 0.001
DEPR2 -0.157 0.059 -2.691 0.007
DEPR3 -0.185 0.056 -3.307 0.001
ISD -0.152 0.087 -1.741 0.082

DEPR2 WITH
COMP2 -0.570 0.051 -11.108 0.000

DEPR3 WITH
COMP3 -0.570 0.049 -11.625 0.000

COMP1 WITH
DEPR1 -0.675 0.032 -21.093 0.000

Intercepts
D11 0.954 0.051 18.870 0.000
D21 0.909 0.050 18.254 0.000
D12 0.859 0.049 17.521 0.000
D22 0.842 0.049 17.264 0.000
D13 0.895 0.050 18.049 0.000
D23 0.845 0.049 17.323 0.000
C11 3.579 0.114 31.374 0.000
C21 3.446 0.110 31.414 0.000
C12 3.702 0.117 31.514 0.000
C22 3.487 0.111 31.443 0.000
C13 3.454 0.111 31.226 0.000
C23 3.598 0.114 31.647 0.000

Variances
DEPR1 1.000 0.000 999.000 999.000
ISD 1.000 0.000 999.000 999.000
COMP1 1.000 0.000 999.000 999.000
ISC 1.000 0.000 999.000 999.000

Residual Variances
D11 0.174 0.023 7.652 0.000
D21 0.107 0.022 4.907 0.000
D12 0.136 0.021 6.538 0.000
D22 0.152 0.021 7.249 0.000
D13 0.101 0.020 5.060 0.000
D23 0.091 0.025 3.703 0.000
C11 0.331 0.037 9.004 0.000
C21 0.227 0.032 7.146 0.000
C12 0.265 0.037 7.219 0.000
C22 0.256 0.033 7.652 0.000
C13 0.241 0.036 6.702 0.000
C23 0.230 0.032 7.209 0.000
142   Data Analysis with Mplus

DEPR2 0.450 0.034 13.290 0.000

DEPR3 0.425 0.031 13.604 0.000
COMP2 0.410 0.043 9.547 0.000
COMP3 0.369 0.038 9.642 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.826 0.023 36.229 0.000

D21 0.893 0.022 41.054 0.000
D12 0.864 0.021 41.489 0.000
D22 0.848 0.021 40.347 0.000
D13 0.899 0.020 45.115 0.000
D23 0.909 0.025 36.837 0.000
C11 0.669 0.037 18.218 0.000
C21 0.773 0.032 24.327 0.000
C12 0.735 0.037 20.074 0.000
C22 0.744 0.033 22.263 0.000
C13 0.759 0.036 21.141 0.000
C23 0.770 0.032 24.147 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

DEPR2 0.550 0.034 16.267 0.000

DEPR3 0.575 0.031 18.403 0.000
COMP2 0.590 0.043 13.765 0.000
COMP3 0.631 0.038 16.494 0.000

The parameter estimates pertaining to the structural model show

some interesting differences compared to the manifest autoregressive
model. Overall, we can see that the estimated R 2 values for the structural
model (R-SQUARE—Latent Variable) are much larger in the LAM
(.55 ≤ R 2 ≤ .63) than the corresponding values in the manifest model (.39
≤ R 2 ≤ .47). In other words, the LAM can explain a much larger amount of
variability in the T2 and T3 depression and competence constructs than
the manifest model.
On the other hand, the cross-­lagged effects from competence to
depression that were found to be significant in the manifest path model
turn out to be nonsignificant in the LAM (whereas the cross-­lagged effects
from depression to competence are still significant). The larger R 2 values
in the LAM can primarily be explained by the fact that the first-order
autoregressive effects are estimated to be larger in this model [standard-
ized first-order latent autoregressive path coefficients: 0.55 ≤ bˆ standardized
 Structural Equation Models for Measuring Variability and Change   143

IS21
(isd)
λIS211 λIS213
λIS212
ε 111 ε 211 ε 112 ε 212 ε 113 ε 213

Y111 Y211 Y112 Y212 Y113 Y213

(d11) (d21) (d12) (d22) (d13) (d23)

λ111 λ211 λ112 λ212 λ113 λ213

β9
ζ12 ζ13
State 11 State 12 State 13
(Depr1) β1 (Depr2) β2 (Depr3)
β5 β7
β6 β8
State 21 State 22 State 23
(Komp1) β3 (Komp2) β4 (Komp3)
β10 ζ22 ζ23
λ121 λ221 λ122 λ222 λ123 λ223

Y121 Y221 Y122 Y222 Y123 Y223

(c11) (c21) (c12) (c22) (c13) (c23)

ε 121 ε 221 ε 122 ε 222 ε 123 ε 223

λIS221 λIS222 λIS223
IS22
(isc)

FIGURE 4.20. Latent autoregressive model for 12 manifest indicators Yijk (i =

indicator, j = construct, k = time point) measuring two constructs (j = 1, 2; depres-
sion and competence) on three measurement occasions (k = 1, 2, 3). lijk = state factor
loadings; lIS2jk = loadings on the indicator-­specific factors IS2j; b1–b4 = latent path
coefficients representing first-order autoregressive effects; b9 and b10 = latent path
coefficients representing second-­order autoregressive effects; b5 –b8 = latent path
coefficients representing cross-­lagged effects; eijk = measurement error variables; zjk
= latent state residual variable. In contrast to the model shown in Figure 4.18, this
model allows for second-­order autoregressive effects, correlated residual variables
at the same measurement occasion, and indicator-­specific factors IS2j for the sec-
ond indicator. To avoid cluttering, not all possible correlations between variables are
included in the picture. The variable names used in Mplus appear in parentheses.
144   Data Analysis with Mplus

FIGURE 4.21. Mplus model command for the specification of the extended
latent autoregressive model with second-­ order autoregressive effects, correlated
latent residual variables, and indicator-­specific factors shown in Figure 4.20.

≤ 0.71; see results reported under STANDARDIZED MODEL RESULTS

(STDYX Standardization)] compared to the manifest model (0.36
≤ bˆ standardized ≤ 0.53). This finding indicates an underestimation of the
actual stability of individual differences in the manifest model, which is
likely due to the fact that measurement error was not explicitly taken into
account in the manifest model.
In summary, the analyses again illustrate the advantages of using
latent variable models for analyzing longitudinal data. These mod-
els correct for measurement error and therefore allow for a more accu-
rate analysis of complex relationships between variables. In the present
example, especially, the scales measuring competencies show relatively
low reliabilities (.67 ≤ R 2 ≤ .77; see values reported under R-SQUARE—
Observed Variable, variables C11–C23). Likewise, the indicators
of depression have high, albeit less than perfect, reliabilities (.83 ≤ R 2
≤ .91; see R-SQUARE—Observed Variable, variables D11–D23).
 Structural Equation Models for Measuring Variability and Change   145

Imperfect reliabilities are one reason for the differences seen in the results
between the manifest and the latent autoregressive models. An additional
reason may be the relatively large amounts of specific variance in the
competence test halves (the competence scale is composed of rather het-
erogeneous items), as indicated by the relatively large occasion-­specific
effects (standardized loadings on the indicator-­ specific factor ISC:
.377 ≤ lˆ standardized
≤ .446).
As in an LS analysis, the degree of invariance of the parameters of
the measurement model over time should be examined for the LAM in
subsequent analysis steps. (The possibility to test for measurement invari-
ance is another advantage of the LAM over the manifest model.) I do not
discuss such an analysis here in detail given that the principle of invari-
ance testing was already illustrated in detail for the LS model in Section
4.1.4. The steps shown for the LS model can be transferred to the LAM.
The companion website contains a sample input file for a version of the
LAM assuming strong factorial invariance across time.

4.4 Latent Change Models

In latent change (LC) models (McArdle & Hamagami, 2001; Raykov, 1993;
Steyer et al., 1997, 2000), change is not measured “indirectly” through
the residual variables in autoregressions (as in autoregressive models), but
directly through so-­called latent difference variables. Latent difference vari-
ables represent interindividual differences in true intraindividual change
over time—that is, change scores corrected for random measurement
error. For this reason, these types of models have also been referred to as
true change models (Steyer et al., 1997). Because latent difference variables
are of key importance in this approach, LC models are sometimes also
called latent difference models.
Latent difference variables can be included in SEMs in a relatively
simple way by reformulating LS models (see Section 4.1). Figure 4.22 illus-
trates the reparameterization of an LS model (A) to an LC model (B) based
on a single construct measured by two indicators on two measurement
occasions. The basic idea of the LC model is that the LS variable at time
2 (State 2) is a function of (1 times) the initial LS variable (State 1) plus (1
times) change (in terms of the latent difference variable State 2 – State 1):

State 2 = 1 State 1 + 1 (State 2 – State 1)

146   Data Analysis with Mplus

A
State 1 State 2
(state1) (state2)

λ1 λ2 λ1 λ2

Y11 Y21 Y12 Y22

(d11) (d21) (d12) (d22)

ε 11 ε 21 ε 12 ε 22

B State 2–State 1
(diff2_1)
1 0
State 1 1 State 2
(state1) (state2)

λ1 λ2 λ1 λ2

Y11 Y21 Y12 Y22

(d11) (d21) (d12) (d22)

ε 11 ε 21 ε 12 ε 22

FIGURE 4.22. LS model and corresponding latent change (LC) model for depres-
sion measured by two indicators Yik (i = indicator, k = time point) on two measure-
ment occasions: li = time-­invariant state factor loadings; eik = measurement error
variable. A: LS model. B: LC model. The variable names used in the Mplus appear
in parentheses.

Note that this decomposition is not at all restrictive: We add one variable
(the factor State 1) to another variable (here the factor State 2) and subtract
this variable again at the same time. This decomposition into initial state
plus change can be accomplished in an LS model through the inclusion
of an additional latent variable, the latent difference variable (State 2 –
State 1), as shown in Figure 4.22B. This means that the latent difference
becomes an explicit part of the model in the form of an additional latent
variable, which makes it possible to study interindividual differences in
intraindividual change at the latent (error-free) level.
A latent difference variable can be used like any other latent vari-
able in an SEM. For example, the latent difference variable can be cor-
related with other variables or it can serve as an exogenous (predictor)
 Structural Equation Models for Measuring Variability and Change   147

or endogenous (criterion) variable in extended SEMs. Often, researchers

are interested in explaining individual differences in change over time
(i.e., why do some people change more than others?). In such a case, the
latent difference model could be used as an endogenous variable that is
predicted (or “explained”) by other variables. In addition, the mean of the
latent difference variable can be estimated and used as a measure of mean
differences across time (e.g., does significant mean change occur?).
Note that in the above equation, the variable State 2 can be seen as
being “regressed on” State 1 as well as the latent difference variable (State
2 – State 1). Technically, one can see this as a “perfect” (or deterministic)
regression analysis, because there is no residual variable in the equation.
This can also be seen by the fact that the path diagram in Figure 4.22B
explicitly includes a residual of zero for the endogenous variable State 2. It
is important to emphasize this aspect because it is relevant for the model
specification in Mplus, in which the residual variance of each endogenous
variable is estimated by default and thus has to be set to zero by the user,
in this case.
Another important issue to consider when using LC models is that
they require at least strong factorial invariance with regard to the mea-
surement model (cf. Section 4.1.4). Strong factorial invariance is a prereq-
uisite for a meaningful interpretation of latent difference scores, because
it ensures that both State 1 and State 2 are measured in the same metric
(equal loadings) and have the same origin of measurement (equal inter-
cepts). Strong measurement invariance thus ensures that we are not sub-
tracting apples from oranges.

BOX 4.10. Example of LC Models

As a substantive example, we once again consider the LS model for depres-

sion from Section 4.1. We might be interested in decomposing the Time 2 LS
factor for depression (Depression T2) into initial state (factor Depression T1)
plus change (Depression T2 – Depression T1):

Depression T2 = Depression T1 + (Depression T2 – Depression T1)

This decomposition allows us to examine the degree of individual differences

in change over time with regard to depression (by means of the variance of
the latent difference factor). It also allows us to relate individual differences
in change to external variables; for example, to explain why some people’s
depression scores changed more than others.
148   Data Analysis with Mplus

The assumption of time-­invariant loadings (as one required aspect of

strong factorial invariance) is reflected in Figure 4.22 by the fact that the
loading parameters li carry an index only for the variable, but not for the
occasion of measurement (hence, they do not depend on the time point
but are the same at each measurement occasion). The time-­invariant inter-
cepts are not shown in the figure. Implementing the assumption of strong
factorial invariance is critical in the model specification in Mplus. Equal-
ity constraints, as discussed in detail in Section 4.1.4, have to be imposed
to obtain time-­invariant loadings and intercepts.
For more than two time points, Steyer et al. (1997, 2000) distinguish
between a baseline and a neighbor change version of the LC model. In
the baseline change version, change is assessed with regard to a specific
baseline situation or time point (often the first time point). In this model,
for example, change from the first to the second, the first to the third, the
first to the fourth time point, and so on, is examined.
In contrast, neighbor change models always consider change between
adjacent time points; for example, change between the first and the sec-
ond, the second and the third, and the third and the fourth time points.
Figure 4.23 shows a baseline (A) and a neighbor (B) LC model for three
time points.
In this case, the difference between the baseline and the neighbor
change version is that the third LS variable (State 3) is decomposed into
initial state (State 1) plus change from time 1 to time 3 (State 3 – State 1) in
the baseline change model:

BOX 4.11. Data Example of LC Models in Mplus

We use the LS model for depression from Section 4.1.4 (data set depression.
dat) as the basis for our illustration of the analysis of LC models in Mplus.
As the starting point, we use the LS model version with an indicator-­specific
factor for the second indicator and strong factorial invariance (in which the
factor loadings and intercepts are assumed to be equal across time). Given
that we have already tested the assumption of strong factorial invariance in
Section 4.1.4 (and found that we did not have to reject it), we can directly
start with the analysis of LC models. In general, the assumption of strong
factorial invariance has to be tested first before one can start an LC analysis.
The LS model is equivalent to both a baseline and a neighbor change model;
that is, all three models imply the same mean and covariance structure. The
reason is that the LC models are just reparameterizations of the LS model.
Therefore, the fit of all three models is also identical.
 Structural Equation Models for Measuring Variability and Change   149

A
State 2–State 1 State 3–State 1
(diff2_1) (diff3_1)
1 1
State 1 State 2 0 State 3
0
1
(state1) (state2) 1 (state3)

λ1 λ2 λ1 λ2 λ1 λ2

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

B
State 2–State 1 State 3–State 2
(diff2_1) (diff3_2)
1 1 1
State 1 State 2 0 State 3
0
1
(state1) (state2) 1 (state3)

λ1 λ2 λ1 λ2 λ1 λ2

Y11 Y21 Y12 Y22 Y13 Y23

(d11) (d21) (d12) (d22) (d13) (d23)

ε 11 ε 21 ε 12 ε 22 ε 13 ε 23

FIGURE 4.23. LC model for two indicators Yik (i = indicator, k = time point) and
three measurement occasions. li = time-­invariant state factor loadings; eik = mea-
surement error variable. A: Baseline change model. B: Neighbor change model. The
models are equivalent and equivalent to an LS model with invariant parameters. The
(often required) indicator-­specific factor for the second indicator is not shown here
to save space. The variable names used in the Mplus input appear in parentheses.

State 3 = 1 · State 1 + 1 · (State 3 – State 1)

whereas in the neighbor change model, State 3 is decomposed into the

initial state plus two separate change components:

State 3 = 1 · State 1 + 1 · (State 2 – State 1) + 1 · (State 3 – State 2)

Figure 4.24 shows the Mplus model command for the specification
of the baseline change model. (The specification of the neighbor change
150   Data Analysis with Mplus

model is shown in Figure 4.25, page 158.) The remaining input does not
differ from the input for the LS model (cf. Figure 4.9).
The measurement models for the three LS factors are formulated in
analogy to the LS model with time-­invariant loadings. According to the
parameterization in Figure 4.23, the latent difference variables are not
directly connected with observed indicators. Steyer et al. (1997, 2000)
discuss an alternative equivalent parameterization of LC models in which
the latent difference variables are directly connected to observed indica-
tors. For didactic reasons, I did not choose this parameterization here.

FIGURE 4.24. Mplus model command for the specification of the baseline
change model shown in Figure 4.23A. An additional indicator-­specific factor for the
second indicator was included in the model. (This factor is not shown in Figure
4.23.)
 Structural Equation Models for Measuring Variability and Change   151

However, both types of parameterizations can be found in separate input

files on the companion website.
Given that, in the parameterization chosen here, the latent differ-
ence variables diff2_1 and diff3_1 are not directly connected to
the observed variables, we have to use a trick to include them in the
model. The reason is that Mplus requires all latent variables to be defined
via a by statement. In order to define the latent difference variables, we
therefore simply choose an arbitrary observed variable that is listed under
usevariables (e.g., d11) and specify that this variable has a loading
of 0 on each latent difference variable:

diff2_1 by d11@0;
diff3_1 by d11@0;

These commands thus merely serve to introduce the names of the latent
variables diff2_1 and diff3_1—they have no deeper implication for
the model specification other than defining the names of the latent differ-
ence variables so that they can be used in the model. The relevant com-
mands that assign the meaning as difference scores to these variables are
shown next. We use on commands to specify that the LS variables at T2
and T3 are functions of the initial LS variable (state1) plus the latent
difference, respectively:

state2 on state1@1 diff2_1@1;

state3 on state1@1 diff3_1@1;

The “regression coefficients” of these regressions are fixed to 1 using the

command @1, because the implicit specification of the LC model is that,
for example, state2 is equal to one times the initial LS variable (state1)
plus 1 times the latent difference diff2_1.
Furthermore, we have to fix the residuals of these regressions to zero
because Mplus would otherwise automatically add a residual term to each
regression equation (which would not be correct, given that state2 and
state3 are fully determined by initial status and change in this model):

state2@0;
state3@0;

The latent difference variables diff2_1 and diff3_1 obtain their

meaning as latent difference variables through the above specifications.
152   Data Analysis with Mplus

Given that the endogenous variables state2 and state3 are per-
fectly determined by initial state and change in this model, their residual
variables were fixed to zero. As a consequence, the residual terms can-
not be correlated with other variables. These correlations have to be sup-
pressed in Mplus, because otherwise the model would be misspecified
and would also, very likely, produce estimation problems:

state2 with state3@0 diff3_1@0;

state3 with diff2_1@0;

The following commands serve to specify the indicator-­specific factor

for the second indicator (is2), as in the LS model (cf. Section 4.1.3). All
loadings on this factor are fixed to 1 to keep them invariant across time.
The with command is used to set the correlations of the is2 factor with
all other latent variables to zero.

is2 by d21 d22@1 d23@1;

is2 with state1@0 state2@0 state3@0 diff2_1@0 diff3_1@0;

All remaining commands concern the mean structure of the model. We

first set the intercepts of the reference indicators to 0 to identify the latent
means (cf. Section 4.1.4):

[d11@0 d12@0 d13@0];

Next, we request the estimation of the latent means for state1 as well
as for the two latent difference variables diff2_1 and diff3_1 (other-
wise, Mplus would keep these means fixed at 0 by default):

[state1 diff2_1 diff3_1];

Finally, the intercepts of the remaining (non-­reference) indicators are set

equal across time to establish strong factorial invariance:

[d21 d22 d23] (2);

As expected, the fit of the baseline change model is identical to the

fit of the LS model with strong factorial invariance, c2 = 6.656, df = 9, p
= .6729, RMSEA = .00, p(RMSEA ≤ .05 = .95, SRMR = .014, CFI = 1.00
(cf. Section 4.1.4; the complete output can be found on the companion
website).
 Structural Equation Models for Measuring Variability and Change   153

In the following MODEL RESULTS, we can see that the parameters

of the measurement model are also the same as in the LS model with
strong factorial invariance. The output now additionally contains the
covariances (under MODEL RESULTS) and correlations [under STAN-
DARDIZED MODEL RESULTS (STDYX Standardization)] of
the latent difference variables diff2_1 and diff3_1 with the initial
LS factor (state1) as well as between the two latent difference variables.

MODEL RESULTS (Baseline Change Model)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 1.000 0.000 999.000 999.000
D21 0.738 0.030 24.997 0.000

STATE2 BY
D12 1.000 0.000 999.000 999.000
D22 0.738 0.030 24.997 0.000

STATE3 BY
D13 1.000 0.000 999.000 999.000
D23 0.738 0.030 24.997 0.000

DIFF2_1 BY
D11 0.000 0.000 999.000 999.000

DIFF3_1 BY
D11 0.000 0.000 999.000 999.000

IS2 BY
D21 1.000 0.000 999.000 999.000
D22 1.000 0.000 999.000 999.000
D23 1.000 0.000 999.000 999.000

STATE2 ON
STATE1 1.000 0.000 999.000 999.000
DIFF2_1 1.000 0.000 999.000 999.000

STATE3 ON
STATE1 1.000 0.000 999.000 999.000
DIFF3_1 1.000 0.000 999.000 999.000

STATE2 WITH
STATE3 0.000 0.000 999.000 999.000
DIFF3_1 0.000 0.000 999.000 999.000

STATE3 WITH
DIFF2_1 0.000 0.000 999.000 999.000
154   Data Analysis with Mplus

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000
DIFF2_1 0.000 0.000 999.000 999.000
DIFF3_1 0.000 0.000 999.000 999.000

DIFF2_1 WITH
STATE1 -0.077 0.015 -5.324 0.000

DIFF3_1 WITH
STATE1 -0.088 0.015 -5.856 0.000
DIFF2_1 0.099 0.015 6.451 0.000

Means
STATE1 1.583 0.026 60.157 0.000
DIFF2_1 0.056 0.027 2.078 0.038
DIFF3_1 0.052 0.027 1.900 0.057

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.187 0.049 3.861 0.000
D12 0.000 0.000 999.000 999.000
D22 0.187 0.049 3.861 0.000
D13 0.000 0.000 999.000 999.000
D23 0.187 0.049 3.861 0.000

Variances
STATE1 0.195 0.019 10.510 0.000
DIFF2_1 0.193 0.021 9.385 0.000
DIFF3_1 0.198 0.021 9.453 0.000
IS2 0.007 0.002 3.433 0.001

Residual Variances
D11 0.040 0.009 4.477 0.000
D21 0.028 0.005 5.425 0.000
D12 0.022 0.009 2.453 0.014
D22 0.035 0.006 6.299 0.000
D13 0.019 0.009 2.192 0.028
D23 0.037 0.006 6.608 0.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 0.911 0.021 43.988 0.000
D21 0.867 0.022 39.895 0.000
 Structural Equation Models for Measuring Variability and Change   155

STATE2 BY
D12 0.956 0.018 52.311 0.000
D22 0.867 0.020 43.500 0.000

STATE3 BY
D13 0.959 0.019 50.014 0.000
D23 0.854 0.021 41.331 0.000

DIFF2_1 BY
D11 0.000 0.000 999.000 999.000

DIFF3_1 BY
D11 0.000 0.000 999.000 999.000

IS2 BY
D21 0.220 0.033 6.732 0.000
D22 0.202 0.030 6.720 0.000
D23 0.205 0.031 6.733 0.000

STATE2 ON
STATE1 0.915 0.049 18.583 0.000
DIFF2_1 0.909 0.049 18.466 0.000

STATE3 ON
STATE1 0.946 0.052 18.195 0.000
DIFF3_1 0.953 0.052 18.430 0.000

STATE2 WITH
STATE3 999.000 999.000 999.000 999.000
DIFF3_1 999.000 999.000 999.000 999.000

STATE3 WITH
DIFF2_1 999.000 999.000 999.000 999.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 999.000 999.000 999.000 999.000
STATE3 999.000 999.000 999.000 999.000
DIFF2_1 0.000 0.000 999.000 999.000
DIFF3_1 0.000 0.000 999.000 999.000

DIFF2_1 WITH
STATE1 -0.399 0.054 -7.345 0.000

DIFF3_1 WITH
STATE1 -0.446 0.052 -8.646 0.000
DIFF2_1 0.508 0.050 10.167 0.000

Means
STATE1 3.583 0.181 19.792 0.000
DIFF2_1 0.127 0.061 2.069 0.039
DIFF3_1 0.116 0.061 1.892 0.058
156   Data Analysis with Mplus

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.498 0.134 3.706 0.000
D12 0.000 0.000 999.000 999.000
D22 0.456 0.124 3.677 0.000
D13 0.000 0.000 999.000 999.000
D23 0.465 0.126 3.686 0.000

Variances
STATE1 1.000 0.000 999.000 999.000
DIFF2_1 1.000 0.000 999.000 999.000
DIFF3_1 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000

Residual Variances
D11 0.171 0.038 4.532 0.000
D21 0.200 0.036 5.535 0.000
D12 0.086 0.035 2.448 0.014
D22 0.208 0.033 6.332 0.000
D13 0.081 0.037 2.191 0.028
D23 0.228 0.034 6.767 0.000
STATE2 0.000 999.000 999.000 999.000
STATE3 0.000 999.000 999.000 999.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.829 0.038 21.994 0.000

D21 0.800 0.036 22.155 0.000
D12 0.914 0.035 26.155 0.000
D22 0.792 0.033 24.165 0.000
D13 0.919 0.037 25.007 0.000
D23 0.772 0.034 22.941 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

STATE2 1.000 999.000 999.000 999.000

STATE3 1.000 999.000 999.000 999.000

In addition, we obtain the latent means of state1, diff2_1, and

diff3_1. If we compare these values to the LS factor means obtained in
the LS model with strong factorial invariance (M1 = 1.583, M2 = 1.639, M3
= 1.634), we can see that (1) the mean of state1 is identical, as expected,
and that (2) the mean of diff2_1 (0.056) is exactly equal to the differ-
ence between the means of state2 and state1:
 Structural Equation Models for Measuring Variability and Change   157

1.639 – 1.583 = 0.056

Likewise, the mean of diff3_1 (0.052) is equal, within rounding error,

to the mean difference between state3 and state1:

1.634 – 1.583 = 0.051

The positive signs of the latent difference factor means indicate that
the latent depression scores have (slightly) increased between time 1 and
time 2 as well as between time 1 and time 3.
The estimated p-values for the latent difference factor means indicate
that the mean difference between T1 and T2 is significantly different from
zero at the .05 level for a two-­tailed test, z = 2.078, p = .038, indicating
that the increase in depression from T1 to T2 was statistically significant.
The mean difference between T1 and T3 failed to reach statistical signifi-
cance at the .05 level (at least for a two-­tailed test), z = 1.900, p = 0.057.
The estimated variances of state1 (0.195), diff2_1 (0.193), and
diff3_1 (0.198) are significantly different from zero (all ps < .001), show-
ing that there are individual differences in the initial latent depression
state as well as in the LC scores. Notice that the R 2 values for state2 and
state3 equal 1, reflecting the fact that these state factors are, by defini-
tion, perfectly determined by initial state plus change in this model.
Figure 4.25 shows the model statement for the specification of the
neighbor change model for the same data example. The specification of
the latent difference variable diff2_1 is identical to the specification of
this variable in the baseline change model. This variable still represents
change from time 1 to time 2. The second latent difference variable in
the model is now labeled diff3_2 to make clear that in the neighbor
change version, we consider change between the adjacent time points 2
and 3 (rather than between time 1 and 3). According to the path diagram
in Figure 4.23B, the latent state variable state3 is now determined by
three components: (1) the initial state factor (state1), (2) LC between
time 1 and time 2 (diff2_1), and (3) LC between time 2 and time 3
(diff3_2). Therefore, the on statement for state3 now lists all three
components and sets the paths of all these components to 1 (using the
@1 command). The remaining commands are analogous to the baseline
change model.
The model fit of the neighbor change model is again identical to the
fit of the LS model with strong factorial invariance as well as to the fit of
the baseline change model (see Mplus outputs on the companion website
158   Data Analysis with Mplus

FIGURE 4.25. Mplus model command for the specification of the neighbor
change model shown in Figure 4.23B. An additional indicator-­specific factor for the
second indicator was included in the model. (This factor is not shown in Figure
4.23.)

for details). The parameter estimates of the measurement model (loadings,

intercepts, and residual variances) are also identical to the corresponding
loadings in the LS and baseline change model. In the structural model,
we now obtain the covariances and correlations between the initial state
factor state1 and the latent difference variables diff2_1 (as in the
baseline change model) and diff3_2. Furthermore, the covariances/
correlations between diff2_1 and diff3_2 are estimated, as are the
means of state1, diff2_1, and diff3_2. The means of state1 and
 Structural Equation Models for Measuring Variability and Change   159

diff2_1 are identical to the means in the baseline change model (M1 =
1.583, M2–1 = 0.056). The mean of the latent difference variable diff3_2
is estimated to M3–2 = –0.004. Within rounding error, this is equal to the
difference in means between state3 and state2 in the LS model with
strong factorial invariance. The mean difference of –0.004 is not statisti-
cally significant, z = –0.163, p = .871. We can conclude that the average
level of depression has not significantly changed between T2 and T3.
In addition, we again obtain estimates for the variances of state1
(0.195), diff2_1 (0.193), and diff3_2 (0.192). These variances indi-
cate the degree to which individuals differ in their initial latent depression
scores as well as in their LC scores. In additional analysis steps, one could
add external variables (covariates) to the model that may be able to partly
or fully explain individual differences in initial status and change (for a
substantive example, see Steyer et al., 2000).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 1.000 0.000 999.000 999.000
D21 0.738 0.030 24.997 0.000

STATE2 BY
D12 1.000 0.000 999.000 999.000
D22 0.738 0.030 24.997 0.000

STATE3 BY
D13 1.000 0.000 999.000 999.000
D23 0.738 0.030 24.997 0.000

DIFF2_1 BY
D11 0.000 0.000 999.000 999.000

DIFF3_2 BY
D11 0.000 0.000 999.000 999.000

IS2 BY
D21 1.000 0.000 999.000 999.000
D22 1.000 0.000 999.000 999.000
D23 1.000 0.000 999.000 999.000

STATE2 ON
STATE1 1.000 0.000 999.000 999.000
DIFF2_1 1.000 0.000 999.000 999.000
160   Data Analysis with Mplus

STATE3 ON
STATE1 1.000 0.000 999.000 999.000
DIFF2_1 1.000 0.000 999.000 999.000
DIFF3_2 1.000 0.000 999.000 999.000

STATE2 WITH
STATE3 0.000 0.000 999.000 999.000
DIFF3_2 0.000 0.000 999.000 999.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000
DIFF2_1 0.000 0.000 999.000 999.000
DIFF3_2 0.000 0.000 999.000 999.000

DIFF2_1 WITH
STATE1 -0.077 0.015 -5.324 0.000

DIFF3_2 WITH
STATE1 -0.010 0.012 -0.832 0.405
DIFF2_1 -0.093 0.015 -6.142 0.000

Means
STATE1 1.583 0.026 60.157 0.000
DIFF2_1 0.056 0.027 2.078 0.038
DIFF3_2 -0.004 0.026 -0.163 0.871

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.187 0.049 3.861 0.000
D12 0.000 0.000 999.000 999.000
D22 0.187 0.049 3.861 0.000
D13 0.000 0.000 999.000 999.000
D23 0.187 0.049 3.861 0.000

Variances
STATE1 0.195 0.019 10.510 0.000
DIFF2_1 0.193 0.021 9.385 0.000
DIFF3_2 0.192 0.021 9.341 0.000
IS2 0.007 0.002 3.433 0.001

Residual Variances
D11 0.040 0.009 4.477 0.000
D21 0.028 0.005 5.425 0.000
D12 0.022 0.009 2.453 0.014
D22 0.035 0.006 6.299 0.000
D13 0.019 0.009 2.192 0.028
D23 0.037 0.006 6.608 0.000
STATE2 0.000 0.000 999.000 999.000
STATE3 0.000 0.000 999.000 999.000
 Structural Equation Models for Measuring Variability and Change   161

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

STATE1 BY
D11 0.911 0.021 43.988 0.000
D21 0.867 0.022 39.895 0.000

STATE2 BY
D12 0.956 0.018 52.311 0.000
D22 0.867 0.020 43.501 0.000

STATE3 BY
D13 0.959 0.019 50.014 0.000
D23 0.854 0.021 41.331 0.000

DIFF2_1 BY
D11 0.000 0.000 999.000 999.000

DIFF3_2 BY
D11 0.000 0.000 999.000 999.000

IS2 BY
D21 0.220 0.033 6.732 0.000
D22 0.202 0.030 6.720 0.000
D23 0.205 0.031 6.733 0.000

STATE2 ON
STATE1 0.915 0.049 18.583 0.000
DIFF2_1 0.909 0.049 18.466 0.000

STATE3 ON
STATE1 0.946 0.052 18.195 0.000
DIFF2_1 0.939 0.061 15.318 0.000
DIFF3_2 0.939 0.054 17.503 0.000

STATE2 WITH
STATE3 999.000 999.000 999.000 999.000
DIFF3_2 999.000 999.000 999.000 999.000

IS2 WITH
STATE1 0.000 0.000 999.000 999.000
STATE2 999.000 999.000 999.000 999.000
STATE3 999.000 999.000 999.000 999.000
DIFF2_1 0.000 0.000 999.000 999.000
DIFF3_2 0.000 0.000 999.000 999.000

DIFF2_1 WITH
STATE1 -0.399 0.054 -7.345 0.000
162   Data Analysis with Mplus

DIFF3_2 WITH
STATE1 -0.053 0.064 -0.835 0.403
DIFF2_1 -0.485 0.052 -9.383 0.000

Means
STATE1 3.583 0.181 19.792 0.000
DIFF2_1 0.127 0.061 2.069 0.039
DIFF3_2 -0.010 0.060 -0.163 0.871

Intercepts
D11 0.000 0.000 999.000 999.000
D21 0.498 0.134 3.706 0.000
D12 0.000 0.000 999.000 999.000
D22 0.456 0.124 3.677 0.000
D13 0.000 0.000 999.000 999.000
D23 0.465 0.126 3.686 0.000

Variances
STATE1 1.000 0.000 999.000 999.000
DIFF2_1 1.000 0.000 999.000 999.000
DIFF3_2 1.000 0.000 999.000 999.000
IS2 1.000 0.000 999.000 999.000

Residual Variances
D11 0.171 0.038 4.532 0.000
D21 0.200 0.036 5.535 0.000
D12 0.086 0.035 2.448 0.014
D22 0.208 0.033 6.332 0.000
D13 0.081 0.037 2.191 0.028
D23 0.228 0.034 6.767 0.000
STATE2 0.000 999.000 999.000 999.000
STATE3 0.000 999.000 999.000 999.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

D11 0.829 0.038 21.994 0.000

D21 0.800 0.036 22.155 0.000
D12 0.914 0.035 26.155 0.000
D22 0.792 0.033 24.165 0.000
D13 0.919 0.037 25.007 0.000
D23 0.772 0.034 22.941 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

STATE2 1.000 999.000 999.000 999.000

STATE3 1.000 999.000 999.000 999.000
 Structural Equation Models for Measuring Variability and Change   163

The specific appeal of LC models lies in the fact that these models
allow us to analyze “true” change over time (i.e., change scores that are
corrected for measurement error) and to relate latent change variables to
other variables. For example, we may try to explain why depression scores
changed more in some individuals than in others. In addition, latent dif-
ference variables may serve as independent or mediator variables.
From a technical point of view, such extended analyses are easily car-
ried out in Mplus by adding additional variables to the model and includ-
ing additional on statements to either regress latent difference variables
on external variables or use the latent difference variables themselves as
predictors of other variables. For example, if age was another variable in
the data set (which is not the case here), we could examine whether indi-
vidual differences in latent change between time 1 and time 2 could be
explained by age differences. The additional specification would look as
follows, where age is an additional variable in the data set:

diff2_1 on age;

A substantive example of an LC analysis with multiple constructs can be

found in Reuter et al. (2010).

4.5 Latent Growth Curve Models

LGCMs (Bollen & Curran, 2006; Duncan et al., 2006; Meredith & Tisak,
1984, 1990) are increasingly used in the social sciences to analyze lon-
gitudinal data. In contrast to the LC models that were discussed in the
previous section, LGCMs explicitly focus on modeling the form of change
over time. One key question in growth curve modeling is which math-
ematical function adequately describes individual latent growth trajecto-
ries over time (e.g., linear or quadratic growth). Important research ques-
tions that can be answered by LGCMs are (among others):

• How large is the latent mean with regard to an attribute on the first
occasion of measurement (mean initial value)?
• Are there individual differences in the true initial value (“inter-
cept”) and, if yes, how large are they?
• Does the attribute change over time and, if yes, what is the form
and strength of this change? On average, is there an increase or
164   Data Analysis with Mplus

decline in the attribute? Is change, for example, linear or curvilin-

ear (quadratic)?
• How large is the mean slope (rate of the growth or decline)?
• Are there individual differences in the rate of change (e.g., do some
individuals show a steeper slope than others?) and, if yes, how large
are these differences?
• Are intercept and slope correlated? If yes, what is the direction and
strength of this relationship?
• Are there other variables (covariates) that predict individual differ-
ences in intercept and/or slope?

Obviously, the research questions that can be addressed with LGCMs

are similar (or even identical) to those addressed with LC models (see Sec-
tion 4.4). For example, also with LC models, one is typically interested in
examining (1) mean changes over time, (2) individual differences in initial
value and change scores, and (3) relationships between initial value and
change as well as external variables.
The main difference between LC models and LGCMs is that LGCMs
typically imply explicit hypotheses as to the form of change over time (lin-
ear, quadratic, etc.), which is not the case for LC models. In this sense,
LC models are less restrictive, because in these models, change over time
does not have to follow a specific functional form. We first consider the
formal background and application of first-order LGCMs, then in Section
4.5.2, second-­order LGCMs.

4.5.1 First‑Order LGCMs

First-order LGCMs are based on one single repeatedly measured indica-
tor per construct (e.g., the repeatedly measured sum score of an anxiety
scale). This indicator is used to directly measure one or more latent growth
factors. In contrast, second-­order LGCMs can (and should!) be used when
multiple indicators (e.g., multiple items, parallel tests, or item parcels) are
available for each time point. As discussed in Section 4.5.2, second-­order
LGCMs have a number of important advantages over first-order LGCMs.

4.5.1.1 Analysis of First‑Order Linear LGCMs

For simplicity, we first consider linear LGCMs. In Section 4.5.1.3, I show
an extension to a quadratic LGCM. Figure 4.26 shows a path diagram
of a first-order LGCM for one construct measured by the same observed
 Structural Equation Models for Measuring Variability and Change   165

BOX 4.12. Example of First-Order LGCM

To illustrate the analysis of longitudinal data with first-order LGCMs, we

use data on the self-­reported anxiety of children (N = 485) measured on
four (equally spaced) time points T1–T4.* There was a time interval of
approximately 6 months between the measurement occasions. With regard
to the model shown in Figure 4.26, the variables Y1–Y4 represent repeatedly
assessed anxiety sum scores.
The data are provided in the file anxiety.dat. This data set contains
the individual scores of the children on four manifest variables a1–a4. The
variables a1–a4 represent the anxiety sum scores at T1–T4, respectively. The
data set does not contain missing data.
*I would like to thank David A. Cole for providing the data used for this example.

variable Yk on four measurement occasions, where k indicates the occa-

sion of measurement.
In the linear LGCM in Figure 4.26, two latent factors have a direct
effect on the observed variables Y1–Y4: the intercept factor and the linear
slope factor. In addition, the model assumes that all measurements are
influenced by random measurement error, as represented by the variables
e1–e4.

Intercept Linear Slope

(interc) (linear)

1 1 1 1 0 1 2 3

Y1 (a1) Y2 (a2) Y3 (a3) Y4 (a4)

ε1 ε2 ε3 ε4

FIGURE 4.26. First-order LGCM for one observed variable Yk (k = time point)
measured on four occasions of measurement (k = 1, 2, 3, 4). ek = error variable. The
model assumes linear growth and equal spacing between time points. The variable
names used in the Mplus input appear in parentheses.
166   Data Analysis with Mplus

It can be seen in Figure 4.26 that all factor loadings in the model are
set to specific values (i.e., the loadings are not estimated as free param-
eters, but are fixed a priori). This is done so that the factors can be inter-
preted as intercept (or “initial status”) factor and linear slope factor.
Specifically, all Yk variables have unit loadings on the intercept fac-
tor and zero intercepts (all constant intercept parameters ak are set to 0;
this is not explicitly shown in the figure). Furthermore, Y1 loads only on
the intercept factor (but not on the slope factor). The intercept factor can
therefore be interpreted as the true score variable (t1) pertaining to Y1 in
the sense of classical test theory: Y1 = t1 + e1 = Intercept + e1. In other
words, the intercept factor represents individual differences in the true
(error-free) initial (T1) anxiety scores.
The linear slope factor can be interpreted as the true score difference
variable between the true score variable pertaining to Y2 and the true
score variable pertaining to Y1:

Linear Slope = t2 – t1

This is because Y2 has unit loadings on both the intercept and linear slope
factors (for more details on how to define factors in LGCMs based on true
score theory see Geiser, Keller, & Lockhart, in press): Y2 = t1 + (t2 – t1)
+ e2 = Intercept + Linear Slope + e2. This definition parallels the defini-
tion of latent difference factors in LC models (see Section 4.4). In contrast
to the LC factors considered in the previous section, however, the lin-
ear slope factor in LGCM represents true individual differences in linear
change over time, because the loadings of the (equally spaced) measure-
ments are set to 0, 1, 2, and 3, respectively.
With regard to our example, children with larger positive scores on
the slope factor show a stronger linear increase in their anxiety scores
than children with smaller slope factor scores. Children with negative
slope factor scores show decreasing anxiety values over time. The double-­
headed arrow between the intercept and slope factors indicates that inter-
cept and slope can be correlated. (Their correlation can be estimated as a
separate model parameter in Mplus.) For example, individuals with larger
initial anxiety scores may show different trajectories than those with
smaller initial anxiety scores.
How can we answer the research questions discussed above with
the linear LGCMs? The estimated mean of the intercept factor provides
an answer to the question of what the mean true anxiety level was at T1.
A large intercept factor mean indicates a high level of anxiety at T1. The
 Structural Equation Models for Measuring Variability and Change   167

variance of the intercept factor indicates to which extent individuals dif-

fered in their latent anxiety scores at T1. A large intercept factor variance
indicates that there is a large amount of variability in the latent anxiety
scores at T1.
The mean of the slope factor informs us about the average trajec-
tory in the sample. Positive values indicate that, on average, there is an
increase in latent anxiety scores over time, whereas negative values indi-
cate a decline. A value of zero means that, on average, there is no change.
The variance of the slope factor shows to which extent individual growth
trajectories differ in their slopes. The larger the variance of the slope fac-
tor, the more the individual growth curves differ in their steepness.
The question about the relationship between initial status and change
can be answered by estimating the correlation between intercept and
slope. A positive correlation indicates that children with larger initial
anxiety scores tend to show larger slope values than children with smaller
initial anxiety scores. A negative correlation would indicate that children
with larger initial values tend to show smaller slope factor values than
children with smaller initial values.
The question concerning the influence of external variables on change
(e.g., age, gender, personality, socioeconomic status) can be addressed
by adding external variables (covariates) to the model. It is possible, for
example, to regress the intercept and slope factor on covariates—­similar
to what can be done in LC models (see Section 4.4). In addition, the growth
factors can also serve as independent variables predicting other variables.
It could, for example, be interesting to study whether changes in anxiety
can predict changes in other constructs, such as depression.
Figure 4.27 shows the Mplus input file for the specification of the
linear first-order LCGM from Figure 4.26 for the four anxiety measure-
ments as represented by the variables a1, a2, a3, and a4. From the
model command, it can be seen that Mplus uses a special syntax for
growth models that simplifies the specification of these types of models.
With this special syntax, only one line of code is needed to specify the
entire model:

model: interc linear | a1@0 a2@1 a3@2 a4@3;

In this command, the latent variables interc and linear repre-

sent the intercept and linear slope factors, respectively. The vertical line (|)
is the sign that Mplus uses to indicate random slopes in multilevel models
(see Chapter 5). In the context of LGCMs, random slope means that the
168   Data Analysis with Mplus

FIGURE 4.27. Mplus input file for the specification of the linear first-order LGCM
for anxiety measured by one indicator across four measurement occasions shown in
Figure 4.26. In this input file, the simplified syntax for LGCMs is used.

slope of the growth curve is a random effect that can vary across individu-
als (i.e., the slope is a variable, not a constant; individuals can differ in
how much they change over time). The same is true for the intercept fac-
tor that also represents a random effect (i.e., individuals can differ in their
true initial anxiety scores). (The distinction of random vs. fixed effects is
revisited in Chapter 5, in which we discuss multilevel models.)
Behind the | sign, the repeatedly measured observed anxiety vari-
ables are listed that serve as indicators of the growth factors. The @ sign
is used to fix the loadings of the indicators on the slope factor so that the
slope factor properly represents the assumed form of the growth trajecto-
ries. In line with the assumption of linear growth over time, the loadings
are fixed to 0, 1, 2, and 3 on the slope factor (cf. Figure 4.26).
In the default specification of growth curve models with the spe-
cial syntax shown above, Mplus automatically fixes all loadings on the
intercept factor to 1 and sets all constant intercept terms (ak) to 0. In
addition, Mplus automatically estimates the means of the intercept and
slope factors. It is important to be aware of these default settings, because
growth models can also be specified using the conventional Mplus syntax
for standard CFA. If the conventional CFA syntax is used, the intercept
parameters ak have to be explicitly set to 0 by the user and the estimation
 Structural Equation Models for Measuring Variability and Change   169

BOX 4.13. Specification of LGCMs Using the Conventional Mplus

Syntax for CFA Models

LGCMs can also be specified using the standard Mplus syntax for confirma-
tory factor models. This alternative specification is slightly more complicated
than using the special Mplus syntax for LGCMs, because the loading pat-
tern for both the intercept and slope factors has to be explicitly specified,
the intercepts ak have to be explicitly set to 0, and the estimation of the
latent growth factor means has to be explicitly requested by the user. On the
other hand, this method makes the underlying specification of LGCMs more
explicit and shows that these models are special cases of confirmatory factor
models. In the following, we show the alternative equivalent specification
of the same first-order LGCM for the anxiety data using the conventional
Mplus syntax:

model: interc by a1@1 a2@1 a3@1 a4@1; ! Intercept factor

linear by a1@0 a2@1 a3@2 a4@3; ! Slope factor
[a1-a4@0]; ! set constant intercepts to zero
[interc linear]; ! estimate growth factor means

The complete input files for both types of specifications are provided on the
companion website.

of the latent growth factor means has to be explicitly requested (see Box
4.13).
The Mplus plot option is very useful in the analysis of LGCMs,
because it can be used to analyze, for example, the individual observed
and model-­implied growth trajectories. The plot option also allows the
plotting of the average observed trajectory along with the average model-­
implied trajectories. These curves can be useful, for example, to exam-
ine model fit. If, for instance, many individuals appear to deviate from a
close-to-­linear trajectory, this might be a sign that a linear growth model
may not be appropriate for the data. Note that the plot option can be used
only when individual data (and not summary data) are used as input for
the analyses. (The difference between individual and summary data is
explained in Chapter 2.)
In order to obtain plots of the growth curves, we request the Mplus
plot type 3 (type = plot3;). Using the series option, we specify
that the values of the variables a1–a4 are to be plotted as a function of
time (linear here refers to the name chosen for the linear slope factor in
the model statement):
170   Data Analysis with Mplus

plot: type = plot3;

series = a1 (linear) a2 (linear) a3 (linear) a4 (linear);

One obtains a set of graphs that can be opened in the Mplus output
by selecting the menu option Graph → View Graphs, as described in
detail in Section 4.5.1.2. Before looking at the graphs, we first examine the
numerical output provided by Mplus. The model fit statistics indicate that
the linear LGCM shows a rather mediocre fit to the anxiety data. This is
shown especially by the large chi-­square and RMSEA values: c2 = 27.288,
df = 5, p < .001, CFI = 0.98, RMSEA = 0.096, SRMR = 0.061.
For comparison, we can estimate a so-­called intercept-­only model
in which only an intercept, but no slope factor, is assumed. This model
implies that there is no true change in the latent anxiety scores over time.
(According to the intercept-­only model, any change seen in the observed
scores over time is due only to measurement error.) By comparing the
linear LGCM to the intercept-­only model, we can see whether the LGCM
at least fits better than a no-­change model. Figure 4.28 shows the Mplus
input file for the specification of the intercept-­only model.

FIGURE 4.28. Mplus input file for the specification of the first-order intercept-­
only model for the anxiety data using the simplified syntax for LGCMs.
 Structural Equation Models for Measuring Variability and Change   171

The basic model specification for LGCMs can be retained for the
specification of the intercept-­only model. Three additional commands are
included to set the mean, variance, and covariance of the linear slope fac-
tor to zero, given that these parameters are not estimated in the intercept-­
only model. According to this specification, only the intercept factor
remains in the model.
The plot option is unchanged, so that we can examine the model-­
implied latent growth curves in the intercept-­only model as well and com-
pare them to the model-­implied growth curves in the linear LGCM. The
model fit statistics for the intercept-­only model reveal that (as expected)
it shows a very poor fit to the data: c2 = 176.355, df = 8, p < .001, CFI =
0.857, RMSEA = 0.208, SRMR = 0.141. (The full output for the intercept-­
only model can be found on the companion website.) A comparison of the
AIC values indicates that the intercept-­only model (AIC = 1296.065) fits
worse than the linear LGCM (AIC = 1152.997). Hence, the assumption
of linear change is at least preferred over the assumption of no change.
(Nonetheless, the fit of the linear model is not optimal, indicating that the
assumption of linear change for every individual in the sample may be
too restrictive.)
In the following, we discuss the most important parameters for the
first-order linear LGCM, despite its modest fit. Under MODEL RESULTS
we see that the mean and variance of both the intercept and slope factor
are significantly different from zero (assuming an alpha level of .05). This
finding indicates that the mean initial latent level of anxiety is significantly
different from zero (MINTERCEPT = 0.698, z = 34.866, p < .001). There also
is significant variability in the initial latent anxiety scores (VarINTERCEPT
= 0.151, z = 11.996, p < .001).
With regard to the slope factor, we can see that, on average, there is
a significant linear decline in latent anxiety scores across time, because
the estimated slope factor mean has a negative sign (MLINEAR = –0.062,
z = –10.405, p < 0.001). This means that the mean model-­implied lin-
ear growth curve has a negative slope. In addition, there is a signifi-
cant amount of individual differences in the slope values around the
mean growth curve, as indicated by the significant slope factor variance
(VarLINEAR = 0.007, z = 4.667, p < .001). This means that the children
differ not only in their initial latent anxiety scores, but also in their tra-
jectories over time: Some children show a stronger increase (or decrease)
in their anxiety scores than others. (This finding will be made clearer
shortly when we look at individual growth trajectories in the requested
plots.)
172   Data Analysis with Mplus

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

INTERC |
A1 1.000 0.000 999.000 999.000
A2 1.000 0.000 999.000 999.000
A3 1.000 0.000 999.000 999.000
A4 1.000 0.000 999.000 999.000

LINEAR |
A1 0.000 0.000 999.000 999.000
A2 1.000 0.000 999.000 999.000
A3 2.000 0.000 999.000 999.000
A4 3.000 0.000 999.000 999.000

L NEAR WITH
INTERC -0.011 0.003 -3.485 0.000

Means
INTERC 0.698 0.020 34.866 0.000
LINEAR -0.062 0.006 -10.405 0.000

Intercepts
A1 0.000 0.000 999.000 999.000
A2 0.000 0.000 999.000 999.000
A3 0.000 0.000 999.000 999.000
A4 0.000 0.000 999.000 999.000

Variances
INTERC 0.151 0.013 11.996 0.000
LINEAR 0.007 0.001 4.667 0.000

Residual Variances
A1 0.067 0.008 8.675 0.000
A2 0.048 0.004 10.965 0.000
A3 0.048 0.004 11.180 0.000
A4 0.040 0.006 6.545 0.000

The standardized model results [STANDARDIZED MODEL

RESULTS (STDYX Standardization)] provide us with the stan-
dardized covariance (i.e., the correlation) between the intercept and slope
factors. In our example, the intercept and slope factors are significantly
negatively correlated (r = –.349, z = –5.001, p < .001). Furthermore, the
values reported under R-SQUARE—Observed Variable show us that
between 69.4 and 78.5% of observed individual differences are accounted
for by the latent growth factors. The remaining variance is due to mea-
surement error and systematic occasion-­specific variance in this model.
 Structural Equation Models for Measuring Variability and Change   173

(Second-order LGCMs, discussed in Section 4.5.2, allow us to separate

error variance from occasion-­specific variance; for a more detailed com-
parison of first- and second-­order LGCMs, see Geiser et al., in press.)

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

INTERC |
A1 0.833 0.019 43.608 0.000
A2 0.906 0.023 40.204 0.000
A3 0.911 0.030 29.912 0.000
A4 0.903 0.040 22.597 0.000

LINEAR |
A1 0.000 0.000 999.000 999.000
A2 0.193 0.021 9.160 0.000
A3 0.389 0.042 9.259 0.000
A4 0.578 0.065 8.861 0.000

LINEAR WITH
INTERC -0.349 0.070 -5.001 0.000

Means
INTERC 1.796 0.091 19.789 0.000
LINEAR -0.753 0.108 -6.965 0.000

Intercepts
A1 0.000 0.000 999.000 999.000
A2 0.000 0.000 999.000 999.000
A3 0.000 0.000 999.000 999.000
A4 0.000 0.000 999.000 999.000

Variances
INTERC 1.000 0.000 999.000 999.000
LINEAR 1.000 0.000 999.000 999.000

Residual Variances
A1 0.306 0.032 9.622 0.000
A2 0.264 0.024 11.037 0.000
A3 0.265 0.023 11.747 0.000
A4 0.215 0.031 6.812 0.000

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

A1 0.694 0.032 21.804 0.000

A2 0.736 0.024 30.789 0.000
174   Data Analysis with Mplus

A3 0.735 0.023 32.541 0.000

A4 0.785 0.031 24.941 0.000

The model could be extended by adding external variables as predic-

tors of the intercept and slope factors that might explain why the chil-
dren differ in their initial anxiety scores and in their rate of change. For
example, variables such as parental support or changes in other variables
such as personal competencies could be related to changes in anxiety. The
inclusion of independent variables predicting initial state and/or growth is
done by specifying one or more on statements, in addition to the growth
specification; for example:

interc linear on x y z;

Using this command, the intercept and slope factors are regressed on
three independent variables named x, y, and z. The growth factors could
also serve as independent variables themselves. In this case, they would
appear on the right-hand side of an on statement.

4.5.1.2 Graphical Analysis of Growth Curves

This section deals with the Mplus options used to retrieve graphics
requested through the plot option. These graphics are useful for visual-
izing individual and average growth curves. Graphs are retrieved in the
output window either by clicking on Graph → View Graphs (see Figure
4.29) or by clicking on the

symbol in the Mplus menu. Figure 4.30 shows different options avail-
able for viewing graphs. The option Sample means provides a plot of
the observed anxiety means as a function of time. The option Estimated
means results in a plot of the mean model-­implied growth curve. When
choosing Sample and estimated means, both the average sample and
the average model-­implied means are plotted in a single graph. Observed
individual values yields the observed individual growth curves for all
individuals in the sample. This option was chosen to produce the graph
shown in Figure 4.32.
After clicking on Observed individual values, the window Proper-
ties for viewing individual curves appears (see Figure 4.31). Here the
user can choose between different options for viewing individual growth
 Structural Equation Models for Measuring Variability and Change   175

FIGURE 4.29. Opening requested graphs in the Mplus output window.

FIGURE 4.30. Choosing a specific graph in Mplus. In this case, we want to look
at the observed individual growth curves.
176   Data Analysis with Mplus

FIGURE 4.31. Choosing the type, number, and order of individual growth curves
to be displayed. Here: curves based on observed data (Individual data), 10 curves per
graph (Number of curves: 10), and consecutive order as in the data file, beginning
with the first individual in the data set (Consecutive order/Starting at index: 1).

trajectories. The default is that growth curves are given in consecutive

order (Order to view individual curves → Consecutive order), as
extracted from the individual data file, and that only 10 curves are shown
at a time to avoid cluttering (Number of curves: 10). In addition, as the
default, the observed (“raw”) data are shown without a specific curve-­
fitting procedure (Type of curves for observed data → Individual data).
Specific curve-­fitting procedures can be selected by choosing the option
Type of curves for observed data → Individually-­fitted curves (e.g.,
linear or quadratic fit).
By keeping the default options as shown in Figure 4.31, we obtain
the graph in Figure 4.32 that contains the observed trajectories of the first
10 individuals in the data set anxiety.dat. By clicking on a specific data
point, the curve of the corresponding individual is highlighted in bold
and a label is shown that informs us about the number of the individual as
well as the individual’s actual observed numerical scores for all four time
points. In Figure 4.32, the curve of the second individual in the data set
is highlighted and the observed anxiety scores for this individual (1.643
at T1, 1.643 at T2, 1.214 at T3, and 1.929 at T4) are shown. This feature
 Structural Equation Models for Measuring Variability and Change   177

FIGURE 4.32. Individual growth curves of the first 10 individuals in the data
set anxiety.dat. By moving the cursor over a specific data point, the number of the
individual as well as his or her numeric scores for all time points are shown.

is useful, for example, to identify extreme cases or cases with aberrant

growth trajectories relative to the remaining individuals in the sample.
The growth curves of the remaining individuals in the sample can be
viewed by clicking on the symbol

that allows us to proceed to the next 10 curves. The symbol

allows us to return to the Properties for viewing individual curves

options to change settings. By clicking on

a different type of plot can be selected; for example, the model-­implied

individual or the average growth curves. The graphs in Figures 4.34 and
4.35 were produced by choosing the option Sample and estimated means
178   Data Analysis with Mplus

FIGURE 4.33. Choosing a specific graph in Mplus. In this case, we want to look
at the observed and estimated mean growth curves.

FIGURE 4.34. Mean observed (Sample) and mean model-­implied (Estimated)

growth curves in the first-order LGCM for the anxiety data.
 Structural Equation Models for Measuring Variability and Change   179

FIGURE 4.35. Mean observed (Sample) and mean model-­implied (Estimated)

growth curves in the first-order intercept-­only model for the anxiety data.

(see Figure 4.33). Figure 4.34 shows the mean observed and the mean
model-­implied growth curves for the linear LGCM. Figure 4.35 shows
the same curves for the intercept-­only model, which assumes no change
over time. Note that the model-­implied curve in the intercept-­only model
is exactly parallel to the x-axis, a pattern that does not correspond well
to the observed means (which clearly show a decline over time). Mplus
graphs can be saved as a picture file in either *.DIB, *.EMF, or *.JPEG
format. This can be done by choosing the menu option Graph → Export
plot to, as shown in Figure 4.36.

4.5.1.3 Analysis of First‑Order Quadratic LGCMs

The latent growth curve analysis discussed so far only contained a linear
slope factor, restricting change to be linear for all individuals. To model
other, nonlinear types of growth, the model can be extended by including
additional growth factors. If one wants to test, for example, the hypoth-
esis that growth is curvilinear, an additional factor can be included in the
model that represents quadratic growth (see Figure 4.37).
180   Data Analysis with Mplus

FIGURE 4.36. Exporting an Mplus graph (here as a *.JPEG file).

Linear Quadratic
Intercept Factor
Slope Factor Slope Factor
(interc)
(linear) (quadr)

1 0 3 0 1 4 9
1 1 1 1 2

Y1 (a1) Y2 (a2) Y3 (a3) Y4 (a4)

ε1 ε2 ε3 ε4

FIGURE 4.37. First-order LGCM for one observed variable Yk (k = time point)
measured on four occasions of measurement (k = 1, 2, 3, 4). ek = error variable.
The model assumes curvilinear (quadratic) growth and equal spacing between time
points. The variable names used in the Mplus input appear in parentheses.
 Structural Equation Models for Measuring Variability and Change   181

The quadratic growth factor is again defined by a specific pattern of

fixed loadings. For quadratic growth, the loadings are set to 0, 1, 4, and
9, respectively, as shown in Figure 4.37. If there are theoretical reasons to
assume curvilinear change, and the quadratic growth model shows a bet-
ter fit to the data than the linear growth model, then the quadratic model
may be preferred. Other forms of change can also be tested. LGCMs are
rather flexible in general and can be extended for various special cases
(e.g., for nonequally spaced time points as well as specific growth compo-
nents; e.g., see Mayer, Steyer, & Mueller, in press). Detailed discussions
of special models can be found in Bollen and Curran (2006) as well as in
Duncan et al. (2006).
Figure 4.38 shows the Mplus specification of a first-order quadratic
LGCM for the anxiety data. The only difference in the model specifica-
tion, compared to the linear LGCM, is that an additional growth factor
(quadr) is introduced before the | sign:

model: interc linear quadr | a1@0 a2@1 a3@2 a4@3;

Mplus automatically treats the third factor quadr as a quadratic

growth factor. In the estimation of the model parameters of the quadratic

FIGURE 4.38. Mplus input file for the specification of the quadratic first-order
LGCM for anxiety measured by one indicator across four measurement occasions
shown in Figure 4.37. In this input file, the simplified syntax for LGCMs is used.
182   Data Analysis with Mplus

growth model for the anxiety data, an estimation problem is indicated by

Mplus through the following warning message in the output:

WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT

POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE
VARIANCE/RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION
GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR
DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES. CHECK THE TECH4
OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE QUADR.

This warning message indicates that the covariance matrix of the

latent growth factors is not a proper covariance matrix. This is often a sign
that one or more factor variances were estimated to be zero or negative,
that a correlation above |1| was estimated, or that a linear dependency
among latent variables occurred. In our case, the problem is associated
with the quadratic growth factor quadr. The table with the parameter
estimates reveals that the variance of this variable is very small and not
statistically significant (VarQUADR = 0.001, z = 0.673, p = .501). In addi-
tion, the correlation between the linear and the quadratic growth fac-
tor was estimated to 1.205 (see the standardized model results), which
is an inadmissible value for a correlation coefficient. This is the cause of
the warning message output by Mplus. Even though the quadratic model
shows a better global model fit than the linear model (c2 = 1.148, df = 1, p
= .284, CFI = 1.00, RMSEA = 0.017, SRMR = 0.006), this result shows us
that the assumption of quadratic growth may not be reasonable for these
data.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

INTERC |
A1 1.000 0.000 999.000 999.000
A2 1.000 0.000 999.000 999.000
A3 1.000 0.000 999.000 999.000
A4 1.000 0.000 999.000 999.000

LINEAR |
A1 0.000 0.000 999.000 999.000
A2 1.000 0.000 999.000 999.000
A3 2.000 0.000 999.000 999.000
A4 3.000 0.000 999.000 999.000

QUADR |
A1 0.000 0.000 999.000 999.000
A2 1.000 0.000 999.000 999.000
A3 4.000 0.000 999.000 999.000
A4 9.000 0.000 999.000 999.000
 Structural Equation Models for Measuring Variability and Change   183

LINEAR WITH
INTERC -0.026 0.021 -1.201 0.230

QUADR WITH
INTERC 0.002 0.005 0.374 0.709
LINEAR -0.008 0.005 -1.565 0.117

Means
INTERC 0.720 0.020 35.617 0.000
LINEAR -0.114 0.018 -6.452 0.000
QUADR 0.016 0.005 3.113 0.002

Intercepts
A1 0.000 0.000 999.000 999.000
A2 0.000 0.000 999.000 999.000
A3 0.000 0.000 999.000 999.000
A4 0.000 0.000 999.000 999.000

Variances
INTERC 0.163 0.022 7.441 0.000
LINEAR 0.048 0.023 2.110 0.035
QUADR 0.001 0.001 0.673 0.501

Residual Variances
A1 0.036 0.019 1.870 0.061
A2 0.053 0.007 7.720 0.000
A3 0.035 0.006 5.498 0.000
A4 0.072 0.019 3.890 0.000

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

...

LINEAR WITH
INTERC -0.290 0.166 -1.746 0.081

QUADR WITH
INTERC 0.154 0.370 0.415 0.678
LINEAR -1.205 0.471 -2.558 0.011

Means
INTERC 1.784 0.128 13.928 0.000
LINEAR -0.518 0.146 -3.552 0.000
QUADR 0.514 0.415 1.237 0.216

4.5.2 Second‑Order LGCMs

An important assumption in the application of LGCMs is measurement
invariance of the measures over time (cf. Section 4.1.4). Measurement invari-
ance means that properties of the measurement model (i.e., factor load-
ings, intercepts, and/or residual variances) have not changed over time
184   Data Analysis with Mplus

(Meredith, 1993; Meredith & Horn, 2001; Millsap & Meredith, 2007).
As in LC models, at least strong measurement invariance is required for a
meaningful interpretation of growth parameters. Otherwise, changes over
time would be confounded with potential changes in the properties of the
measurement instrument.
In first-order LGCMs discussed in the previous section, the assump-
tion of strong measurement invariance is implicitly made, but is not test-
able in these models, because there is only a single indicator at each time
point. One of the strengths of second-­order LGCMs (curve-of-­factors models;
Geiser et al., in press; Hancock, Kuo, & Lawrence, 2001; McArdle, 1988;
Sayer & Cumsille, 2001) is that these models use multiple indicators at
each time point, so that the assumption of measurement invariance can
be tested in the same way as, for example, in LS and LC models.
Figure 4.39 shows a second-­order LGCM for anxiety measured at
four time points using two indicators (e.g., two different scales or parallel
test halves) at each time point. Note that the second-­order model is based
on more than just one observed variable per time point. Furthermore,
in contrast to first-order LGCMs, the indicators in second-­order LGCMs
do not directly measure growth factors, but are indirectly related to the
intercept and slope factors through LS factors. The LS factors represent the
first-order factors in this model, whereas the growth factors are specified
as second-­order latent variables similar to the latent trait factor in LST
models. In fact, Geiser et al. (in press) showed that second-­order LGCMs
can be formulated based on the fundamental concepts of LST theory.
The use of multiple indicators per time points makes it possible to
empirically test the assumption of measurement invariance by constrain-
ing factor loadings, intercepts, and residual variables to be equal across
time, as demonstrated for the LS model in Section 4.1.4. At least strong
factorial invariance (equal loadings and intercepts across time) should
hold for the growth parameters to be interpretable.
Another advantage of second-­order LGCMs is that the use of multiple
indicators makes it possible to separate reliable occasion-­specific vari-
ance from variance due to the growth factors and measurement error. In
second-­order LGCMs, occasion-­specific effects are explicitly represented
by the LS residual variables (indicated as z1–z4 in Figure 4.39), whereas
they are confounded with the error variables ek in first-order models. The
reliabilities of the indicators therefore tend to be underestimated in first-
order models (Geiser et al., in press). In the model shown in Figure 4.39,
an indicator-­specific residual factor (IS2) was included to reflect stable
specific variance in the second indicator not shared with the first indica-
tor (cf. Section 4.1.3).
 Structural Equation Models for Measuring Variability and Change   185

Intercept Linear Slope

(interc) (linear)

1 1 1 1 0 1 2 3

ζ1 ζ2 ζ3 ζ4
State 1 State 2 State 3 State 4
(anx1) (anx2) (anx3) (anx4)

λ1 λ2 λ1 λ2 λ1 λ2 λ1 λ2

Y11 Y21 Y12 Y22 Y13 Y23 Y14 Y24

(a11) (a21) (a12) (a22) (a13) (a23) (a14) (a24)

ε11 ε21 ε12 ε22 ε13 ε23 ε14 ε24

λIS21 λIS22 λIS23 λIS24

IS2

FIGURE 4.39. Second-order LGCM for two observed variables Yik (i = indicator, k
= time point) measured on four occasions of measurement. eik = measurement error
variable; li = time-­invariant first-order state factor loading; zk = LS residual variable.
The model contains an indicator-­specific (method) factor (IS2) for the second indica-
tor. The parameters lIS2k indicate factor loadings on the indicator-­specific factor. The
model assumes linear growth and equal spacing between time points. The change
process is modeled at the level of the LS variables via second-­order intercept and
slope factors. The variable names used in Mplus appear in parentheses.

BOX 4.14. Example of Second-Order LGCMs

We illustrate the use of second-­order LGCMs based on the anxiety data

example that we used also for analyzing first-order LGCMs in Mplus. Instead
of using just a single observed variable at each time point, we use two test
halves (“parcels”) constructed from the same scale. These test halves were
constructed by assigning half of the items of the anxiety questionnaire to
one test half and the other half to the second test half. The data of N = 484
children are in the file anxiety_parcels.dat on the companion website. This
data set contains 2 × 4 = 8 variables (a11–a24).
186   Data Analysis with Mplus

Figure 4.40 shows the Mplus input for the specification of a second-­
order linear LGCM for the anxiety test halves. The variable names corre-
spond to the names shown in parentheses in Figure 4.39. The first num-
ber indicates the test half (i = 1, 2), whereas the second index indicates the
time point (k = 1, 2, 3, 4).
In practice, it is useful to first study the degree of measurement invari-
ance of the indicators using a simple LS model (cf. discussion in Section
4.1) before imposing a restrictive growth structure on the LS variables.
Given that the steps in testing measurement invariance were discussed
in detail for the LS model, we proceed directly to the analysis of LGCMs,
assuming strong measurement invariance for the indicators.
The first part of the model specification is identical to an LS model
with equal loadings:

anx1 by a11
a21 (1);
anx2 by a12
a22 (1);
anx3 by a13
a23 (1);
anx4 by a14
a24 (1);

In the next part, the intercepts pertaining to the first indicator are fixed to
zero to identify the means of the LS factors:

[a11@0 a12@0 a13@0 a14@0];

The intercepts pertaining to the second indicator are set equal across
time so that—in conjunction with the constraints placed on the LS factor
loadings—­strong factorial invariance is maintained:

[a21 a22 a23 a24] (2);

The second-­ order factor structure that involves the growth factors
interc and linear can be specified using the same syntax as for first-
order LGCMs. The only difference is that the indicators of the growth fac-
tors now are the LS factors rather than the observed variables:

interc linear | anx1@0 anx2@1 anx3@2 anx4@3;

 Structural Equation Models for Measuring Variability and Change   187

FIGURE 4.40. Mplus input file for the specification of linear second-­order LGCM
for anxiety measured by two indicators across four measurement occasions shown in
Figure 4.39. In this input file, the simplified syntax for LGCMs is used.
188   Data Analysis with Mplus

In the next step, we request the estimation of the intercept and slope fac-
tor means:

[interc linear];

An indicator-­specific factor is included for the second test half to account

for indicator-­specific variance:

is2 by a21 a22@1 a23@1 a24@1;

Defined as a residual factor, this factor is not allowed to correlate with any
of the remaining latent variables in the model. Therefore, these correla-
tions have to be set to zero:

is2 with anx1-anx4@0 interc@0 linear@0;

Even though the chi-­square test is significant, the second-­order linear

LGCM shows an otherwise acceptable fit: c2 = 55.372, df = 24, p = .0003,
CFI = 0.993, RMSEA = 0.052, SRMR = 0.050. Interestingly, it appears to fit
better than the first-order linear LGCM that did not account for occasion-­
specific fluctuations in the anxiety ratings.
The parameter estimates reveal similar results with regard to the
growth process as the first-order LGCM. The mean latent anxiety value is
significantly different from zero: MINTERC = 0.693, z = 35.827, p < .001.
There is also significant variability in the initial latent scores (VarINTERC
= 0.141, z = 11.993, p < .001). As in the first-order LGCM, the mean of
the latent slope factor is negative and significantly different from zero
(MLINEAR = –0.059, z = _10.593, p < .001). This again shows that there is
a tendency for anxiety scores to decrease over time. The slope factor vari-
ance is also significant, VarLINEAR = 0.007, z = 5.408, p < .001, indicating
significant variability in the slopes of the individual growth curves (i.e.,
anxiety is not reduced to the same extent in all children).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

INTERC |
ANX1 1.000 0.000 999.000 999.000
ANX2 1.000 0.000 999.000 999.000
ANX3 1.000 0.000 999.000 999.000
ANX4 1.000 0.000 999.000 999.000
 Structural Equation Models for Measuring Variability and Change   189

LINEAR |
ANX1 0.000 0.000 999.000 999.000
ANX2 1.000 0.000 999.000 999.000
ANX3 2.000 0.000 999.000 999.000
ANX4 3.000 0.000 999.000 999.000

ANX1 BY
A11 1.000 0.000 999.000 999.000
A21 0.987 0.018 55.720 0.000

ANX2 BY
A12 1.000 0.000 999.000 999.000
A22 0.987 0.018 55.720 0.000

ANX3 BY
A13 1.000 0.000 999.000 999.000
A23 0.987 0.018 55.720 0.000

ANX4 BY
A14 1.000 0.000 999.000 999.000
A24 0.987 0.018 55.720 0.000

IS2 BY
A21 1.000 0.000 999.000 999.000
A22 1.000 0.000 999.000 999.000
A23 1.000 0.000 999.000 999.000
A24 1.000 0.000 999.000 999.000

IS2 WITH
ANX1 0.000 0.000 999.000 999.000
ANX2 0.000 0.000 999.000 999.000
ANX3 0.000 0.000 999.000 999.000
ANX4 0.000 0.000 999.000 999.000
INTERC 0.000 0.000 999.000 999.000
LINEAR 0.000 0.000 999.000 999.000

LINEAR WITH
INTERC -0.008 0.003 -2.875 0.004

Means
INTERC 0.693 0.019 35.827 0.000
LINEAR -0.059 0.006 -10.593 0.000

Intercepts
A11 0.000 0.000 999.000 999.000
A21 0.028 0.012 2.263 0.024
A12 0.000 0.000 999.000 999.000
A22 0.028 0.012 2.263 0.024
A13 0.000 0.000 999.000 999.000
A23 0.028 0.012 2.263 0.024
A14 0.000 0.000 999.000 999.000
A24 0.028 0.012 2.263 0.024
190   Data Analysis with Mplus

ANX1 0.000 0.000 999.000 999.000

ANX2 0.000 0.000 999.000 999.000
ANX3 0.000 0.000 999.000 999.000
ANX4 0.000 0.000 999.000 999.000

Variances
IS2 0.009 0.001 7.336 0.000
INTERC 0.141 0.012 11.993 0.000
LINEAR 0.007 0.001 5.408 0.000

Residual Variances
A11 0.021 0.004 5.561 0.000
A21 0.025 0.004 6.720 0.000
A12 0.019 0.003 6.491 0.000
A22 0.023 0.003 7.844 0.000
A13 0.019 0.003 7.209 0.000
A23 0.017 0.003 6.718 0.000
A14 0.020 0.003 7.352 0.000
A24 0.016 0.003 6.003 0.000
ANX1 0.048 0.007 7.063 0.000
ANX2 0.031 0.004 7.922 0.000
ANX3 0.030 0.004 8.199 0.000
ANX4 0.017 0.005 3.339 0.001

The standardized parameters under STANDARDIZED MODEL

RESULTS (STDYX Standardization) show us that the correlation
between the intercept and slope factors was estimated to –.259, which is
somewhat lower than the estimate in the first-order LGCM.
Overall, the second-­order growth factors account for between 74.6
and 90.3% of the variance in the latent anxiety states, as represented
by the LS factors anx1–anx4 (see values reported under R-SQUARE—
Latent Variable). This indicates a good fit of the linear growth model
to the data. The remaining 25.4 (first time point) to 9.7% (last time point)
of the LS factor variance is due to reliable occasion-­specific variability.
This shows that occasion-­specific fluctuations were more important at the
beginning than toward the end of the study.
With regard to the measurement model, we find that both indicators
have high loadings on the LS factors on all measurement occasions with
values between .911 and .947. These values suggest that the indicators
are rather homogeneous and have high reliabilities. The loadings of the
second indicator on the indicator-­specific factor are fairly small (between
.208 and .224), albeit statistically significant. Therefore, only 4.3–5.0% of
the observed score variability is due to indicator-­specific effects.
The reliabilities of the observed variables (R-SQUARE—Observed
Variable) were estimated to values between .88 and .918. These are
substantially higher than in the first-order LGCM, in which the estimated
 Structural Equation Models for Measuring Variability and Change   191

reliabilities varied between .694 and .785. Furthermore, reliabilities

seemed to increase over time in the first-order LGCM, whereas they are
almost perfectly stable according to the second-­order LGCM. These dis-
crepancies are explained by the fact that the first-order model confounds
unreliability with occasion specificity. First, this leads to an underesti-
mation of indicator reliabilities in first-order LGCMs whenever occasion-­
specific effects are present. Second, changes in the amount of occasion
specificity (as observed in the present data set in which occasion speci-
ficity decreased over time) are erroneously reflected in “changes” in the
indicator reliabilities in first-order LGCMs. These findings illustrate that
first-order LGCMs should be used only when constructs are perfectly
trait-like, whereas second-­ order models are less restrictive and yield
meaningful results also when constructs contain both state and trait
components (for a detailed discussion of this issue, see Geiser et al., in
press).

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

INTERC |
ANX1 0.864 0.019 46.518 0.000
ANX2 0.932 0.021 43.447 0.000
ANX3 0.922 0.030 31.217 0.000
ANX4 0.909 0.039 23.222 0.000

LINEAR |
ANX1 0.000 0.000 999.000 999.000
ANX2 0.205 0.020 10.387 0.000
ANX3 0.405 0.038 10.669 0.000
ANX4 0.600 0.059 10.193 0.000

ANX1 BY
A11 0.949 0.010 99.540 0.000
A21 0.918 0.010 90.264 0.000

ANX2 BY
A12 0.947 0.009 110.162 0.000
A22 0.911 0.009 97.051 0.000

ANX3 BY
A13 0.947 0.008 125.147 0.000
A23 0.927 0.009 108.099 0.000

ANX4 BY
A14 0.946 0.008 123.294 0.000
A24 0.932 0.009 106.026 0.000
192   Data Analysis with Mplus

IS2 BY
A21 0.208 0.015 13.758 0.000
A22 0.223 0.016 13.928 0.000
A23 0.224 0.016 13.857 0.000
A24 0.223 0.016 13.532 0.000

IS2 WITH
ANX1 0.000 0.000 999.000 999.000
ANX2 0.000 0.000 999.000 999.000
ANX3 0.000 0.000 999.000 999.000
ANX4 0.000 0.000 999.000 999.000
INTERC 0.000 0.000 999.000 999.000
LINEAR 0.000 0.000 999.000 999.000

LINEAR WITH
INTERC -0.259 0.071 -3.659 0.000

Means
INTERC 1.844 0.092 20.014 0.000
LINEAR -0.720 0.095 -7.582 0.000

Intercepts
A11 0.000 0.000 999.000 999.000
A21 0.059 0.027 2.230 0.026
A12 0.000 0.000 999.000 999.000
A22 0.064 0.029 2.226 0.026
A13 0.000 0.000 999.000 999.000
A23 0.064 0.029 2.221 0.026
A14 0.000 0.000 999.000 999.000
A24 0.064 0.029 2.219 0.027
ANX1 0.000 0.000 999.000 999.000
ANX2 0.000 0.000 999.000 999.000
ANX3 0.000 0.000 999.000 999.000
ANX4 0.000 0.000 999.000 999.000

Variances
IS2 1.000 0.000 999.000 999.000
INTERC 1.000 0.000 999.000 999.000
LINEAR 1.000 0.000 999.000 999.000

Residual Variances
A11 0.099 0.018 5.460 0.000
A21 0.115 0.017 6.746 0.000
A12 0.104 0.016 6.363 0.000
A22 0.120 0.015 7.793 0.000
A13 0.103 0.014 7.176 0.000
A23 0.090 0.014 6.465 0.000
A14 0.106 0.015 7.310 0.000
A24 0.082 0.014 5.696 0.000
ANX1 0.254 0.032 7.910 0.000
ANX2 0.188 0.023 8.188 0.000
 Structural Equation Models for Measuring Variability and Change   193

ANX3 0.180 0.021 8.648 0.000

ANX4 0.097 0.028 3.446 0.001

R-SQUARE

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value
A11 0.901 0.018 49.770 0.000
A21 0.885 0.017 52.066 0.000
A12 0.896 0.016 55.081 0.000
A22 0.880 0.015 57.118 0.000
A13 0.897 0.014 62.574 0.000
A23 0.910 0.014 65.336 0.000
A14 0.894 0.015 61.647 0.000
A24 0.918 0.014 63.977 0.000

Latent Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

ANX1 0.746 0.032 23.259 0.000

ANX2 0.812 0.023 35.357 0.000
ANX3 0.820 0.021 39.311 0.000
ANX4 0.903 0.028 32.101 0.000

Using the plot option, we generated a plot of the mean model-­implied

mean growth curve that can be viewed using the Graph → View graphs
option (see Figure 4.41 on page 194). Graphs of individual growth curves
are currently not available for second-­order LGCMs in Mplus.
In the same way as first-order LGCMs, second-­order LGCMs can be
extended in various ways. For example, in addition to linear growth, we
can also study quadratic or other forms of growth by adding additional
growth factors to the model. The specification of a quadratic growth fac-
tor is analogous to the first-order LGCMs and is therefore not described
in detail here.
In addition, covariates of change can also be included in second-­
order LGCMs. In this regard, second-­order models are, once again, more
flexible than first-order LGCMs. Second-order LGCMs allow researchers
not only to include covariates of change, but also of occasion-­specific vari-
ability, as represented by the LS residual variables.
194   Data Analysis with Mplus

FIGURE 4.41. Mean model-­implied growth curve for the linear second-­order
LGCM for the anxiety data generated via the Mplus plot option.
5

Multilevel Regression Analysis

5.1 Introduction to Multilevel Analysis

Multilevel models (also referred to as hierarchical linear models, random

coefficient models, or mixed models) are used to analyze clustered data, that
is, data with a hierarchical (or cluster) structure. Clustered data occur, for
example, when students from different school classes are selected for a
study in such a way that some students share the same school class (with
the associated common peer environment, teachers, etc.). In this case, the
observational units (the students) are nested (or “clustered”) within school
classes. In a two-level sampling design, one distinguishes between level-1
and level-2 units. In our example, the students represent the level-1 units
(units of the lower, so-­called micro level) and the school classes represent
the level-2 units (units of the macro level). Depending on the sampling
design, there could be additional, higher levels (e.g., schools, districts,
cities, counties, states, countries). In our examples, we consider only two
levels.
Nested data structures also occur with longitudinal data in which
measurement occasions (or time points) are nested within individuals.
Therefore, multilevel analysis can also be used to analyze longitudinal
data (e.g., see Bryk & Raudenbush, 1987; Singer & Willett, 2003).
Taking the hierarchical structure of clustered data into account is
important for at least two reasons. First, clustered sampling, by definition,
violates the assumption of independence of observations that is made by
many traditional statistical approaches such as ordinary least squares
(OLS) regression analysis and analysis of variance (ANOVA; Cohen et al.,
195
196   Data Analysis with Mplus

2003). Second, independent variables or predictors may act at different

levels of analysis. For example, teacher-­related variables such as teaching
style or gender would be variables at level-2 and should be properly mod-
eled as such.
In random samples (without cluster structure) the assumption of
independence of observations is met. In clustered samples, however, there
often are dependencies among observations from the same cluster. With
regard to our student example, we might expect students within the same
class to be more similar to each other with respect to certain attributes
(e.g., knowledge about certain topics, school achievement in general) than
students who belong to a different class. The reason is that students in the
same class are, in part, subject to the same influences. For example, they
typically share the same teacher(s) and are exposed to the same group
dynamics within the class.
Using conventional statistical approaches to analyze clustered data
without proper adjustment can lead to biased results. In particular, the
standard errors of the model parameters (e.g., regression coefficients)
may be underestimated, because dependencies in the data due to cluster-
ing lead to an overestimation of the effective sample size (Cohen et al.,
2003; Snijders & Bosker, 1999). Undesirable consequences of the under-
estimation of standard errors are biased statistical inference (specifically,
an increase in the alpha error rate) and incorrectly estimated confidence
intervals around the parameter estimates. The p-values associated with
tests of statistical significance will generally be too small, leading to a too-­
frequent rejection of the null hypothesis that a parameter is equal to zero
in the population. Moreover, confidence intervals will be too small, sug-
gesting a greater level of precision in parameter estimation than is actually
present.
The second reason why the multilevel structure of the data should be
taken into account is that modeling variables at different levels—­as well
as their interaction—is often of substantive interest. For example, in many
cases, variables at the individual level (e.g., attributes of the students) as
well as variables at the cluster level (e.g., attributes at the level of the class)
are relevant for the prediction of an outcome variable (e.g., school achieve-
ment). At the level of the students, variables such as intelligence, indus-
triousness, and motivation could be relevant for explaining school perfor-
mance. At the class level, variables such as the size of the class, the class
climate, and teacher attributes (teaching style, gender of the teacher, etc.)
could be interesting predictor variables. In addition to studying the main
effects of predictors at different levels, multilevel models allow us to model
 Multilevel Regression Analysis   197

BOX 5.1. Summary of the Most Important Advantages of Multilevel

Regression Analysis

The most important advantages of multilevel regression analysis can be sum-

marized as follows:

• Multilevel models appropriately account for the hierarchical data structure

that causes dependencies in the data.
• Multilevel modeling avoids standard error bias due to clustering that leads
to inflated type-I error rates and incorrect confidence intervals.
• Multilevel models allow analyzing variables at different levels as well as
the analysis of cross-level interactions.
• Multilevel analysis is more flexible and requires fewer assumptions than
traditional statistical methods such as, for example, repeated measures of
ANOVA.

interactions between variables at different levels (so-­called cross-level inter-

actions; e.g., see Luke, 2004, as well as Section 5.5.2).
More detailed descriptions of multilevel analysis can be found in
the textbooks by Hox (2002), Kreft and de Leeuw (1998), Luke (2004),
Raudenbush and Bryk (2002), Singer and Willett (2003), as well as
­Snijders and Bosker (1999). A superb introductory chapter on multilevel
analysis is also available in the book by Cohen et al. (2003, their Chapter
14). The use of multilevel models to analyze longitudinal data is dis-
cussed in Bryk and Raudenbush (1987) as well as in Singer and Willett
(2003).

5.2 Specification of Multilevel Models in Mplus

The specification of multilevel models in Mplus is carried out by using the

analysis: type = twolevel; option. In order to use this option,
either the “multilevel add-on” or the “combination add-on” upgrades are
required in addition to the Mplus base program (for details, see www.
statmodel.com). As of version 6, Mplus is able to model variables at only
two levels. An exception is longitudinal data, for which a multivariate
approach can be used. Multivariate means that the two-level structure
caused by repeated measurement (time points nested within individuals)
is accounted for by using (single-­level) LGCMs in which there are one or
198   Data Analysis with Mplus

BOX 5.2. Example Multilevel Regression Analysis

We use a simple cross-­sectional data set to illustrate the analysis of multilevel

regression models in Mplus. In this example, the outcome (or dependent)
variable is mathematics achievement of German students. In our example,
there are N = 503 students (level-1 units) nested within 34 school classes
(level-2 units). Each student took a mathematics test (variable math) as well
as the German Kognitiver Fähigkeitstest (cognitive ability test KFT; Heller
et al., 1976; variable kft). The variables math and kft represent level-1 vari-
ables, because they were measured at the level of the students (each student
can potentially have a different score on each test).
In addition, the school type (conventional German high school vs. Ger-
man gymnasium) was recorded and dummy-coded (variable stype; 0 = con-
ventional high school, 1 = gymnasium). School type is a level-2 variable
because it can vary for different classes, but not for students within the same
class.
Figure 5.1 shows part of the SPSS data set named math_kft.sav. Note
that an individual data set is required for the multilevel analysis in Mplus
and that this individual data set needs to contain a so-­called cluster variable
to indicate membership in a specific level-2 unit (here, the school classes). In
our example, the cluster variable is named class and enumerates the school
classes from 1 through 34.

more outcome variables at each time point (see Section 4.5). A third level
(e.g., individuals nested within organizations or school classes) can then
again be accounted for using the type = twolevel; option. There-
fore, a total of three levels can be modeled in Mplus if one of the levels is
defined by repeated observations.

5.3 Option TwoLevel BASIC

A useful option in Mplus is the so-­called twolevel basic option that

can be used at the beginning of a multilevel analysis to check whether
the data are read correctly by Mplus (see also Chapter 2). The twolevel
basic option also provides an overview of the cluster structure in the
data and the degree of dependence of observations. The degree of depen-
dence of observations is often measured by the intraclass correlation coef-
ficient (ICC). An ICC of zero indicates that observations are independent
of cluster membership. The larger the ICC, the more individual differences
 Multilevel Regression Analysis   199

BOX 5.3. Formal Notation for Multilevel Regression Models Used

in This Chapter

The following common notation is used in this chapter to indicate the param-
eters estimated in multilevel regression models. Note that actual estimates
based on sample data are indicated with a hat (^) to make clear that they are
sample estimates of population parameters.

Yij = value of a person i belonging to cluster j on the dependent variable

b0j = random intercept in the level-1 regression equation

b1j = random slope in the level-1 regression equation

Xij = value of a person i belonging to cluster j on a level-1 predictor variable

r ij = level-1 residual score

σ 2r = level-1 residual variance (also denoted as σ 2W for within in the

ij
intercept-­only model)

g00 = grand mean/intercept for the random intercept coefficient on level 2

g01 = slope coefficient for the random intercept on level 2

g10 = intercept for the random slope on level 2

g11 = slope coefficient for the random slope on level 2

Wj = value of a level-2 unit (e.g., a school class) on a level-2 predictor vari-

able

u0j = residual score for the random intercept on level 2

u1j = residual score for the random slope on level 2

σ 2u = level-2 residual variance for the random intercept (also denoted as
0j
σ 2B for between in the intercept-­only model)

σ 2u =   level-2 residual variance for the random slope

1j
t01 = level-2 (residual) covariance between random intercept and random
slope

rIC = intraclass correlation coefficient

200   Data Analysis with Mplus

are due to differences between clusters. Mplus automatically outputs the

ICC in a multilevel analysis. In other multilevel regression analysis pro-
grams, the ICC sometimes has to be calculated manually based on vari-
ance parameter estimates of the so-­called null or intercept-­only model (see
Section 5.4.1).
The ICC rIC is defined as the ratio of the variance between the clus-
ters (level-2 variance σ 2B , “B” for “between clusters”) to the total of vari-
ance (σ 2B + σ 2W ), where σ 2W indicates the (level-1) variance within the
clusters (“W” for “within clusters”):

σ 2B
rIC =
( σ 2W + σ 2B )
Even very small ICC values of, say, .05 or .10 can lead to considerable
bias in the results of tests of statistical significance in conventional OLS
regression analyses and ANOVA, which do not take the dependence of
observations into account (e.g., Cohen et al., 2003, p. 538). In a two-
level basic analysis, Mplus outputs the ICC as well as the estimated
level-1 (within clusters) and level-2 (between clusters) variances for all
variables included in the analysis (i.e., for all variables listed under use-
variables).
The individual data shown in Figure 5.1 were exported to an ASCII
file named math_kft.dat for use in Mplus following the procedure
described in Section 2.1. The Mplus input file for the specification of the
twolevel basic analysis for the variables math and kft is shown in
Figure 5.2. The cluster variable class has to be specifically defined using
the subcommand cluster = class; to inform Mplus that the level-2
units (here the school classes) are defined based on this variable. The
cluster variable does not have to appear under the usevariables list,
because this variable is a variable with a special function and is neither
used as an independent nor a dependent variable in the model.
Below, the most important part of the Mplus output for the two-
level basic analysis is shown. Under SUMMARY OF ANALYSIS,
we can see that N = 503 students (level-1 units) and two variables (math
and kft) were used in the analysis. Mplus also informs us that the vari-
able class served as the cluster variable in this analysis, as intended. This
allows us to once again check that our level-2 units are correctly defined
(the level-2 variable stype is included in the analysis in Sections 5.4.3
and 5.5.2).
FIGURE 5.1. Screen shot of the sample data set math_kft.sav. The variable class
indicates the different school classes and is used as cluster variable in Mplus. The
variable math is the continuous dependent variable, kft is a continuous level-1 pre-
dictor variable, and stype is a dummy-coded dichotomous level-2 predictor variable.

FIGURE 5.2. Mplus input file for a twolevel basic analysis for the vari-
ables math and kft.
202   Data Analysis with Mplus

SUMMARY OF ANALYSIS

Number of groups 1
Number of observations 503

Number of dependent variables 2

Number of independent variables 0
Number of continuous latent variables 0

Observed dependent variables

Continuous
MATH KFT

Variables with special functions

Cluster variable CLASS

Under the rubric SUMMARY OF DATA we obtain more detailed informa-

tion about the cluster structure present in the data. Here, we are dealing
with 34 clusters (school classes = level-2 units). Mplus informs us about
the distribution of the cluster sizes in a stem-and-leaf type of plot. We
can see that there is one cluster that contributed five students (the school
class with the number 22), one cluster that contributed eight students
(class number 18), and so on. The mode is 13 students per cluster (school
classes number 14, 2, 5, 23, and 13). The class that provides the most stu-
dents is class number 7 (24 students stem from this class). Under the plot,
Mplus reports the average cluster size (here, 14.794 students per class).

SUMMARY OF DATA

Number of clusters 34

Size (s) Cluster ID with Size s

5 22
8 18
9 19
10 16 21 12 25
11 20 3
12 8
13 14 2 5 23 13
14 4
15 17 24 1
16 6 28 30
17 11 27
18 29 26 35 36
19 33
 Multilevel Regression Analysis   203

20 31
21 34
22 10
23 32
24 7

Average cluster size 14.794

Mplus next outputs the estimated ICCs for both the kft and math
variables. The values of rIC = .44 for kft and rIC = .378 for math indicate
that class membership accounts for a substantial portion of individual
differences in both the math and KFT test scores. In other words, a non-
trivial amount of the variance in math and KFT scores is due to average
performance differences between the school classes. This also shows us
that in this example it makes sense to consider predictor variables at the
class level, because there is a considerable amount of variability at this
level. The inclusion of level-2 predictors would not make sense if there
were no differences between the classes.

Estimated Intraclass Correlations for the Y Variables

Intraclass Intraclass
Variable Correlation Variable Correlation

MATH 0.440 KFT 0.378

Below is the Mplus output of descriptive statistics (means, covari-

ances, and correlations) separately for level 1 and level 2. Note that the
(grand) means of the variables are estimated at level 2 (the between level).
The estimated grand means are gˆ 00 = 11.930 for math and gˆ 00 = 66.716
for kft. Given that the level-1 means (or cluster means) are a function of
level-2 parameters (the grand means plus the cluster-­specific deviations
from the grand means), the means at level 1 (within level) are output as
0.000 in Mplus. The estimated variances of the variables at the individual
level are σˆ 2W = 26.484 for math and σˆ 2W = 235.735 for kft. At the class
level, the estimated variances are σˆ 2B = 20.806 for math and σˆ 2B = 142.985
for kft.

RESULTS FOR BASIC ANALYSIS

NOTE: The sample statistics for within and
between refer to the maximum-likelihood estimated
within and between covariance matrices,
respectively.
204   Data Analysis with Mplus

ESTIMATED SAMPLE STATISTICS FOR WITHIN

Means
MATH KFT
________ ________
1 0.000 0.000

Covariances
MATH KFT
________ ________
MATH 26.484
KFT 52.327 235.735

Correlations
MATH KFT
________ ________
MATH 1.000
KFT 0.662 1.000

ESTIMATED SAMPLE STATISTICS FOR BETWEEN

Means
MATH KFT
________ ________
1 11.930 66.716

Covariances
MATH KFT
________ ________
MATH 20.806
KFT 49.983 142.985

Correlations
MATH KFT
________ ________
MATH 1.000
KFT 0.916 1.000

The variance estimates can be used to calculate the estimated ICCs (this
is not really necessary, given that Mplus automatically outputs the rIC
values):

σˆ 2B 20.806
For math: rˆ IC = = = 0.44
σˆ W + σˆ B 26.484 + 20.806
2 2

σˆ 2B 142.985
For kft: rˆ IC = = = 0.378
σˆ W + σˆ B 235.735 + 142.985
2 2
 Multilevel Regression Analysis   205

In addition, we see that the correlation between math and kft at the indi-
vidual level is r = .662, whereas the relationship at the level of the school
classes is much higher (r = .916). In summary, we find that the use of
a multilevel analysis to analyze the present data must be highly recom-
mended, given the large amount of variability between school classes. If
we used a conventional OLS regression analysis without taking the cluster
structure into account, we would have to expect a significant underesti-
mation of the standard errors of the regression coefficients and therefore
a substantial increase of our type-I error rate as well as biased confidence
intervals. We now consider different types of multilevel regression models
following the taxonomy of Luke (2004; see his Table 2.1). We begin with
so-­called random coefficient regression models, the most basic of which is the
null or intercept-­only model.

5.4 Random Intercept Models

5.4.1 Null Model (Intercept‑Only Model)
The null or intercept-­only model is often estimated in the first step of a
multilevel regression analysis because it can be used to estimate the ICC
for the dependent variable. In Mplus, this step is not strictly necessary,
because the level-1 and level-2 variances as well as the ICC are directly
output whenever the analysis option type = twolevel (or two-
level basic) is used. We still discuss the intercept-­only model here
for didactic reasons because it is one of the simplest multilevel models,
and having an understanding of this basic model and its specification in
Mplus facilitates understanding of the analysis of more complex multi-
level models in Mplus.
In the intercept-­only model, there is no predictor variable of the
dependent variable on either level 1 or level 2. At level 1, the individual
scores Yij are decomposed into the cluster mean b0j plus the individual-­
specific deviation (or residual score) from the cluster mean r ij:

Level 1: Yij = b0j + rij

In terms of our data example, we decompose the math score of an indi-

vidual student i (e.g., David) who belongs to school class j into the mean
of David’s school class (cluster mean) plus the deviation of David’s math
score from the class average. The variance σ 2rij characterizes the amount
of within-­class (level-1) variability:
206   Data Analysis with Mplus

σ 2rij = σ 2W

Large values of σ 2rij indicate that there are substantial individual dif-
ferences in the math scores among students who belong to the same class.
From the twolevel basic analysis (see Section 5.3) we already know
that σˆ 2rij = σˆ 2W = 26.484 for the math test score. The same estimate will
also be obtained based on the intercept-­only model. At level 2, variability
in the cluster means b0j is modeled. The cluster means are decomposed
into the grand mean across all school classes (g00) plus the deviations of
the cluster means from the grand mean (u0j):

Level 2: b0j = g00 + u0j

The level-2 equation shows that in the intercept-­only model, the cluster
means are variables that have a mean g00 and a variance σ 2u0 j, both of
which are parameters at level 2. This also explains the name random inter-
cept model: Whereas in conventional OLS regression analysis the intercept
is a constant (fixed effect), the intercept b0j in the random intercept model
can vary across clusters and therefore represents a random effect. In our
example, the math test means are not constant, but can vary across school
classes.
The variance σ 2u0 j characterizes differences in the cluster means
across level-2 units (between clusters):

σ 2u0 j = σ 2B

that is, the degree of differences between the school classes. Large val-
ues of σ 2u0 j indicate that there are large differences in the mean math
scores between school classes. From the twolevel basic analysis, we
know that σˆ 2u0 j = σˆ 2B = 20.806 for the math test score. A similar estimate
is also obtained in the intercept-­only model. Because in this model the
variability in the math test means across classes is not explained by exter-
nal variables (the level-2 equation does not contain any predictor vari-
ables), the intercept-­only model is sometimes referred to as unconditional
cell means model, unconstrained model, empty model, or null model.
In summary, we obtain Mplus estimates for the following three types
of parameters in the intercept-­only model:
 Multilevel Regression Analysis   207

• At level 1 (WITHIN):
—The
— variance of the individual scores around the cluster means,
σ 2rij = σ 2W
• At level 2 (BETWEEN):
—The
— grand mean g00
—The
— variance of the cluster means around the grand mean,
σ u0 j = σ 2B
2

Figure 5.3 shows the Mplus input file for the estimation of the intercept-­
only model using the outcome variable math. It is only necessary to list
the variable math under usevariables, specify the relevant cluster
variable (here, class), and request analysis: type = twolevel.
A specific model statement is not required for the “null model.” After run-
ning the input file, the following Mplus warning message occurs in the
output:

*** WARNING in MODEL command

All variables are uncorrelated with all other
variables in the model. Check that this is
what is intended.
1 WARNING(S) FOUND IN THE INPUT INSTRUCTIONS

FIGURE 5.3. Mplus input file for the specification of the intercept-­only model for
the variable math.
208   Data Analysis with Mplus

This message occurs due to the fact that we did not specify any relations
between variables. This is in line with our null model, however, which
does not relate either the individual scores or the cluster means to any
external variables. Therefore, the warning message can be ignored in this
particular case. (Note that in other models, the same or similar messages
may indicate a potentially serious model misspecification and should
therefore not be generally ignored.)
The Mplus output provides us with the estimate of the level-1 vari-
ance under MODEL RESULTS—Within Level (σˆ 2rij = 26.480). Under
Between Level we obtain estimates of the grand mean (gˆ 00 = 11.933)
and the level-2 variance σˆ 2u0 j = 20.853. The estimated values are only
slightly different from the values that were estimated in the twolevel
basic analysis (see Section 5.3). In addition, the intercept-­only model
also provides us with standard errors of the model parameter estimates as
well as test statistics (z-scores) and corresponding p-values.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

Variances
MATH 26.480 2.450 10.809 0.000

Between Level

Means
MATH 11.933 0.808 14.776 0.000

Variances
MATH 20.853 6.017 3.466 0.001

BOX 5.4. Estimation of Standard Errors under type = twolevel

in Mplus
Under type = twolevel the standard errors of the model parameters are
estimated by default using robust maximum likelihood estimation (esti-
mator = mlr;) in Mplus. Conventional maximum likelihood estimation
can be requested by using the following specification:

analysis: estimator = ml;

 Multilevel Regression Analysis   209

The intercept-­only model provides only the level-1 and level-2 variances
that can be used for the calculation of the ICC. The model does not con-
tain any additional independent variables (“empty model”). Therefore,
variability in the scores at the individual level or the group means is not
accounted for by other variables in this model. In the following, we dis-
cuss more complex multilevel models in which additional independent
variables are present (1) only at level 1 (so-­called random effects analysis
of covariance [ANCOVA] model), (2) only at level 2 (so-­called means-as-­
outcomes models), or (3) at both level 1 and level 2 (so-­called intercepts-­and-­
slopes-­as-­outcomes models).

5.4.2 One‑Way Random Effects ANCOVA

We now discuss the case that a researcher (1) wants to predict an out-
come variable (e.g., math score) using a continuous level-1 predictor vari-
able only (e.g., KFT score) and (2) wants to control for differences in the
KFT means between school classes without assuming differences in the
slopes of the level-1 regressions across clusters. The resulting model con-
sists of a level-1 regression equation with a random intercept (b0j) and
a constant slope (b1j = g10). This model is equivalent to a one-­factorial
analysis of covariance with a random factor (so-­called one-way random
effects ANCOVA; Luke, 2004), where the random factor, in our case, is the
school class:

Level 1: Yij = b0j + b1j · Xij + rij

Level 2: b0j = g00 + u0j

b1j = g10

The level-2 equations show that in this model, the intercept b0j from the
level-1 regression of math score on KFT score can vary across school
classes (level-2 units). In contrast, the level-1 slope coefficient b1j is
assumed to be constant across school classes. This can be seen from the
fact that there is no residual score in the level-2 equation for b1j, whereas
for the intercept b0j, a residual term u0j is included. Note that this model
is too restrictive if the slopes of the level-1 regressions actually do vary
across school classes, because no such variation is allowed in the model.
In Section 5.5 we consider models that allow for variation in both the
intercept and slope coefficients across clusters.
210   Data Analysis with Mplus

In Mplus, we obtain the parameters of the one-way random effects

ANCOVA model in the following way:

• At level 1 (WITHIN):
—The
— constant slope (fixed effect) b1j = g10
—The
— residual variance σ 2rij
• At level 2 (BETWEEN):
—The
— mean intercept g00 across clusters
—The
— variance σ 2u0 j of the intercepts across clusters

Before we consider the results for this model in detail, I introduce a con-
cept that is of great importance in multilevel analysis: that of centering
predictor variables. Centering was briefly discussed in the Section 3.2 on
OLS regression analysis (cf. Box 3.4). For simplicity, we return to simple
OLS regression analysis to revisit the issue of centering:

Y = b0 + b1 · X + e

where b0 indicates a fixed intercept and b1 indicates a fixed slope coef-

ficient in the regression of a criterion variable Y on a single predictor vari-
able X, and e denotes a residual variable. The estimated OLS regression
equation for our example is

math = –6.486 + 0.279 · kft + residual

Note that the estimated OLS regression coefficients do not take cluster-
ing into account and therefore should not be used. The negative intercept
term in this equation (bˆ 0 = –6.486) means that for students who obtained
a KFT score of 0, we would expect a math test score of –6.486. There are
two problems associated with this estimate for our interpretation of the
results. First, negative scores lie outside of the possible range of values for
the outcome variable math (the smallest possible score on the math test is
0 points). Second, a KFT raw score of 0 is very unlikely to be observed for
any student (and, in fact, was not observed in the present study). There-
fore, a substantively meaningful interpretation of the intercept is not pos-
sible in this example—­at least not with uncentered variables. By centering
the independent variable kft before the analysis, the zero point of this
variable becomes more meaningful. Centering means that a meaningful
constant (usually a type of mean) is subtracted from each raw score. In
the analysis, the centered (or deviation) scores are then used instead of
the raw scores.
 Multilevel Regression Analysis   211

A frequently used form of centering is grand mean centering, in which

the grand mean X.. is subtracted from all raw scores Xij:

XijC = Xij − X..

where XijC indicates the centered variable. If the student, for example, has
a score of 0 on the grand-mean-­centered KFT variable, this means that the
KFT score of this student is exactly identical to the average in the entire
sample. Negative values on the centered variable indicate that a student
scored below average on the KFT, whereas positive values indicate scores
above average in the sample.
In the OLS regression model, centering leads to a different intercept
estimate that typically has a more meaningful interpretation when inde-
pendent variables have no meaningful zero point or when the value of 0 is
meaningful, but did not occur in the data (cf. Section 3.2). The intercept
in the regression equation with a centered predictor variable provides the
expected value of the dependent variable for an individual with an average
value on the independent variable. The slope is unaffected by centering in
the simple regression model. The OLS regression equation for the math–
KFT example changes as follows when centered, instead of uncentered,
KFT scores are used:

math = 12.533 + 0.279 · kft + residual

We see that only the intercept, but not the slope of the regression equa-
tion, has changed. The new intercept is 12.533, which means that for
students with an average KFT score, we expect a math test score of 12.533
based on the linear regression of math on kft. Substantively, this repre-
sents a more meaningful result than the negative intercept estimate in the
uncentered regression equation.
In Mplus, centering of variables is performed by adding the single
additional subcommand centering in the variable command. This
means that the independent variables do not have to be manually cen-
tered prior to the analysis in Mplus. Mplus offers two options for centering
variables: (1) centering at the grand mean (grand mean centering) and (2)
centering at the cluster mean (group mean centering). Grand mean center-
ing can be requested in Mplus with the following specification:

variable: centering = grandmean (kft);

Group mean centering is requested as follows:

212   Data Analysis with Mplus

variable: centering = groupmean (kft);

Detailed recommendations on the use of different forms of centering

in multilevel models can be found in Enders and Tofighi (2007), Hofmann
and Gavin (1998), as well as Kreft, de Leeuw, and Aiken (1995). Figure 5.4
shows the Mplus input for the specification of the one-way random effects
ANCOVA model for our example. Of importance, the predictor variable
kft is defined as a level-1 variable, because in our example, this variable
is only used at level 1:

within = kft;

In addition, we use grand mean centering for the variable kft, as demon-
strated above. In the model command, we specify the regression equa-
tion at level 1. Since level 1 is referred to as the within level in Mplus,
we have to include a command that makes clear that this equation refers
to this level:

FIGURE 5.4. Mplus input file for the specification of the one-way random effects
ANCOVA model for the outcome variable math and the level-1 predictor variable kft.
 Multilevel Regression Analysis   213

model: %within%
math on kft;

Because in our model, the level-2 equations only contain a random inter-
cept and a fixed slope with no additional structure, nothing has to be
specified for the between level. Mplus automatically estimates the
parameters g00 and σ 2u0 j at level 2. (The fixed slope b1j = g10 is treated
as a level-1 parameter in Mplus.) The subcommand stdyx in the out-
put command is used to request the fully standardized solution. Under
MODEL RESULTS—Within Level Mplus provides the estimated
slope coefficient bˆ 1 j = gˆ 10 = 0.234 as well as the estimated level-1 residual
­variance σˆ 2r = 14.883. Under Between Level, the mean intercept
ij
(gˆ 00 = 12.310) and the level-2 residual variance for the random intercept
(σˆ 2u0 j = 5.403) are given.

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

MATH ON
KFT 0.234 0.012 20.042 0.000

Residual Variances
MATH 14.883 1.052 14.141 0.000

Between Level

Means
MATH 12.310 0.437 28.142 0.000

Variances
MATH 5.403 1.612 3.352 0.001

It can be seen that the level-1 residual variance (14.883) is substan-

tially smaller than the corresponding (unconditional) level-1 variance
estimated in the intercept-­only model (26.48) that did not contain any
level-1 predictor variable (see Section 5.4.1). This shows that a substantial
amount of the variability in the math test scores can be explained by dif-
ferences in individual KFT scores. (Since we used grand mean centering,
these differences reflect both within and between class differences; see
214   Data Analysis with Mplus

Enders & Tofighi, 2007, for details.) More detailed information on the
size of this effect is provided under the standardized solution (STDYX
Standardization). We can see that the standardized level-1 regres-
sion coefficient for the regression of math on KFT was estimated to 0.773.
The estimated R 2 value for the level-1 regression is .598, which shows that
almost 60% of the variability in individual math test scores is accounted
for by individual and group differences in the KFT scores.

STANDARDIZED MODEL RESULTS (STDYX Standardization)

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

MATH ON
KFT 0.773 0.020 38.617 0.000

Residual Variances
MATH 0.402 0.031 12.987 0.000

Between Level

Means
MATH 5.296 0.763 6.938 0.000

Variances
MATH 1.000 0.000 999.000 999.000

R-SQUARE

Within Level

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

MATH 0.598 0.031 19.309 0.000

5.4.3 Means‑as‑Outcomes Model

In the previous section, we discussed a random intercept model with one
level-1 predictor and no level-2 predictors. We now consider a model in
which there is no level-1 predictor but one level-2 predictor. Such a model
can be useful when one is interested in only explaining differences in
the cluster means on the dependent variable (Luke, 2004). For example,
one might be interested in explaining differences in the math test cluster
 Multilevel Regression Analysis   215

means through macro variables such as, for example, characteristics of the
teacher or the type of school students attend. The equations for a model
with no level-1 and one level-2 predictor, Wj, can be written as follows:

Level 1: Yij = b0j + rij

Level 2: b0j = g00 + g01 · Wj + u0j

The level-1 equation contains only the simple decomposition of an

observed score into the cluster mean plus an individual deviation from
the cluster mean. This part of the model is identical to the intercept-­
only model (see Section 5.4.1). In contrast to the intercept-­only model,
the level-2 equation in the means-as-­outcomes model is structured such
that at level 2, differences in the cluster means are explained through the
level-2 predictor Wj. The parameters g00 and g01 are fixed level-2 regres-
sion coefficients for the regression of the cluster means on the level-2 pre-
dictor. The term u0j reflects that part of the cluster mean b0j that cannot
be predicted by Wj. Because this model focuses on the explanation of dif-
ferences between clusters in terms of level-2 variables, it has been referred
to as means-as-­outcomes model by Luke (2004).
With regard to our data example, the means-as-­outcomes model
allows us to examine the question of whether differences in the math
scores at the level of the school classes can be explained by school type
(standard high school vs. gymnasium). The variable stype is a level-2
variable because it varies between, but not within, school classes. We can
use stype in our model as a level-2 predictor (Wj). Mplus provides us with
the following parameter estimates for the means-as-­outcomes model:

• At level 1 (WITHIN):
—The
— residual variance σ 2rij
• At level 2 (BETWEEN):
—The
— fixed intercept g00 for the regression of b0j on Wj
—The
— fixed slope g01 for the regression of b0j on Wj
—The
— residual variance σ 2u0 j

In our example, the coefficients g00 and g01 have a special meaning,
because our level-2 predictor school type is a dichotomous variable with
the categories 0 = standard high school and 1 = gymnasium. For this rea-
son, g00 gives the expected mean math score for the standard high school
group. The slope coefficient g01 represents the expected mean difference
216   Data Analysis with Mplus

in math scores (gymnasium minus high school). Note that the interpreta-
tion of the coefficients would change if we used a different coding scheme
for the categories of the variable school type (e.g., effect or contrast cod-
ing; for details see Chapter 8 in Cohen et al., 2003).
Figure 5.5 shows the Mplus input for the means-as-­outcomes model
for math test scores. It is important to note that the variable stype has to
be defined as a level-2 only variable. This is accomplished through the fol-
lowing specification in the variable command:

between = stype;

Note that between level variables can vary only between, but not
within clusters. If variables defined as between level variables show
variation within any of the clusters, Mplus issues a warning message in
the output. Further note that it is not necessary to center the variable
stype, because the value of 0 is generally meaningful for a dummy-coded
variable (in our case, it refers to the group conventional high school). In the
model statement, we specify that the math test score (variable math) is
predicted by the variable school type at level 2:

FIGURE 5.5. Mplus input file for the specification of the means-as-­outcomes
model for the outcome variable math and the level-2 predictor variable stype.
 Multilevel Regression Analysis   217

model: %between%
math on stype;

Because there is only a random intercept and otherwise no structure

at level 1, no statement for the within level is required in this model.
Mplus automatically estimates the residual variance for the within level.
We again request that the fully standardized solution be output by includ-
ing the stdyx keyword in the output command. In the output we
obtained the following parameter estimates:

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

Variances
MATH 26.467 2.450 10.802 0.000

Between Level

MATH ON
STYPE 10.641 1.188 8.957 0.000

Intercepts
MATH 10.003 0.441 22.671 0.000

Residual Variances
MATH 3.980 0.887 4.488 0.000

For level 1, Mplus estimates only the level-1 residual variance (σˆ 2rij =
26.467). For level 2, the slope coefficient gˆ 10 is estimated at 10.641. This
means that the expected mean difference in math test scores between the
group gymnasium and the group standard high school is 10.641. This
value is statistically significant (z = 8.957, p < .001). We can conclude that
the mean math score in group gymnasium is significantly larger than
in group standard high school. Hence, the variable stype accounts for a
significant portion of the differences seen in the mean math scores across
school classes.
The intercept term of the level-2 regression is estimated to be gˆ 00 =
10.223. This means that the expected average math test score in the group
standard high school is 10.003 points. The estimated mean for group
gymnasium can be calculated as the sum of the level-2 intercept plus the
level-2 slope coefficient: gˆ 00 + gˆ 01 = 10.003 + 10.641 = 20.644.
218   Data Analysis with Mplus

The level-2 residual variance in this model is estimated at σˆ 2u0 j =

3.98. This shows that the level-2 residual variance has decreased con-
siderably through adding the predictor school type in comparison to the
intercept-­only model, in which the estimated (unconditional) level-2 vari-
ance was 20.853 (see Section 5.4.1). This again indicates that the level-2
predictor school type accounts for a considerable amount of difference
in the math test scores between school classes. The graph in Figure 5.6
illustrates this finding. The circles and boxes indicate the (unweighted)
means in each of the 34 school classes (the unweighted cluster means).
The error bars indicate one standard deviation above and below each clus-
ter mean, respectively. Circles indicate the cluster means in the standard
high school group, whereas boxes indicate the cluster means for the group
gymnasium. The solid horizontal line represents the estimated intercept
coefficient gˆ 00 , that is, the estimated mean for the group high school
based on the Mplus multilevel analysis. The dotted horizontal line shows
the estimated mean for the group gymnasium (the sum of the estimated
coefficients gˆ 00 + gˆ 01 = 20.644). The graph illustrates the large difference

32

30 Conventional high school

Gymnasium
28
Error bars show ± 1 SD
26

24

22
γˆ 00 + γˆ 01 = 20.644
20
Math Score

18

16 γˆ 01 = 10.641
14

12

10 γˆ 00 = 10.003
8

0
1 3 5 7 10 12 14 17 19 21 23 25 27 29 31 33 35
2 4 6 8 11 13 16 18 20 22 24 26 28 30 32 34 36

School Class Number

FIGURE 5.6. Error bar plot for illustrating the relationship between cluster-­level
math performance and school type. The figure also illustrates the meaning of the
level-2 regression coefficients estimated in the means-as-­outcomes model.
 Multilevel Regression Analysis   219

in the estimated means of 10.641 points in the math test performance

between high school and gymnasium. This mean difference leads to the
significant slope coefficient gˆ 01 in the means-as-­outcomes model and
results in a large reduction of the level-2 variance after inclusion of the
predictor school type.
The estimated proportion of explained variability can be found in
the Mplus output under the standardized solution (stdyx). We can
see that the standardized level-2 regression coefficient of 0.897 is very
large. The R 2 value is estimated to be 0.805 by Mplus. This indicates that
approximately 80.5% of the differences in the math test means between
school classes are accounted for by differences between school types. The
remaining variability in math scores across school classes remains unex-
plained in the model.

STANDARDIZED MODEL RESULTS

STDYX Standardization

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

Variances
MATH 1.000 0.000 999.000 999.000

Between Level

MATH ON
STYPE 0.897 0.033 27.184 0.000

Intercepts
MATH 2.213 0.246 9.013 0.000

Residual Variances
MATH 0.195 0.059 3.287 0.001

R-SQUARE

Within Level

Between Level

Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value

MATH 0.805 0.059 13.592 0.000

220   Data Analysis with Mplus

5.5 Random Intercept and Slope Models

So far, we have discussed multilevel regression models in which only the

intercept (but not the slope) of a level-1 regression was allowed to vary
across level-2 units. Often, however, it is also of interest to analyze pos-
sible variation in the level-1 slope coefficients across clusters—­that is, to
analyze models with random slopes.
Figure 5.7 demonstrates at a descriptive level that simple OLS regres-
sion analyses of math test scores on KFT by school class show regression
lines that differ with regard to both the intercepts (the regression lines inter-
sect with the y-axis at different levels) and slopes (the regression lines have
different rates of change), even though at this point, it is not clear whether
these differences are statistically significant. It seems interesting to exam-
ine variability in the slopes across level-2 units as well. Models in which
both the intercepts and slopes of level-1 regressions can vary across level-2
units are often referred to as random coefficient regression models (Cohen et
al., 2003; Luke, 2004). Models in which differences in the intercepts and

32
30
28
26
24
22
20
Math Score

18
16
14
12
10
8
6
4
2
0
0 20 40 60 80 100 120
KFT Score

FIGURE 5.7. OLS regression of math score on KFT score by school class. Each line
refers to the regression in one of the 34 school classes. The thick dotted line repre-
sents the (uncentered) average OLS regression line across all school classes.
 Multilevel Regression Analysis   221

slopes across clusters are explained by level-2 predictor variables have been
referred to as intercepts-­and-­slopes-­as-­outcomes models by Luke (2004).

5.5.1 Random Coefficient Regression Analysis

In order to allow for differences in both the intercept and slope of a level-1
regression across clusters, the one-way random effects ANCOVA model
(see Section 5.4.2) is extended as follows:

Level 1: Yij = b0j + b1j · Xij + rij

Level 2: b0j = g00 + u0j

b1j = g10 + u1j

The level-2 equations show that variability is now allowed for both the
intercept and the slope, since residual terms are now included in both
level-2 equations. Such a random coefficient regression model, without
any level-2 predictors, is useful, for example, when a researcher is exclu-
sively interested in level-1 predictor variables, but in addition wants to
account for the effects of clustered data. If, in such a case, the researcher
used a conventional OLS regression analysis without taking the clustering
of observations into account, he or she would have to expect bias in the
estimated standard errors of the regression coefficients and, consequently,
inaccurate tests of significance and confidence intervals. In the random
coefficient regression model, in contrast, the dependencies among obser-
vations are taken into account and the standard errors, test statistics,
p-values, and confidence intervals are more accurate.
Note that the only difference between the one-way random effects
ANCOVA model and the random coefficient regression model is that now
there is also a residual term in the level-2 equation for the level-1 slope
coefficient. For our example, this means that the regression lines for the
regressions of math scores on KFT scores can now have a different inter-
cept and a different slope in each school class, as illustrated in Figure 5.7.
Mplus provides the following parameter estimates for the random coef-
ficient regression model:

• At Level 1 (WITHIN):
—The
— residual variance σ 2rij
• At level 2 (BETWEEN):
—The
— mean intercept g00 across clusters
—The
— variance σ 2u0 j in the intercepts across clusters
222   Data Analysis with Mplus

—The
— mean slope coefficient g10 across clusters
—The
— variance σ 2u1 j in the slopes across clusters
—The
— covariance t01 between intercept and slope (needs to be
specifically requested in Mplus)

Figure 5.8 shows the specification of the random coefficient regres-

sion model for our example. The predictor variable KFT is again defined
as a level-1 variable using the within subcommand and is entered into
the model in grand-mean-­centered form to allow for a more straightfor-
ward interpretation of the results (see Section 5.4.2):

within = kft;
centering = grandmean (kft);

At level 1, math test performance is predicted by KFT test perfor-

mance. Intercept and slope of the level-1 regression are considered ran-
dom coefficients that can vary across level-2 units. In order to analyze

FIGURE 5.8. Mplus input file for the specification of the random coefficient regres-
sion model for the outcome variable math and the level-1 predictor variable kft.
 Multilevel Regression Analysis   223

random slopes in Mplus, the analysis command has to be changed to

type = twolevel random. Subsequently, the relevant random slope
has to be defined in the model statement. Mplus treats a random slope as
if it were a latent variable, such that the random slope automatically has a
variance at level 2. In the syntax file, this is specified as follows:

model: %within%
beta1j | math on kft;

In this specification, beta1j denotes the (arbitrary) label chosen for the
random slope coefficient. This line of command means that the slope of
the level-1 regression of math test scores on KFT scores is treated as a
latent variable that can vary across level-2 units. At level 2, we include a
command for the estimation of the covariance t01 between the random
intercept and the random slope, because this parameter is not estimated
by default in Mplus:

%between%
math with beta1j;

In the Mplus output, we obtain the following parameter estimates

(note that for analysis: type = twolevel random, Mplus cur-
rently does not provide standardized parameter estimates):

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

Residual Variances
MATH 14.862 1.091 13.627 0.000

Between Level

MATH WITH
BETA1J 0.050 0.032 1.572 0.116

Means
MATH 12.195 0.456 26.771 0.000
BETA1J 0.233 0.012 19.401 0.000

Variances
MATH 4.801 1.356 3.542 0.000
BETA1J 0.001 0.001 0.672 0.502
224   Data Analysis with Mplus

The level-1 residual variance is estimated at σˆ 2rij = 14.862. This is a con-

siderably smaller value than in the intercept-­only model (20.853) in which
the predictor of KFT was not included. For the between level, we first
obtain estimates of the mean intercept (gˆ 00 = 12.195) and the mean slope
(gˆ 10 = 0.233) across clusters. Therefore, the “average” regression equation
across all school classes can be written as:

Math = 12.195 + 0.233 · kft_centered + residual

For comparison purposes, we again consider the OLS regression analysis

that does not account for the cluster structure in the data:

Math = 12.533 + 0.279 · kft_centered + residual

We can see that the results are very similar. Nonetheless, the random
coefficient regression model is the preferred method in this case because
it takes the dependencies among observations into account that arise due
to clustering, such that the standard errors are correctly estimated and
potential differences in the intercepts and slopes across classes are per-
mitted.
Variability in the intercepts and slopes can be analyzed by the vari-
ance estimates for the between level. Once again, we can see that there
is variability in the intercepts (σˆ 2u0 j = 4.801). It can be seen that the vari-
ance in the slopes (σ 2u1 j = 0.001) is rather small. Nonetheless, one should
not rule out the possibility that differences in the slopes between school
classes may be explained through level-2 predictor variables if such effects
are theoretically predicted. Predicting differences in the slopes across clus-
ters can be done by specifying so-­called intercepts-­and-­slopes-­as-­outcomes
models, which are described in the next section. Finally, the estimated
covariance between intercept and slope is tˆ 01 = 0.050, which is statisti-
cally nonsignificant (z = 1.572, p = .116).

5.5.2 Intercepts‑and‑Slopes‑as‑Outcomes Model

Often, researchers are interested in the influence of variables at both the
individual level (micro variables, level-1 predictors) and the cluster level
(macro variables, level-2 predictors). Thus far, we have discussed only
models in which there were either no predictor variables at all (intercept-­
only model, Section 5.4.1) or there were predictor variables only at level
1 (one-way random effects ANCOVA, Section 5.4.2; random coefficient
 Multilevel Regression Analysis   225

regression model, Section 5.5.1) or only at level 2 (means-as-­outcomes

model, Section 5.4.3). We now want to learn about a more general model
in which both random intercepts and random slopes are specified and in
which there are both level-1 and level-2 predictor variables at the same
time. Following Luke (2004), we refer to this model as intercepts-­and-­
slopes-­as-­outcomes model. As an example, we use the KFT score as a level-1
predictor and school type as a level-2 predictor of math performance.
Below are the formal equations for an intercepts-­and-­slopes-­as-­outcomes
model with one level-1 and one level-2 predictor:

Level 1: Yij = b0j + b1j · Xij + rij

Level 2: b0j = g00 + g01 · Wj + u0j

b1j = g10 + g11 · Wj + u1j

In our example, we again consider the regression of the individual

math test scores on the individual KFT scores at level 1. The intercept
b0j and the slope b1j in this regression are allowed to vary across school
classes (as in the random coefficient regression model discussed in Sec-
tion 5.5.1). The residual term rij refers to differences in the math test per-
formance between students that cannot be explained through the level-1
predictor.
At level 2, the intercept and slope coefficients from level 1 are
regressed on the level-2 predictor school type (Wj). (Note that it is also
possible to include a predictor variable for only one of the two level-1
coefficients—­e.g., only for the intercept term.) The parameter g00 repre-
sents the expected intercept when the level-2 predictor takes on the value
of zero. In our example, g00 therefore indicates the intercept in the group
that is coded zero, which is the conventional high school group.
The parameter g01 gives the expected change in the intercept for a
1-unit change in the level-2 predictor variable. In our case, the level-2 pre-
dictor stype is dichotomous with values 0 = conventional high school
and 1 = gymnasium. Therefore, g01 gives us the expected difference in the
intercepts between group gymnasium and group high school. If g01 = 0,
there would be no difference in the intercepts between school types. If
g01 > 0, the intercept in group gymnasium would be larger than the inter-
cept in group high school. If g01 < 0, the intercept in group high school
would be larger than in group gymnasium.
The interpretation of the parameters g10 and g11 follows a similar
principle, except that those coefficients refer to the slope of the regression
226   Data Analysis with Mplus

line. The parameter g10 gives the estimated slope in the group that is
coded zero (high school). The parameter g11 indicates the difference in
the slopes (gymnasium minus high school). If g11 = 0, there is no differ-
ence in the slopes between the school types. If g11 > 0, the slope in group
gymnasium is larger than the slope in group high school. If g11 < 0, the
slope in group high school is larger than the slope in group gymnasium.
The parameter g11 is often of particular interest because this coef-
ficient reflects a potential interaction effect between the level-1 predictor
(here, KFT score) and the level-2 predictor (here, school type). This is
often referred to as a cross-level interaction, and it means that the rela-
tionship between the level-1 predictor and the outcome varies depending
upon the values of the level-2 predictor. In our example, this would mean
that if a cross-level interaction is present, the regression of math on kft
would be different for different school types. A significant g11 coefficient
would indicate a significant cross-level interaction.
The residual terms u0j and u1j reflect differences in the intercept and
slope coefficients between school classes that cannot be explained through
the level-2 predictor school type. In summary, we obtain the following
estimates for the intercepts-­and-­slopes-­as-­outcomes model in Mplus:

• At level 1 (WITHIN):
—The
— residual variance σ 2rij
• At level 2 (BETWEEN):
—The
— fixed intercept g00 for the regression of the random inter-
cept b0j on Wj
—The
— fixed slope g01 for the regression of the random intercept
b0j on Wj
—The
— fixed intercept g10 for the regression of the random slope
b1j on Wj
—The
— fixed slope g11 for the regression of the random slope b1j on
Wj
—The
— residual variance σˆ 2u0 j for the random intercepts
—The
— residual variance σ 2u1 j for the random slopes
—The
— residual covariance t01 between the random intercept and
random slope (provided only upon request in Mplus)

The Mplus input file for the specification of the intercepts-­ and-­
slopes-­as-­outcomes model is shown in Figure 5.9. Note that the variable
kft was again defined as a level-1 variable. Furthermore, the variable kft
 Multilevel Regression Analysis   227

FIGURE 5.9. Mplus input file for the specification of the intercepts-­and-­slopes-­
as-­outcomes model for the outcome variable math, the level-1 predictor variable kft,
and the level-2 predictor variable stype.

was centered at the grand mean to allow for a meaningful interpretation

of the results. The variable stype was defined as a level-2 variable, so that
this variable can only be used at level 2. In the model statement for level 1
(%within%) the slope for the regression of math on kft is again defined
as a random effect (named beta1j):

model: %within%
beta1j | math on kft;

At level 2 (%between%), the regression of the random intercept on the

level-2 predictor stype is requested by the command
228   Data Analysis with Mplus

%between%
math on stype;

Subsequently, it is specified that the random slope beta1j is also

regressed on school type:

beta1j on stype;

The command math with beta1j; is included to request estimation

of the residual covariance t01 between the random intercept and slope. All
relevant residual variances are estimated by default. In the Mplus output
we obtain the following parameter estimates for our example:

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Within Level

Residual Variances
MATH 14.608 1.070 13.658 0.000

Between Level

BETA1J ON
STYPE -0.058 0.018 -3.139 0.002

MATH ON
STYPE 5.394 0.969 5.568 0.000

MATH WITH
BETA1J 0.035 0.022 1.567 0.117

Intercepts
MATH 11.536 0.396 29.149 0.000
BETA1J 0.226 0.013 17.258 0.000

Residual Variances
MATH 3.022 0.724 4.171 0.000
BETA1J 0.001 0.001 0.509 0.611

The level-1 residual variance is estimated to be σˆ 2rij = 14.608. At level

2, Mplus first outputs the estimated regression coefficient gˆ 11 for the
regression of the random slope b1j on school type. In the present example,
the estimated coefficient gˆ 11 gives the difference in the estimated mean
 Multilevel Regression Analysis   229

regression slopes (gymnasium minus high school). The value of gˆ 11 is

–0.058 which is statistically significant at the .01 level (z = –3.139, p =
.002). This means that the slope of the regression line in the regression of
math score on KFT score differs significantly between group gymnasium
and group high school.
This cross-level interaction effect is illustrated in Figure 5.10 in a
scatter plot that shows the conventional OLS regression lines by school
type. Figure 5.10 shows that the regression line for the group gymnasium
(dotted line) has a flatter slope than the corresponding regression line for
the group high school (solid line). According to the z-test for gˆ 11, this dif-
ference in the slopes of the regression lines is statistically significant. The
negative sign of gˆ 11 shows us that the average slope coefficient is smaller
(less steep) in group gymnasium than in group high school, because gym-
nasium is coded 1 and high school is coded 0. (A positive sign of gˆ 11
would have indicated a steeper slope in group gymnasium than in group
high school.)
Next, Mplus reports the estimated regression coefficient gˆ 01 for the
regression of the random intercept on school type. In our case, the coeffi-
cient gˆ 01 gives the difference in the estimated mean intercepts (gymnasium

32

30 Conventional high school

Gymnasium
28 ___ Conventional high school
------ Gymnasium
26

24

22

20
Math Score

γˆ 00 + γˆ 01 = 16.93
18

16

14 γˆ 01 = 5.394
12 γˆ 00 = 11.536
10

0
–50.00 –25.00 .00 25.00 50.00

KFT Score (grand-mean-centered)

FIGURE 5.10. OLS regression of math score on grand-mean-­centered KFT score

by school class and illustration of the estimated level-2 regression intercept coeffi-
cients in the intercepts-­and-­slopes-­as-­outcomes model.
230   Data Analysis with Mplus

minus high school). The value of gˆ 01 is 5.394, which is also statistically sig-
nificant (z = 5.568, p < .001). In summary, the level-2 predictor school type
explains differences in both the intercept and slope coefficients between
school classes. Given that the predictor variable kft has been centered at
the grand mean, the estimated value of 5.394 indicates that students with
an average KFT score (relative to the entire sample) are expected to have
5.394 points more on the math test if they belong to the group gymna-
sium than if they belong to the group high school. The difference in the
intercepts is also illustrated in Figure 5.10. The regression lines for the
two school types intersect the zero point of the grand-mean-­centered kft
variable at different levels (the dotted line for the gymnasium cuts the zero
point above the solid line of the high school group).
Under Intercepts we first obtain the estimated intercept, gˆ 00 , for
the regression of the random intercept, b0j, on school type (gˆ 00 = 11.536).
This value tells us that the estimated intercept for the regression of math
performance on KFT score in group high school is 11.536 (cf. Figure
5.10). Therefore, for high school students with an average KFT value, a
math test score of 11.536 is expected. The sum of the coefficients gˆ 00 +
gˆ 01 equals the expected intercept for students in the group gymnasium:
11.536 + 5.394 = 16.93. Consequently, for students in the gymnasium
with an average KFT score, a math test score of 16.93 is expected.
In addition, the estimated intercept gˆ 10 for the regression of the
random slope, b1j, on school type is estimated (gˆ 10 = 0.226). This value
gives the estimated slope of the regression line for the regression of the
math score on the KFT score in group high school (gˆ 10; this coefficient is
approximately equal to the slope of the solid OLS regression line in Figure
5.10). Therefore, we can now also calculate the expected slope in group
gymnasium, which is equal to the sum of gˆ 10 + gˆ 11 = 0.226 + (–0.058)
= 0.168. This shows that the slope of the regression line is flatter for stu-
dents in the group gymnasium than in the group high school. At level
2, Mplus additionally estimated the residual variances for the intercept
(σˆ 2u0 j = 3.022) and the slope (σˆ 2u1 j = 0.001) as well as the residual covari-
ance between intercept and slope (tˆ 01 = 0.035).
Note that Mplus can also be used to analyze multilevel regression
models with more than one level-1 and/or more than one level-2 predic-
tor. In such extended models, some or all intercepts and slopes can be
specified as random coefficients. For predictors for which the slopes are
not supposed to be random coefficients, a separate on statement needs to
be included in the model command, in which the | sign as well as the
name of the random slope is omitted. Consider the following example:
 Multilevel Regression Analysis   231

model: %within%
beta1J | y on x1;
y on x2;

In this example, only the slope of the regression of the variable Y on X1

(beta1J | y on x1;) would be a random slope, whereas the slope
of the regression of Y on X2 (y on x2;) would be fixed. The slope of the
regression of Y on X2 is a constant (fixed) effect in this example that does
not vary between clusters.
A particular strength of Mplus is that not only univariate multilevel
regression models (models with a single dependent variable), as discussed
in this book, can be analyzed, but also multivariate regression models,
in which multiple dependent variables are considered at the same time.
This means that in case of a multilevel data structure, researchers can also
analyze path analyses, CFAs, SEMs (see Chapter 3), LGCMs (see Section
4.5), and latent class models (see Chapter 6) as two-level models using
the option analysis: type = twolevel. The analysis of more
complex multivariate multilevel models in Mplus is discussed in Geiser
et al. (2012).
6

Latent Class Analysis

6.1 Introduction to Latent Class Analysis

Latent class analysis (LCA) or latent structure analysis (Goodman, 1974;

Lazarsfeld & Henry, 1968) is a statistical procedure that can be used to
classify individuals into homogeneous subgroups (latent classes, latent
types). The starting points for the classification are the observed response
patterns of individuals across a set of categorical (nominal or ordinal)
test or questionnaire items (e.g., intelligence test items solved/not solved;
symptoms present vs. absent; questionnaire items measured on a Likert
scale). In an LCA, the relationships between the items are explained by
the presence of a priori unknown subpopulations (the latent classes). In
other words, individual differences in observed item response patterns
are explained by differences in latent class membership, where each class
shows a characteristic, class-­specific response profile.
As an example, consider individuals who solve only certain item
types for measuring spatial ability with a high probability, but not oth-
ers. If different items can be solved with different solution strategies, this
might explain these differences. Hence, individuals showing this specific
response profile could be assigned to a latent class that represents the
specific solution strategy used by these individuals. Similarly, individuals
who guess and thus give random answers in a test may form a “guessing
class” in an LCA. In the field of clinical psychology, patients with distinct
symptom patterns could be assigned to different latent classes represent-
ing different types of disorders or diseases. As an example from personal-
ity research, individuals could also be classified according to their person-
ality type. The main goals of LCA are summarized in Box 6.1.
232
 Latent Class Analysis   233

BOX 6.1. Important Goals of an LCA

• Determine the number of classes necessary to sufficiently explain differ-

ences in the observed response patterns and summarize the data in a sub-
stantively meaningful way (exploratory LCA). The model should be parsi-
monious, that is, use as few latent classes as possible to explain the data.
• Examine what the classes are like: How can the classes be interpreted
substantively, and in which way do they differ from one another?
• Test hypotheses about typological structures (confirmatory LCA).
• Evaluate the reliability of the classification based on a specific LCA solu-
tion.
• Determine the most likely latent class membership of one or more indi-
viduals for diagnostic purposes.
• Analyze relationships between class membership and external variables
(e.g., age, sex, personality variables, intelligence).

The above examples point to similarities between LCA and explor-

atory factor analysis (EFA). Both approaches can be used for data reduction.
In EFA, the relationships between continuous variables are “explained”
by a smaller number of continuous latent variables (factors). In classical
LCA, the relationships between categorical variables are “explained” by
membership in one of several latent classes. One of the main differences
between LCA and EFA is that EFA uses continuous latent variables (“fac-
tors”) and metrical indicators, whereas LCA uses one categorical latent
variable (a latent class variable with several categories = latent classes) to
explain the relationships between categorical indicators.
LCA cannot only be used in an exploratory way, but also as a con-
firmatory analysis technique. For example, LCA allows us to test theo-
ries about typological differences between individuals (e.g., differences
in attachment style, personality types, solution strategies) by assuming a
specific number of classes and/or imposing parameter restrictions (e.g.,
with regard to the properties or sizes of the classes). Using model tests
and/or model comparisons, we can test whether the hypothesized typo-
logical model fits the observed data. This is similar to using model tests in
CFA and SEMs (see Chapter 3).
For simplicity, we consider only the analysis of dichotomous variables
using exploratory LCA in this book. The procedure can be easily general-
ized to polytomous variables and analyses that are more confirmatory in
234   Data Analysis with Mplus

nature. For a dichotomous variable, the LCA model for a single item can
be written as follows (e.g., see Collins & Lanza, 2010; Rost, 2004):

G
p( X vi = 1) = ∑ p g pig .
g =1

In this equation, p(Xvi = 1) denotes the unconditional probability that a

randomly selected individual v obtained a score of X = 1 on item i, i =
1, . . . , I (e.g., correctly solved a test item or endorsed a questionnaire
item). The class size parameter pg indicates the unconditional probability
of belonging to latent class g (g = 1, . . . , G). The model assumes that each
individual belongs to one and only one latent class. Therefore, the sum of
all class-size parameters equals 1:

G
∑ pg = 1
g =1

The parameter pig indicates the conditional probability of a score of 1

on item i given membership in class g:

pig = p(Xvi = 1|G = g)

This probability is often referred to as a conditional response or conditional

solution probability. The LCA model therefore states that an individual’s
response probability for item i depends on both the latent class preva-
lences and the class-­specific response probabilities of the item. Box 6.2
gives an overview of the parameters that are estimated in an LCA model.
Note that the number of classes that is necessary to appropriately
account for the observed response patterns is not a model parameter to
be estimated. Similarly to EFA, the number of classes in exploratory LCA
is usually determined by means of certain fit criteria and model compari-
sons. Models with different numbers of classes are fit to the same data
set. Using statistical indices, the fit of the models is compared. The model
with the best fit is selected, if this solution is also proper and easily inter-
pretable in terms of the estimated parameters. Criteria for assessing model
fit and for statistical model comparisons are discussed in Section 6.3.
More detailed descriptions of LCA as well as extensions of the basic
model can be found in Clogg (1995), Collins and Lanza (2010), Hagenaars
 Latent Class Analysis   235

BOX 6.2. Summary of Parameters Estimated in an LCA Model for

Dichotomous Variables

• pg = class size or class proportion parameter. G – 1 class size parameters

pg are estimated. The size of the last class is not an independent model
G
parameter because of the assumption that ∑ p g = 1.
g=1
• pig = conditional response probabilities for item i in class g. For dichoto-
mous indicators, G · I conditional response probabilities pig are estimated
(the corresponding probability for the second category of each item is
given by 1 – pig).
• Mplus also outputs all parameters of the LCA model in an equivalent logis-
tic threshold parameterization (e.g., see Hagenaars, 1993). The conver-
sion of the logistic threshold parameters into probabilities is discussed in
Chapter 13 of the Mplus User’s Guide (Muthén & Muthén, 1998–2007).

(1993), Hagenaars and McCutcheon (2002), Langeheine and Rost (1988),

McCutcheon (1987), Rost and Langeheine (1997), as well as on the
homepage of John Uebersax (http://www.john-­uebersax.com/stat/soft.htm).
The use of LCA for comparisons of multiple groups (multigroup LCA) is
discussed in Eid, Langeheine, and Diener (2003) in the context of cross-­
cultural research. An applied example of a multigroup LCA, including
sample Mplus syntax, can be found in Geiser, Lehmann, and Eid (2006).

6.2 Specification of LCA Models in Mplus

The specification of LCA models in Mplus requires the module for the
analysis of mixture distribution models (option analysis: type =
mixture). Either the mixture or the combination add-on upgrades are
needed to analyze LCA models in Mplus in addition to the Mplus base
version. Box 6.3 introduces the data example used in this chapter.
Figure 6.1 shows the Mplus input file for an LCA model with three
latent classes. (My rationale for preferring a three-class solution will
become clear in later sections when we discuss model fit.) The data set
used here has the name computergames.dat. It contains the subject ID
number (variable id) as well as the responses on the eight computer
game questionnaire items (variable names: c1–c8). Using the command
variable: auxiliary = id; we tell Mplus that the variable id
236   Data Analysis with Mplus

BOX 6.3. Data Example Used in This Chapter

We illustrate the analysis of an LCA in Mplus using the data set computer-
games.dat. This data set contains the item responses of N = 861 German
children, who answered a questionnaire on their computer game preferences.
In the following, an English translation of this questionnaire is shown:

How often do you play the following computer games?

Never or rarely Often or very often
1. Adventure games
2. Action games
3. Sport games
4. Fantasy role-playing games
5. Logic games
6. Skill-­training games
7. Simulation games
8. Driving simulator games

Quaiser-­Pohl, Geiser, and Lehmann (2006) applied LCA to the dichotomous

item responses to find out whether different types of computer game play-
ers could be identified based on the eight computer game items above. More
specifically, they wanted to find out how many distinct latent classes would
emerge and how each latent class would be interpreted. In addition, they
were interested in finding out how reliably children could be classified into
distinct computer game player types. For information on additional research
questions examined in this study, see Quaiser-­Pohl et al. (2006).

is used as an auxiliary variable that does not play a role in the statistical
analysis as such, but that should still be included in a newly generated
data set containing the individual class assignment probabilities pro-
duced by Mplus on request (see description of the savedata command
on p. 239). Without the auxiliary command, the subject ID would not be
saved to the new data set. This could be a disadvantage when, for exam-
ple, LCA is used for individual diagnostic purposes. Specific individuals
in the data set could no longer be linked to the estimated class member-
ship information if there is no ID variable. The command categorical
 Latent Class Analysis   237

FIGURE 6.1. Mplus input file for the specification of an LCA with three classes for
the eight computer game items.

= c1–c8; defines the eight computer game items as (ordered) categorical

(here, dichotomous). The same command would also be used if ordered
categorical items with more than two categories (ordered polytomous
items) were used. (For unordered categorical [nominal] variables, the
command nominal = would be used.) Using the command classes
= c(3); we specify that the latent categorical (or latent class) variable
has the name c and three categories (i.e., three latent classes). This means
that a three-class solution will be extracted. Using the command analy-
sis: type = mixture; we tell Mplus that a mixture distribution
analysis (here LCA) is to be carried out. Mixture distribution means that
the population of students is assumed to consist of distinct subgroups
(latent classes reflecting distinct computer game types). The goal of LCA
is to “unmix” the population into homogeneous subgroups.
Similarly to CFA and SEMs, the parameter estimates as well as model
fit statistics (see Section 6.3) for an LCA are determined using an itera-
tive estimation procedure that is based on the maximum likelihood
(ML) estimation method by default in Mplus. The goal is to maximize
238   Data Analysis with Mplus

the likelihood function to find the solution with the largest possible log
likelihood value. The log likelihood value is a measure of the probability of
the observed data given the model and is used as the basis for calculating
various fit statistics (see Section 6.3).
One complication for iterative estimation procedures that is particu-
larly relevant to LCA (and mixture distribution analyses in general) is that,
often, a single set of parameter starting values will not allow us to find the
solution with the best possible log likelihood value. The consequence may
be that the model estimation terminates at a so-­called local likelihood maxi-
mum, which is often associated with inaccurate parameter estimates. Such
a local solution can, for example, be associated with incorrect fit statistics
and biased parameter estimates (e.g., see Uebersax, 2000). Local solutions
should therefore be avoided (cf. Box 6.4).
If only a few sets of starting values are tried in an LCA, there is a
relatively high chance of encountering a local likelihood maximum.
Moreover, studies indicate that local maxima occur frequently in empiri-
cal applications, even if only four or five classes are extracted (Uebersax,
2000). The use of a sufficient number of random sets of starting values
increases the chances that an LCA program will find the optimal solution
with the highest log likelihood value. In fact, for relatively simple, clearly
identified LCA models with few classes, the use of a large number of start
values usually guarantees that the true maximum will be found.
The Mplus default setting is that 10 sets of random starting values
are generated by the program and used in the model estimation. This
relatively small number of start value sets can be insufficient in practice,
especially for more complex models when more than two or three latent
classes are extracted. To be on the safe side, we encourage users to specify
a larger number of starting value sets. A relatively safe setting is to request
at least 500 sets of start values in the first step of the optimization (for
complex models, an even larger number may be necessary). The com-
mand analysis: starts = can be used to increase the number of
starting values in Mplus. The specification starts = 500 50; causes
Mplus to use 500 sets of start values in the first step of the optimization
(instead of only 10) and to pick the 50 starting value sets that show the
largest log likelihood values in the first step for the second step of the opti-
mization process, until a convergence criterion (in Mplus the default is a
parameter change of less than 0.000001) is reached. The solution with the
highest log likelihood value is then output by Mplus as the final solution.
In addition to increasing the number of start value sets, it is also
recommended to choose a sufficient number of iterations for the initial
stage of the optimization (Uebersax, 2000). In Mplus, the command
 Latent Class Analysis   239

analysis: stiterations = can be used to increase the number

of iterations in the first step of the optimization (the Mplus default is only
10 iterations in the first step). The command stiterations = 50;
increases the number of initial stage iterations from the default of 10 to 50.
In general, if the best log likelihood value was not found by Mplus
for at least two sets of starting values, Mplus issues a warning message.
The model should be rerun with a larger number of starting values in this
case, before any parameter estimates are interpreted. If the best log likeli-
hood value cannot be replicated by at least one other starting value set,
even for a very large number of starting value sets (say, 2,000), the results
should be interpreted with great caution. The model may not be identified
in this case or it may not be well defined for the data at hand (e.g., due to
a too-small sample). Box 6.4 summarizes the most important strategies for
avoiding local maxima in an LCA.
The plot option is useful to visualize the estimated conditional
response probabilities in different classes. In LCA this is often done by
means of a line graph, in which the items are shown on the x-axis and
the conditional response probabilities are shown on the y-axis. Different
classes are represented by separate lines. In this way, one obtains a profile
of the items in each class that often allows for an easier interpretation of
the classes. Such a diagram can be obtained in Mplus using the option
plot: type = plot3;. Using the series option, it is specified in
which order the items will be shown on the x-axis:

series = c1 (1) c2 (2) c3 (3) c4 (4) c5 (5) c6 (6)

c7 (7) c8 (8);

The resulting plot is shown below, where we discuss the Mplus output for
an LCA.
Based on the estimated LCA model parameters, the probability with
which each individual belongs to each of the extracted classes can be
estimated. On the basis of these individual class assignment probabilities,
one can estimate the most likely class membership for each individual for
diagnostic purposes. An individual is assigned to the class for which his
or her assignment probability is highest. In Mplus, the savedata com-
mand can be used to save individual class assignment probabilities and
the most likely latent class membership information to an external file. In
order to save these values to an external file, we specify:

savedata: file = computergames_3classes.dat;

save = cprobabilities;
240   Data Analysis with Mplus

BOX 6.4. Avoiding Local Likelihood Maxima in LCA

Local likelihood maxima are frequently encountered in practical applica-

tions of LCA. Following the recommendations below can help to avoid local
solutions (see also Uebersax, 2000):

• Use a sufficient number of different start values. In the estimation of LCA

models with more than two classes, a sufficient number of random sets of
start values should be used (say, at least 500 in the first and the best 50 in
the second step of the optimization; with complex models 1,000/100 or
more should be used) → analysis: starts = 500 50; or starts
= 1000 100;
• Use a sufficient number of initial stage iterations. It is also recommended to
choose a larger number of initial stage iterations than Mplus uses as the
default (e.g., analysis: stiterations = 50;).
• Check the replicability of the best log likelihood value. Users of LCA should
check that the largest (best) log likelihood value was the result of at least
two different sets of start values. This can easily be checked in the Mplus
output (see discussion in the text).
• Prefer models with fewer latent classes. If possible, users should avoid the
extraction of a large number of latent classes, because local maxima are
more likely to occur in models with more classes as compared to models
with fewer classes. Already with four or more classes, local maxima tend
to be more frequently observed.
• Use a tight convergence criterion. The Mplus criterion for convergence is
0.000001. If in doubt, one should choose the smaller convergence cri-
terion to avoid local maxima (e.g., analysis: convergence =
0.0000001;).
• Avoid solutions that contain many boundary parameter estimates. If many con-
ditional response probabilities are estimated to exactly 0 or 1 in a solution
(so-­called boundary estimates; cf. Box 6.5), this can be a sign of an invalid
solution and/or indicate that too many classes were extracted.
• Replicate the model parameter estimates in one or more additional LCA pro-
grams. It can be reassuring to compare the estimates across different soft-
ware programs, especially when analyzing complex models with many
classes. Ideally, the best log likelihood value should be found by all pro-
grams.
• Replicate the model with additional data. Another sign of the validity of a
class solution is if this solution can be replicated with a fresh data set.
 Latent Class Analysis   241

Through this command, Mplus saves a new file with the name
computergames_3classes.dat to the same directory that also contains
the input file. The new file contains the eight computer game variables,
the ID variable id (because we defined this variable as an auxiliary vari-
able earlier under the variable command), the individual class assignment
probabilities (new variables named CPROB1, CPROB2, and CPROB3 to be
generated by Mplus in the analysis), and the estimated class membership
(new variable C to be generated by Mplus).
For an LCA, Mplus by default estimates the parameters using ML esti-
mation with robust standard errors. Conventional ML standard errors can
be requested using the command analysis: estimator = ml;.
In the following, the most important part of the Mplus output for the
three-class solution of the computer game example is shown. First of all,
we obtain the unconditional item response probabilities for each category
in the total sample. It can be seen that overall, the category never/rarely
was chosen more frequently in all items. Of interest to us is whether the

BOX 6.5. Boundary Parameter Estimates

Boundary parameter estimates are estimates at the border of the theoretically

admissible parameter space. For probabilities (such as the pig parameters in
LCA) the admissible range lies between 0 and 1 (which corresponds to val-
ues of +15 and –15 in the Mplus threshold parameterization, respectively).
Boundary estimates are more likely to occur when the data are sparse (e.g.,
due to the use of many items and/or a small sample), as well as in solutions
with many classes. Boundary estimates are often a sign that too many classes
were extracted, but can also sometimes be a symptom of a local maximum
or an identification problem. For this reason, boundary solutions should be
interpreted with caution. In Mplus, the occurrence of a boundary solution is
indicated through the following warning message:
IN THE OPTIMIZATION, ONE OR MORE LOGIT THRESHOLDS APPROACHED AND
WERE SET AT THE EXTREME VALUES. EXTREME VALUES ARE -15.000
AND 15.000.
THE FOLLOWING THRESHOLDS WERE SET AT THESE VALUES:
* THRESHOLD 1 OF CLASS INDICATOR C1 FOR CLASS 2 AT ITERATION 35
* THRESHOLD 1 OF CLASS INDICATOR C4 FOR CLASS 2 AT ITERATION 38
* THRESHOLD 1 OF CLASS INDICATOR C2 FOR CLASS 2 AT ITERATION 58
* THRESHOLD 1 OF CLASS INDICATOR C3 FOR CLASS 3 AT ITERATION 66
* THRESHOLD 1 OF CLASS INDICATOR C8 FOR CLASS 3 AT ITERATION 69

This message indicates that in the estimation process, one or more condi-
tional response probabilities were estimated and set to boundary values of
0 or 1 (which corresponds to threshold values of +15 and –15 in Mplus).
242   Data Analysis with Mplus

conditional response probabilities show a different pattern in subgroups,

that is, in the three extracted latent classes.

SUMMARY OF CATEGORICAL DATA PROPORTIONS

C1
Category 1 0.822
Category 2 0.178
C2
Category 1 0.727
Category 2 0.273
C3
Category 1 0.746
Category 2 0.254
C4
Category 1 0.816
Category 2 0.184
C5
Category 1 0.739
Category 2 0.261
C6
Category 1 0.746
Category 2 0.254
C7
Category 1 0.649
Category 2 0.351
C8
Category 1 0.674
Category 2 0.326

Next is a summary of the log likelihood values for the 50 sets of start
values that showed the largest log likelihood values after the initial step
of the optimization (ordered from the best to the worst log likelihood
values), together with the associated seeds (seed = random start number
for generating start values) and the number of the corresponding starting
value set. It can be seen that the same log likelihood value of –3463.814
was found for each of the 50 best sets of start values from the first step.
This is a good sign because it shows that the solution is replicable across
a variety of different start values. This indicates that local maxima are
apparently not an issue for this solution. Such a result is typical for well-
defined models with a relatively small number of classes.

RANDOM STARTS RESULTS RANKED FROM THE BEST TO THE

WORST LOGLIKELIHOOD VALUES

Final stage loglikelihood values at local maxima,

seeds, and initial stage start numbers:
 Latent Class Analysis   243

-3463.814 605358 321

-3463.814 57226 208
-3463.814 714455 476
-3463.814 965639 463
-3463.814 798821 423
-3463.814 85734 411
-3463.814 926797 406
-3463.814 973369 202
-3463.814 7959 256
-3463.814 987090 70
-3463.814 137305 379
-3463.814 732596 320
-3463.814 900268 327
-3463.814 319575 499
-3463.814 576596 99
-3463.814 772131 407
-3463.814 100874 108
-3463.814 377504 294
-3463.814 694303 282
-3463.814 741888 138
-3463.814 582296 452
-3463.814 344422 296
-3463.814 340112 126
-3463.814 937225 394
-3463.814 848890 95
-3463.814 695155 150
-3463.814 455617 242
-3463.814 168762 200
-3463.814 846194 93
-3463.814 347515 24
-3463.814 970689 266
-3463.814 21345 199
-3463.814 652266 490
-3463.814 971853 402
-3463.814 81117 305
-3463.814 417035 149
-3463.814 570908 98
-3463.814 655497 376
-3463.814 637345 19
-3463.814 736574 414
-3463.814 923437 398
-3463.814 782821 272
-3463.814 315029 471
-3463.814 850840 232
-3463.814 972873 157
-3463.814 860102 495
-3463.814 793035 187
-3463.814 414828 322
-3463.814 483369 270
-3463.814 529496 343
244   Data Analysis with Mplus

By default Mplus informs the user of the issue of local maxima with the
following message, which is always printed under the list of log likelihood
values:

WARNING: WHEN ESTIMATING A MODEL WITH MORE THAN TWO CLASSES,

IT MAY BE NECESSARY TO INCREASE THE NUMBER OF RANDOM STARTS
USING THE STARTS OPTION TO AVOID LOCAL MAXIMA.

In the present example, this message can be ignored, because we have

clearly used a sufficient number of random starts for the three-class
model, as seen above.
In models with more latent classes, one often finds that there are
slight differences among the log likelihood values in the Mplus list. In this
case, it is reassuring when the best log likelihood value was found for at
least two different sets of start values. If this is not the case, Mplus issues
the following warning message:

WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED.

THE SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA.
INCREASE THE NUMBER OF RANDOM STARTS.

As recommended by Mplus, in this case one should increase the

number of random starts and/or the number of iterations in the first step
of the optimization (see above) until the best log likelihood value is reli-
ably identified by multiple sets of start values. If the best log likelihood
value cannot be replicated even with an increased number of start values,
this may indicate that the model is not well defined for the data. Subse-
quently, only the following message (and no additional error messages)
should appear:

THE MODEL ESTIMATION TERMINATED NORMALLY

If the model converged properly, the best log likelihood value for the final
solution is output once again under MODEL FIT INFORMATION:

MODEL FIT INFORMATION

Loglikelihood

H0 Value -3463.814

This log likelihood value is identical to the best value that appears on the
very top of the previous listing of log likelihood values for different sets
 Latent Class Analysis   245

BOX 6.6. Excursion optseed Option

The listing of the log likelihood values in Mplus can also be useful for repli-
cating the solution with the best log likelihood value in subsequent analyses
steps. This allows the user to reproduce the best solution without having
to generate a large number of start value sets again, which can be time-­
consuming for complex models. This approach can be useful, for example,
when one wants to request additional output for a solution that has not been
requested initially. Using the so-­called optseed option in Mplus, one can
replicate a specific solution on the basis of the specific seed number associ-
ated with this solution. In the current example, we could specify the follow-
ing in the input:

analysis: optseed = 605358;

where the number 605358 refers to the seed number that pertains to the
starting value set with the number 321, which is the starting value set that is
ranked first in the list of log likelihood values:

RANDOM STARTS RESULTS RANKED FROM THE BEST TO THE WORST

LOGLIKELIHOOD VALUES

Final stage loglikelihood values at local maxima, seeds,

and initial stage start numbers:

-3463.814 605358 321

This starting value set is among those sets that produced the best log likeli-
hood value of –3463.814. Note that in this example one could have chosen
any one of the 50 seed numbers, because all 50 start value sets resulted in
the same best log likelihood value. This, however, is not always the case,
especially for models with more classes. Using the optseed command as
specified above, Mplus will replicate the exact same solution that has been
found based on the seed 605358, although this should be verified in the
subsequent output file.

of start values. Following are different model fit statistics, which are dis-
cussed in detail in Section 6.3:

Information Criteria

Number of Free Parameters 26

Akaike (AIC) 6979.628
Bayesian (BIC) 7103.338
246   Data Analysis with Mplus

Sample-Size Adjusted BIC 7020.769

(n* = (n + 2) / 24)

Chi-Square Test of Model Fit for the Binary and

Ordered Categorical
(Ordinal) Outcomes

Pearson Chi-Square

Value 386.883
Degrees of Freedom 229
P-Value 0.0000

Likelihood Ratio Chi-Square

Value 340.605
Degrees of Freedom 229
P-Value 0.0000

Next, Mplus provides information about the estimated sizes of the latent
classes in the model. The following table contains the class sizes in terms
of the estimated p̂ g parameters:

FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT

CLASSES BASED ON THE ESTIMATED MODEL

Latent
Classes

1 149.48776 0.17362
2 327.71918 0.38063
3 383.79305 0.44575

According to these estimates, Class 3 is the largest class with approx-

imately 44.6% (pˆ 3 = .44575). Relative to our sample, this corresponds to
approximately N = 384 participants. Class 2 has a size of approximately
38.1% (pˆ 2 = .38063, approximately N = 328 participants). Class 1 is the
smallest class with approximately 17.4% (pˆ 1 = .17362, approximately N
= 150). Note that the order of the classes is arbitrary. This means, for
example, that if we run the same three-class solution on the same data but
different computers, we might obtain solutions that differ with regard to
the order of the classes. However, the solution would still show the same
log likelihood value, and the parameter estimates would not be different.
Mplus also outputs class size estimates based on the estimated poste-
rior probabilities. In this example, these estimates are close to identical to
the class proportions based on the estimated model:
 Latent Class Analysis   247

FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT

CLASS PATTERNS BASED ON ESTIMATED POSTERIOR
PROBABILITIES

Latent
Classes

1 149.48771 0.17362
2 327.71923 0.38063
3 383.79307 0.44575

The entropy statistic is a summary measure for the quality of the clas-
sification. Values close to 1 indicate high classification accuracy, whereas
values close to 0 indicate low classification certainty.

CLASSIFICATION QUALITY

Entropy 0.731

Somewhat more informative are the average class assignment prob-

abilities for individuals that were assigned to a specific class. For each
person, LCA allows estimating the probability with which each individ-
ual belongs to each of the three extracted classes. As mentioned above,
for diagnostic (or other) purposes, a person can be assigned to the class
for which he or she has the greatest probability according to his or her
observed response pattern. Mplus first outputs the class sizes that are
based on this kind of manifest classification:

CLASSIFICATION OF INDIVIDUALS BASED ON THEIR MOST

LIKELY LATENT CLASS MEMBERSHIP

Class Counts and Proportions

Latent
Classes

1 152 0.17654
2 317 0.36818
3 392 0.45528

The class sizes that are estimated based on the manifest classification
deviate slightly from the estimated class portions p̂ g in our example. A
certain amount of discrepancy between these two kinds of class propor-
tion estimates is usually observed, because the manifest classification of
persons using the maximum class assignment probability is almost always
248   Data Analysis with Mplus

associated with some amount of estimation error. The reason is that each
person has a certain (nonzero) membership probability for each class,
which is not taken into account when people are assigned to the one latent
class for which the probability is highest. For this reason, the p̂ g param-
eters should generally be preferred as estimates of the latent class sizes.
Following are the average latent class assignment probabilities for
individuals assigned to each class. These are important indicators of the
quality of an LCA solution. Values close to 1 on the main diagonal of the
classification matrix indicate a high precision or reliability of the clas-
sification. High values on the main diagonal of this matrix mean that
individuals, on average, are classified with high certainty into their most
likely latent class. In contrast, moderate to low values indicate that there
is a high uncertainty of the class assignment, on average. According to
Rost (2006), the values on the main diagonal should be .8 or larger for a
good class solution. In our example, we obtained high class assignment
probabilities for Classes 2 and 3 (.913 and .922, respectively), whereas the
mean assignment probability for members of Class 1 is only .76, which
is relatively low. Apparently, there is somewhat more uncertainty with
regard to the assignment to Class 1 compared to Classes 2 and 3. Mem-
bers of Class 1 also have a relatively high average probability of belonging
to Class 2 (.146). Therefore, there is a certain degree of overlap between
these two classes. Nonetheless, the classification quality for Class 1 can
still be seen as satisfactory.

Average Latent Class Probabilities for Most Likely

Latent Class Membership (Row) by Latent Class
(Column)

1 2 3

1 0.760 0.146 0.094

2 0.061 0.913 0.025
3 0.037 0.041 0.922

The more latent classes are extracted, the lower in general are the
mean class assignment probabilities. The reason is that with many
extracted classes, there are more possibilities for an incorrect classifica-
tion. The classes often are not as clearly separated any more.
Under MODEL RESULTS, Mplus first provides the parameters of the
LCA model in the logistic threshold parameterization. First are the esti-
mated threshold parameters that correspond to the conditional response
probabilities p̂ig . Under Categorical Latent Variables—­Means,
 Latent Class Analysis   249

Mplus outputs the threshold parameters that correspond to the class sizes
p̂ g . The conversion of these parameters into probabilities is explained in
Chapter 13 of the Mplus User’s Guide (Muthén & Muthén, 1998–2012).

MODEL RESULTS

Two-Tailed
Estimate S.E. Est./S.E. P-Value

Latent Class 1

Thresholds
C1$1 1.398 0.506 2.761 0.006
C2$1 1.920 0.706 2.717 0.007
C3$1 1.675 0.423 3.961 0.000
C4$1 1.489 0.523 2.850 0.004
C5$1 -0.981 0.406 -2.413 0.016
C6$1 -1.480 0.654 -2.263 0.024
C7$1 0.187 0.357 0.524 0.600
C8$1 0.685 0.334 2.052 0.040

Latent Class 2

Thresholds
C1$1 0.555 0.144 3.852 0.000
C2$1 -0.591 0.191 -3.101 0.002
C3$1 -0.097 0.163 -0.593 0.553
C4$1 0.463 0.143 3.244 0.001
C5$1 1.154 0.247 4.681 0.000
C6$1 1.542 0.291 5.290 0.000
C7$1 -0.628 0.152 -4.134 0.000
C8$1 -0.576 0.159 -3.633 0.000

Latent Class 3

Thresholds
C1$1 4.582 0.780 5.872 0.000
C2$1 4.334 0.804 5.388 0.000
C3$1 2.723 0.322 8.455 0.000
C4$1 4.583 0.927 4.947 0.000
C5$1 2.216 0.300 7.391 0.000
C6$1 2.166 0.366 5.924 0.000
C7$1 2.873 0.326 8.817 0.000
C8$1 2.845 0.410 6.937 0.000

Categorical Latent Variables

Means
C#1 -0.943 0.247 -3.817 0.000
C#2 -0.158 0.151 -1.048 0.294
250   Data Analysis with Mplus

More readily interpretable is the output of the LCA results in terms of

the estimated probability parameters p̂ g and p̂ig . The estimates of the p̂ g
parameters were already output earlier under FINAL CLASS COUNTS
AND PROPORTIONS FOR THE LATENT CLASSES BASED ON
THE ESTIMATED MODEL. Estimates of the conditional response prob-
abilities p̂ig (as well as 1 – p̂ig ) are obtained under the rubric RESULTS
IN PROBABILITY SCALE. Under Category 1, Mplus outputs the
conditional response probabilities for the lower category (1 – p̂ig ), respec-
tively. In our case, the lower category is coded zero and labeled “never/
rarely.” Under Category 2, the conditional response probabilities p̂ig
for the higher category (in our case, the category that is coded 1: “often/
very often”) are output.

RESULTS IN PROBABILITY SCALE

Latent Class 1

C1
Category 1 0.802 0.080 9.968 0.000
Category 2 0.198 0.080 2.463 0.014
C2
Category 1 0.872 0.079 11.068 0.000
Category 2 0.128 0.079 1.623 0.105
C3
Category 1 0.842 0.056 14.991 0.000
Category 2 0.158 0.056 2.808 0.005
C4
Category 1 0.816 0.078 10.399 0.000
Category 2 0.184 0.078 2.345 0.019
C5
Category 1 0.273 0.081 3.383 0.001
Category 2 0.727 0.081 9.020 0.000
C6
Category 1 0.185 0.099 1.878 0.060
Category 2 0.815 0.099 8.247 0.000
C7
Category 1 0.547 0.088 6.179 0.000
Category 2 0.453 0.088 5.125 0.000
C8
Category 1 0.665 0.074 8.941 0.000
Category 2 0.335 0.074 4.508 0.000

Latent Class 2

C1
Category 1 0.635 0.033 19.021 0.000
Category 2 0.365 0.033 10.915 0.000
 Latent Class Analysis   251

C2
Category 1 0.356 0.044 8.152 0.000
Category 2 0.644 0.044 14.721 0.000
C3
Category 1 0.476 0.041 11.727 0.000
Category 2 0.524 0.041 12.916 0.000
C4
Category 1 0.614 0.034 18.144 0.000
Category 2 0.386 0.034 11.423 0.000
C5
Category 1 0.760 0.045 16.916 0.000
Category 2 0.240 0.045 5.332 0.000
C6
Category 1 0.824 0.042 19.466 0.000
Category 2 0.176 0.042 4.165 0.000
C7
Category 1 0.348 0.034 10.094 0.000
Category 2 0.652 0.034 18.917 0.000
C8
Category 1 0.360 0.037 9.847 0.000
Category 2 0.640 0.037 17.521 0.000

Latent Class 3

C1
Category 1 0.990 0.008 126.531 0.000
Category 2 0.010 0.008 1.295 0.195
C2
Category 1 0.987 0.010 96.062 0.000
Category 2 0.013 0.010 1.259 0.208
C3
Category 1 0.938 0.019 50.383 0.000
Category 2 0.062 0.019 3.309 0.001
C4
Category 1 0.990 0.009 106.680 0.000
Category 2 0.010 0.009 1.090 0.276
C5
Category 1 0.902 0.027 33.919 0.000
Category 2 0.098 0.027 3.699 0.000
C6
Category 1 0.897 0.034 26.596 0.000
Category 2 0.103 0.034 3.049 0.002
C7
Category 1 0.947 0.017 57.362 0.000
Category 2 0.053 0.017 3.242 0.001
C8
Category 1 0.945 0.021 44.386 0.000
Category 2 0.055 0.021 2.580 0.010
252   Data Analysis with Mplus

On the basis of the estimated conditional response probabilities, we

can see that children who were assigned to Class 1 have fairly low prob-
abilities to endorse the category “often/very often,” except for items C5
(logic games) and C6 (skill-­training games). Apparently, Class 1 repre-
sents the group of children who prefer logic and skill-­training games,
whereas they report playing other types of games rarely or never. As seen
above, approximately 17.4% of children belong to this latent class.
Children in Class 2 show a different pattern of conditional response
probabilities. They generally show a higher probability for the category
“often/very often” than do children in Class 1. Specifically, members of
Class 2 have high response probabilities for items C2 (action games),
C3 (sport games), C7 (simulation games), and C8 (driving simulator
games). Therefore, Class 2 represents a group of children who play
computer games more often in general, and in particular prefer action
and simulation games. As we have seen, based on the estimated class
proportion parameters, approximately 38.1% of children belong to this
group.
The largest class (Class 3, 44.6%) is characterized by a pattern of very
low probabilities for the category “often/very often” for all items. There-
fore, we can interpret Class 3 as a class of nonplayers, that is, as a group
of children who do not play any of the computer games listed in the ques-
tionnaire often or very often.
For a clear and easy interpretation, it is useful to visualize the con-
ditional response probabilities p̂ig in each class using a so-­called class
profile plot. In this way, it is easier to assess the meaning of the classes as
well as the quality of a class solution. For example, a good class solution
is typically characterized by classes that show mostly low or high con-
ditional response probabilities, so that the classes have an unequivocal
interpretation. We can view the line graph that we requested through the
plot option via the Mplus point-and-click menu using GRAPH → View
Graphs in the output. Under Select a plot to view, we choose Estimated
probabilities (see Figure 6.2).
Subsequently, one has to specify whether the probabilities for the
first or the second category should be shown in the graph (see Figure 6.3).
Although arbitrary for dichotomous variables, the plot is usually easier
to interpret if the probabilities for the higher category (p̂ig ; here, “often/
very often”) are plotted (because this category indicates agreement) rather
than the probabilities for the lower category (1 – p̂ig , which indicates dis-
agreement). One obtains the line graph shown in Figure 6.4, in which the
estimated class proportions are also given again in terms of percentages.
 Latent Class Analysis   253

FIGURE 6.2. First step in viewing the item profile plot in Mplus.

Given that there are limited options to edit the Mplus line graph (e.g.,
it does not contain labels for the y- and x-axes), it is useful for a first
inspection of the class solution, but not suitable for inclusion in publica-
tions, presentations, etc. It is therefore usually a good idea to plot the
relevant conditional response probabilities again in an external graph-
ics program that allows for more nicely designed graphs. As an example,
Figure 6.5 shows the solution in a Microsoft Excel® line graph. In this

FIGURE 6.3. Specification of the types of conditional response probabilities to be

shown in the profile plot. In this example, the probabilities for the second category
(“often/very often”) are selected.
254   Data Analysis with Mplus

FIGURE 6.4. Mplus profile plot showing the conditional response probabilities for
the second category of the computer game items in each of the three latent classes.
The items are on the x-axis, whereas the y-axis shows the conditional probability of
endorsing the category “often/very often.”

0.9
Conditional response probability

0.8
for category "often/very often"

0.7

0.6
Class 1 (44.6%)
0.5 Class 2 (38.1%)
Class 3 (17.4%)
0.4

0.3

0.2

0.1

FIGURE 6.5. Microsoft Excel line graph for the same data shown in Figure 6.4.
 Latent Class Analysis   255

example, the order of classes was changed from “largest first” to “smallest
last,” so that the smallest class in the Mplus solution is now Class 3 and
the largest class is now Class 1. In addition, the x-axis was labeled accord-
ing to the item content to facilitate the interpretation of the solution.
Using the class profile plot, we can once again see how the classes
differ from one another: The largest class (now Class 1) is the class of
nonplayers, in which members have very low probabilities for the cat-
egory “often/very often” across all items. The second largest class (Class
2) represents the group of children who play more often in general and
who have a specific preference for action, sport, and simulation games.
The smallest class (now Class 3) consists of logic and skill-­training game
players.
In addition to the class profile plot, we have also requested that the
individual class assignment probabilities and estimated class member-
ship based on the maximum class assignment probability be saved to an
external file. At the end of the Mplus output file, we obtain a list of the
old and new variables that were saved to the new file by Mplus as well as
information on the format of the variables in the new data file. This infor-
mation is important, because it tells us the order in which the variables
(columns) appear in the new data file, so that the data can be processed
correctly in further analyses in external statistic programs such as, for
example, SPSS. As an example, note that the variable ID (subject ID) now
is no longer the first variable in the data set, but is now the ninth vari-
able:

SAVEDATA INFORMATION

Order and format of variables

C1 F10.3
C2 F10.3
C3 F10.3
C4 F10.3
C5 F10.3
C6 F10.3
C7 F10.3
C8 F10.3
ID F10.3
CPROB1 F10.3
CPROB2 F10.3
CPROB3 F10.3
C F10.3
256   Data Analysis with Mplus

Save file
computergames_3classes.dat

Save file format

13F10.3

Save file record length 5000

The four new variables that Mplus added to the data file are
CPROB1, CPROB2, CPROB3, and C. The newly created data file is named
computergames_3classes.dat (the name was specified in the input file)
and can be used for further analyses. For instance, one could use the class
membership information saved by Mplus to find out about the prefer-
ences of specific children in the data set. To which computer game class
is a specific child most likely to belong? This and similar questions can
be answered based on the individual class assignment probabilities that
Mplus saved for each child. Figure 6.6 shows an excerpt from that part of
the newly created data file (after it was read into SPSS) that contains the
four newly generated variables CPROB1, CPROB2, CPROB3, and C. An
SPSS syntax file for reading the data set generated by Mplus can be found
on the companion website.
The variables CPROB1, CPROB2, and CPROB3 indicate the esti-
mated probability of class membership for each child for Classes 1, 2, and
3, respectively. For example, based on his or her response pattern, the
first child in the data set (first row) has a probability of .218 of belong-
ing to Class 1, .028 of belonging to Class 2, and .754 of belonging to
Class 3. Note that the sum of the probabilities is 1. This is because LCA
assumes that each child belongs to one (and only one) of the extracted
latent classes. Given that this child’s probability is largest for Class 3, the
child is assigned to this class and therefore obtains the value of 3 on the
variable C, which indicates the most likely latent class membership for
each child.
The second child (second row) has the highest assignment probabil-
ity for Class 1 (.988) and is therefore assigned to Class 1, etc. In this
way, one can (1) examine how reliably each individual can be classified
and (2) obtain an estimate of the most likely latent class membership for
each individual based on the estimated class assignment probabilities.
The average latent class assignment probabilities for the most likely latent
class membership that Mplus reports in the standard output are calcu-
lated based on these individual class assignment probabilities.
 Latent Class Analysis   257

FIGURE 6.6. Excerpt from the data set computergames_3classes.sav, which

contains the individual estimated class assignment probabilities (variables CPROB1–
CPROB3) and the most likely class membership (variable C).

6.3 Model Fit Assessment and Model Comparisons

Often, researchers do not have a strong theory that allows them to make
a firm assumption about the number of classes to extract. Given that the
number of classes is not a model parameter to be estimated in LCA, the
number of classes has to be determined indirectly through model tests
and/or model comparisons. In practice, this means that researchers usu-
ally estimate a series of LCA models with different numbers of classes
(e.g., the two- to five-class solutions) and then compare the absolute and
relative fit of these models. The model with the best model fit is usually
selected.
For the selection of the best fitting model, a number of different cri-
teria are available. Here I discuss only those criteria that are available
in Mplus. One can distinguish between indices for evaluating absolute
fit (how well does a model fit the observed data in absolute terms?) and
258   Data Analysis with Mplus

indices for evaluating relative fit (how well does a model fit compared to
other models?). More detailed information on indices for evaluating model
fit in LCA can be found in Collins, Fidler, Wugalter, and Long (1993),
Formann (2003), and Rost (2004).

6.3.1 Absolute Model Fit

6.3.1.1Likelihood Ratio and Pearson X2 Statistic
Absolute model fit refers to the degree to which a class solution can repro-
duce the observed response patterns. In Mplus, two different test statistics
are available for this purpose: the likelihood ratio (LR) test and the Pear-
son X2 test. Both statistics are based on the comparison of the observed
response pattern frequencies to the model-­implied response pattern fre-
quencies and are output by default when analyzing LCA. Under optimal
conditions (see discussion below), both statistics follow a chi-­square dis-
tribution under the null hypothesis that the model is true in the popu-
lation. Significant values indicate a statistically significant discrepancy
between the model and the data. In this case, the model does not perfectly
reproduce the observed data (e.g., because too few classes were extracted).
Unfortunately, the LR and Pearson statistics follow a chi-­square dis-
tribution only under fairly restrictive conditions. Specifically, relatively
large sample sizes in combination with a relatively small number of items
are needed to obtain reliable p-values from these statistics (Collins et al.,
1993; Langeheine, Pannekoek, & van de Pol, 1996; von Davier, 1997). In
practice, however, one often has to deal with a relatively large number of
items, but a rather small sample size, so that the asymptotic conditions
of the LR and Pearson statistics are not met. In these cases, one speaks of
“data sparseness” or “sparse tables.” One consequence of data sparseness
may be that the p-values (which are based on the theoretical chi-­square
distribution) obtained for the LR and Pearson statistics may not be reliable
(Rost, 2004).
As a rule of thumb, the p-values should not be trusted if there is a
large discrepancy between the values of the LR and the Pearson test statis-
tics for the same model. If the values are very different, this indicates that
at least one of the two statistics does not follow the theoretical chi-­square
distribution. Collins et al. (1993) as well as Langeheine et al. (1996) pro-
posed the use of a parametric bootstrap method in these cases, which can
be used to approximate the empirical distribution of the test statistic that
is valid for the application at hand.
 Latent Class Analysis   259

Unfortunately, as of version 6.1, the parametric bootstrap has been

implemented in Mplus only for relative model fit assessment (bootstrap
LR difference test; see Section 6.3.2.1), but not for absolute fit assessment.
Currently, a parametric bootstrap for global model fit assessment can be
obtained only indirectly in Mplus through a Monte Carlo simulation study
(see Muthén & Muthén, 2002, as well as Geiser et al., 2012). For our data
example, the Pearson X2 and LR statistics with conventional p-values are
as follows:

MODEL FIT INFORMATION

Chi-Square Test of Model Fit for the Binary and Ordered

Categorical (Ordinal) Outcomes

Pearson Chi-Square

Value 386.883
Degrees of Freedom 229
P-Value 0.0000

Likelihood Ratio Chi-Square

Value 340.605
Degrees of Freedom 229
P-Value 0.0000

The values of the test statistics are relatively similar (LR = 386.883;
Pearson X2 = 340.605), so that we can suspect that the asymptotic con-
ditions for the two tests are approximately met. As a consequence, the
p-values can be interpreted, albeit with caution.
Both statistics show significant p-values in our example. Therefore,
the three-class solution would have to be rejected based on these sta-
tistics. However, neither the LR nor the Pearson X2 value is extremely
large relative to their degrees of freedom (df = 229). This indicates that the
degree of misfit of the model is not extremely large. In addition, we are
using a relatively large sample size (N = 861), so that there is considerable
statistical power to detect even relatively minor misspecifications. Addi-
tional model fit indices as well as model comparisons should be used to
obtain a more precise picture of the fit of the model.

6.3.1.2 Residual Statistics (tech10 Output)

Residual statistics are useful to assess which response patterns contribute
particularly to the misfit of the model in terms of their contribution to the
260   Data Analysis with Mplus

Pearson X2 statistic and log likelihood value. Residual statistics also allow
us to detect “outliers,” that is, cases with unique or aberrant response
patterns. Residual statistics indicate the contribution of each observed
response pattern to the global Pearson X2 value and the log likelihood.
In addition, Mplus outputs standardized residuals (z-scores) for each pat-
tern. Large absolute z-values indicate response pattern frequencies that
are particularly under- or overestimated by the model. Residual statistics
are not part of the Mplus standard output, but can be requested using the
following line of code in the Mplus input file:

output: tech10;

To save time, we use the optseed function (see Box 6.6) to reestimate
the model with the additional request for the tech10 output. For this
purpose, we specify the seed that was associated with the best log likeli-
hood value in our previous analysis, so that Mplus automatically repro-
duces the best solution (without having to iterate through all 500 sets of
start values again):

analysis: optseed = 605358;

The complete input file can be found on the companion website. We now
obtain additional output (TECHNICAL 10 OUTPUT) under the esti-
mated model parameters. Mplus first provides us with a detailed overview
of the observed response patterns. Each observed response pattern across
the eight computer game items is listed and enumerated by Mplus. In our
example, a total of 163 (of the total of 28 = 256 possible) response pat-
terns were actually observed. For example, the pattern with the number 1
(00100001) refers to the pattern in which the children selected the second
category (“often/very often”) only for the third (C3) and last item (C8). In
pattern number 2 (00000000), the category “often/very often” was not
selected for any of the eight items, and so on.

TECHNICAL 10 OUTPUT

MODEL FIT INFORMATION FOR THE LATENT CLASS INDICATOR MODEL PART

RESPONSE PATTERNS

No. Pattern No. Pattern No. Pattern No. Pattern

1 00100001 2 00000000 3 11000101 4 00001100
5 00000100 6 00001000 7 00000001 8 00101000
9 00001110 10 10000001 11 00000010 12 00010000
13 10011111 14 01001100 15 01001010 16 11101011
 Latent Class Analysis   261

17 11001100 18 01011110 19 00011110 20 00001101

21 10011100 22 10001100 23 01001111 24 00101100
25 10111111 26 11100111 27 01000011 28 00010010
29 10000000 30 00011000 31 10001110 32 10101110
33 00001010 34 00011111 35 01000110 36 00100101
37 10010111 38 00011100 39 10010100 40 11011100
41 00100011 42 01000000 43 00000110 44 00000101
45 00001001 46 00001111 47 10000110 48 00100000
49 00010001 50 11111111 51 10111011 52 00100100
53 01000010 54 11000000 55 10101111 56 11010001
57 01100010 58 01110000 59 01100000 60 00111111
61 01011001 62 00110000 63 01110111 64 01001011
65 01100011 66 00101011 67 01010001 68 11100010
69 11000111 70 11000001 71 11100001 72 01110011
73 01100111 74 01101000 75 11000010 76 11111011
77 11010111 78 11110011 79 11100011 80 11101111
81 01100001 82 10010010 83 01101111 84 11101001
85 11001000 86 00101001 87 10110010 88 11110001
89 00011010 90 00000011 91 11101010 92 01010011
93 11110111 94 10001011 95 00000111 96 11000011
97 01010010 98 11010011 99 11001001 100 00100111
101 10010011 102 00110010 103 10001010 104 00011011
105 01010100 106 11110010 107 11111000 108 00101110
109 01101011 110 11011110 111 11001011 112 01100110
113 01000100 114 00001011 115 01100101 116 11011011
117 01010000 118 01001110 119 10100011 120 00010100
121 00101111 122 11010010 123 11011001 124 10000010
125 10110111 126 10010000 127 00110011 128 10010110
129 00110001 130 10100000 131 10000011 132 10011000
133 10111101 134 10100010 135 00101101 136 10001111
137 01011000 138 01110010 139 11011000 140 10100001
141 10111010 142 00010110 143 01000001 144 01101001
145 10000100 146 01010111 147 00010111 148 11110110
149 00100010 150 01111010 151 11001111 152 01011011
153 00101010 154 01011111 155 11011111 156 11010000
157 00111011 158 11011010 159 01110101 160 01110001
161 10010001 162 11010101 163 01001000

Subsequently, Mplus provides us with a list of the observed and

corresponding model-­implied response pattern frequencies for all 163
observed response patterns. In order to save space, only the first 44 pat-
terns are shown here (the full output is available on the companion web-
site). For example, the first pattern (00100001) was observed for 13 chil-
dren in our sample, but was expected for about only four to five children
under the three-class model. This discrepancy between the observed and
the model-­implied pattern frequencies results in a relatively large z-score
of 4.09 for pattern number 1. This shows that the observed frequency of
this response pattern is “too high” and thus contributes significantly to
the global misfit of the model.
Overall, the largest z-value in our example is observed for the response
pattern number 40 (11011100; z = 8.47). Apparently, children who show
this pattern are particularly hard to classify based on the three-class solu-
tion. In addition to the z-statistics, we also obtain the contributions of
262   Data Analysis with Mplus

each pattern to the overall Pearson X2 statistic and log likelihood value.
For our pattern number 1, we find relatively large contributions of 16.68
Pearson X2 statistic and 28.07 for the log likelihood. These values have
a similar interpretation as do modification indices in SEMs (cf. Box 3.7).
If, for example, pattern number 1 had not been observed, the overall
Pearson X2 value would decrease by 16.68. Residual statistics can provide
valuable information about outliers (e.g., “unscalable” cases) and other
peculiarities in the data. However, these statistics should never be used
to exclude cases with “inconvenient” response patterns in a purely data-­
driven way.

RESPONSE PATTERN FREQUENCIES AND CHI-SQUARE CONTRIBUTIONS

Response Frequency Standardized Chi-square Contribution

Pattern Observed Estimated Residual Pearson Loglikelihood
(z-score)
1 13.00 4.42 4.09 16.68 28.07
2 261.00 255.07 0.44 0.14 12.00
3 1.00 0.78 0.25 0.06 0.51
4 27.00 18.73 1.93 3.65 19.75
5 32.00 35.06 -0.53 0.27 -5.84
6 29.00 31.54 -0.46 0.20 -4.87
7 14.00 18.34 -1.02 1.03 -7.56
8 5.00 3.05 1.12 1.24 4.93
9 11.00 13.22 -0.62 0.37 -4.04
10 2.00 2.05 -0.03 0.00 -0.10
11 19.00 18.53 0.11 0.01 0.96
12 3.00 3.94 -0.48 0.23 -1.64
13 1.00 0.50 0.71 0.50 1.39
14 1.00 2.52 -0.96 0.91 -1.85
15 3.00 2.26 0.49 0.24 1.70
16 1.00 2.05 -0.73 0.54 -1.43
17 4.00 0.68 4.03 16.22 14.18
18 2.00 0.67 1.63 2.65 4.38
19 2.00 3.03 -0.59 0.35 -1.66
20 9.00 8.18 0.29 0.08 1.71
21 2.00 0.90 1.15 1.33 3.18
22 2.00 3.92 -0.97 0.94 -2.69
23 4.00 1.64 1.85 3.40 7.14
24 5.00 3.23 0.99 0.97 4.37
25 1.00 0.22 1.67 2.77 3.03
26 2.00 1.40 0.51 0.26 1.43
27 9.00 10.32 -0.41 0.17 -2.47
28 5.00 2.40 1.68 2.83 7.35
29 4.00 3.88 0.06 0.00 0.24
30 1.00 1.41 -0.35 0.12 -0.69
31 4.00 3.30 0.39 0.15 1.55
32 2.00 0.73 1.49 2.21 4.03
33 4.00 5.48 -0.63 0.40 -2.52
34 3.00 1.70 1.00 1.00 3.41
35 2.00 1.96 0.03 0.00 0.09
36 4.00 1.37 2.25 5.04 8.57
37 2.00 0.57 1.89 3.55 5.00
38 1.00 3.59 -1.37 1.87 -2.56
39 1.00 0.46 0.80 0.64 1.56
40 4.00 0.20 8.47 71.73 23.92
 Latent Class Analysis   263

41 8.00 6.40 0.64 0.40 3.57

42 7.00 6.57 0.17 0.03 0.89
43 9.00 7.12 0.71 0.49 4.21
44 8.00 5.25 1.20 1.44 6.74
. . . . . . . . . . . . . . . . . .

At the end of the residual statistics table, Mplus also provides the over-
all contribution to the global Pearson X2 value that stems from model-­
implied response patterns that were not observed in the actual data (so-­
called “empty cells”):

THE TOTAL PEARSON CHI-SQUARE CONTRIBUTION FROM EMPTY CELLS IS 77.39

In addition, tech10 includes univariate and bivariate model fit informa-

tion (not shown here) that can also be useful to examine model fit in LCA.

6.3.2 Relative Model Fit

Given the complications of evaluating absolute model fit using the Pear-
son X2 and LR statistic, due to the frequent violation of the asymptotic
conditions for those statistics (cf. Section 6.3.1.1), researchers often evalu-
ate model fit in LCA more in terms of the relative fit of a model compared
to competing models (often models with more or fewer classes than the
target model). Upon request, Mplus offers two different statistical tests
that can be used to directly compare the fit of models with different num-
bers of classes: the bootstrap LR difference test (tech14) and the Vuong–
Lo–­Mendell–­Rubin test (tech11). In addition, similar to other statisti-
cal models, LCA models can be compared using information criteria (in
Mplus: AIC, BIC, and sample-­size-­adjusted BIC).

6.3.2.1 Bootstrap LR Difference Test (tech14 Output)

Using the bootstrap LR difference test, one can compare a model with G
latent classes against a model with G – 1 classes. Hereby, the difference
in the LR values between the G and the G – 1 class model is calculated.
Using a parametric bootstrap procedure (e.g., Langeheine et al., 1996; von
Davier, 1997), approximate p-values for the LR difference are estimated
by fitting both the G and the G – 1 model to each bootstrap sample and
computing LR difference values for each bootstrap sample. This proce-
dure results in an empirical distribution of LR difference values under the
null hypothesis that the G – 1 class model is the true model. A significant
264   Data Analysis with Mplus

p-value indicates that the model with G classes fits the data better than
the more parsimonious model with one class less. For a nonsignificant
p-value, the more parsimonious model with one class less is preferred.
In our example, we first test the three-class model against a two-class
solution. The bootstrap LR difference test procedure is requested in Mplus
by adding the following command to the input file for the three-class
model in Figure 6.1:

output: tech14;

To select the desired number of bootstrap samples (e.g., 500), we addition-

ally specify:

analysis: lrtbootstrap = 500;

Five hundred bootstrap samples are often sufficient to obtain a reasonably

precise estimate of the approximate p-value for the LR difference test. To
ensure that no local maxima occur during the analysis of each parametric
bootstrap sample, we increase the number of random start value sets for
the estimation of both the G and the G – 1 class model to 50 in the first
and 20 in the second step of the optimization:

analysis: lrtstarts = 50 20 50 20;

For more complex models, it is highly recommended to further increase

the number of random starts used in the bootstrap analysis, even if this
may lead to a considerable increase in the required computing time. The
Mplus default (lrtstarts = 0 0 20 5) is often not sufficient to
obtain valid bootstrap results due to the problem of local maxima (cf. Box
6.4). We obtain the following additional output in Mplus:

TECHNICAL 14 OUTPUT

Random Starts Specifications for the k-1 Class Analysis Model

Number of initial stage random starts 10
Number of final stage optimizations 2

Random Starts Specification for the k-1 Class Model for Generated Data
Number of initial stage random starts 50
Number of final stage optimizations 20
Random Starts Specification for the k Class Model for Generated Data
Number of initial stage random starts 50
Number of final stage optimizations 20
Number of bootstrap draws requested 500
 Latent Class Analysis   265

PARAMETRIC BOOTSTRAPPED LIKELIHOOD RATIO TEST FOR 2 (H0) VERSUS 3 CLASSES

H0 Loglikelihood Value -3522.517

2 Times the Loglikelihood Difference 117.407
Difference in the Number of Parameters 9
Approximate P-Value 0.0000
Successful Bootstrap Draws 500

The log likelihood value for the G – 1 model (here, the two-class model
that is tested against the three-class model) is –3522.517 (it can be found
under H0 Loglikelihood Value). It is important to compare the log
likelihood value for the G – 1 model to the log likelihood value found in a
previous run of the G – 1 model that used a large number of start values
to make sure that the model estimated in the bootstrap procedure does
not represent a local maximum. The output for the separate analysis of the
two-class model is not shown here, but can be found on the companion
website. It turns out that the separate estimation of the two-class solution
with 500 sets of start values resulted in the same overall log likelihood
value, which indicates that we can trust the present bootstrap results.
Mplus also reports the LR difference value, which is equal to the dif-
ference in LR values between the two models. For the two-class model, the
LR value is 458.012 (see output on the companion website). For the three-
class model, we obtained an LR value of 340.605 (see Section 6.3.1.1).
Hence, the LR difference is 117.407. In the three-class model, nine addi-
tional parameters are estimated, compared to the two-class model (one
additional class size parameter and eight additional conditional response
probabilities). The LR difference of 117.407 is significant according to the
bootstrap p-value (Approximate P-Value), which is smaller than
.0001. Therefore, according to this criterion, the three-class model fits
significantly better than the two-class model.
For the comparison of three versus four classes, we additionally have
to estimate the four-class solution, including tech14 output (see com-
panion website for the detailed input specification). We obtain the follow-
ing result (only the most relevant output is shown):

PARAMETRIC BOOTSTRAPPED LIKELIHOOD RATIO TEST FOR 3 (H0) VERSUS 4 CLASSES

H0 Loglikelihood Value -3463.814

2 Times the Loglikelihood Difference 54.239
Difference in the Number of Parameters 9
Approximate P-Value 0.0000
Successful Bootstrap Draws 500

We can see that the loglikelihood value for the three-class model
(–3463.814) has again been replicated exactly, so that we can trust the
266   Data Analysis with Mplus

bootstrap results. It turns out that the difference in the LR values between
the three- and the four-class solution is also significant, LR D = 54.239, df
= 9, p < .0001. According to this criterion, we would prefer the four-class
model over the three-class solution. However, the decision as to the num-
ber of classes to retain should not be based only on statistical, but also on
substantive criteria and interpretability of a solution. In our example, it
turns out that the three-class solution is easier to interpret than the four-
class solution (see Section 6.3.3).

6.3.2.2 Vuong–Lo–Mendell–Rubin Test (tech11 Output)

The Vuong–Lo–­ Mendell–­Rubin (VLMR) test (Lo, Mendell, & Rubin,
2001) is based on a principle similar to the LR difference test. A significant
value of the VLMR test shows that the estimated model fits significantly
better than the model with one class less. Two versions of the VLMR test
are obtained through the command output: tech11; in Mplus. In
our example, the VLMR test provides the following result for two versus
three latent classes:

TECHNICAL 11 OUTPUT

Random Starts Specifications for the k-1 Class Analysis Model

Number of initial stage random starts 10
Number of final stage optimizations 2

VUONG-LO-MENDELL-RUBIN LIKELIHOOD RATIO TEST FOR 2 (H0) VERSUS 3 CLASSES

H0 Loglikelihood Value -3522.517

2 Times the Loglikelihood Difference 117.407
Difference in the Number of Parameters 9
Mean 19.334
Standard Deviation 20.559
P-Value 0.0035

LO-MENDELL-RUBIN ADJUSTED LRT TEST

Value 115.508
P-Value 0.0038

It is again important to check that the likelihood value listed under

HO Loglikelihood Value is compared to the log likelihood value
obtained for the G – 1 class model in a separate run with a sufficient num-
ber of sets of start values (here the values are identical, as desired). In this
case, like the bootstrap LR difference test, the VLMR test favors the three-
class model over the two-class model (p = .0035 for the VLMR test and
p = .0038 for the adjusted version). In contrast, the comparison of three
 Latent Class Analysis   267

versus four classes (using an extended input file for the four-class model
with tech11; see companion website) yields a result different from the
bootstrap LR difference test:

TECHNICAL 11 OUTPUT

VUONG-LO-MENDELL-RUBIN LIKELIHOOD RATIO TEST FOR 3 (H0) VERSUS 4 CLASSES

H0 Loglikelihood Value -3463.814

2 Times the Loglikelihood Difference 54.239
Difference in the Number of Parameters 9
Mean 18.229
Standard Deviation 28.354
P-Value 0.0896

LO-MENDELL-RUBIN ADJUSTED LRT TEST

Value 53.362
P-Value 0.0925

The p-values associated with both the VLMR test and its adjusted version
are nonsignificant (p = .0896 and p = .0925, respectively). This means
that according to these tests, the four-class model does not fit significantly
better than the more parsimonious three-class solution. (In contrast, the
bootstrap LR test found the four-class model to fit significantly better than
the three-class model.)
The results of a simulation study by Nylund, Asparouhov, and Muthén
(2007) indicate that, in general, the bootstrap LR test is a more accurate
indicator of the number of classes than the VLMR test. Therefore, the
bootstrap LR test should be preferred in practical applications.

6.3.2.3 Information Criteria

Information criteria (IC) are descriptive indices for model comparisons.
They take into account both the goodness of fit of a model to the data
and model parsimony (the number of estimated parameters). Models with
more classes tend to fit the data better than models with fewer classes. On
the other hand, with each additional class, the number of model param-
eters to be estimated increases (and parsimony decreases). In terms of
the IC, the best model is the one that fits well and uses as few parameters
as possible. In a series of models, the model with the smallest AIC, BIC,
or sample-­size-­adjusted BIC (aBIC) value is preferred. In our example,
we obtain the following IC values for the three-class model in the Mplus
default output:
268   Data Analysis with Mplus

MODEL FIT INFORMATION

Information Criteria

Number of Free Parameters 26

Akaike (AIC) 6979.628
Bayesian (BIC) 7103.338
Sample-Size Adjusted BIC 7020.769
(n* = (n + 2) / 24)

These values are not meaningful in an absolute sense. However, they can
be compared to the corresponding IC values calculated for other mod-
els. For example, we can compare them to the IC values obtained for the
two- and four-class solutions. The IC values for all three solutions are
shown in Table 6.1. It can be seen that the BIC index is smallest for three
classes, whereas AIC and aBIC favor four (or even more) classes. Based on
their simulation study, Nylund et al. (2007) recommend the BIC index for
determining the number of classes in LCA (in addition to the bootstrap
LR test; see Section 6.3.2.1).

6.3.3 Interpretability
For model fit evaluation and model comparisons in LCA, not only statis-
tical criteria play an important role. Equally important is the interpret-
ability of a solution (Collins & Lanza, 2010). This point can be illustrated
using the data sample discussed in this chapter. Even though the boot-
strap LR test indicated that the four-class solution fit the data better than
the three-class solution, close inspection of the latent class profiles in the
four-class solution reveals that one of the classes in the four-class solution

TABLE 6.1. IC for Different Class Solutions

for the Computer Game Example

Model AIC BIC aBIC

Two classes 7079.035 7159.923 7105.935

Three classes 6979.628 7103.338 7020.769
Four classes 6943.389 7109.922 6998.771

Note. AIC, Akaike’s information criterion; BIC, Bayesian information cri-

terion; aBIC, sample-­size-­adjusted BIC. Smallest values are printed in
boldface.
 Latent Class Analysis   269

BOX 6.7. Overview of the Most Important Criteria for Evaluating

Model Fit in LCA

Tests of Significance for Evaluating Absolute Fit. The LR and Pear-

son X2 statistics are reported by default in Mplus. A practical problem with
these statistics is that they follow the theoretical chi-­square distribution only
when the number of items is small and the sample size is large. A parametric
bootstrap can be used to generate an empirical test distribution that may
yield more reliable p-values in cases where data are sparse (in Mplus indi-
rectly available via a Monte Carlo simulation study).

Analysis of Model Residuals. Model residuals (output: tech10;)

can help a researcher to figure out why a model shows a bad overall fit. Spe-
cifically, a researcher can use residuals to find out which response patterns
contribute most to the global misfit. This can be helpful, for example, to
detect outliers or individuals with aberrant response patterns.

Statistical Model Comparisons. Information criteria (IC; e.g., AIC,

BIC, aBIC) can be used for descriptive model comparisons. The model with
the smallest IC value is selected. BIC tends to perform well for determining
the number of classes. Statistical tests (VLMR test → tech11 output and
bootstrap LR test → tech14 output) are also available in Mplus. The boot-
strap LR test is especially recommended.

Mean Class Assignment Probabilities. For a good class solution, the

mean class assignment probabilities should be equal to or larger than .8.

Entropy. Summary measure for the quality of the classification in an

LCA model. Values close to 1 indicate good classification accuracy, whereas
values close to 0 indicate lack of accuracy. The entropy index is part of the
Mplus default output.

Number of Classes. A parsimonious solution uses as few classes as

possible. For solutions with one or more very small classes, one might con-
sider selecting a model with fewer classes (unless the small classes are of
substantive interest).

Number of Boundary Parameter Estimates. If many conditional

response probabilities are estimated to exactly 0 or 1, this can be a sign for
the extraction of too many classes, a local maximum, an identification prob-
lem, or an otherwise unreliable solution. Mplus issues a warning message in
the output if boundary parameters occurred in an analysis.
270   Data Analysis with Mplus

Interpretation. Each latent class should have a clear interpretation. In

general, interpretation is easiest when all items show either high or low con-
ditional response probabilities within a class. Interpretable and theoretically
meaningful solutions should be preferred to uninterpretable solutions.

Replicability/Stability of a Solution. It is a good sign for the validity

and generalizability of the findings if an exploratory LCA solution can be
replicated with fresh data, preferably samples that show different character-
istics from the original sample (e.g., different age group or nationality).

is difficult to interpret, because this class shows primarily medium-­size

conditional response probabilities for the computer game items. There-
fore, it is not clear how one would interpret this class.
In general, an easy-to-­interpret class solution is typically character-
ized by the majority of conditional response probabilities being either
close to 1 or close to 0, whereas medium-­size conditional response proba-
bilities do not occur. Classes that show medium-­size conditional response
probabilities for many items are often difficult to interpret, because indi-
viduals in this class would have a “fifty-fifty” chance of endorsing the
higher category of an item. An exception are dichotomous test items, for
which the answer can be either right or wrong. For such items, a probabil-
ity of .5 corresponds to the probability of guessing an answer. Therefore,
classes with around .5 conditional response probabilities for some or all
items might be meaningfully interpreted as a class of individuals who use
a guessing strategy under certain conditions.
Another argument against the four-class solution in the present
example is that this solution contains a boundary estimate (one condi-
tional response probability is estimated to be exactly 0). Therefore, the
three-class solution appears to provide a more reasonable representation
of the data, because it is easier to interpret (and also more parsimonious).
Box 6.7 summarizes the most important criteria for evaluating model fit
in LCA.
Appendix A

Summary of Key Mplus Commands

Discussed in This Book

Meaning/ Chapters/
Command explanation Comments Section

title: Title of the Not a required 2

analysis command

data: file = Specification of the The complete path 2

<name_of_data_file. name and location does not have to be
dat >; of the file that specified if the data
contains the data set is located in the
to be analyzed same directory as
the input file

data: listwise Using listwise The Mplus default, 2

= on; deletion of missing as of version 5, is
values FIML estimation
with missing data

data: type = Specification Example 1 2

<type of summary of the kind of (means, standard
data >; summary data to deviations, and
be analyzed correlation matrix):
data: type
= means std
corr;

271
272  Appendix A

Meaning/ Chapters/
Command explanation Comments Section

Example 2
(covariance
matrix): data:
type = cova;

data: Specification of Not required when 2

nobservations = the sample size individual data are
<sample size >; (required when analyzed; in this
summary data book, I also use
are used in the the abbreviation
analysis) nobs instead of
nobservations

variable: names Specification of the Variable names 2

= <list of all variable variable names cannot be longer
names >; than eight
characters

variable: Specification of 6
categorical = variables in the
<list of ordered analysis that are
categorical variables ordered categorical
to be used in the (ordinal) or
analysis >; dichotomous

variable: Specification 6
nominal = <list of variables
of nominal variables as unordered
to be used in the categorical
analysis >; (nominal)

variable: Specification of Not required when 2

usevariables = all variables that all variables in the
<list of all variables are used in the data set are used
used in the statistical statistical model in the statistical
model >; model; in the
book, I mainly use
the abbreviation
usevar
 Appendix A  273

Meaning/ Chapters/
Command explanation Comments Section

variable: Specification of the missing = 2

missing = missing value code all(Missing-­
<variable name Value-Code);
(missing value defines the same
code) >; missing value code
for all variables

variable: Definition of Required when 5

cluster = <cluster a variable that nested data
variable >; indicates cluster (e.g., children
membership nested within
school classes)
are analyzed
using multilevel
modeling
techniques

variable: Definition of a 5
within = <name of variable that is
the level-1 variable(s) used only at level
in the analysis >; 1 of a multilevel
analysis

variable: Definition of a 5
between = variable that is
< name of the level-2 used only at level
variable(s) in the 2 of a multilevel
analysis>; analysis

variable: Centering of a 3; 5
centering = variable at the
grandmean × grand mean
(<variable name >);

variable: Centering of a 5
centering = variable at the
groupmean × cluster mean
(<variable name >); (group mean
centering)
274  Appendix A

Meaning/ Chapters/
Command explanation Comments Section

variable: Specification of the Used in LCA 6

classes = <name name of a latent and other types
of the latent class class variable of mixture
variable > (<number and the number distribution
of latent classes >); of classes to be analyses
extracted

variable: Specification of an Can be used to 6

auxiliary = auxiliary variable save additional
<name of the auxiliary (e.g., the subject ID variables to an
variable >; variable) external file
generated by
Mplus, even if
those variables are
not used in the
actual analysis

analysis: type Basic analysis Used to check the 2

= basic; data; provides
only descriptive
statistics

analysis: type Type of analysis Default type of 3; 4

= general; used for analysis in Mplus
conventional
regression, path,
and SEM analyses

analysis: type Two-level analysis Used in connection 5

= twolevel; with the
specification of a
cluster variable
to carry out a
multilevel analysis
with two levels

analysis: type Two-level analysis Used in multilevel 5.5

= twolevel with random analyses with
random; slopes random slopes
 Appendix A  275

Meaning/ Chapters/
Command explanation Comments Section

analysis: type Mixture Used to specify 6

= mixture; distribution an LCA or other
analysis type of mixture
distribution
analysis in
connection with
variable:
classes =
<name of the latent
class variable >
(<number of latent
classes >);

analysis: Specification of Can be used to Various

estimator = the method of select a different chapters
<type of estimator >; estimation to estimator rather
be used in the than the default
analysis (e.g., ml
for maximum
likelihood
estimation)

model: Specification of a 3–6

model

model: y on x; Regression of 3–5

variable y on
variable x

model: y with Covariance/ 3; 4

x; correlation
between y and x

model: f by y; Factor loading of The loading of the 3

an indicator y on a first indicator listed
latent factor f after by is fixed to
1 for identification
as the default; in
the example, the
loading of y on f
would be fixed to 1
276  Appendix A

Meaning/ Chapters/
Command explanation Comments Section

<parameter >* Setting a fixed Example: f 3

parameter free by y*; this
command sets the
loading of y on f
free

<parameter >* Assigning a Also sets a fixed 3

<starting value > user-­defined parameter free
starting value to a (and at the same
parameter time assigns a
starting value)

<parameter >@ Fixation of a Example: f by 3

<number > parameter to a y1*1 y2@1; this
specific value command sets
the loading of y1
on f free (and
assigns the starting
value of 1 to this
loading) and fixes
the loading of y2
on the same factor
to 1

<parameter >(label) Assigning a label The label can be a 4

to a parameter numeric or string;
this command is
often used to set
parameters equal
to other parameters
or to refer to
parameters in the
so-­called model
constraint command

Specification of Used in growth 4.5; 5.5

growth factors or curve modeling
random slopes and multilevel
modeling with
random slopes
 Appendix A  277

Meaning/ Chapters/
Command explanation Comments Section

model: %within% Specification of the 5

<model specification > level-1 part of a
two-level analysis

model: Specification of the 5

%between% level-2 part of a
<model specification > two-level analysis

<name of the random Specification of Used in multilevel 5.5

slope > | y on x; a random slope modeling with
for the regression random slopes
of y on x in a
multilevel model

model indirect: Request for a Relevant especially 3.5

summary of direct, in statistical
indirect, and total mediation analysis
effects in a path (path analysis with
model indirect effects)

model indirect: Output of all 3.5

y ind x; indirect effects
from x to y

output: Request for Various

additional output chapters
that is not
provided as the
default

output: Descriptive Various

sampstat; statistics chapters

output: Observed missing Not available 2

patterns; data patterns under data:
and missing data listwise =
pattern frequencies on;
(when FIML
estimation is used)
278  Appendix A

Meaning/ Chapters/
Command explanation Comments Section

output: All standardized Various

standardized; solutions chapters

output: stdyx; Only the fully Various

standardized chapters
solution

output: Confidence 3.5

cinterval; intervals for
all parameter
estimates

output: Model-­implied
residual; covariance and
mean structure
and residual
statistics

output: Model
modindices; modification
indices

plot: Output of graphs 4.5; 6

savedata: file Saves new or 6

= <name of the new modified data sets/
data set >; external files

savedata: save Specification of Can be used, for 6

= <information to be the information example, to request
saved >; to be saved to an that the individual
external file class assignment
probabilities be
saved as new
variables in LCA
Appendix B

Common Mistakes in the Mplus

Input Setup and Troubleshooting

Problem Possible cause Potential remedy

Error messages • Data set could not be • Check that the name
in the output; no found by Mplus. and location of the
model computed. data set were correctly
specified.
• Data set could not be • Check that the number
correctly read by Mplus. and order of the
variables were correctly
specified.
• Input line exceeds 80 or • Use return to break
90 characters. too-long input lines into
smaller lines.
• One or more semicolons • Insert a semicolon
(;) missing in the input (;) at the end of each
file. command.
• Incorrect specification of • Check whether the
the usevar command usevar command is
(e.g., too many or too specified correctly.
few variables listed
under usevar) or
usevar command not
specified in the input.

279
280  Appendix B

Problem Possible cause Potential remedy

• Variable names • Check whether the
incorrectly spelled (in model command is
the usevar or model specified correctly.
command).

Incorrect or • Missing values not • Properly code missing

implausible properly coded or values and specify
descriptive missing value code missing value code
statistics/model incorrectly specified. in the variable
results; estimation command using
or convergence the subcommand
problems, etc. missing =.
• Incorrect processing of • Check the data set; if
the data set by Mplus. necessary save new data
set from SPSS; run a
basic analysis to check
proper processing of the
data in Mplus.
• Order or number of the • Check the number
variables in the data and correct order of
set is different from the variables in the
what is specified under variable command.
variable: names.

The following error • Usevar command not • Add usevar command

message appears in specified (in this case unless all variables in
the output: At least Mplus assumes that all the data set should be
one variable is variables are used in the used in the model.
uncorrelated with model).
all other variables • Usevar command • List only those variables
in the model. incorrectly specified under usevar that
Check that this is (more variables listed are actually used in the
what is intended. than are actually used in analysis.
the model).
 Appendix B  281

Problem Possible cause Potential remedy

Data could not be • Data set contains • Make sure that decimal
read by Mplus or commas instead of dots points are indicated by
were incorrectly as decimal signs. dots.
read; for this • Data set contains string • Avoid string variables
reason no output variables that cannot be or transform them into
or incorrect/ processed by Mplus. numeric variables prior
implausible results. to saving a data set for
Mplus.
• Missing values have • Assign a missing value
not been coded with a code to all missing
missing value code. values. A numeric
missing value code (e.g.,
–99) seems to work best.
Can be easily done in
SPSS using the RECODE
command.
• Variable names appear • Delete variable names in
in the data set. the data file or save data
file again without adding
the variable names to
the file.
• Data were not saved in • Save data in tab-­
tab-­delimited format but delineated format.
in fixed ASCII format.
• Data set was not found • Make sure the data set is
because it is not in the located in the directory
same directory as the that you specified.
input file or because
an incorrect path was
specified under data.
• Order or number of • Make sure the order
variables under and number of variables
variable: names listed in the names
= does not match command are correct.
the order/number of
variables in the data set.
282  Appendix B

Problem Possible cause Potential remedy

• One or more input • Break input lines that
lines exceed 80 or 90 are too long using the
characters. return key.

Mplus does not • analysis: type = • Replace type =

estimate the model basic; is specified. basic; by type =
even though all Under the basic option, general; or another
input commands no actual model relevant type of analysis.
are correct and the parameters are estimated
data are correctly by Mplus.
processed by Mplus.
 Appendix C

Further Readings

A .pdf version of the Mplus User’s Guide can be downloaded for free from the
Mplus homepage (http://www.statmodel.com).
Muthén, L. K., & Muthén, B. O. (1998–2007). Mplus user’s guide, fifth edition. Los
Angeles: Muthén & Muthén. http://www.statmodel.com/ugexcerpts.shtml.

The following book by Barbara Byrne introduces basic and advanced structural
equation modeling with Mplus:
Byrne, B. M. (2011). Structural equation modeling with Mplus: Basic concepts, appli-
cations, and programming. New York: Routledge.

The following book deals with more complex analyses in Mplus:

Geiser, C., Crayen, C., & Enders, K. (2012). Advanced data analysis with Mplus.
Wiesbaden: VS Verlag für Sozialwissenschaften. (Manuscript in prepara-
tion)

Specific readings that deal with the specific statistical details of methods dis-
cussed in the book can be found at the beginning of each chapter.

283
References

Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting inter-
actions. Newbury Park, CA: Sage.
Baron, R. M., & Kenny, D. A. (1986). The moderator–­mediator variable distinc-
tion in social psychological research: Conceptual, strategic, and statistical
considerations. Journal of Personality and Social Psychology, 51, 1173–1182.
Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley.
Bollen, K. A., & Curran, P. J. (2006). Latent curve models: A structural equation
perspective. New York: Wiley.
Bollen, K. A., & Long, S. (Eds.). (1993). Testing structural equation models. New-
bury Park, CA: Sage.
Bryk, A. S., & Raudenbush, S. W. (1987). Application of hierarchical linear mod-
els to assessing change. Psychological Bulletin, 101, 147–158.
Byrne, B. M., Shavelson, R. J., & Muthén, B. O. (1989). Testing for the equiva-
lence of factor covariance and mean structures: The issue of partial mea-
surement invariance. Psychological Bulletin, 105, 456–466.
Chen, F., Bollen, K. A., Paxton, P., Curran, P., & Kirby, J. (2001). Improper solu-
tions in structural equation models: Causes, consequences, and strategies.
Sociological Methods and Research, 29, 468–508.
Clogg, C. C. (1995). Latent class models: Recent developments and prospects
for the future. In G. Arminger, C. C. Clogg, & M. E. Sobel (Eds.), Handbook
of statistical modeling in the social sciences (pp. 311–359). New York: Plenum
Press.
Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regres-
sion/correlation analysis for the behavioral sciences. Mahwah, NJ: Erlbaum.
Cole, D. A., Martin, J. M., & Powers, B. (1997). A competency-­based model of
child depression: A longitudinal study of peer, parent, teacher, and self-­
evaluations. Journal of Child Psychology and Psychiatry and Allied Disciplines,
38, 505–514.

285
286  References

Cole, D. A., Martin, J. M., Powers, B., & Truglio, R. (1996). Modeling causal
relations between academic and social competence and depression: A
multitrait–­multimethod longitudinal study of children. Journal of Abnormal
Psychology, 105, 258–270.
Cole, D. A., & Maxwell, S. E. (2003). Testing mediational models with longitu-
dinal data: Questions and tips in the use of structural equation modeling.
Journal of Abnormal Psychology, 112, 558–577.
Collins, L. M., Fidler, P. L., Wugalter, S. E., & Long, J. D. (1993). Goodness-­of-fit
testing for latent class models. Multivariate Behavioral Research, 28, 375–389.
Collins, L. M., & Lanza, S. T. (2010). Latent class and latent transition analysis with
applications in the social, behavioral, and health sciences. New York: Wiley.
Crayen, C. (2010). Chi-­square distributions calculator version 3 [Computer soft-
ware]. Berlin: Freie Universitat.
Duncan, T. E., Duncan, S. C., & Strycker, L. A. (2006). An introduction to latent
variable growth curve modeling: Concepts, issues, and applications (2nd ed.).
Mahwah, NJ: Erlbaum.
Eid, M. (2000). A multitrait–­multimethod model with minimal assumptions.
Psychometrika, 65, 241–261.
Eid, M., Courvoisier, D. S., & Lischetzke, T. (2011). Structural equation model-
ing of ambulatory assessment data. In M. R. Mehl & T. S. Connor (Eds.),
Handbook of research methods for studying daily life (pp. 384–406). New York:
Guilford Press.
Eid, M., Gollwitzer, M., & Schmitt, M. (2010). Forschungsmethoden und Statistik
[Research methods and statistics]. Weinheim, Germany: Beltz.
Eid, M., Langeheine, R., & Diener, E. (2003). Comparing typological structures
across cultures by multigroup latent class analysis. Journal of Cross-­Cultural
Psychology, 34, 195–210.
Eid, M., Lischetzke, T., & Nussbeck, F. W. (2006). Structural equation models
for multitrait–­multimethod data. In M. Eid & E. Diener (Eds.), Handbook
of multimethod measurement in psychology (pp. 283–299). Washington, DC:
American Psychological Association.
Eid, M., Nussbeck, F. W., Geiser, C., Cole, D. A., Gollwitzer, M., & Lischetzke,
T. (2008). Structural equation modeling of multitrait–­multimethod data:
Different models for different types of methods. Psychological Methods, 13,
230–253.
Eid, M., Schneider, C., & Schwenkmezger, P. (1999). Do you feel better or worse?:
The validity of perceived deviations of mood states from mood traits. Euro-
pean Journal of Personality, 13, 283–306.
Enders, C. K. (2010). Applied missing data analysis. New York: Guilford Press.
Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in cross-­
sectional multilevel models: A new look at an old issue. Psychological Meth-
ods, 12, 121–138.
Finney, S. J., & DiStefano, C. (2006). Non-­normal and categorical data in struc-
tural equation modeling. In G. R. Hancock & R. O. Mueller (Eds.), Structural
 References  287

equation modeling: A second course (pp. 269–314). Greenwich, CT: Informa-

tion Age.
Formann, A. K. (2003). Latent class model diagnostics: A review and some pro-
posals. Computational Statistics and Data Analysis, 41, 549–559.
Geiser, C. (2009). Multitrait–­multimethod–­multioccasion modeling. Munich: AVM.
Geiser, C., Crayen, C., & Enders, C. K. (2012). Datenanalyse mit Mplus für Fort-
geschrittene [Advanced data analysis with Mplus]. Wiesbaden, Germany: VS
Verlag für Sozialwissenschaften. (Manuscript in preparation)
Geiser, C., Eid, M., & Nussbeck, F. W. (2008). On the meaning of the latent vari-
ables in the CT-C(M–1) model: A comment on Maydeu-­Olivares & Coffman
(2006). Psychological Methods, 13, 49–57.
Geiser, C., Keller, B. T., & Lockhart, G. (in press). First- versus second-­order
latent growth curve models: Some insights from latent state–trait theory.
Structural Equation Modeling.
Geiser, C., Lehmann, W., & Eid, M. (2006). Separating “rotators” from “non-­
rotators” in the mental rotations test: A multigroup latent class analysis.
Multivariate Behavioral Research, 41, 261–293.
Geiser, C., & Lockhart, G. (2012). A comparison of four approaches to account
for method effects in latent state–trait analyses. Psychological Methods, 17,
255–283.
Goodman, L. A. (1974). Exploratory latent structure analysis using both identifi-
able and unidentifiable models. Biometrika, 61, 215–231.
Hagenaars, J. A. (1993). Loglinear models with latent variables. Beverly Hills, CA:
Sage.
Hagenaars, J. A., & McCutcheon, A. L. (Eds.). (2002). Applied latent class analysis.
Cambridge, UK: Cambridge University Press.
Hancock, G. R., Kuo, W., & Lawrence, F. R. (2001). An illustration of second-­
order latent growth models. Structural Equation Modeling, 8, 470–489.
Heller, K., Gaedicke, A.-K., & Weinläder, H. (1976). Kognitiver Fahigkeitstest (KFT
4–13) [Cognitive Ability Test (KFT 4-13)]. Weinheim, Germany: Beltz.
Hertzog, C., & Nesselroade, J. R. (1987). Beyond autoregressive models: Some
implications of the trait–state distinction for the structural modeling of
developmental change. Child Development, 58, 93–109.
Hofmann, D. A., & Gavin, M. B. (1998). Centering decisions in hierarchical
linear models: Implications for research in organizations. Journal of Manage-
ment, 24, 623–641.
Hox, J. J. (2002). Multilevel analysis: Techniques and applications. Mahwah, NJ: Erl-
baum.
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance struc-
ture analysis: Conventional criteria versus new alternatives. Structural Equa-
tion Modeling, 6, 1–55.
Jöreskog, K. G. (1979a). Statistical models and methods for analysis of longitu-
dinal data. In K. G. Jöreskog & D. Sörbom (Eds.), Advances in factor analysis
and structural equation models (pp. 129–169). Cambridge, MA: Abt.
288  References

Jöreskog, K. G. (1979b). Statistical estimation of structural models in longitudinal–­

developmental investigations. In J. R. Nesselroade & P. B. Baltes (Eds.), Lon-
gitudinal research in the study of behavior and development (pp. 303–351). New
York: Academic Press.
Kaplan, D. (2009). Structural equation modeling: Foundations and extensions (2nd
ed.). Newbury Park, CA: Sage.
Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.).
New York: Guilford Press.
Kovacs, M. (1985). The Children’s Depression Inventory (CDI). Psychopharmacol-
ogy Bulletin, 21, 995–998.
Kreft, I. G. G., & de Leeuw, J. (1998). Introducing multilevel modeling. London: Sage.
Kreft, I. G. G., de Leeuw, J., & Aiken, L. S. (1995). The effect of different forms
of centering in hierarchical linear models. Multivariate Behavioral Research,
30, 1–21.
Lance, C. E., Noble, C. L., & Scullen, S. E. (2002). A critique of the correlated
trait–­correlated method and correlated uniqueness models for multitrait–­
multimethod data. Psychological Methods, 7, 228–244.
Langeheine, R., Pannekoek, J., & van de Pol, F. (1996). Bootstrapping goodness-­
of-fit measures in categorical data analysis. Sociological Methods and Research,
24, 249–264.
Langeheine, R., & Rost, J. (Eds.). (1988). Latent trait and latent class models. New
York: Plenum.
Lazarsfeld, P. F., & Henry, N. W. (1968). Latent structure analysis. Boston: Hough-
ton Mifflin.
Lo, Y., Mendell, N. R., & Rubin, D. B. (2001). Testing the number of components
in a normal mixture. Biometrika, 88, 767–778.
Loehlin, J. C. (1998). Latent variable models: An introduction to factor, path, and
structural analysis. Mahwah, NJ: Erlbaum.
Luke, D. A. (2004). Multilevel modeling. Thousand Oaks, CA: Sage.
MacKinnon, D. P. (2008). Introduction to statistical mediation analysis. Mahwah,
NJ: Erlbaum.
MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V.
(2002). A comparison of methods to test mediation and other intervening
variable effects. Psychological Methods, 7, 83–104.
MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence limits
for the indirect effect: Distribution of the product and resampling methods.
Multivariate Behavioral Research, 39, 99–128.
Marsh, H. W., Wen, Z., & Hau, K. T. (2006). Structural equation models of
latent interaction and quadratic effects. In G. Hancock & R. Mueller (Eds.),
Structural equation modeling: A second course (pp. 225–265). Greenwich, CT:
Information Age.
Mayer, A., Steyer, R., & Mueller, H. (in press). A general approach to defining
latent growth components. Structural Equation Modeling.
McArdle, J. J. (1988). Dynamic but structural equation modeling of repeated measures
 References  289

data. In R. B. Cattell & J. Nesselroade (Eds.), Handbook of multivariate experi-

mental psychology (pp. 561–614). New York: Plenum Press.
McArdle, J. J., & Hamagami, F. (2001). Latent difference score structural mod-
els for linear dynamic analysis with incomplete longitudinal data. In L. M.
Collins & A. G. Sayer (Eds.), New methods for the analysis of change (pp. 137–
175). Washington, DC: American Psychological Association.
McCutcheon, A. L. (1987). Latent class analysis. Beverly Hills, CA: Sage.
Meredith, W. (1993). Measurement invariance, factor analysis and factorial
invariance. Psychometrika, 58, 525–543.
Meredith, W., & Horn, J. (2001). The role of factorial invariance in modeling
growth and change. In L. M. Collins & A. G. Sayer (Eds.), New methods for
the analysis of change (pp. 203–240). Washington, DC: American Psychologi-
cal Association.
Meredith, W., & Tisak, J. (1984, June). On “Tuckerizing” curves. Paper presented
at the annual meeting of the Psychometric Society, Santa Barbara, CA.
Meredith, W., & Tisak, J. (1990). Latent curve analysis. Psychometrika, 55, 107–
122.
Millsap, R. E., & Meredith, W. (2007). Factorial invariance: Historical perspec-
tives and new problems. In R. Cudeck & R. MacCallum (Eds.), Factor analy-
sis at 100 (pp. 131–152). Mahwah, NJ: Erlbaum.
Mulaik, S. A., & Millsap, R. E. (2000). Doing the four-step right. Structural Equa-
tion Modeling, 7, 36–74.
Muthén, B. O. (2002). Beyond SEM: General latent variable modeling. Behavior-
metrika, 29, 81–117.
Muthén, L. K., & Muthén, B. O. (2002). How to use a Monte Carlo study
to decide on sample size and power. Structural Equation Modeling, 9, 599–
620.
Muthén, L. K., & Muthén, B. O. (2010). Mplus 6 [Computer software]. Los Ange-
les: Muthén & Muthén. http://statmodel.com.
Muthén, L. K., & Muthén, B. O. (1998–2012). Mplus user’s guide, sixth edition. Los
Angeles: Muthén & Muthén. http://statmodel.com/ugexcerpts.shtml.
Nylund, K., Asparouhov, T., & Muthén, B. O. (2007). Deciding on the number
of classes in latent class analysis and growth mixture modeling: A Monte
Carlo simulation study. Structural Equation Modeling, 14, 535–569.
Pinquart, M. (2001). Correlates of subjective health in older adults: A meta-­
analysis. Psychology and Aging, 16, 414–426.
Quaiser-­Pohl, C., Geiser, C., & Lehmann, W. (2006). The relationship between
computer-­game preference, gender, and mental rotation ability. Personality
and Individual Differences, 40, 609–619.
Radloff, L. S. (1977). The CES-D Scale: A self-­report depression scale for research
in the general population. Applied Psychological Measurement, 1, 385–401.
Raffalovich, L. E., & Bohrnstedt, G. W. (1987). Common, specific, and error
variance components of factor models: Estimation with longitudinal data.
Sociological Methods and Research, 15, 385–405.
290  References

Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models (2nd ed.).
Thousand Oaks, CA: Sage.
Raykov, T. (1993). On estimating true change interrelationships with other vari-
ables. Quality and Quantity, 27, 353–370.
Raykov, T., & Marcoulides, G. A. (2006). A first course in structural equation model-
ing (2nd ed.). Mahwah, NJ: Erlbaum.
Reuter, T., Ziegelmann, J. P., Wiedemann, A. U., Geiser, C., Lippke, S., Schüz,
B., et al. (2010). Changes in intentions, planning, and self-­efficacy predict
changes in behaviors: An application of latent true change modeling. Jour-
nal of Health Psychology, 15, 935–947.
Roosa, M. W., Liu, F., Torres, M., Gonzales, N., Knight, G., & Saenz, D. (2008).
Sampling and recruitment in studies of cultural influences on adjustment:
A case study with Mexican Americans. Journal of Family Psychology, 22,
293–302.
Rost, J. (2004). Lehrbuch Testtheorie–Teskonstruktion (2nd ed.) [Textbook test the-
ory and test construction]. Bern, Germany: Huber.
Rost, J. (2006). Latent-Class-Analyse [Latent class analysis]. In F. Petermann &
M. Eid (Eds.), Handbuch de Psycologischen Dianostik [Handbook of Psycho-
logical Assessment] (pp. 275–287). Göttingen, Germany: Hogrefe.
Rost, J., & Langeheine, R. (Eds.). (1997). Applications of latent trait and latent class
models in the social sciences. Munster, Germany: Waxmann.
Sayer, A. G., & Cumsille, P. E. (2001). Second-order latent growth models. In
L. M. Collins & A. G. Sayer (Eds.), New methods for the analysis of change
(pp. 177–200). Washington, DC: American Psychological Association.
Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the
art. Psychological Methods, 7, 147–177.
Schermelleh-­Engel, K., Moosbrugger, H., & Müller, H. (2003). Evaluating the fit
of structural equation models: Test of significance and descriptive goodness-­
of-fit measures. Methods of Psychological Research—­Online, 8, 23–74. http://
www.dgps.de/fachgruppen/methoden/mpr-­online.
Schumacker, R. E., & Lomax, R. G. (1996). A beginner’s guide to structural equation
modeling. Mahwah, NJ: Erlbaum.
Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling
change and event occurrence. New York: Oxford University Press.
Snijders, T. A. B., & Bosker, R. J. (1999). Multilevel analysis: An introduction to basic
and advanced multilevel modeling. London: Sage.
Sobel, M. E. (1982). Asymptotic confidence intervals for indirect effects in struc-
tural equation models. In S. Leinhardt (Ed.), Sociological methodology 1982
(pp. 290–312). Washington, DC: American Sociological Association.
Sörbom, D. (1975). Detection of correlated errors in longitudinal data. British
Journal of Mathematical and Statistical Psychology, 28, 138–151.
Steyer, R. (1988). Experiment, Regression und Kausalitat: Die logische Struktur kau-
saler Regressionsmodelle [Experiment, regression, and causality: On the logi-
cal structure of causal regression models]. Unpublished habitation thesis,
University of Trier, Trier, Germany.
 References  291

Steyer, R., Eid, M., & Schwenkmezger, P. (1997). Modeling true intraindividual
change: True change as a latent variable. Methods of Psychological Research—­
Online, 2, 21–33. http://www.dgps.de/fachgruppen/methoden/mpr-­online.
Steyer, R., Ferring, D., & Schmitt, M. J. (1992). States and traits in psychological
assessment. European Journal of Psychological Assessment, 8, 79–98.
Steyer, R., Partchev, I., & Shanahan, M. (2000). Modeling true intra-­individual
change in structural equation models: The case of poverty and children’s
psychosocial adjustment. In T. D. Little, K. U. Schnabel, & J. Baumert
(Eds.), Modeling longitudinal and multiple-­group data: Practical issues, applied
approaches, and specific examples (pp. 109–126). Hillsdale, NJ: Erlbaum.
Steyer, R., Schmitt, M. J., & Eid, M. (1999). Latent state–trait theory and research
in personality and individual differences. European Journal of Personality, 13,
389–408.
Uebersax, J. (2000). A brief study of local maximum solutions in latent class analysis.
Available at http://www.johnuebersax.com/stat/local.htm.
von Davier, M. (1997). Bootstrapping goodness-­of-fit statistics for sparse categor-
ical data: Results of a Monte Carlo study. Methods of Psychological Research—­
Online, 2, 29–48.
Ware, J. E., & Sherbourne, C. D. (1992). The MOS 36-item short-form health
survey (SF-36): I. Conceptual framework and item selection. Medical Care,
30, 473–483.
Whitelaw, N. A., & Liang, J. (1991). The structure of the OARS physical health
measures. Medical Care, 29, 332–347.
Widaman, K. F., & Reise, S. P. (1997). Exploring the measurement invariance
of psychological instruments: Applications in the substance use domain. In
K. J. Bryant, M. Windle, & S. G. West (Eds.), The science of prevention: Meth-
odological advances from alcohol and substance abuse research (pp. 281–324).
Washington, DC: American Psychological Association.
Author Index

Aiken, L. S., 28, 38, 64, 212 DiStefano, C., viii

Asparouhov, T., 267 Duncan, S. C., 26
Duncan, T. E., 26, 82, 163, 181
Eid, M., 26, 81, 90, 91, 95, 118, 235
B Enders, C. K., 2, 15, 19, 28, 212, 214

Baron, R. M., 63, 64

Bentler, P. M., 45 F
Bohrnstedt, G. W., 85
Bollen, K. A., 26, 27, 35, 42, 59, 67, 82, 89, Ferring, D., 26
163, 181 Fidler, P. L., 258
Bosker, R. J., 196, 197 Finney, S. J., viii
Bryk, A. S., 195, 197 Formann, A. K., 258
Byrne, B. M., 100

G
C
Gaedicke, A.-K., 2
Chen, F., 59, 61 Gavin, M. B., 212
Clogg, C. C., 234 Geiser, C., 2, 26, 28, 32, 89, 91, 95, 166, 184,
Cohen, J., 28, 38, 64, 195, 196, 197, 216, 220 191, 231, 235, 236, 259
Cohen, P., 28 Goodman, L. A., 232
Cole, D. A., 88, 128 Graham, J. W., 19, 28
Collins, L. M., 234, 258, 268
Courvoisier, D. S., 81
Crayen, C., 45, 103 H
Cumsille, P. E., 184
Curran, P. J, 26, 59, 82, 163, 181 Hagenaars, J. A., 234–235
Hamagami, F., 145
Hancock, G. R., 184
D Hau, K. T., 64
Heller, K., 2
de Leeuw, J., 197, 212 Henry, N. W., 232
Diener, E., 235 Hertzog, C., 82, 126, 134

293
294  Author Index

Hoffman, J. M., 65 McCutcheon, A.L., 235

Hofmann, D. A., 212 Mendell, N. R., 266
Horn, J., 100, 184 Meredith, W., 100, 163, 184
Hox, J. J., 197 Millsap, R. E., 30, 184
Hu, L., 45 Moosbrugger, H., 35
Mueller, H., 181
Mulaik, S. A., 30
J Müller, H., 35
Muthén, B. O., vii, 1, 32, 100, 235, 249, 259,
Jöreskog, K. G., 82, 126, 134 267
Muthén, L. K., 1, 32, 235, 249, 259

K
N
Kaplan, D., 27
Keller, B.T., 166 Nesselroade, J. R., 82, 126, 134
Kenny, D. A., 63, 64 Noble, C. L., 88
Kirby, J., 59 Nussbeck, F. W., 26, 95
Kline, R. B., 27 Nylund, K., 267, 268
Kovacs, M., 128
Kreft, I. G. G., 197, 212
Kuo, W., 184 P

Pannekoek, J., 258

L Partchev, I., 82
Paxton, P., 59
Lance, C. E., 88 Pinquart, M., 63
Langeheine, R., 235, 258, 263 Powers, B., 128
Lanza, S. T., 234, 268
Lawrence, F. R., 184
Lazarsfeld, P. F., 232 Q
Lehmann, W., 235, 236
Liang, J., 62, 63, 73, 74 Quaiser-Pohl, C., 236
Lischetzke, T., 26, 81
Lo, Y., 266
Lockhart, G., 89, 166 R
Lockwood, C. M., 65
Loehlin, J. C., 27 Radloff, L. S., 83
Lomax, R. G, 27 Raffalovich, L. E., 85
Long, J. D., 258 Raudenbush, S. W., 195, 197
Long, S., 35 Raykov, T., 27, 82, 145
Luke, D. A., 197, 205, 209, 214, 215, 220, Reise, S. P., 100
221 Reuter, T., 163
Roosa, M. W., 83
Rost, J., 235, 248, 258
M Rubin, D. B., 266

MacKinnon, D. P., 63, 64, 65, 71, 72

Marcoulides, G. A., 27 S
Marsh, H. W., 64
Martin, J. M., 128 Sayer, A. G., 184
Maxwell, S. E., 88 Schafer, J. L., 19, 28
Mayer, A., 181 Schermelleh-Engel, K., 35, 45
McArdle, J. J., 82, 145, 184 Schmitt, M. J., 26
 Author Index  295

Schneider, C., 90 U
Schumacker, R. E., 27
Schwenkmezger, P., 26, 90 Uebersax, J., 238, 240
Scullen, S. E., 88
Shanahan, M., 82
Shavelson, R. J., 100 V
Sheets, V., 65
Sherbourne, C. D., 63 van de Pol, F., 258
Singer, J. D., 195, 197 von Davier, M., 258, 263
Snijders, T. A. B., 196, 197
Sobel, M. E., 70
Sörbom, D., 88 W
Steyer, R., 26, 81, 82, 85, 87, 116, 145, 148,
150, 159, 181 Ware, J. E., 63
Strycker, L. A., 26 Weinlader, H., 2
Wen, Z., 64
West, S. G., 28, 38, 64, 65
T Whitelaw, N. A., 62, 63, 73, 74
Widaman, K. F., 100
Tisak, J., 163 Willett, J. B, 195, 197
Tofighi, D., 212, 214 Williams, J, 65
Truglio, R., 128 Wugalter, S. E., 258
Subject Index

Note. Page numbers followed by b, t, and f indicate boxes, tables, and figures.
Commands and subcommands appear in Courier font.

Absolute model fit, 258–268, 269b manifest (See Manifest autoregressive

AIC (Akaike information criterion). See also model)
model fit information under specific model fit information, 129, 130f
models path diagram, 127f
Chi-Square Difference Calculator, 103, 105b single-indicator, 127, 127f
in one- and three-factor models, 55, 55t
in output, 46b
analysis: bootstrap =, 72 B
analysis: convergence =, 240b
analysis: estimator =, 241, 275 basic analysis
analysis: estimator = ml;, 208b, analysis: type = basic;, 15,
241 31–32, 274, 282
analysis: opseed =, 245b, 260–261 input file, 28
analysis: 1rtstarts =, 264 output, 14–20, 20f, 21f
analysis: starts =, 238, 240b purpose of, 10
analysis: stiterations =, 238–239, syntax commands, 10, 11f
240b twolevel, 198, 200–205, 201f, 208b
analysis: 1trbootstrap =, 264 Bayesian information criterion. See BIC
analysis: type = basic;, 15, 31, between = stype;, 216–217, 216f
274, 282 Bias-corrected bootstrap methods, 71–73
analysis: type = general;, 274, 282 BIC (Bayesian information criterion), 46b,
analysis: type = mixture;, 235, 104b–105b, 263, 267, 268, 268t, 269b
237, 275 Bootstrap LR difference test (tech14), 259,
analysis: type = twolevel;, 263–266
197–198, 207, 231, 274 Bootstrap methods, 65b, 71–73
analysis: type = twolevel Boundary estimates, 240b
random;, 223–224, 274 by statements, 41
Analysis of model residuals, 269b
ASCII data file, exporting from SPSS to Mplus,
7, 7f–8f C
Autoregressive models, 126–145
cross-lagged effects, 126–127 categorical =, 32, 236–237
goals of, 126–127 Categorical Latent Variables-
latent (See Latent autoregressive model) Means, 249

296
 Subject Index  297

CDC (Chi-Square Difference Calculator), 103, first-order, 51–52, 52f–54f, 54–58, 55t
104b–105b for longitudinal data, 82 (See also Latent
Center for Epidemiological Studies Depression state model)
Scale (CES-D), 83b second-order, 58–60, 59f–60f, 61b
Centering Convergence criterion, 240b
definition of, 210 Correlated uniqueness model, 88
grand mean, 211 Correlation matrix, 21–23, 22f, 23f
group mean, 211–212 Correlations
of independent variables, 38b between residual variables, 133
of predictor variables, 210 between state factors, 95
variable: centering =, 38b, 211 between state variables, 84
CES-D (Center for Epidemiological Studies between two factors, setting to zero, 52,
Depression Scale), 83b 54
CFA. See Confirmatory factor analysis Covariance matrix, 9, 21–23, 22f, 23f
CFI. See Comparative Fit Index Covariance residuals, 46b
Change models Covariances
functional form over time (See Latent coverage of data, 17–18
growth curve models) standardized, 55
goal of, 81 between two factors, setting to zero, 52,
LAM (See Latent autoregressive model) 54
latent (See Latent change model (latent Covariance structure models, 24
difference model)) Cross-lagged effects, 126–127, 133, 134,
Chi-Square Difference Calculator (CDC), 103, 144
104b–105b Cross-level interactions, 197
Chi-square difference test
automated using CDC software, 103,
104b–105b D
description of, 45b
Chi-square test data: file =, 271
description of, 45b data: listwise = on, 271
for independence model, 35 data: nobservations =, 272
latent path analysis, 76 data: type =, 271–272
for latent regression model, 44 data command, 12
latent state model, 87 Data set
manifest path model, 67 sample, 2
classes =, 237 unreadable in Mplus, 281–282
Class profile plot, 252–253, 253f .dat file format, 1, 7f
cluster = class, 200 Deletion, 2
Clustered data, 195–196 Dependent (endogenous) variables, 24, 26
Coding missing values, 2, 3f–6f, 4 Descriptive statistics
Coefficient of determination, 37 implausible or incorrect, 280
Command lines, 14b SPSS, 19, 21f
Comments, 12 Directional relationships (regressions), in path
Comparative Fit Index (CFI) diagrams, 26
description of, 45b–46b
latent state model, 88
one-factor and three-factor models, 55, E
55t
Conditional response or conditional solution EFA (exploratory factor analysis), vs. latent
probability, 234 class analysis, 233
Configural invariance, 100, 100b Empty cells, 263
Confirmatory factor analysis (CFA models) Empty model, 209. See also Null model
definition of, 25–26 (intercept-only model)
example of, 51b Endogenous (dependent) variables, 24, 26
298  Subject Index

Entropy, 269b G
Equality constraint, 99–100
Equality of latent means across time, testing General factor (g-factor model), 51, 52f–53f
for, 114b General factor model (g-factor model), 51,
Error messages, 279–280 52f–53f
missing semicolons, 12 G–1 model, 265
related to actual estimation of model, 34 Goodness of fit. See also model fit information,
WARNING in Model command, under specific models
30–31 in structural equation modeling, 45b–47b
warnings (See Warning messages) Grand mean centering, 38, 211
Error variables (residual variables), 27 Green font, for comments, 12
Exclamation mark, 14b Group mean centering, 211–212
Exogenous variables, 24, 26
Exploratory factor analysis (EFA), vs. latent
class analysis, 233 H
Exporting
coding missing values, 2, 3f–6f, 4 Heywood case, 59–61, 61b
Mplus graph, 179, 180f Hierarchical linear models. See Mplus,
SPSS data to Mplus, 1–2, 7, 7f–8f multilevel models
Hierarchical structure, of clustered data,
195–196
F

Factorial invariance I
strict, 100b, 114–115, 115f
strong, 100, 100b, 109–114 ICC (interclass correlation coefficient), 198,
weak, 100b, 102 200, 201f, 202–205
Factor loadings, 25 Implementation of constraints, checking, 105b
File formats, 1–2, 7f Importing, into Mplus, 7, 8f, 10
FIML estimation, 15–16, 19, 20f Incremental fit indices, 45b
First-order latent growth curve models, Independence model (null model), 23, 35, 45b
164–183 Independent variables, 38b, 196
example of, 165 Indicator-specific effects, 88–89
external variables, influence of, 167 Indicator-specific factor or method, 89–92,
intercept factor mean, 166–167 89f–93f, 102
linear Indicator specificity, 85–86
graphical analysis of growth curves, 174, Indirect effects, 63, 65b, 70–71
175f–179f, 176–177, 179 Individual data format, 9–10
intercept-only model, 170, 170f Information Criteria (IC), 46b, 267–268. See
model fit information, 170–171 also AIC; BIC
MODEL RESULTS, 171–172 *.inp, 14
path diagram, 164–166, 165f INPUT READING TERMINATED
plot option, 169–170, 171, 174 NORMALLY, 32–33
slope factor, 166 Intercept-only model (null model), 23,
specification, 167–168, 168f, 169b, 170f 205–209, 208b
STANDARDIZED MODEL RESULTS, Intercepts, 230
172–174 Intercepts-and-slopes-as-outcomes model,
quadratic, 179, 180f–181f, 181–183 221, 224–231
exporting Mplus graph, 179, 180f cross-level interaction effect, 228–229, 229f
MODEL RESULTS, 182–183 output, 228–231, 229f
path diagram, 180f parameters, 225–226
specification, 181–182, 181f specification of, 226–228, 227f
vs. second-order LGCMs, 164, 184 Interclass correlation coefficient (ICC), 198,
Formative measurement models, 25 200, 201f, 202–205
 Subject Index  299

Intervening variables (mediator variables), 24 for dichotomous variable, 234, 235b

Item parcels, 63b example of, 236b
goals of, 233b
interpretability, 268, 268t, 269b
J log likelihood values, 242–245
model fit assessment, 257–268
Just identified models (Saturated path models), absolute, 258–263
67, 68b evaluation criteria, 269b
relative, 263–268
output file, 254–256, 255f, 257f
K parameters, 237–239
specification, 235–237, 237f
KFT scales, 51 vs. exploratory factor analysis, 233
Latent difference factor, positive signs of,
157
L Latent difference variables, 145–147, 151
Latent factors. See Latent variables
LAM. See Latent autoregressive model Latent growth curve models (LGCMs),
Latent autoregressive model (LAM), 133–145 163–194
advantages of, 144–145 autoregressive models and, 126
examples of, 134b first-order (See First-order latent growth
input file, 135–136, 136f curve models)
LS model and, 134–135 measurement invariance, 183–184
measurement invariance over time, 145 research questions, 163–164
model fit information, 136–137 second-order, 183–194
MODEL RESULTS, 137–142 specification, 197–198
Mplus model command, 144f vs. LC models, 163, 164
path diagrams, 134–135, 135f, 143f Latent mean structures, in latent state-trait
STANDARDIZED MODEL RESULTS, models, 120b–121b
139–142 Latent moderator models, 64b
Latent change model (latent difference model), Latent path analysis, 26, 74–80
145–163 input file, 74–76, 75f
advantages of, 163 output file, 76–80
baseline vs. neighbor change versions, path diagram, 73, 74f
148–149, 149f STANDARDIZED MODEL RESULTS,
data example in Mplus, 148b 76–78
examples of, 147b Latent regression model, 39–51
MODEL RESULTS, 153–157, 159–162 chi-square value, 44
parameterization, 150–151 example of, 40b
path diagrams, 146f, 149f input file, 43f
specification model fit information, 43–44
of baseline change, 149–150, 150f model results, 44, 47–48
of neighbor change model, 157–158, path analysis, 41f
158f path diagram, 25f, 26
strong factorial invariance, 147, 152 R-SQUARE-Latent Variable, 50–51
time-invariant loadings, 146–148, 146f specification, 41–42, 41f
vs. latent growth curve models, 163, 164 standardized model results, 48–50
Latent class analysis (LCA), 232–270 Latent state model (LS model), 81–116
applications, 235 analysis in Mplus, 85–88, 86f, 87b
avoiding local likelihood maxima in, 240b with autocorrelated error variables, 88–89,
boundary parameter estimates, 241b 89f–90f
class profile plot, 252–254, 253f, 254f characteristics of, 82
conditional response or conditional solution default for mean structure, 87b
probability, 234 example of, 83b
300  Subject Index

Latent state model (cont.) measurement model parameters, 27b

with indicator-specific factor or method, 89, path diagrams, 25f, 26
91–92, 92f simple linear regression analysis with
MODEL RESULTS, 95–96 manifest variables, 28–37, 28b, 29f,
STANDARDIZED MODEL RESULTS, 30b, 31f, 38b, 39
94, 96–98 STANDARDIZED MODEL RESULTS,
limitations of, 85, 116 37, 39
for longitudional data analysis, 85 structural model parameters, 27b
as LST precursors, 82 SUMMARY OF ANALYSIS, 33
model fit information, 87–88, 90, 110, 114, Listwise deletion, 2
114b Listwise = on;, 19
Mplus analysis, 85–88, 86f, 87b Local likelihood maximum, 238
path analysis, 83f Lo-Mendell-Rubin Test. See Vuong-Lo-
specificiation, 86–87, 86f, 87–88 Mendell-Rubin Test
structural model, saturation of, 84 Loglikelihood value, 265
variables, correlation of, 83f, 84 LS factors, 99, 101
vs. latent state-trait models, 85 LS model. See Latent state model
Latent state-trait model (LST model), 81, 82, LS variables, 99, 109
116–126
applications, 125–126
latent mean structures in, 120b–121b M
model fit statistics, 121–122
MODEL RESULTS, 122–125 Macro level, 195
path diagram, 117–118, 117f–118f Manifest autoregressive model, 127–129
specification of, 119, 121 examples of, 128b
vs. latent state models, 85, 116 model fit statistics, 129
Latent structure analysis. See Latent class MODEL RESULTS, 131–133
analysis Mplus input files, 128–129, 128f
Latent variable means, 109. See also Latent second-order autoregressive effects, 129,
mean structures 130f
Latent variables (latent factors), 25–27, 40 situation-specific effects, 129
autoregressive models with (See Latent Manifest path analysis, 62, 62f, 65–73, 66f, 68b
autoregressive model) Manifest path model, 80
path analysis (See Latent path analysis) input file, 65–66, 66f
LGCMs. See Latent growth curve models output file, 66–67
Likelihood ratio (LR) statistic, 258–259, 260, 262 STANDARDIZED MODEL RESULTS,
Linear factor analysis (CFA), 51–61 67–68
Linear regression analysis, simple Manifest variables, 25
example of, 28b path analysis (See Manifest path analysis)
with manifest variables, 28–37, 28b, 29f, residual variances of, 41, 43
30b, 31f, 38b, 39 Marker variables, 77, 92
first model specification, 30b Maximum likelihood (ML) estimation,
SUMMARY OF ANALYSIS, 33 237–238
Linear regression model Mean class assignment probabilities, 269b
bivariate Mean residuals, 46b
estimated, path diagram of, 29f Means-as-outcomes model, 209, 214–219
specification of, 30, 31f input file, 216–217, 216f
Linear structural equation models (linear output file, 217–219, 218f
SEMs), 24–80 parameter estimates, 215–216
applications, 26, 81 Mean structure, Mplus default for, 87b.See also
latent regression analysis, 39–51 Latent mean structures, Latent variable
with latent variables, 24–25, 25f (See also means
Covariance structure models) Measurement equivalence (measurement
advantages of, 26 invariance), 99, 100b
 Subject Index  301

Measurement model Multilevel regression analysis, 195–231

in linear structural equation models, 24–25, 25f advantages of, 197b
parameters of, 27b example of, 198b
Mediated effects, 63, 64, 64b formal notation, 199b
Mediator variables (intervening variables), 24, interclass correlation coefficient, 198, 200,
63–64 201f, 202–205
Metric, of latent variable or factor, 42 random intercept and slope models (See
Micro level, 195 Random intercept and slope models)
missing;, 12 Multimethod data, 26
MISSING DATA PATTERN Multivariate, 197–198
FREQUENCIES, 16–17
Missing values, 15
Mixed models. See Mplus, multilevel models N
Mixture distribution, 237
model:, 32, 41, 102–103, 103f, 129, 130f, 275 <name of the random slope>|y on
model: %between%, 223, 228, 277 x:, 277
model: f by y;, 275 names =, 12, 31
model: %within%, 223, 227–228, 231, 277 Nested data structures (clustered data
model: y on x;, 275 structures), 195
model: y with x;, 275 Nested models, 45b, 103
<model specification>, 277 Nondirectional relationships. See also
MODEL ESTIMATION TERMINATED Correlations; Covariances
NORMALLY, 34–35 in path diagrams, 26
Model fit information, 35, 67. See also under Nonsaturated path models, vs. saturated path
specific models models, 68b
model indirect, 66, 66f, 80 nobservations= *;, 22–23
model indirect:, 66, 277 Null model (intercept-only model), 23,
model indirect: y ind x;, 277 205–209, 208b
Model Modification indices, 47b Number of latent classes, 269b
Model results, 36–37. See also under specific
models
estimation or convergence problems, 280 O
latent path analysis, 76–78
manifest path model, 67 OLS regression, 220–221, 220f, 229, 229f
Moderated regression, 64b One-factor models
Moderator variable, 64b fit statistics, 54–55, 55t
Mplus. See also subcommands; specific commands general factor or g-factor model, 51, 52f–53f
default for mean structure, 87b Mplus input files, 52, 53f
editor, opening, 10 RMSEA value, 54–55, 55t
importing into, 7, 8f, 10 One-way random effects ANCOVA, 209–214,
input files (See also under specific models) 212f
saving under new name, 28, 29f input file, 212–213, 212f
language generator, 10 intercept and slope differences and, 221
multilevel models, 197–198 MODEL RESULTS, 213–214
output files (See under specific models) STANDARDIZED MODEL RESULTS, 214
specification of models, 195–198 (See also vs. random coefficient regression model, 221
under specific models) optseed option, 245b
specifying correlated error variables, 89–90 *.out, 14
syntax output:, 32, 277
basic structure, 10–12, 11f output: cinterval;, 71–72, 278
to specify correlated error variables, 90, 91f output: modindices;, 47b, 278
to specify LS model with indicator- output: patterns;, 278
specific factor for second indicator, output: residual;, 47b, 278
92–93, 93f output: sampstat;, 23, 23f, 32, 277
302  Subject Index

output: sampstat stdyx;, 32 Random intercept and slope models, 220–231

output: standardized;, 278 intercept-only model (See Null model)
output: stdyx;, 278 intercepts-and-slopes-as-outcomes model,
output: tech4, 54 221, 224–231
output: tech10;, 260–263, 269b means-as-outcomes model, 214–219
output: tech11;, 266 null model (See Null model)
output: tech14;, 263–268 one-way random effects ANCOVA, 209–214,
212f
random coefficient regression analysis,
P 221–224, 222f
Random slope, 167–168
<parameter>, 276 Reference indicators, 91, 94–95
<parameter>(label), 276 Reflective measurement model, 25
<parameter>@ <number>, 276 Regression analysis, 62
<parameter>* <starting value>, Regression analysis with manifest variables,
276 40
Partial measurement invariance, 100b Regression coefficients, 151
Path analysis Relative model fit, 263–268
example of, 63b Residuals, 46b–47b
latent, 62, 73, 74f Residual statistics, 259–263
latent state model, 83f Residual variances, 60–61
LS model with autocorrelated error Root Mean Square Error of
variables, 89f–90f Approximation (RMSEA), 46b, 93.
manifest, 62 See also model fit information under
Path coefficients, 128 specific models
Path diagrams, 25f, 26–27 latent state model, 88
of estimated bivariate regresson model, for one- and three-factor models, 54–55,
29f 55t
latent path analysis, 73, 74f
latent state-trait model, 117f
second-order factor model, 59f S
Pearson X 2 statistic, 258–259, 260, 262, 263
plot;, 193, 194f, 239, 278 Sample-size adjusted BIC (aBIC), 46b
p-values, 196, 208, 258–259, 264 Sample statistics, 23, 33–34, 37
Saturated models, 35
Saturated path models (just identified models),
Q 67, 68b
savedata: file =, 278
Quadratic growth factor, 181 savedata: save =, 278
.sav file format, 1–2
Second-order factor, 58
R Second-order latent growth curve model,
183–194
Random coefficient models. See Mplus, advantages of, 184
multilevel models example of, 184, 185f
Random coefficient regression model, 205, mean model-implied growth curve, 193,
220–221, 220f 194f
parameter estimates, 221–222 MODEL RESULTS, 188–190
specification, input file for, 222–223, 222f Mplus input file, 186, 187f, 188
vs. one-way random effects ANCOVA, path diagram, 184, 185f
221 STANDARDIZED MODEL RESULTS,
Random effects analysis of covariance model, 190–193
209 vs. first-order LGCMs, 164, 184
 Subject Index  303

Semicolons, 12, 14b Syntax rules, 14b

SEMs. See Structural equation models System-defined missing values, 2
Significance testing, of indirect effects, 65b
Sobel test, 70–71
SPSS T
ASCII data file, exporting to Mplus, 7,
7f–8f Tab-delimited format, 1, 8f
coding missing values, 2, 3f–6f, 4 tech11 output, 266–267
descriptive statistics, 19, 21f Tests of model fit. See Model fit information
options, 12, 13f Text file format, 1
preparing raw data set for Mplus, 1–2 Three-factor models, 52, 53f–54f
syntax, 11, 13f fit statistics, 54–55, 55t
SRMR (Standardized Root Mean Square MODEL RESULTS, 55–56
Residual), 46b, 93. See also model fit Mplus input files, 52, 54f
information under specific models STANDARDIZED MODEL RESULTS, 56–58
latent state model, 88 title:, 11–12, 271
for one- and three-factor models, 55, 55t TLI (Tucker-Lewis Index), 46b, 55, 55t
Standard errors, estimation under type = Trait component, 82
twolevel, 208b Tucker-Lewis Index (TLI), 46b, 55, 55t
Standardized residuals, 46b–47b twolevel basic analysis, 198, 200–205,
Standardized Root Mean Square Residual. See 201f
SRMR .txt file format, 1
State factor loading, 101–102 type= means std corr;, 22
State residual component, 82 type = twolevel, 198, 200–205, 201f, 208b
State variables, correlations between, 84
Statistical model comparisons, 269b
Strong factorial invariance, latent change U
models and, 147
Structural equation analyses, 9 Unconditional cell means model. See Null
Structural equation models (SEMs) model (intercept-only model)
goodness of fit criteria, 45b–47b Unconstrained model. See Null model
inadmissible parameter estimates or (intercept-only model)
Heywood cases, 59–61, 61b User-defined missing values, 2, 3f–6f, 4
with latent variables, advantages of, 40 usevar=, 30
for longitudinal data, 99 usevariables, 41
for measuring change (See Change
models)
for measuring variability (See Variability V
models)
Structural (latent variable) model Variability models
in linear structural equation models, 24–25, analysis, goals of, 82
25f definition of, 81
parameters of, 27b LS model (See Latent state model)
Subcommands. See also specific subcommands LST (See Latent state-trait model)
semicolons for, 12 variable: auxiliary =, 235–236, 274
Summary data, importing and analyzing, variable: between =, 273
21–23, 22f, 23f variable: categorical =, 272
Summary data format, 9 variable: centering =, 38b, 211
SUMMARY OF DATA, 16. See also under variable: centering = grandmean
specific models x, 273
SUMMARY OF MISSING DATA variable: centering = groupmean
PATTERNS, 16–17. See also under x, 273
specific models variable: classes =, 274
304  Subject Index

variable: cluster =, 273 W

variable: missing =, 273
variable: names =, 13f, 272 Warning messages
variable: nominal =, 272 latent class analysis, 244
variable: usevariables =, 272 for missing values, 15
variable: within =, 273 for nonadmissible parameter estimate,
Variable command, 12 59–60
Variables null model, 207–208
interactions between, 64b
listed under names =, 31
listed under usevar, 30 Z
Vuong–Lo–Mendell–Rubin Test (VLMR),
266–267 z-statistics, 261–262
About the Author

Christian Geiser, PhD, is Assistant Professor in the Department of

­Psychology at Utah State University in Logan. His methodological research
focuses on the development, evaluation, and application of latent variable
psychometric models for longitudinal and multimethod data. In his sub-
stantive research, he focuses on individual differences in spatial abilities
and how they can be explained.

305