Statistical Analysis With Latent Variables

User’s Guide

Linda K. Muthén
Bengt O. Muthén
 Following is the correct citation for this document:

Muthén, L.K. and Muthén, B.O. (1998-2012). Mplus User’s Guide. Seventh Edition.
Los Angeles, CA: Muthén & Muthén

Copyright © 1998-2012 Muthén & Muthén

Program Copyright © 1998-2012 Muthén & Muthén
Version 7
November 2012

The development of this software has been funded in whole or in part with Federal funds
from the National Institute on Alcohol Abuse and Alcoholism, National Institutes of
Health, under Contract No. N44AA52008 and Contract No. N44AA92009.

Muthén & Muthén

3463 Stoner Avenue
Los Angeles, CA 90066
Tel: (310) 391-9971
Fax: (310) 391-8971
Web: www.StatModel.com
Support@StatModel.com
 TABLE OF CONTENTS
Chapter 1: Introduction 1

Chapter 2: Getting started with Mplus 13

Chapter 3: Regression and path analysis 19

Chapter 4: Exploratory factor analysis 43

Chapter 5: Confirmatory factor analysis and structural equation modeling 55

Chapter 6: Growth modeling and survival analysis 111

Chapter 7: Mixture modeling with cross-sectional data 153

Chapter 8: Mixture modeling with longitudinal data 209

Chapter 9: Multilevel modeling with complex survey data 251

Chapter 10: Multilevel mixture modeling 339

Chapter 11: Missing data modeling and Bayesian analysis 387

Chapter 12: Monte Carlo simulation studies 409

Chapter 13: Special features 443

Chapter 14: Special modeling issues 459

Chapter 15: TITLE, DATA, VARIABLE, and DEFINE commands 503

Chapter 16: ANALYSIS command 587

Chapter 17: MODEL command 641

Chapter 18: OUTPUT, SAVEDATA, and PLOT commands 713

Chapter 19: MONTECARLO command 775

Chapter 20: A summary of the Mplus language 805

 PREFACE
We started to develop Mplus seventeen years ago with the goal of providing researchers
with powerful new statistical modeling techniques. We saw a wide gap between new
statistical methods presented in the statistical literature and the statistical methods used
by researchers in substantively-oriented papers. Our goal was to help bridge this gap
with easy-to-use but powerful software. Version 1 of Mplus was released in November
1998; Version 2 was released in February 2001; Version 3 was released in March 2004;
Version 4 was released in February 2006; Version 5 was released in November 2007, and
Version 6 was released in April 2010. We are now proud to present the new and unique
features of Version 7. With Version 7, we have gone a considerable way toward
accomplishing our goal, and we plan to continue to pursue it in the future.

The new features that have been added between Version 6 and Version 7 would never
have been accomplished without two very important team members, Tihomir
Asparouhov and Thuy Nguyen. It may be hard to believe that the Mplus team has only
two programmers, but these two programmers are extraordinary. Tihomir has developed
and programmed sophisticated statistical algorithms to make the new modeling possible.
Without his ingenuity, they would not exist. His deep insights into complex modeling
issues and statistical theory are invaluable. Thuy has developed the post-processing
graphics module, the Mplus editor and language generator, and the Mplus Diagrammer
based on a framework designed by Delian Asparouhov. In addition, Thuy has
programmed the Mplus language and is responsible for keeping control of the entire code
which has grown enormously. Her unwavering consistency, logic, and steady and calm
approach to problems keep everyone on target. We feel fortunate to work with such a
talented team. Not only are they extremely bright, but they are also hard-working, loyal,
and always striving for excellence. Mplus Version 7 would not have been possible
without them.

Another important team member is Michelle Conn. Michelle was with us at the
beginning when she was instrumental in setting up the Mplus office and has been
managing the office for the past ten years. In addition, Michelle is responsible for
creating the pictures of the models in the example chapters of the Mplus User’s Guide.
She has patiently and quickly changed them time and time again as we have repeatedly
changed our minds. She is also responsible for keeping the website updated and
interacting with customers. She was the driving force behind the design of the new
shopping cart. With the vastly increased customer base, her efficiency in multi-tasking
and calm under pressure are much appreciated. Sarah Hastings recently joined the
Mplus team. She is responsible for testing the Graphics Module and the Mplus
Diagrammer in addition to providing assistance to Bengt. She has proven to be a
valuable team member.

We would also like to thank all of the people who have contributed to the development of
Mplus in past years. These include Stephen Du Toit, Shyan Lam, Damir Spisic, Kerby
Shedden, and John Molitor.

Initial work on Mplus was supported by SBIR contracts and grants from NIAAA that we
acknowledge gratefully. We thank Bridget Grant for her encouragement in this work.

Linda K. Muthén
Bengt O. Muthén
Los Angeles, California
September 2012
 Introduction

CHAPTER 1
INTRODUCTION

Mplus is a statistical modeling program that provides researchers with a

flexible tool to analyze their data. Mplus offers researchers a wide
choice of models, estimators, and algorithms in a program that has an
easy-to-use interface and graphical displays of data and analysis results.
Mplus allows the analysis of both cross-sectional and longitudinal data,
single-level and multilevel data, data that come from different
populations with either observed or unobserved heterogeneity, and data
that contain missing values. Analyses can be carried out for observed
variables that are continuous, censored, binary, ordered categorical
(ordinal), unordered categorical (nominal), counts, or combinations of
these variable types. In addition, Mplus has extensive capabilities for
Monte Carlo simulation studies, where data can be generated and
analyzed according to most of the models included in the program.

The Mplus modeling framework draws on the unifying theme of latent

variables. The generality of the Mplus modeling framework comes from
the unique use of both continuous and categorical latent variables.
Continuous latent variables are used to represent factors corresponding
to unobserved constructs, random effects corresponding to individual
differences in development, random effects corresponding to variation in
coefficients across groups in hierarchical data, frailties corresponding to
unobserved heterogeneity in survival time, liabilities corresponding to
genetic susceptibility to disease, and latent response variable values
corresponding to missing data. Categorical latent variables are used to
represent latent classes corresponding to homogeneous groups of
individuals, latent trajectory classes corresponding to types of
development in unobserved populations, mixture components
corresponding to finite mixtures of unobserved populations, and latent
response variable categories corresponding to missing data.

THE Mplus MODELING FRAMEWORK

The purpose of modeling data is to describe the structure of data in a
simple way so that it is understandable and interpretable. Essentially,
the modeling of data amounts to specifying a set of relationships

1
CHAPTER 1

between variables. The figure below shows the types of relationships

that can be modeled in Mplus. The rectangles represent observed
variables. Observed variables can be outcome variables or background
variables. Background variables are referred to as x; continuous and
censored outcome variables are referred to as y; and binary, ordered
categorical (ordinal), unordered categorical (nominal), and count
outcome variables are referred to as u. The circles represent latent
variables. Both continuous and categorical latent variables are allowed.
Continuous latent variables are referred to as f. Categorical latent
variables are referred to as c.

The arrows in the figure represent regression relationships between

variables. Regressions relationships that are allowed but not specifically
shown in the figure include regressions among observed outcome
variables, among continuous latent variables, and among categorical
latent variables. For continuous outcome variables, linear regression
models are used. For censored outcome variables, censored (tobit)
regression models are used, with or without inflation at the censoring
point. For binary and ordered categorical outcomes, probit or logistic
regressions models are used. For unordered categorical outcomes,
multinomial logistic regression models are used. For count outcomes,
Poisson and negative binomial regression models are used, with or
without inflation at the zero point.

2
 Introduction

A
f y

c u
B

Within
Between

Models in Mplus can include continuous latent variables, categorical

latent variables, or a combination of continuous and categorical latent
variables. In the figure above, Ellipse A describes models with only
continuous latent variables. Ellipse B describes models with only
categorical latent variables. The full modeling framework describes
models with a combination of continuous and categorical latent
variables. The Within and Between parts of the figure above indicate
that multilevel models that describe individual-level (within) and cluster-
level (between) variation can be estimated using Mplus.

MODELING WITH CONTINUOUS LATENT

VARIABLES
Ellipse A describes models with only continuous latent variables.
Following are models in Ellipse A that can be estimated using Mplus:

3
CHAPTER 1

• Regression analysis
• Path analysis
• Exploratory factor analysis
• Confirmatory factor analysis
• Item response theory modeling
• Structural equation modeling
• Growth modeling
• Discrete-time survival analysis
• Continuous-time survival analysis

Observed outcome variables can be continuous, censored, binary,

ordered categorical (ordinal), unordered categorical (nominal), counts,
or combinations of these variable types.

Special features available with the above models for all observed
outcome variables types are:

• Single or multiple group analysis

• Missing data under MCAR, MAR, and NMAR and with multiple
imputation
• Complex survey data features including stratification, clustering,
unequal probabilities of selection (sampling weights), subpopulation
analysis, replicate weights, and finite population correction
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcomes types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities
• Factor scores and plausible values for latent variables

MODELING WITH CATEGORICAL LATENT

VARIABLES
Ellipse B describes models with only categorical latent variables.
Following are models in Ellipse B that can be estimated using Mplus:

4
 Introduction

• Regression mixture modeling

• Path analysis mixture modeling
• Latent class analysis
• Latent class analysis with covariates and direct effects
• Confirmatory latent class analysis
• Latent class analysis with multiple categorical latent variables
• Loglinear modeling
• Non-parametric modeling of latent variable distributions
• Multiple group analysis
• Finite mixture modeling
• Complier Average Causal Effect (CACE) modeling
• Latent transition analysis and hidden Markov modeling including
mixtures and covariates
• Latent class growth analysis
• Discrete-time survival mixture analysis
• Continuous-time survival mixture analysis

Observed outcome variables can be continuous, censored, binary,

ordered categorical (ordinal), unordered categorical (nominal), counts,
or combinations of these variable types. Most of the special features
listed above are available for models with categorical latent variables.
The following special features are also available.

• Analysis with between-level categorical latent variables

• Tests to identify possible covariates not included in the analysis that
influence the categorical latent variables
• Tests of equality of means across latent classes on variables not
included in the analysis
• Plausible values for latent classes

MODELING WITH BOTH CONTINUOUS AND

CATEGORICAL LATENT VARIABLES
The full modeling framework includes models with a combination of
continuous and categorical latent variables. Observed outcome variables
can be continuous, censored, binary, ordered categorical (ordinal),
unordered categorical (nominal), counts, or combinations of these
variable types. Most of the special features listed above are available for
models with both continuous and categorical latent variables. Following

5
CHAPTER 1

are models in the full modeling framework that can be estimated using
Mplus:

• Latent class analysis with random effects

• Factor mixture modeling
• Structural equation mixture modeling
• Growth mixture modeling with latent trajectory classes
• Discrete-time survival mixture analysis
• Continuous-time survival mixture analysis

Most of the special features listed above are available for models with
both continuous and categorical latent variables. The following special
features are also available.

• Analysis with between-level categorical latent variables

• Tests to identify possible covariates not included in the analysis that
influence the categorical latent variables
• Tests of equality of means across latent classes on variables not
included in the analysis

MODELING WITH COMPLEX SURVEY DATA

There are two approaches to the analysis of complex survey data in
Mplus. One approach is to compute standard errors and a chi-square test
of model fit taking into account stratification, non-independence of
observations due to cluster sampling, and/or unequal probability of
selection. Subpopulation analysis, replicate weights, and finite
population correction are also available. With sampling weights,
parameters are estimated by maximizing a weighted loglikelihood
function. Standard error computations use a sandwich estimator. For
this approach, observed outcome variables can be continuous, censored,
binary, ordered categorical (ordinal), unordered categorical (nominal),
counts, or combinations of these variable types.

A second approach is to specify a model for each level of the multilevel

data thereby modeling the non-independence of observations due to
cluster sampling. This is commonly referred to as multilevel modeling.
The use of sampling weights in the estimation of parameters, standard
errors, and the chi-square test of model fit is allowed. Both individual-
level and cluster-level weights can be used. With sampling weights,

6
 Introduction

parameters are estimated by maximizing a weighted loglikelihood

function. Standard error computations use a sandwich estimator. For
this approach, observed outcome variables can be continuous, censored,
binary, ordered categorical (ordinal), unordered categorical (nominal),
counts, or combinations of these variable types.

The multilevel extension of the full modeling framework allows random

intercepts and random slopes that vary across clusters in hierarchical
data. Random slopes include the special case of random factor loadings.
These random effects can be specified for any of the relationships of the
full Mplus model for both independent and dependent variables and both
observed and latent variables. Random effects representing across-
cluster variation in intercepts and slopes or individual differences in
growth can be combined with factors measured by multiple indicators on
both the individual and cluster levels. In line with SEM, regressions
among random effects, among factors, and between random effects and
factors are allowed.

The two approaches described above can be combined. In addition to

specifying a model for each level of the multilevel data thereby
modeling the non-independence of observations due to cluster sampling,
standard errors and a chi-square test of model fit are computed taking
into account stratification, non-independence of observations due to
cluster sampling, and/or unequal probability of selection. When there is
clustering due to both primary and secondary sampling stages, the
standard errors and chi-square test of model fit are computed taking into
account the clustering due to the primary sampling stage and clustering
due to the secondary sampling stage is modeled.

Most of the special features listed above are available for modeling of
complex survey data.

MODELING WITH MISSING DATA

Mplus has several options for the estimation of models with missing
data. Mplus provides maximum likelihood estimation under MCAR
(missing completely at random), MAR (missing at random), and NMAR
(not missing at random) for continuous, censored, binary, ordered
categorical (ordinal), unordered categorical (nominal), counts, or
combinations of these variable types (Little & Rubin, 2002). MAR
means that missingness can be a function of observed covariates and

7
CHAPTER 1

observed outcomes. For censored and categorical outcomes using

weighted least squares estimation, missingness is allowed to be a
function of the observed covariates but not the observed outcomes
(Asparouhov & Muthén, 2010a). When there are no covariates in the
model, this is analogous to pairwise present analysis. Non-ignorable
missing data (NMAR) modeling is possible using maximum likelihood
estimation where categorical outcomes are indicators of missingness and
where missingness can be predicted by continuous and categorical latent
variables (Muthén, Jo, & Brown, 2003; Muthén et al., 2011).

In all models, missingness is not allowed for the observed covariates

because they are not part of the model. The model is estimated
conditional on the covariates and no distributional assumptions are made
about the covariates. Covariate missingness can be modeled if the
covariates are brought into the model and distributional assumptions
such as normality are made about them. With missing data, the standard
errors for the parameter estimates are computed using the observed
information matrix (Kenward & Molenberghs, 1998). Bootstrap
standard errors and confidence intervals are also available with missing
data.

Mplus provides multiple imputation of missing data using Bayesian

analysis (Rubin, 1987; Schafer, 1997). Both the unrestricted H1 model
and a restricted H0 model can be used for imputation. Multiple data sets
generated using multiple imputation can be analyzed using a special
feature of Mplus. Parameter estimates are averaged over the set of
analyses, and standard errors are computed using the average of the
standard errors over the set of analyses and the between analysis
parameter estimate variation (Rubin, 1987; Schafer, 1997). A chi-square
test of overall model fit is provided (Asparouhov & Muthén, 2008c;
Enders, 2010).

ESTIMATORS AND ALGORITHMS

Mplus provides both Bayesian and frequentist inference. Bayesian
analysis uses Markov chain Monte Carlo (MCMC) algorithms. Posterior
distributions can be monitored by trace and autocorrelation plots.
Convergence can be monitored by the Gelman-Rubin potential scaling
reduction using parallel computing in multiple MCMC chains. Posterior
predictive checks are provided.

8
 Introduction

Frequentist analysis uses maximum likelihood and weighted least

squares estimators. Mplus provides maximum likelihood estimation for
all models. With censored and categorical outcomes, an alternative
weighted least squares estimator is also available. For all types of
outcomes, robust estimation of standard errors and robust chi-square
tests of model fit are provided. These procedures take into account non-
normality of outcomes and non-independence of observations due to
cluster sampling. Robust standard errors are computed using the
sandwich estimator. Robust chi-square tests of model fit are computed
using mean and mean and variance adjustments as well as a likelihood-
based approach. Bootstrap standard errors are available for most
models. The optimization algorithms use one or a combination of the
following: Quasi-Newton, Fisher scoring, Newton-Raphson, and the
Expectation Maximization (EM) algorithm (Dempster et al., 1977).
Linear and non-linear parameter constraints are allowed. With
maximum likelihood estimation and categorical outcomes, models with
continuous latent variables and missing data for dependent variables
require numerical integration in the computations. The numerical
integration is carried out with or without adaptive quadrature in
combination with rectangular integration, Gauss-Hermite integration, or
Monte Carlo integration.

MONTE CARLO SIMULATION CAPABILITIES

Mplus has extensive Monte Carlo facilities both for data generation and
data analysis. Several types of data can be generated: simple random
samples, clustered (multilevel) data, missing data, discrete- and
continuous-time survival data, and data from populations that are
observed (multiple groups) or unobserved (latent classes). Data
generation models can include random effects and interactions between
continuous latent variables and between categorical latent variables.
Outcome variables can be generated as continuous, censored, binary,
ordered categorical (ordinal), unordered categorical (nominal), counts,
or combinations of these variable types. In addition, two-part
(semicontinuous) variables and time-to-event variables can be generated.
Independent variables can be generated as binary or continuous. All or
some of the Monte Carlo generated data sets can be saved.

The analysis model can be different from the data generation model. For
example, variables can be generated as categorical and analyzed as
continuous or generated as a three-class model and analyzed as a two-

9
CHAPTER 1

class model. In some situations, a special external Monte Carlo feature

is needed to generate data by one model and analyze it by a different
model. For example, variables can be generated using a clustered design
and analyzed ignoring the clustering. Data generated outside of Mplus
can also be analyzed using this special external Monte Carlo feature.

Other special Monte Carlo features include saving parameter estimates

from the analysis of real data to be used as population and/or coverage
values for data generation in a Monte Carlo simulation study. In
addition, analysis results from each replication of a Monte Carlo
simulation study can be saved in an external file.

GRAPHICS
Mplus includes a dialog-based, post-processing graphics module that
provides graphical displays of observed data and analysis results
including outliers and influential observations.

These graphical displays can be viewed after the Mplus analysis is

completed. They include histograms, scatterplots, plots of individual
observed and estimated values, plots of sample and estimated means and
proportions/probabilities, plots of estimated probabilities for a
categorical latent variable as a function of its covariates, plots of item
characteristic curves and information curves, plots of survival and
hazard curves, plots of missing data statistics, plots of user-specified
functions, and plots related to Bayesian estimation. These are available
for the total sample, by group, by class, and adjusted for covariates. The
graphical displays can be edited and exported as a DIB, EMF, or JPEG
file. In addition, the data for each graphical display can be saved in an
external file for use by another graphics program.

DIAGRAMMER
The Diagrammer can be used to draw an input diagram, to automatically
create an output diagram, and to automatically create a diagram using an
Mplus input without an analysis or data. To draw an input diagram, the
Diagrammer is accessed through the Open Diagrammer menu option of
the Diagram menu in the Mplus Editor. The Diagrammer uses a set of
drawing tools and pop-up menus to draw a diagram. When an input
diagram is drawn, a partial input is created which can be edited before
the analysis. To automatically create an output diagram, an input is

10
 Introduction

created in the Mplus Editor. The output diagram is automatically

created when the analysis is completed. This diagram can be edited and
used in a new analysis. The Diagrammer can be used as a drawing tool
by using an input without an analysis or data.

LTA CALCULATOR
Conditional probabilities, including latent transition probabilities, for
different values of a set of covariates can be computed using the LTA
Calculator. It is accessed by choosing LTA calculator from the Mplus
menu of the Mplus Editor.

LANGUAGE GENERATOR
Mplus includes a language generator to help users create Mplus input
files. The language generator takes users through a series of screens that
prompts them for information about their data and model. The language
generator contains all of the Mplus commands except DEFINE,
MODEL, PLOT, and MONTECARLO. Features added after Version 2
are not included in the language generator.

THE ORGANIZATION OF THE USER’S GUIDE

The Mplus User’s Guide has 20 chapters. Chapter 2 describes how to
get started with Mplus. Chapters 3 through 13 contain examples of
analyses that can be done using Mplus. Chapter 14 discusses special
issues. Chapters 15 through 19 describe the Mplus language. Chapter
20 contains a summary of the Mplus language. Technical appendices
that contain information on modeling, model estimation, model testing,
numerical algorithms, and references to further technical information
can be found at www.statmodel.com.

It is not necessary to read the entire User’s Guide before using the
program. A user may go straight to Chapter 2 for an overview of Mplus
and then to one of the example chapters.

11
CHAPTER 1

12
 Getting Started With Mplus

CHAPTER 2
GETTING STARTED WITH Mplus

After Mplus is installed, the program can be run from the Mplus editor.
The Mplus Editor for Windows includes a language generator and a
graphics module. The graphics module provides graphical displays of
observed data and analysis results.

In this chapter, a brief description of the user language is presented

along with an overview of the examples and some model estimation
considerations.

THE Mplus LANGUAGE

The user language for Mplus consists of a set of ten commands each of
which has several options. The default options for Mplus have been
chosen so that user input can be minimized for the most common types
of analyses. For most analyses, only a small subset of the Mplus
commands is needed. Complicated models can be easily described using
the Mplus language. The ten commands of Mplus are:

• TITLE
• DATA (required)
• VARIABLE (required)
• DEFINE
• ANALYSIS
• MODEL
• OUTPUT
• SAVEDATA
• PLOT
• MONTECARLO

The TITLE command is used to provide a title for the analysis. The
DATA command is used to provide information about the data set to be
analyzed. The VARIABLE command is used to provide information
about the variables in the data set to be analyzed. The DEFINE
command is used to transform existing variables and create new
variables. The ANALYSIS command is used to describe the technical

13
CHAPTER 2

details of the analysis. The MODEL command is used to describe the

model to be estimated. The OUTPUT command is used to request
additional output not included as the default. The SAVEDATA
command is used to save the analysis data, auxiliary data, and a variety
of analysis results. The PLOT command is used to request graphical
displays of observed data and analysis results. The MONTECARLO
command is used to specify the details of a Monte Carlo simulation
study.

The Mplus commands may come in any order. The DATA and
VARIABLE commands are required for all analyses. All commands
must begin on a new line and must be followed by a colon. Semicolons
separate command options. There can be more than one option per line.
The records in the input setup must be no longer than 90 columns. They
can contain upper and/or lower case letters and tabs.

Commands, options, and option settings can be shortened for

convenience. Commands and options can be shortened to four or more
letters. Option settings can be referred to by either the complete word or
the part of the word shown in bold type in the command boxes in each
chapter.

Comments can be included anywhere in the input setup. A comment is

designated by an exclamation point. Anything on a line following an
exclamation point is treated as a user comment and is ignored by the
program.

The keywords IS, ARE, and = can be used interchangeably in all

commands except DEFINE, MODEL CONSTRAINT, and MODEL
TEST. Items in a list can be separated by blanks or commas.

Mplus uses a hyphen (-) to indicate a list of variables or numbers. The

use of this feature is discussed in each section for which it is appropriate.
There is also a special keyword ALL which can be used to indicate all
variables. This keyword is discussed with the options that use it.

Following is a set of Mplus input files for a few prototypical examples.

The first example shows the input file for a factor analysis with
covariates (MIMIC model).

14
 Getting Started With Mplus

TITLE: this is an example of a MIMIC model

with two factors, six continuous factor
indicators, and three covariates
DATA: FILE IS mimic.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;

The second example shows the input file for a growth model with time-
invariant covariates. It illustrates the new simplified Mplus language for
specifying growth models.

TITLE: this is an example of a linear growth

model for a continuous outcome at four
time points with the intercept and slope
growth factors regressed on two time-
invariant covariates
DATA: FILE IS growth.dat;
VARIABLE: NAMES ARE y1-y4 x1 x2;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
i s ON x1 x2;

The third example shows the input file for a latent class analysis with
covariates and a direct effect.

TITLE: this is an example of a latent class

analysis with two classes, one covariate,
and a direct effect
DATA: FILE IS lcax.dat;
VARIABLE: NAMES ARE u1-u4 x;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
c ON x;
u4 ON x;

The fourth example shows the input file for a multilevel regression
model with a random intercept and a random slope varying across
clusters.

15
CHAPTER 2

TITLE: this is an example of a multilevel

regression analysis with one individual-
level outcome variable regressed on an
individual-level background variable where
the intercept and slope are regressed on a
cluster-level variable
DATA: FILE IS reg.dat;
VARIABLE: NAMES ARE clus y x w;
CLUSTER = clus;
WITHIN = x;
BETWEEN = w;
MISSING = .;
DEFINE: CENTER x (GRANDMEAN);
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL:
%WITHIN%
s | y ON x;
%BETWEEN%
y s ON w;

OVERVIEW OF Mplus EXAMPLES

The next eleven chapters contain examples of prototypical input setups
for several different types of analyses. The input, data, and output, as
well as the corresponding Monte Carlo input and Monte Carlo output for
most of the examples are on the CD that contains the Mplus program.
The Monte Carlo input is used to generate the data for each example.
They are named using the example number. For example, the names of
the files for Example 3.1 are ex3.1.inp; ex3.1.dat; ex3.1.out;
mcex3.1.inp, and mcex3.1.out. The data in ex3.1.dat are generated using
mcex3.1.inp.

The examples presented do not cover all models that can be estimated
using Mplus but do cover the major areas of modeling. They can be
seen as building blocks that can be put together as needed. For example,
a model can combine features described in an example from one chapter
with features described in an example from another chapter. Many
unique and unexplored models can therefore be created. In each chapter,
all commands and options for the first example are discussed. After that,
only the highlighted parts of each example are discussed.

For clarity, certain conventions are used in the input setups. Program
commands, options, settings, and keywords are written in upper case.
Information provided by the user is written in lower case. Note,

16
 Getting Started With Mplus

however, that Mplus is not case sensitive. Upper and lower case can be
used interchangeably in the input setups.

For simplicity, the input setups for the examples are generic. Observed
continuous and censored outcome variable names start with a y;
observed binary or ordered categorical (ordinal), unordered categorical
(nominal), and count outcome variable names start with a u; time-to-
event variables in continuous-time survival analysis start with a t;
observed background variable names start with an x; observed time-
varying background variables start with an a; observed between-level
background variables start with a w; continuous latent variable names
start with an f; categorical latent variable names start with a c; intercept
growth factor names start with an i; and slope growth factor names and
random slope names start with an s or a q. Note, however, that variable
names are not limited to these choices.

Following is a list of the example chapters:

• Chapter 3: Regression and path analysis

• Chapter 4: Exploratory factor analysis
• Chapter 5: Confirmatory factor analysis and structural equation
modeling
• Chapter 6: Growth modeling and survival analysis
• Chapter 7: Mixture modeling with cross-sectional data
• Chapter 8: Mixture modeling with longitudinal data
• Chapter 9: Multilevel modeling with complex survey data
• Chapter 10: Multilevel mixture modeling
• Chapter 11: Missing data modeling and Bayesian analysis
• Chapter 12: Monte Carlo simulation studies
• Chapter 13: Special features

The Mplus Base program covers the analyses described in Chapters 3, 5,

6, 11, 13, and parts of Chapters 4 and 12. The Mplus Base program does
not include analyses with TYPE=MIXTURE, TYPE=TWOLEVEL,
TYPE=THREELEVEL, or TYPE=CROSSCLASSIFIED.

The Mplus Base and Mixture Add-On program covers the analyses
described in Chapters 3, 5, 6, 7, 8, 11, 13, and parts of Chapters 4 and
12. The Mplus Base and Mixture Add-On program does not include
analyses with TYPE=TWOLEVEL, TYPE=THREELEVEL, or
TYPE=CROSSCLASSIFIED.

17
CHAPTER 2

The Mplus Base and Multilevel Add-On program covers the analyses
described in Chapters 3, 5, 6, 9, 11, 13, and parts of Chapters 4 and 12.
The Mplus Base and Multilevel Add-On program does not include
analyses with TYPE=MIXTURE.

The Mplus Base and Combination Add-On program covers the analyses
described in all chapters. There are no restrictions on the analyses that
can be requested.

18
 Examples: Regression And Path Analysis

CHAPTER 3
EXAMPLES: REGRESSION AND
PATH ANALYSIS

Regression analysis with univariate or multivariate dependent variables

is a standard procedure for modeling relationships among observed
variables. Path analysis allows the simultaneous modeling of several
related regression relationships. In path analysis, a variable can be a
dependent variable in one relationship and an independent variable in
another. These variables are referred to as mediating variables. For both
types of analyses, observed dependent variables can be continuous,
censored, binary, ordered categorical (ordinal), counts, or combinations
of these variable types. In addition, for regression analysis and path
analysis for non-mediating variables, observed dependent variables can
be unordered categorical (nominal).

For continuous dependent variables, linear regression models are used.

For censored dependent variables, censored-normal regression models
are used, with or without inflation at the censoring point. For binary and
ordered categorical dependent variables, probit or logistic regression
models are used. Logistic regression for ordered categorical dependent
variables uses the proportional odds specification. For unordered
categorical dependent variables, multinomial logistic regression models
are used. For count dependent variables, Poisson regression models are
used, with or without inflation at the zero point. Both maximum
likelihood and weighted least squares estimators are available.

All regression and path analysis models can be estimated using the
following special features:

• Single or multiple group analysis

• Missing data
• Complex survey data
• Random slopes
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals

19
CHAPTER 3

• Wald chi-square test of parameter equalities

For continuous, censored with weighted least squares estimation, binary,

and ordered categorical (ordinal) outcomes, multiple group analysis is
specified by using the GROUPING option of the VARIABLE command
for individual data or the NGROUPS option of the DATA command for
summary data. For censored with maximum likelihood estimation,
unordered categorical (nominal), and count outcomes, multiple group
analysis is specified using the KNOWNCLASS option of the
VARIABLE command in conjunction with the TYPE=MIXTURE
option of the ANALYSIS command. The default is to estimate the
model under missing data theory using all available data. The
LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. The
SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed. Random slopes are
specified by using the | symbol of the MODEL command in conjunction
with the ON option of the MODEL command. Linear and non-linear
parameter constraints are specified by using the MODEL
CONSTRAINT command. Indirect effects are specified by using the
MODEL INDIRECT command. Maximum likelihood estimation is
specified by using the ESTIMATOR option of the ANALYSIS
command. Bootstrap standard errors are obtained by using the
BOOTSTRAP option of the ANALYSIS command. Bootstrap
confidence intervals are obtained by using the BOOTSTRAP option of
the ANALYSIS command in conjunction with the CINTERVAL option
of the OUTPUT command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, and plots
of sample and estimated means and proportions/probabilities. These are

20
 Examples: Regression And Path Analysis

available for the total sample, by group, by class, and adjusted for
covariates. The PLOT command includes a display showing a set of
descriptive statistics for each variable. The graphical displays can be
edited and exported as a DIB, EMF, or JPEG file. In addition, the data
for each graphical display can be saved in an external file for use by
another graphics program.

Following is the set of regression examples included in this chapter:

• 3.1: Linear regression

• 3.2: Censored regression
• 3.3: Censored-inflated regression
• 3.4: Probit regression
• 3.5: Logistic regression
• 3.6: Multinomial logistic regression
• 3.7: Poisson regression
• 3.8: Zero-inflated Poisson and negative binomial regression
• 3.9: Random coefficient regression
• 3.10: Non-linear constraint on the logit parameters of an unordered
categorical (nominal) variable

Following is the set of path analysis examples included in this chapter:

• 3.11: Path analysis with continuous dependent variables

• 3.12: Path analysis with categorical dependent variables
• 3.13: Path analysis with categorical dependent variables using the
Theta parameterization
• 3.14: Path analysis with a combination of continuous and
categorical dependent variables
• 3.15: Path analysis with a combination of censored, categorical, and
unordered categorical (nominal) dependent variables
• 3.16: Path analysis with continuous dependent variables,
bootstrapped standard errors, indirect effects, and confidence
intervals
• 3.17: Path analysis with a categorical dependent variable and a
continuous mediating variable with missing data*
• 3.18: Moderated mediation with a plot of the indirect effect

21
CHAPTER 3

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

EXAMPLE 3.1: LINEAR REGRESSION

TITLE: this is an example of a linear regression

for a continuous observed dependent
variable with two covariates
DATA: FILE IS ex3.1.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1 x1 x3;
MODEL: y1 ON x1 x3;

In this example, a linear regression is estimated.

TITLE: this is an example of a linear regression

for a continuous observed dependent
variable with two covariates

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex3.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex3.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y1-y6 x1-x4;

USEVARIABLES ARE y1 x1 x3;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains ten variables: y1, y2, y3, y4, y5, y6, x1, x2, x3, and
x4. Note that the hyphen can be used as a convenience feature in order
to generate a list of names. If not all of the variables in the data set are
used in the analysis, the USEVARIABLES option can be used to select a
subset of variables for analysis. Here the variables y1, x1, and x3 have

22
 Examples: Regression And Path Analysis

been selected for analysis. Because the scale of the dependent variable
is not specified, it is assumed to be continuous.

MODEL: y1 ON x1 x3;

The MODEL command is used to describe the model to be estimated.

The ON statement describes the linear regression of y1 on the covariates
x1 and x3. It is not necessary to refer to the means, variances, and
covariances among the x variables in the MODEL command because the
parameters of the x variables are not part of the model estimation.
Because the model does not impose restrictions on the parameters of the
x variables, these parameters can be estimated separately as the sample
values. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator.

EXAMPLE 3.2: CENSORED REGRESSION

TITLE: this is an example of a censored

regression for a censored dependent
variable with two covariates
DATA: FILE IS ex3.2.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1 x1 x3;
CENSORED ARE y1 (b);
ANALYSIS: ESTIMATOR = MLR;
MODEL: y1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a censored variable instead of a continuous
variable. The CENSORED option is used to specify which dependent
variables are treated as censored variables in the model and its
estimation, whether they are censored from above or below, and whether
a censored or censored-inflated model will be estimated. In the example
above, y1 is a censored variable. The b in parentheses following y1
indicates that y1 is censored from below, that is, has a floor effect, and
that the model is a censored regression model. The censoring limit is
determined from the data. The default estimator for this type of analysis
is a robust weighted least squares estimator. By specifying
ESTIMATOR=MLR, maximum likelihood estimation with robust
standard errors is used. The ON statement describes the censored

23
CHAPTER 3

regression of y1 on the covariates x1 and x3. An explanation of the

other commands can be found in Example 3.1.

EXAMPLE 3.3: CENSORED-INFLATED REGRESSION

TITLE: this is an example of a censored-inflated

regression for a censored dependent
variable with two covariates
DATA: FILE IS ex3.3.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1 x1 x3;
CENSORED ARE y1 (bi);
MODEL: y1 ON x1 x3;
y1#1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a censored variable instead of a continuous
variable. The CENSORED option is used to specify which dependent
variables are treated as censored variables in the model and its
estimation, whether they are censored from above or below, and whether
a censored or censored-inflated model will be estimated. In the example
above, y1 is a censored variable. The bi in parentheses following y1
indicates that y1 is censored from below, that is, has a floor effect, and
that a censored-inflated regression model will be estimated. The
censoring limit is determined from the data.

With a censored-inflated model, two regressions are estimated. The first

ON statement describes the censored regression of the continuous part of
y1 on the covariates x1 and x3. This regression predicts the value of the
censored dependent variable for individuals who are able to assume
values of the censoring point and above. The second ON statement
describes the logistic regression of the binary latent inflation variable
y1#1 on the covariates x1 and x3. This regression predicts the
probability of being unable to assume any value except the censoring
point. The inflation variable is referred to by adding to the name of the
censored variable the number sign (#) followed by the number 1. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 3.1.

24
 Examples: Regression And Path Analysis

EXAMPLE 3.4: PROBIT REGRESSION

TITLE: this is an example of a probit regression

for a binary or categorical observed
dependent variable with two covariates
DATA: FILE IS ex3.4.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
CATEGORICAL = u1;
MODEL: u1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a binary or ordered categorical (ordinal) variable
instead of a continuous variable. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the
example above, u1 is a binary or ordered categorical variable. The
program determines the number of categories. The ON statement
describes the probit regression of u1 on the covariates x1 and x3. The
default estimator for this type of analysis is a robust weighted least
squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 3.1.

EXAMPLE 3.5: LOGISTIC REGRESSION

TITLE: this is an example of a logistic

regression for a categorical observed
dependent variable with two covariates
DATA: FILE IS ex3.5.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
CATEGORICAL IS u1;
ANALYSIS: ESTIMATOR = ML;
MODEL: u1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a binary or ordered categorical (ordinal) variable
instead of a continuous variable. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the

25
CHAPTER 3

example above, u1 is a binary or ordered categorical variable. The

program determines the number of categories. By specifying
ESTIMATOR=ML, a logistic regression will be estimated. The ON
statement describes the logistic regression of u1 on the covariates x1 and
x3. An explanation of the other commands can be found in Example 3.1.

EXAMPLE 3.6: MULTINOMIAL LOGISTIC REGRESSION

TITLE: this is an example of a multinomial

logistic regression for an unordered
categorical (nominal) dependent variable
with two covariates
DATA: FILE IS ex3.6.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
NOMINAL IS u1;
MODEL: u1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is an unordered categorical (nominal) variable
instead of a continuous variable. The NOMINAL option is used to
specify which dependent variables are treated as unordered categorical
variables in the model and its estimation. In the example above, u1 is a
three-category unordered variable. The program determines the number
of categories. The ON statement describes the multinomial logistic
regression of u1 on the covariates x1 and x3 when comparing categories
one and two of u1 to the third category of u1. The intercept and slopes
of the last category are fixed at zero as the default. The default estimator
for this type of analysis is maximum likelihood with robust standard
errors. The ESTIMATOR option of the ANALYSIS command can be
used to select a different estimator. An explanation of the other
commands can be found in Example 3.1.

Following is an alternative specification of the multinomial logistic

regression of u1 on the covariates x1 and x3:

u1#1 u1#2 ON x1 x3;

where u1#1 refers to the first category of u1 and u1#2 refers to the
second category of u1. The categories of an unordered categorical
variable are referred to by adding to the name of the unordered

26
 Examples: Regression And Path Analysis

categorical variable the number sign (#) followed by the number of the
category. This alternative specification allows individual parameters to
be referred to in the MODEL command for the purpose of giving starting
values or placing restrictions.

EXAMPLE 3.7: POISSON REGRESSION

TITLE: this is an example of a Poisson regression

for a count dependent variable with two
covariates
DATA: FILE IS ex3.7.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
COUNT IS u1;
MODEL: u1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u1 is a count variable that is not inflated. The ON
statement describes the Poisson regression of u1 on the covariates x1
and x3. The default estimator for this type of analysis is maximum
likelihood with robust standard errors. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 3.1.

27
CHAPTER 3

EXAMPLE 3.8: ZERO-INFLATED POISSON AND NEGATIVE

BINOMIAL REGRESSION

TITLE: this is an example of a zero-inflated

Poisson regression for a count dependent
variable with two covariates
DATA: FILE IS ex3.8a.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
COUNT IS u1 (i);
MODEL: u1 ON x1 x3;
u1#1 ON x1 x3;

The difference between this example and Example 3.1 is that the
dependent variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the first
part of this example, a zero-inflated Poisson regression is estimated. In
the example above, u1 is a count variable. The i in parentheses
following u1 indicates that a zero-inflated Poisson model will be
estimated. In the second part of this example, a negative binomial model
is estimated.

With a zero-inflated Poisson model, two regressions are estimated. The

first ON statement describes the Poisson regression of the count part of
u1 on the covariates x1 and x3. This regression predicts the value of the
count dependent variable for individuals who are able to assume values
of zero and above. The second ON statement describes the logistic
regression of the binary latent inflation variable u1#1 on the covariates
x1 and x3. This regression predicts the probability of being unable to
assume any value except zero. The inflation variable is referred to by
adding to the name of the count variable the number sign (#) followed by
the number 1. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 3.1.

An alternative way of specifying this model is presented in Example

7.25. In Example 7.25, a categorical latent variable with two classes is

28
 Examples: Regression And Path Analysis

used to represent individuals who are able to assume values of zero and
above and individuals who are unable to assume any value except zero.
This approach allows the estimation of the probability of being in each
class and the posterior probabilities of being in each class for each
individual.

TITLE: this is an example of a negative binomial

model for a count dependent variable with
two covariates
DATA: FILE IS ex3.8b.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1 x1 x3;
COUNT IS u1 (nb);
MODEL: u1 ON x1 x3;

The difference between this part of the example and the first part is that
a regression for a count outcome using a negative binomial model is
estimated instead of a zero-inflated Poisson model. The negative
binomial model estimates a dispersion parameter for each of the
outcomes (Long, 1997; Hilbe, 2011).

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and which type
of model is estimated. The nb in parentheses following u1 indicates that
a negative binomial model will be estimated. The dispersion parameter
can be referred to using the name of the count variable. An explanation
of the other commands can be found in the first part of this example and
in Example 3.1.

EXAMPLE 3.9: RANDOM COEFFICIENT REGRESSION

TITLE: this is an example of a random coefficient

regression
DATA: FILE IS ex3.9.dat;
VARIABLE: NAMES ARE y x1 x2;
DEFINE: CENTER x1 x2 (GRANDMEAN);
ANALYSIS: TYPE = RANDOM;
MODEL: s | y ON x1;
s WITH y;
y s ON x2;

29
CHAPTER 3

x1 y

x2 s

In this example a regression with random coefficients shown in the

picture above is estimated. Random coefficient regression uses random
slopes to model heterogeneity in the residual variance as a function of a
covariate that has a random slope (Hildreth & Houck, 1968; Johnston,
1984). The s shown in a circle represents the random slope. The broken
arrow from s to the arrow from x1 to y indicates that the slope in this
regression is random. The random slope is predicted by the covariate
x2.

The CENTER option is used to specify the type of centering to be used

in an analysis and the variables that will be centered. Centering
facilitates the interpretation of the results. In this example, the
covariates are centered using the grand means, that is, the sample means
of x1 and x2 are subtracted from the values of the covariates x1 and x2.
The TYPE option is used to describe the type of analysis that is to be
performed.

By selecting RANDOM, a model with random slopes will be estimated.

The | symbol is used in conjunction with TYPE=RANDOM to name and
define the random slope variables in the model. The name on the left-
hand side of the | symbol names the random slope variable. The
statement on the right-hand side of the | symbol defines the random slope
variable. The random slope s is defined by the linear regression of y on
the covariate x1. The residual variance in the regression of y on x is
estimated as the default. The residual covariance between s and y is
fixed at zero as the default. The WITH statement is used to free this
parameter. The ON statement describes the linear regressions of the
dependent variable y and the random slope s on the covariate x2. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The estimator option of the ANALYSIS

30
 Examples: Regression And Path Analysis

command can be used to select a different estimator. An explanation of

the other commands can be found in Example 3.1.

EXAMPLE 3.10: NON-LINEAR CONSTRAINT ON THE LOGIT

PARAMETERS OF AN UNORDERED CATEGORICAL
(NOMINAL) VARIABLE

TITLE: this is an example of non-linear

constraint on the logit parameters of an
unordered categorical (nominal) variable
DATA: FILE IS ex3.10.dat;
VARIABLE: NAMES ARE u;
NOMINAL = u;
MODEL: [u#1] (p1);
[u#2] (p2);
[u#3] (p2);
MODEL CONSTRAINT:
p2 = log ((exp (p1) – 1)/2 – 1);

In this example, theory specifies the following probabilities for the four
categories of an unordered categorical (nominal) variable: ½ + ¼ p, ¼
(1-p), ¼ (1-p), ¼ p, where p is a probability parameter to be estimated.
These restrictions on the category probabilities correspond to non-linear
constraints on the logit parameters for the categories in the multinomial
logistic model. This example is based on Dempster, Laird, and Rubin
(1977, p. 2).

The NOMINAL option is used to specify which dependent variables are

treated as unordered categorical (nominal) variables in the model and its
estimation. In the example above, u is a four-category unordered
variable. The program determines the number of categories. The
categories of an unordered categorical variable are referred to by adding
to the name of the unordered categorical variable the number sign (#)
followed by the number of the category. In this example, u#1 refers to
the first category of u, u#2 refers to the second category of u, and u#3
refers to the third category of u.

In the MODEL command, parameters are given labels by placing a name

in parentheses after the parameter. The logit parameter for category one
is referred to as p1; the logit parameter for category two is referred to as
p2; and the logit parameter for category three is also referred to as p2.

31
CHAPTER 3

When two parameters are referred to using the same label, they are held
equal. The MODEL CONSTRAINT command is used to define linear
and non-linear constraints on the parameters in the model. The non-
linear constraint for the logits follows from the four probabilities given
above after some algebra. The default estimator for this type of analysis
is maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 3.1.

EXAMPLE 3.11: PATH ANALYSIS WITH CONTINUOUS

DEPENDENT VARIABLES

TITLE: this is an example of a path analysis

with continuous dependent variables
DATA: FILE IS ex3.11.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1-y3 x1-x3;
MODEL: y1 y2 ON x1 x2 x3;
y3 ON y1 y2 x2;

x1

y1

x2 y3

y2

x3

In this example, the path analysis model shown in the picture above is
estimated. The dependent variables in the analysis are continuous. Two
of the dependent variables y1 and y2 mediate the effects of the
covariates x1, x2, and x3 on the dependent variable y3.

32
 Examples: Regression And Path Analysis

The first ON statement describes the linear regressions of y1 and y2 on

the covariates x1, x2, and x3. The second ON statement describes the
linear regression of y3 on the mediating variables y1 and y2 and the
covariate x2. The residual variances of the three dependent variables are
estimated as the default. The residuals are not correlated as the default.
As in regression analysis, it is not necessary to refer to the means,
variances, and covariances among the x variables in the MODEL
command because the parameters of the x variables are not part of the
model estimation. Because the model does not impose restrictions on
the parameters of the x variables, these parameters can be estimated
separately as the sample values. The default estimator for this type of
analysis is maximum likelihood. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 3.1.

EXAMPLE 3.12: PATH ANALYSIS WITH CATEGORICAL

DEPENDENT VARIABLES

TITLE: this is an example of a path analysis

with categorical dependent variables
DATA: FILE IS ex3.12.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1-u3 x1-x3;
CATEGORICAL ARE u1-u3;
MODEL: u1 u2 ON x1 x2 x3;
u3 ON u1 u2 x2;

The difference between this example and Example 3.11 is that the
dependent variables are binary and/or ordered categorical (ordinal)
variables instead of continuous variables. The CATEGORICAL option
is used to specify which dependent variables are treated as binary or
ordered categorical (ordinal) variables in the model and its estimation.
In the example above, u1, u2, and u3 are binary or ordered categorical
variables. The program determines the number of categories for each
variable. The first ON statement describes the probit regressions of u1
and u2 on the covariates x1, x2, and x3. The second ON statement
describes the probit regression of u3 on the mediating variables u1 and
u2 and the covariate x2. The default estimator for this type of analysis is
a robust weighted least squares estimator. The ESTIMATOR option of
the ANALYSIS command can be used to select a different estimator. If
the maximum likelihood estimator is selected, the regressions are

33
CHAPTER 3

logistic regressions. An explanation of the other commands can be

found in Example 3.1.

EXAMPLE 3.13: PATH ANALYSIS WITH CATEGORICAL

DEPENDENT VARIABLES USING THE THETA
PARAMETERIZATION

TITLE: this is an example of a path analysis

with categorical dependent variables using
the Theta parameterization
DATA: FILE IS ex3.13.dat;
VARIABLE: NAMES ARE u1-u6 x1-x4;
USEVARIABLES ARE u1-u3 x1-x3;
CATEGORICAL ARE u1-u3;
ANALYSIS: PARAMETERIZATION = THETA;
MODEL: u1 u2 ON x1 x2 x3;
u3 ON u1 u2 x2;

The difference between this example and Example 3.12 is that the Theta
parameterization is used instead of the default Delta parameterization.
In the Delta parameterization, scale factors for continuous latent
response variables of observed categorical dependent variables are
allowed to be parameters in the model, but residual variances for
continuous latent response variables are not. In the Theta
parameterization, residual variances for continuous latent response
variables of observed categorical dependent variables are allowed to be
parameters in the model, but scale factors for continuous latent response
variables are not. An explanation of the other commands can be found
in Examples 3.1 and 3.12.

34
 Examples: Regression And Path Analysis

EXAMPLE 3.14: PATH ANALYSIS WITH A COMBINATION

OF CONTINUOUS AND CATEGORICAL DEPENDENT
VARIABLES

TITLE: this is an example of a path analysis

with a combination of continuous and
categorical dependent variables
DATA: FILE IS ex3.14.dat;
VARIABLE: NAMES ARE y1 y2 u1 y4-y6 x1-x4;
USEVARIABLES ARE y1-u1 x1-x3;
CATEGORICAL IS u1;
MODEL: y1 y2 ON x1 x2 x3;
u1 ON y1 y2 x2;

The difference between this example and Example 3.11 is that the
dependent variables are a combination of continuous and binary or
ordered categorical (ordinal) variables instead of all continuous
variables. The CATEGORICAL option is used to specify which
dependent variables are treated as binary or ordered categorical (ordinal)
variables in the model and its estimation. In the example above, y1 and
y2 are continuous variables and u1 is a binary or ordered categorical
variable. The program determines the number of categories. The first
ON statement describes the linear regressions of y1 and y2 on the
covariates x1, x2, and x3. The second ON statement describes the probit
regression of u1 on the mediating variables y1 and y2 and the covariate
x2. The default estimator for this type of analysis is a robust weighted
least squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. If a maximum
likelihood estimator is selected, the regression for u1 is a logistic
regression. An explanation of the other commands can be found in
Example 3.1.

35
CHAPTER 3

EXAMPLE 3.15: PATH ANALYSIS WITH A COMBINATION

OF CENSORED, CATEGORICAL, AND UNORDERED
CATEGORICAL (NOMINAL) DEPENDENT VARIABLES

TITLE: this is an example of a path analysis

with a combination of censored,
categorical, and unordered categorical
(nominal) dependent variables
DATA: FILE IS ex3.15.dat;
VARIABLE: NAMES ARE y1 u1 u2 y4-y6 x1-x4;
USEVARIABLES ARE y1-u2 x1-x3;
CENSORED IS y1 (a);
CATEGORICAL IS u1;
NOMINAL IS u2;
MODEL: y1 u1 ON x1 x2 x3;
u2 ON y1 u1 x2;

The difference between this example and Example 3.11 is that the
dependent variables are a combination of censored, binary or ordered
categorical (ordinal), and unordered categorical (nominal) variables
instead of continuous variables. The CENSORED option is used to
specify which dependent variables are treated as censored variables in
the model and its estimation, whether they are censored from above or
below, and whether a censored or censored-inflated model will be
estimated. In the example above, y1 is a censored variable. The a in
parentheses following y1 indicates that y1 is censored from above, that
is, has a ceiling effect, and that the model is a censored regression
model. The censoring limit is determined from the data. The
CATEGORICAL option is used to specify which dependent variables
are treated as binary or ordered categorical (ordinal) variables in the
model and its estimation. In the example above, u1 is a binary or
ordered categorical variable. The program determines the number of
categories. The NOMINAL option is used to specify which dependent
variables are treated as unordered categorical (nominal) variables in the
model and its estimation. In the example above, u2 is a three-category
unordered variable. The program determines the number of categories.

The first ON statement describes the censored regression of y1 and the

logistic regression of u1 on the covariates x1, x2, and x3. The second
ON statement describes the multinomial logistic regression of u2 on the
mediating variables y1 and u1 and the covariate x2 when comparing

36
 Examples: Regression And Path Analysis

categories one and two of u2 to the third category of u2. The intercept
and slopes of the last category are fixed at zero as the default. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 3.1.

Following is an alternative specification of the multinomial logistic

regression of u2 on the mediating variables y1 and u1 and the covariate
x2:

u2#1 u2#2 ON y1 u1 x2;

where u2#1 refers to the first category of u2 and u2#2 refers to the
second category of u2. The categories of an unordered categorical
variable are referred to by adding to the name of the unordered
categorical variable the number sign (#) followed by the number of the
category. This alternative specification allows individual parameters to
be referred to in the MODEL command for the purpose of giving starting
values or placing restrictions.

EXAMPLE 3.16: PATH ANALYSIS WITH CONTINUOUS

DEPENDENT VARIABLES, BOOTSTRAPPED STANDARD
ERRORS, INDIRECT EFFECTS, AND CONFIDENCE
INTERVALS

TITLE: this is an example of a path analysis

with continuous dependent variables,
bootstrapped standard errors, indirect
effects, and confidence intervals
DATA: FILE IS ex3.16.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1-y3 x1-x3;
ANALYSIS: BOOTSTRAP = 1000;
MODEL: y1 y2 ON x1 x2 x3;
y3 ON y1 y2 x2;
MODEL INDIRECT:
y3 IND y1 x1;
y3 IND y2 x1;
OUTPUT: CINTERVAL;

37
CHAPTER 3

The difference between this example and Example 3.11 is that

bootstrapped standard errors, indirect effects, and confidence intervals
are requested. The BOOTSTRAP option is used to request
bootstrapping and to specify the number of bootstrap draws to be used in
the computation. When the BOOTSTRAP option is used alone,
bootstrap standard errors of the model parameter estimates are obtained.
When the BOOTSTRAP option is used in conjunction with the
CINTERVAL option of the OUTPUT command, bootstrap standard
errors of the model parameter estimates and bootstrap confidence
intervals for the model parameter estimates are obtained. The
BOOTSTRAP option can be used in conjunction with the MODEL
INDIRECT command to obtain bootstrap standard errors for indirect
effects. When both MODEL INDIRECT and CINTERVAL are used,
bootstrap standard errors and bootstrap confidence intervals are obtained
for the indirect effects. By selecting BOOTSTRAP=1000, bootstrapped
standard errors will be computed using 1000 draws.

The MODEL INDIRECT command is used to request indirect effects

and their standard errors. Total indirect, specific indirect, and total
effects are obtained using the IND and VIA options of the MODEL
INDIRECT command. The IND option is used to request a specific
indirect effect or a set of indirect effects. In the IND statements above,
the variable on the left-hand side of IND is the dependent variable. The
last variable on the right-hand side of IND is the independent variable.
Other variables on the right-hand side of IND are mediating variables.
The first IND statement requests the specific indirect effect from x1 to
y1 to y3. The second IND statement requests the specific indirect effect
from x1 to y2 to y3. Total effects are computed for all IND statements
that start and end with the same variables.

The CINTERVAL option is used to request confidence intervals for

parameter estimates of the model, indirect effects, and standardized
indirect effects. When the BOOTSTRAP option is requested in the
ANALYSIS command, bootstrapped standard errors are computed.
When both the CINTERVALS and BOOTSTRAP options are used,
bootstrapped confidence intervals are computed. An explanation of the
other commands can be found in Examples 3.1 and 3.11.

38
 Examples: Regression And Path Analysis

EXAMPLE 3.17: PATH ANALYSIS WITH A CATEGORICAL

DEPENDENT VARIABLE AND A CONTINUOUS MEDIATING
VARIABLE WITH MISSING DATA

TITLE: this is an example of a path analysis

with a categorical dependent variable and
a continuous mediating variable with
missing data
DATA: FILE IS ex3.17.dat;
VARIABLE: NAMES ARE u y x;
CATEGORICAL IS u;
MISSING IS y (999);
ANALYSIS: ESTIMATOR = MLR;
INTEGRATION = MONTECARLO;
MODEL: y ON x;
u ON y x;
OUTPUT: TECH1 TECH8;

In this example, the dependent variable is binary or ordered categorical

(ordinal) and the continuous mediating variable has missing values. The
CATEGORICAL option is used to specify which dependent variables
are treated as binary or ordered categorical (ordinal) variables in the
model and its estimation. In the example above, u is a binary or ordered
categorical variable. The program determines the number of categories.
The MISSING option is used to identify the values or symbols in the
analysis data set that will be treated as missing or invalid. In this
example, the number 999 is the missing value flag. By specifying
ESTIMATOR=MLR, a maximum likelihood estimator with robust
standard errors using a numerical integration algorithm will be used.
Note that numerical integration becomes increasingly more
computationally demanding as the number of mediating variables with
missing data and the sample size increase. In this example, Monte Carlo
integration with 500 integration points is used. The ESTIMATOR
option can be used to select a different estimator.

The first ON statement describes the linear regression of y on the

covariate x. The second ON statement describes the logistic regression
of u on the mediating variable y and the covariate x. The OUTPUT
command is used to request additional output not included as the default.
The TECH1 option is used to request the arrays containing parameter
specifications and starting values for all free parameters in the model.

39
CHAPTER 3

The TECH8 option is used to request that the optimization history in

estimating the model be printed in the output. TECH8 is printed to the
screen during the computations as the default. TECH8 screen printing is
useful for determining how long the analysis takes. An explanation of
the other commands can be found in Example 3.1.

EXAMPLE 3.18: MODERATED MEDIATION WITH A PLOT

OF THE INDIRECT EFFECT

TITLE: this is an example of moderated mediation

with a plot of the indirect effect
DATA: FILE = ex3.18.dat;
VARIABLE: NAMES = y m x z;
USEVARIABLES = y m x xz;
DEFINE: xz = x*z;
ANALYSIS: ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (30000);
MODEL: y ON m (b);
m ON x (gamma1)
z
xz (gamma2);
MODEL CONSTRAINT:
PLOT(indirect);
LOOP(mod,-2,2,0.1);
indirect = b*(gamma1+gamma2*mod);
PLOT: TYPE = PLOT2;
OUTPUT: TECH8;

In this example, a moderated mediation analysis with a plot of the

indirect effect is carried out (Preacher, Rucker, & Hayes, 2007). The
variable z moderates the relationship between the mediator m and the
covariate x. The DEFINE command is used to create the variable xz
which is the interaction between the moderator z and the covariate x.
The variable xz must be included on the USEVARIABLES list after the
original variables in order to be used in the analysis.

By specifying ESTIMATOR=BAYES, a Bayesian analysis will be

carried out. In Bayesian estimation, the default is to use two
independent Markov chain Monte Carlo (MCMC) chains. If multiple
processors are available, using PROCESSORS=2 will speed up
computations. The BITERATIONS option is used to specify the
maximum and minimum number of iterations for each Markov chain

40
 Examples: Regression And Path Analysis

Monte Carlo (MCMC) chain when the potential scale reduction (PSR)
convergence criterion (Gelman & Rubin, 1992) is used. Using a number
in parentheses, the BITERATIONS option specifies that a minimum of
30,000 and a maximum of the default of 50,000 iterations will be used.
The large minimum value is chosen to obtain a smooth plot.

In the MODEL command, the first ON statement describes the linear

regression of y on the mediator m. The second ON statement describes
the linear regression of the mediator m on the covariate x, the moderator
z, and the interaction xz. The intercepts and residual variances of y and
m are estimated and the residuals are not correlated as the default.

In MODEL CONSTRAINT, the LOOP option is used in conjunction

with the PLOT option to create plots of variables. In this example, the
indirect effect defined in MODEL CONSTRAINT will be plotted. The
PLOT option names the variable that will be plotted on the y-axis. The
LOOP option names the variable that will be plotted on the x-axis, gives
the numbers that are the lower and upper values of the variable, and the
incremental value of the variable to be used in the computations. In this
example, the variable indirect will be on the y-axis and the variable mod
will be on the x-axis. The variable mod, as in moderation, varies over
the range of z that is of interest such as two standard deviations away
from its mean. Corresponding to the case of z being standardized, the
lower and upper values of mod are -2 and 2 and 0.1 is the incremental
value of mod to use in the computations. When mod appears in a
MODEL CONSTRAINT statement involving a new parameter, that
statement is evaluated for each value of mod specified by the LOOP
option. For example, the first value of mod is -2; the second value of
mod is -2 plus 0.1 or -1.9; the third value of mod is -1.9 plus 0.1 or -1.8;
the last value of mod is 2.

Using TYPE=PLOT2 in the PLOT command, the plot of indirect and

mod can be viewed by choosing Loop plots from the Plot menu of the
Mplus Editor. The plot presents the computed values along with a 95%
confidence interval. For Bayesian estimation, the default is credibility
intervals of the posterior distribution with equal tail percentages. The
CINTERVAL option of the OUTPUT command can be used to obtain
credibility intervals of the posterior distribution that give the highest
posterior density. An explanation of the other commands can be found
in Example 3.1.

41
CHAPTER 3

42
 Examples: Exploratory Factor Analysis

CHAPTER 4
EXAMPLES: EXPLORATORY
FACTOR ANALYSIS

Exploratory factor analysis (EFA) is used to determine the number of

continuous latent variables that are needed to explain the correlations
among a set of observed variables. The continuous latent variables are
referred to as factors, and the observed variables are referred to as factor
indicators. In EFA, factor indicators can be continuous, censored,
binary, ordered categorical (ordinal), counts, or combinations of these
variable types. EFA can also be carried out using exploratory structural
equation modeling (ESEM; Asparouhov & Muthén, 2009a) when factor
indicators are continuous, censored, binary, ordered categorical
(ordinal), and combinations of these variable types. ESEM examples are
shown under Confirmatory Factor Analysis.

Several rotations are available using both orthogonal and oblique

procedures. The algorithms used in the rotations are described in
Jennrich and Sampson (1966), Browne (2001), Bernaards and Jennrich
(2005), Browne et al. (2004), and Jennrich and Bentler (2011, 2012).
Standard errors for the rotated solutions are available using algorithms
described in Jennrich (1973, 1974, 2007). Cudeck and O’Dell (1994)
discuss the benefits of standard errors for rotated solutions.

All EFA models can be estimated using the following special features:

• Missing data
• Complex survey data
• Mixture modeling

The default is to estimate the model under missing data theory using all
available data. The LISTWISE option of the DATA command can be
used to delete all observations from the analysis that have missing values
on one or more of the analysis variables. Corrections to the standard
errors and chi-square test of model fit that take into account
stratification, non-independence of observations, and unequal probability
of selection are obtained by using the TYPE=COMPLEX option of the
ANALYSIS command in conjunction with the STRATIFICATION,

43
CHAPTER 4

CLUSTER, and WEIGHT options of the VARIABLE command. The

SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of eigenvalues, individual observed and estimated
values, and plots of sample and estimated means and
proportions/probabilities. These are available for the total sample, by
group, by class, and adjusted for covariates. The PLOT command
includes a display showing a set of descriptive statistics for each
variable. The graphical displays can be edited and exported as a DIB,
EMF, or JPEG file. In addition, the data for each graphical display can
be saved in an external file for use by another graphics program.

Following is the set of EFA examples included in this chapter.

• 4.1: Exploratory factor analysis with continuous factor indicators

• 4.2: Exploratory factor analysis with categorical factor indicators
• 4.3: Exploratory factor analysis with continuous, censored,
categorical, and count factor indicators*
• 4.4: Exploratory factor mixture analysis with continuous latent class
indicators
• 4.5: Two-level exploratory factor analysis with continuous factor
indicators
• 4.6: Two-level exploratory factor analysis with both individual- and
cluster-level factor indicators
• 4.7: Bi-factor exploratory factor analysis with continuous factor
indicators

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

44
 Examples: Exploratory Factor Analysis

EXAMPLE 4.1: EXPLORATORY FACTOR ANALYSIS WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of an exploratory

factor analysis with continuous factor
indicators
DATA: FILE IS ex4.1a.dat;
VARIABLE: NAMES ARE y1-y12;
ANALYSIS: TYPE = EFA 1 4;
OUTPUT: MODINDICES;

In the first part of this example, an exploratory factor analysis with

continuous factor indicators is carried out. Rotated solutions with
standard errors are obtained for each number of factors. Modification
indices are requested for the residual correlations. In the second part of
this example, the same exploratory factor analysis for four factors is
carried out using exploratory structural equation modeling (ESEM).

TITLE: this is an example of an exploratory

factor analysis with continuous factor
indicators

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex4.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex4.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y1-y12;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains 12 variables: y1, y2, y3, y4, y5, y6, y7, y8, y9, y10,
y11, and y12. Note that the hyphen can be used as a convenience feature
in order to generate a list of names.

45
CHAPTER 4

ANALYSIS: TYPE = EFA 1 4;

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis that
is to be performed. By specifying TYPE=EFA, an exploratory factor
analysis will be carried out. The numbers following EFA give the lower
and upper limits on the number of factors to be extracted. The default
rotation is the oblique rotation of GEOMIN. The ROTATION option of
the ANALYSIS command can be used to select a different rotation. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

OUTPUT: MODINDICES;

The MODINDICES option is used with EFA to request modification

indices and expected parameter change indices for the residual
correlations which are fixed at zero in EFA.

TITLE: this is an example of an exploratory

factor analysis with continuous factor
indicators using exploratory structural
equation modeling (ESEM)
DATA: FILE IS ex4.1b.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1-f4 BY y1-y12 (*1);
OUTPUT: MODINDICES;

The difference between this part of the example and the first part is that
an exploratory factor analysis for four factors is carried out using
exploratory structural equation modeling (ESEM). In the MODEL
command, the BY statement specifies that the factors f1 through f4 are
measured by the continuous factor indicators y1 through y12. The label
1 following an asterisk (*) in parentheses following the BY statement is
used to indicate that f1, f2, f3, and f4 are a set of EFA factors. When no
rotation is specified using the ROTATION option of the ANALYSIS
command, the default oblique GEOMIN rotation is used. The intercepts
and residual variances of the factor indicators are estimated and the
residuals are not correlated as the default. The variances of the factors
are fixed at one as the default. The factors are correlated under the
default oblique GEOMIN rotation. The results are the same as for the
four-factor EFA in the first part of the example.

46
 Examples: Exploratory Factor Analysis

EXAMPLE 4.2: EXPLORATORY FACTOR ANALYSIS WITH

CATEGORICAL FACTOR INDICATORS

TITLE: this is an example of an exploratory

factor analysis with categorical factor
indicators
DATA: FILE IS ex4.2.dat;
VARIABLE: NAMES ARE u1-u12;
CATEGORICAL ARE u1-u12;
ANALYSIS: TYPE = EFA 1 4;

The difference between this example and Example 4.1 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. Estimation of factor analysis models with binary
variables is discussed in Muthén (1978) and Muthén et al. (1997). The
CATEGORICAL option is used to specify which dependent variables
are treated as binary or ordered categorical (ordinal) variables in the
model and its estimation. In the example above, all twelve factor
indicators are binary or ordered categorical variables. Categorical
variables can be binary or ordered categorical. The program determines
the number of categories for each variable. The default estimator for
this type of analysis is a robust weighted least squares estimator. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. With maximum likelihood estimation, numerical
integration is used with one dimension of integration for each factor. To
reduce computational time with several factors, the number of
integration points per dimension can be reduced from the default of 7 for
exploratory factor analysis to as few as 3 for an approximate solution.
An explanation of the other commands can be found in Example 4.1.

47
CHAPTER 4

EXAMPLE 4.3: EXPLORATORY FACTOR ANALYSIS WITH

CONTINUOUS, CENSORED, CATEGORICAL, AND COUNT
FACTOR INDICATORS

TITLE: this is an example of an exploratory

factor analysis with continuous, censored,
categorical, and count factor indicators
DATA: FILE = ex4.3.dat;
VARIABLE: NAMES = u4-u6 y4-y6 u1-u3 y1-y3;
CENSORED = y4-y6(b);
CATEGORICAL = u1-u3;
COUNT = u4-u6;
ANALYSIS: TYPE = EFA 1 4;

The difference between this example and Example 4.1 is that the factor
indicators are a combination of continuous, censored, binary or ordered
categorical (ordinal), and count variables instead of all continuous
variables. The CENSORED option is used to specify which dependent
variables are treated as censored variables in the model and its
estimation, whether they are censored from above or below, and whether
a censored or censored-inflated model will be estimated. In the example
above, y4, y5, and y6 are censored variables. The b in parentheses
indicates that they are censored from below, that is, have a floor effect,
and that the model is a censored regression model. The censoring limit
is determined from the data. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the
example above, the factor indicators u1, u2, and u3 are binary or ordered
categorical variables. The program determines the number of categories
for each variable. The COUNT option is used to specify which
dependent variables are treated as count variables in the model and its
estimation and whether a Poisson or zero-inflated Poisson model will be
estimated. In the example above, u4, u5, and u6 are count variables.
The variables y1, y2, and y3 are continuous variables.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, the four-factor solution requires four

48
 Examples: Exploratory Factor Analysis

dimensions of integration. Using the default of 7 integration points per

factor for exploratory factor analysis, a total of 2,401 integration points
is required for this analysis. To reduce computational time with several
factors, the number of integration points per dimension can be reduced
from the default of 7 for exploratory factor analysis to as few as 3 for an
approximate solution. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 4.1.

EXAMPLE 4.4: EXPLORATORY FACTOR MIXTURE

ANALYSIS WITH CONTINUOUS LATENT CLASS
INDICATORS

TITLE: this is an example of an exploratory

factor mixture analysis with continuous
latent class indicators
DATA: FILE = ex4.4.dat;
VARIABLE: NAMES = y1-y8;
CLASSES = c(2);
ANALYSIS: TYPE = MIXTURE EFA 1 2;

In this example, an exploratory factor mixture analysis with continuous

latent class indicators is carried out. Factor mixture analysis uses a
combination of categorical and continuous latent variables. Mixture
modeling refers to modeling with categorical latent variables that
represent subpopulations where population membership is not known
but is inferred from the data. With continuous latent class indicators, the
means of the latent class indicators vary across the classes as the default.
The continuous latent variables describe within-class correlations among
the latent class indicators. The within-class correlations follow an
exploratory factor analysis model that varies across the latent classes.
This is the mixtures of factor analyzers model discussed in McLachlan
and Peel (2000) and McLachlan et al. (2004). Rotated solutions with
standard errors are obtained for each latent class. See Example 7.27 for
a confirmatory factor mixture analysis.

The CLASSES option is used to assign names to the categorical latent

variables in the model and to specify the number of latent classes in the
model for each categorical latent variable. In the example above, there
is one categorical latent variable c that has two latent classes. The

49
CHAPTER 4

ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis that
is to be performed. By specifying TYPE=MIXTURE EFA, an
exploratory factor mixture analysis will be carried out. The numbers
following EFA give the lower and upper limits on the number of factors
to be extracted. The default rotation is the oblique rotation of GEOMIN.
The ROTATION option of the ANALYSIS command can be used to
select a different rotation. The default estimator for this type of analysis
is maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 4.1.

EXAMPLE 4.5: TWO-LEVEL EXPLORATORY FACTOR

ANALYSIS WITH CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a two-level

exploratory factor analysis with
continuous factor indicators
DATA: FILE IS ex4.5.dat;
VARIABLE: NAMES ARE y1-y6 x1 x2 w clus;
USEVARIABLES = y1-y6;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL EFA 1 2 UW 1 1 UB;

In this example, a two-level exploratory factor analysis model with

individual-level continuous factor indicators is carried out. Two-level
analysis models non-independence of observations due to cluster
sampling. An exploratory factor analysis is specified for both the within
and between parts of the model. Rotated solutions with standard errors
are obtained for both the within and between parts of the model. See
Example 9.6 for a two-level confirmatory factor analysis.

The CLUSTER option is used to identify the variable that contains

clustering information. The ANALYSIS command is used to describe
the technical details of the analysis. The TYPE option is used to
describe the type of analysis that is to be performed. By specifying
TYPE=TWOLEVEL EFA, a two-level exploratory factor analysis will
be carried out. The numbers following EFA give the lower and upper
limits on the number of factors to be extracted. The first set of numbers
are for the within part of the model. The second set of numbers are for

50
 Examples: Exploratory Factor Analysis

the between part of the model. In both parts of the model, one- and two-
factors solutions and an unrestricted solution will be obtained. The
unrestricted solution for the within part of the model is specified by UW
and the unrestricted solution for the between part of the model is
specified by UB. The within and between specifications are crossed.
Factor solutions will be obtained for one factor within and one factor
between, two factors within and one factor between, unrestricted within
and one factor between, one factor within and unrestricted between, and
two factors within and unrestricted between. Rotations are not given for
unrestricted solutions. The default rotation is the oblique rotation of
GEOMIN. The ROTATION option of the ANALYSIS command can be
used to select a different rotation. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 4.1.

EXAMPLE 4.6: TWO-LEVEL EXPLORATORY FACTOR

ANALYSIS WITH BOTH INDIVIDUAL- AND CLUSTER-
LEVEL FACTOR INDICATORS

TITLE: this is an example of a two-level

exploratory factor analysis with both
individual- and cluster-level factor
indicators
DATA: FILE = ex4.6.dat;
VARIABLE: NAMES = u1-u6 y1-y4 x1 x2 w clus;
USEVARIABLES = u1-u6 y1-y4;
CATEGORICAL = u1-u6;
CLUSTER = clus;
BETWEEN = y1-y4;
ANALYSIS: TYPE = TWOLEVEL EFA 1 2 UW 1 2 UB;
SAVEDATA: SWMATRIX = ex4.6sw.dat;

The difference between this example and Example 4.5 is that there is a
combination of individual-level categorical factor indicators and
between-level continuous factor indicators. The exploratory factor
analysis structure for the within part of the model includes only the
individual-level factor indicators whereas the exploratory factor analysis
structure for the between part of the model includes the between part of
the individual-level factor indicators and the between-level factor

51
CHAPTER 4

indicators. Rotated solutions with standard errors are obtained for both
the within and between parts of the model.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level and modeled only on the between
level. Variables not mentioned on the WITHIN or the BETWEEN
statements are measured on the individual level and can be modeled on
both the within and between levels. The default rotation is the oblique
rotation of GEOMIN. The ROTATION option of the ANALYSIS
command can be used to select a different rotation. The default
estimator for this type of analysis is a robust weighted least squares
estimator using a diagonal weight matrix (Asparouhov & Muthén, 2007).
The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. The SWMATRIX option of the
SAVEDATA command is used with TYPE=TWOLEVEL and weighted
least squares estimation to specify the name and location of the file that
contains the within- and between-level sample statistics and their
corresponding estimated asymptotic covariance matrix. It is
recommended to save this information and use it in subsequent analyses
along with the raw data to reduce computational time during model
estimation. An explanation of the other commands can be found in
Examples 4.1, 4.3, and 4.5.

EXAMPLE 4.7: BI-FACTOR EXPLORATORY FACTOR

ANALYSIS WITH CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a bi-factor

exploratory factor analysis with
continuous factor indicators
DATA: FILE = ex4.7.dat;
VARIABLE: NAMES = y1-y10;
ANALYSIS: TYPE = EFA 2 3;
ROTATION = BI-GEOMIN;

In this example, a bi-factor exploratory factor analysis (Jennrich &

Bentler, 2011, 2012) with continuous factor indicators is carried out
using a Geomin rotation. By specifying TYPE=EFA, an exploratory
factor analysis will be carried out. The number 2 is the lower limit and
the number 3 is the upper limit on the number of factors to be extracted.
By specifying BI-GEOMIN, a bi-factor EFA will be carried out using a

52
 Examples: Exploratory Factor Analysis

bi-factor Geomin rotation. Because this is a bi-factor analysis, the two-

factor solution will have one general factor and one specific factor. The
three-factor solution will have one general factor and two specific
factors. The default for the BI-GEOMIN rotation is an oblique rotation
where the specific factors are correlated with the general factor and are
correlated with each other. In the orthogonal rotation, the specific
factors are uncorrelated with the general factor and are uncorrelated with
each other. An orthogonal rotation is obtained by specifying
ROTATION=BI-GEOMIN(ORTHOGONAL). An alternative bi-factor
rotation can be obtained using the BI-CF-QUARTIMAX setting of the
ROTATION option. The default estimator for this type of analysis is
maximum likelihood. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 4.1.

53
CHAPTER 4

54
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

CHAPTER 5
EXAMPLES: CONFIRMATORY
FACTOR ANALYSIS AND
STRUCTURAL EQUATION
MODELING

Confirmatory factor analysis (CFA) is used to study the relationships

between a set of observed variables and a set of continuous latent
variables. When the observed variables are categorical, CFA is also
referred to as item response theory (IRT) analysis (Baker & Kim, 2004;
du Toit, 2003; Fox, 2010). CFA with covariates (MIMIC) includes
models where the relationship between factors and a set of covariates are
studied to understand measurement invariance and population
heterogeneity. These models can include direct effects, that is, the
regression of a factor indicator on a covariate in order to study
measurement non-invariance. Structural equation modeling (SEM)
includes models in which regressions among the continuous latent
variables are estimated (Bollen, 1989; Browne & Arminger, 1995;
Joreskog & Sorbom, 1979). In all of these models, the latent variables
are continuous. Observed dependent variable variables can be
continuous, censored, binary, ordered categorical (ordinal), unordered
categorical (nominal), counts, or combinations of these variable types.

CFA is a measurement model. SEM has two parts: a measurement

model and a structural model. The measurement model for both CFA
and SEM is a multivariate regression model that describes the
relationships between a set of observed dependent variables and a set of
continuous latent variables. The observed dependent variables are
referred to as factor indicators and the continuous latent variables are
referred to as factors. The relationships are described by a set of linear
regression equations for continuous factor indicators, a set of censored
normal or censored-inflated normal regression equations for censored
factor indicators, a set of probit or logistic regression equations for
binary or ordered categorical factor indicators, a set of multinomial
logistic regression equations for unordered categorical factor indicators,

55
CHAPTER 5

and a set of Poisson or zero-inflated Poisson regression equations for

count factor indicators.

The structural model describes three types of relationships in one set of

multivariate regression equations: the relationships among factors, the
relationships among observed variables, and the relationships between
factors and observed variables that are not factor indicators. These
relationships are described by a set of linear regression equations for the
factors that are dependent variables and for continuous observed
dependent variables, a set of censored normal or censored-inflated
normal regression equations for censored observed dependent variables,
a set of probit or logistic regression equations for binary or ordered
categorical observed dependent variables, a set of multinomial logistic
regression equations for unordered categorical observed dependent
variables, and a set of Poisson or zero-inflated Poisson regression
equations for count observed dependent variables. For logistic
regression, ordered categorical variables are modeled using the
proportional odds specification. Both maximum likelihood and weighted
least squares estimators are available.

All CFA, MIMIC and SEM models can be estimated using the following
special features:

• Single or multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities

For continuous, censored with weighted least squares estimation, binary,

and ordered categorical (ordinal) outcomes, multiple group analysis is
specified by using the GROUPING option of the VARIABLE command
for individual data or the NGROUPS option of the DATA command for
summary data. For censored with maximum likelihood estimation,
unordered categorical (nominal), and count outcomes, multiple group

56
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

analysis is specified using the KNOWNCLASS option of the

VARIABLE command in conjunction with the TYPE=MIXTURE
option of the ANALYSIS command. The default is to estimate the
model under missing data theory using all available data. The
LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. The
SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed. Latent variable interactions
are specified by using the | symbol of the MODEL command in
conjunction with the XWITH option of the MODEL command. Random
slopes are specified by using the | symbol of the MODEL command in
conjunction with the ON option of the MODEL command. Linear and
non-linear parameter constraints are specified by using the MODEL
CONSTRAINT command. Indirect effects are specified by using the
MODEL INDIRECT command. Maximum likelihood estimation is
specified by using the ESTIMATOR option of the ANALYSIS
command. Bootstrap standard errors are obtained by using the
BOOTSTRAP option of the ANALYSIS command. Bootstrap
confidence intervals are obtained by using the BOOTSTRAP option of
the ANALYSIS command in conjunction with the CINTERVAL option
of the OUTPUT command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, plots of
sample and estimated means and proportions/probabilities, and plots of
item characteristic curves and information curves. These are available
for the total sample, by group, by class, and adjusted for covariates. The
PLOT command includes a display showing a set of descriptive statistics
for each variable. The graphical displays can be edited and exported as a
DIB, EMF, or JPEG file. In addition, the data for each graphical display
can be saved in an external file for use by another graphics program.

57
CHAPTER 5

Following is the set of CFA examples included in this chapter:

• 5.1: CFA with continuous factor indicators

• 5.2: CFA with categorical factor indicators
• 5.3: CFA with continuous and categorical factor indicators
• 5.4: CFA with censored and count factor indicators*
• 5.5: Two-parameter logistic item response theory (IRT) model*
• 5.6: Second-order factor analysis
• 5.7: Non-linear CFA*
• 5.8: CFA with covariates (MIMIC) with continuous factor
indicators
• 5.9: Mean structure CFA for continuous factor indicators
• 5.10: Threshold structure CFA for categorical factor indicators

Following is the set of SEM examples included in this chapter:

• 5.11: SEM with continuous factor indicators

• 5.12: SEM with continuous factor indicators and an indirect effect
for factors
• 5.13: SEM with continuous factor indicators and an interaction
between two factors*

Following is the set of multiple group examples included in this chapter:

• 5.14: Multiple group CFA with covariates (MIMIC) with

continuous factor indicators and no mean structure
• 5.15: Multiple group CFA with covariates (MIMIC) with
continuous factor indicators and a mean structure
• 5.16: Multiple group CFA with covariates (MIMIC) with
categorical factor indicators and a threshold structure
• 5.17: Multiple group CFA with covariates (MIMIC) with
categorical factor indicators and a threshold structure using the
Theta parameterization
• 5.18: Two-group twin model for continuous outcomes where factors
represent the ACE components
• 5.19: Two-group twin model for categorical outcomes where factors
represent the ACE components

58
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

Following is the set of examples included in this chapter that estimate

models with parameter constraints:

• 5.20: CFA with parameter constraints

• 5.21: Two-group twin model for continuous outcomes using
parameter constraints
• 5.22: Two-group twin model for categorical outcomes using
parameter constraints
• 5.23: QTL sibling model for a continuous outcome using parameter
constraints

Following is the set of exploratory structural equation modeling (ESEM)

examples included in this chapter:

• 5.24: EFA with covariates (MIMIC) with continuous factor

indicators and direct effects
• 5.25: SEM with EFA and CFA factors with continuous factor
indicators
• 5.26: EFA at two time points with factor loading invariance and
correlated residuals across time
• 5.27: Multiple-group EFA with continuous factor indicators
• 5.28: EFA with residual variances constrained to be greater than
zero
• 5.29: Bi-factor EFA using ESEM
• 5.30: Bi-factor EFA with two items loading on only the general
factor
• 5.31: Bayesian bi-factor CFA with two items loading on only the
general factor and cross-loadings with zero-mean and small-variance
priors
• 5.32: Bayesian MIMIC model with cross-loadings and direct effects
with zero-mean and small-variance priors
• 5.33: Bayesian multiple group model with approximate
measurement invariance using zero-mean and small-variance priors

• Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

59
CHAPTER 5

EXAMPLE 5.1: CFA WITH CONTINUOUS FACTOR

INDICATORS

TITLE: this is an example of a CFA with

continuous factor indicators
DATA: FILE IS ex5.1.dat;
VARIABLE: NAMES ARE y1-y6;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;

y1

f1 y2

y3

y4

f2 y5

y6

In this example, the confirmatory factor analysis (CFA) model with

continuous factor indicators shown in the picture above is estimated.
The model has two correlated factors that are each measured by three
continuous factor indicators.

60
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling
TITLE: this is an example of a CFA with
continuous factor indicators

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex5.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex5.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y1-y6;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains six variables: y1, y2, y3, y4, y5, y6. Note that the
hyphen can be used as a convenience feature in order to generate a list of
names.

MODEL: f1 BY y1-y3;
f2 BY y4-y6;

The MODEL command is used to describe the model to be estimated.

Here the two BY statements specify that f1 is measured by y1, y2, and
y3, and f2 is measured by y4, y5, and y6. The metric of the factors is set
automatically by the program by fixing the first factor loading in each
BY statement to 1. This option can be overridden. The intercepts and
residual variances of the factor indicators are estimated and the residuals
are not correlated as the default. The variances of the factors are
estimated as the default. The factors are correlated as the default
because they are independent (exogenous) variables. The default
estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

61
CHAPTER 5

EXAMPLE 5.2: CFA WITH CATEGORICAL FACTOR

INDICATORS

TITLE: this is an example of a CFA with

categorical factor indicators
DATA: FILE IS ex5.2.dat;
VARIABLE: NAMES ARE u1-u6;
CATEGORICAL ARE u1-u6;
MODEL: f1 BY u1-u3;
f2 BY u4-u6;

The difference between this example and Example 5.1 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. The CATEGORICAL option is used to specify
which dependent variables are treated as binary or ordered categorical
(ordinal) variables in the model and its estimation. In the example
above, all six factor indicators are binary or ordered categorical
variables. The program determines the number of categories for each
factor indicator. The default estimator for this type of analysis is a
robust weighted least squares estimator (Muthén, 1984; Muthén, du Toit,
& Spisic, 1997). With this estimator, probit regressions for the factor
indicators regressed on the factors are estimated. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 5.1.

With maximum likelihood estimation, logistic regressions for the factor

indicators regressed on the factors are estimated using a numerical
integration algorithm. This is shown in Example 5.5. Note that
numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase.

62
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.3: CFA WITH CONTINUOUS AND

CATEGORICAL FACTOR INDICATORS

TITLE: this is an example of a CFA with

continuous and categorical factor
indicators
DATA: FILE IS ex5.3.dat;
VARIABLE: NAMES ARE u1-u3 y4-y6;
CATEGORICAL ARE u1 u2 u3;
MODEL: f1 BY u1-u3;
f2 BY y4-y6;

The difference between this example and Example 5.1 is that the factor
indicators are a combination of binary or ordered categorical (ordinal)
and continuous variables instead of all continuous variables. The
CATEGORICAL option is used to specify which dependent variables
are treated as binary or ordered categorical (ordinal) variables in the
model and its estimation. In the example above, the factor indicators u1,
u2, and u3 are binary or ordered categorical variables whereas the factor
indicators y4, y5, and y6 are continuous variables. The program
determines the number of categories for each factor indicator. The
default estimator for this type of analysis is a robust weighted least
squares estimator. With this estimator, probit regressions are estimated
for the categorical factor indicators, and linear regressions are estimated
for the continuous factor indicators. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. With
maximum likelihood estimation, logistic regressions are estimated for
the categorical dependent variables using a numerical integration
algorithm. Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. An explanation of the other commands can be found in
Example 5.1.

63
CHAPTER 5

EXAMPLE 5.4: CFA WITH CENSORED AND COUNT FACTOR

INDICATORS

TITLE: this is an example of a CFA with censored

and count factor indicators
DATA: FILE IS ex5.4.dat;
VARIABLE: NAMES ARE y1-y3 u4-u6;
CENSORED ARE y1-y3 (a);
COUNT ARE u4-u6;
MODEL: f1 BY y1-y3;
f2 BY u4-u6;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 5.1 is that the factor
indicators are a combination of censored and count variables instead of
all continuous variables. The CENSORED option is used to specify
which dependent variables are treated as censored variables in the model
and its estimation, whether they are censored from above or below, and
whether a censored or censored-inflated model will be estimated. In the
example above, y1, y2, and y3 are censored variables. The a in
parentheses following y1-y3 indicates that y1, y2, and y3 are censored
from above, that is, have ceiling effects, and that the model is a censored
regression model. The censoring limit is determined from the data. The
COUNT option is used to specify which dependent variables are treated
as count variables in the model and its estimation and whether a Poisson
or zero-inflated Poisson model will be estimated. In the example above,
u4, u5, and u6 are count variables. Poisson regressions are estimated for
the count dependent variables and censored regressions are estimated for
the censored dependent variables.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. The
OUTPUT command is used to request additional output not included as
the default. The TECH1 option is used to request the arrays containing
parameter specifications and starting values for all free parameters in the
model. The TECH8 option is used to request that the optimization

64
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

history in estimating the model be printed in the output. TECH8 is

printed to the screen during the computations as the default. TECH8
screen printing is useful for determining how long the analysis takes. An
explanation of the other commands can be found in Example 5.1.

EXAMPLE 5.5: TWO-PARAMETER LOGISTIC ITEM

RESPONSE THEORY (IRT) MODEL

TITLE: this is an example of a two-parameter

logistic item response theory (IRT) model
DATA: FILE IS ex5.5.dat;
VARIABLE: NAMES ARE u1-u20;
CATEGORICAL ARE u1-u20;
ANALYSIS: ESTIMATOR = MLR;
MODEL: f BY u1-u20*;
f@1;
OUTPUT: TECH1 TECH8;
PLOT: TYPE = PLOT3;

In this example, a logistic IRT model is estimated. With binary factor

indicators, this is referred to as a two-parameter logistic model. With
ordered categorical (ordinal) factor indicators, this is referred to as
Samejima’s graded response model (Baker & Kim, 2004; du Toit, 2003).
A single continuous factor is measured by 20 categorical factor
indicators.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, the factor
indicators u1 through u20 are binary or ordered categorical variables.
The program determines the number of categories for each factor
indicator. By specifying ESTIMATOR=MLR, a maximum likelihood
estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
of the ANALYSIS command can be used to select a different estimator.

In the MODEL command, the BY statement specifies that f is measured

by u1 through u20. The asterisk (*) frees the first factor loading which

65
CHAPTER 5

is fixed at one as the default to define the metric of the factor. Instead
the metric of the factor is defined by fixing the factor variance at one in
line with IRT. For one-factor models with no covariates, results are
presented both in a factor model parameterization and in a conventional
IRT parameterization. The OUTPUT command is used to request
additional output not included as the default. The TECH1 option is used
to request the arrays containing parameter specifications and starting
values for all free parameters in the model. The TECH8 option is used
to request that the optimization history in estimating the model be
printed in the output. TECH8 is printed to the screen during the
computations as the default. TECH8 screen printing is useful for
determining how long the analysis takes. The PLOT command is used to
request graphical displays of observed data and analysis results. These
graphical displays can be viewed after the analysis is completed using a
post-processing graphics module. Item characteristic curves and
information curves are available. When covariates are included in the
model with direct effects on one or more factor indicators, item
characteristic curves can be plotted for each value of the covariate to
show differential item functioning (DIF). An explanation of the other
commands can be found in Example 5.1.

EXAMPLE 5.6: SECOND-ORDER FACTOR ANALYSIS

TITLE: this is an example of a second-order

factor analysis
DATA: FILE IS ex5.6.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f5 BY f1-f4;

66
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12

f1 f2 f3 f4

f5

In this example, the second-order factor analysis model shown in the

figure above is estimated. The factor indicators of the first-order factors
f1, f2, f3, and f4 are continuous. The first-order factors are indicators of
the second-order factor f5.

The first four BY statements specify that f1 is measured by y1, y2, and
y3; f2 is measured by y4, y5, and y6; f3 is measured by y7, y8, and y9;
and f4 is measured by y10, y11, and y12. The fifth BY statement
specifies that the second-order factor f5 is measured by f1, f2, f3, and f4.
The metrics of the first- and second-order factors are set automatically
by the program by fixing the first factor loading in each BY statement to
1. This option can be overridden. The intercepts and residual variances
of the first-order factor indicators are estimated and the residuals are not
correlated as the default. The residual variances of the first-order factors
are estimated as the default. The residuals of the first-order factors are
not correlated as the default. The variance of the second-order factor is
estimated as the default. The default estimator for this type of analysis
is maximum likelihood. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 5.1.

67
CHAPTER 5

EXAMPLE 5.7: NON-LINEAR CFA

TITLE: this is an example of a non-linear CFA

DATA: FILE IS ex5.7.dat;
VARIABLE: NAMES ARE y1-y5;
ANALYSIS: TYPE = RANDOM;
ALGORITHM = INTEGRATION;
MODEL: f BY y1-y5;
fxf | f XWITH f;
y1-y5 ON fxf;
OUTPUT: TECH1 TECH8;

In this example, a non-linear CFA model is estimated (McDonald, 1967).

The factor indicators are quadratic functions of the factor. The TYPE
option is used to describe the type of analysis that is to be performed.
By selecting RANDOM, a model with a random effect will be estimated.
By specifying ALGORITHM=INTEGRATION, a maximum likelihood
estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
of the ANALYSIS command can be used to select a different estimator.

The BY statement specifies that f is measured by y1 through y5. This

specifies the linear part of the quadratic function. The | statement in
conjunction with the XWITH option of the MODEL command is used to
define the quadratic factor term. The name on the left-hand side of the |
symbol names the quadratic factor term. The XWITH statement on the
right-hand side of the | symbol defines the quadratic factor term fxf. The
ON statement specifies the quadratic part of the quadratic function. The
OUTPUT command is used to request additional output not included as
the default. The TECH1 option is used to request the arrays containing
parameter specifications and starting values for all free parameters in the
model. The TECH8 option is used to request that the optimization
history in estimating the model be printed in the output. TECH8 is
printed to the screen during the computations as the default. TECH8
screen printing is useful for determining how long the analysis takes. An
explanation of the other commands can be found in Example 5.1.

68
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.8: CFA WITH COVARIATES (MIMIC) WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a CFA with

covariates (MIMIC) with continuous factor
indicators
DATA: FILE IS ex5.8.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;

y1

x1 f1 y2

y3
x2

y4

x3 f2 y5

y6

In this example, the CFA model with covariates (MIMIC) shown in the
figure above is estimated. The two factors are regressed on three
covariates.

69
CHAPTER 5

The first BY statement specifies that f1 is measured by y1, y2, and y3.
The second BY statement specifies that f2 is measured by y4, y5, and y6.
The metric of the factors is set automatically by the program by fixing
the first factor loading in each BY statement to 1. This option can be
overridden. The intercepts and residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The residual variances of the factors are estimated as the
default. The residuals of the factors are correlated as the default because
residuals are correlated for latent variables that do not influence any
other variable in the model except their own indicators. The ON
statement describes the linear regressions of f1 and f2 on the covariates
x1, x2, and x3. The ESTIMATOR option of the ANALYSIS command
can be used to select a different estimator. An explanation of the other
commands can be found in Example 5.1.

EXAMPLE 5.9: MEAN STRUCTURE CFA FOR CONTINUOUS

FACTOR INDICATORS

TITLE: this is an example of a mean structure CFA

for continuous factor indicators
DATA: FILE IS ex5.9.dat;
VARIABLE: NAMES ARE y1a-y1c y2a-y2c;
MODEL: f1 BY y1a y1b@1 y1c@1;
f2 BY y2a y2b@1 y2c@1;
[y1a y1b y1c] (1);
[y2a y2b y2c] (2);

70
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

y1a

f1 y1b

y1c

y2a

f2 y2b

y2c

In this example, the CFA model in which two factors are measured by
three equivalent tests forms shown in the picture above is estimated.
The three equivalent test forms are referred to as a, b, and c.

The first BY statement specifies that f1 is measured by y1a, y1b, and

y1c. The second BY statement specifies that f2 is measured by y2a, y2b,
and y2c. The letters a, b, and c are used to represent three equivalent test
forms, and 1 and 2 represent two different topics. The metric of the
factors is set automatically by the program by fixing the first factor
loading in each BY statement to 1. This option can be overridden. The
second and third factor loadings for both factors are fixed at one using
the @ option to reflect the hypothesis that the two test forms are
equivalent. The intercepts and residual variances of the factor indicators
are estimated and the residuals are not correlated as the default. The
variances of the factors are estimated as the default. The covariance
between f1 and f2 is estimated as the default because f1 and f2 are
independent (exogenous) variables.

To reflect the hypothesis that the three test forms are equivalent with
respect to their measurement intercepts, the first bracket statement
specifies that the intercepts for y1a, y1b, and y1c are equal and the

71
CHAPTER 5

second bracket statement specifies that the intercepts for y2a, y2b, and
y2c are equal. Equalities are designated by a number in parentheses. All
parameters in a statement followed by the same number in parentheses
are held equal. The means of the two factors are fixed at zero as the
default. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 5.1.

EXAMPLE 5.10: THRESHOLD STRUCTURE CFA FOR

CATEGORICAL FACTOR INDICATORS

TITLE: this is an example of a threshold

structure CFA for categorical factor
indicators
DATA: FILE IS ex5.10.dat;
VARIABLE: NAMES ARE u1a-u1c u2a-u2c;
CATEGORICAL ARE u1a-u1c u2a-u2c;
MODEL: f1 BY u1a u1b@1 u1c@1;
f2 BY u2a u2b@1 u2c@1;
[u1a$1 u1b$1 u1c$1] (1);
[u2a$1 u2b$1 u2c$1] (2);

The difference between this example and Example 5.9 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. The CATEGORICAL option is used to specify
which dependent variables are treated as binary or ordered categorical
(ordinal) variables in the model and its estimation. In the example
above, all six factor indicators are binary or ordered categorical
variables. The program determines the number of categories for each
factor indicator. In this example, it is assumed that the factor indicators
are binary variables with one threshold each.

For binary and ordered categorical factor indicators, thresholds are

modeled rather than intercepts or means. The number of thresholds for a
categorical variable is equal to the number of categories minus one. In
the example above, the categorical variables are binary so they have one
threshold. Thresholds are referred to by adding to the variable name a $
followed by a number. The thresholds of the factor indicators are
referred to as u1a$1, u1b$1, u1c$1, u2a$1, u2b$1, and u2c$1.
Thresholds are referred to in square brackets. To reflect the hypothesis

72
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

that the three test forms are equivalent with respect to their measurement
thresholds, the (1) after the first bracket statement specifies that the
thresholds for u1a, u1b, and u1c are constrained to be equal and the (2)
after the second bracket statement specifies that the thresholds for u2a,
u2b, and u2c are constrained to be equal. The default estimator for this
type of analysis is a robust weighted least squares estimator. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. With maximum likelihood, logistic regressions are
estimated using a numerical integration algorithm. Note that numerical
integration becomes increasingly more computationally demanding as
the number of factors and the sample size increase. An explanation of
the other commands can be found in Examples 5.1 and 5.9.

EXAMPLE 5.11: SEM WITH CONTINUOUS FACTOR

INDICATORS

TITLE: this is an example of a SEM with

continuous factor indicators
DATA: FILE IS ex5.11.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f4 ON f3;
f3 ON f1 f2;

73
CHAPTER 5

y7 y8 y9 y10 y11 y12

y1
f1
y2

y3
f3 f4
y4

y5
f2
y6

In this example, the SEM model with four continuous latent variables
shown in the picture above is estimated. The factor indicators are
continuous variables.

The first BY statement specifies that f1 is measured by y1, y2 and y3.

The second BY statement specifies that f2 is measured by y4, y5, and y6.
The third BY statement specifies that f3 is measured by y7, y8, and y9.
The fourth BY statement specifies that f4 is measured by y10, y11, and
y12. The metric of the factors is set automatically by the program by
fixing the first factor loading in each BY statement to 1. This option can
be overridden. The intercepts and residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The variances of the factors are estimated as the default. The
covariance between f1 and f2 is estimated as the default because f1 and
f2 are independent (exogenous) variables. The other factor covariances
are not estimated as the default.

The first ON statement describes the linear regression of f4 on f3. The

second ON statement describes the linear regression of f3 on f1 and f2.
The default estimator for this type of analysis is maximum likelihood.
The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. An explanation of the other commands can
be found in Example 5.1.

74
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.12: SEM WITH CONTINUOUS FACTOR

INDICATORS AND AN INDIRECT EFFECT FOR FACTORS

TITLE: this is an example of a SEM with

continuous factor indicators and an
indirect effect for factors
DATA: FILE IS ex5.12.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f4 ON f3;
f3 ON f1 f2;
MODEL INDIRECT:
f4 IND f3 f1;

The difference between this example and Example 5.11 is that an

indirect effect is estimated. Indirect effects and their standard errors can
be requested using the MODEL INDIRECT command. Total indirect,
specific indirect, and total effects are specified by using the IND and
VIA statements. Total effects include all indirect effects and the direct
effect. The IND statement is used to request a specific indirect effect or
set of indirect effects. The VIA statement is used to request a set of
indirect effects that include specific mediators.

In the IND statement above, the variable on the left-hand side of IND is
the dependent variable. The last variable on the right-hand side of IND
is the independent variable. Other variables on the right-hand side of
IND are mediating variables. The IND statement requests the specific
indirect effect from f1 to f3 to f4. The default estimator for this type of
analysis is maximum likelihood. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 5.1 and
5.11.

75
CHAPTER 5

EXAMPLE 5.13: SEM WITH CONTINUOUS FACTOR

INDICATORS AND AN INTERACTION BETWEEN TWO
LATENT VARIABLES

TITLE: this is an example of a SEM with

continuous factor indicators and an
interaction between two latent variables
DATA: FILE IS ex5.13.dat;
VARIABLE: NAMES ARE y1-y12;
ANALYSIS: TYPE = RANDOM;
ALGORITHM = INTEGRATION;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f4 ON f3;
f3 ON f1 f2;
f1xf2 | f1 XWITH f2;
f3 ON f1xf2;
OUTPUT: TECH1 TECH8;

y7 y8 y9 y10 y11 y12

y1
f1
y2

y3
f3 f4
y4

y5
f2
y6

The difference between this example and Example 5.11 is that an

interaction between two latent variables is included in the model. The

76
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

interaction is shown in the picture above as a filled circle. The model is

estimated using maximum likelihood (Klein & Moosbrugger, 2000).

The TYPE option is used to describe the type of analysis that is to be

performed. By selecting RANDOM, a model with a random effect will
be estimated. By specifying ALGORITHM=INTEGRATION, a
maximum likelihood estimator with robust standard errors using a
numerical integration algorithm will be used. Note that numerical
integration becomes increasingly more computationally demanding as
the number of factors and the sample size increase. In this example, two
dimensions of integration are used with a total of 225 integration points.
The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator.

Latent variable interactions are specified by using the | statement in

conjunction with the XWITH option of the MODEL command. The
name on the left-hand side of the | symbol names the latent variable
interaction. The XWITH statement on the right-hand side of the |
symbol defines the latent variable interaction. The latent variable f1xf2
is the interaction between f1 and f2. The last ON statement uses the
latent variable interaction as an independent variable. The OUTPUT
command is used to request additional output not included as the default.
The TECH1 option is used to request the arrays containing parameter
specifications and starting values for all free parameters in the model.
The TECH8 option is used to request that the optimization history in
estimating the model be printed in the output. TECH8 is printed to the
screen during the computations as the default. TECH8 screen printing is
useful for determining how long the analysis takes. An explanation of
the other commands can be found in Examples 5.1 and 5.11.

77
CHAPTER 5

EXAMPLE 5.14: MULTIPLE GROUP CFA WITH

COVARIATES (MIMIC) WITH CONTINUOUS FACTOR
INDICATORS AND NO MEAN STRUCTURE

TITLE: this is an example of a multiple group CFA

with covariates (MIMIC) with continuous
factor indicators and no mean structure
DATA: FILE IS ex5.14.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3 g;
GROUPING IS g (1 = male 2 = female);
ANALYSIS: MODEL = NOMEANSTRUCTURE;
INFORMATION = EXPECTED;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;
MODEL female:
f1 BY y3;

The difference between this example and Example 5.8 is that this is a
multiple group rather than a single group analysis. The GROUPING
option is used to identify the variable in the data set that contains
information on group membership when the data for all groups are
stored in a single data set. The information in parentheses after the
grouping variable name assigns labels to the values of the grouping
variable found in the data set. In the example above, observations with g
equal to 1 are assigned the label male, and individuals with g equal to 2
are assigned the label female. These labels are used in conjunction with
the MODEL command to specify model statements specific to each
group.

The NOMEANSTRUCTURE setting for the MODEL option of the

ANALYSIS command is used with TYPE=GENERAL to specify that
means, intercepts, and thresholds are not included in the analysis model.
As a result, a covariance structure model is estimated. The
INFORMATION option is used to select the estimator of the information
matrix to be used in computing standard errors when the ML or MLR
estimators are used for analysis. The default is the observed information
matrix. In this example, the expected information matrix is used in line
with conventional covariance structure analysis.

78
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

In multiple group analysis, two variations of the MODEL command are

used. They are MODEL and MODEL followed by a label. MODEL
describes the overall model to be estimated for each group. The factor
loading measurement parameters are held equal across groups as the
default to specify measurement invariance. MODEL followed by a label
describes differences between the overall model and the model for the
group designated by the label. In the group-specific MODEL command
for females, the factor loading for variable y3 and factor f1 is specified
to be free and not equal to the same factor loading for males. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 5.1 and 5.8.

EXAMPLE 5.15: MULTIPLE GROUP CFA WITH

COVARIATES (MIMIC) WITH CONTINUOUS FACTOR
INDICATORS AND A MEAN STRUCTURE

TITLE: this is an example of a multiple group CFA

with covariates (MIMIC) with continuous
factor indicators and a mean structure
DATA: FILE IS ex5.15.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3 g;
GROUPING IS g (1 = male 2 = female);
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;
MODEL female:
f1 BY y3;
[y3];

The difference between this example and Example 5.14 is that means are
included in the model. In multiple group analysis, when a model
includes a mean structure, both the intercepts and factor loadings of the
continuous factor indicators are held equal across groups as the default
to specify measurement invariance. The intercepts of the factors are
fixed at zero in the first group and are free to be estimated in the other
groups as the default. The group-specific MODEL command for
females specifies that the intercept of y3 for females is free and not
equal to the intercept for males. Intercepts are referred to by using
square brackets. The default estimator for this type of analysis is

79
CHAPTER 5

maximum likelihood. The ESTIMATOR option of the ANALYSIS

command can be used to select a different estimator. An explanation of
the other commands can be found in Examples 5.1, 5.8, and 5.14.

EXAMPLE 5.16: MULTIPLE GROUP CFA WITH

COVARIATES (MIMIC) WITH CATEGORICAL FACTOR
INDICATORS AND A THRESHOLD STRUCTURE

TITLE: this is an example of a multiple group CFA

with covariates (MIMIC) with categorical
factor indicators and a threshold
structure
DATA: FILE IS ex5.16.dat;
VARIABLE: NAMES ARE u1-u6 x1-x3 g;
CATEGORICAL ARE u1-u6;
GROUPING IS g (1 = male 2 = female);
MODEL: f1 BY u1-u3;
f2 BY u4-u6;
f1 f2 ON x1-x3;
MODEL female:
f1 BY u3;
[u3$1];
{u3@1};

The difference between this example and Example 5.15 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. For multiple-group CFA with categorical factor
indicators, see Muthén and Christoffersson (1981) and Muthén and
Asparouhov (2002).

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, all six factor
indicators are binary or ordered categorical variables. The program
determines the number of categories for each factor indicator.

For binary and ordered categorical factor indicators, thresholds are

modeled rather than intercepts or means. The number of thresholds for a
categorical variable is equal to the number of categories minus one. In
the above example, u3 is a binary variable with two categories.
Thresholds are referred to by adding to the variable name a $ followed
by a number. The threshold for u3 is u3$1. Thresholds are referred to in

80
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

square brackets. When a model includes a mean structure, the

thresholds of the factor indicators are held equal across groups as the
default to specify measurement invariance. In the group-specific
MODEL command for females, the threshold and factor loading of u3
for females are specified to be free and not equal to the threshold and
factor loading for males.

Because the factor indicators are categorical, scale factors are required
for multiple group analysis when the default Delta parameterization is
used. Scale factors are referred to using curly brackets ({}). By default,
scale factors are fixed at one in the first group and are free to be
estimated in the other groups. When a threshold and a factor loading for
a categorical factor indicator are free across groups, the scale factor for
that variable must be fixed at one in all groups for identification
purposes. Therefore, the scale factor for u3 is fixed at one for females.

The default estimator for this type of analysis is a robust weighted least
squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. With maximum
likelihood, logistic regressions are estimated using a numerical
integration algorithm. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. An explanation of the other commands can
be found in Examples 5.1, 5.8, 5.14, and 5.15.

81
CHAPTER 5

EXAMPLE 5.17: MULTIPLE GROUP CFA WITH

COVARIATES (MIMIC) WITH CATEGORICAL FACTOR
INDICATORS AND A THRESHOLD STRUCTURE USING THE
THETA PARAMETERIZATION

TITLE: this is an example of a multiple group CFA

with covariates (MIMIC) with categorical
factor indicators and a threshold
structure using the Theta parameterization
DATA: FILE IS ex5.17.dat;
VARIABLE: NAMES ARE u1-u6 x1-x3 g;
CATEGORICAL ARE u1-u6;
GROUPING IS g (1 = male 2 = female);
ANALYSIS: PARAMETERIZATION = THETA;
MODEL: f1 BY u1-u3;
f2 BY u4-u6;
f1 f2 ON x1-x3;
MODEL female:
f1 BY u3;
[u3$1];
u3@1;

The difference between this example and Example 5.16 is that the Theta
parameterization is used instead of the Delta parameterization. In the
Delta parameterization, scale factors are allowed to be parameters in the
model, but residual variances for latent response variables of observed
categorical dependent variables are not. In the alternative Theta
parameterization, residual variances for latent response variables are
allowed to be parameters in the model but scale factors are not. The
Theta parameterization is selected by specifying
PARAMETERIZATION=THETA in the ANALYSIS command.

When the Theta parameterization is used, the residual variances for the
latent response variables of the observed categorical dependent variables
are fixed at one in the first group and are free to be estimated in the other
groups as the default. When a threshold and a factor loading for a
categorical factor indicator are free across groups, the residual variance
for the variable must be fixed at one in these groups for identification
purposes. In the group-specific MODEL command for females, the
residual variance for u3 is fixed at one. An explanation of the other
commands can be found in Examples 5.1, 5.8, 5.14, 5.15, and 5.16.

82
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.18: TWO-GROUP TWIN MODEL FOR

CONTINUOUS OUTCOMES WHERE FACTORS REPRESENT
THE ACE COMPONENTS

TITLE: this is an example of a two-group twin

model for continuous outcomes where
factors represent the ACE components
DATA: FILE = ex5.18.dat;
VARIABLE: NAMES = y1 y2 g;
GROUPING = g (1 = mz 2 = dz);
ANALYSIS: MODEL = NOCOVARIANCES;
MODEL: [y1-y2] (1);
y1-y2@0;
a1 BY y1* (2);
a2 BY y2* (2);
c1 BY y1* (3);
c2 BY y2* (3);
e1 BY y1* (4);
e2 BY y2* (4);
a1-e2@1;
[a1-e2@0];
a1 WITH a2@1;
c1 WITH c2@1;
MODEL dz: a1 WITH a2@.5;

y1 y2

A1 C1 E1 A2 C2 E2

In this example, the univariate twin model shown in the picture above is
estimated. This is a two-group twin model for a continuous outcome
where factors represent the ACE components (Neale & Cardon, 1992).

83
CHAPTER 5

The variables y1 and y2 represent a univariate outcome for each member

of the twin pair. The A factors represent the additive genetic
components which correlate 1.0 for monozygotic twin pairs and 0.5 for
dizygotic twin pairs. The C factors represent common environmental
effects which correlate 1.0 for all twin pairs. The E factors represent
uncorrelated environmental effects. A simpler alternative way of
specifying this model is shown in Example 5.21 where parameter
constraints are used instead of the A, C, and E factors.

Exogenous factors are correlated as the default. By specifying

MODEL=NOCOVARIANCES in the ANALYSIS command, all
covariances in the model are fixed at zero. The WITH option of the
MODEL command can be used to override the default for selected
covariances as shown in the three WITH statements. In the MODEL
command, the (1) following the first bracket statement specifies that the
intercepts of y1 and y2 are held equal across twins. The second
statement fixes the residual variances of y1 and y2 to zero. The residual
variances of y1 and y2 are instead captured by the loadings of the E
factors. The six BY statements are used to define the six factors. The
asterisk (*) is used to free the factor loadings because the default is that
the factor loading for the first factor indicator is fixed at one. The
loadings for the A, C, and E factors are held equal across twins by
placing (2) following the two BY statements for the A factors, (3)
following the two BY statements for the C factors, and (4) following the
two BY statements for the E factors. In the next two statements, the A,
C, and E factor variances are fixed at one and the A, C, and E factor
means are fixed at zero. Because the factor means are fixed at zero, the
intercepts of y1 and y2 are their means.

The WITH statement for the A factors is used to fix the covariance
(correlation) between the A factors to 1.0 for monozygotic twin pairs.
The group-specific MODEL command is used to fix the covariance
between the A factors to 0.5 for the dizygotic twin pairs. The WITH
statement for the C factors is used to fix the covariance between the C
factors to 1. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Examples 5.1 and 5.14.

84
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.19: TWO-GROUP TWIN MODEL FOR

CATEGORICAL OUTCOMES WHERE FACTORS
REPRESENT THE ACE COMPONENTS

TITLE: this is an example of a two-group twin

model for categorical outcomes where
factors represent the ACE components
DATA: FILE = ex5.19.dat;
VARIABLE: NAMES = u1 u2 g;
CATEGORICAL = u1-u2;
GROUPING = g (1 = mz 2 = dz);
ANALYSIS: MODEL = NOCOVARIANCES;
MODEL: [u1$1-u2$1] (1);
a1 BY u1* (2);
a2 BY u2* (2);
c1 BY u1* (3);
c2 BY u2* (3);
a1-c2@1;
[a1-c2@0];
a1 WITH a2@1;
c1 WITH c2@1;
MODEL dz: a1 WITH a2@.5;
{u1-u2@1};

The difference between this example and Example 5.18 is that the
outcomes are binary or ordered categorical instead of continuous
variables. Because of this, the outcomes have no freely estimated
residual variances and therefore the E factors are not part of the model.
With categorical outcomes, the twin model is formulated for normally-
distributed latent response variables underlying the categorical outcomes
which are also called liabilities. This model is referred to as the
threshold model for liabilities (Neale & Cardon, 1992). More complex
examples of such models are given in Prescott (2004). A simpler
alternative way of specifying this model is shown in Example 5.22
where parameter constraints are used instead of the A and C factors.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, u1 and u2 are
binary or ordered categorical variables. The program determines the
number of categories for each variable.

85
CHAPTER 5

For binary and ordered categorical outcomes, thresholds are modeled

rather than intercepts or means. The number of thresholds for a
categorical variable is equal to the number of categories minus one. In
the example above, the categorical variables are binary so they have one
threshold. Thresholds are referred to by adding to the variable name a $
followed by a number. The thresholds of u1 and u2 are referred to as
u1$1 and u2$1. Thresholds are referred to in square brackets. The (1)
after the first bracket statement specifies that the thresholds for u1$1 and
u2$1 are constrained to be equal.

Because the outcomes are categorical, scale factors are required for
multiple group analysis when the default Delta parameterization is used.
Scale factors are referred to using curly brackets ({}). By default, scale
factors are fixed at one in the first group and are free to be estimated in
the other groups. In this model where the variance contributions from
the A and C factors are assumed equal across the two groups, the scale
factors are fixed at one in both groups to represent the equality of
variance for latent response variables underlying u1 and u2. The
statement in curly brackets in the group-specific MODEL command
specifies that the scale factors are fixed at one. The variance
contribution from the E factor is a remainder obtained by subtracting the
variance contributions of the A and C factors from the unit variance of
the latent response variables underlying u1 and u2. These are obtained
as part of the STANDARDIZED option of the OUTPUT command.

The default estimator for this type of analysis is a robust weighted least
squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. With maximum
likelihood and categorical factor indicators, numerical integration is
required. Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. An explanation of the other commands can be found in
Examples 5.1, 5.14, and 5.18.

86
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.20: CFA WITH PARAMETER CONSTRAINTS

TITLE: this is an example of a CFA with parameter

constraints
DATA: FILE = ex5.20.dat;
VARIABLE: NAMES = y1-y6;
MODEL: f1 BY y1
y2-y3(lam2-lam3);
f2 BY y4
y5-y6(lam5-lam6);
f1 (vf1);
f2 (vf2);
y1-y3 (ve1-ve3);
y4-y6 (ve4-ve6);
MODEL CONSTRAINT:
NEW(rel2 rel5 stan3 stan6);
rel2 = lam2**2*vf1/(lam2**2*vf1 + ve2);
rel5 = lam5**2*vf2/(lam5**2*vf2 + ve5);
rel5 = rel2;
stan3 = lam3*SQRT(vf1)/SQRT(lam3**2*vf1 +
ve3);
stan6 = lam6*SQRT(vf2)/SQRT(lam6**2*vf2 +
ve6);
0 = stan6 - stan3;
ve2 > ve5;
ve4 > 0;
OUTPUT: STANDARDIZED;

In this example, parameter constraints are used to estimate reliabilities,

estimate standardized coefficients, constrain functions of parameters to
be equal, and constrain parameters to be greater than a value. This
example uses the model from Example 5.1.

The MODEL CONSTRAINT command specifies parameter constraints

using labels defined for parameters in the MODEL command, labels
defined for parameters not in the MODEL command using the NEW
option of the MODEL CONSTRAINT command, and names of observed
variables that are identified using the CONSTRAINT option of the
VARIABLE command. This example illustrates constraints using labels
defined for parameters in the MODEL command and labels defined
using the NEW option. The NEW option is used to assign labels and
starting values to parameters not in the analysis model. Parameters in
the analysis model are given labels by placing a name in parentheses
after the parameter in the MODEL command.

87
CHAPTER 5

In the MODEL command, labels are defined for twelve parameters. The
list function can be used when assigning labels to a list of parameters.
The labels lam2, lam3, lam5, and lam6 are assigned to the factor
loadings for y2, y3, y5, and y6. The labels vf1 and vf2 are assigned to
the factor variances for f1 and f2. The labels ve1, ve2, ve3, ve4, ve5,
and ve6 are assigned to the residual variances of y1, y2, y3, y4, y5, and
y6.

In the MODEL CONSTRAINT command, the NEW option is used to

assign labels to four parameters that are not in the analysis model: rel2,
rel5, stan3, and stan6. The parameters rel2 and rel6 estimate the
reliability of y2 and y6 where reliability is defined as variance explained
divided by total variance. The parameters stan3 and stan6 estimate the
standardized coefficients for y3 and y6 using conventional
standardization formulas. In the statement that begins 0=, two
parameters are held equal to each other by defining their difference as
zero. In the last two statements, the residual variance of y2 is
constrained to be greater than the residual variance of y5, and the
residual variance of y4 is constrained to be greater than zero. The
STANDARDIZED option of the OUTPUT command is requested to
illustrate that the R-square values found in the output are the same as the
estimated reliabilities, and the standardized values found in the output
are the same as the estimated standardized values. Standard errors for
parameters named using the NEW option are given. The default
estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 5.1.

88
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.21: TWO-GROUP TWIN MODEL FOR

CONTINUOUS OUTCOMES USING PARAMETER
CONSTRAINTS

TITLE: this is an example of a two-group twin

model for continuous outcomes using
parameter constraints
DATA: FILE = ex5.21.dat;
VARIABLE: NAMES = y1 y2 g;
GROUPING = g(1 = mz 2 = dz);
MODEL: [y1-y2] (1);
y1-y2 (var);
y1 WITH y2 (covmz);
MODEL dz: y1 WITH y2 (covdz);
MODEL CONSTRAINT:
NEW(a c e h);
var = a**2 + c**2 + e**2;
covmz = a**2 + c**2;
covdz = 0.5*a**2 + c**2;
h = a**2/(a**2 + c**2 + e**2);

y1 y2

In this example, the model shown in the picture above is estimated using
parameter constraints. The model estimated is the same as the model in
Example 5.18.

In the MODEL command, labels are defined for three parameters. The
label var is assigned to the variances of y1 and y2. Because they are
given the same label, these parameters are held equal. In the overall
MODEL command, the label covmz is assigned to the covariance
between y1 and y2 for the monozygotic twins. In the group-specific
MODEL command, the label covdz is assigned to the covariance
between y1 and y2 for the dizygotic twins.

In the MODEL CONSTRAINT command, the NEW option is used to

assign labels to four parameters that are not in the analysis model: a, c,

89
CHAPTER 5

e, and h. The three parameters a, c, and e are used to decompose the

variances and covariances of y1 and y2 into genetic and environmental
components. The parameter h does not impose restrictions on the model
parameters but is used to compute the heritability estimate and its
standard error. The default estimator for this type of analysis is
maximum likelihood. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Examples 5.1, 5.14, 5.18, and 5.20.

EXAMPLE 5.22: TWO-GROUP TWIN MODEL FOR

CATEGORICAL OUTCOMES USING PARAMETER
CONSTRAINTS

TITLE: this is an example of a two-group twin

model for categorical outcomes using
parameter constraints
DATA: FILE = ex5.22.dat;
VARIABLE: NAMES = u1 u2 g;
GROUPING = g(1 = mz 2 = dz);
CATEGORICAL = u1 u2;
MODEL: [u1$1-u2$1](1);
u1 WITH u2(covmz);
MODEL dz: u1 WITH u2(covdz);
MODEL CONSTRAINT:
NEW(a c e h);
covmz = a**2 + c**2;
covdz = 0.5*a**2 + c**2;
e = 1 - (a**2 + c**2);
h = a**2/1;

The difference between this example and Example 5.21 is that the
outcomes are binary or ordered categorical instead of continuous
variables. Because of this, the outcomes have no freely estimated
residual variances. The ACE variance and covariance restrictions are
placed on normally-distributed latent response variables underlying the
categorical outcomes which are also called liabilities. This model is
referred to as the threshold model for liabilities (Neale & Cardon, 1992).
The model estimated is the same as the model in Example 5.19.

The variance contribution from the E factor is not a freely estimated

parameter with categorical outcomes. It is a remainder obtained by
subtracting the variance contributions of the A and C factors from the

90
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

unit variance of the latent response variables underlying u1 and u2 as

shown in the MODEL CONSTRAINT command. The denominator for
the heritability estimate is one with categorical outcomes because the
latent response variables have unit variances.

The default estimator for this type of analysis is a robust weighted least
squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. With maximum
likelihood, logistic or probit regressions are estimated using a numerical
integration algorithm. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. An explanation of the other commands can
be found in Examples 5.1, 5.14, 5.19 and 5.21.

EXAMPLE 5.23: QTL SIBLING MODEL FOR A CONTINUOUS

OUTCOME USING PARAMETER CONSTRAINTS

TITLE: this is an example of a QTL sibling model

for a continuous outcome using parameter
constraints
DATA: FILE = ex5.23.dat;
VARIABLE: NAMES = y1 y2 pihat;
USEVARIABLES = y1 y2;
CONSTRAINT = pihat;
MODEL: [y1-y2] (1);
y1-y2 (var);
y1 WITH y2 (cov);
MODEL CONSTRAINT:
NEW(a e q);
var = a**2 + e**2 + q**2;
cov = 0.5*a**2 + pihat*q**2;

y1 y2

pihat

91
CHAPTER 5

In this example, the model shown in the picture above is estimated. This
is a QTL model for two siblings (Marlow et al. 2003; Posthuma et al.
2004) for continuous outcomes where parameter constraints are used to
represent the A, E, and Q components. The A component represents the
additive genetic effects which correlate 0.5 for siblings. The E
component represents uncorrelated environmental effects. The Q
component represents a quantitative trait locus (QTL). The observed
variable pihat contains the estimated proportion alleles shared identity-
by-descent (IBD) by the siblings and moderates the effect of the Q
component on the covariance between the outcomes.

The CONSTRAINT option in the VARIABLE command is used to

identify the variables that can be used in the MODEL CONSTRAINT
command. These can be not only variables used in the MODEL
command but also other variables. In this example, the variable pihat is
used in the MODEL CONSTRAINT command although it is not used in
the MODEL command.

In the MODEL command, the (1) following the first bracket statement
specifies that the intercepts of y1 and y2 are held equal across the two
siblings. In addition, labels are defined for two parameters. The label
var is assigned to the variances of y1 and y2. Because they are given the
same label, these parameters are held equal. The label cov is assigned to
the covariance between y1 and y2.

In the MODEL CONSTRAINT command, the NEW option is used to

assign labels to three parameters that are not in the analysis model: a, e,
and q. The three parameters a, e, and q and the variable pihat are used to
decompose the variances and covariances of y1 and y2 into genetic,
environmental, and QTL components. The default estimator for this
type of analysis is maximum likelihood. The ESTIMATOR option of
the ANALYSIS command can be used to select a different estimator.
An explanation of the other commands can be found in Examples 5.1
and 5.20.

92
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.24: EFA WITH COVARIATES (MIMIC) WITH

CONTINUOUS FACTOR INDICATORS AND DIRECT
EFFECTS

TITLE: this is an example of an EFA with

covariates (MIMIC) with continuous factor
indicators and direct effects
DATA: FILE IS ex5.24.dat;
VARIABLE: NAMES ARE y1-y8 x1 x2;
MODEL: f1-f2 BY y1-y8(*1);
f1-f2 ON x1-x2;
y1 ON x1;
y8 ON x2;
OUTPUT: TECH1;

y1

y2

x1 f1 y3

y4

y5

x2 f2 y6

y7

y8

In this example, the EFA with covariates (MIMIC) with continuous

factor indicators and direct effects shown in the picture above is

93
CHAPTER 5

estimated. This is an exploratory structural equation model (ESEM;

Asparouhov & Muthén, 2009a). The factors f1 and f2 are EFA factors
which have the same factor indicators. Unlike CFA, no factor loadings
are fixed at zero. Instead, the four restrictions on the factor loadings,
factor variances, and factor covariances necessary for identification are
imposed by rotating the factor loading matrix and fixing the factor
residual variances at one.

In the MODEL command, the BY statement specifies that the factors f1

and f2 are measured by the continuous factor indicators y1 through y8.
The label 1 following an asterisk (*) in parentheses following the BY
statement is used to indicate that f1 and f2 are a set of EFA factors.
When no rotation is specified using the ROTATION option of the
ANALYSIS command, the default oblique GEOMIN rotation is used.
The intercepts and residual variances of the factor indicators are
estimated and the residuals are not correlated as the default. The
residual variances of the factors are fixed at one as the default. The
residuals of the factors are correlated under the default oblique
GEOMIN rotation. The first ON statement describes the linear
regressions of f1 and f2 on the covariates x1 and x2. The second and
third ON statements describe the linear regressions of y1 on x1 and y8
on x2. These regressions represent direct effects used to test for
measurement non-invariance.

The default estimator for this type of analysis is maximum likelihood.

The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. An explanation of the other commands can
be found in Example 5.1.

94
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.25: SEM WITH EFA AND CFA FACTORS WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a SEM with EFA and

CFA factors with continuous factor
indicators
DATA: FILE IS ex5.25.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1-f2 BY y1-y6 (*1);
f3 BY y7-y9;
f4 BY y10-y12;
f3 ON f1-f2;
f4 ON f3;

y7 y8 y9 y10 y11 y12

y1

y2 f1

y3
f3 f4
y4

y5 f2

y6

In this example, the SEM with EFA and CFA factors with continuous
factor indicators shown in the picture above is estimated. This is an
exploratory structural equation model (ESEM; Asparouhov & Muthén,
2009a). The factors f1 and f2 are EFA factors which have the same
factor indicators. Unlike CFA, no factor loadings are fixed at zero.
Instead, the four restrictions on the factor loadings, factor variances, and
factor covariances necessary for identification are imposed by rotating

95
CHAPTER 5

the factor loading matrix and fixing the factor variances at one. The
factors f3 and f4 are CFA factors.

In the MODEL command, the first BY statement specifies that the

factors f1 and f2 are measured by the continuous factor indicators y1
through y6. The label 1 following an asterisk (*) in parentheses
following the BY statement is used to indicate that f1 and f2 are a set of
EFA factors. When no rotation is specified using the ROTATION
option of the ANALYSIS command, the default oblique GEOMIN
rotation is used. For EFA factors, the intercepts and residual variances
of the factor indicators are estimated and the residuals are not correlated
as the default. The variances of the factors are fixed at one as the
default. The factors are correlated under the default oblique GEOMIN
rotation. The second BY statement specifies that f3 is measured by y7,
y8, and y9. The third BY statement specifies that f4 is measured by y10,
y11, and y12. The metric of the factors is set automatically by the
program by fixing the first factor loading in each BY statement to 1.
This option can be overridden. The intercepts and residual variances of
the factor indicators are estimated and the residual are not correlated as
the default. The residual variances of the factors are estimated as the
default.

The first ON statement describes the linear regression of f3 on the set of

EFA factors f1 and f2. The second ON statement describes the linear
regression of f4 on f3. The default estimator for this type of analysis is
maximum likelihood. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 5.1.

96
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.26: EFA AT TWO TIME POINTS WITH FACTOR

LOADING INVARIANCE AND CORRELATED RESIDUALS
ACROSS TIME

TITLE: this is an example of an EFA at two time

points with factor loading invariance and
correlated residuals across time
DATA: FILE IS ex5.26.dat;
VARIABLE: NAMES ARE y1-y12;
MODEL: f1-f2 BY y1-y6 (*t1 1);
f3-f4 BY y7-y12 (*t2 1);
y1-y6 PWITH y7-y12;
OUTPUT: TECH1 STANDARDIZED;

y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12

f1 f2 f3 f4

In this example, the EFA at two time points with factor loading
invariance and correlated residuals across time shown in the picture
above is estimated. This is an exploratory structural equation model
(ESEM; Asparouhov & Muthén, 2009a). The factor indicators y1
through y6 and y7 through y12 are the same variables measured at two
time points. The factors f1 and f2 are one set of EFA factors which have
the same factor indicators and the factors f3 and f4 are a second set of
EFA factors which have the same factor indicators. Unlike CFA, no
factor loadings are fixed at zero in either set. Instead, for each set, the
four restrictions on the factor loadings, factor variances, and factor
covariances necessary for identification are imposed by rotating the
factor loading matrix and fixing the factor variances at one at the first
time point. For the other time point, factor variances are free to be

97
CHAPTER 5

estimated as the default when factor loadings are constrained to be equal

across time.

In the MODEL command, the first BY statement specifies that the

factors f1 and f2 are measured by the continuous factor indicators y1
through y6. The label t1 following an asterisk (*) in parentheses
following the BY statement is used to indicate that f1 and f2 are a set of
EFA factors. The second BY statement specifies that the factors f3 and
f4 are measured by the continuous factor indicators y7 through y12. The
label t2 following an asterisk (*) in parentheses following the BY
statement is used to indicate that f3 and f4 are a set of EFA factors. The
number 1 following the labels t1 and t2 specifies that the factor loadings
matrices for the two sets of EFA factors are constrained to be equal.
When no rotation is specified using the ROTATION option of the
ANALYSIS command, the default oblique GEOMIN rotation is used.

For EFA factors, the intercepts and residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The intercepts are not held equal across time as the default.
The means of the factors are fixed at zero at both time points and the
variances of the factors are fixed at one as the default. In this example
because the factor loadings are constrained to be equal across time, the
factor variances are fixed at one at the first time point and are free to be
estimated at the other time point. The factors are correlated as the
default under the oblique GEOMIN rotation. The PWITH statement
specifies that the residuals for each factor indicator are correlated over
time. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 5.1.

98
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.27: MULTIPLE-GROUP EFA WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of multiple-group EFA

with continuous factor indicators with no
measurement invariance
DATA: FILE IS ex5.27.dat;
VARIABLE: NAMES ARE y1-y10 group;
GROUPING IS group (1 = g1 2 = g2);
MODEL: f1-f2 BY y1-y10 (*1);
[f1-f2@0];
MODEL g2: f1-f2 BY y1-y10 (*1);
[y1-y10];
OUTPUT: TECH1;

y1 y2 y3 y4 y5 y6 y7 y8 y9 y10

f1 f2

In this example, the multiple-group EFA with continuous indicators

shown in the picture above is estimated. This is an exploratory
structural equation model (ESEM; Asparouhov & Muthén, 2009a). The
factors f1 and f2 are EFA factors which have the same factor indicators.
Unlike CFA, no factor loadings are fixed at zero. Instead, for the first
group the four restrictions on the factor loadings, factor variances, and
factor covariances necessary for model identification are imposed by
rotating the factor loading matrix and fixing the factor variances at one
in one group. For the other group, factor variances are free to be
estimated. The first model in this example imposes no equality
constraints on the model parameters across the two groups. Four
subsequent models impose varying degrees of invariance on the model
parameters.

99
CHAPTER 5

In the MODEL command, the BY statement specifies that the factors f1

and f2 are measured by the continuous factor indicators y1 through y10.
The label 1 following an asterisk (*) in parentheses following the BY
statement is used to indicate that f1 and f2 are a set of EFA factors.
When no rotation is specified using the ROTATION option of the
ANALYSIS command, the default oblique GEOMIN rotation is used.

The intercepts and residual variances of the factor indicators are

estimated and the residuals are not correlated as the default. The
variances of the factors are fixed at one in both groups. The factors are
correlated under the default oblique GEOMIN rotation. The bracket
statement specifies that the factor means are fixed at zero in both groups
to override the default of the factor means being fixed at zero in the first
group and being free in the other group.

In the group-specific MODEL command for g2, the BY statement

relaxes the default equality constraint on the factor loading matrices in
the two groups. The bracket statement relaxes the default equality
constraint on the intercepts of the factor indicators y1 through y10 in the
two groups. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 5.1

Following is the second part of the example where equality of factor

loading matrices across the two groups is imposed. The variances of the
factors are fixed at one in the first group and are free to be estimated in
the other group.

MODEL: f1-f2 BY y1-y10 (*1);

[f1-f2@0];
MODEL g2: [y1-y10];

Equality of factor loading matrices is accomplished by removing the BY

statement from the group-specific MODEL command for g2. Equality
of factor loading matrices is the default.

Following is the third part of the example where equality of factor

loading matrices and intercepts of the factor indicators across the two
groups is imposed.

100
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling
MODEL: f1-f2 by y1-y10 (*1);

Equality of factor indicator intercepts is accomplished by removing the

bracket statement for y1 through y10 from the group-specific MODEL
command for g2. Equality of factor indicator intercepts is the default.
This specification is the default setting in multiple group analysis,
specifying measurement invariance of the intercepts of the factor
indicators and the factor loading matrices. The factor means are fixed at
zero in the first group and are free to be estimated in the other group as
the default.

Following is the fourth part of the example where equality of factor

variances and the factor covariance is imposed in addition to
measurement invariance of the intercepts and factor loading matrices.

MODEL: f1-f2 by y1-y10 (*1);

f1 WITH f2 (1);
f1-f2@1;

In the MODEL command, the number one in parentheses following the

WITH statement specifies that the covariance between f1 and f2 is held
equal across the two groups. The default in multiple group EFA when
factor loading matrices are held equal across groups is that the factor
variances are fixed to one in the first group and are free to be estimated
in the other group. The third statement in the MODEL command
overrides this default by specifing that the factor variances are fixed at
one in both groups.

Following is the fifth part of the example where in addition to equality

of factor variances and the factor covariance, equality of the factor
means is imposed in addition to measurement invariance of the
intercepts and factor loading matrices.

MODEL: f1-f2 by y1-y10 (*1);

f1 WITH f2 (1);
f1-f2@1;
[f1-f2@0];

The default in multiple group EFA is that the factor means are fixed to
zero in the first group and are free to be estimated in the other groups.
The bracket statement in the MODEL command specifies that the factor
means are fixed at zero in both groups.

101
CHAPTER 5

EXAMPLE 5.28: EFA WITH RESIDUAL VARIANCES

CONSTRAINED TO BE GREATER THAN ZERO

TITLE: this is an example of an EFA with residual

variances constrained to be greater than
zero
DATA: FILE = ex5.28.dat;
VARIABLE: NAMES = y1-y10;
ANALYSIS: ROTATION = GEOMIN;
MODEL: f1-f2 BY y1-y10 (*1);
y1-y10 (v1-v10);
MODEL CONSTRAINT:
DO(1,10) v#>0;
OUTPUT: STDY;

In this example, an exploratory factor analysis with residual variances

constrained to be greater than zero is carried out using a Geomin
rotation. This is an exploratory structural equation model (ESEM;
Asparouhov & Muthén, 2009a). The factors f1 and f2 are EFA factors
which have the same factor indicators. By specifying GEOMIN, an EFA
will be carried out using the Geomin rotation. The default is an oblique
rotation. An orthogonal rotation can be obtained by specifying
ROTATION=GEOMIN(ORTHOGONAL). The ROTATION option can
be used to specify other rotations.

In the MODEL command, the BY statement specifies that the factors f1

and f2 are measured by the continuous factor indicators y1 through y10.
The label 1 following an asterisk (*) in parentheses following the BY
statement is used to indicate that f1 and f2 are a set of EFA factors. The
intercepts and residual variances of the factor indicators are estimated
and the residuals are not correlated as the default. The variances of the
factors are fixed at one as the default. The DO option of the MODEL
CONSTRAINT command is used to constrain the residual variances of
the factor indicators to be greater than zero. The DO option provides a
do loop to facilitate specifying the same expression for a set of
parameters. The parameters are given labels in the MODEL command.
In the DO option, the numbers in parentheses give the range of values
for the do loop. The number sign (#) is replaced by these values during
the execution of the do loop. In the OUTPUT command, the STDY
option is chosen for standardization with respect to y. This puts the
results in the metric of an EFA. The default estimator for this type of

102
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

analysis is maximum likelihood. The ESTIMATOR option of the

ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 5.1.

EXAMPLE 5.29: BI-FACTOR EFA USING ESEM

TITLE: this is an example of a bi-factor EFA

using ESEM
DATA: FILE = ex5.29.dat;
VARIABLE: NAMES = y1-y10;
ANALYSIS: ROTATION = BI-GEOMIN;
MODEL: fg f1 f2 BY y1-y10 (*1);
OUTPUT: STDY;

In this example, a bi-factor exploratory factor analysis (Jennrich &

Bentler, 2011, 2012) using ESEM with continuous factor indicators is
carried out using a bi-factor Geomin rotation. This is an exploratory
structural equation model (ESEM; Asparouhov & Muthén, 2009a). The
factors fg, f1, and f2 are EFA factors which have the same factor
indicators. By specifying BI-GEOMIN, a bi-factor EFA will be carried
out using a bi-factor Geomin rotation. The default is an oblique solution
where the specific factors are correlated with the general factor and are
correlated with each other. In the orthogonal solution, the specific
factors are uncorrelated with the general factor and are uncorrelated with
each other. An orthogonal rotation is obtained by specifying
ROTATION=BI-GEOMIN(ORTHOGONAL). An alternative bi-factor
rotation can be obtained using the BI-CF-QUARTIMAX setting of the
ROTATION option.

In the MODEL command, the BY statement specifies that the factors fg,
f1, and f2 are measured by the continuous factor indicators y1 through
y10. The factor fg is a general factor and f1 and f2 are specific factors.
The label 1 following an asterisk (*) in parentheses following the BY
statement is used to indicate that fg, f1, and f2 are a set of EFA factors.
The intercepts and residual variances of the factor indicators are
estimated and the residuals are not correlated as the default. The
variances of the factors are fixed at one as the default. In the OUTPUT
command, the STDY option is chosen for standardization with respect to
y. This puts the results in the metric of an EFA. The default estimator
for this type of analysis is maximum likelihood. The ESTIMATOR
option of the ANALYSIS command can be used to select a different

103
CHAPTER 5

estimator. An explanation of the other commands can be found in

Example 5.1.

EXAMPLE 5.30: BI-FACTOR EFA WITH TWO ITEMS

LOADING ON ONLY THE GENERAL FACTOR

TITLE: this is an example of bi-factor EFA with

two items loading on only the general
factor
DATA: FILE = ex5.30.dat;
VARIABLE: NAMES = y1-y10;
ANALYSIS: ROTATION = GEOMIN;
MODEL: fg BY y1-y10*;
fg@1;
f1-f2 BY y1-y8 (*1);
fg WITH f1-f2@0;
OUTPUT: STDY;

In this example, a bi-factor exploratory factor analysis with continuous

factor indicators is carried out using a Geomin rotation. This is an
exploratory structural equation model (ESEM; Asparouhov & Muthén,
2009a). The general factor fg is a CFA factor. The specific factors f1
and f2 are EFA factors which have the same factor indicators. Only the
specific factors are rotated. By specifying GEOMIN, an EFA will be
carried out using the Geomin rotation for the specific factors. The
default is an oblique rotation. An orthogonal rotation can be obtained by
specifying ROTATION=GEOMIN(ORTHOGONAL). The ROTATION
option can be used to specify other rotations.

In the MODEL command, the first BY statement specifies that the

general factor fg is measured by y1 through y10. The asterisk (*) frees
the first factor loading which is fixed at one as the default to define the
metric of the factor. Instead the metric of the factor is defined by fixing
the factor variance at one. The second BY statement specifies that the
specific factors f1 and f2 are measured by the continuous factor
indicators y1 through y8. The label following an asterisk (*) in
parentheses following the BY statement is used to indicate that f1 and f2
are a set of EFA factors. The variances of the factors are fixed at one as
the default. The specific factors are correlated under the oblique
Geomin rotation. The WITH statement specifies that the general and
specific factors are not correlated. The intercepts and residual variances

104
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

of the factor indicators are estimated and the residuals are not correlated
as the default. In the OUTPUT command, the STDY option is chosen
for standardization with respect to y. This puts the results in the metric
of an EFA. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 5.1.

EXAMPLE 5.31: BAYESIAN BI-FACTOR CFA WITH TWO

ITEMS LOADING ON ONLY THE GENERAL FACTOR AND
CROSS-LOADINGS WITH ZERO-MEAN AND SMALL-
VARIANCE PRIORS

TITLE: this is an example of a Bayesian bi-factor

CFA with two items loading on only the
general factor and cross-loadings with
zero-mean and small-variance priors
DATA: FILE = ex5.31.dat;
VARIABLE: NAMES = y1-y10;
ANALYSIS: ESTIMATOR = BAYES;
PROCESSORS = 2;
MODEL: fg BY y1-y10*;
fg@1;
f1 BY y1-y4
y5-y10 (f1xlam5-f1xlam10);
f2 BY y5-y8
y1-y4 y9-y10(f2xlam1-f2xlam6);
fg WITH f1-f2@0;
MODEL PRIORS:
f1xlam5-f2xlam6~N(0,0.01);
PLOT: TYPE = PLOT2;

In this example, a bi-factor CFA with two items loading on only the
general factor and cross-loadings with zero-mean and small-variance
priors is carried out using the Bayes estimator. This is a Bayesian
structural equation model (BSEM; Muthén & Asparouhov, 2012). By
specifying ESTIMATOR=BAYES, a Bayesian analysis will be carried
out. In Bayesian estimation, the default is to use two independent
Markov chain Monte Carlo (MCMC) chains. If multiple processors are
available, using PROCESSORS=2 will speed up computations.

105
CHAPTER 5

In the MODEL command, the first BY statement specifies that the

general factor, fg, is measured by the continuous factor indicators y1
through y10. The asterisk (*) frees the first factor loading which is fixed
at one as the default to define the metric of the factor. Instead the metric
of the factor is defined by fixing the factor variance at one. The second
and third BY statements specify that the specific factors, f1 and f2, are
measured by y1 through y10. The first factor loadings are fixed at one to
set the metric of the factors. The first line of each BY statement shows
the major loadings for each factor. The second line shows the cross-
loadings which are assigned labels. For f1, labels are assigned to cross-
loadings for y5 through y10. For f2, labels are assigned to cross-
loadings for y1 through y4, y9, and y10. The WITH statement specifies
that the general and specific factors are not correlated. The intercepts
and residual variances of the factor indicators are estimated and the
residuals are not correlated as the default. In MODEL PRIORS, the
labels assigned in the MODEL command are used to assign zero-mean
and small-variance priors to the cross-loadings. By specifying
TYPE=PLOT2 in the PLOT command, the following plots are available:
posterior parameter distributions, posterior parameter trace plots,
autocorrelation plots, posterior predictive checking scatterplots, and
posterior predictive checking distribution plots. An explanation of the
other commands can be found in Example 5.1.

106
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

EXAMPLE 5.32: BAYESIAN MIMIC MODEL WITH CROSS-

LOADINGS AND DIRECT EFFECTS WITH ZERO-MEAN AND
SMALL-VARIANCE PRIORS

TITLE: this is an example of a Bayesian MIMIC

model with cross-loadings and direct
effects with zero-mean and small-variance
priors
DATA: FILE = ex5.32.dat;
VARIABLE: NAMES = y1-y6 x1-x3;
ANALYSIS: ESTIMATOR = BAYES;
PROCESSORS = 2;
MODEL: f1 BY y1-y3
y4-y6 (xload4-xload6);
f2 BY y4-y6
y1-y3 (xload1-xload3);
f1-f2 ON x1-x3;
y1-y6 ON x1-x3 (dir1-dir18);
MODEL PRIORS:
xload1-xload6~N(0,0.01);
dir1-dir18~N(0,0.01);
PLOT: TYPE = PLOT2;

In this example, a MIMIC model with cross-loadings and direct effects

with zero-mean and small-variance priors is carried out using the Bayes
estimator. This is a Bayesian structural equation model (BSEM; Muthén
& Asparouhov, 2012).

In the MODEL command, the first BY statement specifies that f1 is

measured by the continuous factor indicators y1 through y6. The second
BY statements specifies that f2 is measured by the continuous factor
indicators y1 through y6. The first factor loadings are fixed at one to set
the metric of the factors. The first line of each BY statement shows the
major loadings for each factor. The second line shows the cross-
loadings which are assigned labels. For f1, labels are assigned to y4
through y6. For f2, labels are assigned to y1 through y3. The intercepts
and residual variances of the factor indicators are estimated and the
residuals are not correlated as the default. In MODEL PRIORS, the
labels assigned in the MODEL command are used to assign zero-mean
and small-variance priors to the factor loadings. The first ON statement
describes the linear regressions of f1 and f2 on the covariates x1, x2, and
x3. The residual variances of f1 and f2 are estimated and the residuals

107
CHAPTER 5

are correlated as the default. The second ON statement describes the

linear regressions of y1 through y6 on the covariates x1, x2, and x3.
These are direct effects which are assumed to be small. Labels are
assigned to these regression coefficients. In MODEL PRIORS, the
labels assigned in the MODEL command are used to assign zero-mean
and small-variance priors to the factor loadings and regression
coefficients. By specifying TYPE=PLOT2 in the PLOT command, the
following plots are available: posterior parameter distributions,
posterior parameter trace plots, autocorrelation plots, posterior
predictive checking scatterplots, and posterior predictive checking
distribution plots. An explanation of the other commands can be found
in Examples 5.1 and 5.31.

EXAMPLE 5.33: BAYESIAN MULTIPLE GROUP MODEL

WITH APPROXIMATE MEASUREMENT INVARIANCE
USING ZERO-MEAN AND SMALL-VARIANCE PRIORS

TITLE: this is an example of a Bayesian

multiple group model with approximate
measurement invariance using zero-mean and
small-variance priors
DATA: FILE = ex5.33.dat;
VARIABLE: NAMES = u y1-y6 group;
USEVARIABLES = y1-y6 group;
CLASSES = c(10);
KNOWNCLASS = c(group = 1-10);
ANALYSIS: TYPE = MIXTURE;
ESTIMATOR = BAYES;
PROCESSORS = 2;
MODEL = ALLFREE;
MODEL: %OVERALL%
f1 BY y1-y3* (lam#_1-lam#_3);
f2 BY y4-y6* (lam#_4-lam#_6);
[y1-y6] (nu#_1-nu#_6);
%c#10%
f1-f2@1;
[f1-f2@0];
MODEL PRIORS:
DO(1,6) DIFF(lam1_#-lam10_#)~N(0,0.01);
DO(1,6) DIFF(nu1_#-nu10_#)~N(0,0.01);
PLOT: TYPE = PLOT2;
OUTPUT: TECH1 TECH8;

108
 Examples: Confirmatory Factor Analysis And
Structural Equation Modeling

In this example, a multiple group model with approximate measurement

invariance using zero-mean and small-variance priors is carried out using
the Bayes estimator. This is a Bayesian structural equation model
(BSEM; Muthén & Asparouhov, 2012). In Bayesian estimation,
multiple group analysis is carried out using the CLASSES and
KNOWNCLASS options and TYPE=MIXTURE. The CLASSES option
is used to assign names to the categorical latent variables in the model
and to specify the number of latent classes in the model for each
categorical latent variable. In the example above, there is one
categorical latent variable c that has ten latent classes. The
KNOWNCLASS option identifies c as the categorical latent variable for
which latent class membership is known. The information in
parentheses following the categorical latent variable name defines the
known classes using an observed variable. In this example, the observed
variable group is used to define the known classes. The first class
consists of individuals with the value 1 on the variable group. The
second class consists of individuals with the value 2 on the variable
group etc.

MODEL=ALLFREE is used with TYPE=MIXTURE, the

KNOWNCLASS option, ESTIMATOR=BAYES, and a special labeling
function to assign zero-mean and small-variance priors to differences in
intercepts, thresholds, and factor loadings across groups. By specifying
MODEL=ALLFREE, factor means, variances, and covariances are free
across groups except for factor means in the last group which are fixed
at zero. In addition, intercepts, thresholds, factor loadings, and residual
variances of the factor indicators are free across the groups.

In the overall model, the first BY statement specifies that f1 is measured

by the continuous factor indicators y1 through y3. The second BY
statement specifies that f2 is measured by the continuous factor
indicators y4 through y6. In both BY statements the asterisk (*) frees
the first factor loadings which are fixed at one as the default to set the
metric of the factors. The metric of the factors is set instead by fixing
the factor variances to one in class 10. The residual variances of the
factor indicators are estimated and the residuals are not correlated as the
default.

In the overall part of the model, labels are assigned to the factor loadings
and the intercepts using automatic labeling for groups. The labels must
include the number sign (#) followed by the underscore (_) symbol

109
CHAPTER 5

followed by a number. The number sign (#) refers to a group and the
number refers to a parameter. The label lam#_1 is assigned to the factor
loading for y1; the label lam#_2 is assigned to the factor loading for y2;
and the label lam#_3 is assigned to the factor loading for y3. These
labels are expanded to include group information. For example, the
label for parameter 1 is expanded across the ten groups to give labels
lam1_1, lam2_1 through lam10_1. In MODEL PRIORS, these expanded
labels are used to assign zero-mean and small-variance priors to the
differences across groups of the factor loadings and intercepts using the
DO and DIFFERENCE options. They can be used together to simplify
the assignment of priors to a large set of difference parameters for
models with multiple groups and multiple time points. For the DO
option, the numbers in parentheses give the range of values for the do
loop. The number sign (#) is replaced by these values during the
execution of the do loop. The numbers refer to the six factor indicators.

By specifying TYPE=PLOT2 in the PLOT command, the following

plots are available: posterior parameter distributions, posterior
parameter trace plots, autocorrelation plots, posterior predictive
checking scatterplots, and posterior predictive checking distribution
plots. An explanation of the other commands can be found in Example
5.1 and 5.31.

110
 Examples: Growth Modeling And Survival Analysis

CHAPTER 6
EXAMPLES: GROWTH
MODELING AND SURVIVAL
ANALYSIS

Growth models examine the development of individuals on one or more

outcome variables over time. These outcome variables can be observed
variables or continuous latent variables. Observed outcome variables
can be continuous, censored, binary, ordered categorical (ordinal),
counts, or combinations of these variable types if more than one growth
process is being modeled. In growth modeling, random effects are used
to capture individual differences in development. In a latent variable
modeling framework, the random effects are reconceptualized as
continuous latent variables, that is, growth factors.

Mplus takes a multivariate approach to growth modeling such that an

outcome variable measured at four occasions gives rise to a four-variate
outcome vector. In contrast, multilevel modeling typically takes a
univariate approach to growth modeling where an outcome variable
measured at four occasions gives rise to a single outcome for which
observations at the different occasions are nested within individuals,
resulting in two-level data. Due to the use of the multivariate approach,
Mplus does not consider a growth model to be a two-level model as in
multilevel modeling but a single-level model. With longitudinal data,
the number of levels in Mplus is one less than the number of levels in
conventional multilevel modeling. The multivariate approach allows
flexible modeling of the outcomes such as differences in residual
variances over time, correlated residuals over time, and regressions
among the outcomes over time.

In Mplus, there are two options for handling the relationship between the
outcome and time. One approach allows time scores to be parameters in
the model so that the growth function can be estimated. This is the
approach used in structural equation modeling. The second approach
allows time to be a variable that reflects individually-varying times of
observations. This variable has a random slope. This is the approach
used in multilevel modeling. Random effects in the form of random

111
CHAPTER 6

slopes are also used to represent individual variation in the influence of

time-varying covariates on outcomes.

Growth modeling in Mplus allows the analysis of multiple processes,

both parallel and sequential; regressions among growth factors and
random effects; growth modeling of factors measured by multiple
indicators; and growth modeling as part of a larger latent variable model.

Survival modeling in Mplus includes both discrete-time and continuous-

time analyses. Both types of analyses consider the time to an event.
Discrete-time survival analysis is used when the outcome is recorded
infrequently such as monthly or annually, typically leading to a limited
number of measurements. Continuous-time survival analysis is used
when the outcome is recorded more frequently such as hourly or daily,
typically leading to a large number of measurements. Survival modeling
is integrated into the general latent variable modeling framework so that
it can be part of a larger model.

All growth and survival models can be estimated using the following
special features:

• Single or multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities

For continuous, censored with weighted least squares estimation, binary,

and ordered categorical (ordinal) outcomes, multiple group analysis is
specified by using the GROUPING option of the VARIABLE command
for individual data or the NGROUPS option of the DATA command for
summary data. For censored with maximum likelihood estimation,
unordered categorical (nominal), and count outcomes, multiple group
analysis is specified using the KNOWNCLASS option of the

112
 Examples: Growth Modeling And Survival Analysis

VARIABLE command in conjunction with the TYPE=MIXTURE

option of the ANALYSIS command. The default is to estimate the
model under missing data theory using all available data. The
LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. The
SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed. Latent variable interactions
are specified by using the | symbol of the MODEL command in
conjunction with the XWITH option of the MODEL command. Random
slopes are specified by using the | symbol of the MODEL command in
conjunction with the ON option of the MODEL command. Individually-
varying times of observations are specified by using the | symbol of the
MODEL command in conjunction with the AT option of the MODEL
command and the TSCORES option of the VARIABLE command.
Linear and non-linear parameter constraints are specified by using the
MODEL CONSTRAINT command. Indirect effects are specified by
using the MODEL INDIRECT command. Maximum likelihood
estimation is specified by using the ESTIMATOR option of the
ANALYSIS command. Bootstrap standard errors are obtained by using
the BOOTSTRAP option of the ANALYSIS command. Bootstrap
confidence intervals are obtained by using the BOOTSTRAP option of
the ANALYSIS command in conjunction with the CINTERVAL option
of the OUTPUT command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, and plots
of sample and estimated means and proportions/probabilities. These are
available for the total sample, by group, by class, and adjusted for
covariates. The PLOT command includes a display showing a set of
descriptive statistics for each variable. The graphical displays can be
edited and exported as a DIB, EMF, or JPEG file. In addition, the data

113
CHAPTER 6

for each graphical display can be saved in an external file for use by
another graphics program.

Following is the set of growth modeling examples included in this

chapter:

• 6.1: Linear growth model for a continuous outcome

• 6.2: Linear growth model for a censored outcome using a censored
model*
• 6.3: Linear growth model for a censored outcome using a censored-
inflated model*
• 6.4: Linear growth model for a categorical outcome
• 6.5: Linear growth model for a categorical outcome using the Theta
parameterization
• 6.6: Linear growth model for a count outcome using a Poisson
model*
• 6.7: Linear growth model for a count outcome using a zero-inflated
Poisson model*
• 6.8: Growth model for a continuous outcome with estimated time
scores
• 6.9: Quadratic growth model for a continuous outcome
• 6.10: Linear growth model for a continuous outcome with time-
invariant and time-varying covariates
• 6.11: Piecewise growth model for a continuous outcome
• 6.12: Growth model with individually-varying times of observation
and a random slope for time-varying covariates for a continuous
outcome
• 6.13: Growth model for two parallel processes for continuous
outcomes with regressions among the random effects
• 6.14: Multiple indicator linear growth model for continuous
outcomes
• 6.15: Multiple indicator linear growth model for categorical
outcomes
• 6.16: Two-part (semicontinuous) growth model for a continuous
outcome*
• 6.17: Linear growth model for a continuous outcome with first-
order auto correlated residuals using non-linear constraints
• 6.18: Multiple group multiple cohort growth model

114
 Examples: Growth Modeling And Survival Analysis

Following is the set of survival analysis examples included in this

chapter:

• 6.19: Discrete-time survival analysis

• 6.20: Continuous-time survival analysis using the Cox regression
model
• 6.21: Continuous-time survival analysis using a parametric
proportional hazards model
• 6.22: Continuous-time survival analysis using a parametric
proportional hazards model with a factor influencing survival*

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

EXAMPLE 6.1: LINEAR GROWTH MODEL FOR A

CONTINUOUS OUTCOME

TITLE: this is an example of a linear growth

model for a continuous outcome
DATA: FILE IS ex6.1.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;

y11 y12 y13 y14

i s

In this example, the linear growth model for a continuous outcome at

four time points shown in the picture above is estimated.

115
CHAPTER 6

TITLE: this is an example of a linear growth

model for a continuous outcome

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex6.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex6.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;

USEVARIABLES ARE y11-y14;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains ten variables: y11, y12, y13, y14, x1, x2, x31, x32,
x33, and x34. Note that the hyphen can be used as a convenience feature
in order to generate a list of names. If not all of the variables in the data
set are used in the analysis, the USEVARIABLES option can be used to
select a subset of variables for analysis. Here the variables y11, y12,
y13, and y14 have been selected for analysis. They represent the
outcome measured at four equidistant occasions.

MODEL: i s | y11@0 y12@1 y13@2 y14@3;

The MODEL command is used to describe the model to be estimated.

The | symbol is used to name and define the intercept and slope factors
in a growth model. The names i and s on the left-hand side of the |
symbol are the names of the intercept and slope growth factors,
respectively. The statement on the right-hand side of the | symbol
specifies the outcome and the time scores for the growth model. The
time scores for the slope growth factor are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
growth factor as an initial status factor. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are

116
 Examples: Growth Modeling And Survival Analysis

estimated and allowed to be different across time and the residuals are
not correlated as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables at the four time points are fixed at zero as the
default. The means and variances of the growth factors are estimated as
the default, and the growth factor covariance is estimated as the default
because the growth factors are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

EXAMPLE 6.2: LINEAR GROWTH MODEL FOR A

CENSORED OUTCOME USING A CENSORED MODEL

TITLE: this is an example of a linear growth

model for a censored outcome using a
censored model
DATA: FILE IS ex6.2.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
CENSORED ARE y11-y14 (b);
ANALYSIS: ESTIMATOR = MLR;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 6.1 is that the
outcome variable is a censored variable instead of a continuous variable.
The CENSORED option is used to specify which dependent variables
are treated as censored variables in the model and its estimation, whether
they are censored from above or below, and whether a censored or
censored-inflated model will be estimated. In the example above, y11,
y12, y13, and y14 are censored variables. They represent the outcome
variable measured at four equidistant occasions. The b in parentheses
following y11-y14 indicates that y11, y12, y13, and y14 are censored
from below, that is, have floor effects, and that the model is a censored
regression model. The censoring limit is determined from the data. The
residual variances of the outcome variables are estimated and allowed to
be different across time and the residuals are not correlated as the
default.

117
CHAPTER 6

The default estimator for this type of analysis is a robust weighted least
squares estimator. By specifying ESTIMATOR=MLR, maximum
likelihood estimation with robust standard errors using a numerical
integration algorithm is used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of
integration are used with a total of 225 integration points. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables at the four time points are fixed at zero as the
default. The means and variances of the growth factors are estimated as
the default, and the growth factor covariance is estimated as the default
because the growth factors are independent (exogenous) variables. The
OUTPUT command is used to request additional output not included as
the default. The TECH1 option is used to request the arrays containing
parameter specifications and starting values for all free parameters in the
model. The TECH8 option is used to request that the optimization
history in estimating the model be printed in the output. TECH8 is
printed to the screen during the computations as the default. TECH8
screen printing is useful for determining how long the analysis takes. An
explanation of the other commands can be found in Example 6.1.

118
 Examples: Growth Modeling And Survival Analysis

EXAMPLE 6.3: LINEAR GROWTH MODEL FOR A

CENSORED OUTCOME USING A CENSORED-INFLATED
MODEL

TITLE: this is an example of a linear growth

model for a censored outcome using a
censored-inflated model
DATA: FILE IS ex6.3.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
CENSORED ARE y11-y14 (bi);
ANALYSIS: INTEGRATION = 7;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;
ii si | y11#1@0 y12#1@1 y13#1@2 y14#1@3;
si@0;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 6.1 is that the
outcome variable is a censored variable instead of a continuous variable.
The CENSORED option is used to specify which dependent variables
are treated as censored variables in the model and its estimation, whether
they are censored from above or below, and whether a censored or
censored-inflated model will be estimated. In the example above, y11,
y12, y13, and y14 are censored variables. They represent the outcome
variable measured at four equidistant occasions. The bi in parentheses
following y11-y14 indicates that y11, y12, y13, and y14 are censored
from below, that is, have floor effects, and that a censored-inflated
regression model will be estimated. The censoring limit is determined
from the data. The residual variances of the outcome variables are
estimated and allowed to be different across time and the residuals are
not correlated as the default.

With a censored-inflated model, two growth models are estimated. The

first | statement describes the growth model for the continuous part of
the outcome for individuals who are able to assume values of the
censoring point and above. The residual variances of the outcome
variables are estimated and allowed to be different across time and the
residuals are not correlated as the default. The second | statement
describes the growth model for the inflation part of the outcome, the
probability of being unable to assume any value except the censoring
point. The binary latent inflation variable is referred to by adding to the

119
CHAPTER 6

name of the censored variable the number sign (#) followed by the
number 1.

In the parameterization of the growth model for the continuous part of

the outcome, the intercepts of the outcome variables at the four time
points are fixed at zero as the default. The means and variances of the
growth factors are estimated as the default, and the growth factor
covariance is estimated as the default because the growth factors are
independent (exogenous) variables.

In the parameterization of the growth model for the inflation part of the
outcome, the intercepts of the outcome variable at the four time points
are held equal as the default. The mean of the intercept growth factor is
fixed at zero. The mean of the slope growth factor and the variances of
the intercept and slope growth factors are estimated as the default, and
the growth factor covariance is estimated as the default because the
growth factors are independent (exogenous) variables.

In this example, the variance of the slope growth factor si for the
inflation part of the outcome is fixed at zero. Because of this, the
covariances among si and all of the other growth factors are fixed at zero
as the default. The covariances among the remaining three growth
factors are estimated as the default.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, three dimensions of integration are used with
a total of 343 integration points. The INTEGRATION option of the
ANALYSIS command is used to change the number of integration points
per dimension from the default of 15 to 7. The ESTIMATOR option of
the ANALYSIS command can be used to select a different estimator.
The OUTPUT command is used to request additional output not
included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis

120
 Examples: Growth Modeling And Survival Analysis

takes. An explanation of the other commands can be found in Example

6.1.

EXAMPLE 6.4: LINEAR GROWTH MODEL FOR A

CATEGORICAL OUTCOME

TITLE: this is an example of a linear growth

model for a categorical outcome
DATA: FILE IS ex6.4.dat;
VARIABLE: NAMES ARE u11-u14 x1 x2 x31-x34;
USEVARIABLES ARE u11-u14;
CATEGORICAL ARE u11-u14;
MODEL: i s | u11@0 u12@1 u13@2 u14@3;

The difference between this example and Example 6.1 is that the
outcome variable is a binary or ordered categorical (ordinal) variable
instead of a continuous variable. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the
example above, u11, u12, u13, and u14 are binary or ordered categorical
variables. They represent the outcome variable measured at four
equidistant occasions.

In the parameterization of the growth model shown here, the thresholds

of the outcome variable at the four time points are held equal as the
default. The mean of the intercept growth factor is fixed at zero. The
mean of the slope growth factor and the variances of the intercept and
slope growth factors are estimated as the default, and the growth factor
covariance is estimated as the default because the growth factors are
independent (exogenous) variables.

The default estimator for this type of analysis is a robust weighted least
squares estimator. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. With the weighted
least squares estimator, the probit model and the default Delta
parameterization for categorical outcomes are used. The scale factor for
the latent response variable of the categorical outcome at the first time
point is fixed at one as the default, while the scale factors for the latent
response variables at the other time points are free to be estimated. If a
maximum likelihood estimator is used, the logistic model for categorical

121
CHAPTER 6

outcomes with a numerical integration algorithm is used (Hedeker &

Gibbons, 1994). Note that numerical integration becomes increasingly
more computationally demanding as the number of factors and the
sample size increase. An explanation of the other commands can be
found in Example 6.1.

EXAMPLE 6.5: LINEAR GROWTH MODEL FOR A

CATEGORICAL OUTCOME USING THE THETA
PARAMETERIZATION

TITLE: this is an example of a linear growth

model for a categorical outcome using the
Theta parameterization
DATA: FILE IS ex6.5.dat;
VARIABLE: NAMES ARE u11-u14 x1 x2 x31-x34;
USEVARIABLES ARE u11-u14;
CATEGORICAL ARE u11-u14;
ANALYSIS: PARAMETERIZATION = THETA;
MODEL: i s | u11@0 u12@1 u13@2 u14@3;

The difference between this example and Example 6.4 is that the Theta
parameterization instead of the default Delta parameterization is used.
In the Delta parameterization, scale factors for the latent response
variables of the observed categorical outcomes are allowed to be
parameters in the model, but residual variances for the latent response
variables are not. In the Theta parameterization, residual variances for
latent response variables are allowed to be parameters in the model, but
scale factors are not. Because the Theta parameterization is used, the
residual variance for the latent response variable at the first time point is
fixed at one as the default, while the residual variances for the latent
response variables at the other time points are free to be estimated. An
explanation of the other commands can be found in Examples 6.1 and
6.4.

122
 Examples: Growth Modeling And Survival Analysis

EXAMPLE 6.6: LINEAR GROWTH MODEL FOR A COUNT

OUTCOME USING A POISSON MODEL

TITLE: this is an example of a linear growth

model for a count outcome using a Poisson
model
DATA: FILE IS ex6.6.dat;
VARIABLE: NAMES ARE u11-u14 x1 x2 x31-x34;
USEVARIABLES ARE u11-u14;
COUNT ARE u11-u14;
MODEL: i s | u11@0 u12@1 u13@2 u14@3;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 6.1 is that the
outcome variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u11, u12, u13, and u14 are count variables. They
represent the outcome variable measured at four equidistant occasions.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables at the four time points are fixed at zero as the
default. The means and variances of the growth factors are estimated as
the default, and the growth factor covariance is estimated as the default
because the growth factors are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors using a numerical integration algorithm. Note that
numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase. In this
example, two dimensions of integration are used with a total of 225
integration points. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. The OUTPUT
command is used to request additional output not included as the default.
The TECH1 option is used to request the arrays containing parameter
specifications and starting values for all free parameters in the model.
The TECH8 option is used to request that the optimization history in
estimating the model be printed in the output. TECH8 is printed to the
screen during the computations as the default. TECH8 screen printing is
useful for determining how long the analysis takes. An explanation of
the other commands can be found in Example 6.1.

123
CHAPTER 6

EXAMPLE 6.7: LINEAR GROWTH MODEL FOR A COUNT

OUTCOME USING A ZERO-INFLATED POISSON MODEL

TITLE: this is an example of a linear growth

model for a count outcome using a zero-
inflated Poisson model
DATA: FILE IS ex6.7.dat;
VARIABLE: NAMES ARE u11-u14 x1 x2 x31-x34;
USEVARIABLES ARE u11-u14;
COUNT ARE u11-u14 (i);
ANALYSIS: INTEGRATION = 7;
MODEL: i s | u11@0 u12@1 u13@2 u14@3;
ii si | u11#1@0 u12#1@1 u13#1@2 u14#1@3;
s@0 si@0;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 6.1 is that the
outcome variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u11, u12, u13, and u14 are count variables. They
represent the outcome variable u1 measured at four equidistant
occasions. The i in parentheses following u11-u14 indicates that a zero-
inflated Poisson model will be estimated.

With a zero-inflated Poisson model, two growth models are estimated.

The first | statement describes the growth model for the count part of the
outcome for individuals who are able to assume values of zero and
above. The second | statement describes the growth model for the
inflation part of the outcome, the probability of being unable to assume
any value except zero. The binary latent inflation variable is referred to
by adding to the name of the count variable the number sign (#) followed
by the number 1.

In the parameterization of the growth model for the count part of the
outcome, the intercepts of the outcome variables at the four time points
are fixed at zero as the default. The means and variances of the growth
factors are estimated as the default, and the growth factor covariance is
estimated as the default because the growth factors are independent
(exogenous) variables.

124
 Examples: Growth Modeling And Survival Analysis

In the parameterization of the growth model for the inflation part of the
outcome, the intercepts of the outcome variable at the four time points
are held equal as the default. The mean of the intercept growth factor is
fixed at zero. The mean of the slope growth factor and the variances of
the intercept and slope growth factors are estimated as the default, and
the growth factor covariance is estimated as the default because the
growth factors are independent (exogenous) variables.

In this example, the variance of the slope growth factor s for the count
part and the slope growth factor si for the inflation part of the outcome
are fixed at zero. Because of this, the covariances among s, si, and the
other growth factors are fixed at zero as the default. The covariance
between the i and ii intercept growth factors is estimated as the default.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 49 integration points. The INTEGRATION option of the
ANALYSIS command is used to change the number of integration points
per dimension from the default of 15 to 7. The ESTIMATOR option of
the ANALYSIS command can be used to select a different estimator.
The OUTPUT command is used to request additional output not
included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes. An explanation of the other commands can be found in Example
6.1.

125
CHAPTER 6

EXAMPLE 6.8: GROWTH MODEL FOR A CONTINUOUS

OUTCOME WITH ESTIMATED TIME SCORES

TITLE: this is an example of a growth model for a

continuous outcome with estimated time
scores
DATA: FILE IS ex6.8.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
MODEL: i s | y11@0 y12@1 y13*2 y14*3;

The difference between this example and Example 6.1 is that two of the
time scores are estimated. The | statement highlighted above shows how
to specify free time scores by using the asterisk (*) to designate a free
parameter. Starting values are specified as the value following the
asterisk (*). For purposes of model identification, two time scores must
be fixed for a growth model with two growth factors. In the example
above, the first two time scores are fixed at zero and one, respectively.
The third and fourth time scores are free to be estimated at starting
values of 2 and 3, respectively. The default estimator for this type of
analysis is maximum likelihood. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 6.1.

EXAMPLE 6.9: QUADRATIC GROWTH MODEL FOR A

CONTINUOUS OUTCOME

TITLE: this is an example of a quadratic growth

model for a continuous outcome
DATA: FILE IS ex6.9.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
MODEL: i s q | y11@0 y12@1 y13@2 y14@3;

126
 Examples: Growth Modeling And Survival Analysis

y11 y12 y13 y14

i s q

The difference between this example and Example 6.1 is that the
quadratic growth model shown in the picture above is estimated. A
quadratic growth model requires three random effects: an intercept
factor (i), a linear slope factor (s), and a quadratic slope factor (q). The |
symbol is used to name and define the intercept and slope factors in the
growth model. The names i, s, and q on the left-hand side of the |
symbol are the names of the intercept, linear slope, and quadratic slope
factors, respectively. In the example above, the linear slope factor has
equidistant time scores of 0, 1, 2, and 3. The time scores for the
quadratic slope factor are the squared values of the linear time scores.
These time scores are automatically computed by the program.

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The means and variances of the three growth factors are
estimated as the default, and the three growth factors are correlated as
the default because they are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 6.1.

127
CHAPTER 6

EXAMPLE 6.10: LINEAR GROWTH MODEL FOR A

CONTINUOUS OUTCOME WITH TIME-INVARIANT AND
TIME-VARYING COVARIATES

TITLE: this is an example of a linear growth

model for a continuous outcome with time-
invariant and time-varying covariates
DATA: FILE IS ex6.10.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 a31-a34;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;
i s ON x1 x2;
y11 ON a31;
y12 ON a32;
y13 ON a33;
y14 ON a34;

y11 y12 y13 y14

i
s

x1 x2 a31 a32 a33 a34

The difference between this example and Example 6.1 is that time-
invariant and time-varying covariates as shown in the picture above are
included in the model.

128
 Examples: Growth Modeling And Survival Analysis

The first ON statement describes the linear regressions of the two

growth factors on the time-invariant covariates x1 and x2. The next four
ON statements describe the linear regressions of the outcome variable on
the time-varying covariates a31, a32, a33, and a34 at each of the four
time points. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 6.1.

EXAMPLE 6.11: PIECEWISE GROWTH MODEL FOR A

CONTINUOUS OUTCOME

TITLE: this is an example of a piecewise growth

model for a continuous outcome
DATA: FILE IS ex6.11.dat;
VARIABLE: NAMES ARE y1-y5;
MODEL: i s1 | y1@0 y2@1 y3@2 y4@2 y5@2;
i s2 | y1@0 y2@0 y3@0 y4@1 y5@2;

y1 y2 y3 y4 y5

i s1 s2

In this example, the piecewise growth model shown in the picture above
is estimated. In a piecewise growth model, different phases of
development are captured by more than one slope growth factor. The
first | statement specifies a linear growth model for the first phase of
development which includes the first three time points. The second |
statement specifies a linear growth model for the second phase of
development which includes the last three time points. Note that there is

129
CHAPTER 6

one intercept growth factor i. It must be named in the specification of

both growth models when using the | symbol.

In the parameterization of the growth models shown here, the intercepts

of the outcome variable at the five time points are fixed at zero as the
default. The means and variances of the three growth factors are
estimated as the default, and the three growth factors are correlated as
the default because they are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 6.1.

EXAMPLE 6.12: GROWTH MODEL WITH INDIVIDUALLY-

VARYING TIMES OF OBSERVATION AND A RANDOM
SLOPE FOR TIME-VARYING COVARIATES FOR A
CONTINUOUS OUTCOME

TITLE: this is an example of a growth model with

individually-varying times of observation
and a random slope for time-varying
covariates for a continuous outcome
DATA: FILE IS ex6.12.dat;
VARIABLE: NAMES ARE y1-y4 x a11-a14 a21-a24;
TSCORES = a11-a14;
ANALYSIS: TYPE = RANDOM;
MODEL: i s | y1-y4 AT a11-a14;
st | y1 ON a21;
st | y2 ON a22;
st | y3 ON a23;
st | y4 ON a24;
i s st ON x;

130
 Examples: Growth Modeling And Survival Analysis

y1 y2 y3 y4

st

x a21 a22 a23 a24

In this example, the growth model with individually-varying times of

observation, a time-invariant covariate, and time-varying covariates with
random slopes shown in the picture above is estimated. The st shown in
a circle represents the random slope. The broken arrows from st to the
arrows from a21 to y1, a22 to y2, a23 to y3, and a24 to y4 indicate that
the slopes in these regressions are random.

The TSCORES option is used to identify the variables in the data set that
contain information about individually-varying times of observation for
the outcomes. The TYPE option is used to describe the type of analysis
that is to be performed. By selecting RANDOM, a growth model with
random slopes will be estimated.

The | symbol is used in conjunction with TYPE=RANDOM to name and

define the random effect variables in the model. The names on the left-
hand side of the | symbol name the random effect variables. In the first |
statement, the AT option is used on the right-hand side of the | symbol to
define a growth model with individually-varying times of observation for

131
CHAPTER 6

the outcome variable. Two growth factors are used in the model, a
random intercept, i, and a random slope, s.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables are fixed at zero as the default. The residual
variances of the outcome variables are free to be estimated as the
default. The residual covariances of the outcome variables are fixed at
zero as the default. The means, variances, and covariances of the
intercept and slope growth factors are free as the default.

The second, third, fourth, and fifth | statements use the ON option to
name and define the random slope variables in the model. The name on
the left-hand side of the | symbol names the random slope variable. The
statement on the right-hand side of the | symbol defines the random slope
variable. In the second | statement, the random slope st is defined by the
linear regression of the dependent variable y1 on the time-varying
covariate a21. In the third | statement, the random slope st is defined by
the linear regression of the dependent variable y2 on the time-varying
covariate a22. In the fourth | statement, the random slope st is defined
by the linear regression of the dependent variable y3 on the time-varying
covariate a23. In the fifth | statement, the random slope st is defined by
the linear regression of the dependent variable y4 on the time-varying
covariate a24. Random slopes with the same name are treated as one
variable during model estimation. The ON statement describes the linear
regressions of the intercept growth factor i, the slope growth factor s,
and the random slope st on the covariate x. The intercepts and residual
variances of, i, s, and st, are free as the default. The residual covariance
between i and s is estimated as the default. The residual covariances
between st and i and s are fixed at zero as the default. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors. The estimator option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 6.1.

132
 Examples: Growth Modeling And Survival Analysis

EXAMPLE 6.13: GROWTH MODEL FOR TWO PARALLEL

PROCESSES FOR CONTINUOUS OUTCOMES WITH
REGRESSIONS AMONG THE RANDOM EFFECTS

TITLE: this is an example of a growth model for

two parallel processes for continuous
outcomes with regressions among the random
effects
DATA: FILE IS ex6.13.dat;
VARIABLE: NAMES ARE y11 y12 y13 y14 y21 y22 y23 y24;
MODEL: i1 s1 | y11@0 y12@1 y13@2 y14@3;
i2 s2 | y21@0 y22@1 y23@2 y24@3;
s1 ON i2;
s2 ON i1;

y11 y12 y13 y14

il sl

i2 s2

y21 y22 y23 y24

133
CHAPTER 6

In this example, the model for two parallel processes shown in the
picture above is estimated. Regressions among the growth factors are
included in the model.

The | statements are used to name and define the intercept and slope
growth factors for the two linear growth models. The names i1 and s1
on the left-hand side of the first | statement are the names of the intercept
and slope growth factors for the first linear growth model. The names i2
and s2 on the left-hand side of the second | statement are the names of
the intercept and slope growth factors for the second linear growth
model. The values on the right-hand side of the two | statements are the
time scores for the two slope growth factors. For both growth models,
the time scores of the slope growth factors are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
factors as initial status factors. The coefficients of the intercept growth
factors are fixed at one as part of the growth model parameterization.
The residual variances of the outcome variables are estimated and
allowed to be different across time, and the residuals are not correlated
as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables at the four time points are fixed at zero as the
default. The means and variances of the intercept growth factors are
estimated as the default, and the intercept growth factor covariance is
estimated as the default because the intercept growth factors are
independent (exogenous) variables. The intercepts and residual
variances of the slope growth factors are estimated as the default, and
the slope growth factors are correlated as the default because residuals
are correlated for latent variables that do not influence any other variable
in the model except their own indicators.

The two ON statements describe the regressions of the slope growth

factor for each process on the intercept growth factor of the other
process. The default estimator for this type of analysis is maximum
likelihood. The ESTIMATOR option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 6.1.

134
 Examples: Growth Modeling And Survival Analysis

EXAMPLE 6.14: MULTIPLE INDICATOR LINEAR GROWTH

MODEL FOR CONTINUOUS OUTCOMES

TITLE: this is an example of a multiple indicator

linear growth model for continuous
outcomes
DATA: FILE IS ex6.14.dat;
VARIABLE: NAMES ARE y11 y21 y31 y12 y22 y32 y13
y23 y33;
MODEL: f1 BY y11
y21-y31 (1-2);
f2 BY y12
y22-y32 (1-2);
f3 BY y13
y23-y33 (1-2);
[y11 y12 y13] (3);
[y21 y22 y23] (4);
[y31 y32 y33] (5);
i s | f1@0 f2@1 f3@2;

y11 y21 y31 y12 y22 y32 y13 y23 y33

fl f2 f3

i s

In this example, the multiple indicator linear growth model for

continuous outcomes shown in the picture above is estimated. The first
BY statement specifies that f1 is measured by y11, y21, and y31. The
second BY statement specifies that f2 is measured by y12, y22, and y32.
The third BY statement specifies that f3 is measured by y13, y23, and

135
CHAPTER 6

y33. The metric of the three factors is set automatically by the program
by fixing the first factor loading in each BY statement to one. This
option can be overridden. The residual variances of the factor indicators
are estimated and the residuals are not correlated as the default.

A multiple indicator growth model requires measurement invariance of

the three factors across time. Measurement invariance is specified by
holding the intercepts and factor loadings of the factor indicators equal
over time. The (1-2) following the factor loadings in the three BY
statements uses the list function to assign equality labels to these
parameters. The label 1 is assigned to the factor loadings of y21, y22,
and y23 which holds these factor loadings equal across time. The label 2
is assigned to the factor loadings of y31, y32, and y33 which holds these
factor loadings equal across time. The factor loadings of y11, y21, and
y31 are fixed at one as described above. The bracket statements refer to
the intercepts. The (3) holds the intercepts of y11, y12, and y13 equal.
The (4) holds the intercepts of y21, y22, and y23 equal. The (5) holds
the intercepts of y31, y32, and y33 equal.

The | statement is used to name and define the intercept and slope factors
in the growth model. The names i and s on the left-hand side of the | are
the names of the intercept and slope growth factors, respectively. The
values on the right-hand side of the | are the time scores for the slope
growth factor. The time scores of the slope growth factor are fixed at 0,
1, and 2 to define a linear growth model with equidistant time points.
The zero time score for the slope growth factor at time point one defines
the intercept growth factor as an initial status factor. The coefficients of
the intercept growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the factors f1, f2, and f3 are
estimated and allowed to be different across time, and the residuals are
not correlated as the default.

In the parameterization of the growth model shown here, the intercepts

of the factors f1, f2, and f3 are fixed at zero as the default. The mean of
the intercept growth factor is fixed at zero and the mean of the slope
growth factor is estimated as the default. The variances of the growth
factors are estimated as the default, and the growth factors are correlated
as the default because they are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select

136
 Examples: Growth Modeling And Survival Analysis

a different estimator. An explanation of the other commands can be

found in Example 6.1.

EXAMPLE 6.15: MULTIPLE INDICATOR LINEAR GROWTH

MODEL FOR CATEGORICAL OUTCOMES

TITLE: this is an example of a multiple indicator

linear growth model for categorical
outcomes
DATA: FILE IS ex6.15.dat;
VARIABLE: NAMES ARE u11 u21 u31 u12 u22 u32
u13 u23 u33;
CATEGORICAL ARE u11 u21 u31 u12 u22 u32
u13 u23 u33;

MODEL: f1 BY u11
u21-u31 (1-2);
f2 BY u12
u22-u32 (1-2);
f3 BY u13
u23-u33 (1-2);
[u11$1 u12$1 u13$1] (3);
[u21$1 u22$1 u23$1] (4);
[u31$1 u32$1 u33$1] (5);
{u11-u31@1 u12-u33};
i s | f1@0 f2@1 f3@2;

The difference between this example and Example 6.14 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. The CATEGORICAL option is used to specify
which dependent variables are treated as binary or ordered categorical
(ordinal) variables in the model and its estimation. In the example
above, all of the factor indicators are categorical variables. The program
determines the number of categories for each indicator.

For binary and ordered categorical factor indicators, thresholds are

modeled rather than intercepts or means. The number of thresholds for a
categorical variable is equal to the number of categories minus one. In
the example above, the categorical variables are binary so they have one
threshold. Thresholds are referred to by adding to the variable name a $
followed by a number. The thresholds of the factor indicators are
referred to as u11$1, u12$1, u13$1, u21$1, u22$1, u23$1, u31$1, u32$1,
and u33$1. Thresholds are referred to in square brackets.

137
CHAPTER 6

The growth model requires measurement invariance of the three factors

across time. Measurement invariance is specified by holding the
thresholds and factor loadings of the factor indicators equal over time.
The (3) after the first bracket statement holds the thresholds of u11, u12,
and u13 equal. The (4) after the second bracket statement holds the
thresholds of u21, u22, and u23 equal. The (5) after the third bracket
statement holds the thresholds of u31, u32, and u33 equal. A list of
observed variables in curly brackets refers to scale factors. The scale
factors for the latent response variables of the categorical outcomes for
the first factor are fixed at one, while the scale factors for the latent
response variables for the other factors are free to be estimated. An
explanation of the other commands can be found in Examples 6.1 and
6.14.

EXAMPLE 6.16: TWO-PART (SEMICONTINUOUS) GROWTH

MODEL FOR A CONTINUOUS OUTCOME

TITLE: this is an example of a two-part

(semicontinuous) growth model for a
continuous outcome
DATA: FILE = ex6.16.dat;
DATA TWOPART:
NAMES = y1-y4;
BINARY = bin1-bin4;
CONTINUOUS = cont1-cont4;
VARIABLE: NAMES = x y1-y4;
USEVARIABLES = bin1-bin4 cont1-cont4;
CATEGORICAL = bin1-bin4;
MISSING = ALL(999);
ANALYSIS: ESTIMATOR = MLR;
MODEL: iu su | bin1@0 bin2@1 bin3@2 bin4@3;
iy sy | cont1@0 cont2@1 cont3@2 cont4@3;
su@0; iu WITH sy@0;
OUTPUT: TECH1 TECH8;

138
 Examples: Growth Modeling And Survival Analysis

u1 u2 u3 u4

iu su

iy sy

y1 y2 y3 y4

In this example, the two-part (semicontinuous) growth model (Olsen &

Schafer, 2001) for a continuous outcome shown in the picture above is
estimated. This is one type of model that can be considered when a
variable has a floor effect, for example, a preponderance of zeroes. The
analysis requires that one binary variable and one continuous variable be
created from the outcome being studied.

The DATA TWOPART command is used to create a binary and a

continuous variable from a variable with a floor effect. In this example,
a set of binary and continuous variables are created using the default
value of zero as the cutpoint. The CUTPOINT option of the DATA
TWOPART command can be used to select another value. The two
variables are created using the following rules:

1. If the value of the original variable is missing, both the new binary
and the new continuous variable values are missing.

139
CHAPTER 6

2. If the value of the original variable is greater than the cutpoint value,
the new binary variable value is one and the new continuous variable
value is the log of the original variable as the default.
3. If the value of the original variable is less than or equal to the
cutpoint value, the new binary variable value is zero and the new
continuous variable value is missing.

The TRANSFORM option of the DATA TWOPART command can be

used to select an alternative to the default log transformation of the new
continuous variables. One choice is no transformation.

The NAMES option of the DATA TWOPART command is used to

identify the variables from the NAMES option of the VARIABLE
command that are used to create a set of binary and continuous variables.
Variables y1, y2, y3, and y4 are used. The BINARY option is used to
assign names to the new set of binary variables. The names for the new
binary variables are bin1, bin2, bin3, and bin4. The CONTINUOUS
option is used to assign names to the new set of continuous variables.
The names for the new continuous variables are cont1, cont2, cont3, and
cont4. The new variables must be placed on the USEVARIABLES
statement of the VARIABLE command if they are used in the analysis.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, bin1, bin2, bin3,
and bin4 are binary variables. The MISSING option is used to identify
the values or symbols in the analysis data set that are to be treated as
missing or invalid. In this example, the number 999 is the missing value
flag. The default is to estimate the model under missing data theory
using all available data. By specifying ESTIMATOR=MLR, a
maximum likelihood estimator with robust standard errors using a
numerical integration algorithm will be used. Note that numerical
integration becomes increasingly more computationally demanding as
the number of growth factors and the sample size increase. In this
example, one dimension of integration is used with a total of 15
integration points. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator.

The first | statement specifies a linear growth model for the binary
outcome. The second | statement specifies a linear growth model for the
continuous outcome. In the parameterization of the growth model for

140
 Examples: Growth Modeling And Survival Analysis

the binary outcome, the thresholds of the outcome variable at the four
time points are held equal as the default. The mean of the intercept
growth factor is fixed at zero. The mean of the slope growth factor and
the variances of the intercept and slope growth factors are estimated as
the default. In this example, the variance of the slope growth factor is
fized at zero for simplicity. In the parameterization of the growth model
for the continuous outcome, the intercepts of the outcome variables at
the four time points are fixed at zero as the default. The means and
variances of the growth factors are estimated as the default, and the
growth factors are correlated as the default because they are independent
(exogenous) variables.

It is often the case that not all growth factor covariances are significant
in two-part growth modeling. Fixing these at zero stabilizes the
estimation. This is why the growth factor covariance between iu and sy
is fixed at zero. The OUTPUT command is used to request additional
output not included as the default. The TECH1 option is used to request
the arrays containing parameter specifications and starting values for all
free parameters in the model. The TECH8 option is used to request that
the optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes. An explanation of the other commands can be found in Example
6.1.

141
CHAPTER 6

EXAMPLE 6.17: LINEAR GROWTH MODEL FOR A

CONTINUOUS OUTCOME WITH FIRST-ORDER AUTO
CORRELATED RESIDUALS USING NON-LINEAR
CONSTRAINTS

TITLE: this is an example of a linear growth

model for a continuous outcome with first-
order auto correlated residuals using non-
linear constraints
DATA: FILE = ex6.17.dat;
VARIABLE: NAMES = y1-y4;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
y1-y4 (resvar);
y1-y3 PWITH y2-y4 (p1);
y1-y2 PWITH y3-y4 (p2);
y1 WITH y4 (p3);
MODEL CONSTRAINT:
NEW (corr);
p1 = resvar*corr;
p2 = resvar*corr**2;
p3 = resvar*corr**3;

The difference between this example and Example 6.1 is that first-order
auto correlated residuals have been added to the model. In a model with
first-order correlated residuals, one residual variance parameter and one
residual auto-correlation parameter are estimated.

In the MODEL command, the label resvar following the residual

variances serves two purposes. It specifies that the residual variances
are held equal to each other and gives that residual variance parameter a
label to be used in the MODEL CONSTRAINT command. The labels
p1, p2, and p3 specify that the residual covariances at adjacent time
points, at adjacent time points once removed, and at adjacent time points
twice removed are held equal. The MODEL CONSTRAINT command
is used to define linear and non-linear constraints on the parameters in
the model. In the MODEL CONSTRAINT command, the NEW option
is used to introduce a new parameter that is not part of the MODEL
command. This residual auto-correlation parameter is referred to as
corr. The p1 parameter constraint specifies that the residual covariances
at adjacent time points are equal to the residual variance parameter
multiplied by the auto-correlation parameter. The p2 parameter

142
 Examples: Growth Modeling And Survival Analysis

constraint specifies that the residual covariances at adjacent time points

once removed are equal to the residual variance parameter multiplied by
the auto-correlation parameter to the power of two. The p3 parameter
constraint specifies that the residual covariance at adjacent time points
twice removed is equal to the residual variance parameter multiplied by
the auto-correlation parameter to the power of three. An explanation of
the other commands can be found in Example 6.1.

EXAMPLE 6.18: MULTIPLE GROUP MULTIPLE COHORT

GROWTH MODEL

TITLE: this is an example of a multiple group

multiple cohort growth model
DATA: FILE = ex6.18.dat;
VARIABLE: NAMES = y1-y4 x a21-a24 g;
GROUPING = g (1 = 1990 2 = 1989 3 = 1988);
MODEL: i s |y1@0 y2@.2 y3@.4 y4@.6;
[i] (1); [s] (2);
i (3); s (4);
i WITH s (5);
i ON x (6);
s ON x (7);
y1 ON a21;
y2 ON a22 (12);
y3 ON a23 (14);
y4 ON a24 (16);
y2-y4 (22-24);
MODEL 1989:
i s |y1@.1 y2@.3 y3@.5 y4@.7;
y1 ON a21;
y2 ON a22;
y3 ON a23;
y4 ON a24;
y1-y4;
MODEL 1988:
i s |y1@.2 y2@.4 y3@.6 y4@.8;
y1 ON a21 (12);
y2 ON a22 (14);
y3 ON a23 (16);
y4 ON a24;
y1-y3 (22-24);
y4;
OUTPUT: TECH1 MODINDICES(3.84);

143
CHAPTER 6

y1 y2 y3 y4

i
s

x a21 a22 a23 a24

In this example, the multiple group multiple cohort growth model shown
in the picture above is estimated. Longitudinal research studies often
collect data on several different groups of individuals defined by their
birth year or cohort. This allows the study of development over a wider
age range than the length of the study and is referred to as an accelerated
or sequential cohort design. The interest in these studies is the
development of an outcome over age not measurement occasion. This
can be handled by rearranging the data so that age is the time axis using
the DATA COHORT command or using a multiple group approach as
described in this example. The advantage of the multiple group
approach is that it can be used to test assumptions of invariance of
growth parameters across cohorts.

In the multiple group approach the variables in the data set represent the
measurement occasions. In this example, there are four measurement
occasions: 2000, 2002, 2004, and 2006. Therefore there are four
variables to represent the outcome. In this example, there are three
cohorts with birth years 1988, 1989, and 1990. It is the combination of
the time of measurement and birth year that determines the ages
represented in the data. This is shown in the table below where rows
represent cohort and columns represent measurement occasion. The

144
 Examples: Growth Modeling And Survival Analysis

entries in the table represent the ages. In this example, ages 10 to 18 are
represented.

M.O./ 2000 2002 2004 2006

Cohort
1988 12 14 16 18
1989 11 13 15 17
1990 10 12 14 16

The model that is estimated uses the time axis of age as shown in the
table below where rows represent cohort and columns represent age.
The entries for the first three rows in the table are the years of the
measurement occasions. The entries for the last row are the time scores
for a linear model.

Age/
10 11 12 13 14 15 16 17 18
Cohort
1988 2000 2002 2004 2006
1989 2000 2002 2004 2006
1990 2000 2002 2004 2006
Time
0 .1 .2 .3 .4 .5 .6 .7 .8
Score

As shown in the table, three ages are represented by more than one
cohort. Age 12 is represented by cohorts 1988 and 1990 measured in
2000 and 2002; age 14 is represented by cohorts 1988 and 1990
measured in 2002 and 2004; and age 16 is represented by cohorts 1988
and 1990 measured in 2004 and 2006. This information is needed to
constrain parameters to be equal in the multiple group model.

The table also provides information about the time scores for each
cohort. The time scores are obtained as the difference in age between
measurement occasions divided by ten. The division is used to avoid
large time scores which can lead to convergence problems. Cohort 1990
provides information for ages 10, 12, 14, and 16. The time scores for
cohort 2000 are 0, .2, .4, and .6. Cohort 1989 provides information for
ages 11, 13, 15, and 17. The time scores for cohort 1989 are .1, .3, .5,
and .7. Cohort 1988 provides information for ages 12, 14, 16, and 18.
The time scores for cohort 1988 are .2, .4, .6, and .8.

145
CHAPTER 6

The GROUPING option is used to identify the variable in the data set
that contains information on group membership when the data for all
groups are stored in a single data set. The information in parentheses
after the grouping variable name assigns labels to the values of the
grouping variable found in the data set. In the example above,
observations with g equal to 1 will be assigned the label 1990,
individuals with g equal to 2 will be assigned the label 1989, and
individuals with g equal to 3 will be assigned the label 1988. These
labels are used in conjunction with the MODEL command to specify
model statements specific to each group.

In multiple group analysis, two variations of the MODEL command are

used. They are MODEL and MODEL followed by a label. MODEL
describes the overall model to be estimated for each group. MODEL
followed by a label describes differences between the overall model and
the model for the group designated by the label. In the MODEL
command, the | symbol is used to name and define the intercept and
slope factors in a growth model. The names i and s on the left-hand side
of the | symbol are the names of the intercept and slope growth factors,
respectively. The statement on the right-hand side of the | symbol
specifies the outcome and the time scores for the growth model. The
time scores for the slope growth factor are fixed at 0, .2, .4, and .6.
These are the time scores for cohort 1990. The zero time score for the
slope growth factor at time point one defines the intercept growth factor
as an initial status factor for age 10. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are
estimated and allowed to be different across age and the residuals are not
correlated as the default. The time scores for the other two cohorts are
specified in the group-specific MODEL commands. The group-specific
MODEL command for cohort 1989 fixes the time scores at .1, .3, .5, and
.7. The group-specific MODEL command for cohort 1988 fixes the time
scores at .2, .4, .6, and .8.

The equalities specified by the numbers in parentheses represent the

baseline assumption that the cohorts come from the same population.
Equalities specified in the overall MODEL command constrain
parameters to be equal across all groups. All parameters related to the
growth factors are constrained to be equal across all groups. Other
parameters are held equal when an age is represented by more than one
cohort. For example, the ON statement with the (12) equality in the

146
 Examples: Growth Modeling And Survival Analysis

overall MODEL command describes the linear regression of y2 on the

time-varying covariate a22 for cohort 1990 at age 12. In the group-
specific MODEL command for cohort 1988, the ON statement with the
(12) equality describes the linear regression of y1 on the time-varying
covariate a21 for cohort 1988 at age 12. Other combinations of cohort
and age do not involve equality constraints. Cohort 1990 is the only
cohort that represents age 10; cohort 1989 is the only cohort that
represents ages 11, 13, 15, 17; and cohort 1988 is the only cohort that
represents age 18. Statements in the group-specific MODEL commands
relax equality constraints specified in the overall MODEL command.
An explanation of the other commands can be found in Example 6.1.

EXAMPLE 6.19: DISCRETE-TIME SURVIVAL ANALYSIS

TITLE: this is an example of a discrete-time

survival analysis
DATA: FILE IS ex6.19.dat;
VARIABLE: NAMES ARE u1-u4 x;
CATEGORICAL = u1-u4;
MISSING = ALL (999);
ANALYSIS: ESTIMATOR = MLR;
MODEL: f BY u1-u4@1;
f ON x;
f@0;

u1 u2 u3 u4

x f

In this example, the discrete-time survival analysis model shown in the

picture above is estimated. Each u variable represents whether or not a
single non-repeatable event has occurred in a specific time period. The
value 1 means that the event has occurred, 0 means that the event has not

147
CHAPTER 6

occurred, and a missing value flag means that the event has occurred in a
preceding time period or that the individual has dropped out of the study
(Muthén & Masyn, 2005). The factor f is used to specify a proportional
odds assumption for the hazards of the event.

The MISSING option is used to identify the values or symbols in the

analysis data set that are to be treated as missing or invalid. In this
example, the number 999 is the missing value flag. The default is to
estimate the model under missing data theory using all available data.
The default estimator for this type of analysis is a robust weighted least
squares estimator. By specifying ESTIMATOR=MLR, maximum
likelihood estimation with robust standard errors is used. The BY
statement specifies that f is measured by u1, u2, u3, and u4 where the
factor loadings are fixed at one. This represents a proportional odds
assumption where the covariate x has the same influence on u1, u2, u3,
and u4. The ON statement describes the linear regression of f on the
covariate x. The residual variance of f is fixed at zero to correspond to a
conventional discrete-time survival model. An explanation of the other
commands can be found in Example 6.1.

EXAMPLE 6.20: CONTINUOUS-TIME SURVIVAL ANALYSIS

USING THE COX REGRESSION MODEL

TITLE: this is an example of a continuous-time

survival analysis using the Cox regression
model
DATA: FILE = ex6.20.dat;
VARIABLE: NAMES = t x tc;
SURVIVAL = t (ALL);
TIMECENSORED = tc (0 = NOT 1 = RIGHT);
ANALYSIS: BASEHAZARD = OFF;
MODEL: t ON x;

x t

In this example, the continuous-time survival analysis model shown in

the picture above is estimated. This is the Cox regression model (Singer
& Willett, 2003). The profile likelihood method is used for model
estimation (Asparouhov et al., 2006).

148
 Examples: Growth Modeling And Survival Analysis

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option must be used in conjunction with the
TIMECENSORED option. In this example, t is the variable that
contains time-to-event information. By specifying the keyword ALL in
parenthesis following the time-to-event variable, the time intervals are
taken from the data. The TIMECENSORED option is used to identify
the variables that contain information about right censoring. In this
example, the variable is named tc. The information in parentheses
specifies that the value zero represents no censoring and the value one
represents right censoring. This is the default. The BASEHAZARD
option of the ANALYSIS command is used with continuous-time
survival analysis to specify if a non-parametric or a parametric baseline
hazard function is used in the estimation of the model. The setting OFF
specifies that a non-parametric baseline hazard function is used. This is
the default.

In the MODEL command, the ON statement describes the loglinear

regression of the time-to-event variable t on the covariate x. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors. The estimator option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 6.1.

EXAMPLE 6.21: CONTINUOUS-TIME SURVIVAL ANALYSIS

USING A PARAMETRIC PROPORTIONAL HAZARDS MODEL

TITLE: this is an example of a continuous-time

survival analysis using a parametric
proportional hazards model
DATA: FILE = ex6.21.dat;
VARIABLE: NAMES = t x tc;
SURVIVAL = t(20*1);
TIMECENSORED = tc (0 = NOT 1 = RIGHT);
ANALYSIS: BASEHAZARD = ON;
MODEL: [t#1-t#21];
t ON x;

The difference between this example and Example 6.20 is that a

parametric proportional hazards model is used instead of a Cox

149
CHAPTER 6

regression model. In contrast to the Cox regression model, the

parametric model estimates parameters and their standard errors for the
baseline hazard function (Asparouhov et al., 2006).

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option must be used in conjunction with the
TIMECENSORED option. In this example, t is the variable that
contains time-to-event information. The numbers in parentheses
following the time-to-event variable specify that twenty time intervals of
length one are used in the analysis for the baseline hazard function. The
TIMECENSORED option is used to identify the variables that contain
information about right censoring. In this example, this variable is
named tc. The information in parentheses specifies that the value zero
represents no censoring and the value one represents right censoring.
This is the default.

The BASEHAZARD option of the ANALYSIS command is used with

continuous-time survival analysis to specify if a non-parametric or a
parametric baseline hazard function is used in the estimation of the
model. The setting ON specifies that a parametric baseline hazard
function is used. When the parametric baseline hazard function is used,
the baseline hazard parameters can be used in the MODEL command.
There are as many baseline hazard parameters are there are time
intervals plus one. These parameters can be referred to in the MODEL
command by adding to the name of the time-to-event variable the
number sign (#) followed by a number. In the MODEL command, the
bracket statement specifies that the 21 baseline hazard parameters are
part of the model.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors. The estimator option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Examples 6.1 and 6.20.

150
 Examples: Growth Modeling And Survival Analysis

EXAMPLE 6.22: CONTINUOUS-TIME SURVIVAL ANALYSIS

USING A PARAMETRIC PROPORTIONAL HAZARDS MODEL
WITH A FACTOR INFLUENCING SURVIVAL

TITLE: this is an example of a continuous-time

survival analysis using a parametric
proportional hazards model with a factor
influencing survival
DATA: FILE = ex6.22.dat;
VARIABLE: NAMES = t u1-u4 x tc;
SURVIVAL = t (20*1);
TIMECENSORED = tc;
CATEGORICAL = u1-u4;
ANALYSIS: ALGORITHM = INTEGRATION;
BASEHAZARD = ON;
MODEL: f BY u1-u4;
[t#1-t#21];
t ON x f;
f ON x;
OUTPUT: TECH1 TECH8;

u1 u2 u3 u4 t

In this example, the continuous-time survival analysis model shown in

the picture above is estimated. The model is similar to Larsen (2005)

151
CHAPTER 6

although in this example the analysis uses a parametric baseline hazard

function (Asparouhov et al., 2006).

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with a total of 15 integration points. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

In the MODEL command the BY statement specifies that f is measured

by the binary indicators u1, u2, u3, and u4. The bracket statement
specifies that the 21 baseline hazard parameters are part of the model.
The first ON statement describes the loglinear regression of the time-to-
event variable t on the covariate x and the factor f. The second ON
statement describes the linear regression of f on the covariate x. An
explanation of the other commands can be found in Examples 6.1 and
6.21.

152
 Examples: Mixture Modeling With Cross-Sectional Data

CHAPTER 7
EXAMPLES: MIXTURE
MODELING WITH CROSS-
SECTIONAL DATA

Mixture modeling refers to modeling with categorical latent variables

that represent subpopulations where population membership is not
known but is inferred from the data. This is referred to as finite mixture
modeling in statistics (McLachlan & Peel, 2000). A special case is
latent class analysis (LCA) where the latent classes explain the
relationships among the observed dependent variables similar to factor
analysis. In contrast to factor analysis, however, LCA provides
classification of individuals. In addition to conventional exploratory
LCA, confirmatory LCA and LCA with multiple categorical latent
variables can be estimated. In Mplus, mixture modeling can be applied
to any of the analyses discussed in the other example chapters including
regression analysis, path analysis, confirmatory factor analysis (CFA),
item response theory (IRT) analysis, structural equation modeling
(SEM), growth modeling, survival analysis, and multilevel modeling.
Observed dependent variables can be continuous, censored, binary,
ordered categorical (ordinal), unordered categorical (nominal), counts,
or combinations of these variable types. LCA and general mixture
models can be extended to include continuous latent variables. An
overview can be found in Muthén (2008).

LCA is a measurement model. A general mixture model has two parts: a

measurement model and a structural model. The measurement model for
LCA and the general mixture model is a multivariate regression model
that describes the relationships between a set of observed dependent
variables and a set of categorical latent variables. The observed
dependent variables are referred to as latent class indicators. The
relationships are described by a set of linear regression equations for
continuous latent class indicators, a set of censored normal or censored-
inflated normal regression equations for censored latent class indicators,
a set of logistic regression equations for binary or ordered categorical
latent class indicators, a set of multinomial logistic regressions for
unordered categorical latent class indicators, and a set of Poisson or

153
CHAPTER 7

zero-inflated Poisson regression equations for count latent class

indicators.

The structural model describes three types of relationships in one set of

multivariate regression equations: the relationships among the
categorical latent variables, the relationships among observed variables,
and the relationships between the categorical latent variables and
observed variables that are not latent class indicators. These
relationships are described by a set of multinomial logistic regression
equations for the categorical latent dependent variables and unordered
observed dependent variables, a set of linear regression equations for
continuous observed dependent variables, a set of censored normal or
censored normal regression equations for censored-inflated observed
dependent variables, a set of logistic regression equations for binary or
ordered categorical observed dependent variables, and a set of Poisson
or zero-inflated Poisson regression equations for count observed
dependent variables. For logistic regression, ordered categorical
variables are modeled using the proportional odds specification.
Maximum likelihood estimation is used.

The general mixture model can be extended to include continuous latent

variables. The measurement and structural models for continuous latent
variables are described in Chapter 5. In the extended general mixture
model, relationships between categorical and continuous latent variables
are allowed. These relationships are described by a set of multinomial
logistic regression equations for the categorical latent dependent
variables and a set of linear regression equations for the continuous
latent dependent variables.

In mixture modeling, some starting values may result in local solutions

that do not represent the global maximum of the likelihood. To avoid
this, different sets of starting values are automatically produced and the
solution with the best likelihood is reported.

All cross-sectional mixture models can be estimated using the following

special features:

• Single or multiple group analysis

• Missing data
• Complex survey data

154
 Examples: Mixture Modeling With Cross-Sectional Data

• Latent variable interactions and non-linear factor analysis using

maximum likelihood
• Random slopes
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities
• Test of equality of means across latent classes using posterior
probability-based multiple imputations

For TYPE=MIXTURE, multiple group analysis is specified by using the

KNOWNCLASS option of the VARIABLE command. The default is to
estimate the model under missing data theory using all available data.
The LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. The
SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed. Latent variable interactions
are specified by using the | symbol of the MODEL command in
conjunction with the XWITH option of the MODEL command. Random
slopes are specified by using the | symbol of the MODEL command in
conjunction with the ON option of the MODEL command. Linear and
non-linear parameter constraints are specified by using the MODEL
CONSTRAINT command. Indirect effects are specified by using the
MODEL INDIRECT command. Maximum likelihood estimation is
specified by using the ESTIMATOR option of the ANALYSIS
command. Bootstrap standard errors are obtained by using the
BOOTSTRAP option of the ANALYSIS command. Bootstrap
confidence intervals are obtained by using the BOOTSTRAP option of
the ANALYSIS command in conjunction with the CINTERVAL option
of the OUTPUT command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test. The
AUXILIARY option is used to test the equality of means across latent
classes using posterior probability-based multiple imputations.

155
CHAPTER 7

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, plots of
sample and estimated means and proportions/probabilities, and plots of
estimated probabilities for a categorical latent variable as a function of
its covariates. These are available for the total sample, by group, by
class, and adjusted for covariates. The PLOT command includes
a display showing a set of descriptive statistics for each variable. The
graphical displays can be edited and exported as a DIB, EMF, or JPEG
file. In addition, the data for each graphical display can be saved in an
external file for use by another graphics program.

Following is the set of examples included in this chapter.

• 7.1: Mixture regression analysis for a continuous dependent

variable using automatic starting values with random starts
• 7.2: Mixture regression analysis for a count variable using a zero-
inflated Poisson model using automatic starting values with random
starts
• 7.3: LCA with binary latent class indicators using automatic starting
values with random starts
• 7.4: LCA with binary latent class indicators using user-specified
starting values without random starts
• 7.5: LCA with binary latent class indicators using user-specified
starting values with random starts
• 7.6: LCA with three-category latent class indicators using user-
specified starting values without random starts
• 7.7: LCA with unordered categorical latent class indicators using
automatic starting values with random starts
• 7.8: LCA with unordered categorical latent class indicators using
user-specified starting values with random starts
• 7.9: LCA with continuous latent class indicators using automatic
starting values with random starts
• 7.10: LCA with continuous latent class indicators using user-
specified starting values without random starts
• 7.11: LCA with binary, censored, unordered, and count latent class
indicators using user-specified starting values without random starts
• 7.12: LCA with binary latent class indicators using automatic
starting values with random starts with a covariate and a direct effect

156
 Examples: Mixture Modeling With Cross-Sectional Data

• 7.13: Confirmatory LCA with binary latent class indicators and

parameter constraints
• 7.14: Confirmatory LCA with two categorical latent variables
• 7.15: Loglinear model for a three-way table with conditional
independence between the first two variables
• 7.16: LCA with partial conditional independence*
• 7.17: Mixture CFA modeling
• 7.18: LCA with a second-order factor (twin analysis)*
• 7.19: SEM with a categorical latent variable regressed on a
continuous latent variable*
• 7.20: Structural equation mixture modeling
• 7.21: Mixture modeling with known classes (multiple group
analysis)
• 7.22: Mixture modeling with continuous variables that correlate
within class
• 7.23: Mixture randomized trials modeling using CACE estimation
with training data
• 7.24: Mixture randomized trials modeling using CACE estimation
with missing data on the latent class indicator
• 7.25: Zero-inflated Poisson regression carried out as a two-class
model
• 7.26: CFA with a non-parametric representation of a non-normal
factor distribution
• 7.27: Factor (IRT) mixture analysis with binary latent class and
factor indicators*
• 7.28: Two-group twin model for categorical outcomes using
maximum likelihood and parameter constraints*
• 7.29: Two-group IRT twin model for factors with categorical factor
indicators using parameter constraints*
• 7.30: Continuous-time survival analysis using a Cox regression
model to estimate a treatment effect

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

157
CHAPTER 7

EXAMPLE 7.1: MIXTURE REGRESSION ANALYSIS FOR A

CONTINUOUS DEPENDENT VARIABLE USING AUTOMATIC
STARTING VALUES WITH RANDOM STARTS

TITLE: this is an example of a mixture regression

analysis for a continuous dependent
variable using automatic starting values
with random starts
DATA: FILE IS ex7.1.dat;
VARIABLE: NAMES ARE y x1 x2;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
y ON x1 x2;
c ON x1;
%c#2%
y ON x2;
y;
OUTPUT: TECH1 TECH8;

x1 x2

158
 Examples: Mixture Modeling With Cross-Sectional Data

In this example, the mixture regression model for a continuous

dependent variable shown in the picture above is estimated using
automatic starting values with random starts. Because c is a categorical
latent variable, the interpretation of the picture is not the same as for
models with continuous latent variables. The arrow from c to y indicates
that the intercept of y varies across the classes of c. This corresponds to
the regression of y on a set of dummy variables representing the
categories of c. The broken arrow from c to the arrow from x2 to y
indicates that the slope in the regression of y on x2 varies across the
classes of c. The arrow from x1 to c represents the multinomial logistic
regression of c on x1.

TITLE: this is an example of a mixture regression

analysis for a continuous dependent
variable

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex7.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex7.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y x1 x2;

CLASSES = c (2);

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains three variables: y, x1, and x2. The CLASSES option
is used to assign names to the categorical latent variables in the model
and to specify the number of latent classes in the model for each
categorical latent variable. In the example above, there is one
categorical latent variable c that has two latent classes.

ANALYSIS: TYPE = MIXTURE;

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis that

159
CHAPTER 7

is to be performed. By selecting MIXTURE, a mixture model will be

estimated.

When TYPE=MIXTURE is specified, either user-specified or automatic

starting values are used to create randomly perturbed sets of starting
values for all parameters in the model except variances and covariances.
In this example, the random perturbations are based on automatic
starting values. Maximum likelihood optimization is done in two stages.
In the initial stage, 20 random sets of starting values are generated. An
optimization is carried out for ten iterations using each of the 20 random
sets of starting values. The ending values from the 4 optimizations with
the highest loglikelihoods are used as the starting values in the final
stage optimizations which are carried out using the default optimization
settings for TYPE=MIXTURE. A more thorough investigation of
multiple solutions can be carried out using the STARTS and
STITERATIONS options of the ANALYSIS command.

MODEL:
%OVERALL%
y ON x1 x2;
c ON x1;
%c#2%
y ON x2;
y;

The MODEL command is used to describe the model to be estimated.

For mixture models, there is an overall model designated by the label
%OVERALL%. The overall model describes the part of the model that
is in common for all latent classes. The part of the model that differs for
each class is specified by a label that consists of the categorical latent
variable followed by the number sign followed by the class number. In
the example above, the label %c#2% refers to the part of the model for
class 2 that differs from the overall model.

In the overall model, the first ON statement describes the linear

regression of y on the covariates x1 and x2. The second ON statement
describes the multinomial logistic regression of the categorical latent
variable c on the covariate x1 when comparing class 1 to class 2. The
intercept in the regression of c on x1 is estimated as the default.

In the model for class 2, the ON statement describes the linear regression
of y on the covariate x2. This specification relaxes the default equality

160
 Examples: Mixture Modeling With Cross-Sectional Data

constraint for the regression coefficient. By mentioning the residual

variance of y, it is not held equal across classes. The intercepts in class
1 and class 2 are free and unequal as the default. The default estimator
for this type of analysis is maximum likelihood with robust standard
errors. The ESTIMATOR option of the ANALYSIS command can be
used to select a different estimator.

Following is an alternative specification of the multinomial logistic

regression of c on the covariate x1:

c#1 ON x1;

where c#1 refers to the first class of c. The classes of a categorical latent
variable are referred to by adding to the name of the categorical latent
variable the number sign (#) followed by the number of the class. This
alternative specification allows individual parameters to be referred to in
the MODEL command for the purpose of giving starting values or
placing restrictions.

OUTPUT: TECH1 TECH8;

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes.

161
CHAPTER 7

EXAMPLE 7.2: MIXTURE REGRESSION ANALYSIS FOR A

COUNT VARIABLE USING A ZERO-INFLATED POISSON
MODEL USING AUTOMATIC STARTING VALUES WITH
RANDOM STARTS

TITLE: this is an example of a mixture regression

analysis for a count variable using a
zero-inflated Poisson model using
automatic starting values with random
starts
DATA: FILE IS ex7.2.dat;
VARIABLE: NAMES ARE u x1 x2;
CLASSES = c (2);
COUNT = u (i);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
u ON x1 x2;
u#1 ON x1 x2;
c ON x1;
%c#2%
u ON x2;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 7.1 is that the
dependent variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u is a count variable. The i in parentheses following u
indicates that a zero-inflated Poisson model will be estimated.

With a zero-inflated Poisson model, two regressions are estimated. In

the overall model, the first ON statement describes the Poisson
regression of the count part of u on the covariates x1 and x2. This
regression predicts the value of the count dependent variable for
individuals who are able to assume values of zero and above. The
second ON statement describes the logistic regression of the binary
latent inflation variable u#1 on the covariates x1 and x2. This
regression describes the probability of being unable to assume any value
except zero. The inflation variable is referred to by adding to the name
of the count variable the number sign (#) followed by the number 1. The

162
 Examples: Mixture Modeling With Cross-Sectional Data

third ON statement specifies the multinomial logistic regression of the

categorical latent variable c on the covariate x1 when comparing class 1
to class 2. The intercept in the regression of c on x1 is estimated as the
default.

In the model for class 2, the ON statement describes the Poisson

regression of the count part of u on the covariate x2. This specification
relaxes the default equality constraint for the regression coefficient. The
intercepts of u are free and unequal across classes as the default. All
other parameters are held equal across classes as the default. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 7.1.

EXAMPLE 7.3: LCA WITH BINARY LATENT CLASS

INDICATORS USING AUTOMATIC STARTING VALUES
WITH RANDOM STARTS

TITLE: this is an example of a LCA with binary

latent class indicators using automatic
starting values with random starts
DATA: FILE IS ex7.3.dat;
VARIABLE: NAMES ARE u1-u4 x1-x10;
USEVARIABLES = u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
AUXILIARY = x1-x10 (R3STEP);
ANALYSIS: TYPE = MIXTURE;
OUTPUT: TECH1 TECH8 TECH10;

163
CHAPTER 7

u1

u2

u3

u4

In this example, the latent class analysis (LCA) model with binary latent
class indicators shown in the picture above is estimated using automatic
starting values and random starts. Because c is a categorical latent
variable, the interpretation of the picture is not the same as for models
with continuous latent variables. The arrows from c to the latent class
indicators u1, u2, u3, and u4 indicate that the thresholds of the latent
class indicators vary across the classes of c. This implies that the
probabilities of the latent class indicators vary across the classes of c.
The arrows correspond to the regressions of the latent class indicators on
a set of dummy variables representing the categories of c.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, the latent class
indicators u1, u2, u3, and u4, are binary or ordered categorical variables.
The program determines the number of categories for each indicator.
The AUXILIARY option is used to specify variables that are not part of
the analysis that are important predictors of latent classes using a three-
step approach (Vermunt, 2010; Asparouhov & Muthén, 2012b). The
letters R3STEP in parentheses is placed behind the variables in the
AUXILIARY statement that that will be used as covariates in the third
step multinomial logistic regression in a mixture model.

164
 Examples: Mixture Modeling With Cross-Sectional Data

The MODEL command does not need to be specified when automatic

starting values are used. The thresholds of the observed variables and
the mean of the categorical latent variable are estimated as the default.
The thresholds are not held equal across classes as the default. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator.

The TECH10 option is used to request univariate, bivariate, and

response pattern model fit information for the categorical dependent
variables in the model. This includes observed and estimated (expected)
frequencies and standardized residuals. An explanation of the other
commands can be found in Example 7.1.

EXAMPLE 7.4: LCA WITH BINARY LATENT CLASS

INDICATORS USING USER-SPECIFIED STARTING VALUES
WITHOUT RANDOM STARTS

TITLE: this is an example of a LCA with binary

latent class indicators using user-
specified starting values without random
starts
DATA: FILE IS ex7.4.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
STARTS = 0;
MODEL:
%OVERALL%
%c#1%
[u1$1*1 u2$1*1 u3$1*-1 u4$1*-1];
%c#2%
[u1$1*-1 u2$1*-1 u3$1*1 u4$1*1];
OUTPUT: TECH1 TECH8;

The differences between this example and Example 7.3 are that user-
specified starting values are used instead of automatic starting values
and there are no random starts. By specifying STARTS=0 in the
ANALYSIS command, random starts are turned off.

165
CHAPTER 7

In the MODEL command, user-specified starting values are given for the
thresholds of the binary latent class indicators. For binary and ordered
categorical dependent variables, thresholds are referred to by adding to a
variable name a dollar sign ($) followed by a threshold number. The
number of thresholds is equal to the number of categories minus one.
Because the latent class indicators are binary, they have one threshold.
The thresholds of the latent class indicators are referred to as u1$1,
u2$1, u3$1, and u4$1. Square brackets are used to specify starting
values in the logit scale for the thresholds of the binary latent class
indicators. The asterisk (*) is used to assign a starting value. It is placed
after a variable with the starting value following it. In the example
above, the threshold of u1 is assigned the starting value of 1 for class 1
and -1 for class 2. The threshold of u4 is assigned the starting value of -
1 for class 1 and 1 for class 2. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1 and 7.3.

EXAMPLE 7.5: LCA WITH BINARY LATENT CLASS

INDICATORS USING USER-SPECIFIED STARTING VALUES
WITH RANDOM STARTS

TITLE: this is an example of a LCA with binary

latent class indicators using user-
specified starting values with random
starts
DATA: FILE IS ex7.5.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
STARTS = 100 10;
STITERATIONS = 20;
MODEL:
%OVERALL%
%c#1%
[u1$1*1 u2$1*1 u3$1*-1 u4$1*-1];
%c#2%
[u1$1*-1 u2$1*-1 u3$1*1 u4$1*1];
OUTPUT: TECH1 TECH8;

166
 Examples: Mixture Modeling With Cross-Sectional Data

The difference between this example and Example 7.4 is that random
starts are used. In this example, the random perturbations are based on
user-specified starting values. The STARTS option is used to specify
the number of initial stage random sets of starting values to generate and
the number of final stage optimizations to use. The default is 20 random
sets of starting values for the initial stage and 4 optimizations for the
final stage. In the example above, the STARTS option specifies that 100
random sets of starting values for the initial stage and 10 final stage
optimizations will be used. The STITERATIONS option is used to
specify the maximum number of iterations allowed in the initial stage.
In this example, 20 iterations are allowed in the initial stage instead of
the default of 10. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Examples 7.1, 7.3, and 7.4.

EXAMPLE 7.6: LCA WITH THREE-CATEGORY LATENT

CLASS INDICATORS USING USER-SPECIFIED STARTING
VALUES WITHOUT RANDOM STARTS

TITLE: this is an example of a LCA with three-

category latent class indicators using
user-specified starting values without
random starts
DATA: FILE IS ex7.6.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
STARTS = 0;
MODEL:
%OVERALL%
%c#1%
[u1$1*.5 u2$1*.5 u3$1*-.5 u4$1*-.5];
[u1$2*1 u2$2*1 u3$2*0 u4$2*0];
%c#2%
[u1$1*-.5 u2$1*-.5 u3$1*.5 u4$1*.5];
[u1$2*0 u2$2*0 u3$2*1 u4$2*1];
OUTPUT: TECH1 TECH8;

167
CHAPTER 7

The difference between this example and Example 7.4 is that the latent
class indicators are ordered categorical (ordinal) variables with three
categories instead of binary variables. When latent class indicators are
ordered categorical variables, each latent class indicator has more than
one threshold. The number of thresholds is equal to the number of
categories minus one. When user-specified starting values are used, they
must be specified for all thresholds and they must be in increasing order
for each variable within each class. For example, in class 1 the threshold
starting values for latent class indicator u1 are .5 for the first threshold
and 1 for the second threshold. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1, 7.3, and 7.4.

EXAMPLE 7.7: LCA WITH UNORDERED CATEGORICAL

LATENT CLASS INDICATORS USING AUTOMATIC
STARTING VALUES WITH RANDOM STARTS

TITLE: this is an example of a LCA with unordered

categorical latent class indicators using
automatic starting values with random
starts
DATA: FILE IS ex7.7.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
NOMINAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 7.3 is that the latent
class indicators are unordered categorical (nominal) variables instead of
binary variables. The NOMINAL option is used to specify which
dependent variables are treated as unordered categorical (nominal)
variables in the model and its estimation. In the example above, u1, u2,
u3, and u4 are three-category unordered variables. The categories of an
unordered categorical variable are referred to by adding to the name of
the unordered categorical variable the number sign (#) followed by the
number of the category. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different

168
 Examples: Mixture Modeling With Cross-Sectional Data

estimator. An explanation of the other commands can be found in

Examples 7.1 and 7.3.

EXAMPLE 7.8: LCA WITH UNORDERED CATEGORICAL

LATENT CLASS INDICATORS USING USER-SPECIFIED
STARTING VALUES WITH RANDOM STARTS

TITLE: this is an example of a LCA with unordered

categorical latent class indicators using
user-specified starting values with random
starts
DATA: FILE IS ex7.8.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
NOMINAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL: %OVERALL%
%c#1%
[u1#1-u4#1*0];
[u1#2-u4#2*1];
%c#2%
[u1#1-u4#1*-1];
[u1#2-u4#2*-1];
OUTPUT: TECH1 TECH8;

The difference between this example and Example 7.7 is that user-
specified starting values are used instead of automatic starting values.
Means are referred to by using bracket statements. The categories of an
unordered categorical variable are referred to by adding to the name of
the unordered categorical variable the number sign (#) followed by the
number of the category. In this example, u1#1 refers to the first category
of u1 and u1#2 refers to the second category of u1. Starting values of 0
and 1 are given for the means in class 1 and starting values of -1 are
given for the means in class 2. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1, 7.3, and 7.7.

169
CHAPTER 7

EXAMPLE 7.9: LCA WITH CONTINUOUS LATENT CLASS

INDICATORS USING AUTOMATIC STARTING VALUES
WITH RANDOM STARTS

TITLE: this is an example of a LCA with

continuous latent class indicators using
automatic starting values with random
starts
DATA: FILE IS ex7.9.dat;
VARIABLE: NAMES ARE y1-y4;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
OUTPUT: TECH1 TECH8;

y1

y2

y3

y4

The difference between this example and Example 7.3 is that the latent
class indicators are continuous variables instead of binary variables.
When there is no specification in the VARIABLE command regarding
the scale of the dependent variables, it is assumed that they are
continuous. Latent class analysis with continuous latent class indicators
is often referred to as latent profile analysis.

170
 Examples: Mixture Modeling With Cross-Sectional Data

The MODEL command does not need to be specified when automatic

starting values are used. The means and variances of the latent class
indicators and the mean of the categorical latent variable are estimated
as the default. The means of the latent class indicators are not held
equal across classes as the default. The variances are held equal across
classes as the default and the covariances among the latent class
indicators are fixed at zero as the default. The default estimator for this
type of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1 and 7.3.

EXAMPLE 7.10: LCA WITH CONTINUOUS LATENT CLASS

INDICATORS USING USER-SPECIFIED STARTING VALUES
WITHOUT RANDOM STARTS

TITLE: this is an example of a LCA with

continuous latent class indicators using
user-specified starting values without
random starts
DATA: FILE IS ex7.10.dat;
VARIABLE: NAMES ARE y1-y4;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
STARTS = 0;
MODEL:
%OVERALL%
%c#1%
[y1–y4*1];
y1-y4;
%c#2%
[y1–y4*-1];
y1-y4;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 7.4 is that the latent
class indicators are continuous variables instead of binary variables. As
a result, starting values are given for means instead of thresholds.

The means and variances of the latent class indicators and the mean of
the categorical latent variable are estimated as the default. In the models
for class 1 and class 2, by mentioning the variances of the latent class

171
CHAPTER 7

indicators, the default constraint of equality of variances across classes

is relaxed. The covariances among the latent class indicators within
class are fixed at zero as the default. The default estimator for this type
of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1 and 7.4.

EXAMPLE 7.11: LCA WITH BINARY, CENSORED,

UNORDERED, AND COUNT LATENT CLASS INDICATORS
USING USER-SPECIFIED STARTING VALUES WITHOUT
RANDOM STARTS

TITLE: this is an example of a LCA with binary,

censored, unordered, and count latent
class indicators using user-specified
starting values without random starts
DATA: FILE IS ex7.11.dat;
VARIABLE: NAMES ARE u1 y1 u2 u3;
CLASSES = c (2);
CATEGORICAL = u1;
CENSORED = y1 (b);
NOMINAL = u2;
COUNT = u3 (i);
ANALYSIS: TYPE = MIXTURE;
STARTS = 0;
MODEL:
%OVERALL%
%c#1%
[u1$1*-1 y1*3 u2#1*0 u2#2*1 u3*.5
u3#1*1.5];
y1*2;
%c#2%
[u1$1*0 y1*1 u2#1*-1 u2#2*0 u3*1 u3#1*1];
y1*1;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 7.4 is that the latent
class indicators are a combination of binary, censored, unordered
categorical (nominal) and count variables instead of binary variables.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables

172
 Examples: Mixture Modeling With Cross-Sectional Data

in the model and its estimation. In the example above, the latent class
indicator u1 is a binary variable. The CENSORED option is used to
specify which dependent variables are treated as censored variables in
the model and its estimation, whether they are censored from above or
below, and whether a censored or censored-inflated model will be
estimated. In the example above, y1 is a censored variable. The b in
parentheses following y1 indicates that y1 is censored from below, that
is, has a floor effect, and that the model is a censored regression model.
The censoring limit is determined from the data. The NOMINAL option
is used to specify which dependent variables are treated as unordered
categorical (nominal) variables in the model and its estimation. In the
example above, u2 is a three-category unordered variable. The program
determines the number of categories. The categories of an unordered
categorical variable are referred to by adding to the name of the
unordered categorical variable the number sign (#) followed by the
number of the category. In this example, u2#1 refers to the first category
of u2 and u2#2 refers to the second category of u2. The COUNT option
is used to specify which dependent variables are treated as count
variables in the model and its estimation and whether a Poisson or zero-
inflated Poisson model will be estimated. In the example above, u3 is a
count variable. The i in parentheses following u3 indicates that a zero-
inflated model will be estimated. The inflation part of the count variable
is referred to by adding to the name of the count variable the number
sign (#) followed by the number 1. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1 and 7.4.

173
CHAPTER 7

EXAMPLE 7.12: LCA WITH BINARY LATENT CLASS

INDICATORS USING AUTOMATIC STARTING VALUES
WITH RANDOM STARTS WITH A COVARIATE AND A
DIRECT EFFECT

TITLE: this is an example of a LCA with binary

latent class indicators using automatic
starting values with random starts with a
covariate and a direct effect
DATA: FILE IS ex7.12.dat;
VARIABLE: NAMES ARE u1-u4 x;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
c ON x;
u4 ON x;
OUTPUT: TECH1 TECH8;

u1

u2

x c

u3

u4

The difference between this example and Example 7.3 is that the model
contains a covariate and a direct effect. The first ON statement

174
 Examples: Mixture Modeling With Cross-Sectional Data

describes the multinomial logistic regression of the categorical latent

variable c on the covariate x when comparing class 1 to class 2. The
intercepts of this regression are estimated as the default. The second ON
statement describes the logistic regression of the binary indicator u4 on
the covariate x. This is referred to as a direct effect from x to u4. The
regression coefficient is held equal across classes as the default. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Examples 7.1 and 7.3.

EXAMPLE 7.13: CONFIRMATORY LCA WITH BINARY

LATENT CLASS INDICATORS AND PARAMETER
CONSTRAINTS

TITLE: this is an example of a confirmatory LCA

with binary latent class indicators and
parameter constraints
DATA: FILE IS ex7.13.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
%c#1%
[u1$1*-1];
[u2$1-u3$1*-1] (1);
[u4$1*-1] (p1);
%c#2%
[u1$1@-15];
[u2$1-u3$1*1] (2);
[u4$1*1] (p2);
MODEL CONSTRAINT:
p2 = - p1;
OUTPUT: TECH1 TECH8;

In this example, constraints are placed on the measurement parameters

of the latent class indicators to reflect three hypotheses: (1) u2 and u3
are parallel measurements, (2) u1 has a probability of one in class 2, and
(3) the error rate for u4 is the same in the two classes (McCutcheon,
2002, pp. 70-72).

175
CHAPTER 7

The first hypothesis is specified by placing (1) following the threshold

parameters for u2 and u3 in class 1 and (2) following the threshold
parameters for u2 and u3 in class 2. This holds the thresholds for the
two latent class indicators equal to each other but not equal across
classes. The second hypothesis is specified by fixing the threshold of u1
in class 2 to the logit value of -15. The third hypothesis is specified
using the MODEL CONSTRAINT command. The MODEL
CONSTRAINT command is used to define linear and non-linear
constraints on the parameters in the model. Parameters are given labels
by placing a name in parentheses after the parameter in the MODEL
command. In the MODEL command, the threshold of u4 in class 1 is
given the label p1 and the threshold of u4 in class 2 is given the label p2.
In the MODEL CONSTRAINT command, the linear constraint is
defined. The threshold of u4 in class 1 is equal to the negative value of
the threshold of u4 in class 2. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 7.1.

EXAMPLE 7.14: CONFIRMATORY LCA WITH TWO

CATEGORICAL LATENT VARIABLES

TITLE: this is an example of a confirmatory LCA

with two categorical latent variables
DATA: FILE IS ex7.14.dat;
VARIABLE: NAMES ARE u1-u4 y1-y4;
CLASSES = cu (2) cy (3);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
PARAMETERIZATION = LOGLINEAR;
MODEL:
%OVERALL%
cu WITH cy;

176
 Examples: Mixture Modeling With Cross-Sectional Data

MODEL cu:
%cu#1%
[u1$1-u4$1];
%cu#2%
[u1$1-u4$1];
MODEL cy:
%cy#1%
[y1-y4];
%cy#2%
[y1-y4];
%cy#3%
[y1-y4];
OUTPUT: TECH1 TECH8;

u1 u2 u3 u4

cu

cy

y1 y2 y3 y4

In this example, the confirmatory LCA with two categorical latent

variables shown in the picture above is estimated. The two categorical
latent variables are correlated and have their own sets of latent class
indicators.

177
CHAPTER 7

The CLASSES option is used to assign names to the categorical latent

variables in the model and to specify the number of latent classes in the
model for each categorical latent variable. In the example above, there
are two categorical latent variables cu and cy. The categorical latent
variable cu has two latent classes and the categorical latent variable cy
has three latent classes. PARAMETERIZATION=LOGLINEAR is used
to specify associations among categorical latent variables. In the
LOGLINEAR parameterization, the WITH option of the MODEL
command is used to specify the relationships between the categorical
latent variables. When a model has more than one categorical latent
variable, MODEL followed by a label is used to describe the analysis
model for each categorical latent variable. Labels are defined by using
the names of the categorical latent variables. The categorical latent
variable cu has four binary indicators u1 through u4. Their thresholds
are specified to vary only across the classes of the categorical latent
variable cu. The categorical latent variable cy has four continuous
indicators y1 through y4. Their means are specified to vary only across
the classes of the categorical latent variable cy. The default estimator
for this type of analysis is maximum likelihood with robust standard
errors. The ESTIMATOR option of the ANALYSIS command can be
used to select a different estimator. An explanation of the other
commands can be found in Example 7.1.

Following is an alternative specification of the associations among cu

and cy:

cu#1 WITH cy#1 cy#2;

where cu#1 refers to the first class of cu, cy#1 refers to the first class of
cy, and cy#2 refers to the second class of cy. The classes of a
categorical latent variable are referred to by adding to the name of the
categorical latent variable the number sign (#) followed by the number
of the class. This alternative specification allows individual parameters
to be referred to in the MODEL command for the purpose of giving
starting values or placing restrictions.

178
 Examples: Mixture Modeling With Cross-Sectional Data

EXAMPLE 7.15: LOGLINEAR MODEL FOR A THREE-WAY

TABLE WITH CONDITIONAL INDEPENDENCE BETWEEN
THE FIRST TWO VARIABLES

TITLE: this is an example of a loglinear model

for a three-way table with conditional
independence between the first two
variables
DATA: FILE IS ex7.15.dat;
VARIABLE: NAMES ARE u1 u2 u3 w;
FREQWEIGHT = w;
CATEGORICAL = u1-u3;
CLASSES = c1 (2) c2 (2) c3 (2);
ANALYSIS: TYPE = MIXTURE;
STARTS = 0;
PARAMETERIZATION = LOGLINEAR;
MODEL:
%OVERALL%
c1 WITH c3;
c2 WITH c3;
MODEL c1:
%c1#1%
[u1$1@15];
%c1#2%
[u1$1@-15];
MODEL c2:
%c2#1%
[u2$1@15];
%c2#2%
[u2$1@-15];
MODEL c3:
%c3#1%
[u3$1@15];
%c3#2%
[u3$1@-15];
OUTPUT: TECH1 TECH8;

In this example, a loglinear model for a three-way frequency table with

conditional independence between the first two variables is estimated.
The loglinear model is estimated using categorical latent variables that
are perfectly measured by observed categorical variables. It is also
possible to estimate loglinear models for categorical latent variables that
are measured with error by observed categorical variables. The
conditional independence is specified by the two-way interaction

179
CHAPTER 7

between the first two variables being zero for each of the two levels of
the third variable.

PARAMETERIZATION=LOGLINEAR is used to estimate loglinear

models with two- and three-way interactions. In the LOGLINEAR
parameterization, the WITH option of the MODEL command is used to
specify the associations among the categorical latent variables. When a
model has more than one categorical latent variable, MODEL followed
by a label is used to describe the analysis model for each categorical
latent variable. Labels are defined by using the names of the categorical
latent variables. In the example above, the categorical latent variables
are perfectly measured by the latent class indicators. This is specified by
fixing their thresholds to the logit value of plus or minus 15,
corresponding to probabilities of zero and one. The default estimator for
this type of analysis is maximum likelihood with robust standard errors.
The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. An explanation of the other commands can
be found in Examples 7.1 and 7.14.

EXAMPLE 7.16: LCA WITH PARTIAL CONDITIONAL

INDEPENDENCE

TITLE: this is an example of LCA with partial

conditional independence
DATA: FILE IS ex7.16.dat;
VARIABLE: NAMES ARE u1-u4;
CATEGORICAL = u1-u4;
CLASSES = c(2);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL:
%OVERALL%
f by u2-u3@0;
f@1; [f@0];
%c#1%
[u1$1-u4$1*-1];
f by u2@1 u3;
OUTPUT: TECH1 TECH8;

180
 Examples: Mixture Modeling With Cross-Sectional Data

u1

u2

c f
u3

u4

In this example, the LCA with partial conditional independence shown

in the picture above is estimated. A similar model is described in Qu,
Tan, and Kutner (1996).

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
can be used to select a different estimator. In the example above, the
lack of conditional independence between the latent class indicators u2
and u3 in class 1 is captured by u2 and u3 being influenced by the
continuous latent variable f in class 1. The conditional independence
assumption for u2 and u3 is not violated for class 2. This is specified by
fixing the factor loadings to zero in the overall model. The amount of
deviation from conditional independence between u2 and u3 in class 1 is
captured by the u3 factor loading for the continuous latent variable f. An
explanation of the other commands can be found in Example 7.1.

181
CHAPTER 7

EXAMPLE 7.17: MIXTURE CFA MODELING

TITLE: this is an example of mixture CFA modeling

DATA: FILE IS ex7.17.dat;
VARIABLE: NAMES ARE y1-y5;
CLASSES = c(2);
ANALYSIS: TYPE = MIXTURE;
MODEL: %OVERALL%
f BY y1-y5;
%c#1%
[f*1];
OUTPUT: TECH1 TECH8;

y1 y2 y3 y4 y5

c f

In this example, the mixture CFA model shown in the picture above is
estimated (Muthén, 2008). The mean of the factor f varies across the
classes of the categorical latent variable c. The residual arrow pointing
to f indicates that the factor varies within class. This implies that the
distribution of f is allowed to be non-normal. It is possible to allow
other parameters of the CFA model to vary across classes.

The BY statement specifies that f is measured by y1, y2, y3, y4, and y5.
The factor mean varies across the classes. All other model parameters
are held equal across classes as the default. The default estimator for
this type of analysis is maximum likelihood with robust standard errors.
The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. An explanation of the other commands can
be found in Example 7.1.

182
 Examples: Mixture Modeling With Cross-Sectional Data

EXAMPLE 7.18: LCA WITH A SECOND-ORDER FACTOR

(TWIN ANALYSIS)

TITLE: this is an example of a LCA with a second-

order factor (twin analysis)
DATA: FILE IS ex7.18.dat;
VARIABLE: NAMES ARE u11-u13 u21-u23;
CLASSES = c1(2) c2(2);
CATEGORICAL = u11-u23;
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL:
%OVERALL%
f BY;
f@1;
c1 c2 ON f*1 (1);
MODEL c1:
%c1#1%
[u11$1-u13$1*-1];
%c1#2%
[u11$1-u13$1*1];
MODEL c2:
%c2#1%
[u21$1-u23$1*-1];
%c2#2%
[u21$1-u23$1*1];
OUTPUT: TECH1 TECH8;

183
CHAPTER 7

u11 u12 u13 u21 u22 u23

c1 c2

In this example, the second-order factor model shown in the picture

above is estimated. The first-order factors are categorical latent
variables and the second-order factor is a continuous latent variable.
This is a model that can be used for studies of twin associations where
the categorical latent variable c1 refers to twin 1 and the categorical
latent variable c2 refers to twin 2.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
can be used to select a different estimator. When a model has more than
one categorical latent variable, MODEL followed by a label is used to
describe the analysis model for each categorical latent variable. Labels
are defined by using the names of the categorical latent variables.

In the overall model, the BY statement names the second order factor f.
The ON statement specifies that f influences both categorical latent
variables in the same amount by imposing an equality constraint on the
two multinomial logistic regression coefficients. The slope in the
multinomial regression of c on f reflects the strength of association

184
 Examples: Mixture Modeling With Cross-Sectional Data

between the two categorical latent variables. An explanation of the

other commands can be found in Examples 7.1 and 7.14.

EXAMPLE 7.19: SEM WITH A CATEGORICAL LATENT

VARIABLE REGRESSED ON A CONTINUOUS LATENT
VARIABLE

TITLE: this is an example of a SEM with a

categorical latent variable regressed on a
continuous latent variable
DATA: FILE IS ex7.19.dat;
VARIABLE: NAMES ARE u1-u8;
CATEGORICAL = u1-u8;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL:
%OVERALL%
f BY u1-u4;
c ON f;
%c#1%
[u5$1-u8$1];
%c#2%
[u5$1-u8$1];
OUTPUT: TECH1 TECH8;

u1 u5

u2 u6
f c
u3 u7

u4 u8

185
CHAPTER 7

In this example, the model with both a continuous and categorical latent
variable shown in the picture above is estimated. The categorical latent
variable c is regressed on the continuous latent variable f in a
multinomial logistic regression.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
can be used to select a different estimator. In the overall model, the BY
statement specifies that f is measured by the categorical factor indicators
u1 through u4. The categorical latent variable c has four binary latent
class indicators u5 through u8. The ON statement specifies the
multinomial logistic regression of the categorical latent variable c on the
continuous latent variable f. An explanation of the other commands can
be found in Example 7.1.

EXAMPLE 7.20: STRUCTURAL EQUATION MIXTURE

MODELING

TITLE: this is an example of structural equation

mixture modeling
DATA: FILE IS ex7.20.dat;
VARIABLE: NAMES ARE y1-y6;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
f1 BY y1-y3;
f2 BY y4-y6;
f2 ON f1;
%c#1%
[f1*1 f2];
f2 ON f1;
OUTPUT: TECH1 TECH8;

186
 Examples: Mixture Modeling With Cross-Sectional Data

y1 y4

y2 f1 f2 y5

y3 y6

In this example, the structural equation mixture model shown in the

picture above is estimated. A continuous latent variable f2 is regressed
on a second continuous latent variable f1. The solid arrows from the
categorical latent variable c to f1 and f2 indicate that the mean of f1 and
the intercept of f2 vary across classes. The broken arrow from c to the
arrow from f1 to f2 indicates that the slope in the linear regression of f2
on f1 varies across classes. For related models, see Jedidi, Jagpal, and
DeSarbo (1997).

In the overall model, the first BY statement specifies that f1 is measured

by y1 through y3. The second BY statement specifies that f2 is
measured by y4 through y6. The ON statement describes the linear
regression of f2 on f1. In the model for class 1, the mean of f1, the
intercept of f2, and the slope in the regression of f2 on f1 are specified to
be free across classes. All other parameters are held equal across classes
as the default. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 7.1.

187
CHAPTER 7

EXAMPLE 7.21: MIXTURE MODELING WITH KNOWN

CLASSES (MULTIPLE GROUP ANALYSIS)

TITLE: this is an example of mixture modeling

with known classes (multiple group
analysis)
DATA: FILE IS ex7.21.dat;
VARIABLE: NAMES = g y1-y4;
CLASSES = cg (2) c (2);
KNOWNCLASS = cg (g = 0 g = 1);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
c ON cg;
MODEL c:
%c#1%
[y1-y4];
%c#2%
[y1-y4];
MODEL cg:
%cg#1%
y1-y4;
%cg#2%
y1-y4;
OUTPUT: TECH1 TECH8;

y1 y2 y3 y4

cg c

In this example, the multiple group mixture model shown in the picture
above is estimated. The groups are represented by the classes of the
categorical latent variable cg, which has known class (group)
membership.

188
 Examples: Mixture Modeling With Cross-Sectional Data

The KNOWNCLASS option is used for multiple group analysis with

TYPE=MIXTURE. It is used to identify the categorical latent variable
for which latent class membership is known and is equal to observed
groups in the sample. The KNOWNCLASS option identifies cg as the
categorical latent variable for which latent class membership is known.
The information in parentheses following the categorical latent variable
name defines the known classes using an observed variable. In this
example, the observed variable g is used to define the known classes.
The first class consists of individuals with the value 0 on the variable g.
The second class consists of individuals with the value 1 on the variable
g. The means of y1, y2, y3, and y4 vary across the classes of c, while
the variances of y1, y2, y3, and y4 vary across the classes of cg. An
explanation of the other commands can be found in Example 7.1.

EXAMPLE 7.22: MIXTURE MODELING WITH CONTINUOUS

VARIABLES THAT CORRELATE WITHIN CLASS
(MULTIVARIATE NORMAL MIXTURE MODEL)

TITLE: this is an example of mixture modeling

with continuous variables that correlate
within class (multivariate normal mixture
model)
DATA: FILE IS ex7.22.dat;
VARIABLE: NAMES ARE y1-y4;
CLASSES = c (3);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
y1 WITH y2-y4;
y2 WITH y3 y4;
y3 WITH y4;
%c#2%
[y1–y4*-1];
%c#3%
[y1–y4*1];
OUTPUT: TECH1 TECH8;

189
CHAPTER 7

y1

y2

y3

y4

In this example, the mixture model shown in the picture above is

estimated. Because c is a categorical latent variable, the interpretation of
the picture is not the same as for models with continuous latent
variables. The arrows from c to the observed variables y1, y2, y3, and
y4 indicate that the means of the observed variables vary across the
classes of c. The arrows correspond to the regressions of the observed
variables on a set of dummy variables representing the categories of c.
The observed variables correlate within class. This is a conventional
multivariate mixture model (Everitt & Hand, 1981; McLachlan & Peel,
2000).

In the overall model, by specifying the three WITH statements the

default of zero covariances within class is relaxed and the covariances
among y1, y2, y3, and y4 are estimated. These covariances are held
equal across classes as the default. The variances of y1, y2, y3, and y4
are estimated and held equal as the default. These defaults can be
overridden. The means of the categorical latent variable c are estimated
as the default.

When WITH statements are included in a mixture model, starting values

may be useful. In the class-specific model for class 2, starting values of
-1 are given for the means of y1, y2, y3, and y4. In the class-specific
model for class 3, starting values of 1 are given for the means of y1, y2,

190
 Examples: Mixture Modeling With Cross-Sectional Data

y3, and y4. The default estimator for this type of analysis is maximum
likelihood with robust standard errors. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 7.1.

EXAMPLE 7.23: MIXTURE RANDOMIZED TRIALS

MODELING USING CACE ESTIMATION WITH TRAINING
DATA

TITLE: this is an example of mixture randomized

trials modeling using CACE estimation with
training data
DATA: FILE IS ex7.23.dat;
VARIABLE: NAMES ARE y x1 x2 c1 c2;
CLASSES = c (2);
TRAINING = c1 c2;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
y ON x1 x2;
c ON x1;
%c#1%
[y];
y;
y ON x2@0;
%c#2%
[y*.5];
y;
OUTPUT: TECH1 TECH8;

191
CHAPTER 7

x1 x2

In this example, the mixture model for randomized trials using CACE
(Complier-Average Causal Effect) estimation with training data shown
in the picture above is estimated (Little & Yau, 1998). The continuous
dependent variable y is regressed on the covariate x1 and the treatment
dummy variable x2. The categorical latent variable c is compliance
status, with class 1 referring to non-compliers and class 2 referring to
compliers. Compliance status is observed in the treatment group and
unobserved in the control group. Because c is a categorical latent
variable, the interpretation of the picture is not the same as for models
with continuous latent variables. The arrow from c to the y variable
indicates that the intercept of y varies across the classes of c. The arrow
from c to the arrow from x2 to y indicates that the slope in the regression
of y on x2 varies across the classes of c. The arrow from x1 to c
represents the multinomial logistic regression of c on x1.

The TRAINING option is used to identify the variables that contain

information about latent class membership. Because there are two
classes, there are two training variables c1 and c2. Individuals in the
treatment group are assigned values of 1 for c1 and 0 for c2 if they are
non-compliers and 0 for c1 and 1 for c2 if they are compliers.
Individuals in the control group are assigned values of 1 for both c1 and

192
 Examples: Mixture Modeling With Cross-Sectional Data

c2 to indicate that they are allowed to be a member of either class and

that their class membership is estimated.

In the overall model, the first ON statement describes the linear

regression of y on the covariate x1 and the treatment dummy variable x2.
The intercept and residual variance of y are estimated as the default.
The second ON statement describes the multinomial logistic regression
of the categorical latent variable c on the covariate x1 when comparing
class 1 to class 2. The intercept in the regression of c on x1 is estimated
as the default.

In the model for class 1, a starting value of zero is given for the intercept
of y as the default. The residual variance of y is specified to relax the
default across class equality constraint. The ON statement describes the
linear regression of y on x2 where the slope is fixed at zero. This is
done because non-compliers do not receive treatment. In the model for
class 2, a starting value of .5 is given for the intercept of y. The residual
variance of y is specified to relax the default across class equality
constraint. The regression of y ON x2, which represents the CACE
treatment effect, is not fixed at zero for class 2. The default estimator
for this type of analysis is maximum likelihood with robust standard
errors. The ESTIMATOR option of the ANALYSIS command can be
used to select a different estimator. An explanation of the other
commands can be found in Example 7.1.

EXAMPLE 7.24: MIXTURE RANDOMIZED TRIALS

MODELING USING CACE ESTIMATION WITH MISSING
DATA ON THE LATENT CLASS INDICATOR

TITLE: this is an example of mixture randomized

trials modeling using CACE estimation with
missing data on the latent class indicator
DATA: FILE IS ex7.24.dat;
VARIABLE: NAMES ARE u y x1 x2;
CLASSES = c (2);
CATEGORICAL = u;
MISSING = u (999);
ANALYSIS: TYPE = MIXTURE;

193
CHAPTER 7

MODEL:
%OVERALL%
y ON x1 x2;
c ON x1;

%c#1%
[u$1@15];
[y];
y;
y ON x2@0;

%c#2%
[u$1@-15];
[y*.5];
y;
OUTPUT: TECH1 TECH8;

u y

x1 x2

The difference between this example and Example 7.23 is that a binary
latent class indicator u has been added to the model. This binary
variable represents observed compliance status. Treatment compliers
have a value of 1 on this variable; treatment non-compliers have a value
of 0 on this variable; and individuals in the control group have a missing
value on this variable. The latent class indicator u is used instead of
training data.

194
 Examples: Mixture Modeling With Cross-Sectional Data

In the model for class 1, the threshold of the latent class indicator
variable u is set to a logit value of 15. In the model for class 2, the
threshold of the latent class indicator variable u is set to a logit value of
–15. These logit values reflect that c is perfectly measured by u.
Individuals in the non-complier class (class 1) have probability zero of
observed compliance and individuals in the complier class (class 2) have
probability one of observed compliance. The default estimator for this
type of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 7.1 and 7.23.

EXAMPLE 7.25: ZERO-INFLATED POISSON REGRESSION

CARRIED OUT AS A TWO-CLASS MODEL

TITLE: this is an example of a zero-inflated

Poisson regression carried out as a two-
class model
DATA: FILE IS ex3.8.dat;
VARIABLE: NAMES ARE u1 x1 x3;
COUNT IS u1;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
u1 ON x1 x3;
c ON x1 x3;
%c#1%
[u1@-15];
u1 ON x1@0 x3@0;
OUTPUT: TECH1 TECH8;

x1 u1

x3 c

195
CHAPTER 7

In this example, the zero-inflated Poisson regression model shown in the

picture above is estimated. This is an alternative to the way zero-
inflated Poisson regression was carried out in Example 3.8. In the
example above, a categorical latent variable c with two classes is used to
represent individuals who are able to assume values of zero and above
and individuals who are unable to assume any value except zero. The
categorical latent variable c corresponds to the binary latent inflation
variable u1#1 in Example 3.8. This approach has the advantage of
allowing the estimation of the probability of being in each class and the
posterior probabilities of being in each class for each individual.

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u1 is a specified as count variable without inflation
because the inflation is captured by the categorical latent variable c.

In the overall model, the first ON statement describes the Poisson

regression of the count variable u1 on the covariates x1 and x3. The
second ON statement describes the multinomial logistic regression of the
categorical latent variable c on the covariates x1 and x3 when comparing
class 1 to class 2. In this example, class 1 contains individuals who are
unable to assume any value except zero on u1. Class 2 contains
individuals whose values on u1 are distributed as a Poisson variable
without inflation. Mixing the two classes results in u1 having a zero-
inflated Poisson distribution. In the class-specific model for class 1, the
intercept of u1 is fixed at -15 to represent a low log rate at which the
probability of a count greater than zero is zero. Therefore, all
individuals in class 1 have a value of 0 on u1. Because u1 has no
variability, the slopes in the Poisson regression of u1 on the covariates
x1 and x3 in class 1 are fixed at zero. The default estimator for this type
of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 7.1.

196
 Examples: Mixture Modeling With Cross-Sectional Data

EXAMPLE 7.26: CFA WITH A NON-PARAMETRIC

REPRESENTATION OF A NON-NORMAL FACTOR
DISTRIBUTION

TITLE: this is an example of CFA with a non-

parametric representation of a non-normal
factor distribution
DATA: FILE IS ex7.26.dat;
VARIABLE: NAMES ARE y1-y5 c;
USEV = y1-y5;
CLASSES = c (3);
ANALYSIS: TYPE = MIXTURE;
MODEL: %OVERALL%
f BY y1-y5;
f@0;
OUTPUT: TECH1 TECH8;

In this example, a CFA model with a non-parametric representation of a

non-normal factor distribution is estimated. One difference between this
example and Example 7.17 is that the factor variance is fixed at zero in
each class. This is done to capture a non-parametric representation of
the factor distribution (Aitkin, 1999) where the latent classes are used to
represent non-normality not unobserved heterogeneity with substantively
meaningful latent classes. This is also referred to as semiparametric
modeling. The factor distribution is represented by a histogram with as
many bars as there are classes. The bars represent scale steps on the
continuous latent variable. The spacing of the scale steps is obtained by
the factor means in the different classes with a factor mean for one class
fixed at zero for identification, and the percentage of individuals at the
different scale steps is obtained by the latent class percentages. This
means that continuous factor scores are obtained for the individuals
while not assuming normality for the factor but estimating its
distribution. Factor variances can also be estimated to obtain a more
general mixture although this reverts to the parametric assumption of
normality, in this case, within each class. When the latent classes are
used to represent non-normality, the mixed parameter values are of
greater interest than the parameters for each mixture component
(Muthén, 2002, p. 102; Muthén, 2004). An explanation of the other
commands can be found in Example 7.1.

197
CHAPTER 7

EXAMPLE 7.27: FACTOR (IRT) MIXTURE ANALYSIS WITH

BINARY LATENT CLASS AND FACTOR INDICATORS

TITLE: this is an example of a factor (IRT)

mixture analysis with binary latent class
and factor indicators
DATA: FILE = ex7.27.dat;
VARIABLE: NAMES = u1-u8;
CATEGORICAL = u1-u8;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
STARTS = 100 20;
MODEL: %OVERALL%
f BY u1-u8;
[f@0];
%c#1%
f BY u1@1 u2-u8;
f;
[u1$1-u8$1];
%c#2%
f BY u1@1 u2-u8;
f;
[u1$1-u8$1];
OUTPUT: TECH1 TECH8;

u1 u2 u3 u4 u5 u6 u7 u8

198
 Examples: Mixture Modeling With Cross-Sectional Data

In this example, the factor (IRT) mixture model shown in the picture
above is estimated (Muthén, 2008). The model is a generalization of the
latent class model where the latent class model assumption of
conditional independence between the latent class indicators within class
is relaxed using a factor that influences the items within each class
(Muthén, 2006; Muthén & Asparouhov, 2006; Muthén, Asparouhov, &
Rebollo, 2006). The factor represents individual variation in response
probabilities within class. Alternatively, this model may be seen as an
Item Response Theory (IRT) mixture model. The broken arrows from
the categorical latent variable c to the arrows from the factor f to the
latent class indicators u1 to u8 indicate that the factor loadings vary
across classes.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
can be used to select a different estimator. The STARTS option is used
to specify the number of initial stage random sets of starting values to
generate and the number of final stage optimizations to use. The default
is 20 random sets of starting values for the initial stage and 4
optimizations for the final stage. In the example above, the STARTS
option specifies that 100 random sets of starting values for the initial
stage and 20 final stage optimizations will be used.

In the overall model, the BY statement specifies that the factor f is

measured by u1, u2, u3, u4, u5, u6, u7, and u8. The mean of the factor is
fixed at zero which implies that the mean is zero in both classes. The
factor variance is held equal across classes as the default. The
statements in the class-specific parts of the model relax the equality
constraints across classes for the factor loadings, factor variance, and the
thresholds of the indicators. An explanation of the other commands can
be found in Examples 7.1 and 7.3.

199
CHAPTER 7

EXAMPLE 7.28: TWO-GROUP TWIN MODEL FOR

CATEGORICAL OUTCOMES USING MAXIMUM
LIKELIHOOD AND PARAMETER CONSTRAINTS

TITLE: this is an example of a two-group twin

model for categorical outcomes using
maximum likelihood and parameter
constraints
DATA: FILE = ex7.28.dat;
VARIABLE: NAMES = u1 u2 dz;
CATEGORICAL = u1 u2;
CLASSES = cdz (2);
KNOWNCLASS = cdz (dz = 0 dz = 1);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
LINK = PROBIT;
MODEL: %OVERALL%
[u1$1-u2$1] (1);
f1 BY u1;
f2 BY u2;
[f1-f2@0];
f1-f2 (varf);
%cdz#1%
f1 WITH f2(covmz);
%cdz#2%
f1 WITH f2(covdz);
MODEL CONSTRAINT:
NEW(a c h);
varf = a**2 + c**2 + .001;
covmz = a**2 + c**2;
covdz = 0.5*a**2 + c**2;
h = a**2/(a**2 + c**2 + 1);

200
 Examples: Mixture Modeling With Cross-Sectional Data

u1 u2

f1 f2

In this example, the model shown in the picture above is estimated. The
variables u1 and u2 represent a univariate outcome for each member of a
twin pair. Monozygotic and dizygotic twins are considered in a two-
group twin model for categorical outcomes using maximum likelihood
estimation. Parameter constraints are used to represent the ACE model
restrictions. The ACE variance and covariance restrictions are placed on
normally-distributed latent response variables, which are also called
liabilities, underlying the categorical outcomes. This model is referred
to as the threshold model for liabilities (Neale & Cardon, 1992). The
monozygotic and dizygotic twin groups are represented by latent classes
with known class membership.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, the variables u1
and u2 are binary or ordered categorical variables. The program
determines the number of categories for each indicator. The
KNOWNCLASS option identifies cdz as the categorical latent variable
for which latent class membership is known. The information in
parentheses following the categorical latent variable name defines the
known classes using an observed variable. In this example, the observed
variable dz is used to define the known classes. The first class consists
of the monozygotic twins who have the value 0 on the variable dz. The
second class consists of the dizygotic twins who have the value 1 on the
variable dz.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration

201
CHAPTER 7

algorithm will be used. Note that numerical integration becomes

increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of
integration are used with 225 integration points. The ESTIMATOR
option can be used to select a different estimator. The LINK option is
used with maximum likelihood estimation to select a logit or a probit
link for models with categorical outcomes. The default is a logit link. In
this example, the probit link is used because the threshold model for
liabilities uses normally-distributed latent response variables.

In the overall model, the (1) following the first bracket statement
specifies that the thresholds of u1 and u2 are held equal across twins.
The two BY statements define a factor behind each outcome. This is
done because covariances of categorical outcomes are not part of the
model when maximum likelihood estimation is used. The covariances of
the factors become the covariances of the categorical outcomes or more
precisely the covariances of the latent response variables underlying the
categorical outcomes. The means of the factors are fixed at zero and
their variances are held equal across twins. The variance of each
underlying response variable is obtained as the sum of the factor
variance plus one where one is the residual variance in the probit
regression of the categorical outcome on the factor.

In the MODEL command, labels are defined for three parameters. The
label varf is assigned to the variances of f1 and f2. Because they are
given the same label, these parameters are held equal. The label covmz
is assigned to the covariance between f1 and f2 for the monozygotic
twins and the label covdz is assigned to the covariance between f1 and f2
for the dizygotic twins. In the MODEL CONSTRAINT command, the
NEW option is used to assign labels to three parameters that are not in
the analysis model: a, c, and h. The two parameters a and c are used to
decompose the covariances of u1 and u2 into genetic and environmental
components. The value .001 is added to the variance of the factors to
avoid a singular factor covariance matrix which comes about because the
factor variances and covariances are the same. The parameter h does not
impose restrictions on the model parameters but is used to compute the
heritability estimate and its standard error. This heritability estimate
uses the residual variances for the latent response variables which are
fixed at one. An explanation of the other commands can be found in
Example 7.1.

202
 Examples: Mixture Modeling With Cross-Sectional Data

EXAMPLE 7.29: TWO-GROUP IRT TWIN MODEL FOR

FACTORS WITH CATEGORICAL FACTOR INDICATORS
USING PARAMETER CONSTRAINTS

TITLE: this is an example of a two-group IRT twin

model for factors with categorical factor
indicators using parameter constraints
DATA: FILE = ex7.29.dat;
VARIABLE: NAMES = u11-u14 u21-u24 dz;
CATEGORICAL = u11-u24;
CLASSES = cdz (2);
KNOWNCLASS = cdz (dz = 0 dz = 1);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL: %OVERALL%
f1 BY u11
u12-u14 (lam2-lam4);
f2 BY u21
u22-u24 (lam2-lam4);
[f1-f2@0];
f1-f2 (var);
[u11$1-u14$1] (t1-t4);
[u21$1-u24$1] (t1-t4);
%cdz#1%
f1 WITH f2(covmz);
%cdz#2%
f1 WITH f2(covdz);
MODEL CONSTRAINT:
NEW(a c e h);
var = a**2 + c**2 + e**2;
covmz = a**2 + c**2;
covdz = 0.5*a**2 + c**2;
h = a**2/(a**2 + c**2 + e**2);

203
CHAPTER 7

u11 u12 u13 u14 u21 u22 u23 u24

f1 f2

In this example, the model shown in the picture above is estimated. The
factors f1 and f2 represent a univariate variable for each member of the
twin pair. Monozygotic and dizygotic twins are considered in a two-
group twin model for factors with categorical factor indicators using
parameter constraints and maximum likelihood estimation. Parameter
constraints are used to represent the ACE model restrictions. The ACE
variance and covariance restrictions are placed on two factors instead of
two observed variables as in Example 7.28. The relationships between
the categorical factor indicators and the factors are logistic regressions.
Therefore, the factor model for each twin is a two-parameter logistic
Item Response Theory model (Muthén, Asparouhov, & Rebollo, 2006).
The monozygotic and dizygotic twin groups are represented by latent
classes with known class membership.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of
integration are used with 225 integration points. The ESTIMATOR
option can be used to select a different estimator.

In the overall model, the two BY statements specify that f1 is measured

by u11, u12, u13, and u14 and that f2 is measured by u21, u22, u23, and
u24. The means of the factors are fixed at zero. In the class-specific
models, the threshold of the dz variable is fixed at 15 in class one and -
15 in class 2.

204
 Examples: Mixture Modeling With Cross-Sectional Data

In the MODEL command, labels are defined for nine parameters. The
list function can be used when assigning labels. The label lam2 is
assigned to the factor loadings for u12 and u22; the label lam3 is
assigned to the factor loadings for u13 and u23; and the label lam4 is
assigned to the factor loadings for u14 and u24. Factor loadings with the
same label are held equal. The label t1 is assigned to the thresholds of
u11 and u21; the label t2 is assigned to the thresholds of u12 and u22;
the label t3 is assigned to the thresholds of u13 and u23; and the label t4
is assigned to the thresholds of u14 and u24. Parameters with the same
label are held equal. The label covmz is assigned to the covariance
between f1 and f2 for the monozygotic twins and the label covdz is
assigned to the covariance between f1 and f2 for the dizygotic twins.

In the MODEL CONSTRAINT command, the NEW option is used to

assign labels to four parameters that are not in the analysis model: a, c,
e, and h. The three parameters a, c, and e are used to decompose the
variances and covariances of f1 and f2 into genetic and environmental
components. The parameter h does not impose restrictions on the model
parameters but is used to compute the heritability estimate and its
standard error. An explanation of the other commands can be found in
Examples 7.1 and 7.28.

EXAMPLE 7.30: CONTINUOUS-TIME SURVIVAL ANALYSIS

USING A COX REGRESSION MODEL TO ESTIMATE A
TREATMENT EFFECT

TITLE: this is an example of continuous-time

survival analysis using a Cox regression
model to estimate a treatment effect
DATA: FILE = ex7.30.dat;
VARIABLE: NAMES are t u x tcent class;
USEVARIABLES = t-tcent;
SURVIVAL = t;
TIMECENSORED = tcent;
CATEGORICAL = u;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;

205
CHAPTER 7

MODEL:
%OVERALL%
t ON x;
%c#1%
[u$1@15];
[t@0];
%c#2%
[u$1@-15];
[t];
OUTPUT: TECH1 LOGRANK;
PLOT: TYPE = PLOT2;

u t

In this example, the continuous-time survival analysis model shown in

the picture above is estimated. The model is similar to Larsen (2004). A
treatment and a control group are analyzed as two known latent classes.
The baseline hazards are held equal across the classes and the treatment
effect is expressed as the intercept of the survival variable in the
treatment group. For applications of this model, see Muthén et al.
(2009).

The CATEGORICAL option is used to specify that the variable u is a

binary variable. This variable is a treatment dummy variable where zero
represents the control group and one represents the treatment group. In
this example, the categorical latent variable c has two classes. In the
MODEL command, in the model for class 1, the threshold for u is fixed
at 15 so that the probability that u equals one is zero. By this

206
 Examples: Mixture Modeling With Cross-Sectional Data

specification, class 1 is the control group. In the model for class 2, the
threshold for u is fixed at -15 so that the probability that u equals one is
one. By this specification, class 2 is the treatment group. In the overall
model, the ON statement describes the Cox regression for the survival
variable t on the covariate x. In class 1, the intercept in the Cox
regression is fixed at zero. In class 2, it is free. This intercept represents
the treatment effect. The LOGRANK option of the OUTPUT command
provides a logrank test of the equality of the treatment and control
survival curves (Mantel, 1966). By specifying PLOT2 in the PLOT
command, the following plots are obtained:

• Kaplan-Meier curve
• Sample log cumulative hazard curve
• Estimated baseline hazard curve
• Estimated baseline survival curve
• Estimated log cumulative baseline curve
• Kaplan-Meier curve with estimated baseline survival curve
• Sample log cumulative hazard curve with estimated log
cumulative baseline curve

An explanation of the other commands can be found in Example 7.1.

207
CHAPTER 7

208
 Examples: Mixture Modeling With Longitudinal Data

CHAPTER 8
EXAMPLES: MIXTURE
MODELING WITH
LONGITUDINAL DATA

Mixture modeling refers to modeling with categorical latent variables

that represent subpopulations where population membership is not
known but is inferred from the data. This is referred to as finite mixture
modeling in statistics (McLachlan & Peel, 2000). For an overview of
different mixture models, see Muthén (2008). In mixture modeling with
longitudinal data, unobserved heterogeneity in the development of an
outcome over time is captured by categorical and continuous latent
variables. The simplest longitudinal mixture model is latent class
growth analysis (LCGA). In LCGA, the mixture corresponds to
different latent trajectory classes. No variation across individuals is
allowed within classes (Nagin, 1999; Roeder, Lynch, & Nagin, 1999;
Kreuter & Muthén, 2008). Another longitudinal mixture model is the
growth mixture model (GMM; Muthén & Shedden, 1999; Muthén et al.,
2002; Muthén, 2004; Muthén & Asparouhov, 2009). In GMM, within-
class variation of individuals is allowed for the latent trajectory classes.
The within-class variation is represented by random effects, that is,
continuous latent variables, as in regular growth modeling. All of the
growth models discussed in Chapter 6 can be generalized to mixture
modeling. Yet another mixture model for analyzing longitudinal data is
latent transition analysis (LTA; Collins & Wugalter, 1992; Reboussin et
al., 1998), also referred to as hidden Markov modeling, where latent
class indicators are measured over time and individuals are allowed to
transition between latent classes. With discrete-time survival mixture
analysis (DTSMA; Muthén & Masyn, 2005), the repeated observed
outcomes represent event histories. Continuous-time survival mixture
modeling is also available (Asparouhov et al., 2006). For mixture
modeling with longitudinal data, observed outcome variables can be
continuous, censored, binary, ordered categorical (ordinal), counts, or
combinations of these variable types.

209
CHAPTER 8

All longitudinal mixture models can be estimated using the following

special features:

• Single or multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities
• Test of equality of means across latent classes using posterior
probability-based multiple imputations

For TYPE=MIXTURE, multiple group analysis is specified by using the

KNOWNCLASS option of the VARIABLE command. The default is to
estimate the model under missing data theory using all available data.
The LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. The
SUBPOPULATION option is used to select observations for an analysis
when a subpopulation (domain) is analyzed. Latent variable interactions
are specified by using the | symbol of the MODEL command in
conjunction with the XWITH option of the MODEL command. Random
slopes are specified by using the | symbol of the MODEL command in
conjunction with the ON option of the MODEL command. Individually-
varying times of observations are specified by using the | symbol of the
MODEL command in conjunction with the AT option of the MODEL
command and the TSCORES option of the VARIABLE command.
Linear and non-linear parameter constraints are specified by using the
MODEL CONSTRAINT command. Indirect effects are specified by
using the MODEL INDIRECT command. Maximum likelihood

210
 Examples: Mixture Modeling With Longitudinal Data

estimation is specified by using the ESTIMATOR option of the

ANALYSIS command. Bootstrap standard errors are obtained by using
the BOOTSTRAP option of the ANALYSIS command. Bootstrap
confidence intervals are obtained by using the BOOTSTRAP option of
the ANALYSIS command in conjunction with the CINTERVAL option
of the OUTPUT command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test. The
AUXILIARY option is used to test the equality of means across latent
classes using posterior probability-based multiple imputations.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, plots of
sample and estimated means and proportions/probabilities, and plots of
estimated probabilities for a categorical latent variable as a function of
its covariates. These are available for the total sample, by group, by
class, and adjusted for covariates. The PLOT command includes
a display showing a set of descriptive statistics for each variable. The
graphical displays can be edited and exported as a DIB, EMF, or JPEG
file. In addition, the data for each graphical display can be saved in an
external file for use by another graphics program.

Following is the set of GMM examples included in this chapter:

• 8.1: GMM for a continuous outcome using automatic starting values

and random starts
• 8.2: GMM for a continuous outcome using user-specified starting
values and random starts
• 8.3: GMM for a censored outcome using a censored model with
automatic starting values and random starts*
• 8.4: GMM for a categorical outcome using automatic starting values
and random starts*
• 8.5: GMM for a count outcome using a zero-inflated Poisson model
and a negative binomial model with automatic starting values and
random starts*
• 8.6: GMM with a categorical distal outcome using automatic
starting values and random starts
• 8.7: A sequential process GMM for continuous outcomes with two
categorical latent variables

211
CHAPTER 8

• 8.8: GMM with known classes (multiple group analysis)

Following is the set of LCGA examples included in this chapter:

• 8.9: LCGA for a binary outcome

• 8.10: LCGA for a three-category outcome
• 8.11: LCGA for a count outcome using a zero-inflated Poisson
model

Following is the set of hidden Markov and LTA examples included in

this chapter:

• 8.12: Hidden Markov model with four time points

• 8.13: LTA for two time points with a binary covariate influencing
the latent transition probabilities
• 8.14: LTA for two time points with a continuous covariate
influencing the latent transition probabilities
• 8.15: Mover-stayer LTA for three time points using a probability
parameterization

Following are the discrete-time and continuous-time survival mixture

analysis examples included in this chapter:

• 8.16: Discrete-time survival mixture analysis with survival

predicted by growth trajectory classes
• 8.17: Continuous-time survival mixture analysis using a Cox
regression model

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

212
 Examples: Mixture Modeling With Longitudinal Data

EXAMPLE 8.1: GMM FOR A CONTINUOUS OUTCOME

USING AUTOMATIC STARTING VALUES AND RANDOM
STARTS

TITLE: this is an example of a GMM for a

continuous outcome using automatic
starting values and random starts
DATA: FILE IS ex8.1.dat;
VARIABLE: NAMES ARE y1–y4 x;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
STARTS = 40 8;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

y1 y2 y3 y4

i s

In the example above, the growth mixture model (GMM) for a

continuous outcome shown in the picture above is estimated. Because c
is a categorical latent variable, the interpretation of the picture is not the
same as for models with continuous latent variables. The arrows from c

213
CHAPTER 8

to the growth factors i and s indicate that the intercepts in the regressions
of the growth factors on x vary across the classes of c. This corresponds
to the regressions of i and s on a set of dummy variables representing the
categories of c. The arrow from x to c represents the multinomial
logistic regression of c on x. GMM is discussed in Muthén and Shedden
(1999), Muthén (2004), and Muthén and Asparouhov (2009).

TITLE: this is an example of a growth mixture

model for a continuous outcome

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE IS ex8.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex8.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES ARE y1–y4 x;

CLASSES = c (2);

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains five variables: y1, y2, y3, y4, and x. Note that the
hyphen can be used as a convenience feature in order to generate a list of
names. The CLASSES option is used to assign names to the categorical
latent variables in the model and to specify the number of latent classes
in the model for each categorical latent variable. In the example above,
there is one categorical latent variable c that has two latent classes.

ANALYSIS: TYPE = MIXTURE;

STARTS = 40 8;

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis that
is to be performed. By selecting MIXTURE, a mixture model will be
estimated.

214
 Examples: Mixture Modeling With Longitudinal Data

When TYPE=MIXTURE is specified, either user-specified or automatic

starting values are used to create randomly perturbed sets of starting
values for all parameters in the model except variances and covariances.
In this example, the random perturbations are based on automatic
starting values. Maximum likelihood optimization is done in two stages.
In the initial stage, 20 random sets of starting values are generated. An
optimization is carried out for 10 iterations using each of the 20 random
sets of starting values. The ending values from the 4 optimizations with
the highest loglikelihoods are used as the starting values in the final
stage optimizations which is carried out using the default optimization
settings for TYPE=MIXTURE. A more thorough investigation of
multiple solutions can be carried out using the STARTS and
STITERATIONS options of the ANALYSIS command. In this example,
40 initial stage random sets of starting values are used and 8 final stage
optimizations are carried out.

MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON x;

The MODEL command is used to describe the model to be estimated.

For mixture models, there is an overall model designated by the label
%OVERALL%. The overall model describes the part of the model that
is in common for all latent classes. The | symbol is used to name and
define the intercept and slope growth factors in a growth model. The
names i and s on the left-hand side of the | symbol are the names of the
intercept and slope growth factors, respectively. The statement on the
right-hand side of the | symbol specifies the outcome and the time scores
for the growth model. The time scores for the slope growth factor are
fixed at 0, 1, 2, and 3 to define a linear growth model with equidistant
time points. The zero time score for the slope growth factor at time
point one defines the intercept growth factor as an initial status factor.
The coefficients of the intercept growth factor are fixed at one as part of
the growth model parameterization. The residual variances of the
outcome variables are estimated and allowed to be different across time
and the residuals are not correlated as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The intercepts and residual variances of the growth factors are

215
CHAPTER 8

estimated as the default, and the growth factor residual covariance is

estimated as the default because the growth factors do not influence any
variable in the model except their own indicators. The intercepts of the
growth factors are not held equal across classes as the default. The
residual variances and residual covariance of the growth factors are held
equal across classes as the default.

The first ON statement describes the linear regressions of the intercept

and slope growth factors on the covariate x. The second ON statement
describes the multinomial logistic regression of the categorical latent
variable c on the covariate x when comparing class 1 to class 2. The
intercept of this regression is estimated as the default. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors. The ESTIMATOR option of the ANALYSIS command
can be used to select a different estimator.

Following is an alternative specification of the multinomial logistic

regression of c on the covariate x:

c#1 ON x;

where c#1 refers to the first class of c. The classes of a categorical latent
variable are referred to by adding to the name of the categorical latent
variable the number sign (#) followed by the number of the class. This
alternative specification allows individual parameters to be referred to in
the MODEL command for the purpose of giving starting values or
placing restrictions.

OUTPUT: TECH1 TECH8;

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes.

216
 Examples: Mixture Modeling With Longitudinal Data

EXAMPLE 8.2: GMM FOR A CONTINUOUS OUTCOME

USING USER-SPECIFIED STARTING VALUES AND RANDOM
STARTS

TITLE: this is an example of a GMM for a

continuous outcome using user-specified
starting values and random starts
DATA: FILE IS ex8.2.dat;
VARIABLE: NAMES ARE y1–y4 x;
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON x;
%c#1%
[i*1 s*.5];
%c#2%
[i*3 s*1];
OUTPUT: TECH1 TECH8;

The difference between this example and Example 8.1 is that user-
specified starting values are used instead of automatic starting values. In
the MODEL command, user-specified starting values are given for the
intercepts of the intercept and slope growth factors. Intercepts are
referred to using brackets statements. The asterisk (*) is used to assign a
starting value for a parameter. It is placed after the parameter with the
starting value following it. In class 1, a starting value of 1 is given for
the intercept growth factor and a starting value of .5 is given for the
slope growth factor. In class 2, a starting value of 3 is given for the
intercept growth factor and a starting value of 1 is given for the slope
growth factor. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 8.1.

217
CHAPTER 8

EXAMPLE 8.3: GMM FOR A CENSORED OUTCOME USING A

CENSORED MODEL WITH AUTOMATIC STARTING
VALUES AND RANDOM STARTS

TITLE: this is an example of a GMM for a censored

outcome using a censored model with
automatic starting values and random
starts
DATA: FILE IS ex8.3.dat;
VARIABLE: NAMES ARE y1-y4 x;
CLASSES = c (2);
CENSORED = y1-y4 (b);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 8.1 is that the
outcome variable is a censored variable instead of a continuous variable.
The CENSORED option is used to specify which dependent variables
are treated as censored variables in the model and its estimation, whether
they are censored from above or below, and whether a censored or
censored-inflated model will be estimated. In the example above, y1, y2,
y3, and y4 are censored variables. They represent the outcome variable
measured at four equidistant occasions. The b in parentheses following
y1-y4 indicates that y1, y2, y3, and y4 are censored from below, that is,
have floor effects, and that the model is a censored regression model.
The censoring limit is determined from the data.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of
integration are used with a total of 225 integration points. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

218
 Examples: Mixture Modeling With Longitudinal Data

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The intercepts and residual variances of the growth factors are
estimated as the default, and the growth factor residual covariance is
estimated as the default because the growth factors do not influence any
variable in the model except their own indicators. The intercepts of the
growth factors are not held equal across classes as the default. The
residual variances and residual covariance of the growth factors are held
equal across classes as the default. An explanation of the other
commands can be found in Example 8.1.

EXAMPLE 8.4: GMM FOR A CATEGORICAL OUTCOME

USING AUTOMATIC STARTING VALUES AND RANDOM
STARTS

TITLE: this is an example of a GMM for a

categorical outcome using automatic
starting values and random starts
DATA: FILE IS ex8.4.dat;
VARIABLE: NAMES ARE u1–u4 x;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;
MODEL:
%OVERALL%
i s | u1@0 u2@1 u3@2 u4@3;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 8.1 is that the
outcome variable is a binary or ordered categorical (ordinal) variable
instead of a continuous variable. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the
example above, u1, u2, u3, and u4 are binary or ordered categorical
variables. They represent the outcome variable measured at four
equidistant occasions.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration

219
CHAPTER 8

algorithm will be used. Note that numerical integration becomes

increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of
integration are used with a total of 225 integration points. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

In the parameterization of the growth model shown here, the thresholds

of the outcome variable at the four time points are held equal as the
default. The intercept of the intercept growth factor is fixed at zero in
the last class and is free to be estimated in the other classes. The
intercept of the slope growth factor and the residual variances of the
intercept and slope growth factors are estimated as the default, and the
growth factor residual covariance is estimated as the default because the
growth factors do not influence any variable in the model except their
own indicators. The intercepts of the growth factors are not held equal
across classes as the default. The residual variances and residual
covariance of the growth factors are held equal across classes as the
default. An explanation of the other commands can be found in
Example 8.1.

EXAMPLE 8.5: GMM FOR A COUNT OUTCOME USING A

ZERO-INFLATED POISSON MODEL AND A NEGATIVE
BINOMIAL MODEL WITH AUTOMATIC STARTING VALUES
AND RANDOM STARTS

TITLE: this is an example of a GMM for a count

outcome using a zero-inflated Poisson
model with automatic starting values and
random starts
DATA: FILE IS ex8.5a.dat;
VARIABLE: NAMES ARE u1–u8 x;
CLASSES = c (2);
COUNT ARE u1-u8 (i);
ANALYSIS: TYPE = MIXTURE;
STARTS = 40 8;
STITERATIONS = 20;
ALGORITHM = INTEGRATION;

220
 Examples: Mixture Modeling With Longitudinal Data

MODEL:
%OVERALL%
i s q | u1@0 u2@.1 u3@.2 u4@.3 u5@.4 u6@.5
u7@.6 u8@.7;
ii si qi | u1#1@0 u2#1@.1 u3#1@.2 u4#1@.3
u5#1@.4 u6#1@.5 u7#1@.6 u8#1@.7;
s-qi@0;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 8.1 is that the
outcome variable is a count variable instead of a continuous variable. In
addition, the outcome is measured at eight occasions instead of four and
a quadratic rather than a linear growth model is estimated. The COUNT
option is used to specify which dependent variables are treated as count
variables in the model and its estimation and the type of model that will
be estimated. In the first part of this example a zero-inflated Poisson
model is estimated. In the example above, u1, u2, u3, u4, u5, u6, u7, and
u8 are count variables. They represent the outcome variable measured at
eight equidistant occasions. The i in parentheses following u1-u8
indicates that a zero-inflated Poisson model will be estimated.

A more thorough investigation of multiple solutions can be carried out

using the STARTS and STITERATIONS options of the ANALYSIS
command. In this example, 40 initial stage random sets of starting
values are used and 8 final stage optimizations are carried out. In the
initial stage analyses, 20 iterations are used instead of the default of 10
iterations. By specifying ALGORITHM=INTEGRATION, a maximum
likelihood estimator with robust standard errors using a numerical
integration algorithm will be used. Note that numerical integration
becomes increasingly more computationally demanding as the number of
factors and the sample size increase. In this example, one dimension of
integration is used with 15 integration points. The ESTIMATOR option
of the ANALYSIS command can be used to select a different estimator.

With a zero-inflated Poisson model, two growth models are estimated.

The first | statement describes the growth model for the count part of the
outcome for individuals who are able to assume values of zero and
above. The second | statement describes the growth model for the
inflation part of the outcome, the probability of being unable to assume
any value except zero. The binary latent inflation variable is referred to

221
CHAPTER 8

by adding to the name of the count variable the number sign (#) followed
by the number 1.

In the parameterization of the growth model for the count part of the
outcome, the intercepts of the outcome variable at the eight time points
are fixed at zero as the default. The intercepts and residual variances of
the growth factors are estimated as the default, and the growth factor
residual covariances are estimated as the default because the growth
factors do not influence any variable in the model except their own
indicators. The intercepts of the growth factors are not held equal across
classes as the default. The residual variances and residual covariances
of the growth factors are held equal across classes as the default. In this
example, the variances of the slope growth factors s and q are fixed at
zero. This implies that the covariances between i, s, and q are fixed at
zero. Only the variance of the intercept growth factor i is estimated.

In the parameterization of the growth model for the inflation part of the
outcome, the intercepts of the outcome variable at the eight time points
are held equal as the default. The intercept of the intercept growth factor
is fixed at zero in all classes as the default. The intercept of the slope
growth factor and the residual variances of the intercept and slope
growth factors are estimated as the default, and the growth factor
residual covariances are estimated as the default because the growth
factors do not influence any variable in the model except their own
indicators. The intercept of the slope growth factor, the residual
variances of the growth factors, and residual covariance of the growth
factors are held equal across classes as the default. These defaults can
be overridden, but freeing too many parameters in the inflation part of
the model can lead to convergence problems. In this example, the
variances of the intercept and slope growth factors are fixed at zero.
This implies that the covariances between ii, si, and qi are fixed at zero.
An explanation of the other commands can be found in Example 8.1.

TITLE: this is an example of a GMM for a count

outcome using a negative binomial model
with automatic starting values and random
starts
DATA: FILE IS ex8.5b.dat;
VARIABLE: NAMES ARE u1-u8 x;
CLASSES = c(2);
COUNT = u1-u8(nb);
ANALYSIS: TYPE = MIXTURE;
ALGORITHM = INTEGRATION;

222
 Examples: Mixture Modeling With Longitudinal Data

MODEL:
%OVERALL%
i s q | u1@0 u2@.1 u3@.2 u4@.3 u5@.4 u6@.5
u7@.6 u8@.7;
s-q@0;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

The difference between this part of the example and the first part is that
a growth mixture model (GMM) for a count outcome using a negative
binomial model is estimated instead of a zero-inflated Poisson model.
The negative binomial model estimates a dispersion parameter for each
of the outcomes (Long, 1997; Hilbe, 2011).

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and which type
of model is estimated. The nb in parentheses following u1-u8 indicates
that a negative binomial model will be estimated. The dispersion
parameters for each of the outcomes are held equal across classes as the
default. The dispersion parameters can be referred to using the names of
the count variables. An explanation of the other commands can be
found in the first part of this example and in Example 8.1.

EXAMPLE 8.6: GMM WITH A CATEGORICAL DISTAL

OUTCOME USING AUTOMATIC STARTING VALUES AND
RANDOM STARTS

TITLE: this is an example of a GMM with a

categorical distal outcome using automatic
starting values and random starts
DATA: FILE IS ex8.6.dat;
VARIABLE: NAMES ARE y1–y4 u x;
CLASSES = c(2);
CATEGORICAL = u;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON x;
OUTPUT: TECH1 TECH8;

223
CHAPTER 8

y1 y2 y3 y4

i s

x c u

The difference between this example and Example 8.1 is that a binary or
ordered categorical (ordinal) distal outcome has been added to the model
as shown in the picture above. The distal outcome u is regressed on the
categorical latent variable c using logistic regression. This is
represented as the thresholds of u varying across classes.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, u is a binary or
ordered categorical variable. The program determines the number of
categories for each indicator. The default is that the thresholds of u are
estimated and vary across the latent classes. Because automatic starting
values are used, it is not necessary to include these class-specific
statements in the model command. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 8.1.

224
 Examples: Mixture Modeling With Longitudinal Data

EXAMPLE 8.7: A SEQUENTIAL PROCESS GMM FOR

CONTINUOUS OUTCOMES WITH TWO CATEGORICAL
LATENT VARIABLES

TITLE: this is an example of a sequential

process GMM for continuous outcomes with
two categorical latent variables
DATA: FILE IS ex8.7.dat;
VARIABLE: NAMES ARE y1-y8;
CLASSES = c1 (3) c2 (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i1 s1 | y1@0 y2@1 y3@2 y4@3;
i2 s2 | y5@0 y6@1 y7@2 y8@3;
c2 ON c1;
MODEL c1:
%c1#1%
[i1 s1];

%c1#2%
[i1*1 s1];

%c1#3%
[i1*2 s1];
MODEL c2:
%c2#1%
[i2 s2];

%c2#2%
[i2*-1 s2];
OUTPUT: TECH1 TECH8;

225
CHAPTER 8

y1 y2 y3 y4 y5 y6 y7 y8

i1 s1 i2 s2

c1 c2

In this example, the sequential process growth mixture model for

continuous outcomes shown in the picture above is estimated. The latent
classes of the second process are related to the latent classes of the first
process. This is a type of latent transition analysis. Latent transition
analysis is shown in Examples 8.12, 8.13, and 8.14.

The | statements in the overall model are used to name and define the
intercept and slope growth factors in the growth models. In the first |
statement, the names i1 and s1 on the left-hand side of the | symbol are
the names of the intercept and slope growth factors, respectively. In the
second | statement, the names i2 and s2 on the left-hand side of the |
symbol are the names of the intercept and slope growth factors,
respectively. In both | statements, the values on the right-hand side of
the | symbol are the time scores for the slope growth factor. For both
growth processes, the time scores of the slope growth factors are fixed at
0, 1, 2, and 3 to define linear growth models with equidistant time
points. The zero time scores for the slope growth factors at time point
one define the intercept growth factors as initial status factors. The
coefficients of the intercept growth factors i1 and i2 are fixed at one as
part of the growth model parameterization. In the parameterization of
the growth model shown here, the means of the outcome variables at the
four time points are fixed at zero as the default. The intercept and slope
growth factor means are estimated as the default. The variances of the
growth factors are also estimated as the default. The growth factors are

226
 Examples: Mixture Modeling With Longitudinal Data

correlated as the default because they are independent (exogenous)

variables. The means of the growth factors are not held equal across
classes as the default. The variances and covariances of the growth
factors are held equal across classes as the default.

In the overall model, the ON statement describes the probabilities of

transitioning from a class of the categorical latent variable c1 to a class
of the categorical latent variable c2. The ON statement describes the
multinomial logistic regression of c2 on c1 when comparing class 1 of c2
to class 2 of c2. In this multinomial logistic regression, coefficients
corresponding to the last class of each of the categorical latent variables
are fixed at zero. The parameterization of models with more than one
categorical latent variable is discussed in Chapter 14. Because c1 has
three classes and c2 has two classes, two regression coefficients are
estimated. The means of c1 and the intercepts of c2 are estimated as the
default.

When there are multiple categorical latent variables, each one has its
own MODEL command. The MODEL command for each latent
variable is specified by MODEL followed by the name of the latent
variable. For each categorical latent variable, the part of the model that
differs for each class is specified by a label that consists of the
categorical latent variable followed by the number sign followed by the
class number. In the example above, the label %c1#1% refers to the part
of the model for class one of the categorical latent variable c1 that
differs from the overall model. The label %c2#1% refers to the part of
the model for class one of the categorical latent variable c2 that differs
from the overall model. The class-specific part of the model for each
categorical latent variable specifies that the means of the intercept and
slope growth factors are free to be estimated for each class. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors. The ESTIMATOR option of the ANALYSIS command
can be used to select a different estimator. An explanation of the other
commands can be found in Example 8.1.

Following is an alternative specification of the multinomial logistic

regression of c2 on c1:

c2#1 ON c1#1 c1#2;

227
CHAPTER 8

where c2#1 refers to the first class of c2, c1#1 refers to the first class of
c1, and c1#2 refers to the second class of c1. The classes of a
categorical latent variable are referred to by adding to the name of the
categorical latent variable the number sign (#) followed by the number
of the class. This alternative specification allows individual parameters
to be referred to in the MODEL command for the purpose of giving
starting values or placing restrictions.

EXAMPLE 8.8: GMM WITH KNOWN CLASSES (MULTIPLE

GROUP ANALYSIS)

TITLE: this is an example of GMM with known

classes (multiple group analysis)
DATA: FILE IS ex8.8.dat;
VARIABLE: NAMES ARE g y1-y4 x;
USEVARIABLES ARE y1-y4 x;
CLASSES = cg (2) c (2);
KNOWNCLASS = cg (g = 0 g = 1);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i s ON x;
c ON cg x;
%cg#1.c#1%
[i*2 s*1];
%cg#1.c#2%
[i*0 s*0];
%cg#2.c#1%
[i*3 s*1.5];
%cg#2.c#2%
[i*1 s*.5];
OUTPUT: TECH1 TECH8;

228
 Examples: Mixture Modeling With Longitudinal Data

y1 y2 y3 y4

i s

cg

x c

The difference between this example and Example 8.1 is that this
analysis includes a categorical latent variable for which class
membership is known resulting in a multiple group growth mixture
model. The CLASSES option is used to assign names to the categorical
latent variables in the model and to specify the number of latent classes
in the model for each categorical latent variable. In the example above,
there are two categorical latent variables cg and c. Both categorical
latent variables have two latent classes. The KNOWNCLASS option is
used for multiple group analysis with TYPE=MIXTURE to identify the
categorical latent variable for which latent class membership is known
and is equal to observed groups in the sample. The KNOWNCLASS
option identifies cg as the categorical latent variable for which class
membership is known. The information in parentheses following the
categorical latent variable name defines the known classes using an
observed variable. In this example, the observed variable g is used to
define the known classes. The first class consists of individuals with the
value 0 on the variable g. The second class consists of individuals with
the value 1 on the variable g.

In the overall model, the second ON statement describes the multinomial

logistic regression of the categorical latent variable c on the known class
variable cg and the covariate x. This allows the class probabilities to
vary across the observed groups in the sample. In the four class-specific

229
CHAPTER 8

parts of the model, starting values are given for the growth factor
intercepts. The four classes correspond to a combination of the classes
of cg and c. They are referred to by combining the class labels using a
period (.). For example, the combination of class 1 of cg and class 1 of c
is referred to as cg#1.c#1. The default estimator for this type of analysis
is maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 8.1.

EXAMPLE 8.9: LCGA FOR A BINARY OUTCOME

TITLE: this is an example of a LCGA for a binary

outcome
DATA: FILE IS ex8.9.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c (2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | u1@0 u2@1 u3@2 u4@3;
OUTPUT: TECH1 TECH8;

u1 u2 u3 u4

i s

230
 Examples: Mixture Modeling With Longitudinal Data

The difference between this example and Example 8.4 is that a LCGA
for a binary outcome as shown in the picture above is estimated instead
of a GMM. The difference between these two models is that GMM
allows within class variability and LCGA does not (Kreuter & Muthén,
2008; Muthén, 2004; Muthén & Asparouhov, 2009).

When TYPE=MIXTURE without ALGORITHM=INTEGRATION is

selected, a LCGA is carried out. In the parameterization of the growth
model shown here, the thresholds of the outcome variable at the four
time points are held equal as the default. The intercept growth factor
mean is fixed at zero in the last class and estimated in the other classes.
The slope growth factor mean is estimated as the default in all classes.
The variances of the growth factors are fixed at zero as the default
without ALGORITHM=INTEGRATION. Because of this, the growth
factor covariance is fixed at zero. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 8.1 and 8.4.

EXAMPLE 8.10: LCGA FOR A THREE-CATEGORY

OUTCOME

TITLE: this is an example of a LCGA for a three-

category outcome
DATA: FILE IS ex8.10.dat;
VARIABLE: NAMES ARE u1-u4;
CLASSES = c(2);
CATEGORICAL = u1-u4;
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | u1@0 u2@1 u3@2 u4@3;
! [u1$1-u4$1*-.5] (1);
! [u1$2-u4$2* .5] (2);
! %c#1%
! [i*1 s*0];
! %c#2%
! [i@0 s*0];
OUTPUT: TECH1 TECH8;

231
CHAPTER 8

The difference between this example and Example 8.9 is that the
outcome variable is an ordered categorical (ordinal) variable instead of a
binary variable. Note that the statements that are commented out are not
necessary. This results in an input identical to Example 8.9. The
statements are shown to illustrate how starting values can be given for
the thresholds and growth factor means in the model if this is needed.
Because the outcome is a three-category variable, it has two thresholds.
An explanation of the other commands can be found in Examples 8.1,
8.4 and 8.9.

EXAMPLE 8.11: LCGA FOR A COUNT OUTCOME USING A

ZERO-INFLATED POISSON MODEL

TITLE: this is an example of a LCGA for a count

outcome using a zero-inflated Poisson
model
DATA: FILE IS ex8.11.dat;
VARIABLE: NAMES ARE u1-u4;
COUNT = u1-u4 (i);
CLASSES = c (2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | u1@0 u2@1 u3@2 u4@3;
ii si | u1#1@0 u2#1@1 u3#1@2 u4#1@3;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 8.9 is that the
outcome variable is a count variable instead of a continuous variable.
The COUNT option is used to specify which dependent variables are
treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u1, u2, u3, and u4 are count variables and a zero-inflated
Poisson model is used. The count variables represent the outcome
measured at four equidistant occasions.

With a zero-inflated Poisson model, two growth models are estimated.

The first | statement describes the growth model for the count part of the
outcome for individuals who are able to assume values of zero and
above. The second | statement describes the growth model for the
inflation part of the outcome, the probability of being unable to assume
any value except zero. The binary latent inflation variable is referred to

232
 Examples: Mixture Modeling With Longitudinal Data

by adding to the name of the count variable the number sign (#) followed
by the number 1.

In the parameterization of the growth model for the count part of the
outcome, the intercepts of the outcome variable at the four time points
are fixed at zero as the default. The means of the growth factors are
estimated as the default. The variances of the growth factors are fixed
at zero. Because of this, the growth factor covariance is fixed at zero as
the default. The means of the growth factors are not held equal across
classes as the default.

In the parameterization of the growth model for the inflation part of the
outcome, the intercepts of the outcome variable at the four time points
are held equal as the default. The mean of the intercept growth factor is
fixed at zero in all classes as the default. The mean of the slope growth
factor is estimated and held equal across classes as the default. These
defaults can be overridden, but freeing too many parameters in the
inflation part of the model can lead to convergence problems. The
variances of the growth factors are fixed at zero. Because of this, the
growth factor covariance is fixed at zero. The default estimator for this
type of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 8.1 and 8.9.

EXAMPLE 8.12: HIDDEN MARKOV MODEL WITH FOUR

TIME POINTS

TITLE: this is an example of a hidden Markov

model with four time points
DATA: FILE IS ex8.12.dat;
VARIABLE: NAMES ARE u1-u4;
CATEGORICAL = u1-u4;
CLASSES = c1(2) c2(2) c3(2) c4(2);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
[c2#1-c4#1] (1);
c4 ON c3 (2);
c3 ON c2 (2);
c2 ON c1 (2);

233
CHAPTER 8

MODEL c1:
%c1#1%
[u1$1] (3);
%c1#2%
[u1$1] (4);
MODEL c2:
%c2#1%
[u2$1] (3);
%c2#2%
[u2$1] (4);
MODEL c3:
%c3#1%
[u3$1] (3);
%c3#2%
[u3$1] (4);
MODEL c4:
%c4#1%
[u4$1] (3);
%c4#2%
[u4$1] (4);
OUTPUT: TECH1 TECH8;

u1 u2 u3 u4

c1 c2 c3 c4

In this example, the hidden Markov model for a single binary outcome
measured at four time points shown in the picture above is estimated.
Although each categorical latent variable has only one latent class
indicator, this model allows the estimation of measurement error by
allowing latent class membership and observed response to disagree.
This is a first-order Markov process where the transition matrices are
specified to be equal over time (Langeheine & van de Pol, 2002). The
parameterization of this model is described in Chapter 14.

The CLASSES option is used to assign names to the categorical latent

variables in the model and to specify the number of latent classes in the

234
 Examples: Mixture Modeling With Longitudinal Data

model for each categorical latent variable. In the example above, there
are four categorical latent variables c1, c2, c3, and c4. All of the
categorical latent variables have two latent classes. In the overall model,
the transition matrices are held equal over time. This is done by placing
(1) after the bracket statement for the intercepts of c2, c3, and c4 and by
placing (2) after each of the ON statements that represent the first-order
Markov relationships. When a model has more than one categorical
latent variable, MODEL followed by a label is used to describe the
analysis model for each categorical latent variable. Labels are defined
by using the names of the categorical latent variables. The class-specific
equalities (3) and (4) represent measurement invariance across time. An
explanation of the other commands can be found in Example 8.1.

EXAMPLE 8.13: LTA FOR TWO TIME POINTS WITH A

BINARY COVARIATE INFLUENCING THE LATENT
TRANSITION PROBABILITIES

TITLE: this is an example of a LTA for two time

points with a binary covariate influencing
the latent transition probabilities
DATA: FILE = ex8.13.dat;
VARIABLE: NAMES = u11-u15 u21-u25 g;
CATEGORICAL = u11-u15 u21-u25;
CLASSES = cg (2) c1 (3) c2 (3);
KNOWNCLASS = cg (g = 0 g = 1);
ANALYSIS: TYPE = MIXTURE;
MODEL: %OVERALL%
c1 c2 ON cg;
MODEL cg: %cg#1%
c2 ON c1;
%cg#2%
c2 ON c1;
MODEL c1: %c1#1%
[u11$1] (1);
[u12$1] (2);
[u13$1] (3);
[u14$1] (4);
[u15$1] (5);
%c1#2%
[u11$1] (6);
[u12$1] (7);
[u13$1] (8);
[u14$1] (9);
[u15$1] (10);

235
CHAPTER 8

%c1#3%
[u11$1] (11);
[u12$1] (12);
[u13$1] (13);
[u14$1] (14);
[u15$1] (15);
MODEL c2:
%c2#1%
[u21$1] (1);
[u22$1] (2);
[u23$1] (3);
[u24$1] (4);
[u25$1] (5);
%c2#2%
[u21$1] (6);
[u22$1] (7);
[u23$1] (8);
[u24$1] (9);
[u25$1] (10);
%c2#3%
[u21$1] (11);
[u22$1] (12);
[u23$1] (13);
[u24$1] (14);
[u25$1] (15);
OUTPUT: TECH1 TECH8 TECH15;

u11 u12 u13 u14 u15 u21 u22 u23 u24 u25

c1 c2

cg

236
 Examples: Mixture Modeling With Longitudinal Data

In this example, the latent transition analysis (LTA; Mooijaart, 1998;

Reboussin et al., 1998; Kaplan, 2007; Nylund, 2007; Collins & Lanza,
2010) model for two time points with a binary covariate influencing the
latent transition probabilities shown in the picture above is estimated.
The same five latent class indicators are measured at two time points.
The model assumes measurement invariance across time for the five
latent class indicators. The parameterization of this model is described
in Chapter 14.

The KNOWNCLASS option is used for multiple group analysis with

TYPE=MIXTURE to identify the categorical latent variable for which
latent class membership is known and is equal to observed groups in the
sample. The KNOWNCLASS option identifies cg as the categorical
latent variable for which class membership is known. The information
in parentheses following the categorical latent variable name defines the
known classes using an observed variable. In this example, the observed
variable g is used to define the known classes. The first class consists of
individuals with the value 0 on the variable g. The second class consists
of individuals with the value 1 on the variable g.

In the overall model, the first ON statement describes the multinomial

logistic regression of the categorical latent variables c1 and c2 on the
known class variable cg. This allows the class probabilities to vary
across the observed groups in the sample.

When there are multiple categorical latent variables, each one has its
own MODEL command. The MODEL command for each categorical
latent variable is specified by MODEL followed by the name of the
categorical latent variable. In this example, MODEL cg describes the
group-specific parameters of the regression of c2 on c1. This allows the
binary covariate to influence the latent transition probabilities. MODEL
c1 describes the class-specific measurement parameters for variable c1
and MODEL c2 describes the class-specific measurement parameters for
variable c2. The model for each categorical latent variable that differs
for each class of that variable is specified by a label that consists of the
categorical latent variable name followed by the number sign followed
by the class number. For example, in the example above, the label
%c1#1% refers to class 1 of categorical latent variable c1.

In this example, the thresholds of the latent class indicators for a given
class are held equal for the two categorical latent variables. The (1-5),

237
CHAPTER 8

(6-10), and (11-15) following the bracket statements containing the

thresholds use the list function to assign equality labels to these
parameters. For example, the label 1 is assigned to the thresholds u11$1
and u21$1 which holds these thresholds equal over time.

The TECH15 option is used to obtain the transition probabilities for

each of the two known classes. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
estimator option of the ANALYSIS command can be used to select a
different estimator. An explanation of the other commands can be found
in Example 8.1.

Following is the second part of the example that shows an alternative

parameterization. The PARAMETERIZATION option is used to select
a probability parameterization rather than a logit parameterization. This
allows latent transition probabilities to be expressed directly in terms of
probability parameters instead of via logit parameters. In the overall
model, only the c1 on cg regression is specified, not the c2 on cg
regression. Other specifications are the same as in the first part of the
example.

ANALYSIS: TYPE = MIXTURE;

PARAMETERIZATION = PROBABILITY;
MODEL: %OVERALL%
c1 ON cg;
MODEL cg: %cg#1%
c2 ON c1;
%cg#2%
c2 ON c1;

EXAMPLE 8.14: LTA FOR TWO TIME POINTS WITH A

CONTINUOUS COVARIATE INFLUENCING THE LATENT
TRANSITION PROBABILITIES

TITLE: this is an example of a LTA for two time

points with a continuous covariate
influencing the latent transition
probabilities
DATA: FILE = ex8.14.dat;
VARIABLE: NAMES = u11-u15 u21-u25 x;
CATEGORICAL = u11-u15 u21-u25;
CLASSES = c1 (3) c2 (3);

238
 Examples: Mixture Modeling With Longitudinal Data

ANALYSIS: TYPE = MIXTURE;

PROCESSORS = 8;
MODEL: %OVERALL%
c1 ON x;
c2 ON c1;
MODEL c1: %c1#1%
c2 ON x;
[u11$1] (1);
[u12$1] (2);
[u13$1] (3);
[u14$1] (4);
[u15$1] (5);
%c1#2%
c2 ON x;
[u11$1] (6);
[u12$1] (7);
[u13$1] (8);
[u14$1] (9);
[u15$1] (10);
%c1#3%
c2 ON x;
[u11$1] (11);
[u12$1] (12);
[u13$1] (13);
[u14$1] (14);
[u15$1] (15);
MODEL c2: %c2#1%
[u21$1] (1);
[u22$1] (2);
[u23$1] (3);
[u24$1] (4);
[u25$1] (5);
%c2#2%
[u21$1] (6);
[u22$1] (7);
[u23$1] (8);
[u24$1] (9);
[u25$1] (10);
%c2#3%
[u21$1] (11);
[u22$1] (12);
[u23$1] (13);
[u24$1] (14);
[u25$1] (15);
OUTPUT: TECH1 TECH8;

239
CHAPTER 8

u11 u12 u13 u14 u15 u21 u22 u23 u24 u25

c1 c2

In this example, the latent transition analysis (LTA; Reboussin et al.,

1998; Kaplan, 2007; Nylund, 2007; Collins & Lanza, 2010) model for
two time points with a continuous covariate influencing the latent
transition probabilities shown in the picture above is estimated. The
same five latent class indicators are measured at two time points. The
model assumes measurement invariance across time for the five latent
class indicators. The parameterization of this model is described in
Chapter 14.

In the overall model, the first ON statement describes the multinomial

logistic regression of the categorical latent variable c1 on the continuous
covariate x. The second ON statement describes the multinomial logistic
regression of c2 on c1. The multinomial logistic regression of c2 on the
continuous covariate x is specified in the class-specific parts of MODEL
c1. This follows parameterization 2 discussed in Muthén and
Asparouhov (2011). The class-specific regressions of c2 on x allow the
continuous covariate x to influence the latent transition probabilities.
The latent transition probabilities for different values of the covariates
can be computed by choosing LTA calculator from the Mplus menu of
the Mplus Editor.

When there are multiple categorical latent variables, each one has its
own MODEL command. The MODEL command for each categorical
latent variable is specified by MODEL followed by the name of the
categorical latent variable. MODEL c1 describes the class-specific

240
 Examples: Mixture Modeling With Longitudinal Data

multinomial logistic regression of c2 on x and the class-specific

measurement parameters for variable c1. MODEL c2 describes the
class-specific measurement parameters for variable c2. The model for
each categorical latent variable that differs for each class of that variable
is specified by a label that consists of the categorical latent variable
name followed by the number sign followed by the class number. For
example, in the example above, the label %c1#1% refers to class 1 of
categorical latent variable c1.

In this example, the thresholds of the latent class indicators for a given
class are held equal for the two categorical latent variables. The (1-5),
(6-10), and (11-15) following the bracket statements containing the
thresholds use the list function to assign equality labels to these
parameters. For example, the label 1 is assigned to the thresholds u11$1
and u21$1 which holds these thresholds equal over time. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors. The estimator option of the ANALYSIS command can
be used to select a different estimator. An explanation of the other
commands can be found in Example 8.1.

EXAMPLE 8.15: MOVER-STAYER LTA FOR THREE TIME

POINTS USING A PROBABILITY PARAMETERIZATION

TITLE: this is an example of a mover-stayer LTA

for three time points using a probability
parameterization
DATA: FILE = ex8.15.dat;
VARIABLE: NAMES = u11-u15 u21-u25 u31-u35;
CATEGORICAL = u11-u15 u21-u25 u31-u35;
CLASSES = c(2) c1(3) c2(3) c3(3);
ANALYSIS: TYPE = MIXTURE;
PARAMETERIZATION = PROBABILITY;
STARTS = 100 20;
PROCESSORS = 8;
MODEL: %OVERALL%
c1 ON c;
MODEL c: %c#1% !mover class
c2 ON c1;
c3 ON c2;
%c#2% ! stayer class
c2#1 ON c1#1@1; c2#2 ON c1#1@0;
c2#1 ON c1#2@0; c2#2 ON c1#2@1;
c2#1 ON c1#3@0; c2#2 ON c1#3@0;

241
CHAPTER 8

c3#1 ON c2#1@1; c3#2 ON c2#1@0;

c3#1 ON c2#2@0; c3#2 ON c2#2@1;
c3#1 ON c2#3@0; c3#2 ON c2#3@0;
MODEL c1: %c1#1%
[u11$1] (1);
[u12$1] (2);
[u13$1] (3);
[u14$1] (4);
[u15$1] (5);
%c1#2%
[u11$1] (6);
[u12$1] (7);
[u13$1] (8);
[u14$1] (9);
[u15$1] (10);
%c1#3%
[u11$1] (11);
[u12$1] (12);
[u13$1] (13);
[u14$1] (14);
[u15$1] (15);
MODEL c2:
%c2#1%
[u21$1] (1);
[u22$1] (2);
[u23$1] (3);
[u24$1] (4);
[u25$1] (5);
%c2#2%
[u21$1] (6);
[u22$1] (7);
[u23$1] (8);
[u24$1] (9);
[u25$1] (10);
%c2#3%
[u21$1] (11);
[u22$1] (12);
[u23$1] (13);
[u24$1] (14);
[u25$1] (15);
MODEL c3:
%c3#1%
[u31$1] (1);
[u32$1] (2);
[u33$1] (3);
[u34$1] (4);
[u35$1] (5);
%c3#2%
[u31$1] (6);
[u32$1] (7);

242
 Examples: Mixture Modeling With Longitudinal Data

[u33$1] (8);
[u34$1] (9);
[u35$1] (10);
%c3#3%
[u31$1] (11);
[u32$1] (12);
[u33$1] (13);
[u34$1] (14);
[u35$1] (15);
OUTPUT: TECH1 TECH8 TECH15;

u11 ... u15 u21 ... u25 u31 ... u35

c1 c2 c3

In this example, the mover-stayer (Langeheine & van de Pol, 2002)

latent transition analysis (LTA) for three time points using a probability
parameterization shown in the picture above is estimated. The same five
latent class indicators are measured at three time points. The model
assumes measurement invariance across time for the five latent class
indicators. The parameterization of this model is described in Chapter
14.

The PARAMETERIZATION option is used to select a probability

parameterization rather than a logit parameterization. This allows latent
transition probabilities to be expressed directly in terms of probability

243
CHAPTER 8

parameters instead of via logit parameters. The alternative logit

parameterization of mover-stayer LTA is described in the document
LTA With Movers-Stayers (see FAQ, www.statmodel.com).

In the overall model, the ON statement describes the multinomial

logistic regression of the categorical latent variable c1 on the mover-
stayer categorical latent variable c. The multinomial logistic regressions
of c2 on c1 and c3 on c2 are specified in the class-specific parts of
MODEL c.

When there are multiple categorical latent variables, each one has its
own MODEL command. The MODEL command for each categorical
latent variable is specified by MODEL followed by the name of the
categorical latent variable. MODEL c describes the class-specific
multinomial logistic regressions of c2 on c1 and c3 on c2 where the first
c class is the mover class and the second c class is the stayer class.
MODEL c1 describes the class-specific measurement parameters for
variable c1; MODEL c2 describes the class-specific measurement
parameters for variable c2; and MODEL c3 describes the class-specific
measurement parameters for variable c3. The model for each categorical
latent variable that differs for each class of that variable is specified by a
label that consists of the categorical latent variable name followed by the
number sign followed by the class number. For example, in the example
above, the label %c1#1% refers to class 1 of categorical latent variable
c1.

In class 1, the mover class of MODEL c, the two ON statements specify

that the latent transition probabilities are estimated. In class 2, the stayer
class, the ON statements specify that the latent transition probabilities
are fixed at either zero or one. A latent transition probability of one
specifies that an observation stays in the same class across time.

In this example, the thresholds of the latent class indicators for a given
class are held equal for the three categorical latent variables. The (1-5),
(6-10), and (11-15) following the bracket statements containing the
thresholds use the list function to assign equality labels to these
parameters. For example, the label 1 is assigned to the thresholds
u11$1, u21$1, and u31$1 which holds these thresholds equal over time.

The TECH15 option is used to obtain the transition probabilities for both
the mover and stayer classes. The default estimator for this type of

244
 Examples: Mixture Modeling With Longitudinal Data

analysis is maximum likelihood with robust standard errors. The

estimator option of the ANALYSIS command can be used to select a
different estimator. An explanation of the other commands can be found
in Example 8.1.

EXAMPLE 8.16: DISCRETE-TIME SURVIVAL MIXTURE

ANALYSIS WITH SURVIVAL PREDICTED BY GROWTH
TRAJECTORY CLASSES

TITLE: this is an example of a discrete-time

survival mixture analysis with survival
predicted by growth trajectory classes
DATA: FILE IS ex8.16.dat;
VARIABLE: NAMES ARE y1-y3 u1-u4;
CLASSES = c(2);
CATEGORICAL = u1-u4;
MISSING = u1-u4 (999);
ANALYSIS: TYPE = MIXTURE;
MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2;
f BY u1-u4@1;
OUTPUT: TECH1 TECH8;

y1 y2 y3 u1 u2 u3 u4

i s f

245
CHAPTER 8

In this example, the discrete-time survival mixture analysis model shown

in the picture above is estimated. In this model, a survival model for u1,
u2, u3, and u4 is specified for each class of c defined by a growth
mixture model for y1-y3 (Muthén & Masyn, 2005). Each u variable
represents whether or not a single non-repeatable event has occurred in a
specific time period. The value 1 means that the event has occurred, 0
means that the event has not occurred, and a missing value flag means
that the event has occurred in a preceding time period or that the
individual has dropped out of the study. The factor f is used to specify a
proportional odds assumption for the hazards of the event. The arrows
from c to the growth factors i and s indicate that the means of the growth
factors vary across the classes of c.

In the overall model, the | symbol is used to name and define the
intercept and slope growth factors in a growth model. The names i and s
on the left-hand side of the | symbol are the names of the intercept and
slope growth factors, respectively. The statement on the right-hand side
of the | symbol specifies the outcomes and the time scores for the growth
model. The time scores for the slope growth factor are fixed at 0, 1, and
2 to define a linear growth model with equidistant time points. The zero
time score for the slope growth factor at time point one defines the
intercept growth factor as an initial status factor. The coefficients of the
intercept growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are
estimated and allowed to be different across time and the residuals are
not correlated as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The means and variances of the growth factors are estimated as
the default, and the growth factor covariance is estimated as the default
because they are independent (exogenous) variables. The means of the
growth factors are not held equal across classes as the default. The
variances and covariance of the growth factors are held equal across
classes as the default.

In the overall model, the BY statement specifies that f is measured by

u1, u2, u3, and u4 where the factor loadings are fixed at one. This
represents a proportional odds assumption. The mean of f is fixed at
zero in class two as the default. The variance of f is fixed at zero in both

246
 Examples: Mixture Modeling With Longitudinal Data

classes. A variance for f can be estimated by using

ALGORITHM=INTEGRATION as is done in Example 6.19. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 8.1.

EXAMPLE 8.17: CONTINUOUS-TIME SURVIVAL MIXTURE

ANALYSIS USING A COX REGRESSION MODEL

TITLE: this is an example of a continuous-time

survival mixture analysis using a Cox
regression model
DATA: FILE = ex8.17.dat;
VARIABLE: NAMES = t u1-u5 x tc;
CATEGORICAL = u1-u5;
CLASSES = c (2);
SURVIVAL = t (ALL);
TIMECENSORED = tc (0 = NOT 1 = RIGHT);
ANALYSIS: TYPE = MIXTURE;
BASEHAZARD = OFF;
MODEL: %OVERALL%
t ON x;
c ON x;
%c#1%
[u1$1-u5$1];
t ON x;
%c#2%
[u1$1-u5$1];
t ON x;
OUTPUT: TECH1 TECH8;

247
CHAPTER 8

u1 u2 u3 u4 u5 t

In this example, the continuous-time survival analysis model shown in

the picture above is estimated. This is a Cox regression mixture model
similar to the model of Larsen (2004) as discussed in Asparouhov et al.
(2006). The profile likelihood method is used for estimation.

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option must be used in conjunction with the
TIMECENSORED option. In this example, t is the variable that
contains time-to-event information. By specifying the keyword ALL in
parenthesis following the time-to-event variable, the time intervals are
taken from the data. The TIMECENSORED option is used to identify
the variables that contain information about right censoring. In this
example, the variable is named tc. The information in parentheses
specifies that the value zero represents no censoring and the value one
represents right censoring. This is the default. The BASEHAZARD
option of the ANALYSIS command is used with continuous-time
survival analysis to specify if a non-parametric or a parametric baseline
hazard function is used in the estimation of the model. The setting OFF
specifies that a non-parametric baseline hazard function is used. This is
the default.

In the overall model, the first ON statement describes the loglinear

regression of the time-to-event variable t on the covariate x. The second

248
 Examples: Mixture Modeling With Longitudinal Data

ON statement describes the multinomial logistic regression of the

categorical latent variable c on the covariate x. In the class-specific
models, by specifying the thresholds of the latent class indicator
variables and the regression of the time-to-event t on the covariate x,
these parameters will be estimated separately for each class. The non-
parametric baseline hazard function varies across class as the default.
The default estimator for this type of analysis is maximum likelihood
with robust standard errors. The estimator option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 8.1.

249
CHAPTER 8

250
 Examples: Multilevel Modeling With Complex Survey Data

CHAPTER 9
EXAMPLES: MULTILEVEL
MODELING WITH COMPLEX
SURVEY DATA

Complex survey data refers to data obtained by stratification, cluster

sampling and/or sampling with an unequal probability of selection.
Complex survey data are also referred to as multilevel or hierarchical
data. For an overview, see Muthén and Satorra (1995). There are two
approaches to the analysis of complex survey data in Mplus.

One approach is to compute standard errors and a chi-square test of

model fit taking into account stratification, non-independence of
observations due to cluster sampling, and/or unequal probability of
selection. Subpopulation analysis is also available. With sampling
weights, parameters are estimated by maximizing a weighted
loglikelihood function. Standard error computations use a sandwich
estimator. This approach can be obtained by specifying
TYPE=COMPLEX in the ANALYSIS command in conjunction with the
STRATIFICATION, CLUSTER, WEIGHT, and/or SUBPOPULATION
options of the VARIABLE command. Observed outcome variables can
be continuous, censored, binary, ordered categorical (ordinal), unordered
categorical (nominal), counts, or combinations of these variable types.
The implementation of these methods in Mplus is discussed in
Asparouhov (2005, 2006) and Asparouhov and Muthén (2005, 2006a).

A second approach is to specify a model for each level of the multilevel

data thereby modeling the non-independence of observations due to
cluster sampling. This is commonly referred to as multilevel modeling.
The use of sampling weights in the estimation of parameters, standard
errors, and the chi-square test of model fit is allowed. Both individual-
level and cluster-level weights can be used. With sampling weights,
parameters are estimated by maximizing a weighted loglikelihood
function. Standard error computations use a sandwich estimator. This
approach can be obtained for two-level data by specifying
TYPE=TWOLEVEL in the ANALYSIS command in conjunction with
the CLUSTER, WEIGHT, WTSCALE, BWEIGHT, and/or BWTSCALE

251
CHAPTER 9

options of the VARIABLE command. For TYPE=TWOLEVEL,

observed outcome variables can be continuous, censored, binary, ordered
categorical (ordinal), unordered categorical (nominal), counts, or
combinations of these variable types. This approach can also be
obtained for three-level data by specifying TYPE=THREELEVEL in
conjunction with the CLUSTER, WEIGHT, WTSCALE, B2WEIGHT,
B3WEIGHT and/or BWTSCALE options of the VARIABLE command.
For TYPE=THREELEVEL, observed outcome variables can be
continuous. Complex survey features are not available for
TYPE=THREELEVEL with categorical variables or
TYPE=CROSSCLASSIFIED because these models are estimated using
Bayesian analysis for which complex survey features have not been
generally developed.

The approaches described above can be combined by specifying

TYPE=COMPLEX TWOLEVEL in the ANALYSIS command in
conjunction with the STRATIFICATION, CLUSTER, WEIGHT,
WTSCALE, BWEIGHT, and/or BWTSCALE options of the
VARIABLE command or TYPE=COMPLEX THREELEVEL in
conjunction with the STRATIFICATION, CLUSTER, WEIGHT,
WTSCALE, B2WEIGHT, B3WEIGHT, and/or BWTSCALE options of
the VARIABLE command For TYPE=TWOLEVEL, when there is
clustering due to two cluster variables, the standard errors and chi-square
test of model fit are computed taking into account the clustering due to
the highest cluster level using TYPE=COMPLEX whereas clustering
due to the lowest cluster level is modeled using TYPE=TWOLEVEL.
For TYPE=THREELEVEL, when there is clustering due to three cluster
variables, the standard errors and chi-square test of model fit are
computed taking into account the clustering due to the highest cluster
level using TYPE=COMPLEX whereas clustering due to the other
cluster levels is modeled using TYPE=THREELEVEL.

A distinction can be made between cross-sectional data in which non-

independence arises because of cluster sampling and longitudinal data in
which non-independence arises because of repeated measures of the
same individuals across time. With cross-sectional data, the number of
levels in Mplus is the same as the number of levels in conventional
multilevel modeling programs. Mplus allows three-level modeling.
With longitudinal data, the number of levels in Mplus is one less than
the number of levels in conventional multilevel modeling programs
because Mplus takes a multivariate approach to repeated measures

252
 Examples: Multilevel Modeling With Complex Survey Data

analysis. Longitudinal models are two-level models in conventional

multilevel programs, whereas they are single-level models in Mplus.
These models are discussed in Chapter 6. Three-level analysis where
time is the first level, individual is the second level, and cluster is the
third level is handled by two-level modeling in Mplus (see also Muthén,
1997). Four-level analysis where time is the first level, individual is the
second level, classroom is the third level, and school is the fourth level is
handled by three-level modeling in Mplus.

The general latent variable modeling framework of Mplus allows the

integration of random effects and other continuous latent variables
within a single analysis model. Random effects are allowed for both
independent and dependent variables and both observed and latent
variables. Random effects representing across-cluster variation in
intercepts and slopes or individual differences in growth can be
combined with factors measured by multiple indicators on both the
individual and cluster levels. Random factor loadings are available as a
special case of random slopes. In line with SEM, regressions among
random effects, among factors, and between random effects and factors
are allowed.

Multilevel models can include regression analysis, path analysis,

confirmatory factor analysis (CFA), item response theory (IRT) analysis,
structural equation modeling (SEM), latent class analysis (LCA), latent
transition analysis (LTA), latent class growth analysis (LCGA), growth
mixture modeling (GMM), discrete-time survival analysis, continuous-
time survival analysis, and combinations of these models.

For TYPE=TWOLEVEL, there are four estimator options. The first

estimator option is full-information maximum likelihood which allows
continuous, censored, binary, ordered categorical (ordinal), unordered
categorical (nominal), counts, or combinations of these variable types;
random intercepts and slopes; and missing data. With longitudinal data,
maximum likelihood estimation allows modeling of individually-varying
times of observation and random slopes for time-varying covariates.
Non-normality robust standard errors and a chi-square test of model fit
are available. The second estimator option is limited-information
weighted least squares (Asparouhov & Muthén, 2007) which allows
continuous, binary, ordered categorical (ordinal), and combinations of
these variables types; random intercepts; and missing data. The third
estimator option is the Muthén limited information estimator (MUML;

253
CHAPTER 9

Muthén, 1994) which is restricted to models with continuous variables,

random intercepts, and no missing data. The fourth estimator option is
Bayes which allows continuous, categorical, and combinations of these
variable types; random intercepts and slopes; and missing data.

All two-level models can be estimated using the following special

features:

• Multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Wald chi-square test of parameter equalities

For continuous, censored with weighted least squares estimation, binary,

and ordered categorical (ordinal) outcomes, multiple group analysis is
specified by using the GROUPING option of the VARIABLE command
for individual data. For censored with maximum likelihood estimation,
unordered categorical (nominal), and count outcomes, multiple group
analysis is specified using the KNOWNCLASS option of the
VARIABLE command in conjunction with the TYPE=MIXTURE
option of the ANALYSIS command. The default is to estimate the
model under missing data theory using all available data. The
LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. Latent variable
interactions are specified by using the | symbol of the MODEL command
in conjunction with the XWITH option of the MODEL command.
Random slopes are specified by using the | symbol of the MODEL
command in conjunction with the ON option of the MODEL command.

254
 Examples: Multilevel Modeling With Complex Survey Data

Individually-varying times of observations are specified by using the |

symbol of the MODEL command in conjunction with the AT option of
the MODEL command and the TSCORES option of the VARIABLE
command. Linear and non-linear parameter constraints are specified by
using the MODEL CONSTRAINT command. Indirect effects are
specified by using the MODEL INDIRECT command. Maximum
likelihood estimation is specified by using the ESTIMATOR option of
the ANALYSIS command. The MODEL TEST command is used to test
linear restrictions on the parameters in the MODEL and MODEL
CONSTRAINT commands using the Wald chi-square test.

For TYPE=THREELEVEL, there are two estimator options. The first

estimator option is full-information maximum likelihood which allows
continuous variables; random intercepts and slopes; and missing data.
Non-normality robust standard errors and a chi-square test of model fit
are available. The second estimator option is Bayes which allows
continuous, categorical, and combinations of these variable types;
random intercepts and slopes; and missing data.

All three-level models can be estimated using the following special

features:

• Multiple group analysis

• Missing data
• Complex survey data
• Random slopes
• Linear and non-linear parameter constraints
• Maximum likelihood estimation for all outcome types
• Wald chi-square test of parameter equalities

For continuous outcomes, multiple group analysis is specified by using

the GROUPING option of the VARIABLE command. The default is to
estimate the model under missing data theory using all available data.
The LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER, and
WEIGHT options of the VARIABLE command. Random slopes are

255
CHAPTER 9

specified by using the | symbol of the MODEL command in conjunction

with the ON option of the MODEL command. Linear and non-linear
parameter constraints are specified by using the MODEL
CONSTRAINT command. Maximum likelihood estimation is specified
by using the ESTIMATOR option of the ANALYSIS command. The
MODEL TEST command is used to test linear restrictions on the
parameters in the MODEL and MODEL CONSTRAINT commands
using the Wald chi-square test.

For TYPE=CROSSCLASSIFIED, there is one estimator option, Bayes,

which allows continuous, categorical, and combinations of these variable
types; random intercepts and slopes; and missing data.

All cross-classified models can be estimated using the following special

features:

• Missing data
• Random slopes

The default is to estimate the model under missing data theory using all
available data. The LISTWISE option of the DATA command can be
used to delete all observations from the analysis that have missing values
on one or more of the analysis variables. Random slopes are specified
by using the | symbol of the MODEL command in conjunction with the
ON option of the MODEL command.

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, and plots
of sample and estimated means and proportions/probabilities. These are
available for the total sample, by group, by class, and adjusted for
covariates. The PLOT command includes a display showing a set of
descriptive statistics for each variable. The graphical displays can be
edited and exported as a DIB, EMF, or JPEG file. In addition, the data
for each graphical display can be saved in an external file for use by
another graphics program.

Following is the set of cross-sectional multilevel modeling examples

included in this chapter:

256
 Examples: Multilevel Modeling With Complex Survey Data

• 9.1: Two-level regression analysis for a continuous dependent

variable with a random intercept
• 9.2: Two-level regression analysis for a continuous dependent
variable with a random slope
• 9.3: Two-level path analysis with a continuous and a categorical
dependent variable*
• 9.4: Two-level path analysis with a continuous, a categorical, and a
cluster-level observed dependent variable
• 9.5: Two-level path analysis with continuous dependent variables
and random slopes*
• 9.6: Two-level CFA with continuous factor indicators and
covariates
• 9.7: Two-level CFA with categorical factor indicators and
covariates*
• 9.8: Two-level CFA with continuous factor indicators, covariates,
and random slopes
• 9.9: Two-level SEM with categorical factor indicators on the within
level and cluster-level continuous observed and random intercept
factor indicators on the between level
• 9.10: Two-level SEM with continuous factor indicators and a
random slope for a factor*
• 9.11: Two-level multiple group CFA with continuous factor
indicators

Following is the set of longitudinal multilevel modeling examples

included in this chapter:

• 9.12: Two-level growth model for a continuous outcome (three-

level analysis)
• 9.13: Two-level growth model for a categorical outcome (three-
level analysis)*
• 9.14: Two-level growth model for a continuous outcome (three-
level analysis) with variation on both the within and between levels
for a random slope of a time-varying covariate*
• 9.15: Two-level multiple indicator growth model with categorical
outcomes (three-level analysis)
• 9.16: Linear growth model for a continuous outcome with time-
invariant and time-varying covariates carried out as a two-level
growth model using the DATA WIDETOLONG command

257
CHAPTER 9

• 9.17: Two-level growth model for a count outcome using a zero-

inflated Poisson model (three-level analysis)*
• 9.18: Two-level continuous-time survival analysis using Cox
regression with a random intercept
• 9.19: Two-level mimic model with continuous factor indicators,
random factor loadings, two covariates on within, and one covariate
on between with equal loadings across levels
• 9.20: Three-level regression for a continuous dependent variable
• 9.21: Three-level path analysis with a continuous and a categorical
dependent variable
• 9.22: Three-level MIMIC model with continuous factor indicators,
two covariates on within, one covariate on between level 2, one
covariate on between level 3 with random slopes on both within and
between level 2
• 9.23: Three-level growth model with a continuous outcome and one
covariate on each of the three levels
• 9.24: Regression for a continuous dependent variable using cross-
classified data
• 9.25: Path analysis with continuous dependent variables using
cross-classified data
• 9.26: IRT with random binary items using cross-classified data
• 9.27: Multiple indicator growth model with random intercepts and
factor loadings using cross-classified data

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

258
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.1: TWO-LEVEL REGRESSION ANALYSIS FOR A

CONTINUOUS DEPENDENT VARIABLE WITH A RANDOM
INTERCEPT

TITLE: this is an example of a two-level

regression analysis for a continuous
dependent variable with a random intercept
and an observed covariate
DATA: FILE = ex9.1a.dat;
VARIABLE: NAMES = y x w xm clus;
WITHIN = x;
BETWEEN = w xm;
CLUSTER = clus;
DEFINE: CENTER x (GRANDMEAN);
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
y ON x;
%BETWEEN%
y ON w xm;

x y

Within

w Between

xm

259
CHAPTER 9

In this example, the two-level regression model shown in the picture

above is estimated. The dependent variable y in this regression is
continuous. Two ways of treating the covariate x are described. In this
part of the example, the covariate x is treated as an observed variable in
line with conventional multilevel regression modeling. In the second
part of the example, the covariate x is decomposed into two latent
variable parts.

The within part of the model describes the regression of y on an

observed covariate x where the intercept is a random effect that varies
across the clusters. In the within part of the model, the filled circle at the
end of the arrow from x to y represents a random intercept that is
referred to as y in the between part of the model. In the between part of
the model, the random intercept is shown in a circle because it is a
continuous latent variable that varies across clusters. The between part
of the model describes the linear regression of the random intercept y on
observed cluster-level covariates w and xm. The observed cluster-level
covariate xm takes the value of the mean of x for each cluster. The
within and between parts of the model correspond to level 1 and level 2
of a conventional multilevel regression model with a random intercept.

TITLE: this is an example of a two-level

regression analysis for a continuous
dependent variable with a random intercept
and an observed covariate

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE = ex9.1a.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex9.1a.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES = y x w xm clus;

WITHIN = x;
BETWEEN = w xm;
CLUSTER = clus;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to

260
 Examples: Multilevel Modeling With Complex Survey Data

assign names to the variables in the data set. The data set in this
example contains five variables: y, x, w, xm, and clus.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and modeled only on the within
level. They are specified to have no variance in the between part of the
model. The BETWEEN option is used to identify the variables in the
data set that are measured on the cluster level and modeled only on the
between level. Variables not mentioned on the WITHIN or the
BETWEEN statements are measured on the individual level and can be
modeled on both the within and between levels. Because y is not
mentioned on the WITHIN statement, it is modeled on both the within
and between levels. On the between level, it is a random intercept. The
CLUSTER option is used to identify the variable that contains clustering
information. The CENTERING option is used to specify the type of
centering to be used in an analysis and the variables that are to be
centered. In this example, grand-mean centering is chosen.

DEFINE: CENTER x (GRANDMEAN);

The DEFINE command is used to transform existing variables and create

new variables. The CENTER option is used to specify the type of
centering to be used in an analysis and the variables that will be
centered. Centering facilitates the interpretation of the results. In this
example, the covariate is centered using the grand mean, that is, the
sample mean of x is subtracted from the values of the covariate x.

ANALYSIS: TYPE = TWOLEVEL;

The ANALYSIS command is used to describe the technical details of the

analysis. By selecting TWOLEVEL, a multilevel model with random
intercepts will be estimated.

MODEL:
%WITHIN%
y ON x;
%BETWEEN%
y ON w xm;

The MODEL command is used to describe the model to be estimated. In

multilevel models, a model is specified for both the within and between
parts of the model. In the within part of the model, the ON statement

261
CHAPTER 9

describes the linear regression of y on the observed individual-level

covariate x. The within-level residual variance in the regression of y on
x is estimated as the default.

In the between part of the model, the ON statement describes the linear
regression of the random intercept y on the observed cluster-level
covariates w and xm. The intercept and residual variance of y are
estimated as the default. The default estimator for this type of analysis
is maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator.

Following is the second part of the example where the covariate x is

decomposed into two latent variable parts.

TITLE: this is an example of a two-level

regression analysis for a continuous
dependent variable with a random intercept
and a latent covariate
DATA: FILE = ex9.1b.dat;
VARIABLE: NAMES = y x w clus;
BETWEEN = w;
CLUSTER = clus;
DEFINE: CENTER = x (GRANDMEAN);
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
y ON x (gamma10);
%BETWEEN%
y ON w
x (gamma01);
MODEL CONSTRAINT:
NEW(betac);
betac = gamma01 - gamma10;

The difference between this part of the example and the first part is that
the covariate x is decomposed into two latent variable parts instead of
being treated as an observed variable as in conventional multilevel
regression modeling. The decomposition occurs when the covariate x is
not mentioned on the WITHIN statement and is therefore modeled on
both the within and between levels. When a covariate is not mentioned
on the WITHIN statement, it is decomposed into two uncorrelated latent
variables,

262
 Examples: Multilevel Modeling With Complex Survey Data

xij = xwij + xbj ,

where i represents individual, j represents cluster, xwij is the latent

variable covariate used on the within level, and xbj is the latent variable
covariate used on the between level. This model is described in Muthén
(1989, 1990, 1994). The latent variable covariate xb is not used in
conventional multilevel analysis. Using a latent covariate may, however,
be advantageous when the observed cluster-mean covariate xm does not
have sufficient reliability resulting in biased estimation of the between-
level slope (Asparouhov & Muthén, 2006b; Ludtke et al., 2008).

The decomposition can be expressed as,

xwij = xij - xbj ,

which can be viewed as an implicit, latent group-mean centering of the

latent within-level covariate. To obtain results that are not group-mean
centered, a linear transformation of the within and between slopes can be
done as described below using the MODEL CONSTRAINT command.

In the MODEL command, the label gamma10 in the within part of the
model and the label gamma01 in the between part of the model are
assigned to the regression coefficients in the linear regression of y on x
in both parts of the model for use in the MODEL CONSTRAINT
command. The MODEL CONSTRAINT command is used to define
linear and non-linear constraints on the parameters in the model. In the
MODEL CONSTRAINT command, the NEW option is used to
introduce a new parameter that is not part of the MODEL command.
This parameter is called betac and is defined as the difference between
gamma01 and gamma10. It corresponds to a “contextual effect” as
described in Raudenbush and Bryk (2002, p. 140, Table 5.11).

263
CHAPTER 9

EXAMPLE 9.2: TWO-LEVEL REGRESSION ANALYSIS FOR A

CONTINUOUS DEPENDENT VARIABLE WITH A RANDOM
SLOPE

TITLE: this is an example of a two-level

regression analysis for a continuous
dependent variable with a random slope and
an observed covariate
DATA: FILE = ex9.2a.dat;
VARIABLE: NAMES = y x w xm clus;
WITHIN = x;
BETWEEN = w xm;
CLUSTER = clus;
DEFINE: CENTER x (GRANDMEAN);
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL:
%WITHIN%
s | y ON x;
%BETWEEN%
y s ON w xm;
y WITH s;

s
x y

Within

w y Between

xm s

264
 Examples: Multilevel Modeling With Complex Survey Data

The difference between this example and the first part of Example 9.1 is
that the model has both a random intercept and a random slope. In the
within part of the model, the filled circle at the end of the arrow from x
to y represents a random intercept that is referred to as y in the between
part of the model. The filled circle on the arrow from x to y represents a
random slope that is referred to as s in the between part of the model. In
the between part of the model, the random intercept and random slope
are shown in circles because they are continuous latent variables that
vary across clusters. The observed cluster-level covariate xm takes the
value of the mean of x for each cluster. The within and between parts of
the model correspond to level 1 and level 2 of a conventional multilevel
regression model with a random intercept and a random slope.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random slope variables in the
model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined
using the ON option. The random slope s is defined by the linear
regression of the dependent variable y on the observed individual-level
covariate x. The within-level residual variance in the regression of y on
x is estimated as the default.

In the between part of the model, the ON statement describes the linear
regressions of the random intercept y and the random slope s on the
observed cluster-level covariates w and xm. The intercepts and residual
variances of s and y are estimated and the residuals are not correlated as
the default. The WITH statement specifies that the residuals of s and y
are correlated. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 9.1.

Following is the second part of the example that shows how to plot a
cross-level interaction where the cluster-level covariate w moderates the
influence of the within-level covariate x on y.

265
CHAPTER 9

MODEL: %WITHIN%
s | y ON x;
%BETWEEN%
y ON w xm;
[s] (gam0);
s ON w (gam1)
xm;
y WITH s;
MODEL CONSTRAINT:
PLOT(ylow yhigh);
LOOP(level1,-3,3,0.01);
ylow = (gam0+gam1*(-1))*level1;
yhigh = (gam0+gam1*1)*level1;
PLOT: TYPE = PLOT2;

In MODEL CONSTRAINT, the LOOP option is used in conjunction

with the PLOT option to create plots of variables. In this example,
cross-level interaction effects defined in MODEL CONSTRAINT will
be plotted. The PLOT option names the variables that will be plotted on
the y-axis. The LOOP option names the variable that will be plotted on
the x-axis, gives the numbers that are the lower and upper values of the
variable, and the incremental value of the variable to be used in the
computations. In this example, the variables ylow and yhigh will be on
the y-axes and the variable level1 will be on the x-axes. The variable
level1, representing the x covariate, varies over the range of x that is of
interest such as three standard deviations away from its mean. The
lower and upper values of level1 are -3 and 3 and 0.01 is the incremental
value of level1 to use in the computations. When level1 appears in a
MODEL CONSTRAINT statement involving a new parameter, that
statement is evaluated for each value of level1 specified by the LOOP
option. For example, the first value of level1 is -3; the second value of
level1 is -3 plus 0.01 or -2.99; the third value of level1 is -2.99 plus 0.01
or -2.98; the last value of level1 is 3. Ylow and yhigh use the values -1
and 1 of the cluster-level covariate w to represent minus one standard
deviation and plus one standard deviation from the mean for w. The
cross-level interaction effects are evaluated at the value zero for the
cluster-level covariate xm.

Using TYPE=PLOT2 in the PLOT command, the plots of ylow and

yhigh and level1 can be viewed by choosing Loop plots from the Plot
menu of the Mplus Editor. The plots present the computed values along
with a 95% confidence interval. For Bayesian estimation, the default is
credibility intervals of the posterior distribution with equal tail
percentages. The CINTERVAL option of the OUTPUT command can

266
 Examples: Multilevel Modeling With Complex Survey Data

be used to obtain credibility intervals of the posterior distribution that

give the highest posterior density.

Following is the third part of the example that shows an alternative

treatment of the observed covariate x.

TITLE: this is an example of a two-level

regression analysis for a continuous
dependent variable with a random slope and
a latent covariate
DATA: FILE = ex9.2b.dat;
VARIABLE: NAMES = y x w clus;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL:
%WITHIN%
s | y ON x;
%BETWEEN%
y s ON w x;
y WITH s;

The difference between this part of the example and the first part of the
example is that the covariate x is latent instead of observed on the
between level. This is achieved when the individual-level observed
covariate is modeled in both the within and between parts of the model.
This is requested by not mentioning the observed covariate x on the
WITHIN statement in the VARIABLE command. When a random slope
is estimated, the observed covariate x is used on the within level and the
latent variable covariate xbj is used on the between level. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 9.1.

267
CHAPTER 9

EXAMPLE 9.3: TWO-LEVEL PATH ANALYSIS WITH A

CONTINUOUS AND A CATEGORICAL DEPENDENT
VARIABLE

TITLE: this is an example of a two-level path

analysis with a continuous and a
categorical dependent variable
DATA: FILE IS ex9.3.dat;
VARIABLE: NAMES ARE u y x1 x2 w clus;
CATEGORICAL = u;
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER IS clus;
ANALYSIS: TYPE = TWOLEVEL;
ALGORITHM = INTEGRATION;
MODEL:
%WITHIN%
y ON x1 x2;
u ON y x2;
%BETWEEN%
y u ON w;
OUTPUT: TECH1 TECH8;

x1 y

Within
x2 u

y
Between

268
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the two-level path analysis model shown in the picture
above is estimated. The mediating variable y is a continuous variable
and the dependent variable u is a binary or ordered categorical variable.
The within part of the model describes the linear regression of y on x1
and x2 and the logistic regression of u on y and x2 where the intercepts
in the two regressions are random effects that vary across the clusters
and the slopes are fixed effects that do not vary across the clusters. In
the within part of the model, the filled circles at the end of the arrows
from x1 to y and x2 to u represent random intercepts that are referred to
as y and u in the between part of the model. In the between part of the
model, the random intercepts are shown in circles because they are
continuous latent variables that vary across clusters. The between part
of the model describes the linear regressions of the random intercepts y
and u on a cluster-level covariate w.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. The program determines the number of
categories of u. The dependent variable u could alternatively be an
unordered categorical (nominal) variable. The NOMINAL option is
used and a multinomial logistic regression is estimated.

In the within part of the model, the first ON statement describes the
linear regression of y on the individual-level covariates x1 and x2 and
the second ON statement describes the logistic regression of u on the
mediating variable y and the individual-level covariate x2. The slopes in
these regressions are fixed effects that do not vary across the clusters.
The residual variance in the linear regression of y on x1 and x2 is
estimated as the default. There is no residual variance to be estimated in
the logistic regression of u on y and x2 because u is a binary or ordered
categorical variable. In the between part of the model, the ON statement
describes the linear regressions of the random intercepts y and u on the
cluster-level covariate w. The intercept and residual variance of y and u
are estimated as the default. The residual covariance between y and u is
free to be estimated as the default.

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, two dimensions of

269
CHAPTER 9

integration are used with a total of 225 integration points. The

ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. The OUTPUT command is used to request
additional output not included as the default. The TECH1 option is used
to request the arrays containing parameter specifications and starting
values for all free parameters in the model. The TECH8 option is used
to request that the optimization history in estimating the model be
printed in the output. TECH8 is printed to the screen during the
computations as the default. TECH8 screen printing is useful for
determining how long the analysis takes. An explanation of the other
commands can be found in Example 9.1.

EXAMPLE 9.4: TWO-LEVEL PATH ANALYSIS WITH A

CONTINUOUS, A CATEGORICAL, AND A CLUSTER-LEVEL
OBSERVED DEPENDENT VARIABLE

TITLE: this is an example of a two-level path

analysis with a continuous, a categorical,
and a cluster-level observed dependent
variable
DATA: FILE = ex9.4.dat;
VARIABLE: NAMES ARE u z y x w clus;
CATEGORICAL = u;
WITHIN = x;
BETWEEN = w z;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
ESTIMATOR = WLSM;
MODEL:
%WITHIN%
u ON y x;
y ON x;
%BETWEEN%
u ON w y z;
y ON w;
z ON w;
y WITH z;
OUTPUT: TECH1;

270
 Examples: Multilevel Modeling With Complex Survey Data

y
Within

x u

y Between

w u

The difference between this example and Example 9.3 is that the
between part of the model has an observed cluster-level mediating
variable z and a latent mediating variable y that is a random intercept.
The model is estimated using weighted least squares estimation instead
of maximum likelihood.

By specifying ESTIMATOR=WLSM, a robust weighted least squares

estimator using a diagonal weight matrix is used (Asparouhov &
Muthén, 2007). The ESTIMATOR option of the ANALYSIS command
can be used to select a different estimator.

In the between part of the model, the first ON statement describes the
linear regression of the random intercept u on the cluster-level covariate
w, the random intercept y, and the observed cluster-level mediating
variable z. The third ON statement describes the linear regression of the
observed cluster-level mediating variable z on the cluster-level covariate
w. An explanation of the other commands can be found in Examples 9.1
and 9.3.

271
CHAPTER 9

EXAMPLE 9.5: TWO-LEVEL PATH ANALYSIS WITH

CONTINUOUS DEPENDENT VARIABLES AND RANDOM
SLOPES

TITLE: this is an example of two-level path

analysis with continuous dependent
variables and random slopes
DATA: FILE IS ex9.5.dat;
VARIABLE: NAMES ARE y1 y2 x1 x2 w clus;
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER IS clus;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL:
%WITHIN%
s2 | y2 ON y1;
y2 ON x2;
s1 | y1 ON x2;
y1 ON x1;
%BETWEEN%
y1 y2 s1 s2 ON w;
OUTPUT: TECH1 TECH8;

272
 Examples: Multilevel Modeling With Complex Survey Data

x1 y1

s1 s2

Within
x2 y2

y2
Between
s2
w
y1

s1

The difference between this example and Example 9.3 is that the model
includes two random intercepts and two random slopes instead of two
random intercepts and two fixed slopes and the dependent variable is
continuous. In the within part of the model, the filled circle on the arrow
from the covariate x2 to the mediating variable y1 represents a random
slope and is referred to as s1 in the between part of the model. The filled
circle on the arrow from the mediating variable y1 to the dependent
variable y2 represents a random slope and is referred to as s2 in the
between part of the model. In the between part of the model, the
random slopes s1 and s2 are shown in circles because they are
continuous latent variables that vary across clusters.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random slope variables in the
model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined

273
CHAPTER 9

using the ON option. In the first | statement, the random slope s2 is

defined by the linear regression of the dependent variable y2 on the
mediating variable y1. In the second | statement, the random slope s1 is
defined by the linear regression of the mediating variable y1 on the
individual-level covariate x2. The within-level residual variances of y1
and y2 are estimated as the default. The first ON statement describes the
linear regression of the dependent variable y2 on the individual-level
covariate x2. The second ON statement describes the linear regression
of the mediating variable y1 on the individual-level covariate x1.

In the between part of the model, the ON statement describes the linear
regressions of the random intercepts y1 and y2 and the random slopes s1
and s2 on the cluster-level covariate w. The intercepts and residual
variances of y1, y2, s2, and s1 are estimated as the default. The residual
covariances between y1, y2, s2, and s1 are fixed at zero as the default.
This default can be overridden. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 9.1 and 9.3.

EXAMPLE 9.6: TWO-LEVEL CFA WITH CONTINUOUS

FACTOR INDICATORS AND COVARIATES

TITLE: this is an example of a two-level CFA with

continuous factor indicators and
covariates
DATA: FILE IS ex9.6.dat;
VARIABLE: NAMES ARE y1-y4 x1 x2 w clus;
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
fw BY y1-y4;
fw ON x1 x2;
%BETWEEN%
fb BY y1-y4;
y1-y4@0;
fb ON w;

274
 Examples: Multilevel Modeling With Complex Survey Data

y1

x1 y2
fw
x2 y3

y4

Within

Between

y1

y2
w fb
y3

y4

In this example, the two-level CFA model with continuous factor

indicators, a between factor, and covariates shown in the picture above is
estimated. In the within part of the model, the filled circles at the end of
the arrows from the within factor fw to y1, y2, y3, and y4 represent
random intercepts that are referred to as y1, y2, y3, and y4 in the
between part of the model. In the between part of the model, the random
intercepts are shown in circles because they are continuous latent
variables that vary across clusters. They are indicators of the between
factor fb. In this model, the residual variances for the factor indicators
in the between part of the model are fixed at zero. If factor loadings are

275
CHAPTER 9

constrained to be equal across the within and the between levels, this
implies a model where the regression of the within factor on x1 and x2
has a random intercept varying across the clusters.

In the within part of the model, the BY statement specifies that fw is

measured by y1, y2, y3, and y4. The metric of the factor is set
automatically by the program by fixing the first factor loading to one.
This option can be overridden. The residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The ON statement describes the linear regression of fw on the
individual-level covariates x1 and x2. The residual variance of the
factor is estimated as the default. The intercept of the factor is fixed at
zero.

In the between part of the model, the BY statement specifies that fb is

measured by the random intercepts y1, y2, y3, and y4. The metric of the
factor is set automatically by the program by fixing the first factor
loading to one. This option can be overridden. The residual variances
of the factor indicators are set to zero. The ON statement describes the
regression of fb on the cluster-level covariate w. The residual variance
of the factor is estimated as the default. The intercept of the factor is
fixed at zero as the default. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 9.1.

276
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.7: TWO-LEVEL CFA WITH CATEGORICAL

FACTOR INDICATORS AND COVARIATES

TITLE: this is an example of a two-level CFA with

categorical factor indicators and
covariates
DATA: FILE IS ex9.7.dat;
VARIABLE: NAMES ARE u1-u4 x1 x2 w clus;
CATEGORICAL = u1-u4;
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER = clus;
MISSING = ALL (999);
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
fw BY u1-u4;
fw ON x1 x2;
%BETWEEN%
fb BY u1-u4;
fb ON w;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 9.6 is that the factor
indicators are binary or ordered categorical (ordinal) variables instead of
continuous variables. The CATEGORICAL option is used to specify
which dependent variables are treated as binary or ordered categorical
(ordinal) variables in the model and its estimation. In the example
above, all four factor indicators are binary or ordered categorical. The
program determines the number of categories for each indicator. The
default estimator for this type of analysis is maximum likelihood with
robust standard errors using a numerical integration algorithm. Note that
numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase. In this
example, two dimensions of integration are used with a total of 225
integration points. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator.

In the between part of the model, the residual variances of the random
intercepts of the categorical factor indicators are fixed at zero as the
default because the residual variances of random intercepts are often
very small and require one dimension of numerical integration each.
Weighted least squares estimation of between-level residual variances

277
CHAPTER 9

does not require numerical integration in estimating the model. An

explanation of the other commands can be found in Examples 9.1 and
9.6.

EXAMPLE 9.8: TWO-LEVEL CFA WITH CONTINUOUS

FACTOR INDICATORS, COVARIATES, AND RANDOM
SLOPES

TITLE: this is an example of a two-level CFA with

continuous factor indicators, covariates,
and random slopes
DATA: FILE IS ex9.8.dat;
VARIABLE: NAMES ARE y1-y4 x1 x2 w clus;
CLUSTER = clus;
WITHIN = x1 x2;
BETWEEN = w;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL:
%WITHIN%
fw BY y1-y4;
s1 | fw ON x1;
s2 | fw ON x2;
%BETWEEN%
fb BY y1-y4;
y1-y4@0;
fb s1 s2 ON w;

278
 Examples: Multilevel Modeling With Complex Survey Data

y1

x1 s1 y2
s2 fw
x2 y3

y4

Within

Between

y1

y2
w fb
y3
s1
y4
s2

The difference between this example and Example 9.6 is that the model
has random slopes in addition to random intercepts and the random
slopes are regressed on a cluster-level covariate. In the within part of the
model, the filled circles on the arrows from x1 and x2 to fw represent
random slopes that are referred to as s1 and s2 in the between part of the
model. In the between part of the model, the random slopes are shown
in circles because they are latent variables that vary across clusters.

279
CHAPTER 9

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random slope variables in the
model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined
using the ON option. In the first | statement, the random slope s1 is
defined by the linear regression of the factor fw on the individual-level
covariate x1. In the second | statement, the random slope s2 is defined
by the linear regression of the factor fw on the individual-level covariate
x2. The within-level residual variance of f1 is estimated as the default.

In the between part of the model, the ON statement describes the linear
regressions of fb, s1, and s2 on the cluster-level covariate w. The
residual variances of fb, s1, and s2 are estimated as the default. The
residuals are not correlated as the default. The default estimator for this
type of analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 9.1 and 9.6.

280
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.9: TWO-LEVEL SEM WITH CATEGORICAL

FACTOR INDICATORS ON THE WITHIN LEVEL AND
CLUSTER-LEVEL CONTINUOUS OBSERVED AND RANDOM
INTERCEPT FACTOR INDICATORS ON THE BETWEEN
LEVEL

TITLE: this is an example of a two-level SEM with

categorical factor indicators on the
within level and cluster-level continuous
observed and random intercept factor
indicators on the between level
DATA: FILE IS ex9.9.dat;
VARIABLE: NAMES ARE u1-u6 y1-y4 x1 x2 w clus;
CATEGORICAL = u1-u6;
WITHIN = x1 x2;
BETWEEN = w y1-y4;
CLUSTER IS clus;
ANALYSIS: TYPE IS TWOLEVEL;
ESTIMATOR = WLSMV;
MODEL:
%WITHIN%
fw1 BY u1-u3;
fw2 BY u4-u6;
fw1 fw2 ON x1 x2;
%BETWEEN%
fb BY u1-u6;
f BY y1-y4;
fb ON w f;
f ON w;
SAVEDATA: SWMATRIX = ex9.9sw.dat;

281
CHAPTER 9

u1

x1 fw1 u2

u3

u4

x2 fw2 u5

u6

Within

Between

y1 y2 y3 y4

u1

u2
f
u3
w fb
u4

u5

u6

282
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the model with two within factors and two between
factors shown in the picture above is estimated. The within-level factor
indicators are categorical. In the within part of the model, the filled
circles at the end of the arrows from the within factor fw1 to u1, u2, and
u3 and fw2 to u4, u5, and u6 represent random intercepts that are
referred to as u1, u2, u3, u4, u5, and u6 in the between part of the model.
In the between part of the model, the random intercepts are shown in
circles because they are continuous latent variables that vary across
clusters. The random intercepts are indicators of the between factor fb.
This example illustrates the common finding of fewer between factors
than within factors for the same set of factor indicators. The between
factor f has observed cluster-level continuous variables as factor
indicators.

By specifying ESTIMATOR=WLSMV, a robust weighted least squares

estimator using a diagonal weight matrix will be used. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors using a numerical integration algorithm. Note that
numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase. In this
example, three dimensions of integration would be used with a total of
3,375 integration points. For models with many dimensions of
integration and categorical outcomes, the weighted least squares
estimator may improve computational speed. The ESTIMATOR option
of the ANALYSIS command can be used to select a different estimator.

In the within part of the model, the first BY statement specifies that fw1
is measured by u1, u2, and u3. The second BY statement specifies that
fw2 is measured by u4, u5, and u6. The metric of the factors are set
automatically by the program by fixing the first factor loading for each
factor to one. This option can be overridden. Residual variances of the
latent response variables of the categorical factor indicators are not
parameters in the model. They are fixed at one in line with the Theta
parameterization. Residuals are not correlated as the default. The ON
statement describes the linear regressions of fw1 and fw2 on the
individual-level covariates x1 and x2. The residual variances of the
factors are estimated as the default. The residuals of the factors are
correlated as the default because residuals are correlated for latent
variables that do not influence any other variable in the model except
their own indicators. The intercepts of the factors are fixed at zero as
the default.

283
CHAPTER 9

In the between part of the model, the first BY statement specifies that fb
is measured by the random intercepts u1, u2, u3, u4, u5, and u6. The
metric of the factor is set automatically by the program by fixing the first
factor loading to one. This option can be overridden. The residual
variances of the factor indicators are estimated and the residuals are not
correlated as the default. Unlike maximum likelihood estimation,
weighted least squares estimation of between-level residual variances
does not require numerical integration in estimating the model. The
second BY statement specifies that f is measured by the cluster-level
factor indicators y1, y2, y3, and y4. The residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The first ON statement describes the linear regression of fb on
the cluster-level covariate w and the factor f. The second ON statement
describes the linear regression of f on the cluster-level covariate w. The
residual variances of the factors are estimated as the default. The
intercepts of the factors are fixed at zero as the default.

The SWMATRIX option of the SAVEDATA command is used with

TYPE=TWOLEVEL and weighted least squares estimation to specify
the name and location of the file that contains the within- and between-
level sample statistics and their corresponding estimated asymptotic
covariance matrix. It is recommended to save this information and use it
in subsequent analyses along with the raw data to reduce computational
time during model estimation. An explanation of the other commands
can be found in Example 9.1.

284
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.10: TWO-LEVEL SEM WITH CONTINUOUS

FACTOR INDICATORS AND A RANDOM SLOPE FOR A
FACTOR

TITLE: this is an example of a two-level SEM with

continuous factor indicators and a random
slope for a factor
DATA: FILE IS ex9.10.dat;
VARIABLE: NAMES ARE y1-y5 w clus;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
ALGORITHM = INTEGRATION;
INTEGRATION = 10;
MODEL:
%WITHIN%
fw BY y1-y4;
s | y5 ON fw;
%BETWEEN%
fb BY y1-y4;
y1-y4@0;
y5 s ON fb w;
OUTPUT: TECH1 TECH8;

285
CHAPTER 9

y1

y2
s
fw y5
y3

y4 Within

y1 Between

y2
fb y5
y3

y4

w s

In this example, the two-level SEM with continuous factor indicators

shown in the picture above is estimated. In the within part of the model,
the filled circles at the end of the arrows from fw to the factor indicators
y1, y2, y3, and y4 and the filled circle at the end of the arrow from fw to
y5 represent random intercepts that are referred to as y1, y2, y3, y4, and
y5 in the between part of the model. The filled circle on the arrow from
fw to y5 represents a random slope that is referred to as s in the between

286
 Examples: Multilevel Modeling With Complex Survey Data

part of the model. In the between part of the model, the random
intercepts and random slope are shown in circles because they are
continuous latent variables that vary across clusters.

By specifying TYPE=TWOLEVEL RANDOM in the ANALYSIS

command, a multilevel model with random intercepts and random slopes
will be estimated. By specifying ALGORITHM=INTEGRATION, a
maximum likelihood estimator with robust standard errors using a
numerical integration algorithm will be used. Note that numerical
integration becomes increasingly more computationally demanding as
the number of factors and the sample size increase. In this example, four
dimensions of integration are used with a total of 10,000 integration
points. The INTEGRATION option of the ANALYSIS command is
used to change the number of integration points per dimension from the
default of 15 to 10. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator.

In the within part of the model, the BY statement specifies that fw is

measured by the factor indicators y1, y2, y3, and y4. The metric of the
factor is set automatically by the program by fixing the first factor
loading in each BY statement to one. This option can be overridden.
The residual variances of the factor indicators are estimated and the
residuals are uncorrelated as the default. The variance of the factor is
estimated as the default.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random slope variables in the
model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined
using the ON option. In the | statement, the random slope s is defined by
the linear regression of the dependent variable y5 on the within factor
fw. The within-level residual variance of y5 is estimated as the default.

In the between part of the model, the BY statement specifies that fb is

measured by the random intercepts y1, y2, y3, and y4. The metric of the
factor is set automatically by the program by fixing the first factor
loading in the BY statement to one. This option can be overridden. The
residual variances of the factor indicators are fixed at zero. The variance
of the factor is estimated as the default. The ON statement describes the
linear regressions of the random intercept y5 and the random slope s on

287
CHAPTER 9

the factor fb and the cluster-level covariate w. The intercepts and

residual variances of y5 and s are estimated and their residuals are
uncorrelated as the default.

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes. An explanation of the other commands can be found in Example
9.1.

EXAMPLE 9.11: TWO-LEVEL MULTIPLE GROUP CFA WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a two-level

multiple group CFA with continuous
factor indicators
DATA: FILE IS ex9.11.dat;
VARIABLE: NAMES ARE y1-y6 g clus;
GROUPING = g (1 = g1 2 = g2);
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
fw1 BY y1-y3;
fw2 BY y4-y6;
%BETWEEN%
fb1 BY y1-y3;
fb2 BY y4-y6;
MODEL g2: %WITHIN%
fw1 BY y2-y3;
fw2 BY y5-y6;

288
 Examples: Multilevel Modeling With Complex Survey Data

y1 y2 y3 y4 y5 y6

fw1 fw2 Within

y1 y2 y3 y4 y5 y6
Between

fb1 fb2

In this example, the two-level multiple group CFA with continuous

factor indicators shown in the picture above is estimated. In the within
part of the model, the filled circles at the end of the arrows from the
within factors fw1 to y1, y2, and y3 and fw2 to y4, y5, and y6 represent
random intercepts that are referred to as y1, y2, y3, y4, y5, and y6 in the
between part of the model. In the between part of the model, the random
intercepts are shown in circles because they are continuous latent
variables that vary across clusters. The random intercepts are indicators
of the between factors fb1 and fb2.

The GROUPING option of the VARIABLE command is used to identify

the variable in the data set that contains information on group
membership when the data for all groups are stored in a single data set.
The information in parentheses after the grouping variable name assigns
labels to the values of the grouping variable found in the data set. In the
example above, observations with g equal to 1 are assigned the label g1,

289
CHAPTER 9

and individuals with g equal to 2 are assigned the label g2. These labels
are used in conjunction with the MODEL command to specify model
statements specific to each group. The grouping variable should be a
cluster-level variable.

In multiple group analysis, two variations of the MODEL command are

used. They are MODEL and MODEL followed by a label. MODEL
describes the model to be estimated for all groups. The factor loadings
and intercepts are held equal across groups as the default to specify
measurement invariance. MODEL followed by a label describes
differences between the overall model and the model for the group
designated by the label.

In the within part of the model, the BY statements specify that fw1 is
measured by y1, y2, and y3, and fw2 is measured by y4, y5, and y6. The
metric of the factors is set automatically by the program by fixing the
first factor loading in each BY statement to one. This option can be
overridden. The variances of the factors are estimated as the default.
The factors fw1 and fw2 are correlated as the default because they are
independent (exogenous) variables. In the between part of the model,
the BY statements specify that fb1 is measured by y1, y2, and y3, and
fb2 is measured by y4, y5, and y6. The metric of the factor is set
automatically by the program by fixing the first factor loading in each
BY statement to one. This option can be overridden. The variances of
the factors are estimated as the default. The factors fb1 and fb2 are
correlated as the default because they are independent (exogenous)
variables.

In the group-specific MODEL command for group 2, by specifying the

within factor loadings for fw1 and fw2, the default equality constraints
are relaxed and the factor loadings are no longer held equal across
groups. The factor indicators that are fixed at one remain the same, in
this case y1 and y4. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Example 9.1.

290
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.12: TWO-LEVEL GROWTH MODEL FOR A

CONTINUOUS OUTCOME (THREE-LEVEL ANALYSIS)

TITLE: this is an example of a two-level growth

model for a continuous outcome (three-
level analysis)
DATA: FILE IS ex9.12.dat;
VARIABLE: NAMES ARE y1-y4 x w clus;
WITHIN = x;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4 (1);
iw sw ON x;
%BETWEEN%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib sb ON w;

291
CHAPTER 9

y1 y2 y3 y4

iw sw

x
Within

Between

y1 y2 y3 y4

ib sb

292
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the two-level growth model for a continuous outcome

(three-level analysis) shown in the picture above is estimated. In the
within part of the model, the filled circles at the end of the arrows from
the within growth factors iw and sw to y1, y2, y3, and y4 represent
random intercepts that are referred to as y1, y2, y3, and y4 in the
between part of the model. In the between part of the model, the random
intercepts are shown in circles because they are continuous latent
variables that vary across clusters.

In the within part of the model, the | statement names and defines the
within intercept and slope factors for the growth model. The names iw
and sw on the left-hand side of the | symbol are the names of the
intercept and slope growth factors, respectively. The values on the right-
hand side of the | symbol are the time scores for the slope growth factor.
The time scores of the slope growth factor are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
growth factor as an initial status factor. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are
constrained to be equal over time in line with conventional multilevel
growth modeling. This is done by placing (1) after them. The residual
covariances of the outcome variables are fixed at zero as the default.
Both of these restrictions can be overridden. The ON statement
describes the linear regressions of the growth factors on the individual-
level covariate x. The residual variances of the growth factors are free
to be estimated as the default. The residuals of the growth factors are
correlated as the default because residuals are correlated for latent
variables that do not influence any other variable in the model except
their own indicators.

In the between part of the model, the | statement names and defines the
between intercept and slope factors for the growth model. The names ib
and sb on the left-hand side of the | symbol are the names of the intercept
and slope growth factors, respectively. The values on the right-hand side
of the | symbol are the time scores for the slope growth factor. The time
scores of the slope growth factor are fixed at 0, 1, 2, and 3 to define a
linear growth model with equidistant time points. The zero time score
for the slope growth factor at time point one defines the intercept factor
as an initial status factor. The coefficients of the intercept growth factor
are fixed at one as part of the growth model parameterization. The

293
CHAPTER 9

residual variances of the outcome variables are fixed at zero on the

between level in line with conventional multilevel growth modeling.
These residual variances can be estimated. The ON statement describes
the linear regressions of the growth factors on the cluster-level covariate
w. The residual variances and the residual covariance of the growth
factors are free to be estimated as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The intercepts of the growth factors are estimated as the default
in the between part of the model. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Example 9.1.

EXAMPLE 9.13: TWO-LEVEL GROWTH MODEL FOR A

CATEGORICAL OUTCOME (THREE-LEVEL ANALYSIS)

TITLE: this is an example of a two-level

growth model for a categorical outcome
(three-level analysis)
DATA: FILE IS ex9.13.dat;
VARIABLE: NAMES ARE u1-u4 x w clus;
CATEGORICAL = u1-u4;
WITHIN = x;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
INTEGRATION = 7;
MODEL:
%WITHIN%
iw sw | u1@0 u2@1 u3@2 u4@3;
iw sw ON x;
%BETWEEN%
ib sb | u1@0 u2@1 u3@2 u4@3;
ib sb ON w;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 9.12 is that the
outcome variable is a binary or ordered categorical (ordinal) variable
instead of a continuous variable.

294
 Examples: Multilevel Modeling With Complex Survey Data

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, u1, u2, u3, and u4
are binary or ordered categorical variables. They represent the outcome
measured at four equidistant occasions.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, four dimensions of integration are used with a
total of 2,401 integration points. The INTEGRATION option of the
ANALYSIS command is used to change the number of integration points
per dimension from the default of 15 to 7. The ESTIMATOR option of
the ANALYSIS command can be used to select a different estimator.
For models with many dimensions of integration and categorical
outcomes, the weighted least squares estimator may improve
computational speed.

In the parameterization of the growth model shown here, the thresholds

of the outcome variable at the four time points are held equal as the
default and are estimated in the between part of the model. The
intercept of the intercept growth factor is fixed at zero. The intercept of
the slope growth factor is estimated as the default in the between part of
the model. The residual variances of the growth factors are estimated as
the default. The residuals of the growth factors are correlated as the
default because residuals are correlated for latent variables that do not
influence any other variable in the model except their own indicators.
On the between level, the residual variances of the random intercepts u1,
u2, u3, and u4 are fixed at zero as the default.

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes. An explanation of the other commands can be found in Examples
9.1 and 9.12.

295
CHAPTER 9

EXAMPLE 9.14: TWO-LEVEL GROWTH MODEL FOR A

CONTINUOUS OUTCOME (THREE-LEVEL ANALYSIS) WITH
VARIATION ON BOTH THE WITHIN AND BETWEEN
LEVELS FOR A RANDOM SLOPE OF A TIME-VARYING
COVARIATE

TITLE: this is an example of a two-level growth

model for a continuous outcome (three-
level analysis) with variation on both the
within and between levels for a random
slope of a time-varying covariate
DATA: FILE IS ex9.14.dat;
VARIABLE: NAMES ARE y1-y4 x a1-a4 w clus;
WITHIN = x a1-a4;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
ALGORITHM = INTEGRATION;
INTEGRATION = 10;
MODEL:
%WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4 (1);
iw sw ON x;
s* | y1 ON a1;
s* | y2 ON a2;
s* | y3 ON a3;
s* | y4 ON a4;
%BETWEEN%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib sb s ON w;
OUTPUT: TECH1 TECH8;

296
 Examples: Multilevel Modeling With Complex Survey Data

y1 y2 y3 y4

iw

sw

x a1 a2 a3 a4
Within

Between
y1 y2 y3 y4

ib

sb

w s

The difference between this example and Example 9.12 is that the model
includes an individual-level time-varying covariate with a random slope
that varies on both the within and between levels. In the within part of
the model, the filled circles at the end of the arrows from a1 to y1, a2 to
y2, a3 to y3, and a4 to y4 represent random intercepts that are referred to

297
CHAPTER 9

as y1, y2, y3, and y4 in the between part of the model. In the between
part of the model, the random intercepts are shown in circles because
they are continuous latent variables that vary across classes. The broken
arrows from s to the arrows from a1 to y1, a2 to y2, a3 to y3, and a4 to
y4 indicate that the slopes in these regressions are random. The s is
shown in a circle in both the within and between parts of the model to
represent a decomposition of the random slope into its within and
between components.

By specifying TYPE=TWOLEVEL RANDOM in the ANALYSIS

command, a multilevel model with random intercepts and random slopes
will be estimated. By specifying ALGORITHM=INTEGRATION, a
maximum likelihood estimator with robust standard errors using a
numerical integration algorithm will be used. Note that numerical
integration becomes increasingly more computationally demanding as
the number of factors and the sample size increase. In this example, four
dimensions of integration are used with a total of 10,000 integration
points. The INTEGRATION option of the ANALYSIS command is
used to change the number of integration points per dimension from the
default of 15 to 10. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator.

The | symbol is used in conjunction with TYPE=RANDOM to name and

define the random slope variables in the model. The name on the left-
hand side of the | symbol names the random slope variable. The
statement on the right-hand side of the | symbol defines the random slope
variable. The random slope s is defined by the linear regressions of y1
on a1, y2 on a2, y3 on a3, and y4 on a4. Random slopes with the same
name are treated as one variable during model estimation. The random
intercepts for these regressions are referred to by using the name of the
dependent variables in the regressions, that is, y1, y2, y3, and y4. The
asterisk (*) following the s specifies that s will have variation on both
the within and between levels. Without the asterisk (*), s would have
variation on only the between level. An explanation of the other
commands can be found in Examples 9.1 and 9.12.

298
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.15: TWO-LEVEL MULTIPLE INDICATOR

GROWTH MODEL WITH CATEGORICAL OUTCOMES
(THREE-LEVEL ANALYSIS)

TITLE: this is an example of a two-level multiple

indicator growth model with categorical
outcomes (three-level analysis)
DATA: FILE IS ex9.15.dat;
VARIABLE: NAMES ARE u11 u21 u31 u12 u22 u32 u13 u23
u33 clus;
CATEGORICAL = u11-u33;
CLUSTER = clus;
ANALYSIS: TYPE IS TWOLEVEL;
ESTIMATOR = WLSM;
MODEL:
%WITHIN%
f1w BY u11
u21-u31 (1-2);
f2w BY u12
u22-u32 (1-2);
f3w BY u13
u23-u33 (1-2);
iw sw | f1w@0 f2w@1 f3w@2;
%BETWEEN%
f1b BY u11
u21-u31 (1-2);
f2b BY u12
u22-u32 (1-2);
f3b BY u13
u23-u33 (1-2);
[u11$1 u12$1 u13$1] (3);
[u21$1 u22$1 u23$1] (4);
[u31$1 u32$1 u33$1] (5);
ib sb | f1b@0 f2b@1 f3b@2;
[f1b-f3b@0 ib@0 sb];
f1b-f3b (6);
SAVEDATA: SWMATRIX = ex9.15sw.dat;

299
CHAPTER 9

u11 u21 u31 u12 u22 u32 u13 u23 u33

f1w f2w f3w

Within
iw sw

u11 u21 u31 u12 u22 u32 u13 u23 u33

f1b f2b f3b

Between
ib sb

In this example, the two-level multiple indicator growth model with

categorical outcomes (three-level analysis) shown in the picture above is
estimated. The picture shows a factor measured by three indicators at
three time points. In the within part of the model, the filled circles at the
end of the arrows from the within factors f1w to u11, u21, and u31; f2w
to u12, u22, and u32; and f3w to u13, u23, and u33 represent random
intercepts that are referred to as u11, u21, u31, u12, u22, u32, u13, u23,
and u33 in the between part of the model. In the between part of the
model, the random intercepts are continuous latent variables that vary
across clusters. The random intercepts are indicators of the between
factors f1b, f2b, and f3b. In this model, the residual variances of the

300
 Examples: Multilevel Modeling With Complex Survey Data

factor indicators in the between part of the model are estimated. The
residuals are not correlated as the default. Taken together with the
specification of equal factor loadings on the within and the between
parts of the model, this implies a model where the regressions of the
within factors on the growth factors have random intercepts that vary
across the clusters.

By specifying ESTIMATOR=WLSM, a robust weighted least squares

estimator using a diagonal weight matrix will be used. The default
estimator for this type of analysis is maximum likelihood with robust
standard errors using a numerical integration algorithm. Note that
numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase. For
models with many dimensions of integration and categorical outcomes,
the weighted least squares estimator may improve computational speed.

In the within part of the model, the three BY statements define a within-
level factor at three time points. The metric of the three factors is set
automatically by the program by fixing the first factor loading to one.
This option can be overridden. The (1-2) following the factor loadings
uses the list function to assign equality labels to these parameters. The
label 1 is assigned to the factor loadings of u21, u22, and u23 which
holds these factor loadings equal across time. The label 2 is assigned to
the factor loadings of u31, u32, and u33 which holds these factor
loadings equal across time. Residual variances of the latent response
variables of the categorical factor indicators are not free parameters to be
estimated in the model. They are fixed at one in line with the Theta
parameterization. Residuals are not correlated as the default. The |
statement names and defines the within intercept and slope growth
factors for the growth model. The names iw and sw on the left-hand side
of the | symbol are the names of the intercept and slope growth factors,
respectively. The names and values on the right-hand side of the |
symbol are the outcome and time scores for the slope growth factor. The
time scores of the slope growth factor are fixed at 0, 1, and 2 to define a
linear growth model with equidistant time points. The zero time score
for the slope growth factor at time point one defines the intercept growth
factor as an initial status factor. The coefficients of the intercept growth
factor are fixed at one as part of the growth model parameterization.
The variances of the growth factors are free to be estimated as the
default. The covariance between the growth factors is free to be
estimated as the default. The intercepts of the factors defined using BY

301
CHAPTER 9

statements are fixed at zero. The residual variances of the factors are
free and not held equal across time. The residuals of the factors are
uncorrelated in line with the default of residuals for first-order factors.

In the between part of the model, the first three BY statements define a
between-level factor at three time points. The (1-2) following the factor
loadings uses the list function to assign equality labels to these
parameters. The label 1 is assigned to the factor loadings of u21, u22,
and u23 which holds these factor loadings equal across time as well as
across levels. The label 2 is assigned to the factor loadings of u31, u32,
and u33 which holds these factor loadings equal across time as well as
across levels. Time-invariant thresholds for the three indicators are
specified using (3), (4), and (5) following the bracket statements. The
residual variances of the factor indicators are free to be estimated. The |
statement names and defines the between intercept and slope growth
factors for the growth model. The names ib and sb on the left-hand side
of the | symbol are the names of the intercept and slope growth factors,
respectively. The values on the right-hand side of the | symbol are the
time scores for the slope growth factor. The time scores of the slope
growth factor are fixed at 0, 1, and 2 to define a linear growth model
with equidistant time points. The zero time score for the slope growth
factor at time point one defines the intercept growth factor as an initial
status factor. The coefficients of the intercept growth factor are fixed at
one as part of the growth model parameterization. In the
parameterization of the growth model shown here, the intercept growth
factor mean is fixed at zero as the default for identification purposes.
The variances of the growth factors are free to be estimated as the
default. The covariance between the growth factors is free to be
estimated as the default. The intercepts of the factors defined using BY
statements are fixed at zero. The residual variances of the factors are
held equal across time. The residuals of the factors are uncorrelated in
line with the default of residuals for first-order factors.

The SWMATRIX option of the SAVEDATA command is used with

TYPE=TWOLEVEL and weighted least squares estimation to specify
the name and location of the file that contains the within- and between-
level sample statistics and their corresponding estimated asymptotic
covariance matrix. It is recommended to save this information and use it
in subsequent analyses along with the raw data to reduce computational
time during model estimation. An explanation of the other commands
can be found in Example 9.1

302
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.16: LINEAR GROWTH MODEL FOR A

CONTINUOUS OUTCOME WITH TIME-INVARIANT AND
TIME-VARYING COVARIATES CARRIED OUT AS A TWO-
LEVEL GROWTH MODEL USING THE DATA WIDETOLONG
COMMAND

TITLE: this is an example of a linear growth

model for a continuous outcome with time-
invariant and time-varying covariates
carried out as a two-level growth model
using the DATA WIDETOLONG command
DATA: FILE IS ex9.16.dat;
DATA WIDETOLONG:
WIDE = y11-y14 | a31-a34;
LONG = y | a3;
IDVARIABLE = person;
REPETITION = time;
VARIABLE: NAMES ARE y11-y14 x1 x2 a31-a34;
USEVARIABLE = x1 x2 y a3 person time;
CLUSTER = person;
WITHIN = time a3;
BETWEEN = x1 x2;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL: %WITHIN%
s | y ON time;
y ON a3;
%BETWEEN%
y s ON x1 x2;
y WITH s;

303
CHAPTER 9

s
time y

a3
Within

Between
x1 y

x2 s

In this example, a linear growth model for a continuous outcome with

time-invariant and time-varying covariates as shown in the picture above
is estimated. As part of the analysis, the DATA WIDETOLONG
command is used to rearrange the data from a multivariate wide format
to a univariate long format. The model is similar to the one in Example
6.10 using multivariate wide format data. The differences are that the
current model restricts the within-level residual variances to be equal
across time and the within-level influence of the time-varying covariate
on the outcome to be equal across time.

The WIDE option of the DATA WIDETOLONG command is used to

identify sets of variables in the wide format data set that are to be
converted into single variables in the long format data set. These
variables must variables from the NAMES statement of the VARIABLE
command. The two sets of variables y11, y12, y13, and y14 and a31,
a32, a33, and a34 are identified. The LONG option is used to provide
names for the new variables in the long format data set. The names y
and a3 are the names of the new variables. The IDVARIABLE option is
used to provide a name for the variable that provides information about
the unit to which the record belongs. In univariate growth modeling, this
is the person identifier which is used as a cluster variable. In this
example, the name person is used. This option is not required. The

304
 Examples: Multilevel Modeling With Complex Survey Data

default variable name is id. The REPETITION option is used to provide

a name for the variable that contains information on the order in which
the variables were measured. In this example, the name time is used.
This option is not required. The default variable name is rep. The new
variables must be mentioned on the USEVARIABLE statement of the
VARIABLE command if they are used in the analysis. They must be
placed after any original variables. The USEVARIABLES option lists
the original variables x1 and x2 followed by the new variables y, a3,
person, and time.

The CLUSTER option of the VARIABLE command is used to identify

the variable that contains clustering information. In this example, the
cluster variable person is the variable that was created using the
IDVARIABLE option of the DATA WIDETOLONG command. The
WITHIN option is used to identify the variables in the data set that are
measured on the individual level and modeled only on the within level.
They are specified to have no variance in the between part of the model.
The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level and modeled only on the between
level. Variables not mentioned on the WITHIN or the BETWEEN
statements are measured on the individual level and can be modeled on
both the within and between levels.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random slope variables in the
model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined
using the ON option. In the | statement, the random slope s is defined by
the linear regression of the dependent variable y on time. The within-
level residual variance of y is estimated as the default. The ON
statement describes the linear regression of y on the covariate a3.

In the between part of the model, the ON statement describes the linear
regressions of the random intercept y and the random slope s on the
covariates x1 and x2. The WITH statement is used to free the
covariance between y and s. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
estimator option of the ANALYSIS command can be used to select a
different estimator. An explanation of the other commands can be found
in Example 9.1.

305
CHAPTER 9

EXAMPLE 9.17: TWO-LEVEL GROWTH MODEL FOR A

COUNT OUTCOME USING A ZERO-INFLATED POISSON
MODEL (THREE-LEVEL ANALYSIS)

TITLE: this is an example of a two-level growth

model for a count outcome using a zero-
inflated Poisson model (three-level
analysis)
DATA: FILE = ex9.17.dat;
VARIABLE: NAMES = u1-u4 x w clus;
COUNT = u1-u4 (i);
CLUSTER = clus;
WITHIN = x;
BETWEEN = w;
ANALYSIS: TYPE = TWOLEVEL;
ALGORITHM = INTEGRATION;
INTEGRATION = 10;
MCONVERGENCE = 0.01;
MODEL: %WITHIN%
iw sw | u1@0 u2@1 u3@2 u4@3;
iiw siw | u1#1@0 u2#1@1 u3#1@2 u4#1@3;
sw@0;
siw@0;
iw WITH iiw;
iw ON x;
sw ON x;
%BETWEEN%
ib sb | u1@0 u2@1 u3@2 u4@3;
iib sib | u1#1@0 u2#1@1 u3#1@2 u4#1@3;
sb-sib@0;
ib ON w;
OUTPUT: TECH1 TECH8;

The difference between this example and Example 9.12 is that the
outcome variable is a count variable instead of a continuous variable.

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and whether a
Poisson or zero-inflated Poisson model will be estimated. In the
example above, u1, u2, u3, and u4 are count variables. The i in
parentheses following u indicates that a zero-inflated Poisson model will
be estimated.

306
 Examples: Multilevel Modeling With Complex Survey Data

By specifying ALGORITHM=INTEGRATION, a maximum likelihood

estimator with robust standard errors using a numerical integration
algorithm will be used. Note that numerical integration becomes
increasingly more computationally demanding as the number of factors
and the sample size increase. In this example, three dimensions of
integration are used with a total of 1,000 integration points. The
INTEGRATION option of the ANALYSIS command is used to change
the number of integration points per dimension from the default of 15 to
10. The ESTIMATOR option of the ANALYSIS command can be used
to select a different estimator. The MCONVERGENCE option is used
to change the observed-data log likelihood derivative convergence
criterion for the EM algorithm from the default value of .001 to .01
because it is difficult to obtain high numerical precision in this example.

With a zero-inflated Poisson model, two growth models are estimated.

In the within and between parts of the model, the first | statement
describes the growth model for the count part of the outcome for
individuals who are able to assume values of zero and above. The
second | statement describes the growth model for the inflation part of
the outcome, the probability of being unable to assume any value except
zero. The binary latent inflation variable is referred to by adding to the
name of the count variable the number sign (#) followed by the number
1. In the parameterization of the growth model for the count part of the
outcome, the intercepts of the outcome variables at the four time points
are fixed at zero as the default. In the parameterization of the growth
model for the inflation part of the outcome, the intercepts of the outcome
variable at the four time points are held equal as the default. In the
within part of the model, the variances of the growth factors are
estimated as the default, and the growth factor covariances are fixed at
zero as the default. In the between part of the model, the mean of the
growth factors for the count part of outcome are free. The mean of the
intercept growth factor for the inflation part of the outcome is fixed at
zero and the mean for the slope growth factor for the inflation part of the
outcome is free. The variances of the growth factors are estimated as the
default, and the growth factor covariances are fixed at zero as the
default.

In the within part of the model, the variances of the slope growth factors
sw and siw are fixed at zero. The ON statements describes the linear
regressions of the intercept and slope growth factors iw and sw for the
count part of the outcome on the covariate x. In the between part of the

307
CHAPTER 9

model, the variances of the intercept growth factor iib and the slope
growth factors sb and sib are fixed at zero. The ON statement describes
the linear regression of the intercept growth factor ib on the covariate w.
An explanation of the other commands can be found in Examples 9.1
and 9.12.

EXAMPLE 9.18: TWO-LEVEL CONTINUOUS-TIME

SURVIVAL ANALYSIS USING COX REGRESSION WITH A
RANDOM INTERCEPT

TITLE: this is an example of a two-level

continuous-time survival analysis using
Cox regression with a random intercept
DATA: FILE = ex9.18.dat;
VARIABLE: NAMES = t x w tc clus;
CLUSTER = clus;
WITHIN = x;
BETWEEN = w;
SURVIVAL = t (ALL);
TIMECENSORED = tc (0 = NOT 1 = RIGHT);
ANALYSIS: TYPE = TWOLEVEL;
BASEHAZARD = OFF;
MODEL: %WITHIN%
t ON x;
%BETWEEN%
t ON w;
t;

x t

Within

Between

w t

308
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the two-level continuous-time survival analysis model

shown in the picture above is estimated. This is the Cox regression
model with a random intercept (Klein & Moeschberger, 1997; Hougaard,
2000). The profile likelihood method is used for estimation
(Asparouhov et al., 2006).

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option must be used in conjunction with the
TIMECENSORED option. In this example, t is the variable that
contains time to event information. By specifying the keyword ALL in
parenthesis following the time-to-event variable, the time intervals are
taken from the data. The TIMECENSORED option is used to identify
the variables that contain information about right censoring. In this
example, this variable is named tc. The information in parentheses
specifies that the value zero represents no censoring and the value one
represents right censoring. This is the default. The BASEHAZARD
option of the ANALYSIS command is used with continuous-time
survival analysis to specify if a non-parametric or a parametric baseline
hazard function is used in the estimation of the model. The setting OFF
specifies that a non-parametric baseline hazard function is used. This is
the default.

The MODEL command is used to describe the model to be estimated. In

multilevel models, a model is specified for both the within and between
parts of the model. In the within part of the model, the loglinear
regression of the time-to-event t on the covariate x is specified. In the
between part of the model, the linear regression of the random intercept t
on the cluster-level covariate w is specified. The residual variance of t is
estimated. The default estimator for this type of analysis is maximum
likelihood with robust standard errors. The estimator option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 9.1.

309
CHAPTER 9

EXAMPLE 9.19: TWO-LEVEL MIMIC MODEL WITH

CONTINUOUS FACTOR INDICATORS, RANDOM FACTOR
LOADINGS, TWO COVARIATES ON WITHIN, AND ONE
COVARIATE ON BETWEEN WITH EQUAL LOADINGS
ACROSS LEVELS

TITLE: this is an example of a two-level MIMIC

model with continuous factor indicators,
random factor loadings, two covariates on
within, and one covariate on between
with equal loadings across levels
DATA: FILE = ex9.19.dat;
VARIABLE: NAMES = y1-y4 x1 x2 w clus;
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (1000);
MODEL: %WITHIN%
s1-s4 | f BY y1-y4;
f@1;
f ON x1 x2;
%BETWEEN%
f ON w;
f;
PLOT: TYPE = PLOT2;
OUTPUT: TECH1 TECH8;

In this example, a two-level MIMIC model with continuous factor

indicators, random factor loadings, two covariates on within, and one
covariate on between with equal loadings across levels is estimated. In
the ANALYSIS command, TYPE=TWOLEVEL RANDOM is specified
indicating that a two-level model will be estimated. By specifying
ESTIMATOR=BAYES, a Bayesian analysis will be carried out. In
Bayesian estimation, the default is to use two independent Markov chain
Monte Carlo (MCMC) chains. If multiple processors are available,
using PROCESSORS=2 will speed up computations. The
BITERATIONS option is used to specify the maximum and minimum
number of iterations for each Markov chain Monte Carlo (MCMC) chain
when the potential scale reduction (PSR) convergence criterion (Gelman

310
 Examples: Multilevel Modeling With Complex Survey Data

& Rubin, 1992) is used. Using a number in parentheses, the

BITERATIONS option specifies that a minimum of 1000 and a
maximum of the default of 50,000 iterations will be used.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random factor loading
variables in the model. The name on the left-hand side of the | symbol
names the random factor loading variable. The statement on the right-
hand side of the | symbol defines the random factor loading variable.
Random factor loadings are defined using the BY option. The random
factor loading variables s1, s2, s3, and s4 are defined by the linear
regression of the factor indicators y1, y2, y3, and y4 on the factor f. The
factor variance is fixed at one to set the metric of the factor. The
residual variances of y1 through y4 are estimated and the residuals are
not correlated as the default. The ON statement describes the linear
regression of f on the individual-level covariates x1 and x2. In the
between part of the model, the ON statement describes the linear
regression of the random intercept f on the cluster-level covariate w.
The cluster-level residual variance of the factor is estimated. The
intercepts and the cluster-level residual variances of y1 through y4 are
estimated and the residuals are not correlated as the default.

By specifying TYPE=PLOT2 in the PLOT command, the following

plots are available: posterior parameter distributions, posterior
parameter trace plots, autocorrelation plots, posterior predictive
checking scatterplots, and posterior predictive checking distribution
plots. An explanation of the other commands can be found in Example
9.1.

Following is one alternative specification of the MODEL command

where a different factor fb is specified in the between part of the model
using the random intercepts as factor indicators. The residual variance
of fb is estimated as the default.

MODEL: %WITHIN%
s1-s4 | f BY y1-y4;
f@1;
f ON x1 x2;
%BETWEEN%
fb BY y1-y4;
fb ON w;

311
CHAPTER 9

Following is another alternative specification of the MODEL command

where a factor is specified in the between part of the model using the
random intercepts as factors indicators. The factor loadings of this
factor are held equal to the means of the random factor loadings defined
in the within part of the model.

MODEL: %WITHIN%
s1-s4 | f BY y1-y4;
f@1;
f ON x1 x2;
%BETWEEN%
fb BY y1-y4* (lam1-lam4);
fb ON w;
[s1-s4] (lam1-lam4);

EXAMPLE 9.20: THREE-LEVEL REGRESSION FOR A

CONTINUOUS DEPENDENT VARIABLE

TITLE: this is an example of a three-level

regression for a continuous dependent
variable
DATA: FILE = ex9.20.dat;
VARIABLE: NAMES = y x w z level2 level3;
CLUSTER = level3 level2;
WITHIN = x;
BETWEEN =(level2) w (level3) z;
ANALYSIS: TYPE = THREELEVEL RANDOM;
MODEL:
%WITHIN%
s1 | y ON x;
%BETWEEN level2%
s2 | y ON w;
s12 | s1 ON w;
y WITH s1;
%BETWEEN level3%
y ON z;
s1 ON z;
s2 ON z;
s12 ON z;
y WITH s1 s2 s12;
s1 WITH s2 s12;
s2 WITH s12;
OUTPUT: TECH1 TECH8;

312
 Examples: Multilevel Modeling With Complex Survey Data

Within
s1
x y

Between 2
y
s2
w
s12

s1

Between 3

s1
z
s2

s12

In this example, the three-level regression with a continuous dependent

variable shown in the picture above is estimated. The CLUSTER option
is used to identify the variables in the data set that contain clustering
information. Two cluster variables are used for a three-level model.
The CLUSTER option specifies that level3 is the cluster variable for
level 3 and level2 is the cluster variable for level 2. The cluster variable
for the highest level must come first, that is, level 2 is nested in level 3.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which

313
CHAPTER 9

they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2 and 3. If a variable is not
mentioned on the WITHIN list, it is modeled on all levels. The variable
x can be modeled on only level 1. The variable y can be modeled on all
levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. A cluster-level variable can be modeled on
all or some cluster levels. For TYPE=THREELEVEL, if a variable
measured on level 2 is mentioned on the BETWEEN list with a level 2
cluster label, it is modeled on only level 2. It has no variance on level 3.
A variable measured on level 3 must be mentioned on the BETWEEN
list with a level 3 cluster label. The variable w can be modeled on only
level 2. The variable z can be modeled on only level 3.

In the ANALYSIS command, TYPE=THREELEVEL RANDOM is

specified indicating that a three-level model will be estimated. In the
within and level 2 parts of the model, the | symbol is used in conjunction
with TYPE=RANDOM to name and define the random slope variables
in the model. The name on the left-hand side of the | symbol names the
random slope variable. The statement on the right-hand side of the |
symbol defines the random slope variable. Random slopes are defined
using the ON option. In the within part of the model, the random slope
s1 is defined by the linear regression of y on the individual-level
covariate x. The within-level residual variance of y is estimated as the
default. In the level 2 part of the model, two random slopes are defined.
The random slope s2 is defined by the linear regression of the level 2
random intercept y on the level 2 covariate w. The random slope s12 is
defined by the linear regression of the level 2 random slope s1 on the
level 2 covariate w. The level 2 residual variances of y and s1 are
estimated and the residuals are not correlated as the default. The WITH
statement specifies that the level 2 residuals of y and s1 are correlated.

In the level 3 part of the model, the first ON statement describes the
linear regression of the level 3 random intercept y on the level 3
covariate z. The next three ON statements describe the linear

314
 Examples: Multilevel Modeling With Complex Survey Data

regressions of the level 3 random slopes s1, s2, and s12 on the level 3
covariate z. The intercepts and level 3 residual variances of y, s1, s2,
and s12 are estimated and the residuals are not correlated as the default.
The WITH statements specify that the level 3 residuals of y, s1, s2, and
s12 are correlated. The default estimator for this type of analysis is
maximum likelihood with robust standard errors. The ESTIMATOR
option of the ANALYSIS command can be used to select a different
estimator. An explanation of the other commands can be found in
Examples 9.1 and 9.3.

EXAMPLE 9.21: THREE-LEVEL PATH ANALYSIS WITH A

CONTINUOUS AND A CATEGORICAL DEPENDENT
VARIABLE

TITLE: this an example of a three-level path

analysis with a continuous and a
categorical dependent variable
DATA: FILE = ex9.21.dat;
VARIABLE: NAMES = u y2 y y3 x w z level2 level3;
CATEGORICAL = u;
CLUSTER = level3 level2;
WITHIN = x;
BETWEEN = y2 (level2) w (level3) z y3;
ANALYSIS: TYPE = THREELEVEL;
ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (1000);
MODEL: %WITHIN%
u ON y x;
y ON x;
%BETWEEN level2%
u ON w y y2;
y ON w;
y2 ON w;
y WITH y2;
%BETWEEN level3%
u ON y y2;
y ON z;
y2 ON z;
y3 ON y y2;
y WITH y2;
u WITH y3;
OUTPUT: TECH1 TECH8;

315
CHAPTER 9

y
Within

x u

y Between 2

w u

y2

y Between 3

z u

y2 y3

In this example, the three-level path analysis with a continuous and a

categorical dependent variable shown in the picture above is estimated.
The CATEGORICAL option is used to specify which dependent
variables are treated as binary or ordered categorical (ordinal) variables

316
 Examples: Multilevel Modeling With Complex Survey Data

in the model and its estimation. In the example above, the variable u is
binary or ordered categorical.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2 and 3. If a variable is not
mentioned on the WITHIN list, it is modeled on all levels. The variable
x can be modeled on only level 1. The variables u and y can be modeled
on all levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. A cluster-level variable can be modeled on
all or some cluster levels. For TYPE=THREELEVEL, if a variable
measured on level 2 is mentioned on the BETWEEN list without a label,
it is modeled on levels 2 and 3. If a variable measured on level 2 is
mentioned on the BETWEEN list with a level 2 cluster label, it is
modeled on only level 2. It has no variance on level 3. A variable
measured on level 3 must be mentioned on the BETWEEN list with a
level 3 cluster label. The variable y2 can be modeled on levels 2 and 3.
The variable w can be modeled on only level 2. The variables z and y3
can be modeled on only level 3.

In the ANALYSIS command, TYPE=THREELEVEL is specified

indicating that a three-level model will be estimated. By specifying
ESTIMATOR=BAYES, a Bayesian analysis will be carried out. In
Bayesian estimation, the default is to use two independent Markov chain
Monte Carlo (MCMC) chains. If multiple processors are available,
using PROCESSORS=2 will speed up computations. The
BITERATIONS option is used to specify the maximum and minimum
number of iterations for each Markov chain Monte Carlo (MCMC) chain
when the potential scale reduction (PSR) convergence criterion (Gelman
& Rubin, 1992) is used. Using a number in parentheses, the
BITERATIONS option specifies that a minimum of 1000 and a
maximum of the default of 50,000 iterations will be used.

317
CHAPTER 9

In the within part of the model, the first ON statement describes the
probit regression of u on the mediator y and the individual-level
covariate x. The second ON statement describes the linear regression of
the mediator y on the covariate x. The within-level residual variance of
y is estimated as the default. In the level 2 part of the model, the first
ON statement describes the linear regression of the level 2 random
intercept u on the level 2 covariate w, the level 2 random intercept y, and
the level 2 mediator y2. The second ON statement describes the linear
regression of the level 2 random intercept y on the level 2 covariate w.
The third ON statement describes the linear regression of the level 2
mediator y2 on the level 2 covariate w. The level 2 residual variances of
u, y, and y2 are estimated and the residuals are not correlated as the
default. The WITH statement specifies that the level 2 residuals of y and
y2 are correlated. In the level 3 part of the model, the first ON statement
describes the linear regression of the level 3 random intercept u on the
level 3 random intercepts y and y2. The second ON statement describes
the linear regression of the level 3 random intercept y on the level 3
covariate z. The third ON statement describes the linear regression of
the level 3 random intercept y2 on the level 3 covariate z. The fourth
ON statement describes the linear regression of the level 3 variable y3
on the level 3 random intercepts y and y2. The threshold of u; the
intercepts of y, y2, and y3; and the level 3 residual variances of u, y, y2,
and y3 are estimated and the residuals are not correlated as the default.
The first WITH statement specifies that the residuals of y and y2 are
correlated. The second WITH statement specifies that the residuals of u
and y3 are correlated. An explanation of the other commands can be
found in Examples 9.1, 9.3, and 9.20.

318
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.22: THREE-LEVEL MIMIC MODEL WITH

CONTINUOUS FACTOR INDICATORS, TWO COVARIATES
ON WITHIN, ONE COVARIATE ON BETWEEN LEVEL 2, AND
ONE COVARIATE ON BETWEEN LEVEL 3 WITH RANDOM
SLOPES ON BOTH WITHIN AND BETWEEN LEVEL 2

TITLE: this is an example of a three-level MIMIC

model with continuous factor indicators,
two covariates on within, one covariate on
between level 2, one covariate on between
level 3 with random slopes on both within
and between level 2
DATA: FILE = ex9.22.dat;
VARIABLE: NAMES = y1-y6 x1 x2 w z level2 level3;
CLUSTER = level3 level2;
WITHIN = x1 x2;
BETWEEN = (level2) w (level3) z;
ANALYSIS: TYPE = THREELEVEL RANDOM;
MODEL: %WITHIN%
fw1 BY y1-y3;
fw2 BY y4-y6;
fw1 ON x1;
s | fw2 ON x2;
%BETWEEN level2%
fb2 BY y1-y6;
sf2 | fb2 ON w;
ss | s ON w;
fb2 WITH s;
%BETWEEN level3%
fb3 BY y1-y6;
fb3 ON z;
s ON z;
sf2 ON z;
ss ON z;
fb3 WITH s sf2 ss;
s WITH sf2 ss;
sf2 WITH ss;
OUTPUT: TECH1 TECH8;

319
CHAPTER 9

Within
y1

x1 fw1 y2

y3

y4

s
x2 fw2 y5

y6

Between 2
y1

y2

y3
sf2
w fb2
y4
ss

y5
s
y6

Between 3
y1

y2

y3

z fb3
y4

s
y5
sf2
y6

ss

320
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the three-level MIMIC model with continuous factor

indicators, two covariates on within, one covariate on between level 2,
one covariate on between level 3 with random slopes on both within and
between level 2 shown in the picture above is estimated.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2 and 3. If a variable is not
mentioned on the WITHIN list, it is modeled on all levels. The variables
x1 and x2 can be modeled on only level 1. The variables y1 through y6
can be modeled on all levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. A cluster-level variable can be modeled on
all or some cluster levels. For TYPE=THREELEVEL, if a variable
measured on level 2 is mentioned on the BETWEEN list with a level 2
cluster label, it is modeled on only level 2. It has no variance on level 3.
A variable measured on level 3 must be mentioned on the BETWEEN
list with a level 3 cluster label. The variable w can be modeled on only
level 2. The variable z can be modeled on only level 3.

In the ANALYSIS command, TYPE=THREELEVEL RANDOM is

specified indicating that a three-level model will be estimated. In the
within part of the model, the first BY statement specifies that the factor
fw1 is measured by y1 through y3. The second BY statement specifies
that fw2 is measured by y4 through y6. The metric of the factors is set
automatically by the program by fixing the first factor loading in each
BY statement to one. This default can be overridden. The residual
variances of the factor indicators are estimated and the residuals are not
correlated as the default. The residual variances of the factors are
estimated and the residuals are correlated as the default. The ON
statement describes the linear regression of fw1 on the individual-level
covariate x1.

321
CHAPTER 9

In the within and level 2 parts of the model, the | symbol is used in
conjunction with TYPE=RANDOM to name and define the random
slope variables in the model. The name on the left-hand side of the |
symbol names the random slope variable. The statement on the right-
hand side of the | symbol defines the random slope variable. Random
slopes are defined using the ON option. In the within part of the model,
the random slope s is defined by the linear regression of fw2 on the
individual-level covariate x2.

In the level 2 part of the model, the BY statement specifies that the
factor fb2 is measured by the level 2 random intercepts y1 through y6.
The metric of the factors is set automatically by the program by fixing
the first factor loading in each BY statement to one. This default can be
overridden. The level 2 residual variances of the factor indicators are
estimated and the residuals are not correlated as the default. The
variance of the factor is estimated as the default. The random slope sf2
is defined by the linear regression of fb2 on the level 2 covariate w. The
random slope ss is defined by the linear regression of the random slope s
on the level 2 covariate w. The level 2 residual variances of fb2 and s
are estimated and the residuals are not correlated as the default.

In the level 3 part of the model, the BY statement specifies that the
factor fb3 is measured by the level 3 random intercepts y1 through y6.
The metric of the factors is set automatically by the program by fixing
the first factor loading in each BY statement to one. This default can be
overridden. The intercept and level 3 residual variances of the factor
indicators are estimated and the residuals are not correlated as the
default. The residual variance of the factor is estimated as the default.
The first ON statement describes the linear regression of fb3 on the level
3 covariate z. The second ON statement describes the linear regression
of the random slope s on the level 3 covariate z. The third ON statement
describes the linear regression of the random slope sf2 on the level 3
covariate z. The fourth ON statement specifies the linear regression of
the random slope ss on the level 3 covariate z. The intercepts of y1
through y6, s, sf2, and ss; and the level 3 residual variances of fb3, s,
sf2, and ss are estimated and the residuals are not correlated as the
default. The WITH statements specify that the level 3 residuals of fb3,
s, sf2, and ss are correlated. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select

322
 Examples: Multilevel Modeling With Complex Survey Data

a different estimator. An explanation of the other commands can be

found in Examples 9.1, 9.3, and 9.20.

EXAMPLE 9.23: THREE-LEVEL GROWTH MODEL WITH A

CONTINUOUS OUTCOME AND ONE COVARIATE ON EACH
OF THE THREE LEVELS

TITLE: this is an example of a three-level growth

model with a continuous outcome and one
covariate on each of the three levels
DATA: FILE = ex9.23.dat;
VARIABLE: NAMES = y1-y4 x w z level2 level3;
CLUSTER = level3 level2;
WITHIN = x;
BETWEEN = (level2) w (level3) z;
ANALYSIS: TYPE = THREELEVEL;
MODEL: %WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3;
iw sw ON x;
%BETWEEN level2%
ib2 sb2 | y1@0 y2@1 y3@2 y4@3;
ib2 sb2 ON w;
%BETWEEN level3%
ib3 sb3 | y1@0 y2@1 y3@2 y4@3;
ib3 sb3 ON z;
y1-y4@0;
OUTPUT: TECH1 TECH8;

323
CHAPTER 9

y1 y2 y3 y4

iw sw

x
Within

Between 2

y1 y2 y3 y4

ib2 sb2

Between 3

y1 y2 y3 y4

ib3 sb3

324
 Examples: Multilevel Modeling With Complex Survey Data

In this example, the three-level growth model with a continuous outcome

and one covariate on each of the three levels shown in the picture above
is estimated.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2 and 3. If a variable is not
mentioned on the WITHIN list, it is modeled on all levels. The variable
x can be modeled on only level 1. The variables y1 through y4 can be
modeled on all levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. A cluster-level variable can be modeled on
all or some cluster levels. For TYPE=THREELEVEL, if a variable
measured on level 2 is mentioned on the BETWEEN list with a level 2
cluster label, it is modeled on only level 2. It has no variance on level 3.
A variable measured on level 3 must be mentioned on the BETWEEN
list with a level 3 cluster label. The variable w can be modeled on only
level 2. The variable z can be modeled on only level 3.

In the ANALYSIS command, TYPE=THREELEVEL is specified

indicating that a three-level model will be estimated. In the within part
of the model, the | symbol is used to name and define the within
intercept and slope factors in a growth model. The names iw and sw on
the left-hand side of the | symbol are the names of the intercept and slope
growth factors, respectively. The statement on the right-hand side of the
| symbol specifies the outcome and the time scores for the growth model.
The time scores for the slope growth factor are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
growth factor as an initial status factor. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of y1 through y4 are estimated
and allowed to be different across time and the residuals are not
correlated as the default. In the parameterization of the growth model

325
CHAPTER 9

shown here, the intercepts of the outcome variables at the four time
points are fixed at zero as the default. The ON statement describes the
linear regression of the intercept and slope growth factors on the
individual-level covariate x. The residual variances of the growth
factors are estimated and the residuals are correlated as the default. The
level 2 residual variances of y1 through y4 are estimated and allowed to
be different across time and the residuals are not correlated as the
default.

The growth model specified in the within part of the model is also
specified on levels 2 and 3. In the level 2 part of the model, the ON
statement describes the linear regression of the level 2 intercept and
slope growth factors on the level 2 covariate w. The level 2 residual
variances of the growth factors are estimated and the residuals are
correlated as the default. In the level 3 part of the model, the ON
statement describes the linear regression of the level 3 intercept and
slope growth factors on the level 3 covariate z. The intercepts and level
3 residual variances of the growth factors are estimated and the residuals
are correlated as the default. The level 3 residual variances of y1
through y4 are fixed at zero. The default estimator for this type of
analysis is maximum likelihood with robust standard errors. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator. An explanation of the other commands can be
found in Examples 9.1, 9.3, and 9.20.

326
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.24: REGRESSION FOR A CONTINUOUS

DEPENDENT VARIABLE USING CROSS-CLASSIFIED DATA

TITLE: this is an example of a regression for a

continuous dependent variable using cross-
classified data
DATA: FILE = ex9.24.dat;
VARIABLE: NAMES = y x1 x2 w z level2a level2b;
CLUSTER = level2b level2a;
WITHIN = x1 x2;
BETWEEN = (level2a) w (level2b) z;
ANALYSIS: TYPE = CROSSCLASSIFIED RANDOM;
ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (2000);
MODEL: %WITHIN%
y ON x1;
s | y ON x2;
%BETWEEN level2a%
y ON w;
s ON w;
y WITH s;
%BETWEEN level2b%
y ON z;
s ON z;
y WITH s;
OUTPUT: TECH1 TECH8;

327
CHAPTER 9

Within
x1

y
s

x2

Level 2a

Level 2b

In this example, the regression for a continuous dependent variable using

cross-classified data shown in the picture above is estimated. The
CLUSTER option is used to identify the variables in the data set that
contain clustering information. Two cluster variables are used for a
cross-classified model. The CLUSTER option specifies that level2b is
the cluster variable for level 2b and level2a is the cluster variable for
level 2a.

328
 Examples: Multilevel Modeling With Complex Survey Data

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2a and 2b. If a variable is not
mentioned on the WITHIN list, it is modeled on all levels. The variables
x1 and x2 can be modeled on only level 1. The variable y can be
modeled on all levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. For TYPE=CROSSCLASSIFIED, a
variable measured on level 2a must be mentioned on the BETWEEN list
with a level 2a cluster label. It can be modeled on only level 2a. A
variable measured on level 2b must be mentioned on the BETWEEN list
with a level 2b cluster label. It can be modeled on only level 2b. The
variable w can be modeled on only level 2a. The variable z can be
modeled on only level 2b.

In the ANALYSIS command, TYPE=CROSSCLASSIFIED RANDOM

is specified indicating that a cross-classified model will be estimated.
By specifying ESTIMATOR=BAYES, a Bayesian analysis will be
carried out. In Bayesian estimation, the default is to use two
independent Markov chain Monte Carlo (MCMC) chains. If multiple
processors are available, using PROCESSORS=2 will speed up
computations. The BITERATIONS option is used to specify the
maximum and minimum number of iterations for each Markov chain
Monte Carlo (MCMC) chain when the potential scale reduction (PSR)
convergence criterion (Gelman & Rubin, 1992) is used. Using a number
in parentheses, the BITERATIONS option specifies that a minimum of
2000 and a maximum of the default of 50,000 iterations will be used.

In the within part of the model, the ON statement describes the linear
regression of y on the individual-level covariate x1. The residual
variance of y is estimated as the default. The | symbol is used in
conjunction with TYPE=RANDOM to name and define the random
slope variables in the model. The name on the left-hand side of the |
symbol names the random slope variable. The statement on the right-

329
CHAPTER 9

hand side of the | symbol defines the random slope variable. Random
slopes are defined using the ON option. The random slope s is defined
by the linear regression of y on the individual-level covariate x2. In the
level 2a part of the model, the first ON statement describes the linear
regression of the level 2a random intercept y on the level 2a covariate w.
The second ON statement describes the linear regression of the level 2a
random slope s on the level 2a covariate w. The residuals of y and s are
estimated and the residuals are not correlated as the default. The WITH
statement specifies that the residuals of y and s are correlated. In the
level 2b part of the model, the first ON statement describes the linear
regression of the level 2b random intercept y on the level 2b covariate z.
The second ON statement describes the linear regression of the level 2b
random slope s on the level 2b covariate z. The residual variances of y
and s are estimated and the residuals are not correlated as the default.
The WITH statement specifies that the residuals of y and s are
correlated. The intercepts of y and s are estimated as the default on level
2b. An explanation of the other commands can be found in Examples
9.1 and 9.3.

EXAMPLE 9.25: PATH ANALYSIS WITH CONTINUOUS

DEPENDENT VARIABLES USING CROSS-CLASSIFIED DATA

TITLE: this is an example of path analysis with

continuous dependent variables using
cross-classified data
DATA: FILE = ex9.25.dat;
VARIABLE: NAMES = y1 y2 x w z level2a level2b;
CLUSTER = level2b level2a;
WITHIN = x;
BETWEEN = (level2a) w (level2b) z;
ANALYSIS: TYPE = CROSSCLASSIFIED;
ESTIMATOR = BAYES;
PROCESSORS = 2;
MODEL: %WITHIN%
y2 ON y1 x;
y1 ON x;
%BETWEEN level2a%
y1 y2 ON w;
y1 WITH y2;
%BETWEEN level2b%
y1 y2 ON z;
y1 WITH y2;
OUTPUT: TECH1 TECH8;

330
 Examples: Multilevel Modeling With Complex Survey Data

Within
y1

x y2

Level 2a

y1

y2

Level 2b

y1

y2

In this example, the path analysis with continuous dependent variables

using cross-classified data shown in the picture above is estimated. The
WITHIN option is used to identify the variables in the data set that are
measured on the individual level and to specify the levels on which they
are modeled. All variables on the WITHIN list must be measured on the
individual level. An individual-level variable can be modeled on all or
some levels. If a variable measured on the individual level is mentioned

331
CHAPTER 9

on the WITHIN list without a label, it is modeled on only level 1. It has

no variance on levels 2a and 2b. If a variable is not mentioned on the
WITHIN list, it is modeled on all levels. The variable x can be modeled
on only level 1. The variables y1 and y2 can be modeled on all levels.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level(s) and to specify the level(s) on
which they are modeled. All variables on the BETWEEN list must be
measured on a cluster level. For TYPE=CROSSCLASSIFIED, a
variable measured on level 2a must be mentioned on the BETWEEN list
with a level 2a cluster label. It can be modeled on only level 2a. A
variable measured on level 2b must be mentioned on the BETWEEN list
with a level 2b cluster label. It can be modeled on only level 2b. The
variable w can be modeled on only level 2a. The variable z can be
modeled on only level 2b.

In the ANALYSIS command, TYPE=CROSSCLASSIFIED is specified

indicating that a cross-classified model will be estimated. By specifying
ESTIMATOR=BAYES, a Bayesian analysis will be carried out. No
other estimators are available. In Bayesian estimation, the default is to
use two independent Markov chain Monte Carlo (MCMC) chains. If
multiple processors are available, using PROCESSORS=2 will speed up
computations.

In the within part of the model, the first ON statement describes the
linear regression of y2 on the mediator y1 and the individual-level
covariate x. The second ON statement describes the linear regression of
y1 on the individual-level covariate x. The residuals of y1 and y2 are
estimated and the residual are not correlated as the default. In the level
2a part of the model, the first ON statement describes the linear
regressions of the level 2a intercepts y1 and y2 on the level 2a covariate
w. The level 2a residuals are estimated and the residuals are not
correlated as the default. The WITH statement specifies that the level 2a
residuals of y1 and y2 are correlated. In the level 2b part of the model,
the first ON statement describes the linear regression of the level 2b
random intercepts y1 and y2 on the level 2b covariate z. The level 2b
residuals are estimated and the residuals are not correlated as the default.
The WITH statement specifies that the level 2b residuals of y1 and y2
are correlated. The intercepts of y1 and y2 are estimated as the default
on level 2b. An explanation of the other commands can be found in
Examples 9.1, 9.3, and 9.24.

332
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.26: IRT WITH RANDOM BINARY ITEMS USING

CROSS-CLASSIFIED DATA

TITLE: this is an example of IRT with random

binary items using cross-classified
data
DATA: FILE = ex9.26.dat;
VARIABLE: NAMES = u subject item;
CATEGORICAL = u;
CLUSTER = item subject;
ANALYSIS: TYPE = CROSSCLASSIFIED RANDOM;
ESTIMATOR = BAYES;
PROCESSORS = 2;
MODEL: %WITHIN%
%BETWEEN subject%
s | f BY u;
f@1;
u@0;
%BETWEEN item%
u; [u$1];
s; [s];
OUTPUT: TECH1 TECH8;

In this example, an IRT with random binary items using cross-classified

data is estimated (Fox, 2010). Both the intercepts and factor loadings of
the set of items are random. The CATEGORICAL option is used to
specify which dependent variables are treated as binary or ordered
categorical (ordinal) variables in the model and its estimation. In the
example above, the variable u is binary or ordered categorical. The
CLUSTER option is used to identify the variables in the data set that
contain clustering information. Two cluster variables are used for a
cross-classified model. The CLUSTER option specifies that item is the
cluster variable for the item level and subject is the cluster variable for
the subject level. The fastest moving level must come first.

The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. If a variable is not mentioned on the WITHIN list, it
is modeled on all levels. The variable u can be modeled on the subject
and item levels.

In the ANALYSIS command, TYPE=CROSSCLASSIFIED RANDOM

is specified indicating that a cross-classified model will be estimated.

333
CHAPTER 9

By specifying ESTIMATOR=BAYES, a Bayesian analysis will be

carried out. In Bayesian estimation, the default is to use two
independent Markov chain Monte Carlo (MCMC) chains. If multiple
processors are available, using PROCESSORS=2 will speed up
computations.

The within part of the model is not used in this example. In the subject
part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random factor loading
variables in the model. The name on the left-hand side of the | symbol
names the random factor loading variable. The statement on the right-
hand side of the | symbol defines the random factor loading variable.
Random factor loadings are defined using the BY option. The random
factor loading variable s is defined by the probit regression of u on the
factor f. The factor variance is fixed at one to set the metric of the
factor. The across-subject variance of u is fixed at zero. In the item part
of the model, the variance of the random intercept u, the threshold of u,
and the mean and variance of the random factor loading s are estimated
as the default. An explanation of the other commands can be found in
Examples 9.1, 9.3, and 9.24.

334
 Examples: Multilevel Modeling With Complex Survey Data

EXAMPLE 9.27: MULTIPLE INDICATOR GROWTH MODEL

WITH RANDOM INTERCEPTS AND FACTOR LOADINGS
USING CROSS-CLASSIFIED DATA

TITLE: this is an example of a multiple indicator

growth model with random intercepts and
factor loadings using cross-classified
data
DATA: FILE = ex9.27.dat;
VARIABLE: NAMES = y1-y3 time subject;
USEVARIABLES = y1-y3 timescor;
CLUSTER = subject time;
WITHIN = timescor (time) y1-y3;
DEFINE: timescor = (time-1)/100;
ANALYSIS: TYPE = CROSSCLASSIFIED RANDOM;
ESTIMATOR = BAYES;
PROCESSORS = 2;
BITERATIONS = (1000);
MODEL: %WITHIN%
s1-s3 | f BY y1-y3;
f@1;
s | f ON timescor;
y1-y3; [y1-y3@0];
%BETWEEN time%
s1-s3; [s1-s3];
y1-y3; [y1-y3@0];
s@0; [s@0];
%BETWEEN subject%
f; [f];
s1-s3@0; [s1-s3@0];
s; [s];
OUTPUT: TECH1 TECH8;

In this example, a multiple indicator growth model with random

intercepts and factor loadings using cross-classified data is estimated.
The WITHIN option is used to identify the variables in the data set that
are measured on the individual level and to specify the levels on which
they are modeled. All variables on the WITHIN list must be measured
on the individual level. An individual-level variable can be modeled on
all or some levels. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on the time and subject levels. If it is
mentioned on the WITHIN list with a time cluster label, it is modeled on
levels 1 and on the time level. It has no variance on the subject level.

335
CHAPTER 9

The variable timescor can be modeled on only level 1. The variables y1,
y2, and y3 can be modeled on levels 1 and the time level. The DEFINE
command is used to transform existing variables and to create new
variables. The new variable timescor is a time score variable centered at
the first time point.

In the ANALYSIS command, TYPE=CROSSCLASSIFIED RANDOM

is specified indicating that a cross-classified model will be estimated.
By specifying ESTIMATOR=BAYES, a Bayesian analysis will be
carried out. In Bayesian estimation, the default is to use two
independent Markov chain Monte Carlo (MCMC) chains. If multiple
processors are available, using PROCESSORS=2 will speed up
computations. The BITERATIONS option is used to specify the
maximum and minimum number of iterations for each Markov chain
Monte Carlo (MCMC) chain when the potential scale reduction (PSR)
convergence criterion (Gelman & Rubin, 1992) is used. Using a number
in parentheses, the BITERATIONS option specifies that a minimum of
1000 and a maximum of the default of 50,000 iterations will be used.

In the within part of the model, the | symbol is used in conjunction with
TYPE=RANDOM to name and define the random factor loading
variables in the model. The name on the left-hand side of the | symbol
names the random factor loading variable. The statement on the right-
hand side of the | symbol defines the random factor loading variable.
Random factor loadings are defined using the BY option. The random
factor loading variables s1, s2, and s3 are defined by the linear
regression of the factor indicators y1, y2, and y3 on the factor f. The
factor variance is fixed to one to set the metric of the factor. The
intercepts of the factor indicators are fixed at zero as part of the growth
model parameterization. The residual variances are estimated and the
residuals are not correlated as the default.

The | symbol is used in conjunction with TYPE=RANDOM to name and

define the random slope variables in the model. The name on the left-
hand side of the | symbol names the random slope variable. The
statement on the right-hand side of the | symbol defines the random slope
variable. Random slopes are defined using the ON option. The random
slope growth factor s is defined by the linear regression of f on the
individual-level covariate timescor.

336
 Examples: Multilevel Modeling With Complex Survey Data

In the time part of the model, the means and variances of the random
factor loadings s1, s2, and s3 and the variances of the random intercepts
y1, y2, and y3 are estimated. The intercepts of y1, y2, and y3 are fixed
at zero as part of the growth model parameterization. The variances of
the random factor loadings s1, s2, and s3 and the variances of the
random intercepts y1, y2, and y3 represent measurement non-invariance
across time. The mean and variance of the random slope growth factor s
are fixed at zero.

In the subject part of the model, f is a random intercept growth factor.

Its mean and variance are estimated. The means and variances of the
random factor loadings s1, s2, and s3 are fixed at zero. The mean and
variance of the random slope growth factor s are estimated. An
explanation of the other commands can be found in Examples 9.1, 9.3,
and 9.24.

337
CHAPTER 9

338
 Examples: Multilevel Mixture Modeling

CHAPTER 10
EXAMPLES: MULTILEVEL
MIXTURE MODELING

Multilevel mixture modeling (Asparouhov & Muthén, 2008a) combines

the multilevel and mixture models by allowing not only the modeling of
multilevel data but also the modeling of subpopulations where
population membership is not known but is inferred from the data.
Mixture modeling can be combined with the multilevel analyses
discussed in Chapter 9. Observed outcome variables can be continuous,
censored, binary, ordered categorical (ordinal), unordered categorical
(nominal), counts, or combinations of these variable types.

With cross-sectional data, the number of levels in Mplus is the same as

the number of levels in conventional multilevel modeling programs.
Mplus allows two-level modeling. With longitudinal data, the number of
levels in Mplus is one less than the number of levels in conventional
multilevel modeling programs because Mplus takes a multivariate
approach to repeated measures analysis. Longitudinal models are two-
level models in conventional multilevel programs, whereas they are one-
level models in Mplus. Single-level longitudinal models are discussed in
Chapter 6, and single-level longitudinal mixture models are discussed in
Chapter 8. Three-level longitudinal analysis where time is the first level,
individual is the second level, and cluster is the third level is handled by
two-level growth modeling in Mplus as discussed in Chapter 9.

Multilevel mixture models can include regression analysis, path analysis,

confirmatory factor analysis (CFA), item response theory (IRT) analysis,
structural equation modeling (SEM), latent class analysis (LCA), latent
transition analysis (LTA), latent class growth analysis (LCGA), growth
mixture modeling (GMM), discrete-time survival analysis, continuous-
time survival analysis, and combinations of these models.

All multilevel mixture models can be estimated using the following

special features:

• Single or multiple group analysis

• Missing data

339
CHAPTER 10

• Complex survey data

• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Maximum likelihood estimation for all outcome types
• Wald chi-square test of parameter equalities
• Analysis with between-level categorical latent variables
• Test of equality of means across latent classes using posterior
probability-based multiple imputations

For TYPE=MIXTURE, multiple group analysis is specified by using the

KNOWNCLASS option of the VARIABLE command. The default is to
estimate the model under missing data theory using all available data.
The LISTWISE option of the DATA command can be used to delete all
observations from the analysis that have missing values on one or more
of the analysis variables. Corrections to the standard errors and chi-
square test of model fit that take into account stratification, non-
independence of observations, and unequal probability of selection are
obtained by using the TYPE=COMPLEX option of the ANALYSIS
command in conjunction with the STRATIFICATION, CLUSTER,
WEIGHT, WTSCALE, BWEIGHT, and BWTSCALE options of the
VARIABLE command. Latent variable interactions are specified by
using the | symbol of the MODEL command in conjunction with the
XWITH option of the MODEL command. Random slopes are specified
by using the | symbol of the MODEL command in conjunction with the
ON option of the MODEL command. Individually-varying times of
observations are specified by using the | symbol of the MODEL
command in conjunction with the AT option of the MODEL command
and the TSCORES option of the VARIABLE command. Linear and
non-linear parameter constraints are specified by using the MODEL
CONSTRAINT command. Maximum likelihood estimation is specified
by using the ESTIMATOR option of the ANALYSIS command. The
MODEL TEST command is used to test linear restrictions on the
parameters in the MODEL and MODEL CONSTRAINT commands
using the Wald chi-square test. Between-level categorical latent
variables are specified using the CLASSES and BETWEEN options of
the VARIABLE command. The AUXILIARY option is used to test the
equality of means across latent classes using posterior probability-based
multiple imputations.

340
 Examples: Multilevel Mixture Modeling

Graphical displays of observed data and analysis results can be obtained

using the PLOT command in conjunction with a post-processing
graphics module. The PLOT command provides histograms,
scatterplots, plots of individual observed and estimated values, and plots
of sample and estimated means and proportions/probabilities. These are
available for the total sample, by group, by class, and adjusted for
covariates. The PLOT command includes a display showing a set of
descriptive statistics for each variable. The graphical displays can be
edited and exported as a DIB, EMF, or JPEG file. In addition, the data
for each graphical display can be saved in an external file for use by
another graphics program.

Following is the set of cross-sectional examples included in this chapter:

• 10.1: Two-level mixture regression for a continuous dependent

variable*
• 10.2: Two-level mixture regression for a continuous dependent
variable with a between-level categorical latent variable*
• 10.3: Two-level mixture regression for a continuous dependent
variable with between-level categorical latent class indicators for a
between-level categorical latent variable*
• 10.4: Two-level CFA mixture model with continuous factor
indicators*
• 10.5: Two-level IRT mixture analysis with binary factor indicators
and a between-level categorical latent variable*
• 10.6: Two-level LCA with categorical latent class indicators with
covariates*
• 10.7: Two-level LCA with categorical latent class indicators and a
between-level categorical latent variable

Following is the set of longitudinal examples included in this chapter:

• 10.8: Two-level growth model for a continuous outcome (three-

level analysis) with a between-level categorical latent variable*
• 10.9: Two-level GMM for a continuous outcome (three-level
analysis)*
• 10.10: Two-level GMM for a continuous outcome (three-level
analysis) with a between-level categorical latent variable*
• 10.11: Two-level LCGA for a three-category outcome*
• 10.12: Two-level LTA with a covariate*

341
CHAPTER 10

• 10.13: Two-level LTA with a covariate and a between-level

categorical latent variable

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

EXAMPLE 10.1: TWO-LEVEL MIXTURE REGRESSION FOR

A CONTINUOUS DEPENDENT VARIABLE

TITLE: this is an example of a two-level mixture

regression for a continuous dependent
variable
DATA: FILE IS ex10.1.dat;
VARIABLE: NAMES ARE y x1 x2 w class clus;
USEVARIABLES = y x1 x2 w;
CLASSES = c (2);
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
STARTS = 0;
MODEL:
%WITHIN%
%OVERALL%
y ON x1 x2;
c ON x1;
%c#1%
y ON x2;
y;
%BETWEEN%
%OVERALL%
y ON w;
c#1 ON w;
c#1*1;
%c#1%
[y*2];
OUTPUT: TECH1 TECH8;

342
 Examples: Multilevel Mixture Modeling

x1 x2

Within

c#1 y Between

In this example, the two-level mixture regression model for a continuous

dependent variable shown in the picture above is estimated. This
example is the same as Example 7.1 except that it has been extended to
the multilevel framework. In the within part of the model, the filled
circles at the end of the arrows from x1 to c and y represent random
intercepts that are referred to as c#1 and y in the between part of the
model. In the between part of the model, the random intercepts are
shown in circles because they are continuous latent variables that vary

343
CHAPTER 10

across clusters. The random intercepts y and c#1 are regressed on a

cluster-level covariate w.

Because c is a categorical latent variable, the interpretation of the picture

is not the same as for models with continuous latent variables. The
arrow from c to the y variable indicates that the intercept of the y
variable varies across the classes of c. This corresponds to the
regression of y on a set of dummy variables representing the categories
of c. The broken arrow from c to the arrow from x2 to y indicates that
the slope in the linear regression of y on x2 varies across the classes of c.
The arrow from x1 to c represents the multinomial logistic regression of
c on x1.

TITLE: this is an example of a two-level mixture

regression for a continuous dependent
variable

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.
DATA: FILE IS ex10.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex10.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.
VARIABLE: NAMES ARE y x1 x2 w class clus;
USEVARIABLES = y x1 x2 w;
CLASSES = c (2);
WITHIN = x1 x2;
BETWEEN = w;
CLUSTER = clus;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains six variables: y, x1, x2, w, c, and clus. If not all of the
variables in the data set are used in the analysis, the USEVARIABLES
option can be used to select a subset of variables for analysis. Here the
variables y1, x1, x2, and w have been selected for analysis. The
CLASSES option is used to assign names to the categorical latent
variables in the model and to specify the number of latent classes in the
model for each categorical latent variable. In the example above, there

344
 Examples: Multilevel Mixture Modeling

is one categorical latent variable c that has two latent classes. The
WITHIN option is used to identify the variables in the data set that are
measured on the individual level and modeled only on the within level.
They are specified to have no variance in the between part of the model.
The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level and modeled only on the between
level. Variables not mentioned on the WITHIN or the BETWEEN
statements are measured on the individual level and can be modeled on
both the within and between levels. The CLUSTER option is used to
identify the variable that contains cluster information.
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
STARTS = 0;

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis that
is to be performed. By selecting TWOLEVEL MIXTURE, a multilevel
mixture model will be estimated. By specifying STARTS=0 in the
ANALYSIS command, random starts are turned off.
MODEL:
%WITHIN%
%OVERALL%
y ON x1 x2;
c ON x1;
%c#1%
y ON x2;
y;
%BETWEEN%
%OVERALL%
y ON w;
c#1 ON w;
c#1*1;
%c#1%
[y*2];

The MODEL command is used to describe the model to be estimated. In

multilevel models, a model is specified for both the within and between
parts of the model. For mixture models, there is an overall model
designated by the label %OVERALL%. The overall model describes the
part of the model that is in common for all latent classes. The part of the
model that differs for each class is specified by a label that consists of
the categorical latent variable name followed by the number sign (#)
followed by the class number. In the example above, the label %c#2%

345
CHAPTER 10

refers to the part of the model for class 2 that differs from the overall
model.

In the overall model in the within part of the model, the first ON
statement describes the linear regression of y on the individual-level
covariates x1 and x2. The second ON statement describes the
multinomial logistic regression of the categorical latent variable c on the
individual-level covariate x1 when comparing class 1 to class 2. The
intercept in the regression of c on x1 is estimated as the default. In the
model for class 1 in the within part of the model, the ON statement
describes the linear regression of y on the individual-level covariate x2
which relaxes the default equality of regression coefficients across
classes. By mentioning the residual variance of y, it is not held equal
across classes.

In the overall model in the between part of the model, the first ON
statement describes the linear regression of the random intercept y on the
cluster-level covariate w. The second ON statement describes the linear
regression of the random intercept c#1 of the categorical latent variable c
on the cluster-level covariate w. The random intercept c#1 is a
continuous latent variable. Each class of the categorical latent variable c
except the last class has a random intercept. A starting value of one is
given to the residual variance of the random intercept c#1. In the class-
specific part of the between part of the model, the intercept of y is given
a starting value of 2 for class 1.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator.

Following is an alternative specification of the multinomial logistic

regression of c on the individual-level covariate x1 in the within part of
the model:

c#1 ON x1;

346
 Examples: Multilevel Mixture Modeling

where c#1 refers to the first class of c. The classes of a categorical latent
variable are referred to by adding to the name of the categorical latent
variable the number sign (#) followed by the number of the class. This
alternative specification allows individual parameters to be referred to in
the MODEL command for the purpose of giving starting values or
placing restrictions.

OUTPUT: TECH1 TECH8;

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model. The TECH8 option is used to request that the
optimization history in estimating the model be printed in the output.
TECH8 is printed to the screen during the computations as the default.
TECH8 screen printing is useful for determining how long the analysis
takes.

347
CHAPTER 10

EXAMPLE 10.2: TWO-LEVEL MIXTURE REGRESSION FOR

A CONTINUOUS DEPENDENT VARIABLE WITH A
BETWEEN-LEVEL CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level mixture

regression for a continuous dependent
variable with a between-level categorical
latent variable
DATA: FILE = ex10.2.dat;
VARIABLE: NAMES ARE y x1 x2 w dummy clus;
USEVARIABLES = y-w;
CLASSES = cb(2);
WITHIN = x1 x2;
BETWEEN = cb w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE RANDOM;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
s1 | y ON x1;
s2 | y ON x2;
%BETWEEN%
%OVERALL%
cb y ON w; s1-s2@0;
%cb#1%
[s1 s2];
%cb#2%
[s1 s2];

348
 Examples: Multilevel Mixture Modeling

s1 s2

x1 x2

Within

Between

y s1 s2

cb

In this example, the two-level mixture regression model for a continuous

dependent variable shown in the picture above is estimated. This
example is similar to Example 10.1 except that the categorical latent
variable is a between-level variable. This means that latent classes are
formed for clusters (between-level units) not individuals. In addition,
the regression slopes are random not fixed. In the within part of the
model, the random intercept is shown in the picture as a filled circle at

349
CHAPTER 10

the end of the arrow pointing to y. It is referred to as y on the between

level. The random slopes are shown as filled circles on the arrows from
x1 and x2 to y. They are referred to as s1 and s2 on the between level.
The random effects y, s1, and s2 are shown in circles in the between part
of the model because they are continuous latent variables that vary
across clusters (between-level units). In the between part of the model,
the arrows from cb to y, s1, and s2 indicate that the intercept of y and the
means of s1 and s2 vary across the classes of cb. In addition, the random
intercept y and the categorical latent variable cb are regressed on a
cluster-level covariate w. The random slopes s1 and s2 have no within-
class variance. Only their means vary across the classes of cb. This
implies that the distributions of s1 and s2 can be thought of as non-
parametric representations rather than normal distributions (Aitkin,
1999; Muthén & Asparouhov, 2009). Another example of a non-
parametric representation of a latent variable distribution is shown in
Example 7.26.

The BETWEEN option is used to identify the variables in the data set
that are measured on the cluster level and modeled only on the between
level and to identify between-level categorical latent variables. In this
example, the categorical latent variable cb is a between-level variable.
Between-level classes consist of clusters such as schools instead of
individuals. The PROCESSORS option of the ANALYSIS command is
used to specify that 2 processors will be used in the analysis for parallel
computations.

In the overall part of the within part of the model, the | symbol is used in
conjunction with TYPE=RANDOM to name and define the random
slope variables in the model. The name on the left-hand side of the |
symbol names the random slope variable. The statement on the right-
hand side of the | symbol defines the random slope variable. Random
slopes are defined using the ON option. The random slopes s1 and s2
are defined by the linear regressions of the dependent variable y on the
individual-level covariates x1 and x2. The within-level residual variance
in the regression of y on x is estimated as the default.

In the overall part of the between part of the model, the ON statement
describes the multinomial logistic regression of the categorical latent
variable cb on the cluster-level covariate w and the linear regression of
the random intercept y on the cluster-level covariate w. The variances of
the random slopes s1 and s2 are fixed at zero. In the class-specific parts

350
 Examples: Multilevel Mixture Modeling

of the between part of the model, the means of the random slopes are
specified to vary across the between-level classes of cb. The intercept of
the random intercept y varies across the between-level classes of cb as
the default.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, one dimension of integration is used with a
total of 15 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 10.1.

Following is an alternative specification of the MODEL command that is

simpler when the model has many covariates and when the variances of
the random slopes are zero:

MODEL:
%WITHIN%
%OVERALL%
y ON x1 x2;
%cb#1%
y ON x1 x2;
%cb#2%
y ON x1 x2;
%BETWEEN%
%OVERALL%
cb ON w;
y ON w;

In this specification, instead of the | statements, the random slopes are

represented as class-varying slopes in the class-specific parts of the
within part of the model. This specification makes it unnecessary to
refer to the means and variances of the random slopes in the between
part of the model.

351
CHAPTER 10

EXAMPLE 10.3: TWO-LEVEL MIXTURE REGRESSION FOR

A CONTINUOUS DEPENDENT VARIABLE WITH BETWEEN-
LEVEL CATEGORICAL LATENT CLASS INDICATORS FOR
A BETWEEN-LEVEL CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level mixture

regression for a continuous dependent
variable with between-level categorical
latent class indicators for a between-
level categorical latent variable
DATA: FILE = ex10.3.dat;
VARIABLE: NAMES ARE u1-u6 y x1 x2 w dummy clus;
USEVARIABLES = u1-w;
CATEGORICAL = u1-u6;
CLASSES = cb(2);
WITHIN = x1 x2;
BETWEEN = cb w u1-u6;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
y ON x1 x2;
%BETWEEN%
%OVERALL%
cb ON w;
y ON w;
OUTPUT: TECH1 TECH8;

352
 Examples: Multilevel Mixture Modeling

x1 x2

Within

Between

u1 u2 u3 u4 u5 u6 y

cb

In this example, the two-level mixture regression model for a continuous

dependent variable shown in the picture above is estimated. This
example is similar to Example 10.2 except that the between-level
categorical latent variable has between-level categorical latent class
indicators and the slopes are fixed. In the within part of the model, the
random intercept is shown in the picture as a filled circle at the end of
the arrow pointing to y. It is referred to as y on the between level. The

353
CHAPTER 10

random intercept y is shown in a circle in the between part of the model

because it is a continuous latent variable that varies across clusters
(between-level units). In the between part of the model, the arrow from
cb to y indicates that the intercept of y varies across the classes of cb. In
addition, the random intercept y and the categorical latent variable cb are
regressed on a cluster-level covariate w. The arrows from cb to u1, u2,
u3, u4, u5, and u6 indicate that these variables are between-level
categorical latent class indicators of the categorical latent variable cb.

In the overall part of the between part of the model, the first ON
statement describes the multinomial logistic regression of the categorical
latent variable cb on the cluster-level covariate w. The second ON
statement describes the linear regression of the random intercept y on the
cluster-level covariate w. The intercept of the random intercept y and
the thresholds of the between-level latent class indicators u1, u2, u3, u4,
u5, and u6 vary across the between-level classes of cb as the default.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, one dimension of integration is used with a
total of 15 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 10.1 and
10.2.

354
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.4: TWO-LEVEL CFA MIXTURE MODEL WITH

CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a two-level CFA

mixture model with continuous factor
indicators
DATA: FILE IS ex10.4.dat;
VARIABLE: NAMES ARE y1-y5 class clus;
USEVARIABLES = y1-y5;
CLASSES = c (2);
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
STARTS = 0;
MODEL:
%WITHIN%
%OVERALL%
fw BY y1-y5;
%BETWEEN%
%OVERALL%
fb BY y1-y5;
c#1*1;
%c#1%
[fb*2];
OUTPUT: TECH1 TECH8;

355
CHAPTER 10

y1 y2 y3 y4 y5

c fw
Within

Between

y1 y2 y3 y4 y5

c#1 fb

In this example, the two-level confirmatory factor analysis (CFA)

mixture model with continuous factor indicators in the picture above is
estimated. This example is the same as Example 7.17 except that it has
been extended to the multilevel framework. In the within part of the
model, the filled circles at the end of the arrows from the within factor
fw to y1, y2, y3, y4, and y5 represent random intercepts that vary across
clusters. The filled circle on the circle containing c represents the
random mean of c that varies across clusters. In the between part of the
model, the random intercepts are referred to as y1, y2, y3, y4, and y5
and the random mean is referred to as c#1 where they are shown in
circles because they are continuous latent variables that vary across
clusters. In the between part of the model, the random intercepts are
indicators of the between factor fb. In this model, the residual variances

356
 Examples: Multilevel Mixture Modeling

for the factor indicators in the between part of the model are zero. If
factor loadings are constrained to be equal across the within and the
between levels, this implies a model where the mean of the within factor
varies across the clusters. The between part of the model specifies that
the random mean c#1 of the categorical latent variable c and the between
factor fb are uncorrelated. Other modeling possibilities are for fb and
c#1 to be correlated, for fb to be regressed on c#1, or for c#1 to be
regressed on fb. Regressing c#1 on fb, however, leads to an internally
inconsistent model where the mean of fb is influenced by c at the same
time as c#1 is regressed on fb, leading to a reciprocal interaction.

In the overall part of the within part of the model, the BY statement
specifies that fw is measured by the factor indicators y1, y2, y3, y4, and
y5. The metric of the factor is set automatically by the program by
fixing the first factor loading to one. This option can be overridden.
The residual variances of the factor indicators are estimated and the
residuals are not correlated as the default. The variance of the factor is
estimated as the default.

In the overall part of the between part of the model, the BY statement
specifies that fb is measured by the random intercepts y1, y2, y3, y4, and
y5. The residual variances of the random intercepts are fixed at zero as
the default because they are often very small and each residual variance
requires one dimension of numerical integration. The variance of fb is
estimated as the default. A starting value of one is given to the variance
of the random mean of the categorical latent variable c referred to as
c#1. In the model for class 1 in the between part of the model, the mean
of fb is given a starting value of 2.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 10.1.

357
CHAPTER 10

EXAMPLE 10.5: TWO-LEVEL IRT MIXTURE ANALYSIS

WITH BINARY FACTOR INDICATORS AND A BETWEEN-
LEVEL CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level IRT

mixture analysis with binary factor
indicators and a between-level categorical
latent variable
DATA: FILE = ex10.5.dat;
VARIABLE: NAMES ARE u1-u8 dumb dum clus;
USEVARIABLES = u1-u8;
CATEGORICAL = u1-u8;
CLASSES = cb(2) c(2);
BETWEEN = cb;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
ALGORITHM = INTEGRATION;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
f BY u1-u8;
[f@0];
%BETWEEN%
%OVERALL%
%cb#1.c#1%
[u1$1-u8$1];
%cb#1.c#2%
[u1$1-u8$1];
%cb#2.c#1%
[u1$1-u8$1];
%cb#2.c#2%
[u1$1-u8$1];
MODEL c:
%WITHIN%
%c#1%
f;
%c#2%
f;
OUTPUT: TECH1 TECH8;

358
 Examples: Multilevel Mixture Modeling

u1 u2 u3 u4 u5 u6 u7 u8

c f

Within

Between

u1 u2 u3 u4 u5 u6 u7 u8

cb

In this example, the two-level item response theory (IRT) mixture model
with binary factor indicators shown in the picture above is estimated.
The model has both individual-level classes and between-level classes.
Individual-level classes consist of individuals, for example, students.
Between-level classes consist of clusters, for example, schools. The
within part of the model is similar to the single-level model in Example
7.27. In the within part of the model, an IRT mixture model is specified
where the factor indicators u1, u2, u3, u4, u5, u6, u7, and u8 have
thresholds that vary across the classes of the individual-level categorical

359
CHAPTER 10

latent variable c. The filled circles at the end of the arrows pointing to
the factor indicators show that the thresholds of the factor indicators are
random. They are referred to as u1, u2, u3, u4, u5, u6, u7, and u8 on the
between level. The random thresholds u1, u2, u3, u4, u5, u6, u7, and u8
are shown in circles in the between part of the model because they are
continuous latent variables that vary across clusters (between-level
units). The random thresholds have no within-class variance. They vary
across the classes of the between-level categorical latent variable cb.
For related models, see Asparouhov and Muthén (2008a).

In the class-specific part of the between part of the model, the random
thresholds are specified to vary across classes that are a combination of
the classes of the between-level categorical latent variable cb and the
individual-level categorical latent variable c. These classes are referred
to by combining the class labels using a period (.). For example, a
combination of class 1 of cb and class 1 of c is referred to as cb#1.c#1.
This represents an interaction between the two categorical latent
variables in their influence on the thresholds.

When a model has more than one categorical latent variable, MODEL
followed by a label is used to describe the analysis model for each
categorical latent variable. Labels are defined by using the names of the
categorical latent variables. In the model for the individual-level
categorical latent variable c, the variances of the factor f are allowed to
vary across the classes of c.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, one dimension of integration is used with a
total of 15 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 7.27, 10.1,
and 10.2.

360
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.6: TWO-LEVEL LCA WITH CATEGORICAL

LATENT CLASS INDICATORS WITH COVARIATES

TITLE: this is an example of a two-level LCA with

categorical latent class indicators with
covariates
DATA: FILE IS ex10.6.dat;
VARIABLE: NAMES ARE u1-u6 x w class clus;
USEVARIABLES = u1-u6 x w;
CATEGORICAL = u1-u6;
CLASSES = c (3);
WITHIN = x;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
MODEL:
%WITHIN%
%OVERALL%
c ON x;
%BETWEEN%
%OVERALL%
f BY c#1 c#2;
f ON w;
OUTPUT: TECH1 TECH8;

361
CHAPTER 10

u1 u2 u3 u4 u5 u6

Within

Between
c#1 c#2

In this example, the two-level latent class analysis (LCA) with

categorical latent class indicators and covariates shown in the picture
above is estimated (Vermunt, 2003). This example is similar to Example
7.12 except that it has been extended to the multilevel framework. In the

362
 Examples: Multilevel Mixture Modeling

within part of the model, the categorical latent variable c is regressed on

the individual-level covariate x. The filled circles at the end of the
arrow from x to c represent the random intercepts for classes 1 and 2 of
the categorical latent variable c which has three classes. The random
intercepts are referred to as c#1 and c#2 in the between part of the model
where they are shown in circles instead of squares because they are
continuous latent variables that vary across clusters. Because the
random intercepts in LCA are often highly correlated and to reduce the
dimensions of integration, a factor is used to represent the random
intercept variation. This factor is regressed on the cluster-level covariate
w.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, the latent class
indicators u1, u2, u3, u4, u5, and u6 are binary or ordered categorical
variables. The program determines the number of categories for each
indicator.

In the within part of the model, the ON statement describes the

multinomial logistic regression of the categorical latent variable c on the
individual-level covariate x when comparing classes 1 and 2 to class 3.
The intercepts of the random intercepts in the regression of c on x are
estimated as the default. The random intercept for class 3 is zero
because it is the reference class. In the between part of the model, the
BY statement specifies that f is measured by the random intercepts c#1
and c#2. The metric of the factor is set automatically by the program by
fixing the first factor loading to one. The residual variances of the
random intercepts are fixed at zero as the default. The ON statement
describes the linear regression of the between factor f on the cluster-
level covariate w.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, one dimension of integration is used with 15
integration points. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 10.1.

363
CHAPTER 10

EXAMPLE 10.7: TWO-LEVEL LCA WITH CATEGORICAL

LATENT CLASS INDICATORS AND A BETWEEN-LEVEL
CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level LCA with

categorical latent class indicators and a
between-level categorical latent variable
DATA: FILE = ex10.7.dat;
VARIABLE: NAMES ARE u1-u10 dumb dumw clus;
USEVARIABLES = u1-u10;
CATEGORICAL = u1-u10;
CLASSES = cb(5) cw(4);
WITHIN = u1-u10;
BETWEEN = cb;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
PROCESSORS = 2;
STARTS = 100 10;
MODEL:
%WITHIN%
%OVERALL%
%BETWEEN%
%OVERALL%
cw#1-cw#3 ON cb;
MODEL cw:
%WITHIN%
%cw#1%
[u1$1-u10$1];
[u1$2-u10$2];
%cw#2%
[u1$1-u10$1];
[u1$2-u10$2];
%cw#3%
[u1$1-u10$1];
[u1$2-u10$2];
%cw#4%
[u1$1-u10$1];
[u1$2-u10$2];
OUTPUT: TECH1 TECH8;

364
 Examples: Multilevel Mixture Modeling

u1 u2 u3 u4 u5 u6 u7 u8 u9 u10

cw

Within

Between

cw#1 cw#2 cw#3

cb

In this example, the two-level latent class analysis (LCA) with

categorical latent class indicators shown in the picture above is
estimated. This example is similar to Example 10.6 except that the
between level random means are influenced by the between-level
categorical latent variable cb. In the within part of the model, the filled
circles represent the three random means of the four classes of the
individual-level categorical latent variable cw. They are referred to as
cw#1, cw#2, and cw#3 on the between level. The random means are
shown in circles in the between part of the model because they are
continuous latent variables that vary across clusters (between-level
units). The random means have means that vary across the classes of the
categorical latent variable cb but the within-class variances of the
random means are zero (Bijmolt, Paas, & Vermunt, 2004).

365
CHAPTER 10

In the overall part of the between part of the model, the ON statement
describes the linear regressions of cw#1, cw#2, and cw#3 on the
between-level categorical latent variable cb. This regression implies that
the means of these random means vary across the classes of the
categorical latent variable cb.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 10.1, 10.2,
and 10.6.

366
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.8: TWO-LEVEL GROWTH MODEL FOR A

CONTINUOUS OUTCOME (THREE-LEVEL ANALYSIS) WITH
A BETWEEN-LEVEL CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level growth

model for a continuous outcome (three-
level analysis) with a between-level
categorical latent variable
DATA: FILE = ex10.8.dat;
VARIABLE: NAMES ARE y1-y4 x w dummy clus;
USEVARIABLES = y1-w;
CLASSES = cb(2);
WITHIN = x;
BETWEEN = cb w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE RANDOM;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4 (1);
iw sw ON x;
s | sw ON iw;
%BETWEEN%
%OVERALL%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib sb ON w;
cb ON w;
s@0;
%cb#1%
[ib sb s];
%cb#2%
[ib sb s];
OUTPUT: TECH1 TECH8;

367
CHAPTER 10

y1 y2 y3 y4

iw sw
s

Within

Between

y1 y2 y3 y4

ib sb

cb s

368
 Examples: Multilevel Mixture Modeling

In this example, the two-level growth model for a continuous outcome

(three-level analysis) shown in the picture above is estimated. This
example is similar to Example 9.12 except that a random slope is
estimated in the within-level regression of the slope growth factor on the
intercept growth factor and a between-level latent class variable cb is
part of the model. This means that latent classes are formed for clusters
(between-level units) not individuals. In the within part of the model,
the random slope is shown in the picture as a filled circle on the arrow
from iw to sw. It is referred to as s on the between level. The random
slope s is shown in a circle in the between part of the model because it is
a continuous latent variable that varies across clusters (between-level
units). In the between part of the model, the arrows from cb to ib, sb,
and s indicate that the intercepts of ib and sb and the mean of s vary
across the classes of cb. In addition, the categorical latent variable cb is
regressed on a cluster-level covariate w. The random slope s has no
within-class variance. Only its mean varies across the classes of cb.
This implies that the distributions of s can be thought of as a non-
parametric representation rather than a normal distribution (Aitkin,
1999; Muthén & Asparouhov, 2007).

In the overall part of the within part of the model, the | statement is used
to name and define the random slope s which is used in the between part
of the model. In the overall part of the between part of the model, the
second ON statement describes the multinomial logistic regression of the
categorical latent variable cb on a cluster-level covariate w. The
variance of the random slope s is fixed at zero. In the class-specific parts
of the between part of the model, the intercepts of the growth factors ib
and sb and the mean of the random slope s are specified to vary across
the between-level classes of cb.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 9.12, 10.1,
and 10.2.

369
CHAPTER 10

Following is an alternative specification of the MODEL command that is

simpler when the variances of the random slopes are zero:

MODEL:
%WITHIN%
%OVERALL%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4 (1);
iw ON x;
sw ON x iw;
%cb#1%
sw ON iw;
%cb#2%
sw ON iw;
%BETWEEN%
%OVERALL%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib sb ON w;
cb ON w;
%cb#1%
[ib sb];
%cb#2%
[ib sb];

In this specification, instead of the | statement, the random slope is

represented as class-varying slopes in the class-specific parts of the
within part of the model. This specification makes it unnecessary to
refer to the means and variances of the random slopes in the between
part of the model.

370
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.9: TWO-LEVEL GMM FOR A CONTINUOUS

OUTCOME (THREE-LEVEL ANALYSIS)

TITLE: this is an example of a two-level GMM for

a continuous outcome (three-level
analysis)
DATA: FILE IS ex10.9.dat;
VARIABLE: NAMES ARE y1-y4 x w class clus;
USEVARIABLES = y1-y4 x w;
CLASSES = c (2);
WITHIN = x;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
STARTS = 0;
MODEL:
%WITHIN%
%OVERALL%
iw sw | y1@0 y2@1 y3@2 y4@3;
iw sw ON x;
c ON x;
%BETWEEN%
%OVERALL%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib sb ON w;
sb@0;
c#1 ON w;
c#1*1;
%c#1%
[ib sb];
%c#2%
[ib*3 sb*1];
OUTPUT: TECH1 TECH8;

371
CHAPTER 10

y1 y2 y3 y4

iw sw

Within

Between

y1 y2 y3 y4

ib sb

c#1

In this example, the two-level growth mixture model (GMM; Muthén,

2004; Muthén & Asparouhov, 2009) for a continuous outcome (three-
level analysis) shown in the picture above is estimated. This example is
similar to Example 8.1 except that it has been extended to the multilevel

372
 Examples: Multilevel Mixture Modeling

framework. In the within part of the model, the filled circles at the end
of the arrows from the within growth factors iw and sw to y1, y2, y3, and
y4 represent random intercepts that vary across clusters. The filled
circle at the end of the arrow from x to c represents a random intercept.
The random intercepts are referred to in the between part of the model as
y1, y2, y3, y4, and c#1. In the between-part of the model, the random
intercepts are shown in circles because they are continuous latent
variables that vary across clusters.

In the within part of the model, the | statement names and defines the
within intercept and slope factors for the growth model. The names iw
and sw on the left-hand side of the | symbol are the names of the
intercept and slope growth factors, respectively. The values on the right-
hand side of the | symbol are the time scores for the slope growth factor.
The time scores of the slope growth factor are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
growth factor as an initial status factor. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are
estimated and allowed to be different across time and the residuals are
not correlated as the default. The first ON statement describes the linear
regressions of the growth factors on the individual-level covariate x.
The residual variances of the growth factors are free to be estimated as
the default. The residuals of the growth factors are correlated as the
default because residuals are correlated for latent variables that do not
influence any other variable in the model except their own indicators.
The second ON statement describes the multinomial logistic regression
of the categorical latent variable c on the individual-level covariate x
when comparing class 1 to class 2. The intercept in the regression of c
on x is estimated as the default.

In the overall model in the between part of the model, the | statement
names and defines the between intercept and slope factors for the growth
model. The names ib and sb on the left-hand side of the | symbol are the
names of the intercept and slope growth factors, respectively. The
values of the right-hand side of the | symbol are the time scores for the
slope growth factor. The time scores of the slope growth factor are fixed
at 0, 1, 2, and 3 to define a linear growth model with equidistant time
points. The zero time score for the slope growth factor at time point one
defines the intercept growth factor as an initial status factor. The

373
CHAPTER 10

coefficients of the intercept growth factor are fixed at one as part of the
growth model parameterization. The residual variances of the outcome
variables are fixed at zero on the between level in line with conventional
multilevel growth modeling. This can be overridden. The first ON
statement describes the linear regressions of the growth factors on the
cluster-level covariate w. The residual variance of the intercept growth
factor is free to be estimated as the default. The residual variance of the
slope growth factor is fixed at zero because it is often small and each
residual variance requires one dimension of numerical integration.
Because the slope growth factor residual variance is fixed at zero, the
residual covariance between the growth factors is automatically fixed at
zero. The second ON statement describes the linear regression of the
random intercept c#1 of the categorical latent variable c on the cluster-
level covariate w. A starting value of one is given to the residual
variance of the random intercept of the categorical latent variable c
referred to as c#1.

In the parameterization of the growth model shown here, the intercepts

of the outcome variable at the four time points are fixed at zero as the
default. The growth factor intercepts are estimated as the default in the
between part of the model. In the model for class 2 in the between part
of the model, the mean of ib and sb are given a starting value of zero in
class 1 and three and one in class 2.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Example 10.1.

374
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.10: TWO-LEVEL GMM FOR A CONTINUOUS

OUTCOME (THREE-LEVEL ANALYSIS) WITH A BETWEEN-
LEVEL CATEGORICAL LATENT VARIABLE

TITLE: this is an example of a two-level GMM for

a continuous outcome (three-level
analysis) with a between-level categorical
latent variable
DATA: FILE = ex10.10.dat;
VARIABLE: NAMES ARE y1-y4 x w dummyb dummy clus;
USEVARIABLES = y1-w;
CLASSES = cb(2) c(2);
WITHIN = x;
BETWEEN = cb w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
iw sw | y1@0 y2@1 y3@2 y4@3;
iw sw ON x;
c ON x;
%BETWEEN%
%OVERALL%
ib sb | y1@0 y2@1 y3@2 y4@3;
ib2 | y1-y4@1;
y1-y4@0;
ib sb ON w;
c#1 ON w;
sb@0; c#1;
ib2@0;
cb ON w;
MODEL c:
%BETWEEN%
%c#1%
[ib sb];
%c#2%
[ib sb];
MODEL cb:
%BETWEEN%
%cb#1%
[ib2@0];
%cb#2%
[ib2];
OUTPUT: TECH1 TECH8;

375
CHAPTER 10

y1 y2 y3 y4

iw sw

Within

Between

y1 y2 y3 y4

ib sb ib2

c#1 cb

In this example, the two-level growth mixture model (GMM; Muthén &
Asparouhov, 2009) for a continuous outcome (three-level analysis)
shown in the picture above is estimated. This example is similar to
Example 10.9 except that a between-level categorical latent variable cb
has been added along with a second between-level intercept growth
factor ib2. The second intercept growth factor is added to the model so

376
 Examples: Multilevel Mixture Modeling

that the intercept growth factor mean can vary across not only the classes
of the individual-level categorical latent variable c but also across the
classes of the between-level categorical latent variable cb. Individual-
level classes consist of individuals, for example, students. Between-
level classes consist of clusters, for example, schools.

In the overall part of the between part of the model, the second |
statement names and defines the second between-level intercept growth
factor ib2. This growth factor is used to represent differences in
intercept growth factor means across the between-level classes of the
categorical latent variable cb.

When a model has more than one categorical latent variable, MODEL
followed by a label is used to describe the analysis model for each
categorical latent variable. Labels are defined by using the names of the
categorical latent variables. In the model for the individual-level
categorical latent variable c, the intercepts of the intercept and slope
growth factors ib and sb are allowed to vary across the classes of the
individual-level categorical latent variable c. In the model for the
between-level categorical latent variable cb, the means of the intercept
growth factor ib2 are allowed to vary across clusters (between-level
units). The mean in one class is fixed at zero for identification purposes.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 10.1, 10.2,
and 10.4.

377
CHAPTER 10

EXAMPLE 10.11: TWO-LEVEL LCGA FOR A THREE-

CATEGORY OUTCOME

TITLE: this is an example of a two-level LCGA for

a three-category outcome
DATA: FILE IS ex10.11.dat;
VARIABLE: NAMES ARE u1-u4 class clus;
USEVARIABLES = u1-u4;
CATEGORICAL = u1-u4;
CLASSES = c(2);
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
MODEL:
%WITHIN%
%OVERALL%
i s | u1@0 u2@1 u3@2 u4@3;
i-s@0;
%c#1%
[i*1 s*1];
%c#2%
[i@0 s];
%BETWEEN%
%OVERALL%
c#1*1;
[u1$1-u4$1*1] (1);
[u1$2-u4$2*1.5] (2);
OUTPUT: TECH1 TECH8;

378
 Examples: Multilevel Mixture Modeling

u1 u2 u3 u4

i s

c Within

c#1 Between

In this example, the two-level latent class growth analysis (LCGA)

shown in the picture above is estimated. This example is the same as
Example 8.10 except that it has been extended to the multilevel
framework. A growth model is not specified in the between part of the
model because the variances of the growth factors i and s are zero in
LCGA. The filled circle on the circle containing the categorical latent
variable c represents the random mean of c. In the between part of the
model, the random mean is shown in a circle because it is a continuous
latent variable that varies across clusters.

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. In the example above, the latent class
indicators u1, u2, u3, u4, u5, and u6 are binary or ordered categorical
variables. The program determines the number of categories for each
indicator. In this example, u1, u2, u3, and u4 are three-category
variables.

379
CHAPTER 10

In the overall part of the of the within part of the model, the variances of
the growth factors i and s are fixed at zero because latent class growth
analysis has no within class variability. In the overall part of the of the
between part of the model, the two thresholds for the outcome are held
equal across the four time points. The growth factor means are specified
in the within part of the model because there are no between growth
factors.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, one dimension of integration is used with 15
integration points. The ESTIMATOR option of the ANALYSIS
command can be used to select a different estimator. An explanation of
the other commands can be found in Example 10.1.

EXAMPLE 10.12: TWO-LEVEL LTA WITH A COVARIATE

TITLE: this is an example of a two-level LTA with

a covariate
DATA: FILE = ex10.12.dat;
VARIABLE: NAMES ARE u11-u14 u21-u24 x w dum1 dum2
clus;
USEVARIABLES = u11-w;
CATEGORICAL = u11-u14 u21-u24;
CLASSES = c1(2) c2(2);
WITHIN = x;
BETWEEN = w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
c2 ON c1 x;
c1 ON x;
%BETWEEN%
%OVERALL%
c1#1 ON w;
c2#1 ON c1#1 w;
c1#1 c2#1;

380
 Examples: Multilevel Mixture Modeling

MODEL c1:
%BETWEEN%
%c1#1%
[u11$1-u14$1] (1-4);
%c1#2%
[u11$1-u14$1] (5-8);
MODEL c2:
%BETWEEN%
%c2#1%
[u21$1-u24$1] (1-4);
%c2#2%
[u21$1-u24$1] (5-8);
OUTPUT: TECH1 TECH8;

u11 u12 u13 u14 u21 u22 u23 u24

c1 c2

x Within

Between
c1#1 c2#1

381
CHAPTER 10

In this example, the two-level latent transition analysis (LTA) with a

covariate shown in the picture above is estimated. This example is
similar to Example 8.13 except that the categorical latent variables are
allowed to have random intercepts that vary on the between level. This
model is described in Asparouhov and Muthén (2008a). In the within
part of the model, the random intercepts are shown in the picture as
filled circles at the end of the arrows pointing to c1 and c2. They are
referred to as c1#1 and c2#1 on the between level. The random
intercepts c1#1 and c2#1 are shown in circles in the between part of the
model because they are continuous latent variables that vary across
clusters (between-level units).

In the overall part of the between part of the model, the first ON
statement describes the linear regression of the random intercept c1#1 on
a cluster-level covariate w. The second ON statement describes the
linear regression of the random intercept c2#1 on the random intercept
c1#1 and the cluster-level covariate w. The residual variances of the
random intercepts c1#1 and c2#1 are estimated instead of being fixed at
the default value of zero.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors using a numerical integration algorithm.
Note that numerical integration becomes increasingly more
computationally demanding as the number of factors and the sample size
increase. In this example, two dimensions of integration are used with a
total of 225 integration points. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 8.13, 10.1,
and 10.2.

382
 Examples: Multilevel Mixture Modeling

EXAMPLE 10.13: TWO-LEVEL LTA WITH A COVARIATE

AND A BETWEEN-LEVEL CATEGORICAL LATENT
VARIABLE

TITLE: this is an example of a two-level LTA with

a covariate and a between-level
categorical latent variable
DATA: FILE = ex10.13.dat;
VARIABLE: NAMES ARE u11-u14 u21-u24 x w dumb dum1
dum2 clus;
USEVARIABLES = u11-w;
CATEGORICAL = u11-u14 u21-u24;
CLASSES = cb(2) c1(2) c2(2);
WITHIN = x;
BETWEEN = cb w;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL MIXTURE;
PROCESSORS = 2;
MODEL:
%WITHIN%
%OVERALL%
c2 ON c1 x;
c1 ON x;
%BETWEEN%
%OVERALL%
c1#1 ON cb;
c2#1 ON cb;
cb ON w;
MODEL cb:
%WITHIN%
%cb#1%
c2 ON c1;
MODEL c1:
%BETWEEN%
%c1#1%
[u11$1-u14$1] (1-4);
%c1#2%
[u11$1-u14$1] (5-8);
MODEL c2:
%BETWEEN%
%c2#1%
[u21$1-u24$1] (1-4);
%c2#2%
[u21$1-u24$1] (5-8);
OUTPUT: TECH1 TECH8;

383
CHAPTER 10

u11 u12 u13 u14 u21 u22 u23 u24

s
c1 c2

x Within

Between
c1#1 c2#1

cb s

In this example, the two-level latent transition analysis (LTA) with a

covariate shown in the picture above is estimated. This example is
similar to Example 10.12 except that a between-level categorical latent
variable cb has been added, a random slope has been added, and the
random intercepts and random slope have no variance within the classes
of the between-level categorical latent variable cb (Asparouhov &
Muthén, 2008a). In the within part of the model, the random intercepts
are shown in the picture as filled circles at the end of the arrows pointing
to c1 and c2. The random slope is shown as a filled circle on the arrow
from c1 to c2. In the between part of the model, the random intercepts
are referred to as c1#1 and c2#1 and the random slope is referred to as s.
The random intercepts c1#1 and c2#1 and the random slope s are shown
in circles in because they are continuous latent variables that vary across

384
 Examples: Multilevel Mixture Modeling

clusters (between-level units). In the between part of the model, the

arrows from cb to c1#1, c2#1, and s indicate that the means of c1#1,
c2#1, and s vary across the classes of cb.

In the overall part of the between part of the model, the first two ON
statements describe the linear regressions of c1#1 and c2#1 on the
between-level categorical latent variable cb. These regressions imply
that the means of the random intercepts vary across the classes of the
categorical latent variable cb. The variances of c1#1 and c2#1 within
the cb classes are zero as the default.

When a model has more than one categorical latent variable, MODEL
followed by a label is used to describe the analysis model for each
categorical latent variable. Labels are defined by using the names of the
categorical latent variables. In the class-specific part of the within part
of the model for the between-level categorical latent variable cb, the ON
statement describes the multinomial regression of c2 on c1. This implies
that the random slope s varies across the classes of cb. The within-class
variance of s is zero as the default.

The default estimator for this type of analysis is maximum likelihood

with robust standard errors. The ESTIMATOR option of the
ANALYSIS command can be used to select a different estimator. An
explanation of the other commands can be found in Examples 8.13, 10.1,
10.2, and 10.12.

385
CHAPTER 10

386
 Examples: Missing Data Modeling And Bayesian Analysis

CHAPTER 11
EXAMPLES: MISSING DATA
MODELING AND BAYESIAN
ANALYSIS
Mplus provides estimation of models with missing data using both
frequentist and Bayesian analysis. Descriptive statistics and graphics are
available for understanding dropout in longitudinal studies. Bayesian
analysis provides multiple imputation for missing data as well as
plausible values for latent variables.

With frequentist analysis, Mplus provides maximum likelihood

estimation under MCAR (missing completely at random), MAR (missing
at random), and NMAR (not missing at random) for continuous,
censored, binary, ordered categorical (ordinal), unordered categorical
(nominal), counts, or combinations of these variable types (Little &
Rubin, 2002). MAR means that missingness can be a function of
observed covariates and observed outcomes. For censored and
categorical outcomes using weighted least squares estimation,
missingness is allowed to be a function of the observed covariates but
not the observed outcomes. When there are no covariates in the model,
this is analogous to pairwise present analysis. Non-ignorable missing
data (NMAR) modeling is possible using maximum likelihood
estimation where categorical outcomes are indicators of missingness and
where missingness can be predicted by continuous and categorical latent
variables (Muthén, Jo, & Brown, 2003; Muthén et al., 2011). This
includes selection models, pattern-mixture models, and shared-parameter
models (see, e.g., Muthén et al., 2011). In all models, observations with
missing data on covariates are deleted because models are estimated
conditional on the covariates. Covariate missingness can be modeled if
the covariates are brought into the model and distributional assumptions
such as normality are made about them. With missing data, the standard
errors for the parameter estimates are computed using the observed
information matrix (Kenward & Molenberghs, 1998). Bootstrap
standard errors and confidence intervals are also available with missing
data.

387
CHAPTER 11

With Bayesian analysis, modeling with missing data gives

asymptotically the same results as maximum-likelihood estimation under
MAR. Multiple imputation of missing data using Bayesian analysis
(Rubin, 1987; Schafer, 1997) is also available. For an overview, see
Enders (2010). Both unrestricted H1 models and restricted H0 models
can be used for imputation. Several different algorithms are available
for H1 imputation, including sequential regression, also referred to as
chained regression, in line with Raghunathan et al. (2001); see also van
Buuren (2007). Multiple imputation of plausible values for latent
variables is provided. For applications of plausible values in the context
of Item Response Theory, see Mislevy et al. (1992) and von Davier et al.
(2009). Multiple data sets generated using multiple imputation can be
analyzed with frequentist estimators using a special feature of Mplus.
Parameter estimates are averaged over the set of analyses, and standard
errors are computed using the average of the standard errors over the set
of analyses and the between analysis parameter estimate variation
(Rubin, 1987; Schafer, 1997). A chi-square test of overall model fit is
provided with maximum-likelihood estimation (Asparouhov & Muthén,
2008c; Enders, 2010).

Following is the set of frequentist examples included in this chapter:

• 11.1: Growth model with missing data using a missing data

correlate
• 11.2: Descriptive statistics and graphics related to dropout in a
longitudinal study
• 11.3: Modeling with data not missing at random (NMAR) using the
Diggle-Kenward selection model*
• 11.4: Modeling with data not missing at random (NMAR) using a
pattern-mixture model

Following is the set of Bayesian examples included in this chapter:

• 11.5: Multiple imputation for a set of variables with missing values

• 11.6: Multiple imputation followed by the estimation of a growth
model using maximum likelihood
• 11.7: Multiple imputation of plausible values using Bayesian
estimation of a growth model
• 11.8: Multiple imputation using a two-level factor model with
categorical outcomes followed by the estimation of a growth model

388
 Examples: Missing Data Modeling And Bayesian Analysis

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

EXAMPLE 11.1: GROWTH MODEL WITH MISSING DATA

USING A MISSING DATA CORRELATE

TITLE: this is an example of a linear growth

model with missing data on a continuous
outcome using a missing data correlate to
improve the plausibility of MAR
DATA: FILE = ex11.1.dat;
VARIABLE: NAMES = x1 x2 y1-y4 z;
USEVARIABLES = y1-y4;
MISSING = ALL (999);
AUXILIARY = (m) z;
ANALYSIS: ESTIMATOR = ML;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
OUTPUT: TECH1;

y1 y2 y3 y4

i s

In this example, the linear growth model at four time points with missing
data on a continuous outcome shown in the picture above is estimated
using a missing data correlate. The missing data correlate is not part of
the growth model but is used to improve the plausibility of the MAR
assumption of maximum likelihood estimation (Collins, Schafer, &
Kam, 2001; Graham, 2003; Enders, 2010). The missing data correlate is
allowed to correlate with the outcome while providing the correct

389
CHAPTER 11

number of parameters and chi-square test for the analysis model as

described in Asparouhov and Muthén (2008b).

TITLE: this is an example of a linear growth

model with missing data on a continuous
outcome using a missing data correlate to
improve the plausibility of MAR

The TITLE command is used to provide a title for the analysis. The title
is printed in the output just before the Summary of Analysis.

DATA: FILE = ex11.1.dat;

The DATA command is used to provide information about the data set
to be analyzed. The FILE option is used to specify the name of the file
that contains the data to be analyzed, ex11.1.dat. Because the data set is
in free format, the default, a FORMAT statement is not required.

VARIABLE: NAMES = x1 x2 y1-y4 z;

USEVARIABLES = y1-y4;
MISSING = ALL (999);
AUXILIARY = (m) z;

The VARIABLE command is used to provide information about the

variables in the data set to be analyzed. The NAMES option is used to
assign names to the variables in the data set. The data set in this
example contains seven variables: x1, x2, y1, y2, y3, y4, and z. Note
that the hyphen can be used as a convenience feature in order to generate
a list of names. If not all of the variables in the data set are used in the
analysis, the USEVARIABLES option can be used to select a subset of
variables for analysis. Here the variables y1, y2, y3, and y4 have been
selected for analysis. They represent the outcome measured at four
equidistant occasions. The MISSING option is used to identify the
values or symbol in the analysis data set that are treated as missing or
invalid. The keyword ALL specifies that all variables in the analysis
data set have the missing value flag of 999. The AUXILIARY option
using the m setting is used to identify a set of variables that will be used
as missing data correlates in addition to the analysis variables. In this
example, the variable z is a missing data correlate.

390
 Examples: Missing Data Modeling And Bayesian Analysis

ANALYSIS: ESTIMATOR = ML;

The ANALYSIS command is used to describe the technical details of the

analysis. The ESTIMATOR option is used to specify the estimator to be
used in the analysis. By specifying ML, maximum likelihood estimation
is used.

MODEL: i s | y1@0 y2@1 y3@2 y4@3;

The MODEL command is used to describe the model to be estimated.

The | symbol is used to name and define the intercept and slope factors
in a growth model. The names i and s on the left-hand side of the |
symbol are the names of the intercept and slope growth factors,
respectively. The statement on the right-hand side of the | symbol
specifies the outcome and the time scores for the growth model. The
time scores for the slope growth factor are fixed at 0, 1, 2, and 3 to
define a linear growth model with equidistant time points. The zero time
score for the slope growth factor at time point one defines the intercept
growth factor as an initial status factor. The coefficients of the intercept
growth factor are fixed at one as part of the growth model
parameterization. The residual variances of the outcome variables are
estimated and allowed to be different across time and the residuals are
not correlated as the default.

In the parameterization of the growth model shown here, the intercepts

of the outcome variables at the four time points are fixed at zero as the
default. The means and variances of the growth factors are estimated as
the default, and the growth factor covariance is estimated as the default
because the growth factors are independent (exogenous) variables. The
default estimator for this type of analysis is maximum likelihood. The
ESTIMATOR option of the ANALYSIS command can be used to select
a different estimator.

OUTPUT: TECH1;

The OUTPUT command is used to request additional output not

included as the default. The TECH1 option is used to request the arrays
containing parameter specifications and starting values for all free
parameters in the model.

391
CHAPTER 11

EXAMPLE 11.2: DESCRIPTIVE STATISTICS AND GRAPHICS

RELATED TO DROPOUT IN A LONGITUDINAL STUDY

TITLE: this is an example of descriptive

statistics and graphics related to dropout
in a longitudinal study
DATA: FILE = ex11.2.dat;
VARIABLE: NAMES = z1-z5 y0 y1-y5;
USEVARIABLES = z1-z5 y0-y5 d1-d5;
MISSING = ALL (999);
DATA MISSING:
NAMES = y0-y5;
TYPE = DDROPOUT;
BINARY = d1-d5;
DESCRIPTIVE = y0-y5 | * z1-z5;
ANALYSIS: TYPE = BASIC;
PLOT: TYPE = PLOT2;
SERIES = y0-y5(*);

In this example, descriptive statistics and graphics related to dropout in a

longitudinal study are obtained. The descriptive statistics show the
mean and standard deviation for sets of variables related to the outcome
for those who drop out or not before the next time point. These means
are plotted to help in understanding dropout.

The DATA MISSING command is used to create a set of binary

variables that are indicators of missing data or dropout for another set of
variables. Dropout indicators can be scored as discrete-time survival
indicators or dummy dropout indicators. The NAMES option identifies
the set of variables that are used to create a set of binary variables that
are indicators of missing data. In this example, they are y0, y1, y2, y3,
y4, and y5. These variables must be variables from the NAMES
statement of the VARIABLE command. The TYPE option is used to
specify how missingness is coded. In this example, the DDROPOUT
setting specifies that binary dummy dropout indicators will be used. The
BINARY option is used to assign the names d1, d2, d3, d4, and d5 to the
new set of binary variables. There is one less dummy dropout indicator
than there are time points. The DESCRIPTIVE option is used in
conjunction with TYPE=BASIC of the ANALYSIS command and the
DDROPOUT setting to specify the sets of variables for which additional
descriptive statistics are computed. For each variable, the mean and
standard deviation are computed using all observations without missing

392
 Examples: Missing Data Modeling And Bayesian Analysis

on the variable and for those who drop out or not before the next time
point.

The PLOT command is used to request graphical displays of observed

data and analysis results. These graphical displays can be viewed after
the analysis is completed using a post-processing graphics module. The
TYPE option is used to specify the types of plots that are requested. The
setting PLOT2 is used to obtain missing data plots of dropout means and
sample means. The SERIES option is used to list the names of the set of
variables to be used in plots where the values are connected by a line.
The asterisk (*) in parentheses following the variable names indicates
that the values 1, 2, 3, 4, 5, and 6 will be used on the x-axis. An
explanation of the other commands can be found in Example 11.1.

EXAMPLE 11.3: MODELING WITH DATA NOT MISSING AT

RANDOM (NMAR) USING THE DIGGLE-KENWARD
SELECTION MODEL

TITLE: this is an example of modeling with data

not missing at random (NMAR) using the
Diggle-Kenward selection model
DATA: FILE = ex11.3.dat;
VARIABLE: NAMES = z1-z5 y0 y1-y5;
USEVARIABLES = y0-y5 d1-d5;
MISSING = ALL (999);
CATEGORICAL = d1-d5;
DATA MISSING:
NAMES = y0-y5;
TYPE = SDROPOUT;
BINARY = d1-d5;
ANALYSIS: ESTIMATOR = ML;
ALGORITHM = INTEGRATION;
INTEGRATION = MONTECARLO;
PROCESSORS = 2;

393
CHAPTER 11

MODEL: i s | y0@0 y1@1 y2@2 y3@3 y4@4 y5@5;

d1 ON y0 (1)
y1 (2);
d2 ON y1 (1)
y2 (2);
d3 ON y2 (1)
y3 (2);
d4 ON y3 (1)
y4 (2);
d5 ON y4 (1)
y5 (2);
OUTPUT: TECH1;

y0 y1 y2 y3 y4 y5

i s q

d1 d2 d3 d4 d5

In this example, the linear growth model at six time points with missing
data on a continuous outcome shown in the picture above is estimated.
The data are not missing at random because dropout is related to both
past and current outcomes where the current outcome is missing for
those who drop out. In the picture above, y1 through y5 are shown in
both circles and squares where circles imply that dropout has occurred
and squares imply that dropout has not occurred. The Diggle-Kenward
selection model (Diggle & Kenward, 1994) is used to jointly estimate a

394
 Examples: Missing Data Modeling And Bayesian Analysis

growth model for the outcome and a discrete-time survival model for the
dropout indicators (see also Muthén et al, 2011).

In this example, the SDROPOUT setting of the TYPE option specifies

that binary discrete-time (event-history) survival dropout indicators will
be used. In the ANALYSIS command, ALGORITHM=INTEGRATION
is required because latent continuous variables corresponding to missing
data on the outcome influence the binary dropout indicators.
INTEGRATION=MONTECARLO is required because the dimensions
of integration vary across observations. In the MODEL command, the
ON statements specify the logistic regressions of a dropout indicator at a
given time point regressed on the outcome at the previous time point and
the outcome at the current time point. The outcome at the current time
point is latent, missing, for those who have dropped out since the last
time point. The logistic regression coefficients are held equal across
time. An explanation of the other commands can be found in Examples
11.1 and 11.2.

EXAMPLE 11.4: MODELING WITH DATA NOT MISSING AT

RANDOM (NMAR) USING A PATTERN-MIXTURE MODEL

TITLE: this is an example of modeling with data

not missing at random (NMAR) using a
pattern-mixture model
DATA: FILE = ex11.4.dat;
VARIABLE: NAMES = z1-z5 y0 y1-y5;
USEVARIABLES = y0-y5 d1-d5;
MISSING = ALL (999);
DATA MISSING:
NAMES = y0-y5;
TYPE = DDROPOUT;
BINARY = d1-d5;
MODEL: i s | y0@0 y1@1 y2@2 y3@3 y4@4 y5@5;
i ON d1-d5;
s ON d3-d5;
s ON d1 (1);
s ON d2 (1);
OUTPUT: TECH1;

395
CHAPTER 11

y0 y1 y2 y3 y4 y5

i s q

d1 d2 d3 d4 d5

In this example, the linear growth model at six time points with missing
data on a continuous outcome shown in the picture above is estimated.
The data are not missing at random because dropout is related to both
past and current outcomes where the current outcome is missing for
those who drop out. A pattern-mixture model (Little, 1995; Hedeker &
Gibbons, 1997; Demirtas & Schafer, 2003) is used to estimate a growth
model for the outcome with binary dummy dropout indicators used as
covariates (see also Muthén et al, 2011).

The MODEL command is used to specify that the dropout indicators

influence the growth factors. The ON statements specify the linear
regressions of the intercept and slope growth factors on the dropout
indicators. The coefficient in the linear regression of s on d1 is not
identified because the outcome is observed only at the first time point for
the dropout pattern with d1 equal to one. This regression coefficient is
held equal to the linear regression of s on d2 for identification purposes.
An explanation of the other commands can be found in Examples 11.1
and 11.2.

396
 Examples: Missing Data Modeling And Bayesian Analysis

EXAMPLE 11.5: MULTIPLE IMPUTATION FOR A SET OF

VARIABLES WITH MISSING VALUES

TITLE: this is an example of multiple imputation

for a set of variables with missing values
DATA: FILE = ex11.5.dat;
VARIABLE: NAMES = x1 x2 y1-y4 v1-v50 z1-z5;
USEVARIABLES = x1 x2 y1-y4 z1-z5;
AUXILIARY = v1-v10;
MISSING = ALL (999);
DATA IMPUTATION:
IMPUTE = y1-y4 x1 (c) x2;
NDATASETS = 10;
SAVE = missimp*.dat;
ANALYSIS: TYPE = BASIC;
OUTPUT: TECH8;

In this example, multiple imputation for a set of variables with missing

values is carried out using Bayesian analysis (Rubin, 1987; Schafer,
1997). The NAMES option is used to assign names to the variables in
the original data set. The variables on the USEVARIABLES list are
used to create the imputed data sets. The AUXILIARY option is used to
specify the variables that are not used in the data imputation but that will
be saved with the imputed data sets. In the DATA IMPUTATION
command, the IMPUTE option is used to specify the variables for which
missing values will be imputed. A c in parentheses following a variable
indicates that it is categorical. The NDATASETS option is used to
specify the number of imputed data sets to create. In this example, ten
imputed data sets will be created. The SAVE option is used to save the
imputed data sets for further analysis using TYPE=IMPUTATION in the
DATA command. All variables on the USEVARIABLES and
AUXILIARY lists are saved. The asterisk in the data set name is
replaced by the number of the imputation. The data sets saved are
missimp1.dat, missimp2.dat, etc. The imputed data sets will contain the
variables x1, x2, y1-y4, z1-z5, and v1-v10 in that order. The data sets
can be used in a subsequent analysis using TYPE=IMPUTATION in the
DATA command. See Example 13.13. An explanation of the other
commands can be found in Example 11.1.

397
CHAPTER 11

EXAMPLE 11.6: MULTIPLE IMPUTATION FOLLOWED BY

THE ESTIMATION OF A GROWTH MODEL USING
MAXIMUM LIKELIHOOD

TITLE: this is an example of multiple imputation

followed by the estimation of a growth
model using maximum likelihood
DATA: FILE = ex11.6.dat;
VARIABLE: NAMES = x1 y1-y4 z x2;
USEVARIABLES = y1-y4 x1 x2;
MISSING = ALL(999);
DATA IMPUTATION:
IMPUTE = y1-y4 x1 (c) x2;
NDATASETS = 10;
ANALYSIS: ESTIMATOR = ML;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
i s ON x1 x2;
OUTPUT: TECH1 TECH8;

y1 y2 y3 y4

i s

x1 x2

398
 Examples: Missing Data Modeling And Bayesian Analysis

In this example, multiple imputation for a set of variables with missing

values is carried out using Bayesian analysis (Rubin, 1987; Schafer,
1997). The imputed data sets are used in the estimation of the growth
model shown in the figure above using maximum likelihood estimation.

The DATA IMPUTATION command is used when a data set contains

missing values to create a set of imputed data sets using multiple
imputation methodology. Multiple imputation is carried out using
Bayesian estimation. Data are imputed using an unrestricted H1 model.
The IMPUTE option is used to specify the analysis variables for which
missing values will be imputed. In this example, missing values will be
imputed for y1, y2, y3, y4, x1, and x2. The c in parentheses after x1
specifies that x1 is treated as a categorical variable for data imputation.
The NDATASETS option is used to specify the number of imputed data
sets to create. The default is five. In this example, 10 data sets will be
imputed.

The maximum likelihood parameter estimates for the growth model are
averaged over the set of 10 analyses and standard errors are computed
using the average of the standard errors over the set of 10 analyses and
the between analysis parameter estimate variation (Rubin, 1987; Schafer,
1997). A chi-square test of overall model fit is provided (Asparouhov &
Muthén, 2008c; Enders, 2010). The ESTIMATOR option is used to
specify the estimator to be used in the analysis. By specifying ML,
maximum likelihood estimation is used. An explanation of the other
commands can be found in Examples 11.1 and 11.5.

EXAMPLE 11.7: MULTIPLE IMPUTATION OF PLAUSIBLE

VALUES USING BAYESIAN ESTIMATION OF A GROWTH
MODEL

TITLE: this is an example of multiple imputation

of plausible values generated from a
multiple indicator linear growth model for
categorical outcomes using Bayesian
estimation
DATA: FILE = ex11.7.dat;
VARIABLE: NAMES = u11 u21 u31 u12 u22 u32 u13 u23
u33;
CATEGORICAL = u11-u33;

399
CHAPTER 11

ANALYSIS: ESTIMATOR = BAYES;

PROCESSORS = 2;
MODEL: f1 BY u11
u21-u31 (1-2);
f2 BY u12
u22-u32 (1-2);
f3 BY u13
u23-u33 (1-2);
[u11$1 u12$1 u13$1] (3);
[u21$1 u22$1 u23$1] (4);
[u31$1 u32$1 u33$1] (5);
i s | f1@0 f2@1 f3@2;
DATA IMPUTATION:
NDATASETS = 20;
SAVE = ex11.7imp*.dat;
SAVEDATA: FILE = ex11.7plaus.dat;
SAVE = FSCORES (20);
FACTORS = f1-f3 i s;
SAVE = LRESPONSES (20);
LRESPONSES = u11-u33;
OUTPUT: TECH1 TECH8;

u11 u21 u31 u12 u22 u32 u13 u23 u33

fl f2 f3

i s

In this example, plausible values (Mislevy et al., 1992; von Davier et al.,
2009) are obtained by multiple imputation (Rubin, 1987; Schafer, 1997)
based on a multiple indicator linear growth model for categorical
outcomes shown in the picture above using Bayesian estimation. The

400
 Examples: Missing Data Modeling And Bayesian Analysis

plausible values in the multiple imputation data sets can be used for
subsequent analysis.

The ANALYSIS command is used to describe the technical details of the

analysis. The ESTIMATOR option is used to specify the estimator to be
used in the analysis. By specifying BAYES, Bayesian estimation is used
to estimate the model. The DATA IMPUTATION command is used
when a data set contains missing values to create a set of imputed data
sets using multiple imputation methodology. Multiple imputation is
carried out using Bayesian estimation. When a MODEL command is
used with ESTIMATOR=BAYES, data are imputed using the H0 model
specified in the MODEL command. The IMPUTE option is used to
specify the analysis variables for which missing values will be imputed.
When the IMPUTE option is not used, no imputation of missing data for
the analysis variables is done.

In the DATA IMPUTATION command, the NDATASETS option is

used to specify the number of imputed data sets to create. The default is
five. In this example, 20 data sets will be imputed to more fully
represent the variability in the latent variables. The SAVE option is
used to save the imputed data sets for subsequent analysis. The asterisk
(*) is replaced by the number of the imputed data set. A file is also
produced that contains the names of all of the data sets. To name this
file, the asterisk (*) is replaced by the word list. In this example, the file
is called ex11.7implist.dat. The multiple imputation data sets named
using the SAVE option contain the imputed values for each observation
for the observed variables u11 through u33; the continuous latent
response variables u11* through u33* for the categorical outcomes u11
through u33; and the factor scores for the latent variables f1, f2, f3, i,
and s.

In the SAVEDATA command, the FILE option is used to specify the

name of the ASCII file in which the individual-level data used in the
analysis will be saved. In this example, the file is called
ex11.7plaus.dat. When SAVE=FSCORES is used with
ESTIMATOR=BAYES, a distribution of factor scores, called plausible
values, is obtained for each observation. The following summaries are
saved along with the other analysis variables: mean, median, standard
deviation, lower 2.5% limit, and upper 97.5% limit. The number 20 in
parentheses is the number of imputations or draws that are used from the
Bayesian posterior distribution to compute the plausible value

401
CHAPTER 11

distribution for each observation. The FACTORS option is used to

specify the names of the factors for which the plausible value
distributions will also be saved. In this example, the plausible value
distributions will be saved for f1, f2, f3, i, and s.

When SAVE=LRESPONSES is used with ESTIMATOR=BAYES, a

distribution of latent response variable scores is obtained for each
observation. The following summaries are saved along with the other
analysis variables: mean, median, standard deviation, lower 2.5% limit,
and upper 97.5% limit. The number 20 in parentheses is the number of
imputations or draws that are used from the Bayesian posterior
distribution to compute the latent response variable distribution for each
observation. The LRESPONSES option is used to specify the names of
the latent response variables underlying categorical outcomes for which
the latent response variable distributions will also be saved. In this
example, the latent response variable distributions will be saved for u11
through u33. An explanation of the other commands can be found in
Examples 11.1 and 11.2.

EXAMPLE 11.8: MULTIPLE IMPUTATION USING A TWO-

LEVEL FACTOR MODEL WITH CATEGORICAL OUTCOMES
FOLLOWED BY THE ESTIMATION OF A GROWTH MODEL

TITLE: this is an example of multiple imputation

using a two-level factor model with
categorical outcomes
DATA: FILE = ex11.8.dat;
VARIABLE: NAMES are u11 u21 u31 u12 u22 u32 u13 u23
u33 clus;
CATEGORICAL = u11-u33;
CLUSTER = clus;
MISSING = ALL (999);
ANALYSIS: TYPE = TWOLEVEL;
ESTIMATOR = BAYES;
PROCESSORS = 2;

402
 Examples: Missing Data Modeling And Bayesian Analysis

MODEL: %WITHIN%
f1w BY u11
u21 (1)
u31 (2);
f2w BY u12
u22 (1)
u32 (2);
f3w BY u13
u23 (1)
u33 (2);
%BETWEEN%
fb BY u11-u33*1;
fb@1;
DATA IMPUTATION:
IMPUTE = u11-u33(c);
SAVE = ex11.8imp*.dat;
OUTPUT: TECH1 TECH8;

u11 u21 u31 u12 u22 u32 u13 u23 u33

f1w f2w f3w

Within

u11 u21 u31 u12 u22 u32 u13 u23 u33

fb Between

403
CHAPTER 11

In this example, missing values are imputed for a set of variables using
multiple imputation (Rubin, 1987; Schafer, 1997). In the first part of
this example, imputation is done using the two-level factor model with
categorical outcomes shown in the picture above. In the second part of
this example, the multiple imputation data sets are used for a two-level
multiple indicator growth model with categorical outcomes using two-
level weighted least squares estimation.

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis. By
selecting TWOLEVEL, a multilevel model with random intercepts is
estimated. The ESTIMATOR option is used to specify the estimator to
be used in the analysis. By specifying BAYES, Bayesian estimation is
used to estimate the model. The DATA IMPUTATION command is
used when a data set contains missing values to create a set of imputed
data sets using multiple imputation methodology. Multiple imputation is
carried out using Bayesian estimation. When a MODEL command is
used, data are imputed using the H0 model specified in the MODEL
command. The IMPUTE option is used to specify the analysis variables
for which missing values will be imputed. In this example, missing
values will be imputed for u11, u21, u31, u12, u22, u32, u13, u23, and
u33. The c in parentheses after the list of variables specifies that they
are treated as categorical variables for data imputation. An explanation
of the other commands can be found in Examples 11.1, 11.2, and 11.5.

404
 Examples: Missing Data Modeling And Bayesian Analysis

TITLE: this is an example of a two-level multiple

indicator growth model with categorical
outcomes using multiple imputation data
DATA: FILE = ex11.8implist.dat;
TYPE = IMPUTATION;
VARIABLE: NAMES are u11 u21 u31 u12 u22 u32 u13 u23
u33 clus;
CATEGORICAL = u11-u33;
CLUSTER = clus;
ANALYSIS: TYPE = TWOLEVEL;
ESTIMATOR = WLSMV;
PROCESSORS = 2;
MODEL: %WITHIN%
f1w BY u11
u21 (1)
u31 (2);
f2w BY u12
u22 (1)
u32 (2);
f3w BY u13
u23 (1)
u33 (2);
iw sw | f1w@0 f2w@1 f3w@2;
%BETWEEN%
f1b BY u11
u21 (1)
u31 (2);
f2b BY u12
u22 (1)
u32 (2);
f3b BY u13
u23 (1)
u33 (2);
[u11$1 u12$1 u13$1] (3);
[u21$1 u22$1 u23$1] (4);
[u31$1 u32$1 u33$1] (5);
u11-u33;
ib sb | f1b@0 f2b@1 f3b@2;
[f1b-f3b@0 ib@0 sb];
f1b-f3b (6);
OUTPUT: TECH1 TECH8;
SAVEDATA: SWMATRIX = ex11.8sw*.dat;

405
CHAPTER 11

u11 u21 u31 u12 u22 u32 u13 u23 u33

f1w f2w f3w

Within
iw sw

u11 u21 u31 u12 u22 u32 u13 u23 u33

f1b f2b f3b

Between
ib sb

In the second part of this example, the data sets saved in the first part of
the example are used in the estimation of a two-level multiple indicator
growth model with categorical outcomes. The model is the same as in
Example 9.15. The two-level weighted least squares estimator described
in Asparouhov and Muthén (2007) is used in this example. This
estimator does not handle missing data using MAR. By doing Bayesian
multiple imputation as a first step, this disadvantage is avoided given
that there is no missing data for the weighted least squares analysis. To
save computational time in subsequent analyses, the two-level weighted
least squares sample statistics and weight matrix for each of the imputed
data sets are saved.

406
 Examples: Missing Data Modeling And Bayesian Analysis

The ANALYSIS command is used to describe the technical details of the

analysis. The TYPE option is used to describe the type of analysis. By
selecting TWOLEVEL, a multilevel model with random intercepts is
estimated. The ESTIMATOR option is used to specify the estimator to
be used in the analysis. By specifying WLSMV, a robust weighted least
squares estimator is used. The SAVEDATA command is used to save
the analysis data, auxiliary variables, and a variety of analysis results.
The SWMATRIX option is used with TYPE=TWOLEVEL and
weighted least squares estimation to specify the name of the ASCII file
in which the within- and between-level sample statistics and their
corresponding estimated asymptotic covariance matrix will be saved. In
this example, the files are called ex11.8sw*.dat where the asterisk (*) is
replaced by the number of the imputed data set. A file is also produced
that contains the names of all of the imputed data sets. To name this file,
the asterisk (*) is replaced by the word list. The file, in this case
ex11.8swlist.dat, contains the names of the imputed data sets.

To use the saved within- and between-level sample statistics and their
corresponding estimated asymptotic covariance matrix for each
imputation in a subsequent analysis, specify:

DATA:
FILE = ex11.8implist.dat;
TYPE = IMPUTATION;
SWMATRIX = ex11.8swlist.dat;

An explanation of the other commands can be found in Examples 9.15,

11.1, 11.2, and 11.5.

407
CHAPTER 11

408
 Examples: Monte Carlo Simulation Studies

CHAPTER 12
EXAMPLES: MONTE CARLO
SIMULATION STUDIES
Monte Carlo simulation studies are often used for methodological
investigations of the performance of statistical estimators under various
conditions. They can also be used to decide on the sample size needed
for a study and to determine power (Muthén & Muthén, 2002). Monte
Carlo studies are sometimes referred to as simulation studies.

Mplus has extensive Monte Carlo simulation facilities for both data
generation and data analysis. Several types of data can be generated:
simple random samples, clustered (multilevel) data, missing data, and
data from populations that are observed (multiple groups) or unobserved
(latent classes). Data generation models can include random effects,
interactions between continuous latent variables, interactions between
continuous latent variables and observed variables, and between
categorical latent variables. Dependent variables can be continuous,
censored, binary, ordered categorical (ordinal), unordered categorical
(nominal), counts, or combinations of these variable types. In addition,
two-part (semicontinuous) variables and time-to-event variables can be
generated. Independent variables can be binary or continuous. All or
some of the Monte Carlo generated data sets can be saved.

The analysis model can be different from the data generation model. For
example, variables can be generated as categorical and analyzed as
continuous or data can be generated as a three-class model and analyzed
as a two-class model. In some situations, a special external Monte Carlo
feature is needed to generate data by one model and analyze it by a
different model. For example, variables can be generated using a
clustered design and analyzed ignoring the clustering. Data generated
outside of Mplus can also be analyzed using this special Monte Carlo
feature.

Other special features that can be used with Monte Carlo simulation
studies include saving parameter estimates from the analysis of real data
to be used as population parameter and/or coverage values for data
generation in a Monte Carlo simulation study. In addition, analysis
results from each replication of a Monte Carlo simulation study can be

409
CHAPTER 12

saved in an external file for further investigation. Chapter 19 discusses

the options of the MONTECARLO command.

Monte Carlo data generation can include the following special features:

• Single or multiple group analysis for non-mixture models

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations
• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Wald chi-square test of parameter equalities
• Analysis with between-level categorical latent variables

Multiple group data generation is specified by using the NGROUPS

option of the MONTECARLO command and the MODEL
POPULATION-label command. Missing data generation is specified by
using the PATMISS and PATPROBS options of the MONTECARLO
command or the MISSING option of the MONTECARLO command in
conjunction with the MODEL MISSING command. Complex survey
data are generated by using the TYPE=TWOLEVEL option of the
ANALYSIS command in conjunction with the NCSIZES and CSIZES
options of the MONTECARLO command. Latent variable interactions
are generated by using the | symbol of the MODEL POPULATION
command in conjunction with the XWITH option of the MODEL
POPULATION command. Random slopes are generated by using the |
symbol of the MODEL POPULATION command in conjunction with
the ON option of the MODEL POPULATION command. Individually-
varying times of observations are generated by using the | symbol of the
MODEL POPULATION command in conjunction with the AT option of
the MODEL POPULATION command and the TSCORES option of the
MONTECARLO command. Linear and non-linear parameter constraints
are specified by using the MODEL CONSTRAINT command. Indirect
effects are specified by using the MODEL INDIRECT command.
Maximum likelihood estimation is specified by using the ESTIMATOR
option of the ANALYSIS command. The MODEL TEST command is
used to test linear restrictions on the parameters in the MODEL and

410
 Examples: Monte Carlo Simulation Studies

MODEL CONSTRAINT commands using the Wald chi-square test.

Between-level categorical latent variables are generated using the
GENCLASSES option and specified using the CLASSES and
BETWEEN options.

Besides the examples in this chapter, Monte Carlo versions of most of

the examples in the previous example chapters are included on the CD
that contains the Mplus program and at www.statmodel.com. Following
is the set of Monte Carlo examples included in this chapter:

• 12.1: Monte Carlo simulation study for a CFA with covariates

(MIMIC) with continuous factor indicators and patterns of missing
data
• 12.2: Monte Carlo simulation study for a linear growth model for a
continuous outcome with missing data where attrition is predicted by
time-invariant covariates (MAR)
• 12.3: Monte Carlo simulation study for a growth mixture model
with two classes and a misspecified model
• 12.4: Monte Carlo simulation study for a two-level growth model
for a continuous outcome (three-level analysis)
• 12.5: Monte Carlo simulation study for an exploratory factor
analysis with continuous factor indicators
• 12.6 Step 1: Monte Carlo simulation study where clustered data for a
two-level growth model for a continuous outcome (three-level
analysis) are generated, analyzed, and saved
• 12.6 Step 2: External Monte Carlo analysis of clustered data
generated for a two-level growth model for a continuous outcome
using TYPE=COMPLEX for a single-level growth model
• 12.7 Step 1: Real data analysis of a CFA with covariates (MIMIC)
for continuous factor indicators where the parameter estimates are
saved for use in a Monte Carlo simulation study
• 12.7 Step 2: Monte Carlo simulation study where parameter
estimates saved from a real data analysis are used for population
parameter values for data generation and coverage
• 12.8: Monte Carlo simulation study for discrete-time survival
analysis*
• 12.9: Monte Carlo simulation study for a two-part (semicontinuous)
growth model for a continuous outcome*
• 12.10: Monte Carlo simulation study for a two-level continuous-time
survival analysis using Cox regression with a random intercept and a
frailty*

411
CHAPTER 12

• 12.11: Monte Carlo simulation study for a two-level mediation

model with random slopes
• 12:12 Monte Carlo simulation study for a multiple group EFA with
continuous factor indicators with measurement invariance of
intercepts and factor loadings

* Example uses numerical integration in the estimation of the model.

This can be computationally demanding depending on the size of the
problem.

MONTE CARLO DATA GENERATION

Data are generated according to the following steps. First, multivariate
normal data are generated for the independent variables in the model.
Second, the independent variables are categorized if requested. The
third step varies depending on the dependent variable type and the model
used. Data for continuous dependent variables are generated according
to a distribution that is multivariate normal conditional on the
independent variables. For categorical dependent variables under the
probit model using weighted least squares estimation, data for
continuous dependent variables are generated according to a distribution
that is multivariate normal conditional on the independent variables.
These dependent variables are then categorized using the thresholds
provided in the MODEL POPULATION command or the
POPULATION option of the MONTECARLO command. For
categorical dependent variables under the probit model using maximum
likelihood estimation, the dependent variables are generated according to
the probit model using the values of the thresholds and slopes from the
MODEL POPULATION command or the POPULATION option of the
MONTECARLO command. For categorical dependent variables under
the logistic model using maximum likelihood estimation, the dependent
variables are generated according to the logistic model using the values
of the thresholds and slopes from the MODEL POPULATION command
or the POPULATION option of the MONTECARLO command. For
censored dependent variables, the dependent variables are generated
according to the censored normal model using the values of the
intercepts and slopes from the MODEL POPULATION command or the
POPULATION option of the MONTECARLO command. For
unordered categorical (nominal) dependent variables, the dependent
variables are generated according to the multinomial logistic model
using the values of the intercepts and slopes from the MODEL

412
 Examples: Monte Carlo Simulation Studies

POPULATION command or the POPULATION option of the

MONTECARLO command. For count dependent variables, the
dependent variables are generated according to the log rate model using
the values of the intercepts and slopes from the MODEL POPULATION
command or the POPULATION option of the MONTECARLO
command. For time-to-event variables in continuous-time survival
analysis, the dependent variables are generated according to the loglinear
model using the values of the intercepts and slopes from the MODEL
POPULATION command or the POPULATION option of the
MONTECARLO command.

To save the generated data for subsequent analysis without analyzing

them, use the TYPE=BASIC option of the ANALYSIS command in
conjunction with the REPSAVE and SAVE options of the
MONTECARLO command.

MONTE CARLO DATA ANALYSIS

There are two ways to carry out a Monte Carlo simulation study in
Mplus: an internal Monte Carlo simulation study or an external Monte
Carlo simulation study. In an internal Monte Carlo simulation study,
data are generated and analyzed in one step using the MONTECARLO
command. In an external Monte Carlo simulation study, multiple data
sets are generated in a first step using either Mplus or another computer
program. These data are analyzed and the results summarized in a
second step using regular Mplus analysis facilities in conjunction with
the TYPE=MONTECARLO option of the DATA command.

Internal Monte Carlo can be used whenever the analysis type and scales
of the dependent variables remain the same for both data generation and
analysis. Internal Monte Carlo can also be used with TYPE=GENERAL
when dependent variables are generated as categorical and analyzed as
continuous. Internal Monte Carlo can also be used when data are
generated and analyzed for a different number of latent classes. In all
other cases, data from all replications can be saved and subsequently
analyzed using external Monte Carlo.

413
CHAPTER 12

MONTE CARLO OUTPUT

The default output for the MONTECARLO command includes a listing
of the input setup, a summary of the analysis specifications, sample
statistics from the first replication, the analysis results summarized over
replications, and TECH1 which shows the free parameters in the model
and the starting values. Following is an example of the output for tests
of model fit for the chi-square test statistic. The same format is used
with other fit statistics.

Chi-Square Test of Model Fit

Degrees of freedom 5

Mean 5.253
Std Dev 3.325
Number of successful computations 500

Proportions Percentiles
Expected Observed Expected Observed
0.990 0.988 0.554 0.372
0.980 0.976 0.752 0.727
0.950 0.958 1.145 1.193
0.900 0.894 1.610 1.539
0.800 0.804 2.343 2.367
0.700 0.710 3.000 3.090
0.500 0.532 4.351 4.555
0.300 0.330 6.064 6.480
0.200 0.242 7.289 7.870
0.100 0.136 9.236 9.950
0.050 0.062 11.070 11.576
0.020 0.022 13.388 13.394
0.010 0.014 15.086 15.146

The mean and standard deviation of the chi-square test statistic over the
replications of the Monte Carlo analysis are given. The column labeled
Proportions Expected (column 1) should be understood in conjunction
with the column labeled Percentiles Expected (column 3). Each value in
column 1 gives the probability of observing a chi-square value greater
than the corresponding value in column 3. The column 3 percentile
values are determined from a chi-square distribution with the degrees of
freedom given by the model, in this case 5. In this output, the column 1
value of 0.05 gives the probability that the chi-square value exceeds the
column 3 percentile value (the critical value of the chi-square
distribution) of 11.070. Columns 2 and 4 give the corresponding values
observed in the Monte Carlo replications. Column 2 gives the
proportion of replications for which the critical value is exceeded, which

414
 Examples: Monte Carlo Simulation Studies

in this example is 0.062, close to the expected value 0.05 which

indicates that the chi-square distribution is well approximated in this
case. The column 4 value of 11.576 is the chi-square value at this
percentile from the Monte Carlo analysis that has 5% of the values in the
replications above it. The fact that it deviates little from the theoretical
value of 11.070 is again an indication that the chi-square distribution is
well approximated in this case. For the other fit statistics, the normal
distribution is used to obtain the critical values of the test statistic.

The summary of the analysis results includes the population value for
each parameter, the average of the parameter estimates across
replications, the standard deviation of the parameter estimates across
replications, the average of the estimated standard errors across
replications, the mean square error for each parameter (M.S.E.), 95
percent coverage, and the proportion of replications for which the null
hypothesis that a parameter is equal to zero is rejected at the .05 level.

MODEL RESULTS

ESTIMATES S. E. M. S. E. 95% % Sig

Population Average Std. Dev. Average Cover Coeff
I |
Y1 1.000 1.0000 0.0000 0.0000 0.0000 1.000 0.000
Y2 1.000 1.0000 0.0000 0.0000 0.0000 1.000 0.000
Y3 1.000 1.0000 0.0000 0.0000 0.0000 1.000 0.000
Y4 1.000 1.0000 0.0000 0.0000 0.0000 1.000 0.000

S |
Y1 0.000 0.0000 0.0000 0.0000 0.0000 1.000 0.000
Y2 1.000 1.0000 0.0000 0.0000 0.0000 1.000 0.000
Y3 2.000 2.0000 0.0000 0.0000 0.0000 1.000 0.000
Y4 3.000 3.0000 0.0000 0.0000 0.0000 1.000 0.000

I WITH
S 0.000 0.0006 0.0301 0.0306 0.0009 0.958 0.042

Means
I 0.000 -0.0006 0.0473 0.0460 0.0022 0.950 0.050
S 0.200 0.2015 0.0278 0.0274 0.0008 0.946 1.000

Variances
I 0.500 0.4969 0.0704 0.0685 0.0050 0.936 1.000
S 0.200 0.1997 0.0250 0.0237 0.0006 0.930 1.000

Residual Variances
Y1 0.500 0.5016 0.0683 0.0657 0.0047 0.934 1.000
Y2 0.500 0.5018 0.0460 0.0451 0.0021 0.958 1.000
Y3 0.500 0.5025 0.0515 0.0532 0.0027 0.956 1.000
Y4 0.500 0.4991 0.0932 0.0918 0.0087 0.946 1.000

415
CHAPTER 12

The column labeled Population gives the population parameter values

that are given in the MODEL command, the MODEL COVERAGE
command, or using the COVERAGE option of the MONTECARLO
command. The column labeled Average gives the average of the
parameter estimates across the replications of the Monte Carlo
simulation study. These two values are used to evaluate parameter bias.
To determine the percentage of parameter bias, subtract the population
parameter value from the average parameter value, divide this number by
the population parameter value, and multiply by 100. The parameter
bias for the variance of i would be

100 (.4969 - .5000) / .5000 = -0.62.

This results in a bias of -0.62 percent.

The column labeled Std. Dev. gives the standard deviation of the
parameter estimates across the replications of the Monte Carlo
simulation study. When the number of replications is large, this is
considered to be the population standard error. The column labeled S.E.
Average gives the average of the estimated standard errors across
replications of the Monte Carlo simulation study. To determine standard
error bias, subtract the population standard error value from the average
standard error value, divide this number by the population standard error
value, and multiply by 100.

The column labeled M.S.E. gives the mean square error for each
parameter. M.S.E. is equal to the variance of the estimates across the
replications plus the square of the bias. For example, the M.S.E. for the
variance of i is equal to 0.0704 squared plus (0.4969 - 0.5) squared
which is equal to 0.00497 or 0.0050. The column labeled 95% Cover
gives the proportion of replications for which the 95% confidence
interval contains the population parameter value. This gives the
coverage which indicates how well the parameters and their standard
errors are estimated. In this output, all coverage values are close to the
correct value of 0.95.

The column labeled % Sig Coeff gives the proportion of replications for
which the null hypothesis that a parameter is equal to zero is rejected at
the .05 level (two-tailed test with a critical value of 1.96). The statistical
test is the ratio of the parameter estimate to its standard error, an
approximately normally distributed quantity (z-score) in large samples.

416
 Examples: Monte Carlo Simulation Studies

For parameters with population values different from zero, this value is
an estimate of power with respect to a single parameter, that is, the
probability of rejecting the null hypothesis when it is false. For
parameters with population values equal to zero, this value is an estimate
of Type I error, that is, the probability of rejecting the null hypothesis
when it is true. In this output, the power to reject that the slope growth
factor mean is zero is estimated as 1.000, that is, exceeding the standard
of 0.8 power.

MONTE CARLO EXAMPLES

Following is the set of Monte Carlo simulation study examples. Besides
the examples in this chapter, Monte Carlo versions of most of the
examples in the previous example chapters are included on the CD that
contains the Mplus program and at www.statmodel.com.

417
CHAPTER 12

EXAMPLE 12.1: MONTE CARLO SIMULATION STUDY FOR

A CFA WITH COVARIATES (MIMIC) WITH CONTINUOUS
FACTOR INDICATORS AND PATTERNS OF MISSING DATA

TITLE: this is an example of a Monte Carlo

simulation study for a CFA with covariates
(MIMIC) with continuous factor indicators
and patterns of missing data
MONTECARLO:
NAMES ARE y1-y4 x1 x2;
NOBSERVATIONS = 500;
NREPS = 500;
SEED = 4533;
CUTPOINTS = x2(1);
PATMISS = y1(.1) y2(.2) y3(.3) y4(1) |
y1(1) y2(.1) y3(.2) y4(.3);
PATPROBS = .4 | .6;
MODEL POPULATION:
[x1-x2@0];
x1-x2@1;
f BY y1@1 y2-y4*1;
f*.5;
y1-y4*.5;
f ON x1*1 x2*.3;
MODEL: f BY y1@1 y2-y4*1;
f*.5;
y1-y4*.5;
f ON x1*1 x2*.3;
OUTPUT: TECH9;

In this example, data are generated and analyzed according to the CFA
with covariates (MIMIC) model described in Example 5.8. Two factors
are regressed on two covariates and data are generated with patterns of
missing data.

TITLE: this is an example of a Monte Carlo

simulation study for a CFA with covariates
(MIMIC) with continuous factor indicators
and patterns of missing data

The TITLE command is used to provide a title for the output. The title
is printed in the output just before the Summary of Analysis.

418
 Examples: Monte Carlo Simulation Studies

MONTECARLO:
NAMES ARE y1-y4 x1 x2;
NOBSERVATIONS = 500;
NREPS = 500;
SEED = 4533;
CUTPOINTS = x2(1);
PATMISS = y1(.1) y2(.2) y3(.3) y4(1) |
y1(1) y2(.1) y3(.2) y4(.3);
PATPROBS = .4 | .6;

The MONTECARLO command is used to describe the details of a

Monte Carlo simulation study. The NAMES option is used to assign
names to the variables in the generated data sets. The data sets in this
example each have six variables: y1, y2, y3, y4, x1, and x2. Note that a
hyphen can be used as a convenience feature in order to generate a list of
names. The NOBSERVATIONS option is used to specify the sample
size to be used for data generation and for analysis. In this example, the
sample size is 500. The NREPS option is used to specify the number of
replications, that is, the number of samples to draw from a specified
population. In this example, 500 samples will be drawn. The SEED
option is used to specify the seed to be used for the random draws. The
seed 4533 is used here. The default seed value is zero.

The GENERATE option is used to specify the scale of the dependent

variables for data generation. In this example, the dependent variables
are continuous which is the default for the GENERATE option.
Therefore, the GENERATE option is not necessary and is not used here.
The CUTPOINTS option is used to create binary variables from the
multivariate normal independent variables generated by the program. In
this example, the variable x2 is cut at the value of one which is one
standard deviation above the mean because the mean and variance used
for data generation are zero and one. This implies that after the cut x2 is
a 0/1 binary variable where 16 percent of the population have the value
of 1. The mean and variance of x2 for data generation are specified in
the MODEL POPULATION command.

The PATMISS and PATPROBS options are used together to describe

the patterns of missing data to be used in data generation. The
PATMISS option is used to specify the missing data patterns and the
proportion missing for each variable. The patterns are separated using
the | symbol. The PATPROBS option is used to specify the proportion
of individuals for each missing data pattern. In this example, there are
two missing value patterns. In the first pattern, y1 has 10 percent

419
CHAPTER 12

missing, y2 has 20 percent missing, y3 has 30 percent missing, and y4

has 100 percent missing. In the second pattern, y1 has 100 percent
missing, y2 has 10 percent missing, y3 has 20 percent missing, and y4
has 30 percent missing. As specified in the PATPROBS option, 40
percent of the individuals in the generated data have missing data pattern
1 and 60 percent have missing data pattern 2. This may correspond to a
situation of planned missingness where a measurement instrument is
administered in two different versions given to randomly chosen parts of
the population. In this example, some individuals answer items y1, y2,
and y3, while others answer y2, y3, and y4.

MODEL POPULATION:
[x1-x2@0];
x1-x2@1;
f BY y1@1 y2-y4*1;
f*.5;
y1-y4*.5;
f ON x1*1 x2*.3;

The MODEL POPULATION command is used to provide the

population parameter values to be used in data generation. Each
parameter in the model must be specified followed by the @ symbol or
the asterisk (*) and the population parameter value. Any model
parameter not given a population parameter value will be assigned the
value of zero as the population parameter value. The first two lines in
the MODEL POPULATION command refer to the means and variances
of the independent variables x1 and x2. The covariances between the
independent variables can also be specified. Variances of the
independent variables in the model must be specified. Means and
covariances of the independent variables do not need to be specified if
their values are zero.

MODEL: f BY y1@1 y2-y4*1;

f*.5;
y1-y4*.5;
f ON x1*1 x2*.3;

The MODEL command is used to describe the analysis model as in

regular analyses. In Monte Carlo simulation studies, the MODEL
command is also used to provide values for each parameter that are used
as population parameter values for computing coverage and starting
values in the estimation of the model. They are printed in the first
column of the output labeled Population. Population parameter values

420
 Examples: Monte Carlo Simulation Studies

for the analysis model can also be provided using the MODEL
COVERAGE command or the COVERAGE option of the
MONTECARLO command. Alternate starting values can be provided
using the STARTING option of the MONTECARLO command. Note
that the population parameter values for coverage given in the analysis
model are different from the population parameter values used for data
generation if the analysis model is misspecified.

OUTPUT: TECH9;

The OUTPUT command is used to request additional output not

included as the default. The TECH9 option is used to request error
messages related to convergence for each replication of the Monte Carlo
simulation study.

EXAMPLE 12.2: MONTE CARLO SIMULATION STUDY FOR

A LINEAR GROWTH MODEL FOR A CONTINUOUS
OUTCOME WITH MISSING DATA WHERE ATTRITION IS
PREDICTED BY TIME-INVARIANT COVARIATES (MAR)

TITLE: this is an example of a Monte Carlo

simulation study for a linear growth model
for a continuous outcome with missing data
where attrition is predicted by time-
invariant covariates (MAR)
MONTECARLO:
NAMES ARE y1-y4 x1 x2;
NOBSERVATIONS = 500;
NREPS = 500;
SEED = 4533;
CUTPOINTS = x2(1);
MISSING = y1-y4;
MODEL POPULATION:
x1-x2@1;
[x1-x2@0];
i s | y1@0 y2@1 y3@2 y4@3;
[i*1 s*2];
i*1; s*.2; i WITH s*.1;
y1-y4*.5;
i ON x1*1 x2*.5;
s ON x1*.4 x2*.25;

421
CHAPTER 12

MODEL MISSING:
[y1-y4@-1];
y1 ON x1*.4 x2*.2;
y2 ON x1*.8 x2*.4;
y3 ON x1*1.6 x2*.8;
y4 ON x1*3.2 x2*1.6;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
[i*1 s*2];
i*1; s*.2; i WITH s*.1;
y1-y4*.5;
i ON x1*1 x2*.5;
s ON x1*.4 x2*.25;
OUTPUT: TECH9;

In this example, missing data are generated to illustrate both random

missingness and attrition predicted by time-invariant covariates (MAR).
This Monte Carlo simulation study can be used to estimate the power to
detect that the binary covariate x2 has a significant effect on the growth
slope factor s. The binary covariate x2 may correspond to a treatment
variable or a gender variable.

The MISSING option in the MONTECARLO command is used to

identify the dependent variables in the data generation model for which
missing data will be generated. The MODEL MISSING command is
used to provide information about the population parameter values for
the missing data model to be used in the generation of data. The
MODEL MISSING command specifies a logistic regression model for a
set of binary dependent variables that represent not missing (scored as 0)
and missing (scored as 1) for the dependent variables in the data
generation model. The first statement in the MODEL MISSING
command defines the intercepts in the logistic regressions for each of the
binary dependent variables. If the covariates predicting missingness all
have values of zero, the logistic regression intercept value of -1
corresponds to a probability of 0.27 of having missing data on the
dependent variables. This would reflect missing completely at random.
The four ON statements specify the logistic regression of the four binary
dependent variables on the two covariates x1 and x2 to reflect attrition
predicted by the covariates. Because the values of the logistic regression
slopes increase over time as seen in the increase of the slopes from y1 to
y4, attrition also increases over time and becomes more selective over
time. An explanation of the other commands can be found in Example
12.1.

422
 Examples: Monte Carlo Simulation Studies

EXAMPLE 12.3: MONTE CARLO SIMULATION STUDY FOR

A GROWTH MIXTURE MODEL WITH TWO CLASSES AND A
MISSPECIFIED MODEL

TITLE: this is an example of a Monte Carlo

simulation study for a growth mixture
model with two classes and a misspecified
model
MONTECARLO:
NAMES ARE u y1-y4 x;
NOBSERVATIONS = 500;
NREPS = 10;
SEED = 53487;
GENERATE = u (1);
CATEGORICAL = u;
GENCLASSES = c (2);
CLASSES = c (1);
MODEL POPULATION:
%OVERALL%
[x@0];
x@1;
i s | y1@0 y2@1 y3@2 y4@3;
i*.25 s*.04;
i WITH s*0;
y1*.4 y2*.35 y3*.3 y4*.25;
i ON x*.5;
s ON x*.1;
c#1 ON x*.2;
[c#1*0];
%c#1%
[u$1*1 i*3 s*.5];
%c#2%
[u$1*-1 i*1 s*0];

423
CHAPTER 12

ANALYSIS: TYPE = MIXTURE;

MODEL:
%OVERALL%
i s | y1@0 y2@1 y3@2 y4@3;
i*.25 s*.04;
i WITH s*0;
y1*.4 y2*.35 y3*.3 y4*.25;
i ON x*.5;
s ON x*.1;
! c#1 ON x*.2;
! [c#1*0];
u ON x;
%c#1%
[u$1*1 i*3 s*.5];
! %c#2%
! [u$1*-1 i*1 s*0];
OUTPUT: TECH9;

In this example, data are generated according the two class model
described in Example 8.1 and analyzed as a one class model. This
results in a misspecified model. Differences between the parameter
values that generated the data and the estimated parameters can be
studied to determine the extent of the distortion.

The GENERATE option is used to specify the scale of the dependent

variables for data generation. In this example, the dependent variable u
is binary because it has one threshold. For binary variables, this is
specified by placing the number one in parenthesis following the
variable name. The CATEGORICAL option is used to specify which
dependent variables are treated as binary or ordered categorical (ordinal)
variables in the model and its estimation. In the example above, the
variable u is generated and analyzed as a binary variable. The
GENCLASSES option is used to assign names to the categorical latent
variables in the data generation model and to specify the number of
latent classes to be used for data generation. In the example above, there
is one categorical latent variable c that has two latent classes for data
generation. The CLASSES option is used to assign names to the
categorical latent variables in the analysis model and to specify the
number of latent classes to be used for analysis. In the example above,
there is one categorical latent variable c that has one latent class for
analysis. The ANALYSIS command is used to describe the technical
details of the analysis. The TYPE option is used to describe the type of
analysis that is to be performed. By selecting MIXTURE, a mixture
model will be estimated.

424
 Examples: Monte Carlo Simulation Studies

The commented out lines in the MODEL command show how the
MODEL command is changed from a two class model to a one class
model. An explanation of the other commands can be found in
Examples 12.1 and 8.1.

EXAMPLE 12.4: MONTE CARLO SIMULATION STUDY FOR

A TWO-LEVEL GROWTH MODEL FOR A CONTINUOUS
OUTCOME (THREE-LEVEL ANALYSIS)

TITLE: this is an example of a Monte Carlo

simulation study for a two-level growth
model for a continuous outcome (three-
level analysis)
MONTECARLO:
NAMES ARE y1-y4 x w;
NOBSERVATIONS = 1000;
NREPS = 500;
SEED = 58459;
CUTPOINTS = x (1) w (0);
MISSING = y1-y4;
NCSIZES = 3;
CSIZES = 40 (5) 50 (10) 20 (15);
WITHIN = x;
BETWEEN = w;
MODEL POPULATION:
%WITHIN%
x@1;
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4*.5;
iw ON x*1;
sw ON x*.25;
iw*1; sw*.2;
%BETWEEN%
w@1;
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib ON w*.5;
sb ON w*.25;
[ib*1 sb*.5];
ib*.2; sb*.1;

425
CHAPTER 12

MODEL MISSING:
[y1-y4@-1];
y1 ON x*.4;
y2 ON x*.8;
y3 ON x*1.6;
y4 ON x*3.2;
ANALYSIS: TYPE IS TWOLEVEL;
MODEL:
%WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4*.5;
iw ON x*1;
sw ON x*.25;
iw*1; sw*.2;
%BETWEEN%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib ON w*.5;
sb ON w*.25;
[ib*1 sb*.5];
ib*.2; sb*.1;
OUTPUT: TECH9 NOCHISQUARE;

In this example, data for the two-level growth model for a continuous
outcome (three-level analysis) described in Example 9.12 are generated
and analyzed. This Monte Carlo simulation study can be used to
estimate the power to detect that the binary cluster-level covariate w has
a significant effect on the growth slope factor sb.

The NCSIZES option is used to specify the number of unique cluster

sizes to be used in data generation. In the example above, there are three
unique cluster sizes. The CSIZES option is used to specify the number
of clusters and the sizes of the clusters to be used in data generation.
The CSIZES option specifies that 40 clusters of size 5, 50 clusters of
size 10, and 20 clusters of size 15 will be generated. The WITHIN
option is used to identify the variables in the data set that are measured
on the individual level and modeled only on the within level. They are
specified to have no variance in the between part of the model. The
variable x is an individual-level variable. The BETWEEN option is used
to identify the variables in the data set that are measured on the cluster
level and modeled only on the between level. The variable w is a
cluster-level variable. Variables not mentioned on the WITHIN or the
BETWEEN statements are measured on the individual level and can be
modeled on both the within and between levels. The NOCHISQUARE
option of the OUTPUT command is used to request that the chi-square

426
 Examples: Monte Carlo Simulation Studies

fit statistic not be computed. This reduces computational time. An

explanation of the other commands can be found in Examples 12.1 and
12.2 and Example 9.12.

EXAMPLE 12.5: MONTE CARLO SIMULATION STUDY FOR

AN EXPLORATORY FACTOR ANALYSIS WITH
CONTINUOUS FACTOR INDICATORS

TITLE: this is an example of a Monte Carlo

simulation study for an exploratory factor
analysis with continuous factor indicators
MONTECARLO:
NAMES ARE y1-y10;
NOBSERVATIONS = 500;
NREPS = 500;
MODEL POPULATION:
f1 BY y1-y7*.5;
f2 BY y4-y5*.25 y6-y10*.8;
f1-f2@1;
f1 WITH f2*.5;
y1-y10*.36;
MODEL: f1 BY y1-y7*.5 y8-y10*0 (*1);
f2 BY y1-y3*.0 y4-y5*.25 y6-y10*.8 (*1);
f1 WITH f2*.5;
y1-y10*.36;
OUTPUT: TECH9;

In this example, data are generated according to a two-factor CFA model

with continuous outcomes and analyzed as an exploratory factor analysis
using exploratory structural equation modeling (ESEM; Asparouhov &
Muthén, 2009a).

In the MODEL command, the BY statements specify that the factors f1

and f2 are measured by the continuous factor indicators y1 through y10.
The label 1 following an asterisk (*) in parentheses following the BY
statements is used to indicate that f1 and f2 are a set of EFA factors.
When no rotation is specified using the ROTATION option of the
ANALYSIS command, the default oblique GEOMIN rotation is used to
obtain factor loadings and factor correlations. The intercepts and
residual variances of the factor indicators are estimated and the residuals
are not correlated as the default. The variances of the factors are fixed at

427
CHAPTER 12

one as the default. The factors are correlated under the default oblique
GEOMIN rotation.

The default estimator for this type of analysis is maximum likelihood.

The ESTIMATOR option of the ANALYSIS command can be used to
select a different estimator. An explanation of the other commands can
be found in Examples 12.1 and 12.2.

EXAMPLE 12.6 STEP 1: MONTE CARLO SIMULATION

STUDY WHERE CLUSTERED DATA FOR A TWO-LEVEL
GROWTH MODEL FOR A CONTINUOUS OUTCOME
(THREE-LEVEL ANALYSIS) ARE GENERATED, ANALYZED,
AND SAVED

TITLE: this is an example of a Monte Carlo

simulation study where clustered data for
a two-level growth model for a continuous
outcome (three-level) analysis are
generated and analyzed
MONTECARLO:
NAMES ARE y1-y4 x w;
NOBSERVATIONS = 1000;
NREPS = 100;
SEED = 58459;
CUTPOINTS = x(1) w(0);
MISSING = y1-y4;
NCSIZES = 3;
CSIZES = 40 (5) 50 (10) 20 (15);
WITHIN = x; BETWEEN = w;
REPSAVE = ALL;
SAVE = ex12.6rep*.dat;
MODEL POPULATION:
%WITHIN%
x@1;
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4*.5;
iw ON x*1;
sw ON x*.25;
iw*1; sw*.2;
%BETWEEN%
w@1;
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib ON w*.5;

428
 Examples: Monte Carlo Simulation Studies

sb ON w*.25;
[ib*1 sb*.5];
ib*.2; sb*.1;
MODEL MISSING:
[y1-y4@-1];
y1 ON x*.4;
y2 ON x*.8;
y3 ON x*1.6;
y4 ON x*3.2;
ANALYSIS: TYPE = TWOLEVEL;
MODEL:
%WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3;
y1-y4*.5;
iw ON x*1;
sw ON x*.25;
iw*1; sw*.2;
%BETWEEN%
ib sb | y1@0 y2@1 y3@2 y4@3;
y1-y4@0;
ib ON w*.5;
sb ON w*.25;
[ib*1 sb*.5];
ib*.2; sb*.1;
OUTPUT: TECH8 TECH9;

In this example, clustered data are generated and analyzed for the two-
level growth model for a continuous outcome (three-level) analysis
described in Example 9.12. The data are saved for a subsequent external
Monte Carlo simulation study. The REPSAVE and SAVE options of the
MONTECARLO command are used to save some or all of the data sets
generated in a Monte Carlo simulation study. The REPSAVE option
specifies the numbers of the replications for which the data will be
saved. In the example above, the keyword ALL specifies that all of the
data sets will be saved. The SAVE option is used to name the files to
which the data sets will be written. The asterisk (*) is replaced by the
replication number. For example, data from the first replication will be
saved in the file named ex12.6rep1.dat. A file is also produced where
the asterisk (*) is replaced by the word list. The file, in this case
ex12.6replist.dat, contains the names of the generated data sets. The
ANALYSIS command is used to describe the technical details of the
analysis. By selecting TYPE=TWOLEVEL, a multilevel model is
estimated. An explanation of the other commands can be found in
Examples 12.1, 12.2, 12.4 and Example 9.12.

429
CHAPTER 12

EXAMPLE 12.6 STEP 2: EXTERNAL MONTE CARLO

ANALYSIS OF CLUSTERED DATA GENERATED FOR A
TWO-LEVEL GROWTH MODEL FOR A CONTINUOUS
OUTCOME USING TYPE=COMPLEX FOR A SINGLE-LEVEL
GROWTH MODEL

TITLE: this is an example of an external Monte

Carlo analysis of clustered data generated
for a two-level growth model for a
continuous outcome using TYPE=COMPLEX for
a single-level growth model
DATA: FILE = ex12.6replist.dat;
TYPE = MONTECARLO;
VARIABLE: NAMES = y1-y4 x w clus;
USEVARIABLES = y1-w;
MISSING = ALL (999);
CLUSTER = clus;
ANALYSIS: TYPE = COMPLEX;
MODEL: i s | y1@0 y2@1 y3@2 y4@3;
y1-y4*.5;
i ON x*1 w*.5;
s ON x*.25 w*.25;
i*1.2; s*.3;
[i*1 s*.5];
OUTPUT: TECH9;

In this example, an external Monte Carlo simulation study of clustered

data generated for a two-level growth model for a continuous outcome is
carried out using TYPE=COMPLEX for a single-level growth model.
The DATA command is used to provide information about the data sets
to be analyzed. The MONTECARLO setting of the TYPE option is used
when the data sets being analyzed have been generated and saved using
either the REPSAVE option of the MONTECARLO command or by
another computer program. The file named using the FILE option of the
DATA command contains a list of the names of the data sets to be
analyzed and summarized as in a Monte Carlo simulation study. This
file is created when the SAVE and REPSAVE options of the
MONTECARLO command are used to save Monte Carlo generated data
sets. The CLUSTER option of the VARIABLE command is used when
data have been collected under a complex survey data design to identify
the variable that contains cluster information. In the example above, the
variable clus contains cluster information. By selecting

430
 Examples: Monte Carlo Simulation Studies

TYPE=COMPLEX, an analysis is carried out that takes non-

independence of observations into account.

In external Monte Carlo simulation studies, the MODEL command is

also used to provide values for each parameter. These are used as the
population parameter values for the analysis model and are printed in the
first column of the output labeled Population. They are used for
computing coverage and as starting values in the estimation of the
model.

EXAMPLE 12.7 STEP 1: REAL DATA ANALYSIS OF A CFA

WITH COVARIATES (MIMIC) FOR CONTINUOUS FACTOR
INDICATORS WHERE THE PARAMETER ESTIMATES ARE
SAVED FOR USE IN A MONTE CARLO SIMULATION STUDY

TITLE: this is an example of a real data analysis

of a CFA with covariates (MIMIC) for
continuous factor indicators where the
parameter estimates are saved for use in a
Monte Carlo simulation study
DATA: FILE = ex12.7real.dat;
VARIABLE: NAMES = y1-y10 x1 x2;
MODEL: f1 BY y1@1 y2-y5*1;
f2 BY y6@1 y7-y10*1;
f1-f2*.5;
f1 WITH f2*.25;
y1-y5*.5;
[y1-y5*1];
y6-y10*.75;
[y6-y10*2];
f1 ON x1*.3 x2*.5;
f2 ON x1*.5 x2*.3;
OUTPUT: TECH1;
SAVEDATA: ESTIMATES = ex12.7estimates.dat;

In this example, parameter estimates from a real data analysis of a CFA

with covariates (MIMIC) for continuous factor indicators are saved for
use as population parameter values for use in data generation and
coverage in a subsequent internal Monte Carlo simulation study. The
ESTIMATES option of the SAVEDATA command is used to specify the
name of the file in which the parameter estimates of the analysis will be
saved.

431
CHAPTER 12

EXAMPLE 12.7 STEP 2: MONTE CARLO SIMULATION

STUDY WHERE PARAMETER ESTIMATES SAVED FROM A
REAL DATA ANALYSIS ARE USED FOR POPULATION
PARAMETER VALUES FOR DATA GENERATION AND
COVERAGE

TITLE: this is an example of a Monte Carlo

simulation study where parameter estimates
saved from a real data analysis are used
for population parameter values for data
generation and coverage
MONTECARLO:
NAMES ARE y1-y10 x1 x2;
NOBSERVATIONS = 500;
NREPS = 500;
SEED = 45335;
POPULATION = ex12.7estimates.dat;
COVERAGE = ex12.7estimates.dat;
MODEL POPULATION:
f1 BY y1-y5;
f2 BY y6-y10;
f1 ON x1 x2;
f2 ON x1 x2;
MODEL: f1 BY y1-y5;
f2 BY y6-y10;
f1 ON x1 x2;
f2 ON x1 x2;
OUTPUT: TECH9;

In this example, parameter estimates saved from a real data analysis are
used for population parameter values for data generation and coverage
using the POPULATION and COVERAGE options of the
MONTECARLO command. The POPULATION option is used to name
the data set that contains the population parameter values to be used in
data generation. The COVERAGE option is used to name the data set
that contains the parameter values to be used for computing coverage
and are printed in the first column of the output labeled Population. An
explanation of the other commands can be found in Example 12.1.

432
 Examples: Monte Carlo Simulation Studies

EXAMPLE 12.8: MONTE CARLO SIMULATION STUDY FOR

DISCRETE-TIME SURVIVAL ANALYSIS

TITLE: this is an example of a Monte Carlo

simulation study for discrete-time
survival analysis
MONTECARLO:
NAMES = u1-u4 x;
NOBSERVATIONS = 1000;
NREPS = 100;
GENERATE = u1-u4(1);
MISSING = u2-u4;
CATEGORICAL = u1-u4;
MODEL POPULATION:
[x@0]; x@1;
[u1$1*2 u2$1*1.5 u3$1*1 u4$1*1];
f BY u1-u4@1;
f ON x*.5;
f@0;
MODEL MISSING:
[u2-u4@-15];
u2 ON u1@30;
u3 ON u1-u2@30;
u4 ON u1-u3@30;
ANALYSIS: ESTIMATOR = MLR;
MODEL: [u1$1*2 u2$1*1.5 u3$1*1 u4$1*1];
f BY u1-u4@1;
f ON x*.5;
f@0;
OUTPUT: TECH8 TECH9;

In this example, data are generated and analyzed for a discrete-time

survival model like the one shown in Example 6.19. Maximum
likelihood estimation with discrete-time survival analysis for a non-
repeatable event requires that the event history indicators for an
individual are scored as missing after an event has occurred (Muthén &
Masyn, 2005). This is accomplished using the MODEL MISSING
command.

The MISSING option in the MONTECARLO command is used to

identify the dependent variables in the data generation model for which
missing data will be generated. The MODEL MISSING command is
used to provide information about the population parameter values for
the missing data model to be used in the generation of data. The

433
CHAPTER 12

MODEL MISSING command specifies a logistic regression model for a

set of binary dependent variables that represent not missing (scored as 0)
and missing (scored as 1) for the dependent variables in the data
generation model. The binary missing data indicators have the same
names as the dependent variables in the data generation model. The first
statement in the MODEL MISSING command defines the intercepts in
the logistic regressions for the binary dependent variables u2, u3, and u4.
If the covariates predicting missingness all have values of zero, the
logistic regression intercept value of -15 corresponds to a probability of
zero of having missing data on the dependent variables. The variable u1
has no missing values. The first ON statement describes the regression
of the missing value indicator u2 on the event-history variable u1 where
the logistic regression coefficient is fixed at 30 indicating that
observations with the value one on the event-history variable u1 result in
a logit value 15 for the missing value indicator u2 indicating that the
probability that the event-history variable u2 is missing is one. The
second ON statement describes the regression of the missing value
indicator u3 on the event-history variables u1 and u2 where the logistic
regression coefficients are fixed at 30 indicating that observations with
the value one on either or both of the event-history variables u1 and u2
result in a logit value of at least 15 for the missing value indicator u3
indicating that the probability that the event-history variable u3 is
missing is one. The third ON statement describes the regression of the
missing value indicator u4 on the event-history variables u1, u2, and u3
where the logistic regression coefficients are fixed at 30 indicating that
observations with the value one on one or more of the event-history
variables u1, u2, and u3 result in a logit value of at least 15 for the
missing value indicator u4 indicating that the probability that the event-
history variable u4 is missing is one. An explanation of the other
commands can be found in Examples 12.1 and 12.3.

434
 Examples: Monte Carlo Simulation Studies

EXAMPLE 12.9: MONTE CARLO SIMULATION STUDY FOR

A TWO-PART (SEMICONTINUOUS) GROWTH MODEL FOR
A CONTINUOUS OUTCOME

TITLE: this is an example of a Monte Carlo

simulation study for a two-part
(semicontinuous) growth model for a
continuous outcome
MONTECARLO:
NAMES = u1-u4 y1-y4;
NOBSERVATIONS = 500;
NREPS = 100;
GENERATE = u1-u4(1);
MISSING = y1-y4;
CATEGORICAL = u1-u4;
MODEL POPULATION:
iu su | u1@0 u2@1 u3@2 u4@3;
[u1$1-u4$1*-.5] (1);
[iu@0 su*.85];
iu*1.45;
iy sy | y1@0 y2@1 y3@2 y4@3;
[y1-y4@0];
y1-y4*.5;
[iy*.5 sy*1];
iy*1;
sy*.2;
iy WITH sy*.1;
iu WITH iy*0.9;
MODEL MISSING:
[y1-y4@15];
y1 ON u1@-30;
y2 ON u2@-30;
y3 ON u3@-30;
y4 ON u4@-30;
ANALYSIS: ESTIMATOR = MLR;

435
CHAPTER 12

MODEL: iu su | u1@0 u2@1 u3@2 u4@3;

[u1$1-u4$1*-.5] (1);
[iu@0 su*.85];
iu*1.45;
su@0;
iy sy | y1@0 y2@1 y3@2 y4@3;
[y1-y4@0];
y1-y4*.5;
[iy*.5 sy*1];
iy*1;
sy*.2;
iy WITH sy*.1;
iu WITH iy*0.9;
iu WITH sy@0;
OUTPUT: TECH8;

In this example, data are generated and analyzed for a two-part

(semicontinuous) growth model for a continuous outcome like the one
shown in Example 6.16. If these data are saved for subsequent two-part
analysis using the DATA TWOPART command, an adjustment to the
saved data must be made using the DEFINE command as part of the
analysis. If the values of the continuous outcomes y are not 999 which is
the value used as the missing data flag in the saved data, the exponential
function must be applied to the continuous variables. After that
transformation, the value 999 must be changed to zero for the continuous
variables. This represents the floor of the scale.

The MISSING option in the MONTECARLO command is used to

identify the dependent variables in the data generation model for which
missing data will be generated. The MODEL MISSING command is
used to provide information about the population parameter values for
the missing data model to be used in the generation of data. The
MODEL MISSING command specifies a logistic regression model for a
set of binary dependent variables that represent not missing (scored as 0)
and missing (scored as 1) for the dependent variables in the data
generation model. The binary missing data indicators have the same
names as the dependent variables in the data generation model. The first
statement in the MODEL MISSING command defines the intercepts in
the logistic regressions for the binary dependent variables y1, y2, y3, and
y4. If the covariates predicting missingness all have values of zero, the
logistic regression intercept value of 15 corresponds to a probability of
one of having missing data on the dependent variables. The four ON
statements describe the regressions of the missing value indicators y1,
y2, y3, and y4 on the binary outcomes u1, u2, u3, and u4 where the

436
 Examples: Monte Carlo Simulation Studies

logistic regression coefficient is fixed at -30. This results in

observations with the value one on u1, u2, u3, and u4 giving logit values
-15 for the binary missing data indicators. A logit value -15 implies that
the probability that the continuous outcomes y are missing is zero. An
explanation of the other commands can be found in Examples 12.1 and
12.3.

EXAMPLE 12.10: MONTE CARLO SIMULATION STUDY FOR

A TWO-LEVEL CONTINUOUS-TIME SURVIVAL ANALYSIS
USING COX REGRESSION WITH A RANDOM INTERCEPT
AND A FRAILTY

TITLE: this is an example of a Monte Carlo

simulation study for a two-level
continuous-time survival analysis using
Cox regression with a random intercept and
a frailty
MONTECARLO:
NAMES = t x w;
NOBSERVATIONS = 1000;
NREPS = 100;
GENERATE = t(s 20*1);
NCSIZES = 3;
CSIZES = 40 (5) 50 (10) 20 (15);
HAZARDC = t (.5);
SURVIVAL = t (ALL);
WITHIN = x;
BETWEEN = w;
MODEL POPULATION:
%WITHIN%
x@1;
t ON x*.5;
%BETWEEN%
w@1;
[t#1-t#21*1];
t ON w*.2;
t*0.5;
ANALYSIS: TYPE = TWOLEVEL;
BASEHAZARD = OFF;
MODEL: %WITHIN%
t ON x*.5;
%BETWEEN%
t ON w*.2;
t*0.5;

437
CHAPTER 12

In this example, data are generated and analyzed for the two-level
continuous-time survival analysis using Cox regression with a random
intercept and a frailty shown in Example 9.16. Monte Carlo simulation
of continuous-time survival models is described in Asparouhov et al.
(2006).

The GENERATE option is used to specify the scale of the dependent

variables for data generation. In this example, the dependent variable t
is a time-to-event variable. The numbers in parentheses specify that
twenty time intervals of length one will be used for data generation. The
HAZARDC option is used to specify the hazard for the censoring
process in continuous-time survival analysis when time-to-event
variables are generated. This information is used to create a censoring
indicator variable where zero is not censored and one is right censored.
A hazard for censoring of .5 is specified for the time-to-event variable t
by placing the number .5 in parentheses following the variable name.
The SURVIVAL option is used to identify the analysis variables that
contain information about time to event and to provide information
about the time intervals in the baseline hazard function to be used in the
analysis. The keyword ALL is used if the time intervals are taken from
the data.

The ANALYSIS command is used to describe the technical details of the

analysis. By selecting TYPE=TWOLEVEL, a multilevel model will be
estimated. The BASEHAZARD option is used with continuous-time
survival analysis to specify if a non-parametric or a parametric baseline
hazard function is used in the estimation of the model. The default is
OFF which uses the non-parametric baseline hazard function.

The MODEL command is used to describe the analysis model as in

regular analyses. In the within part of the model, the ON statement
describes the loglinear regression of the time-to-event variable t on the
covariate x. In the between part of the model, the ON statement
describes the linear regression of the random intercept of the time-to-
event variable t on the covariate w. The residual variance of t is
estimated and represents a frailty parameter. A detailed explanation of
the MODEL command can be found in Examples 12.1 and 12.4.

438
 Examples: Monte Carlo Simulation Studies

EXAMPLE 12.11: MONTE CARLO SIMULATION STUDY FOR

A TWO-LEVEL MEDIATION MODEL WITH RANDOM
SLOPES

TITLE: this is an example of a Monte Carlo

simulation study for a two-level mediation
model with random slopes
MONTECARLO:
NAMES ARE y m x;
WITHIN = x;
NOBSERVATIONS = 1000;
NCSIZES = 1;
CSIZES = 100 (10);
NREP = 100;
ANALYSIS: TYPE = TWOLEVEL RANDOM;
MODEL POPULATION:
%WITHIN%
x@1;
c | y ON x;
b | y ON m;
a | m ON x;
m*1; y*1;
%BETWEEN%
y WITH m*0.1 b*0.1 a*0.1 c*0.1;
m WITH b*0.1 a*0.1 c*0.1;
a WITH b*0.1 (cab);
a WITH c*0.1;
b WITH c*0.1;
y*1 m*1 a*1 b*1 c*1;
[a*0.4] (ma);
[b*0.5] (mb);
[c*0.6];

439
CHAPTER 12

MODEL:
%WITHIN%
c | y ON x;
b | y ON m;
a | m ON x;
m*1; y*1;
%BETWEEN%
y WITH m*0.1 b*0.1 a*0.1 c*0.1;
m WITH b*0.1 a*0.1 c*0.1;
a WITH b*0.1 (cab);
a WITH c*0.1;
b WITH c*0.1;
y*1 m*1 a*1 b*1 c*1;
[a*0.4] (ma);
[b*0.5] (mb);
[c*0.6];
MODEL CONSTRAINT:
NEW(m*0.3);
m=ma*mb+cab;

In this example, data for a two-level mediation model with a random

slope are generated and analyzed. For related modeling see Bauer et al.
(2006).

The TYPE option is used to describe the type of analysis that is to be

performed. By selecting TWOLEVEL RANDOM, a multilevel model
with random intercepts and random slopes will be estimated. In the
MODEL command, the | statement is used to name and define the
random slopes c, b, and a. The random intercept uses the name of the
dependent variables c, b, and a. The ON statements on the right-hand
side of the | statements describe the linear regressions that have a
random slope.

The label cab is assigned to the covariance between the random slopes a
and b. The labels ma and mb are assigned to the means of the random
slopes a and b. These labels are used in the MODEL CONSTRAINT
command. The MODEL CONSTRAINT command is used to define
linear and non-linear constraints on the parameters in the model. In the
MODEL CONSTRAINT command, the NEW option is used to
introduce a new parameter that is not part of the MODEL command.
The new parameter m is the indirect effect of the covariate x on the
outcome y. The two outcomes y and m can also be categorical. For a
discussion of indirect effects when the outcome y is categorical, see
MacKinnon et al. (2007).

440
 Examples: Monte Carlo Simulation Studies

The default estimator for this type of analysis is maximum likelihood

with robust standard errors. An explanation of the other commands can
be found in Examples 12.1 and 12.4.

EXAMPLE 12:12: MONTE CARLO SIMULATION STUDY

FOR A MULTIPLE GROUP EFA WITH CONTINUOUS
FACTOR INDICATORS WITH MEASUREMENT
INVARIANCE OF INTERCEPTS AND FACTOR LOADINGS

TITLE: this is an example of a Monte Carlo

simulation study for a multiple group EFA
with continuous factor indicators with
measurement invariance of intercepts and
factor loadings
MONTECARLO:
NAMES ARE y1-y10;
NOBSERVATIONS = 500 500;
NREPS = 1;
NGROUPS = 2;
MODEL POPULATION:
f1 BY y1-y5*.8 y6-y10*0;
f2 BY y1-y5*0 y6-y10*.8;
f1-f2@1;
f1 WITH f2*.5;
y1-y10*1;
[y1-y10*1];
[f1-f2@0];
MODEL POPULATION-g2:
f1*1.5 f2*2;
f1 WITH f2*1;
y1-y10*2;
[f1*.5 f2*.8];
MODEL: f1 BY y1-y5*.8 y6-y10*0 (*1);
f2 BY y1-y5*0 y6-y10*.8 (*1);
f1-f2@1;
f1 WITH f2*.5;
y1-y10*1;
[y1-y10*1]; [f1-f2@0];
MODEL g2: f1*1.5 f2*2;
f1 WITH f2*1;
y1-y10*2; [f1*.5 f2*.8];
OUTPUT: TECH9;

In this example, data are generated and analyzed according to a multiple

group EFA model with continuous factor indicators with measurement

441
CHAPTER 12

invariance across groups of intercepts and factor loadings. This model is

described in Example 5.27. The NOBSERVATIONS option specifies
the number of observations for each group. The NGROUPS option
specifies the number of groups. In this study data for two groups of 500
observations are generated and analyzed. One difference between the
MODEL command when EFA factors are involved rather than CFA
factors is that the values given using the asterisk (*) are used only for
coverage. Starting values are not allowed for the factor loading and
factor covariance matrices for EFA factors. An explanation of the other
commands can be found in Example 12.1 and Example 5.27.

442
 Examples: Special Features

CHAPTER 13
EXAMPLES: SPECIAL FEATURES

In this chapter, special features not illustrated in the previous example

chapters are discussed. A cross-reference to the original example is
given when appropriate.

Following is the set of special feature examples included in this chapter:

• 13.1: A covariance matrix as data

• 13.2: Means and a covariance matrix as data
• 13.3: Reading data with a fixed format
• 13.4: Non-numeric missing value flags
• 13.5: Numeric missing value flags
• 13.6: Selecting observations and variables
• 13.7: Transforming variables using the DEFINE command
• 13.8: Freeing and fixing parameters and giving starting values
• 13.9: Equalities in a single group analysis
• 13.10: Equalities in a multiple group analysis
• 13.11: Using PWITH to estimate adjacent residual covariances
• 13.12: Chi-square difference testing for WLSMV and MLMV
• 13.13: Analyzing multiple imputation data sets
• 13.14: Saving data
• 13.15: Saving factor scores
• 13.16: Using the PLOT command
• 13.17: Merging data sets
• 13.18: Using replicate weights
• 13.19: Generating, using, and saving replicate weights

443
CHAPTER 13

EXAMPLE 13.1: A COVARIANCE MATRIX AS DATA

TITLE: this is an example of a CFA with

continuous factor indicators using a
covariance matrix as data
DATA: FILE IS ex5.1.dat;
TYPE = COVARIANCE;
NOBSERVATIONS = 1000;
VARIABLE: NAMES ARE y1-y6;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;

The example above is based on Example 5.1 in which individual data are
analyzed. In this example, a covariance matrix is analyzed. The TYPE
option is used to specify that the input data set is a covariance matrix.
The NOBSERVATIONS option is required for summary data and is
used to indicate how many observations are in the data set used to create
the covariance matrix. Summary data are required to be in an external
data file in free format. Following is an example of the data:

1.0
.86 1.0
.56 .76 1.0
.78 .34 .48 1.0
.65 .87 .32 .56 1.0
.66 .78 .43 .45 .33 1.0

EXAMPLE 13.2: MEANS AND A COVARIANCE MATRIX AS

DATA

TITLE: this is an example of a mean structure CFA

with continuous factor indicators using
means and a covariance matrix as data
DATA: FILE IS ex5.9.dat;
TYPE IS MEANS COVARIANCE;
NOBSERVATIONS = 1000;
VARIABLE: NAMES ARE y1a-y1c y2a-y2c;
MODEL: f1 BY y1a y1b@1 y1c@1;
f2 BY y2a y2b@1 y2c@1;
[y1a y1b y1c] (1);
[y2a y2b y2c] (2);

444
 Examples: Special Features

The example above is based on Example 5.9 in which individual data are
analyzed. In this example, means and a covariance matrix are analyzed.
The TYPE option is used to specify that the input data set contains
means and a covariance matrix. The NOBSERVATIONS option is
required for summary data and is used to indicate how many
observations are in the data set used to create the means and covariance
matrix. Summary data are required to be in an external data file in free
format. Following is an example of the data. The means come first
followed by the covariances. The covariances must start on a new
record.

.4 .6 .3 .5
1.0
.86 1.0
.56 .76 1.0
.78 .34 .48 1.0

EXAMPLE 13.3: READING DATA WITH A FIXED FORMAT

TITLE: this is an example of a CFA with

covariates (MIMIC) with continuous factor
indicators using data in a fixed format
DATA: FILE IS ex5.8.dat;
FORMAT IS 3f4.2 3f2 f1 2f2;
VARIABLE: NAMES ARE y1-y6 x1-x3;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;

The example above is based on Example 5.8 in which individual data

with a free format are analyzed. Because the data are in free format, a
FORMAT statement is not required. In this example, the data have a
fixed format. The inclusion of a FORMAT statement is required in this
situation. The FORMAT statement describes the position of the nine
variables in the data set. In this example, the first three variables take up
four columns each and are read such that two digits follow the decimal
point (3f4.2). The next three variables take three columns with no digits
after the decimal point (3f2). The seventh variable takes one column
with no digits following the decimal point (f1), and the eighth and ninth
variables each take two columns with no digits following the decimal
point (2f2).

445
CHAPTER 13

EXAMPLE 13.4: NON-NUMERIC MISSING VALUE FLAGS

TITLE: this is an example of a SEM with

continuous factor indicators using data
with non-numeric missing value flags
DATA: FILE IS ex5.11.dat;
VARIABLE: NAMES ARE y1-y12;
MISSING = *;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f4 ON f3;
f3 ON f1 f2;

The example above is based on Example 5.11 in which the data contain
no missing values. In this example, there are missing values and the
asterisk (*) is used as a missing value flag. The MISSING option is used
to identify the values or symbol in the analysis data set that will be
treated as missing or invalid. Non-numeric missing value flags are
applied to all variables in the data set.

EXAMPLE 13.5: NUMERIC MISSING VALUE FLAGS

TITLE: this is an example of a SEM with

continuous factor indicators using data
with numeric missing value flags
DATA: FILE IS ex5.11.dat;
VARIABLE: NAMES ARE y1-y12;
MISSING = y1-y3(9) y4(9 99) y5-y12(9-12);
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f3 BY y7-y9;
f4 BY y10-y12;
f4 ON f3;
f3 ON f1 f2;

The example above is based on Example 5.11 in which the data contain
no missing values. In this example, there are missing values and
numeric missing value flags are used. The MISSING option is used to
identify the values or symbol in the analysis data set that will be treated
as missing or invalid. Numeric missing value flags can be applied to a

446
 Examples: Special Features

single variable, to groups of variables, or to all of the variables in a data

set. In the example above, y1, y2, and y3 have a missing value flag of 9;
y4 has missing value flags of 9 and 99; and y5 through y12 have missing
value flags of 9, 10, 11, and 12. If all variables in a data set have the
same missing value flags, the keyword ALL can be used as follows:

MISSING = ALL (9);

to indicate that all variables have the missing value flag of 9.

EXAMPLE 13.6: SELECTING OBSERVATIONS AND

VARIABLES

TITLE: this is an example of a path analysis

with continuous dependent variables using
a subset of the data
DATA: FILE IS ex3.11.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES ARE y1-y3 x1-x3;
USEOBSERVATION ARE (x4 EQ 2);
MODEL: y1 y2 ON x1 x2 x3;
y3 ON y1 y2 x2;

The example above is based on Example 3.11 in which the entire data
set is analyzed. In this example, a subset of variables and a subset of
observations are analyzed. The USEVARIABLES option is used to
select variables for an analysis. In the example above, y1, y2, y3, x1, x2,
and x3 are selected. The USEOBSERVATIONS option is used to select
observations for an analysis by specifying a conditional statement. In
the example above, individuals with the value of 2 on variable x4 are
included in the analysis.

447
CHAPTER 13

EXAMPLE 13.7: TRANSFORMING VARIABLES USING THE

DEFINE COMMAND

TITLE: this is an example of a path analysis

with continuous dependent variables where
two variables are transformed
DATA: FILE IS ex3.11.dat;
DEFINE: y1 = y1/100;
x3 = SQRT(x3);
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEVARIABLES = y1-y3 x1-x3;
MODEL: y1 y2 ON x1 x2 x3;
y3 ON y1 y2 x2;

The example above is based on Example 3.11 where the variables are
not transformed. In this example, two variables are transformed using
the DEFINE command. The variable y1 is transformed by dividing it by
100. The variable x3 is transformed by taking the square root of it. The
transformed variables are used in the estimation of the model. The
DEFINE command can also be used to create new variables.

EXAMPLE 13.8: FREEING AND FIXING PARAMETERS AND

GIVING STARTING VALUES

TITLE: this is an example of a CFA with

continuous factor indicators where
parameters are freed, fixed, and starting
values are given
DATA: FILE IS ex5.1.dat;
VARIABLE: NAMES ARE y1-y6;
MODEL: f1 BY y1* y2*.5 y3;
f2 BY y4* y5 y6*.8;
f1-f2@1;

The example above is based on Example 5.1 where default starting

values are used. In this example, parameters are freed, assigned starting
values, and fixed. In the two BY statements, the factor loadings for y1
and y4 are fixed at one as the default because they are the first variable
following the BY statement. This is done to set the metric of the factors.
To free these parameters, an asterisk (*) is placed after y1 and y4. The
factor loadings for variables y2, y3, y5, and y6 are free as the default

448
 Examples: Special Features

with starting values of one. To assign starting values to y2 and y6, an

asterisk (*) followed by a number is placed after y2 and y6. The starting
value of .5 is assigned to y2, and the starting value of .8 is assigned to
y6. The variances of f1 and f2 are free to be estimated as the default.
To fix these variances to one, an @ symbol followed by 1 is placed after
f1 and f2 in a list statement. This is another way to set the metric of the
factors.

EXAMPLE 13.9: EQUALITIES IN A SINGLE GROUP

ANALYSIS

TITLE: this is an example of a CFA with

continuous factor indicators with
equalities
DATA: FILE IS ex5.1.dat;
VARIABLE: NAMES ARE y1-y6;
MODEL: f1 BY y1
y2-y3 (1-2);
f2 BY y4
y5-y6 (1-2);
y1-y3 (3);
y4-y6 (4);

This example is based on the model in Example 5.1 where there are no
equality constraints on model parameters. In the example above, several
model parameters are constrained to be equal. Equality constraints are
specified by placing the same number in parentheses following the
parameters that are to be held equal. The label (1-2) following the factor
loadings uses the list function to assign equality labels to these
parameters. The label 1 is assigned to the factor loadings of y2 and y5
which holds these factor loadings equal. The label 2 is assigned to the
factor loadings of y3 and y6 which holds these factor loadings equal.
The third equality statement holds the residual variances of y1, y2, and
y3 equal using the label (3), and the fourth equality statement holds the
residual variances of y4, y5, and y6 equal using the label (4).

449
CHAPTER 13

EXAMPLE 13.10: EQUALITIES IN A MULTIPLE GROUP

ANALYSIS

TITLE: this is an example of a multiple group CFA

with covariates (MIMIC) with continuous
factor indicators and a mean structure
with between and within group equalities
DATA: FILE IS ex5.15.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3 g;
GROUPING IS g (1=g1 2=g2 3=g3);
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;
f1 (1);
y1-y3 (2);
y4-y6 (3-5);
MODEL g1: f1 BY y3*;
[y3*];
f2 (6);
MODEL g3: f2 (6);

This example is based on Example 5.15 in which the model has two
groups. In this example, the model has three groups. Parameters are
constrained to be equal by placing the same number in parentheses
following the parameters that will be held equal. In multiple group
analysis, the overall MODEL command is used to set equalities across
groups. The group-specific MODEL commands are used to specify
equalities for specific groups or to relax equalities specified in the
overall MODEL command. In the example above, the first equality
statement holds the variance of f1 equal across the three groups in the
analysis using the equality label 1. The second equality statement holds
the residual variances of y1, y2, and y3 equal to each other and equal
across groups using the equality label 2. The third equality statement
uses the list function to hold the residual variance of y4, y5, and y6 equal
across groups by assigning the equality label 3 to the residual variance of
y4, the label 4 to the residual variance of y5, and the label 5 to the
residual variance of y6. The fourth and fifth equality statements hold
the variance of f2 equal across groups g1 and g3 using the equality label
6.

450
 Examples: Special Features

EXAMPLE 13.11: USING PWITH TO ESTIMATE ADJACENT

RESIDUAL COVARIANCES

TITLE: this is an example of a linear growth

model for a continuous outcome with
adjacent residual covariances
DATA: FILE IS ex6.1.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;
y11-y13 PWITH y12-y14;

The example above is based on Example 6.1 in which a linear growth

model with no residual covariances for the outcome is estimated. In this
example, the PWITH option is used to specify adjacent residual
covariances. The PWITH option pairs the variables on the left-hand side
of the PWITH statement with the variables on the right-hand side of the
PWITH statement. Residual covariances are estimated for the pairs of
variables. In the example above, residual covariances are estimated for
y11 with y12, y12 with y13, and y13 with y14.

EXAMPLE 13.12: CHI-SQUARE DIFFERENCE TESTING FOR

WLSMV AND MLMV

This example shows the two steps needed to do a chi-square difference

test using the WLSMV and MLMV estimators. For these estimators, the
conventional approach of taking the difference between the chi-square
values and the difference in the degrees of freedom is not appropriate
because the chi-square difference is not distributed as chi-square. This
example is based on Example 5.3.

451
CHAPTER 13

TITLE: this is an example of the first step

needed for a chi-square difference test
for the WLSMV or the MLMV estimator
DATA: FILE IS ex5.3.dat;
VARIABLE: NAMES ARE u1-u3 y4-y9;
CATEGORICAL ARE u1 u2 u3;
MODEL: f1 BY u1-u3;
f2 BY y4-y6;
f3 BY y7-y9;
SAVEDATA: DIFFTEST IS deriv.dat;

The input setup above shows the first step needed to do a chi-square
difference test for the WLSMV and MLMV estimators. In this analysis,
the less restrictive H1 model is estimated. The DIFFTEST option of the
SAVEDATA command is used to save the derivatives of the H1 model
for use in the second step of the analysis. The DIFFTEST option is used
to specify the name of the file in which the derivatives from the H1
model will be saved. In the example above, the file name is deriv.dat.

TITLE: this is an example of the second step

needed for a chi-square difference test
for the WLSMV or the MLMV estimator
DATA: FILE IS ex5.3.dat;
VARIABLE: NAMES ARE u1-u3 y4-y9;
CATEGORICAL ARE u1 u2 u3;
ANALYSIS: DIFFTEST IS deriv.dat;
MODEL: f1 BY u1-u3;
f2 BY y4-y6;
f3 BY y7-y9;
f1 WITH f2-f3@0;
f2 WITH f3@0;

The input setup above shows the second step needed to do a chi-square
difference test for the WLSMV and MLMV estimators. In this analysis,
the more restrictive H0 model is estimated. The restriction is that the
covariances among the factors are fixed at zero in this model. The
DIFFTEST option of the ANALYSIS command is used to specify the
name of the file that contains the derivatives of the H1 model that was
estimated in the first step of the analysis. This file is deriv.dat.

452
 Examples: Special Features

EXAMPLE 13.13: ANALYZING MULTIPLE IMPUTATION

DATA SETS

TITLE: this is an example of a CFA with

continuous factor indicators using
multiple imputation data sets
DATA: FILE IS implist.dat;
TYPE = IMPUTATION;
VARIABLE: NAMES ARE y1-y6;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;

The example above is based on Example 5.1 in which a single data set is
analyzed. In this example, data sets generated using multiple imputation
are analyzed. The FILE option of the DATA command is used to give
the name of the file that contains the names of the multiple imputation
data sets to be analyzed. The file named using the FILE option of the
DATA command must contain a list of the names of the multiple
imputation data sets to be analyzed. This file must be created by the
user unless the data are imputed using the DATA IMPUTATION
command in which case the file is created as part of the multiple
imputation. Each record of the file must contain one data set name. For
example, if five data sets are being analyzed, the contents of implist.dat
would be:

imp1.dat
imp2.dat
imp3.dat
imp4.dat
imp5.dat

where imp1.dat, imp2.dat, imp3.dat, imp4.dat, and imp5.dat are the

names of the five data sets created using multiple imputation.

When TYPE=IMPUTATION is specified, an analysis is carried out for

each data set in the file named using the FILE option. Parameter
estimates are averaged over the set of analyses, and standard errors are
computed using the average of the standard errors over the set of
analyses and the between analysis parameter estimate variation (Schafer,
1997).

453
CHAPTER 13

EXAMPLE 13.14: SAVING DATA

TITLE: this is an example of a path analysis

with continuous dependent variables using
a subset of the data which is saved for
future analysis
DATA: FILE IS ex3.11.dat;
VARIABLE: NAMES ARE y1-y6 x1-x4;
USEOBSERVATION ARE (x4 EQ 2);
USEVARIABLES ARE y1-y3 x1-x3;
MODEL: y1 y2 ON x1 x2 x3;
y3 ON y1 y2 x2;
SAVEDATA: FILE IS regress.sav;

The example above is based on Example 3.11 in which the analysis data
are not saved. In this example, the SAVEDATA command is used to
save the analysis data set. The FILE option is used to specify the name
of the ASCII file in which the individual data used in the analysis will be
saved. In this example, the data will be saved in the file regress.sav.
The data are saved in fixed format as the default unless the FORMAT
option of the SAVEDATA command is used.

EXAMPLE 13.15: SAVING FACTOR SCORES

TITLE: this is an example of a CFA with

covariates (MIMIC) with continuous factor
indicators where factor scores are
estimated and saved
DATA: FILE IS ex5.8.dat;
VARIABLE: NAMES ARE y1-y6 x1-x3;
MODEL: f1 BY y1-y3;
f2 BY y4-y6;
f1 f2 ON x1-x3;
SAVEDATA: FILE IS mimic.sav; SAVE = FSCORES;

The example above is based on Example 5.8 in which factor scores are
not saved. In this example, the SAVEDATA command is used to save
the analysis data set and factor scores. The FILE option is used to
specify the name of the ASCII file in which the individual data used in
the analysis will be saved. In this example, the data will be saved in the
file mimic.sav. The SAVE option is used to specify that factor scores
will be saved along with the analysis data. The data are saved in fixed

454
 Examples: Special Features

format as the default unless the FORMAT option of the SAVEDATA

command is used.

EXAMPLE 13.16: USING THE PLOT COMMAND

TITLE: this is an example of a linear growth

model for a continuous outcome
DATA: FILE IS ex6.1.dat;
VARIABLE: NAMES ARE y11-y14 x1 x2 x31-x34;
USEVARIABLES ARE y11-y14;
MODEL: i s | y11@0 y12@1 y13@2 y14@3;
PLOT: SERIES = y11-y14 (s);
TYPE = PLOT3;

The example above is based on Example 6.1 in which no graphical

displays of observed data or analysis results are requested. In this
example, the PLOT command is used to request graphical displays of
observed data and analysis results. These graphical outputs can be
viewed after the Mplus analysis is completed using a post-processing
graphics module. The SERIES option is used to list the names of a set
of variables along with information about the x-axis values to be used in
the graphs. For growth models, the set of variables is the repeated
measures of the outcome over time, and the x-axis values are the time
scores in the growth model. In the example above, the s in parentheses
after the variables listed in the SERIES statement is the name of the
slope growth factor. This specifies that the x-axis values are the time
scores values specified in the growth model. In this example, they are 0,
1, 2, and 3. Other ways to specify x-axis values are described in Chapter
18. The TYPE option is used to request specific plots. The TYPE
option of the PLOT command is described in Chapter 18.

455
CHAPTER 13

EXAMPLE 13.17: MERGING DATA SETS

TITLE: this is an example of merging two data

sets
DATA: FILE IS data1.dat;
VARIABLE: NAMES ARE id y1-y4;
IDVARIABLE IS id;
USEVARIABLES = y1 y2;
MISSING IS *;
ANALYSIS: TYPE = BASIC;
SAVEDATA: MFILE = data2.dat;
MNAMES ARE id y5-y8;
MFORMAT IS F6 4F2;
MSELECT ARE y5 y8;
MMISSING = y5-y8 (99);
FILE IS data12.sav;
FORMAT IS FREE;
MISSFLAG = 999;

This example shows how to merge two data sets using TYPE=BASIC.
Merging can be done with any analysis type. The first data set data1.dat
is named using the FILE option of the DATA command. The second
data set data2.dat is named using the MFILE option of the SAVEDATA
command. The NAMES option of the VARIABLE command gives the
names of the variables in data1.dat. The MNAMES option of the
SAVEDATA command gives the names of the variables in data2.dat.
The IDVARIABLE option of the VARIABLE command gives the name
of the variable to be used for merging. This variable must appear on
both the NAMES and MNAMES statements. The merged data set
data12.dat is saved in the file named using the FILE option of the
SAVEDATA command. The default format for this file is free and the
default missing value flag is the asterisk (*). These defaults can be
changed using the FORMAT and MISSFLAG options as shown above.
In the merged data set data12.dat, the missing value flags of asterisk (*)
in data1.dat and 99 in data2.dat are replaced by 999.

For data1.dat, the USEVARIABLES option of the VARIABLE

command is used to select a subset of the variables to be in the analysis
and for merging. The MISSING option of the VARIABLE command is
used to identify the values or symbol in the data set that are treated as
missing or invalid. In data1.dat, the asterisk (*) is the missing value

456
 Examples: Special Features

flag. If the data are not in free format, the FORMAT statement can be
used to specify a fixed format.

For data2.dat, the MFORMAT option is used to specify a format if the

data are not in the default free format. The MSELECT option is used to
select a subset of the variables to be used for merging. The MMISSING
option is used to identify the values or symbol in the data set that are
treated as missing or invalid.

EXAMPLE 13.18: USING REPLICATE WEIGHTS

TITLE: this is an example of using replicate

weights
DATA: FILE IS rweights.dat;
VARIABLE: NAMES ARE y1-y4 weight r1-r80;
WEIGHT = weight;
REPWEIGHTS = r1-r80;
ANALYSIS: TYPE = COMPLEX;
REPSE = JACKKNIFE1;
MODEL: f BY y1-y4;

This example shows how to use replicate weights in a factor analysis.

Replicate weights summarize information about a complex sampling
design. The WEIGHT option must be used when the REPWEIGHTS
option is used. The WEIGHT option is used to identify the variable that
contains sampling weight information. In this example, the sampling
weight variable is weight. The REPWEIGHTS option is used to identify
the replicate weight variables. These variables are used in the estimation
of standard errors of parameter estimates (Asparouhov & Muthén,
2009b). The data set in this example contains 80 replicate weights
variables, r1 through r80. The STRATIFICATION and CLUSTER
options may not be used in conjunction with the REPWEIGHTS option.
Analysis using replicate weights is available only with
TYPE=COMPLEX. The REPSE option is used to specify the
resampling method that was used to create the replicate weights. The
setting JACKKNIFE1 specifies that Jackknife draws were used.

457
CHAPTER 13

EXAMPLE 13.19: GENERATING, USING, AND SAVING

REPLICATE WEIGHTS

TITLE: this is an example of generating, using,

and saving replicate weights
DATA: FILE IS ex13.19.dat;
VARIABLE: NAMES ARE y1-y4 weight strat psu;
WEIGHT = weight;
STRATIFICATION = strat;
CLUSTER = psu;
ANALYSIS: TYPE = COMPLEX;
REPSE = BOOTSTRAP;
BOOTSTRAP = 100;
MODEL: f BY y1-y4;
SAVEDATA: FILE IS rweights.sav;
SAVE = REPWEIGHTS;

This example shows how to generate, use, and save replicate weights in
a factor analysis. Replicate weights summarize information about a
complex sampling design (Korn & Graubard, 1999; Lohr, 1999;
Asparouhov & Muthén, 2009b). When replicate weights are generated,
the REPSE option of the ANALYSIS command and the WEIGHT option
of the VARIABLE command along with the STRATIFICATION and/or
CLUSTER options of the VARIABLE command are used. The
WEIGHT option is used to identify the variable that contains sampling
weight information. In this example, the sampling weight variable is
weight. The STRATIFICATION option is used to identify the variable
in the data set that contains information about the subpopulations from
which independent probability samples are drawn. In this example, the
variable is strat. The CLUSTER option is used to identify the variable
in the data set that contains clustering information. In this example, the
variable is psu. Replicate weights can be generated and analyzed only
with TYPE=COMPLEX. The REPSE option is used to specify the
resampling method that will be used to create the replicate weights. The
setting BOOTSTRAP specifies that bootstrap draws will be used. The
BOOTSTRAP option specifies that 100 bootstrap draws will be carried
out. When replicate weights are generated, they can be saved for further
analysis using the FILE and SAVE options of the SAVEDATA
command. Replicate weights will be saved along with the other analysis
variables in the file named rweights.sav.

458
 Special Modeling Issues

CHAPTER 14
SPECIAL MODELING ISSUES

In this chapter, the following special modeling issues are discussed:

• Model estimation
• Multiple group analysis
• Missing data
• Categorical mediating variables
• Calculating probabilities from probit regression coefficients
• Calculating probabilities from logistic regression coefficients
• Parameterization of models with more than one categorical latent
variable

In the model estimation section, technical details of parameter

specification and model estimation are discussed. In the multiple group
analysis section, differences in model specification, differences in data
between single-group analysis and multiple group analysis, and testing
for measurement invariance are described. In the missing data section,
estimation of models when there is missing data and special features for
data missing by design are described. There is a section that describes
how categorical mediating variables are treated in model estimation.
There is a section on calculating probabilities for probit regression
coefficients. In the section on calculating probabilities for logistic
regression coefficients, a brief background with examples of converting
logistic regression coefficients to probabilities and odds is given. In the
section on parameterization with multiple categorical latent variables,
conventions related to logistic and loglinear parameterizations of these
models are described.

MODEL ESTIMATION
There are several important issues involved in model estimation beyond
specifying the model. The following general analysis considerations are
discussed below:

• Parameter default settings

• Parameter default starting values

459
CHAPTER 14

• User-specified starting values for mixture models

• Multiple solutions for mixture models
• Convergence problems
• Model identification
• Numerical integration

PARAMETER DEFAULT SETTINGS

Default settings are used to simplify the model specification. In order to
minimize the information provided by the user, certain parameters are
free, constrained to be equal, or fixed at zero as the default. These
defaults are chosen to reflect common practice and to avoid
computational problems. These defaults can be overridden. Because of
the extensive default settings, it is important to examine the analysis
results to verify that the model that is estimated is the intended model.
The output contains parameter estimates for all free parameters in the
model, including those that are free by default and those that are free
because of the model specification. Parameters that are fixed in the
input file are also listed with these results. Parameters fixed by default
are not included. In addition, the TECH1 option of the OUTPUT
command shows which parameters in the model are free to be estimated
and which are fixed.

Following are the default settings for means/intercepts/thresholds in the

model when they are included:

• Means of observed independent variables are estimated as or fixed at

the sample values when they are included in the model estimation.
• In single group analysis, intercepts and thresholds of observed
dependent variables are free.
• In multiple group analysis and multiple class analysis, intercepts and
thresholds of observed dependent variables that are used as factor
indicators for continuous latent variables are free and equal across
groups or classes. Otherwise, they are free and unequal in the other
groups or classes except for the inflation part of censored and count
variables in which case they are free and equal.
• In single group analysis, means and intercepts of continuous latent
variables are fixed at zero.
• In multiple group analysis and multiple class analysis, means and
intercepts of continuous latent variables are fixed at zero in the first

460
 Special Modeling Issues

group and last class and are free and unequal in the other groups or
classes except when a categorical latent variable is regressed on a
continuous latent variable. In this case, the means and intercepts of
continuous latent variables are fixed at zero in all classes.
• Logit means and intercepts of categorical latent variables are fixed at
zero in the last class and free and unequal in the other classes.

Following are the default settings for variances/residual variances/scale

factors:

• Variances of observed independent variables are estimated as or

fixed at the sample values when they are included in the model
estimation.
• In single group analysis and multiple group analysis, variances and
residual variances of continuous and censored observed dependent
variables and continuous latent variables are free. In multiple class
analysis, variances/residual variances of continuous and censored
observed dependent variables and continuous latent variables are
free and equal across classes.
• In single group analysis using the Delta parameterization, scale
factors of latent response variables for categorical observed
dependent variables are fixed at one. In multiple group analysis
using the Delta parameterization, scale factors of latent response
variables for categorical observed dependent variables are fixed at
one in the first group and are free and unequal in the other groups.
• In single group analysis using the Theta parameterization, variances
and residual variances of latent response variables for categorical
observed dependent variables are fixed at one. In multiple group
analysis using the Theta parameterization, variances and residual
variances of latent response variables for categorical observed
dependent variables are fixed at one in the first group and are free
and unequal in the other groups.

Following are the default settings for covariances/residual covariances:

• Covariances among observed independent variables are estimated as

or fixed at the sample values when they are included in the model
estimation.
• In single group analysis and multiple group analysis, covariances
among continuous latent independent variables are free except when
they are random effect variables defined by using ON or XWITH in

461
CHAPTER 14

conjunction with the | symbol. In these cases, the covariances

among continuous latent independent variables are fixed at zero. In
multiple class analysis, free covariances among continuous latent
independent variables are equal across classes.
• In single group analysis and multiple group analysis, covariances
among continuous latent independent variables and observed
independent variables are free except when the continuous latent
variables are random effect variables defined by using ON or
XWITH in conjunction with the | symbol or in multiple class
analysis. In these cases, the covariances among continuous latent
independent variables and observed independent variables are fixed
at zero.
• Covariances among observed variables not explicitly dependent or
independent are fixed at zero.
• Residual covariances among observed dependent variables and
among continuous latent dependent variables are fixed at zero with
the following exceptions:
• In single group analysis and multiple group analysis,
residual covariances among observed dependent variables
are free when neither variable influences any other variable,
when the variables are not factor indicators, and when the
variables are either continuous, censored (using weighted
least squares), or categorical (using weighted least squares).
In multiple class analysis, free residual covariances among
observed dependent variables are equal across classes.
• In single group analysis and multiple group analysis,
residual covariances among continuous latent dependent
variables that are not indicators of a second-order factor are
free when neither variable influences any other variable
except its own indicators, except when they are random
effect variables defined by using ON or XWITH in
conjunction with the | symbol. In these cases, the
covariances among continuous latent independent variables
are fixed at zero. In multiple class analysis, free residual
covariances among continuous latent dependent variables
are equal across classes.

Following are the default settings for regression coefficients:

• Regression coefficients are fixed at zero unless they are explicitly

mentioned in the MODEL command. In multiple group analysis,

462
 Special Modeling Issues

free regression coefficients are unequal in all groups unless they

involve the regression of an observed dependent variable that is used
as a factor indicator on a continuous latent variable. In this case,
they are free and equal across groups. In multiple class analysis,
free regression coefficients are equal across classes.

PARAMETER DEFAULT STARTING VALUES

If a parameter is not free by default, when the parameter is mentioned in
the MODEL command, it is free at the default starting value unless
another starting value is specified using the asterisk (*) followed by a
number or the parameter is fixed using the @ symbol followed by a
number. The exception to this is that variances and residual variances
for latent response variables corresponding to categorical observed
dependent variables cannot be free in the Delta parameterization. They
can be free in the Theta parameterization. In the Theta parameterization,
scale factors for latent response variables corresponding to categorical
observed dependent variables cannot be free. They can be free in the
Delta parameterization.

GENERAL DEFAULTS

Following are the default starting values:

Means/intercepts of continuous and 0 or sample mean

censored observed variables depending on the
analysis
Means/intercepts of count observed variables 0
Thresholds of categorical observed variables 0 or determined by the
sample proportions
depending on the
analysis
Variances/residual variances of .05 or 1 depending on
continuous latent variables the analysis
Variances/residual variances of .5 of the sample
continuous and censored observed variables variance
Variances/residual variances of 1
latent response variables for categorical
observed variables
Scale factors 1

463
CHAPTER 14

Loadings for indicators of continuous 1

latent variables
All other parameters 0

For situations where starting values depend on the analysis, the starting
values can be found using the TECH1 option of the OUTPUT command.

DEFAULTS FOR GROWTH MODELS

When growth models are specified using the | symbol of the MODEL
command and the outcome is continuous or censored, automatic starting
values for the growth factor means and variances are generated based on
individual regressions of the outcome variable on time. For other
outcome types, the defaults above apply.

RANDOM STARTING VALUES FOR MIXTURE MODELS

When TYPE=MIXTURE is specified, the default starting values are

automatically generated values that are used to create randomly
perturbed sets of starting values for all parameters in the model except
variances and covariances.

USER-SPECIFIED STARTING VALUES FOR

MIXTURE MODELS
Following are suggestions for obtaining starting values when random
starts are not used with TYPE=MIXTURE. User-specified starting
values can reduce computation time with STARTS=0. They can be
helpful when there is substantive knowledge of the relationship between
latent classes and the latent class indicators. For example, it may be
well-known that there is a normative class in which individuals have a
very low probability of engaging in any of the behaviors represented by
the latent class indicators. User-specified starting values may also be
used for confirmatory latent class analysis or confirmatory growth
mixture modeling.

LATENT CLASS INDICATORS

Starting values for the thresholds of the categorical latent class

indicators are given in the logit scale. For ordered categorical latent

464
 Special Modeling Issues

class indicators, the threshold starting values for each variable must be
ordered from low to high. The exception to this is when equality
constraints are placed on adjacent thresholds for a variable in which case
the same starting value is used. It is a good idea to start the classes apart
from each other.

Following is a translation of probabilities to logit threshold values that

can be used to help in selecting starting values. Note that logit threshold
values have the opposite sign from logit intercept values. The probability
is the probability of exceeding a threshold. High thresholds are
associated with low probabilities.

Very low probability Logit threshold of +3

Low probability Logit threshold of +1
High probability Logit threshold of -1
Very high probability Logit threshold of -3

GROWTH MIXTURE MODELS

In most analyses, it is sufficient to use the default starting values

together with random starts. If starting values are needed, the following
two strategies are suggested. The first strategy is to estimate the growth
model as either a one-class model or a regular growth model to obtain
means and standard deviations for the intercept and slope growth factors.
These values can be used to compute starting values. For example,
starting values for a 2 class model could be the mean plus or minus half
of a standard deviation.

The second strategy is to estimate a multi-class model with the variances

and covariances of the growth factors fixed at zero. The estimates of the
growth factor means from this analysis can be used as starting values in
an analysis where the growth factor variances and covariances are not
fixed at zero.

MULTIPLE SOLUTIONS FOR MIXTURE

MODELS
With mixture models, multiple maxima of the likelihood often exist. It
is therefore important to use more than one set of starting values to find
the global maximum. If the best (highest) loglikelihood value is not
replicated in at least two final stage solutions and preferably more, it is

465
CHAPTER 14

possible that a local solution has been reached, and the results should not
be interpreted without further investigation. Following is an example of
a set of ten final stage solutions that point to a good solution because all
of the final stage solutions have the same loglikelihood value:

Loglikelihood Seed Initial Stage Starts

-836.899 902278 21
-836.899 366706 29
-836.899 903420 5
-836.899 unperturbed 0
-836.899 27071 15
-836.899 967237 48
-836.899 462953 7
-836.899 749453 33
-836.899 637345 19
-836.899 392418 28

Following is an example of a set of final stage solutions that may point

to a possible local solution because the best loglikelihood value is not
replicated:

Loglikelihood Seed Initial Stage Starts

-835.247 902278 21
-837.132 366706 29
-840.786 903420 5
-840.786 unperturbed 0
-840.786 27071 15
-853.684 967237 48
-867.123 462953 7
-890.442 749453 33
-905.512 637345 19
-956.774 392418 28

Although the loglikelihood value of -840.786 is replicated three times, it

points to a local solution because it is not the best loglikelihood value.
The best loglikelihood value must be replicated for a trustworthy
solution.

When several final stage optimizations result in similar loglikelihood

values that are close to the highest loglikelihood value, the parameter
estimates for these solutions should be studied using the OPTSEED
option of the ANALYSIS command. If the parameter estimates are
different across the solutions, this indicates that the model is not well-
defined for the data. This may be because too many classes are being

466
 Special Modeling Issues

extracted. If the parameter values are very similar across the solutions,
the solution with the highest loglikelihood should be chosen.

Following is a set of recommendations for an increasingly more

thorough investigation of multiple solutions using the STARTS and
STITERATIONS options of the ANALYSIS command. The first
recommendation is:

STARTS = 100 20;

which increases the number of initial stage random sets of starting

values from the default of 20 to 100 and the number of final stage
optimizations from the default of 4 to 20. In this recommendation, the
default of ten STITERATIONS is used.

A second recommendation is:

STARTS = 100 20;

STITERATIONS = 20;

where STITERATIONS is increased from the default of 10 iterations to

20 iterations in addition to increasing the number of initial stage random
sets of starting values and final stage optimizations.

A third recommendation is to increase the initial stage random sets of

starting values further to 500, increase the final stage optimizations to
100, with or without increasing STITERATIONS. Following is the
specification without increasing STITERATIONS.

STARTS = 500 100;

CONVERGENCE PROBLEMS
Some combinations of models and data may cause convergence
problems. A message to this effect is found in the output. Convergence
problems are often related to variables in the model being measured on
very different scales, poor starting values, and/or a model being
estimated that is not appropriate for the data. In addition, certain models
are more likely to have convergence problems. These include mixture
models, two-level models, and models with random effects that have
small variances.

467
CHAPTER 14

GENERAL CONVERGENCE PROBLEMS

It is useful to distinguish between two types of non-convergence. The

type of non-convergence can be determined by examining the
optimization history of the analysis which is obtained by using the
TECH5 and/or TECH8 options of the OUTPUT command. In the first
type of non-convergence, the program stops before convergence because
the maximum number of iterations has been reached. In the second type
of non-convergence, the program stops before the maximum number of
iterations has been reached because of difficulties in optimizing the
fitting function.

For both types of convergence problems, the first thing to check is that
the variables are measured on similar scales. Convergence problems
may occur when the range of sample variance values greatly exceeds 1
to 10. This is particularly important with combinations of categorical
and continuous outcomes.

In the first type of problem, as long as no large negative

variances/residual variances are found in the preliminary parameter
estimates, and each iteration has not had a large number of trys,
convergence may be reached by increasing the number of iterations or
using the preliminary parameter estimates as starting values. If there are
large negative variances/residual variances, new starting values should
be tried. In the second type of problem, the starting values are not
appropriate for the model and the data. New starting values should be
tried. Starting values for variance/residual variance parameters are the
most important to change. If new starting values do not help, the model
should be modified.

A useful way to avoid convergence problems due to poor starting values

is to build up a model by estimating the model parts separately to obtain
appropriate starting values for the full model.

CONVERGENCE PROBLEMS SPECIFIC TO MODELING

WITH RANDOM EFFECTS

Random effect models can have convergence problems when the random
effect variables have small variances. Problems can arise in models in
which random effect variables are defined using the ON or AT options
of the MODEL command in conjunction with the | symbol of the

468
 Special Modeling Issues

MODEL command and in growth models for censored, categorical, and

count outcomes. If convergence problems arise, information in the error
messages identifies the problematic variable. In addition, the output can
be examined to see the size of the random effect variable variance. If it
is close to zero and the random effect variable is a random slope defined
using an ON statement in conjunction with the | symbol, a fixed effect
should be used instead by using a regular ON statement. If it is close to
zero and the random effect variable is a growth factor, the growth factor
variance and corresponding covariances should be fixed at zero.

CONVERGENCE PROBLEMS SPECIFIC TO MIXTURE

MODELS

In mixture models, convergence is determined not only by the

derivatives of the loglikelihood but also by the absolute and relative
changes in the loglikelihood and the changes in the class counts.
Information about changes in the loglikelihood and the class counts can
be found in TECH8.

Even when a mixture model does converge, it is possible to obtain a

local solution. Therefore, it is important to run the model with multiple
sets of starting values to guarantee that the best solution is obtained.
The best solution is the solution with the largest loglikelihood. As
discussed above, the STARTS option of the ANALYSIS command can
be used for automatically generating multiple sets of randomly drawn
starting values that are used to find the best solution.

MODEL IDENTIFICATION
Not all models that can be specified in the program are identified. A
non-identified model is one that does not have meaningful estimates for
all of its parameters. Standard errors cannot be computed for non-
identified models because of a singular Fisher information matrix.
When a model is not identified, an error message is printed in the output.
In most cases, the error message gives the number of the parameter that
contributes to the non-identification. The parameter to which the
number applies is found using the TECH1 option of the OUTPUT
command. Additional restrictions on the parameters of the model are
often needed to make the model identified.

469
CHAPTER 14

Model identification can be complex for mixture models. Mixture

models that are in theory identified can in certain samples and with
certain starting values be empirically non-identified. In this situation,
changing the starting values or changing the model is recommended.

For all models, model identification can be determined by examining

modification indices and derivatives. If a fixed parameter for an
outcome has a modification index or a derivative of zero, it will not be
identified if it is free. For an estimated model that is known to be
identified, the model remains identified if a parameter with a non-zero
modification index or a non-zero derivative is freed. Derivatives are
obtained by using the TECH2 option of the OUTPUT command.
Modification indices are obtained by using the MODINDICES option of
the OUTPUT command.

NUMERICAL INTEGRATION
Numerical integration is required for maximum likelihood estimation
when the posterior distribution of the latent variable does not have a
closed form expression. In the table below, the ON and BY statements
that require numerical integration are designated by a single or double
asterisk (*). A single asterisk (*) indicates that numerical integration is
always required. A double asterisk (*) indicates that numerical
integration is required when the mediating variable has missing data.
Numerical integration is also required for models with interactions
involving continuous latent variables and for certain models with
random slopes such as multilevel mixture models.

470
 Special Modeling Issues

Scale of Scale of Observed Mediating Scale of Latent

Dependent Variable Variable
Variable Continuous Censored, Continuous
Categorical,
and Count
Continuous ON ON** ON
BY
Censored, ON** ON** ON*
Categorical, BY*
and Count
Nominal ON** ON** ON*
Continuous ON ON** ON
Latent BY
Categorical ON** ON** ON*
Latent BY*
Inflation Part of ON** ON** ON*
Censored and BY*
Count

When the posterior distribution does not have a closed form, it is

necessary to integrate over the density of the latent variable multiplied
by the conditional distribution of the outcomes given the latent variable.
Numerical integration approximates this integration by using a weighted
sum over a set of integration points (quadrature nodes) representing
values of the latent variable.

Three types of numerical integration are available in Mplus with or

without adaptive numerical integration. They are rectangular (trapezoid)
numerical integration with a default of 15 integration points per
dimension, Gauss-Hermite integration with a default of 15 integration
points per dimension, and Monte Carlo integration with integration
points generated randomly with a default of 500 integration points in
total. In many cases, all three integration types are available. When
mediating variables have missing data, only the Monte Carlo integration
algorithm is available.

For some analyses it is necessary to increase the number of integration

points to obtain sufficient numerical precision. In these cases, 20-50
integration points per dimension are recommended for rectangular and
Gauss-Hermite integration and 5000 total integration points for Monte
Carlo integration. Going beyond these recommendations is not
advisable because the precision is unlikely to be improved any further,

471
CHAPTER 14

computations will become slower, and numerical instability can arise

from increased round off error.

In most analyses, the default of adaptive numerical integration is

expected to outperform non-adaptive numerical integration. In most
analyses, 15 integration points per dimension are sufficient with adaptive
numerical integration, whereas non-adaptive numerical integration may
require 30-50 integration points per dimension. There are analyses,
however, where adaptive numerical integration leads to numerical
instability. These include analyses with outliers, non-normality in the
latent variable distribution, and small cluster sizes. In such analyses, it
is recommended to turn off the adaptive numerical integration using the
ADAPTIVE option of the ANALYSIS command.

Numerical integration is computationally heavy and thereby time-

consuming because the integration must be done at each iteration, both
when computing the function value and when computing the derivative
values. The computational burden increases as a function of the number
of integration points, increases linearly as a function of the number of
observations, and increases exponentially as a function of the number of
dimensions of integration. For rectangular and Gauss-Hermite
integration, the computational burden also increases exponentially as a
function of the dimensions of integration, that is, the number of latent
variables, random slopes, or latent variable interactions for which
numerical integration is needed. Following is a list that shows the
computational burden in terms of the number of dimensions of
integration using the default number of integration points.

One dimension of integration Light

Two dimensions of integration Moderate
Three to four dimensions of integration Heavy
Five or more dimensions of integration Very heavy

Note that with several dimensions of integration it may be advantageous

to use Monte Carlo integration. Monte Carlo integration may, however,
result in loglikelihood values with low numerical precision making the
testing of nested models using likelihood ratio chi-square tests based on
loglikelihood differences imprecise. To reduce the computational
burden with several dimensions of integration, it is sometimes possible
to get sufficiently precise results by reducing the number of integration
points per dimension from the default of 15 to 10 or 7. For exploratory

472
 Special Modeling Issues

factor analysis, as few as three integration points per dimension may be

sufficient.

PRACTICAL ASPECTS OF NUMERICAL INTEGRATION

Following is a list of suggestions for using numerical integration:

• Start with a model that has a small number of latent variables,

random slopes, or latent variable interactions for which numerical
integration is required and add to this number in small increments
• Start with an analysis using the TECH8 and TECH1 options of the
OUTPUT command in conjunction with the MITERATIONS and
STARTS options of the ANALYSIS command set to 1 and 0,
respectively, to obtain information on the time required for one
iteration and to check that the model specifications are correct
• With more than 3 dimensions of integration, reduce the number of
integration points per dimension to 10 or use Monte Carlo
integration with 5000 total integration points
• If the TECH8 output shows large negative values in the column
labeled ABS CHANGE, increase the number of integration points to
improve the precision of the numerical integration and resolve
convergence problems
• Because non-identification based on a singular information matrix
may be difficult to determine when numerical integration is
involved, it is important to check for a low condition number which
may indicate non-identification, for example, a condition number
less than 1.0E-6

MULTIPLE GROUP ANALYSIS

In this section, special issues related to multiple group or multiple
population analysis are discussed. Multiple group analysis is used when
data from more than one population are being examined to investigate
measurement invariance and population heterogeneity. Measurement
invariance is investigated by testing the invariance of measurement
parameters across groups. Measurement parameters include intercepts
or thresholds of the factor indicators, factor loadings, and residual
variances of the factor indicators. Population heterogeneity is
investigated by testing the invariance of structural parameters across
groups. Structural parameters include factor means, variances, and

473
CHAPTER 14

covariances and regression coefficients. Multiple group analysis is not

available for TYPE=MIXTURE and EFA. Multiple group analysis for
TYPE=MIXTURE can be carried out using the KNOWNCLASS option
of the VARIABLE command. Following are the topics discussed in this
section:

• Requesting a multiple group analysis

• First group in multiple group analysis
• Defaults for multiple group analysis
• MODEL command in multiple group analysis
• Equalities in multiple group analysis
• Means/intercepts/thresholds in multiple group analysis
• Scale factors in multiple group analysis
• Residual variances of latent response variables in multiple group
analysis
• Data in multiple group analysis
• Testing for measurement invariance using multiple group analysis

REQUESTING A MULTIPLE GROUP ANALYSIS

The way to request a multiple group analysis depends on the type of data
that are being analyzed. When individual data stored in one data set are
analyzed, a multiple group analysis is requested by using the
GROUPING option of the VARIABLE command. When individual data
stored in different data sets are analyzed, multiple group analysis is
requested by using multiple FILE statements in the DATA command.
When summary data are analyzed, multiple group analysis is requested
by using the NGROUPS option of the DATA command.

FIRST GROUP IN MULTIPLE GROUP ANALYSIS

In some situations it is necessary to know which group the program
considers to be the first group. How the first group is defined differs
depending on the type of data being analyzed. For individual data in a
single data set, the first group is defined as the group with the lowest
value on the grouping variable. For example if the grouping variable is
gender with males having the value of 1 and females having the value of
0, then the first group is females. For individual data in separate data
sets, the first group is the group represented by the first FILE statement

474
 Special Modeling Issues

listed in the DATA command. For example, if the following FILE

statements are specified in an input setup,

FILE (male) IS male.dat;

FILE (female) IS female.dat;

the first group is males. For summary data, the first group is the group
with the label, g1. This group is the group represented by the first set of
summary data found in the summary data set.

DEFAULTS FOR MULTIPLE GROUP ANALYSIS

In multiple group analysis, some measurement parameters are held equal
across the groups as the default. This is done to reflect measurement
invariance of these parameters. Intercepts, thresholds, and factor
loadings are held equal across groups. The residual variances of the
factor indicators are not held equal across groups.

All structural parameters are free and not constrained to be equal across
groups as the default. Structural parameters include factor means,
variances, and covariances and regressions coefficients. Factor means
are fixed at zero in the first group and are free to be estimated in the
other groups as the default. This is because factor means generally
cannot be identified for all groups. The customary approach is to set the
factor means to zero in a reference group, here the first group.

For observed categorical dependent variables using the default Delta

parameterization, the scale factors of the latent response variables of the
categorical factor indicators are fixed at one in the first group and are
free to be estimated in the other groups as the default. This is because
the latent response variables are not restricted to have across-group
equalities of variances. For observed categorical dependent variables
using the Theta parameterization, the residual variances of the latent
response variables of the categorical factor indicators are fixed at one in
the first group and are free to be estimated in the other groups as the
default.

475
CHAPTER 14

MODEL COMMAND IN MULTIPLE GROUP

ANALYSIS
In multiple group analysis, two variations of the MODEL command are
used. They are MODEL and MODEL followed by a label. MODEL is
used to describe the overall analysis model. MODEL followed by a
label is used to describe differences between the overall analysis model
and the analysis model for each group. These are referred to as group-
specific models. The labels are defined using the GROUPING option of
the VARIABLE command for individual data in a single file, by the
FILE options of the DATA command for individual data in separate
files, and by the program for summary data and Monte Carlo simulation
studies. It is not necessary to describe the full model for each group in
the group-specific models. Group-specific models should contain only
differences from the model described in the overall MODEL command
and the model for that group.

Following is an example of an overall MODEL command for multiple

group analysis:

MODEL: f1 BY y1 y2 y3;
f2 BY y4 y5 y6;

In the above overall MODEL command, the two BY statements specify

that f1 is measured by y1, y2, and y3, and f2 is measured by y4, y5, and
y6. The metric of the factors is set automatically by the program by
fixing the first factor loading in each BY statement to 1. The intercepts
of the factor indicators and the other factor loadings are held equal
across the groups as the default. The residual variances are estimated for
each group and the residual covariances are fixed at zero as the default.
Factor variances and the factor covariance are estimated for each group.

Following is a group-specific MODEL command that relaxes the

equality constraints on the factor loadings in a two-group analysis:

MODEL g2: f1 BY y2 y3;

f2 BY y5 y6;

In the above group-specific MODEL command, the equality constraints

on the factor loadings of y2, y3, y5, and y6 are relaxed by including
them in a group-specific MODEL command. The first factor indicator

476
 Special Modeling Issues

of each factor should not be included because including them frees their
factor loadings which should be fixed at one to set the metric of the
factors.

Factor means are fixed at zero in the first group and are estimated in
each of the other groups. The following group-specific MODEL
command relaxes the equality constraints on the intercepts and
thresholds of the observed dependent variables:

MODEL g2: [y1 y2 y3];

[u4$1 u5$2 u6$3];

Following is a set of MODEL commands for a multiple group analysis in

which three groups are being analyzed: g1, g2, and g3:

MODEL: f1 BY y1-y5;
f2 BY y6-y10;
f1 ON f2;
MODEL g1: f1 BY y5;
MODEL g2: f2 BY y9;

In the overall MODEL command, the first BY statement specifies that f1

is measured by y1, y2, y3, y4, and y5. The second BY statement
specifies that f2 is measured by y6, y7, y8, y9, and y10. The metric of
the factors is set automatically by the program by fixing the first factor
loading in each BY statement to one. The intercepts of the factor
indicators and the other factor loadings are held equal across the groups
as the default. The residual variances for y1 through y10 are estimated
for each group and the residual covariances are fixed at zero as the
default. The variance of the factor f2 and the residual variance of the
factor f1 are estimated for each group. A regression coefficient for the
linear regression of f1 on f2 is estimated for each group.

Differences between the overall model and the group-specific models are
specified using the MODEL command followed by a label. The two
group-specific MODEL commands above specify differences between
the overall model and the group-specific models. In the above example,
the factor loading for y5 in group g1 is not constrained to be equal to the
factor loading for y5 in the other two groups and the factor loading for
y9 in group g2 is not constrained to be equal to the factor loading for y9

477
CHAPTER 14

in the other two groups. The model for g3 is identical to that of the
overall model because there is no group-specific model statement for g3.

EQUALITIES IN MULTIPLE GROUP ANALYSIS

A number or list of numbers in parentheses following a parameter or list
of parameters is used to indicate equality constraints. Constraining
parameters to be equal in a single group analysis is discussed in Chapter
17. In a single group analysis, parameters are constrained to be equal by
placing the same number or list of numbers in parentheses following the
parameters that are to be held equal. For example,

y1 ON x1 (1) ;
y2 ON x2 (1) ;
y3 ON x3 (2) ;
y4 ON x4 (2) ;
y5 ON x5 (2) ;

constrains the regression coefficients of the first two equations to be

equal and the regression coefficients of the last three equations to be
equal.

In multiple group analysis, the interpretation of equality constraints

depends on whether they are part of the overall MODEL command or a
group-specific MODEL command. Equality constraints specified in the
overall MODEL command apply to all groups. Equality constraints
specified in a group-specific MODEL command apply to only that
group.

Following is an example of how to specify across group equality

constraints in the overall MODEL command:

MODEL: f1 BY y1-y5;
y1 (1)
y2 (2)
y3 (3)
y4 (4)
y5 (5);

By placing a different number in parentheses after each residual

variance, each residual variance is held equal across all groups but not

478
 Special Modeling Issues

equal to each other. Note that only one equality constraint can be
specified per line.

Following is another example of how to specify across group equality

constraints in the overall MODEL command:

MODEL: f1 BY y1-y5;
y1-y5 (1);

By placing a one in parentheses after the list of residual variances, y1

through y5, the values of those parameters are held equal to each other
and across groups. If the five residual variances are free to be estimated
across the three groups, there are fifteen parameters. With the equality
constraint, one parameter is estimated.

Following is an example of how to specify an equality constraint in a

group-specific MODEL command:

MODEL g2: y1-y5 (2);

In the group-specific MODEL command for g2, the residual variances of

y1 through y5 are held equal for g2 but are not held equal to the residual
variances of any other group because (2) is not specified in the overall
MODEL command or in any other group-specific MODEL command.
One residual variance is estimated for g2.

Following is an example of how to relax an equality constraint in a

group-specific MODEL command:

MODEL g3: y1-y5;

In this example, by mentioning the residual variances in a group-specific

MODEL command, they are no longer held equal to the residual
variances in groups 1 and 3. Five residual variances are estimated for
g3.

The overall and group-specific MODEL commands discussed above are

shown and interpreted together below:

MODEL: f1 BY y1-y5;
y1-y5 (1);

479
CHAPTER 14

MODEL g2: y1-y5 (2);

MODEL g3: y1-y5;

The overall MODEL command specifies the overall model for the three
groups as described above. Because there is no group-specific MODEL
command for g1, g1 uses the same model as that described in the overall
MODEL command. The group-specific MODEL commands describe
the differences between the overall model and the group-specific
models. The group g2 uses the overall model with the exception that the
one residual variance that is estimated is not constrained to be equal to
the other two groups. The group g3 uses the overall model with the
exception that five residual variances not constrained to be equal to the
other groups are estimated.

MEANS/INTERCEPTS/THRESHOLDS IN
MULTIPLE GROUP ANALYSIS
In multiple group analysis, the intercepts and thresholds of observed
dependent variables that are factor indicators are constrained to be equal
across groups as the default. The means and intercepts of continuous
latent variables are fixed at zero in the first group and are free to be
estimated in the other groups as the default. Means, intercepts, and
thresholds are referred to by the use of square brackets.

Following is an example how to refer to means and intercepts in a

multiple group model.

MODEL: f1 BY y1-y5;
f2 BY y6-y10;
f1 ON f2;
MODEL g1: [f1 f2];
MODEL g2: [f1@0 f2@0];

In the above example, the intercepts and the factor loadings for the
factor indicators y1-y5 are held equal across the three groups as the
default. In the group-specific MODEL command for g1, the mean of f2
and the intercept of f1 are specified to be free. In the group-specific
MODEL command for g2, the mean of f2 and the intercept of f1 are
fixed at zero.

480
 Special Modeling Issues

The following group-specific MODEL command relaxes the equality

constraints on the intercepts of the observed dependent variables:

MODEL g2: [y1-y10];

SCALE FACTORS IN MULTIPLE GROUP

ANALYSIS
Scale factors can be used in multiple group analysis. They are
recommended when observed dependent variables are categorical and a
weighted least squares estimator is used. They capture across group
differences in the variances of the latent response variables for the
observed categorical dependent variables. Scale factors are part of the
model as the default using a weighted least squares estimator when one
or more observed dependent variables are categorical. In this situation,
the first group has scale factors fixed at one. In the other groups, scale
factors are free to be estimated with starting values of one. Scale factors
are referred to using curly brackets. Following is an example of how to
refer to scale factors in a model with multiple groups where u1, u2, u3,
u4, and u5 are observed categorical dependent variables.

MODEL: f BY u1-u5;
MODEL g2: {u1-u5*.5};

In the above example, the scale factors of the latent response variables of
the observed categorical dependent variables in g1 are fixed at one as the
default. Starting values are given for the free scale factors in g2.

RESIDUAL VARIANCES OF LATENT RESPONSE

VARIABLES IN MULTIPLE GROUP ANALYSIS
With the Theta parameterization for observed categorical dependent
variables using a weighted least squares estimator, residual variances of
the latent response variables for the observed categorical dependent
variables are part of the model as the default. In this situation, the first
group has residual variances fixed at one for all observed categorical
dependent variables. In the other groups, residual variances are free to
be estimated with starting values of one. Residual variances of the latent
response variables are referred to using the name of the corresponding
observed variable. Following is an example of how to refer to residual

481
CHAPTER 14

variances in a model with multiple groups where u1, u2, u3, u4, and u5
are observed categorical dependent variables.

MODEL: f BY u1-u5;
MODEL g2: u1-u5*2;

In the above example, the residual variances of the latent response

variables of the observed categorical dependent variables in g1 are fixed
at one as the default. Starting values are given for the free residual
variances in g2.

DATA IN MULTIPLE GROUP ANALYSIS

One difference between single group analysis and multiple group
analysis is related to the data to be analyzed. For individual data, the
data for all groups can be stored in one data set or in different data sets.
If the data are stored in one data set, the data set must include a variable
that identifies the group to which each observation belongs. For
summary data, all data must be stored in the same data set.

INDIVIDUAL DATA, ONE DATA SET

If individual data for several groups are stored in one data set, the data
set must include a variable that identifies the group to which each
observation belongs. The name of this variable is specified using the
GROUPING option of the VARIABLE command. Only one grouping
variable can be specified. If the groups to be analyzed are a combination
of more than one variable, for example, gender and ethnicity, a single
grouping variable can be created using the DEFINE command. An
example of how to specify the GROUPING option is:

GROUPING IS gender (1 = male 2 = female);

The information in parentheses after the grouping variable name assigns

labels to the values of the grouping variable found in the data set. In the
example above, observations with the variable gender equal to 1 are
assigned the label male, and observations with the variable gender equal
to 2 are assigned the label female. These labels are used in group-
specific MODEL commands to specify differences between the overall
model and the group-specific models. If an observation has a value for

482
 Special Modeling Issues

the grouping variable that is not specified using the GROUPING option,
it is not included in the analysis.

INDIVIDUAL DATA, DIFFERENT DATA SETS

For individual data stored in different data sets, the specification of the
FILE option of the DATA command has two differences for multiple
group analysis. First, a FILE statement is required for each data set.
Second, the FILE option allows a label to be specified that can be used
in the group-specific MODEL commands. In the situation where the
data for males are stored in a file named male.dat, and the data for
females are stored in a file named female.dat, the FILE option is
specified as follows:

FILE (male) = male.dat;

FILE (female) = female.dat;

The labels male and female can be used in the group-specific MODEL
commands to specify differences between the group-specific models for
males and females and the overall model.

When individual data are stored in different data sets, all of the data sets
must contain the same number of variables. These variables must be
assigned the same names and be read using the same format.

SUMMARY DATA, ONE DATA SET

Summary data must be stored in one data set with the data for the first
group followed by the data for the second group, etc.. For example, in
an analysis of means and a covariance matrix for two groups with four
observed variables, the data would appear as follows:

0000
2
12
112
1112
1111

483
CHAPTER 14

3
23
223
2223

where the means for group 1 come first, followed by the covariances for
group 1, followed by the means for group 2, followed by the covariances
for group 2.

The NOBSERVATIONS and NGROUPS options have special formats

for multiple group analysis when summary data are analyzed. The
NOBSERVATIONS option requires an entry for each group in the order
that the data appear in the data set. For example, if the summary data for
males appear first in a data set followed by the summary data for
females, the NOBSERVATIONS statement,

NOBSERVATIONS = 180 220;

indicates that the summary data for males come from 180 observations
and the summary data for females come from 220 observations.

In addition, for summary data, it is necessary to specify the number of

groups in the analysis using the NGROUPS option of the DATA
command. The format of this option follows:

NGROUPS = 2;

which indicates that there are two groups in the analysis. For summary
data, the program automatically assigns the label g1 to the first group, g2
to the second group, etc. In this example, males would have the label g1
and females would have the label g2.

TESTING FOR MEASUREMENT INVARIANCE

USING MULTIPLE GROUP ANALYSIS
Multiple group analysis can be used to test measurement invariance of
factors using chi-square difference tests or loglikelihood difference tests
for a set of nested models. For continuous outcomes, the measurement
parameters are the intercepts, factor loadings, and residual variances of
the factor indicators. In many disciplines, invariance of intercepts or
thresholds and factor loadings are considered sufficient for measurement

484
 Special Modeling Issues

invariance. Some disciplines also require invariance of residual

variances. For categorical outcomes, the measurement parameters are
thresholds and factor loadings. For the Delta parameterization of
weighted least squares estimation, scale factors can also be considered.
For the Theta parameterization of weighted least squares estimation,
residual variances can also be considered.

MODELS FOR CONTINUOUS OUTCOMES

Following is a set of models that can be considered for measurement

invariance of continuous outcomes. They are listed from least restrictive
to most restrictive.

1. Intercepts, factor loadings, and residual variances free across groups;

factor means fixed at zero in all groups
2. Factor loadings constrained to be equal across groups; intercepts and
residual variances free; factor means fixed at zero in all groups
3. Intercepts and factors loadings constrained to be equal across
groups; residual variances free; factor means zero in one group and
free in the others (the Mplus default)
4. Intercepts, factor loadings, and residual variances constrained to be
equal across groups; factor means fixed at zero in one group and free
in the others

MODELS FOR CATEGORICAL OUTCOMES

Following is a set of models that can be considered for measurement

invariance of categorical outcomes. They are listed from least restrictive
to most restrictive. For categorical outcomes, measurement invariance
models constrain thresholds and factor loadings in tandem because the
item probability curve is influenced by both parameters.

WEIGHTED LEAST SQUARES ESTIMATOR USING THE

DELTA PARAMETERIZATION

1. Thresholds and factor loadings free across groups; scale factors

fixed at one in all groups; factor means fixed at zero in all groups
2. Thresholds and factor loadings constrained to be equal across
groups; scale factors fixed at one in one group and free in the others;
factor means fixed at zero in one group and free in the others (the
Mplus default)

485
CHAPTER 14

WEIGHTED LEAST SQUARES ESTIMATOR USING THE

THETA PARAMETERIZATION

1. Thresholds and factor loadings free across groups; residual variances

fixed at one in all groups; factor means fixed at zero in all groups
2. Thresholds and factor loadings constrained to be equal across
groups; residual variances fixed at one in one group and free in the
others; factor means fixed at zero in one group and free in the others
(the Mplus default)

MAXIMUM LIKELIHOOD ESTIMATOR WITH

CATEGORICAL OUTCOMES

1. Thresholds and factor loadings free across groups; factor means

fixed at zero in all groups
2. Thresholds and factor loadings constrained to be equal across
groups; factor means fixed at zero in one group and free in the others
(the Mplus default)

PARTIAL MEASUREMENT INVARIANCE

When full measurement invariance does not hold, partial measurement

invariance can be considered. This involves relaxing some equality
constraints on the measurement parameters. For continuous outcomes,
equality constraints can be relaxed for the intercepts, factor loadings,
and residual variances. This is shown in Example 5.15. For categorical
outcomes, equality constraints for thresholds and factor loadings for a
variable should be relaxed in tandem. In addition, for the Delta
parameterization, the scale factor must be fixed at one for that variable.
This is shown in Example 5.16. For the Theta parameterization, the
residual variance must be fixed at one for that variable. This is shown in
Example 5.17.

MODEL DIFFERENCE TESTING

In chi-square difference testing of measurement invariance, the chi-

square value and degrees of freedom of the less restrictive model are
subtracted from the chi-square value and degrees of freedom of the
nested, more restrictive model. The chi-square difference value is
compared to the chi-square value in a chi-square table using the
difference in degrees of freedom between the more restrictive and less

486
 Special Modeling Issues

restrictive models. If the chi-square difference value is significant, it

indicates that constraining the parameters of the nested model
significantly worsens the fit of the model. This indicates measurement
non-invariance. If the chi-square difference value is not significant, this
indicates that constraining the parameters of the nested model did not
significantly worsen the fit of the model. This indicates measurement
invariance of the parameters constrained to be equal in the nested model.

For models where chi-square is not available, difference testing can be

done using -2 times the difference of the loglikelihoods. For the MLR,
MLM, and WLSM estimators, difference testing must be done using the
scaling correction factor printed in the output. A description of how to
do this is posted on the website. For WLSMV and MLMV, difference
testing must be done using the DIFFTEST option of the SAVEDATA
and ANALYSIS commands.

MISSING DATA ANALYSIS

Mplus has several options for the estimation of models with missing
data. Mplus provides maximum likelihood estimation under MCAR
(missing completely at random) and MAR (missing at random; Little &
Rubin, 2002) for continuous, censored, binary, ordered categorical
(ordinal), unordered categorical (nominal), counts, or combinations of
these variable types. MAR means that missingness can be a function of
observed covariates and observed outcomes. For censored and
categorical outcomes using weighted least squares estimation,
missingness is allowed to be a function of the observed covariates but
not the observed outcomes. When there are no covariates in the model,
this is analogous to pairwise present analysis. Non-ignorable missing
data modeling is possible using maximum likelihood estimation where
categorical outcomes are indicators of missingness and where
missingness can be predicted by continuous and categorical latent
variables (Muthén, Jo, & Brown, 2003; Muthén et al., 2011). Robust
standard errors and chi-square are available for all outcomes using the
MLR estimator. For non-normal continuous outcomes, this gives the T2*
chi-square test statistic of Yuan and Bentler (2000).

Mplus provides multiple imputation of missing data using Bayesian

analysis (Rubin, 1987; Schafer, 1997). Both unrestricted H1 and
restricted H0 models can be used for imputation.

487
CHAPTER 14

Multiple data sets generated using multiple imputation (Rubin, 1987;

Schafer, 1997) can be analyzed using a special feature of Mplus.
Parameter estimates are averaged over the set of analyses, and standard
errors are computed using the average of the standard errors over the set
of analyses and the between analysis parameter estimate variation.

In all models, missingness is not allowed for the observed covariates

because they are not part of the model. The model is estimated
conditional on the covariates and no distributional assumptions are made
about the covariates. Covariate missingness can be modeled if the
covariates are brought into the model and distributional assumptions
such as normality are made about them. With missing data, the standard
errors for the parameter estimates are computed using the observed
information matrix (Kenward & Molenberghs, 1998). Bootstrap
standard errors and confidence intervals are also available with missing
data.

With missing data, it is useful to do a descriptive analysis to study the

percentage of missing data as a first step. This can be accomplished by
specifying TYPE=BASIC in the ANALYSIS command. The output for
this analysis produces the number of missing data patterns and the
proportion of non-missing data, or coverage, for variables and pairs of
variables. A default of .10 is used as the minimum coverage proportion
for a model to be estimated. This minimum value can be changed by
using the COVERAGE option of the ANALYSIS command.

DATA MISSING BY DESIGN

Data missing by design occurs when the study determines which subjects
will be observed on which measures. One example is when different
forms of a measurement instrument are administered to randomly
selected subgroups of individuals. A second example is when it is
expensive to collect data on all variables for all individuals and only a
subset of variables is measured for a random subgroup of individuals. A
third example is multiple cohort analysis where individuals who are
measured repeatedly over time represent different birth cohorts. These
types of studies can use the missing data method where all individuals
are used in the analysis, including those who have missing values on
some of the analysis variables by design. This type of analysis is
obtained by identifying the values in the data set that are considered to
be missing value flags using the MISSING option of the VARIABLE

488
 Special Modeling Issues

command and identifying the variables for which individuals should

have a value using the PATTERN option of the VARIABLE command.

MULTIPLE COHORT DESIGN

Longitudinal research studies often collect data on several different

groups of individuals defined by their birth year or cohort. This allows
the study of development over a wider age range than the length of the
study and is referred to as an accelerated or sequential cohort design.
The interest in these studies is the development of an outcome over age
not measurement occasion. When dependent variables are measured
using a continuous scale, options are available for rearranging such a
data set so that age rather than time of measurement is the time variable.
This is available only for TYPE=GENERAL without
ALGORITHM=INTEGRATION.

The DATA COHORT command is used to rearrange longitudinal data

from a format where time points represent measurement occasions to a
format where time points represent age or another time-related variable.
It is necessary to know the cohort (birth year) of each individual and the
year in which each measurement was taken. The difference between
measurement year and cohort year is the age of the individual at the time
of measurement. Age is the variable that is used to determine the pattern
of missing values for each cohort. If an individual does not have
information for a particular age, that value is missing for that individual.
The transformed data set is analyzed using maximum likelihood
estimation for missing data.

REARRANGEMENT OF THE MULTIPLE COHORT DATA

What is of interest in multiple cohort analysis is not how a variable

changes from survey year to survey year, but how it changes with age.
What is needed to answer this question is a data set where age is the time
variable. Following is an example of how a data set is transformed using
the DATA COHORT command. In the following data set, the variable
heavy drinking (HD) is measured in 1982, 1983, 1987, and 1989.
Missing data are indicated with an asterisk (*). The respondents include
individuals born in 1963, 1964, and 1965. Although the respondents
from any one cohort are measured on only four occasions, the cohorts
taken together cover the ages 17 through 26.

489
CHAPTER 14

Observation Cohort HD82 HD83 HD87 HD89

1 63 3 4 5 6
2 63 * 6 7 8
3 63 9 8 * 3
4 63 5 7 6 3
5 63 5 8 7 9
6 64 3 6 5 9
7 64 3 8 * 5
8 64 4 9 8 6
9 64 4 * 6 7
10 64 3 9 8 5
11 65 * 4 5 6
12 65 6 5 5 5
13 65 5 5 5 5
14 65 4 5 6 7
15 65 4 5 5 4

The information in the table above represents how the data look before
they are transformed. As a first step, each observation that does not
have complete data for 1982, 1983, 1987, and 1989 is deleted from the
data set. Following is the data after this step.

Observation Cohort HD82 HD83 HD87 HD89

1 63 3 4 5 6
4 63 5 7 6 3
5 63 5 8 7 9
6 64 3 6 5 9
8 64 4 9 8 6
10 64 3 9 8 5
12 65 6 5 5 5
13 65 5 5 5 5
14 65 4 5 6 7
15 65 4 5 5 4

The second step is to rearrange the data so that age is the time
dimension. This results in the following data set where asterisks (*)
represent values that are missing by design.

490
 Special Modeling Issues

Obs Coh HD17 HD18 HD19 HD20 HD22 HD23 HD24 HD25 HD26
1 63 * * 3 4 * * 5 * 6
4 63 * * 5 7 * * 6 * 3
5 63 * * 5 8 * * 7 * 9
6 64 * 3 6 * * 5 * 9 *
8 64 * 4 9 * * 8 * 6 *
10 64 * 3 9 * * 8 * 5 *
12 65 6 5 * * 5 * 5 * *
13 65 5 5 * * 5 * 5 * *
14 65 4 5 * * 6 * 7 * *
15 65 4 5 * * 5 * 4 * *

The model is specified in the MODEL command using the new variables
hd17 through hd26 instead of the original variables hd82, hd83, hd87,
and hd89. Note that there is no hd21 because no combination of survey
year and birth cohort represents this age. The data are analyzed using
the missing by design feature.

CATEGORICAL MEDIATING VARIABLES

The treatment of categorical mediating variables in model estimation
differs depending on the estimator being used. Consider the following
model:

x -> u -> y

where u is a categorical variable. The issue is how is u treated when it is

a dependent variable predicted by x and how is it treated when it is an
independent variable predicting y. With weighted least squares
estimation, a probit regression coefficient is estimated in the regression
of u on x. In the regression of y on u, the continuous latent response
variable u* is used as the covariate. With maximum likelihood
estimation, either a logistic or probit regression coefficient is estimated
in the regression of u on x. In the regression of y on u, the observed
variable u is used as the covariate. With Bayesian estimation, a probit
regression coefficient is estimated in the regression of u on x. In the
regression of y on u, either the observed variable u or the latent response
variable u* can be used as the covariate using the MEDIATOR option of
the ANALYSIS command.

491
CHAPTER 14

CALCULATING PROBABILITIES FROM PROBIT

REGRESSION COEFFICIENTS
Following is a description of how to translate probit regression
coefficients to probability values. For a treatment of probit regression
for binary and ordered categorical (ordinal) variables, see Agresti (1996,
2002).

For a binary dependent variable, the probit regression model expresses

the probability of u given x as,

P (u = 1 | x) = F (a + b*x)
= F (-t + b*x),

where F is the standard normal distribution function, a is the probit

regression intercept, b is the probit regression slope, t is the probit
threshold where t = -a, and P (u = 0 | x) = 1 – P (u = 1 | x).

Following is an output excerpt that shows the results from the probit
regression of a binary variable u on the covariate age:

Estimates S.E. Est./S.E.

u ON
age 0.055 0.001 43.075

Thresholds
u$1 3.581 0.062 57.866

Using the formula shown above, the probability of u = 1 for age = 62 is

computed as follows:

P (u = 1 | x = 62) = F (-3.581 + 0.055*62)

= F (-0.171).

Using the z table, the value -0.171 corresponds to a probability of

approximately 0.43. This means that the probability of u = 1 at age 62 is
0.43.

For an ordered categorical (ordinal) dependent variable with three

categories, the probit regression model expresses the probability of u

492
 Special Modeling Issues

given x using the two thresholds t1 and t2 and the single probit regression
coefficient b,

P (u = 0 | x) = F (t1 - b*x),
P (u = 1 | x) = F (t2 - b*x) - F (t1 - b*x),
P (u = 2 | x) = F (- t2 + b*x).

CALCULATING PROBABILITIES FROM LOGISTIC

REGRESSION COEFFICIENTS
Following is a description of how to translate logistic regression
coefficients to probability values. Also described is how to interpret the
coefficient estimates in terms of log odds, odds, and odds ratios. For a
treatment of logistic regression for binary, ordered categorical (ordinal),
and unordered categorical (nominal) variables, see Agresti (1996, 2002)
and Hosmer and Lemeshow (2000).

An odds is a ratio of two probabilities. A log odds is therefore the log of

a ratio of two probabilities. The exponentiation of a log odds is an odds.
A logistic regression coefficient is a log odds which is also referred to as
a logit.

For a binary dependent variable u, the logistic regression model

expresses the probability of u given x as,

(1) P (u = 1 | x) = exp (a + b*x) / (1 + exp (a + b*x) )

= 1 / (1 + exp (-a – b*x)),

where P (u = 0 | x) = 1 – P (u = 1 | x). The probability expression in (1)

results in the linear logistic regression expression also referred to as a
log odds or logit,

log [P (u = 1 | x) / P (u = 0 | x)] = log [exp (a + b*x)] = a + b*x,

where b is the logistic regression coefficient which is interpreted as the

increase in the log odds of u = 1 versus u = 0 for a unit increase in x.
For example, consider the x values of x0 and x0 + 1. The corresponding
log odds are,

log odds (x0) = a + b*x0,

493
CHAPTER 14

log odds (x0 +1) = a + b*(x0 + 1) = a + b*x0 + b,

such that the increase from x0 to x0 + 1 in the log odds is b. The

corresponding odds increase is exp (b). For example, consider the
continuous covariate age with a logistic regression coefficient of .75 for
a dependent variable of being depressed (u = 1) or not being depressed
(u = 0). This means that for an increase of one year of age the log odds
of being depressed versus not being depressed increases by .75. The
corresponding odds increase is 2.12.

For a binary covariate x scored as 0 and 1, the log odds for u = 1 versus
u = 0 are,

log odds (x = 0) = a + b*0,

log odds (x = 1) = a + b*1,

such that the increase in the log odds is b as above. Given the
mathematical rule that log y – log z is equal to log (y / z), the difference
in the two log odds,

b = log odds (x = 1) – log odds (x = 0)

= log [odds (x = 1) / odds (x = 0)],

is the log odds ratio for u = 1 versus u = 0 when comparing x = 1 to x =

0. For example, consider the binary covariate gender (1 = female, 0 =
male) with a logistic regression coefficient of 1.0 for a dependent
variable of being depressed (u = 1) or not being depressed (u = 0). This
means that the log odds for females is 1.0 higher than the log odds for
males for being depressed versus not being depressed. The
corresponding odds ratio is 2.72, that is the odds for being depressed
versus not being depressed is 2.72 times larger for females than for
males.

In the case of a binary dependent variable, it is customary to let the first

category u = 0 be the reference category as is done in (1). When a
dependent variable has more than two categories, it is customary to let
the last category be the reference category as is done below. For an
unordered categorical (nominal) variable with more than two categories
R, the probability expression in (1) generalizes to the following
multinomial logistic regression,

494
 Special Modeling Issues

(2) P (u = r | x) = exp (ar + br*x) / (exp (a1 + b1*x) + …

+ exp (aR + bR*x)),

where exp (aR + bR*x) = exp (0 + 0*x) = 1 and the log odds for
comparing category r to category R is

(3) log [P (u = r | x)/P (u = R | x)] = ar + br*x.

With an ordered categorical (ordinal) variable, the logistic regression

slopes br are the same across the categories of u.

Following is an example of an unordered categorical (nominal)

dependent variable that is the categorical latent variable in the model.
The categorical latent variable has four classes and there are three
covariates. The output excerpt shows the results from the multinomial
logistic regression of the categorical latent variable c on the covariates
age94, male, and black:

Estimates S.E. Est./S.E.

C#1 ON
AGE94 -.285 .028 -10.045
MALE 2.578 .151 17.086
BLACK .158 .139 1.141
C#2 ON
AGE94 .069 .022 3.182
MALE .187 .110 1.702
BLACK -.606 .139 -4.357
C#3 ON
AGE94 -.317 .028 -11.311
MALE 1.459 .101 14.431
BLACK .999 .117 8.513
Intercepts
C#1 -1.822 .174 -10.485
C#2 -.748 .103 -7.258
C#3 -.324 .125 -2.600

Using (3), the log odds expression for a particular class compared to the
last class is,

log odds = a + b1*age94 + b2*male + b3*black.

In the first example, the values of the three covariates are all zero so that
only the intercepts contribute to the log odds. Probabilities are
computed using (2). In the first step, the estimated intercept log odds

495
CHAPTER 14

values are exponentiated and summed. In the second step, each

exponentiated value is divided by the sum to compute the probability for
each class of c.
exp probability = exp/sum

log odds (c = 1) = -1.822 0.162 0.069

log odds (c = 2) = -0.748 0.473 0.201
log odds (c = 3) = -0.324 0.723 0.307
log odds (c = 4) = 0 1.0 0.424
_______ ________
sum 2.358 1.001

In the second example, the values of the three covariates are all one so
that both the intercepts and the slopes contribute to the logs odds. In the
first step, the log odds values for each class are computed. In the second
step, the log odds values are exponentiated and summed. In the last step,
the exponentiated value is divided by the sum to compute the probability
for each class of c.

log odds (c = 1) = -1.822 + (-0.285*1) + (2.578*1) + (0.158*1)

= 0.629
log odds (c = 2) = -0.748 + 0.069*1 + 0.187*1 + (-0.606*1)
= -1.098
log odds (c = 3) = -0.324 + (-0.317*1) + 1.459*1 + 0.999*1
= 1.817

exp probability = exp/sum

log odds (c = 1) = 0.629 1.876 0.200

log odds (c = 2) = -1.098 0.334 0.036
log odds (c = 3) = 1.817 6.153 0.657
log odds (c = 4) = 0 1.0 0.107
_______ ________
sum 9.363 1.000

The interpretation of these probabilities is that individuals who have a

value of 1 on each of the covariates have a probability of .200 of being
in class 1, .036 of being in class 2, .657 of being in class 3, and .107 of
being in class 4.

496
 Special Modeling Issues

In the output shown above, the variable male has the value of 1 for males
and 0 for females and the variable black has the value of 1 for blacks and
0 for non-blacks. The variable age94 has the value of 0 for age 16, 1 for
age 17, up to 7 for age 23. An interpretation of the logistic regression
coefficient for class 1 is that comparing class 1 to class 4, the log odds
decreases by -.285 for a unit increase in age, is 2.578 higher for males
than for females, and is .158 higher for blacks than for non-blacks. This
implies that the odds ratio for being in class 1 versus class 4 when
comparing males to females is 13.17 (exp 2.578), holding the other two
covariates constant.

Following is a plot of the estimated probabilities in each of the four

classes where age is plotted on the x-axis and the other covariates take
on the value of one. This plot was created and exported as an EMF file
using the PLOT command in conjunction with the Mplus post-
processing graphics module.

Class 1, 12.7%

Class 2, 20.5%

Class 3, 30.7%
0.6
Class 4, 36.2%

0.4
Probability

0.2

0
0

age94

PARAMETERIZATION OF MODELS WITH MORE THAN

ONE CATEGORICAL LATENT VARIABLE
The parameterization of models with more than one categorical latent
variables is described in this section. There are three parameterizations

497
CHAPTER 14

available for these models. The first parameterization is based on a

series of logistic regressions for non-recursive models. The second
parameterization is that of loglinear modeling of frequency tables. The
third parameterization uses probabilities rather than logits.

LOGIT PARAMETERIZATION
Following is a description of the logistic regression parameterization,
specified using PARAMETERIZATION=LOGIT, for the following
MODEL command for two categorical latent variables with three classes
each:

MODEL:
%OVERALL%
c2#1 ON c1#1;
c2#1 ON c1#2;
c2#2 ON c1#1;
c2#2 ON c1#2;

The set of ON statements describes the logistic regression coefficients in

the conditional distribution of c2 given c1. With three classes for both
c2 and c1, there are a total of six parameters in this conditional
distribution. Two of the parameters are intercepts for c2 and four are the
logistic regression coefficients specified in the MODEL command.

For the c2 classes r = 1, 2, 3, the transition probabilities going from the

classes of c1 to the classes of c2 are given by the following unordered
multinomial logistic regression expressions:

P (c2 = r | c1 = 1) = exp (ar + br1) / sum1,

P (c2 = r | c1 = 2) = exp (ar + br2) / sum2,
P (c2 = r | c1 = 3) = exp (ar + br3) / sum3,

where a3 = 0, b31 = 0, b32 = 0, and b33 = 0 because the last class is the
reference class, and sumj represents the sum of the exponentiations
across the classes of c2 for c1 = j (j = 1, 2, 3). The corresponding log
odds when comparing a c2 class to the last c2 class are summarized in
the table below.

498
 Special Modeling Issues

c2
1 2 3
c1 1 a1 + b11 a2 + b21 0
2 a1 + b12 a2 + b22 0
3 a1 a2 0

The parameters in the table are referred to in the MODEL command

using the following statements:

a1 [c2#1];
a2 [c2#2];
b11 c2#1 ON c1#1;
b12 c2#1 ON c1#2;
b21 c2#2 ON c1#1;
b22 c2#2 ON c1#2;

The TECH15 option is used in conjunction with TYPE=MIXTURE and

PARAMETERIZATION=LOGIT to request marginal and conditional
probabilities, including latent transition probabilities, for the categorical
latent variables in a model. Conditional probabilities, including latent
transition probabilities, for different values of a set of covariates can be
computed using the LTA Calculator. It is accessed by choosing LTA
calculator from the Mplus menu of the Mplus Editor. The logit
parameterization with covariates is described in Muthén and Asparouhov
(2011).

LOGLINEAR PARAMETERIZATION
Following is a description of the loglinear parameterization for the
following MODEL command for two categorical latent variables with
three classes each:

MODEL:
%OVERALL%
c2#1 WITH c1#1;
c2#1 WITH c1#2;
c2#2 WITH c1#1;
c2#2 WITH c1#2;

499
CHAPTER 14

The set of WITH statements describes the relationship between c1 and

c2. The parameters in the table below are referred to in the MODEL
command using the following statements:

a11 [c1#1];
a12 [c1#2];
a21 [c2#1];
a22 [c2#2];
w11 c2#1 WITH c1#1;
w12 c2#1 WITH c1#2;
w21 c2#2 WITH c1#1;
w22 c2#2 WITH c1#2;

The joint probabilities for the classes of c1 and c2 are computed using
the multinomial logistic regression formula (2) in the previous section,
summing over the nine cells shown in the table below.

c2
1 2 3
c1 1 a11 + a21 + a11 + a22 + a11
w11 w21
2 a12 + a21 + a12 + a22 + a12
w12 w22
3 a21 a22 0

PROBABILITY PARAMETERIZATION
Following is a description of the probability parameterization for the
following MODEL command for two categorical latent variables with
three classes each:

MODEL:
%OVERALL%
c2#1 ON c1#1;
c2#1 ON c1#2;
c2#1 ON c1#3;
c2#2 ON c1#1;
c2#2 ON c1#2;
c2#2 ON c1#3;

The set of ON statements describes the probability parameters in the

conditional distribution of c2 given c1. With three classes for both c2

500
 Special Modeling Issues

and c1, there are a total of six probabilities in this conditional

distribution. The last category of c2 is not referred to because the
probabilities sum to one for each row.

c2
1 2 3
c1 1 p11 p12 0
2 p21 p22 0
3 p31 p32 0

The marginal probabilities for c1 are referred to as:

[c1#1];
[c1#2];

The last category of c1 is not referred to because the probabilities sum to

one.

501
CHAPTER 14

502
 TITLE, DATA, VARIABLE, And DEFINE Commands

CHAPTER 15
TITLE, DATA, VARIABLE, AND
DEFINE COMMANDS

In this chapter, the TITLE, DATA, VARIABLE, and DEFINE

commands are discussed. The TITLE command is used to provide a title
for the analysis. The DATA command is used to provide information
about the data set to be analyzed. The VARIABLE command is used to
provide information about the variables in the data set to be analyzed.
The DEFINE command is used to transform existing variables and create
new variables.

THE TITLE COMMAND

The TITLE command is used to provide a title for the analysis.
Following is the general format for the TITLE command:

TITLE: title for the analysis

The TITLE command is not a required command. Note that commands

can be shortened to four or more letters.

The TITLE command can contain any letters and symbols except the
words used as Mplus commands when they are followed by a colon.
These words are: title, data, variable, define, analysis, model, output,
savedata, montecarlo, and plot. These words can be included in the title
if they are not followed by a colon. Colons can be used in the title as
long as they do not follow words that are used as Mplus commands.
Following is an example of how to specify a title:

TITLE: confirmatory factor analysis of diagnostic criteria

The title is printed in the output just before the Summary of Analysis.

503
CHAPTER 15

THE DATA COMMAND

The DATA command is used to provide information about the data set
to be analyzed. The DATA command has options for specifying the
location of the data set to be analyzed, describing the format and type of
data in the data set, specifying the number of observations and number
of groups in the data set if the data are in summary form such as a
correlation or covariance matrix, requesting listwise deletion of
observations with missing data, and specifying whether the data should
be checked for variances of zero.

Data must be numeric except for certain missing value flags and must
reside in an external ASCII file. There is no limit on the number of
variables or observations. The maximum record length is 10,000.
Special features of the DATA command for multiple group analysis are
discussed in Chapter 14. Monte Carlo data generation is discussed in
Chapters 12 and 19. The estimator chosen for an analysis determines the
type of data required for the analysis. Some estimators require a data set
with information for each observation. Some estimators require only
summary information.

There are six DATA transformation commands. They are used to

rearrange data from a wide to long format, to rearrange data from a long
to wide format, to create a binary and a continuous variable from a
semicontinuous variable, to create a set of binary variables that are
indicators of missing data, to create variables for discrete-time survival
modeling, and to rearrange longitudinal data from a format where time
points represent measurement occasions to a format where time points
represent age or another time-related variable.

504
 TITLE, DATA, VARIABLE, And DEFINE Commands

Following are the options for the DATA and the DATA transformation
commands:

DATA:
FILE IS file name;
FORMAT IS format statement; FREE
FREE;
TYPE IS INDIVIDUAL; INDIVIDUAL
COVARIANCE;
CORRELATION;
FULLCOV;
FULLCORR;
MEANS;
STDEVIATIONS;
MONTECARLO;
IMPUTATION;
NOBSERVATIONS ARE number of observations;
NGROUPS = number of groups; 1
LISTWISE = ON; OFF
OFF;
SWMATRIX = file name;
VARIANCES = CHECK; CHECK
NOCHECK;
DATA IMPUTATION:
IMPUTE = names of variables for which missing values
will be imputed;
NDATASETS = number of imputed data sets; 5
SAVE = names of files in which imputed data sets
are stored;
MODEL = COVARIANCE; depends on
SEQUENTIAL; analysis type
REGRESSION;
VALUES = values imputed data can take; no restrictions
ROUNDING = number of decimals for imputed continuous 3
variables;
THIN = k where every k-th imputation is saved; 100
DATA WIDETOLONG:
WIDE = names of old wide format variables;
LONG = names of new long format variables;
IDVARIABLE = name of variable with ID information; ID
REPETITION = name of variable with repetition information; REP

505
CHAPTER 15

DATA LONGTOWIDE:
LONG = names of old long format variables;
WIDE = names of new wide format variables;
IDVARIABLE = name of variable with ID information;
REPETITION = name of variable with repetition information
(values); 0, 1, 2, etc.
DATA TWOPART:
NAMES = names of variables used to create a set of
binary and continuous variables;
CUTPOINT = value used to divide the original variables 0
into a set of binary and continuous
BINARY = variables;
CONTINUOUS = names of new binary variables;
TRANSFORM = names of new continuous variables; LOG
function to use to transform new continuous
variables;
DATA MISSING:
NAMES = names of variables used to create a set of
binary variables;
BINARY = names of new binary variables;
TYPE = MISSING;
SDROPOUT;
DDROPOUT;
DESCRIPTIVE = sets of variables for additional descriptive
statistics separated by the | symbol;
DATA SURVIVAL:
NAMES = names of variables used to create a set of
binary event-history variables;
CUTPOINT = value used to create a set of binary event-
history variables from a set of original
variables;
BINARY = names of new binary variables;
DATA COHORT:
COHORT IS name of cohort variable (values);
COPATTERN IS name of cohort/pattern variable (patterns);
COHRECODE = (old value = new value);
TIMEMEASURES = list of sets of variables separated by the |
symbol;
TNAMES = list of root names for the sets of variables in
TIMEMEASURES separated by the |
symbol;

The DATA command is a required command. The FILE option is a

required option. The NOBSERVATIONS option is required when

506
 TITLE, DATA, VARIABLE, And DEFINE Commands

summary data are analyzed. This option is not required when individual
data are analyzed. Default settings are shown in the last column. If the
default settings are appropriate for the options that are not required,
nothing needs to be specified for these options.

Note that commands and options can be shortened to four or more

letters. Option settings can be referred to by either the complete word or
the part of the word shown above in bold type.

FILE
The FILE option is used to specify the name and location of the ASCII
file that contains the data to be analyzed. The FILE option is required
for each analysis. It is specified for a single group analysis as follows:

FILE IS c:\analysis\data.dat;

where data.dat is the name of the ASCII file containing the data to be
analyzed. In this example, the file data.dat is located in the directory
c:\analysis. If the full path name of the data set contains any blanks, the
full path name must have quotes around it.

If the name of the data set is specified with a path, the directory
specified by the path is checked. If the name of the data set is specified
without a path, the local directory is checked. If the data set is not found
in the local directory, the directory where the input file is located is
checked.

FORMAT
The FORMAT option is used to describe the format of the data set to be
analyzed. Individual data can be in fixed or free format. Free format is
the default. Fixed format is recommended for large data sets because it
is faster to read data using a fixed format. Summary data must be in free
format.

For data in free format, each entry on a record must be delimited by a

comma, space, or tab. When data are in free format, the use of blanks is
not allowed. The number of variables in the data set is determined from

507
CHAPTER 15

information provided in the NAMES option of the VARIABLE

command. Data are read until the number of pieces of information equal
to the number of variables is found. The program then goes to the next
record to begin reading information for the next observation.

For data in fixed format, each observation must have the same number of
records. Information for a given variable must occupy the same position
on the same record for each observation. A FORTRAN-like format
statement describing the position of the variables in the data set is
required. Following is an example of how to specify a format statement:

FORMAT IS 5F4.0, 10x, 6F1.0;

Although any FORTRAN format descriptor (i.e., F, I, G, E, x, t, /, etc.) is

acceptable in a format statement, most format statements use only F, t, x,
and /. Following is an explanation of how to create a FORTRAN-like
format statement using these descriptors.

The F format describes the format for a real variable. F is followed by a

number. It can be a whole number or a decimal, for example, F5.3. The
number before the decimal point describes the number of columns
reserved for the variable; the number after the decimal point specifies
the number of decimal places. If the number 34234 is read with an F5.3
format, it is read as 34.234. If the data contain a decimal point, it is not
necessary to specify information about the position of the decimal point.
For example, the number 34.234 can be read with a F6 format as 34.234.

The F format can also be preceded by a number. This number represents

the number of variables to be read using that format. The statement
5F5.3 is a shorthand way of saying F5.3, F5.3, F5.3, F5.3, F5.3.

There are three options for the format statement related to skipping
columns or records when reading data: x, t, and /. The x option instructs
the program to skip columns. The statement 10x says to skip 10
columns and begin reading in column 11. The t option instructs the
program to go to a particular column and begin reading. For example,
t130 says to go to column 130 and begin reading in column 130. The /
option is used to instruct the program to go to the next record. Consider
the following format statements:

508
 TITLE, DATA, VARIABLE, And DEFINE Commands

1. (20F4, 13F5, 3F2)

2. (3F4.1,25x,5F5)
3. (3F4.1,t38,5F5)
4. (2F4/14F4.2//6F3.1)

1. In the first statement, for each record the program reads 20 four-digit
numbers followed by 13 five-digit numbers, then three two-digit
numbers with a total record length of 151.

2. In the second statement, for each record the program reads three four-
digit numbers with one digit to the right of the decimal, skips 25 spaces,
and then reads five five-digit numbers with a total record length of 62.

3. The third statement is the same as the second but uses the t option
instead of the x option. In the third statement, for each record the
program reads three four-digit numbers with one digit to the right of the
decimal, goes to column 38, and then reads five five-digit numbers.

4. In the fourth statement, each observation has four records. For record
one the program reads two four-digit numbers; for record two the
program reads fourteen four-digit numbers with two digits to the right of
the decimal; record three is skipped; and for record four the program
reads six three-digit numbers with one number to the right of the decimal
point.

Following is an example of a data set with six one-digit numbers with no

numbers to the right of the decimal point:

123234
342765
348765

The format statement for the data set above is:

FORMAT IS 6F1.0;

or

FORMAT IS 6F1;

509
CHAPTER 15

TYPE
The TYPE option is used in conjunction with the FILE option to
describe the contents of the file named using the FILE option. It has the
following settings:

INDIVIDUAL Data matrix where rows represent observations and

columns represent variables
COVARIANCE A lower triangular covariance matrix read row wise
CORRELATION A lower triangular correlation matrix read row wise
FULLCOV A full covariance matrix read row wise
FULLCORR A full correlation matrix read row wise
MEANS Means
STDEVIATIONS Standard deviations
MONTECARLO A list of the names of the data sets to be analyzed
IMPUTATION A list of the names of the imputed data sets to be
analyzed

INDIVIDUAL

The default for the TYPE option is INDIVIDUAL. The TYPE option is
not required if individual data are being analyzed where rows represent
observations and columns represent variables.

SUMMARY DATA

The TYPE option is required when summary data such as a covariance

matrix or a correlation matrix are analyzed. The TYPE option has six
settings related to the analysis of summary data. They are:
COVARIANCE, CORRELATION, FULLCOV, FULLCORR, MEANS,
and STDEVIATIONS. Summary data must reside in a free format
external ASCII file. The number of observations must be specified using
the NOBSERVATIONS option of the DATA command.

When summary data are analyzed and one or more dependent variables
are binary or ordered categorical (ordinal), only a correlation matrix can
be analyzed. When summary data are analyzed and all dependent

510
 TITLE, DATA, VARIABLE, And DEFINE Commands

variables are continuous, a covariance matrix is usually analyzed. In

some cases, a correlation matrix can be analyzed.

A data set with all continuous dependent variables in the form of a

correlation matrix, standard deviations, and means is specified as:

TYPE IS CORRELATION MEANS STDEVIATIONS;

The program creates a covariance matrix using the correlations and

standard deviations and then analyzes the means and covariance matrix.

The external ASCII file for the above example contains the means,
standard deviations, and correlations in free format. Each type of data
must begin on a separate record even if the data fits on less than one
record. The means come first; the standard deviations begin on the
record following the last mean; and the entries of the lower triangular
correlation matrix begin on the record following the last standard
deviation. The data set appears as follows:

.4 .6 .3 .5 .5
.2 .5 .4 .5 .6
1.0
.86 1.0
.56 .76 1.0
.78 .34 .48 1.0
.65 .87 .32 .56 1.0

or alternatively:

.4 .6 .3 .5 .5
.2 .5 .4 .5 .6
1.0 .86 1.0 .56 .76 1.0 .78 .34 .48 1.0 .65 .87 .32 .56 1.0

MONTECARLO

The MONTECARLO setting of the TYPE option is used when the data
sets being analyzed have been generated and saved using either the
REPSAVE option of the MONTECARLO command or by another
computer program. The file named using the FILE option of the DATA
command contains a list of the names of the data sets to be analyzed and

511
CHAPTER 15

summarized as in a Monte Carlo study. This ASCII file is created

automatically when the data sets are generated and saved in a prior
analysis using the REPSAVE option of the MONTECARLO command.
This file must be created by the user when the data sets are generated
and saved using another computer program. Each record of the file must
contain one data set name. For example, if five data sets are being
analyzed, the contents of the file would be:

data1.dat
data2.dat
data3.dat
data4.dat
data5.dat

where data1.dat, data2.dat, data3.dat, data4.dat, and data5.dat are the

names of the five data sets generated and saved using another computer
program. All files must be in the same format. Files saved using the
REPSAVE option are in free format.

When the MONTECARLO option is used, the results are presented in a

Monte Carlo summary format. The output includes the population value
for each parameter, the average of the parameter estimates across
replications, the standard deviation of the parameter estimates across
replications, the average of the estimated standard errors across
replications, the mean square error for each parameter (M.S.E.), 95
percent coverage, and the proportion of replications for which the null
hypothesis that a parameter is equal to zero is rejected at the .05 level.
In addition, the average fit statistics and the percentiles for the fit
statistics are given if appropriate. A description of Monte Carlo output
is given in Chapter 12.

IMPUTATION

The IMPUTATION setting of the TYPE option is used when the data
sets being analyzed have been generated using multiple imputation
procedures. The file named using the FILE option of the DATA
command must contain a list of the names of the multiple imputation
data sets to be analyzed. Parameter estimates are averaged over the set
of analyses. Standard errors are computed using the average of the
squared standard errors over the set of analyses and the between analysis

512
 TITLE, DATA, VARIABLE, And DEFINE Commands

parameter estimate variation (Rubin, 1987; Schafer, 1997). A chi-square

test of overall model fit is provided (Asparouhov & Muthén, 2008c;
Enders, 2010). The ASCII file containing the names of the data sets
must be created by the user. Each record of the file must contain one
data set name. For example, if five data sets are being analyzed, the
contents of the file would be:

imp1.dat
imp2.dat
imp3.dat
imp4.dat
imp5.dat

where imp1.dat, imp2.dat, imp3.dat, imp4.dat, and imp5.dat are the

names of the five data sets created using multiple imputation.

NOBSERVATIONS
The NOBSERVATIONS option is required when summary data are
analyzed. When individual data are analyzed, the program counts the
number of observations. The NOBSERVATIONS option can, however,
be used with individual data to limit the number of records used in the
analysis. For example, if a data set contains 20,000 observations, it is
possible to analyze only the first 1,000 observations by specifying:

NOBSERVATIONS = 1000;

NGROUPS
The NGROUPS option is used for multiple group analysis when
summary data are analyzed. It specifies the number of groups in the
analysis. It is specified as follows:

NGROUPS = 3;

which indicates that the analysis is a three-group analysis. Multiple

group analysis is discussed in Chapter 14.

513
CHAPTER 15

LISTWISE
The LISTWISE option is used to indicate that any observation with one
or more missing values on the set of analysis variables not be used in the
analysis. The default is to estimate the model under missing data theory
using all available data. To turn on listwise deletion, specify:

LISTWISE = ON;

SWMATRIX
The SWMATRIX option is used with TYPE=TWOLEVEL and
weighted least squares estimation to specify the name and location of the
file that contains the within- and between-level sample statistics and
their corresponding estimated asymptotic covariance matrix. The
univariate and bivariate sample statistics are estimated using one- and
two-dimensional numerical integration with a default of 7 integration
points. The INTEGRATION option of the ANALYSIS command can be
used to change the default. It is recommended to save this information
and use it in subsequent analyses along with the raw data to reduce
computational time during model estimation. Analyses using this
information must have the same set of observed dependent and
independent variables, the same DEFINE command, the same
USEOBSERVATIONS statement, and the same USEVARIABLES
statement as the analysis which was used to save the information. It is
specified as follows:

SWMATRIX = swmatrix.dat;

where swmatrix.dat is the file that contains the within- and between-
level sample statistics and their corresponding estimated asymptotic
covariance matrix.

For TYPE=IMPUTATION, the file specified contains a list of file

names. These files contain the within- and between-level sample
statistics and their corresponding estimated asymptotic covariance
matrix for a set of imputed data sets.

514
 TITLE, DATA, VARIABLE, And DEFINE Commands

VARIANCES
The VARIANCES option is used to check that the analysis variables do
not have variances of zero in the sample used for the analysis. Checking
for variances of zero is the default. To turn off this check, specify:

VARIANCES = NOCHECK;

THE DATA IMPUTATION COMMAND

The DATA IMPUTATION command is used when a data set contains
missing values to create a set of imputed data sets using multiple
imputation methodology. Imputation refers to the estimation of missing
values in a data set to create a data set without missing values. Multiple
imputation refers to the creation of several data sets where missing
values have been imputed. Multiple imputation is carried out using
Bayesian estimation. The multiple imputations are random draws from
the posterior distribution of the missing values (Rubin, 1987; Schafer,
1997). For an overview, see Enders (2010). The multiple imputation
data sets can be used for subsequent model estimation using maximum
likelihood or weighted least squares estimation of each data set where
the parameter estimates are averaged over the data sets and the standard
errors are computed using the Rubin formula (Rubin, 1987). A chi-
square test of overall model fit is provided (Asparouhov & Muthén,
2008c; Enders, 2010).

515
CHAPTER 15

Imputation model

H1 H0

Save only or MODEL command

also analyze? ESTIMATOR = BAYES
Example 11.7

Save only Analyze

MODEL command
TYPE = BASIC Not ESTIMATOR = BAYES
Example 11.5 Example 11.6

The figure above shows three ways that data imputation can be done.
The first path in the figure uses an unrestricted H1 imputation model and
saves the imputed data sets for a subsequent analysis. In this case,
TYPE=BASIC is specified in the ANALYSIS command. See Example
11.5. To use the data sets in a subsequent analysis, specify
TYPE=IMPUTATION in the DATA command. See Example 13.13.
The second path in the figure uses an unrestricted H1 imputation model
with an estimator other than BAYES. In this case, the model is
estimated immediately after the data are imputed. See Example 11.6.
The third path in the figure uses an H0 imputation model and
ESTIMATOR=BAYES. The H0 model specified in the MODEL
command is used to impute the data. See Example 11.7.

IMPUTE

The IMPUTE option is used to specify the analysis variables for which
missing values will be imputed. Data can be imputed for all or a subset
of the analysis variables. These variables can be continuous or
categorical. If they are categorical a letter c in parentheses must be
included after the variable name. If a variable is on the
CATEGORICAL list in the VARIABLE command, it must have a c in
parentheses following its name. A variable not on the CATEGORICAL
list can have a c in parentheses following its name. Following is an
example of how to specify the IMPUTE option:

516
 TITLE, DATA, VARIABLE, And DEFINE Commands

IMPUTE = y1-y4 u1-u4 (c) x1 x2;

where values will be imputed for the continuous variables y1, y2, y3, y4,
x1, and x2 and the categorical variables u1, u2, u3, and u4.

The IMPUTE option has an alternative specification that is convenient

when there are several variables that cannot be specified using the list
function. When c in parentheses follows the equal sign, it means that c
applies to all of the variables that follow. For example, the following
IMPUTE statement specifies that the variables x1, x3, x5, x7, and x9 are
categorical:

IMPUTE = (c) x1 x3 x5 x7 x9;

The keyword ALL can be used to indicate that values are to be imputed
for all variables in the dataset. The ALL option can be used with the c
setting, for example,

IMPUTE = ALL (c);

indicates that all of the variables in the data set are categorical.

NDATASETS

The NDATASETS option is used to specify the number of imputed data

sets to create. The default is five. Following is an example of how to
specify the NDATASETS option:

NDATASETS = 20;

where 20 is the number of imputed data sets that will be created. The
default for the NDATASETS option is 5.

SAVE

The SAVE option is used to save the imputed data sets for subsequent
analysis using TYPE=IMPUTATION in the DATA command. It is
specified as follows:

SAVE = impute*.dat;

517
CHAPTER 15

where the asterisk (*) is replaced by the number of the imputed data set.
A file is also produced that contains the names of all of the data sets. To
name this file, the asterisk (*) is replaced by the word list.

MODEL

The MODEL option is used to specify the type of unrestricted H1 model

to use for imputation (Asparouhov & Muthén, 2010b). The MODEL
option has three settings: COVARIANCE, SEQUENTIAL, and
REGRESSION. The default is COVARIANCE. The COVARIANCE
setting uses a model of unrestricted means, variances, and covariances
for a set of continuous variables. The SEQUENTIAL setting uses a
sequential regression method also referred to as the chained equations
algorithm in line with Raghunathan et al. (2001). The REGRESSION
setting uses a model where variables with missing data are regressed on
variables without missing data (Asparouhov & Muthén, 2010b). To
request the sequential regression method, specify:

MODEL = SEQUENTIAL;

VALUES

The VALUES option is used to provide the values for continuous

variables that the imputed data can take. The default is to put no
restrictions on the values that the imputed data can take. The values
must be integers. For example, four five-category variables not declared
as categorical can be restricted to take on only the values of one through
five by specifying:

VALUES = y1-y4 (1-5);

The closest value to the imputed value is used. If the imputed value is
2.7, the value 3 will be used.

ROUNDING

The ROUNDING option is used to specify the number of decimals that

imputed continuous variables will have. The default is three. To request
that five decimals be used, specify:

518
 TITLE, DATA, VARIABLE, And DEFINE Commands

ROUNDING = y1-y10 (5);

The value zero is used to specify no decimals, that is, integer values.

THIN

The THIN option is used to specify which intervals in the draws from
the posterior distribution are used for imputed values. The default is to
use every 100th iteration. To request that every 200th iteration be used,
specify:

THIN = 200;

THE DATA TRANSFORMATION COMMANDS

There are six DATA transformation commands. They are used to
rearrange data from a wide to long format, to rearrange data from a long
to wide format, to create a binary and a continuous variable from a
semicontinuous variable, to create a set of binary variables that are
indicators of missing data for another set of variables, to create variables
for discrete-time survival modeling where a binary variable represents
the occurrence of a single non-repeatable event, and to rearrange
longitudinal data from a format where time points represent
measurement occasions to a format where time points represent age or
another time-related variable.

THE DATA WIDETOLONG COMMAND

In growth modeling an outcome measured at four time points can be
represented in a data set in two ways. In the wide format, the outcome is
represented as four variables on a single record. In the long format, the
outcome is represented as a single variable using four records, one for
each time point. The DATA WIDETOLONG command is used to
rearrange data from a multivariate wide format to a univariate long
format.

When the data are rearranged, the set of outcomes is given a new
variable name and ID and repetition variables are created. These new
variable names must be placed on the USEVARIABLES statement of

519
CHAPTER 15

the VARIABLE command if they are used in the analysis. They must be
placed after any original variables. If the ID variable is used as a cluster
variable, this must be specified using the CLUSTER option of the
VARIABLE command.

The creation of the new variables in the DATA WIDETOLONG

command occurs after any transformations in the DEFINE command and
any of the other DATA transformation commands. If listwise deletion is
used, it occurs after the data have been rearranged. Following is a
description of the options used in the DATA WIDETOLONG command.

WIDE

The WIDE option is used to identify sets of variables in the wide format
data set that will be converted into single variables in the long format
data set. These variables must be variables from the NAMES statement
of the VARIABLE command. The WIDE option is specified as follows:

WIDE = y1-y4 | x1-x4;

where y1, y2, y3, and y4 represent one variable measured at four time
points and x1, x2, x3, and x4 represent another variable measured at four
time points.

LONG

The LONG option is used to provide names for the new variables in the
long format data set. There should be the same number of names as
there are sets of variables in the WIDE statement. The LONG option is
specified as follows:

LONG = y | x;

where y is the name assigned to the set of variables y1-y4 on the WIDE
statement and x is the name assigned to the set of variables x1-x4.

520
 TITLE, DATA, VARIABLE, And DEFINE Commands

IDVARIABLE

The IDVARIABLE option is used to provide a name for the variable that
provides information about the unit to which the record belongs. In
univariate growth modeling, this is the person identifier which is used as
a cluster variable. The IDVARIABLE option is specified as follows:

IDVARIABLE = subject;

where subject is the name of the variable that contains information about
the unit to which the record belongs. If an id variable is specified using
the IDVARIABLE option of the VARIABLE command, the values of
this variable are used for the variable specified using the IDVARIABLE
option. This option is not required.

REPETITION

The REPETITION option is used to provide a name for the variable that
contains information on the order in which the variables were measured.
The REPETITION option is specified as follows:

REPETITION = time;

where time is the variable that contains information on the order in

which the variables were measured. This variable assigns consecutive
values starting with zero to the repetitions. This variable can be used in
a growth model as a time score variable. This option is not required.

THE DATA LONGTOWIDE COMMAND

In growth modeling an outcome measured at four time points can be
represented in a data set in two ways. In the long format, the outcome is
represented as a single variable using four records, one for each time
point. In the wide format, the outcome is represented as four variables
on a single record. The DATA LONGTOWIDE command is used to
rearrange data from a univariate long format to a multivariate wide
format.

521
CHAPTER 15

When the data are rearranged, the outcome is given a set of new variable
names. These new variable names must be placed on the
USEVARIABLES statement of the VARIABLE command if they are
used in the analysis. They must be placed after any original variables.

The creation of the new variables in the DATA LONGTOWIDE

command occurs after any transformations in the DEFINE command and
any of the other DATA transformation commands. Following is a
description of the options used in the DATA LONGTOWIDE command.

LONG

The LONG option is used to identify the variables in the long format
data set that will be used to create sets of variables in the wide format
data set. These variables must be variables from the NAMES statement
of the VARIABLE command. The LONG option is specified as follows:

LONG = y | x;

where y and x are two variables that have been measured at multiple
time points which are represented by multiple records.

WIDE

The WIDE option is used to provide sets of names for the new variables
in the wide format data set. There should be the same number of sets of
names as there are variables in the LONG statement. The number of
names in each set corresponds to the number of time points at which the
variables in the long data set were measured. The WIDE option is
specified as follows:

WIDE = y1-y4 | x1-x4;

where y1, y2, y3, and y4 are the names for the variable y in the wide data
set and x1, x2, x3, and x4 are the names for the variable x in the wide
data set.

522
 TITLE, DATA, VARIABLE, And DEFINE Commands

IDVARIABLE

The IDVARIABLE option is used to identify the variable in the long

data set that contains information about the unit to which each record
belongs. The IDVARIABLE option is specified as follows:

IDVARIABLE = subject;

where subject is the name of the variable that contains information about
the unit to which each record belongs. This variable becomes the
identifier for each observation in the wide data set. The IDVARIABLE
option of the VARIABLE command cannot be used to select a different
identifier.

REPETITION

The REPETITION option is used to identify the variable that contains

information about the times at which the variables in the long data set
were measured. The REPETITION option is specified as follows:

REPETITION = time;

where time is the variable that contains information about the time at
which the variables in the long data set were measured. If the time
variable does not contain consecutive integer values starting at zero, the
time values must be given. For example,

REPETITION = time (4 8 16);

specifies that the values 4, 8, and 16 are the values of the variable time.
The number of values should be equal to the number of variables in the
WIDE option and the order of the values should correspond to the order
of the variables.

THE DATA TWOPART COMMAND

The DATA TWOPART command is used to create a binary and a
continuous variable from a continuous variable with a floor effect for
use in two-part (semicontinuous) modeling (Olsen & Schafer, 2001).

523
CHAPTER 15

One situation where this occurs is when variables have a preponderance

of zeros.

A set of binary and continuous variables are created using the value
specified in the CUTPOINT option of the DATA TWOPART command
or zero which is the default. The two variables are created using the
following rules:

1. If the value of the original variable is missing, both the new binary
and the new continuous variable values are missing.
2. If the value of the original variable is greater than the cutpoint value,
the new binary variable value is one and the new continuous variable
value is the log of the original variable as the default.
3. If the value of the original variable is less than or equal to the
cutpoint value, the new binary variable value is zero and the new
continuous variable value is missing.

The new variables must be placed on the USEVARIABLES statement of

the VARIABLE command if they are used in the analysis. These
variables must come after any original variables. If the binary variables
are used as dependent variables in the analysis, they must be declared as
categorical using the CATEGORICAL option of the VARIABLE
command.

The creation of the new variables in the DATA TWOPART command

occurs after any transformations in the DEFINE command and before
any transformations using the DATA MISSING command. Following is
a description of the options used in the DATA TWOPART command.

NAMES

The NAMES option identifies the variables that are used to create a set
of binary and continuous variables. These variables must be variables
from the NAMES statement of the VARIABLE command. The NAMES
option is specified as follows:

NAMES = smoke1-smoke4;

where smoke1, smoke2, smoke3, and smoke4 are the semicontinuous

variables that are used to create a set of binary and continuous variables.

524
 TITLE, DATA, VARIABLE, And DEFINE Commands

CUTPOINT

The CUTPOINT option is used to provide the value that is used to divide
the original variables into a set of binary and continuous variables. The
default value for the CUTPOINT option is zero. The CUTPOINT
option is specified as follows:

CUTPOINT = 1;

where variables are created based on values being less than or equal to
one or greater than one.

BINARY

The BINARY option is used to assign names to the new set of binary
variables. The BINARY option is specified as follows:

BINARY = u1-u4;

where u1, u2, u3, and u4 are the names of the new set of binary
variables.

CONTINUOUS

The CONTINUOUS option is used to assign names to the new set of

continuous variables. The CONTINUOUS option is specified as
follows:

CONTINUOUS = y1-y4;

where y1, y2, y3, and y4 are the names of the new set of continuous
variables.

TRANSFORM

The TRANSFORM option is used to transform the new continuous

variables. The LOG function is the default. The following functions
can be used with the TRANSFORM option:

525
CHAPTER 15

LOG base e log LOG (y);

LOG10 base 10 log LOG10 (y);
EXP exponential EXP (y);
SQRT square root SQRT (y);
ABS absolute value ABS(y);
SIN sine SIN (y);
COS cosine COS (y);
TAN tangent TAN(y);
ASIN arcsine ASIN (y);
ACOS arccosine ACOS (y);
ATAN arctangent ATAN (y);
NONE no transformation

The TRANSFORM option is specified as follows:

TRANSFORM = NONE;

where specifying NONE results in no transformation of the new

continuous variables.

THE DATA MISSING COMMAND

The DATA MISSING command is used to create a set of binary
variables that are indicators of missing data or dropout for another set of
variables. Dropout indicators can be scored as discrete-time survival
indicators or dropout dummy indicators. The new variables can be used
to study non-ignorable missing data (Little & Rubin, 2002; Muthén et
al., 2011).

The new variables must be placed on the USEVARIABLES statement of

the VARIABLE command if they are used in the analysis. These
variables must come after any original variables. If the binary variables
are used as dependent variables in the analysis, they must be declared as
categorical using the CATEGORICAL option of the VARIABLE
command.

The creation of the new variables in the DATA MISSING command

occurs after any transformations in the DEFINE command and after any
transformations using the DATA TWOPART command. Following is a
description of the options used in the DATA MISSING command.

526
 TITLE, DATA, VARIABLE, And DEFINE Commands

NAMES

The NAMES option identifies the set of variables that are used to create
a set of binary variables that are indicators of missing data. These
variables must be variables from the NAMES statement of the
VARIABLE command. The NAMES option is specified as follows:

NAMES = drink1-drink4;

where drink1, drink2, drink3, and drink4 are the set of variables for
which a set of binary indicators of missing data are created.

BINARY

The BINARY option is used to assign names to the new set of binary
variables. The BINARY option is specified as follows:

BINARY = u1-u4;

where u1, u2, u3, and u4 are the names of the new set of binary
variables.

For TYPE=MISSING, the number of binary indicators is equal to the

number of variables in the NAMES statement. For TYPE=SDROPOUT
and TYPE=DDROPOUT, the number of binary indicators is one less
than the number of variables in the NAMES statement because dropout
cannot occur before the second time point an individual is observed.

TYPE

The TYPE option is used to specify how missingness is coded. It has

three settings: MISSING, SDROPOUT, and DDROPOUT. The default
is MISSING. For the MISSING setting, a binary missing data indicator
variable is created. For the SDROPOUT setting, which is used with
selection missing data modeling, a binary discrete-time (event-history)
survival dropout indicator is created. For the DDROPOUT setting,
which is used with pattern-mixture missing data modeling, a binary
dummy dropout indicator is created. The TYPE option is specified as
follows:

527
CHAPTER 15

TYPE = SDROPOUT;

Following are the rules for creating the set of binary variables for the
MISSING setting:

1. If the value of the original variable is missing, the new binary

variable value is one.
2. If the value of the original variable is not missing, the new binary
variable value is zero.

For the SDROPOUT and DDROPOUT settings, the set of indicator

variables is defined by the last time point an individual is observed.

Following are the rules for creating the set of binary variables for the
SDROPOUT setting:

1. The value one is assigned to the time point after the last time point
an individual is observed.
2. The value missing is assigned to all time points after the value of
one.
3. The value zero is assigned to all time points before the value of one.

Following are the rules for creating the set of binary variables for the
DDROPOUT setting:

1. The value one is assigned to the time point after the last time point
an individual is observed.
2. The value zero is assigned to all other time points.

DESCRIPTIVE

The DESCRIPTIVE option is used in conjunction with TYPE=BASIC

of the ANALYSIS command and the SDROPOUT and DDROPOUT
settings of the TYPE option to specify the sets of variables for which
additional descriptive statistics are computed. For each variable, the
mean and standard deviation are computed using all observations
without missing on the variable. Means and standard deviations are
provided for the following sets of observations whose definitions are
based on missing data patterns:

528
 TITLE, DATA, VARIABLE, And DEFINE Commands

Dropouts after each time point – Individuals who drop out before the
next time point and do not return to the study
Non-dropouts after each time point – Individuals who do not drop out
before the next time point
Total Dropouts – Individuals who are missing at the last time point
Dropouts no intermittent missing – Individuals who do not return to
the study once they have dropped out
Dropouts intermittent missing – Individuals who drop out and return
to the study
Total Non-dropouts – Individuals who are present at the last time point
Non-dropouts complete data – Individuals with complete data
Non-dropouts intermittent missing – Individuals who have missing
data but are present at the last time point
Total sample

The first set of variables given in the DESCRIPTIVE statement is the

outcome variable. This set of variables defines the number of time
points in the model. If the other sets of variables do not have the same
number of time points, the asterisk (*) is used as a placeholder. Sets of
variables are separated by the | symbol. Following is an example of how
to specify the DESCRIPTIVE option:

DESCRIPTIVE = y0-y5 | x0-x5 | * z1-z5;

The first set of variables, y0-y5 defines the number of time points as six.
The last set of variables has only five measures. An asterisk (*) is used
as a placeholder for the first time point.

THE DATA SURVIVAL COMMAND

The DATA SURVIVAL command is used to create variables for
discrete-time survival modeling where a binary discrete-time survival
(event-history) variable represents whether or not a single non-
repeatable event has occurred in a specific time period.

A set of binary discrete-time survival variables is created using the

following rules:

1. If the value of the original variable is missing, the new binary

variable value is missing.

529
CHAPTER 15

2. If the value of the original variable is greater than the cutpoint value,
the new binary variable value is one which represents that the event
has occurred.
3. If the value of the original variable is less than or equal to the
cutpoint value, the new binary variable value is zero which
represents that the event has not occurred.
4. After a discrete-time survival variable for an observation is assigned
the value one, subsequent discrete-time survival variables for that
observation are assigned the value of the missing value flag.

The new variables must be placed on the USEVARIABLES statement of

the VARIABLE command if they are used in the analysis. These
variables must come after any original variables. If the binary variables
are used as dependent variables in the analysis, they must be declared as
categorical using the CATEGORICAL option of the VARIABLE
command.

The creation of the new variables in the DATA SURVIVAL command

occurs after any transformations in the DEFINE command, the DATA
TWOPART command, and the DATA MISSING command. Following
is a description of the options used in the DATA SURVIVAL command.

NAMES

The NAMES option identifies the variables that are used to create a set
of binary event-history variables. These variables must be variables
from the NAMES statement of the VARIABLE command. The NAMES
option is specified as follows:

NAMES = dropout1-dropout4;

where dropout1, dropout2, dropout3, and dropout4 are the variables that
are used to create a set of binary event-history variables.

CUTPOINT

The CUTPOINT option is used provide the value to use to create a set of
binary event-history variables from a set of original variables. The
default value for the CUTPOINT option is zero. The CUTPOINT
option is specified as follows:

530
 TITLE, DATA, VARIABLE, And DEFINE Commands

CUTPOINT = 1;

where variables are created based on values being less than or equal to
one or greater than one.

BINARY

The BINARY option is used to assign names to the new set of binary
event-history variables. The BINARY option is specified as follows:

BINARY = u1-u4;

where u1, u2, u3, and u4 are the names of the new set of binary event-
history variables.

THE DATA COHORT COMMAND

The DATA COHORT command is used to rearrange longitudinal data
from a format where time points represent measurement occasions to a
format where time points represent age or another time-related variable.
It is available only for continuous outcomes. Multiple cohort analysis is
described in Chapter 14.

The new variables must be placed on the USEVARIABLES statement of

the VARIABLE command if they are used in the analysis.

These variables must come after any original variables. The creation of
the new variables in the DATA COHORT command occurs after any
transformations in the DEFINE command. Following is a description of
the options used in the DATA COHORT command.

COHORT

The COHORT option is used when data have been collected using a
multiple cohort design. The COHORT option is used in conjunction
with the TIMEMEASURES and TNAMES options that are described
below. Variables used with the COHORT option must be variables from
the NAMES statement of the VARIABLE command. Following is an
example of how the COHORT option is specified:

531
CHAPTER 15

COHORT IS birthyear (63 64 65);

where birthyear is a variable in the data set to be analyzed, and the

numbers in parentheses following the variable name are the values that
the birthyear variable contains. Birth years of 1963, 1964, and 1965 are
included in the example below. The cohort variable must contain only
integer values.

COPATTERN

The COPATTERN option is used when data are both missing by design
and have been collected using a multiple cohort design. Variables used
with the COPATTERN option must be variables from the NAMES
statement of the VARIABLE command. Following is an example of
how the COPATTERN option is specified:

COPATTERN = cohort (67=y1 y2 y3 68=y4 y5 y6 69=y2 y3 y4);

where cohort is a variable that provides information about both the

cohorts included in the data set and the patterns of variables for each
cohort. In the example above, individuals in cohort 67 should have
information on y1, y2, and y3; individuals in cohort 68 should have
information on y4, y5, and y6; and individuals in cohort 69 should have
information on y2, y3, and y4. Individuals who have missing values on
any variable for which they are expected to have information are
eliminated from the analysis. The copattern variable must contain only
integer values.

COHRECODE

The COHRECODE option is used in conjunction with either the

COHORT or COPATTERN options to recode the values of the cohort or
copattern variable. The COHRECODE option is specified as follows:

COHRECODE = (1=67 2=68 3=69 4=70);

where the original values of 1, 2, 3, and 4 of the cohort or copattern

variable are recoded to 67, 68, 69, and 70, respectively. If the
COHRECODE option is used, all values of the original variable must be

532
 TITLE, DATA, VARIABLE, And DEFINE Commands

recoded to be included in the analysis. Observations with values that are

not recoded will be eliminated from the analysis.

TIMEMEASURES

The TIMEMEASURES option is used with multiple cohort data to

specify the years in which variables to be used in the analysis were
measured. It is used in conjunction with the COHORT and
COPATTERN options to determine the ages that are represented in the
multiple cohort data set. Variables used with the TIMEMEASURES
option must be variables from the NAMES statement of the VARIABLE
command. Following is an example of how the TIMEMEASURES
option is specified:

TIMEMEASURES = y1 (82) y2 (84) y3 (85) y4 (88) y5 (94);

where y1, y2, y3, y4, and y5 are original variables that are to be used in
the analysis, and the numbers in parentheses following each of these
variables represent the years in which they were measured. In this
situation, y1, y2, y3, y4, and y5 are the same measure, for example,
frequency of heavy drinking measured on multiple occasions.

The TIMEMEASURES option can be used to identify more than one

measure that has been measured repeatedly as shown in the following
example:

TIMEMEASURES = y1 (82) y2 (84) y3 (85) y4 (88) y5 (94) |

y6 (82) y7 (85) y8 (90) y9 (95) |
x1 (83) x2(88) x3 (95);

where each set of variables separated by the symbol | represents repeated

measures of that variable. For example, y1, y2, y3, y4, and y5 may
represent repeated measures of heavy drinking; y6, y7, y8, and y9 may
represent repeated measures of alcohol dependence; and x1, x2, and x3
may represent repeated measures of marital status.

TNAMES

The TNAMES option is used to generate variable names for the new
multiple cohort analysis variables. A root name is specified for each set

533
CHAPTER 15

of variables mentioned using the TIMEMEASURES option. The age of

the respondent at the time the variable was measured is attached to the
root name. The age is determined by subtracting the cohort value from
the year the variable was measured. Following is an example of how the
TNAMES option is specified:

TNAMES = hd;

where hd is the root name for the new variables.

Following is an example of how the TNAMES option is specified for the

TIMEMEASURES and COHORT options when multiple outcomes are
measured:

TNAMES = hd | dep | marstat;

Following are the variables that would be created:

hd22, hd24, hd25, hd26, hd27, hd28, hd29, hd30,

hd31, hd32, hd33, hd34, hd36, hd37, hd38, hd39,
dep22, dep24, dep25, dep26, dep27, dep28, dep29, dep30,
dep32, dep33, dep35, dep36, dep37, dep38, dep39, dep40,
marstat23, marstat25, marstat26, marstat27, marstat28
marstat30, marstat31, marstat32, marstat33, marstat35
marstat37, marstat38, marstat39, marstat40.

There is no hd variable for ages 23 and 35, no dep variable for ages 23,
31, and 34, and no marstat variable for ages 24, 29, 34, and 36 because
these ages are not represented by the combination of cohort values and
years of measurement.

THE VARIABLE COMMAND

The VARIABLE command is used to provide information about the
variables in the data set to be analyzed. The VARIABLE command has
options for naming and describing the variables in the data set to be
analyzed, subsetting the data set on observations, subsetting the data set
on variables, and specifying missing values for each variable.

534
 TITLE, DATA, VARIABLE, And DEFINE Commands

Following are the options for the VARIABLE command:

VARIABLE:

NAMES ARE names of variables in the data set;

USEOBSERVATIONS ARE conditional statement to select observations; all observations
in data set
USEVARIABLES ARE names of analysis variables; all variables in
NAMES
MISSING ARE variable (#);
.;
*;
BLANK;
CENSORED ARE names, censoring type, and inflation status for
censored dependent variables;
CATEGORICAL ARE names of binary and ordered categorical
(ordinal) dependent variables;
NOMINAL ARE names of unordered categorical (nominal)
dependent variables;
COUNT ARE names of count variables (model);
DSURVIVAL ARE names of discrete-time survival variables;
GROUPING IS name of grouping variable (labels);
IDVARIABLE IS name of ID variable;
FREQWEIGHT IS name of frequency (case) weight variable;
TSCORES ARE names of observed variables with information
on individually-varying times of observation;
AUXILIARY = names of auxiliary variables;
names of auxiliary variables (M);
names of auxiliary variables (R);
names of auxiliary variables (R3STEP);
names of auxiliary variables (E);
names of auxiliary variables (DU3STEP);
names of auxiliary variables (DE3STEP);
CONSTRAINT = names of observed variables that can be used
in the MODEL CONSTRAINT command;
PATTERN IS name of pattern variable (patterns);
STRATIFICATION IS name of stratification variable;
CLUSTER IS name of cluster variables;
WEIGHT IS name of sampling weight variable;
WTSCALE IS UNSCALED; CLUSTER
CLUSTER;
ECLUSTER;
BWEIGHT name of between-level sampling weight
variable;
B2WEIGHT IS name of the level 2 sampling weight variable;

535
CHAPTER 15

B3WEIGHT IS name of the level 3 sampling weight variable;

BWTSCALE IS UNSCALED; SAMPLE
SAMPLE;
REPWEIGHTS ARE names of replicate weight variables;
SUBPOPULATION IS conditional statement to select subpopulation; all observations
in data set
FINITE = name of variable; FPC
name of variable (FPC);
name of variable (SFRACTION);
name of variable (POPULATION);
CLASSES = names of categorical latent variables (number
of latent classes);
KNOWNCLASS = name of categorical latent variable with known
class membership (labels);
TRAINING = names of training variables; MEMBERSHIP
names of variables (MEMBERSHIP);
names of variables (PROBABILITIES);
names of variables (PRIORS);
WITHIN ARE names of individual-level observed variables;
WITHIN ARE (label) names of individual-level observed variables;

BETWEEN ARE names of cluster-level observed variables;

BETWEEN ARE (label) names of cluster-level observed variables;
SURVIVAL ARE names and time intervals for time-to-event
variables;
TIMECENSORED ARE names and values of variables that contain (0 = NOT
right censoring information; 1 = RIGHT)

The VARIABLE command is a required command. The NAMES option

is a required option. Default settings are shown in the last column. If
the default settings are appropriate for the analysis, nothing needs to be
specified except the NAMES option.

Note that commands and options can be shortened to four or more

letters. Option settings can be referred to by either the complete word or
the part of the word shown above in bold type.

536
 TITLE, DATA, VARIABLE, And DEFINE Commands

ASSIGNING NAMES TO VARIABLES

NAMES

The NAMES option is used to assign names to the variables in the data
set named using the FILE option of the DATA command. This option is
required. The variable names can be separated by blanks or commas and
can be up to 8 characters in length. Variable names must begin with a
letter. They can contain only letters, numbers, and the underscore
symbol. The program makes no distinction between upper and lower
case letters. Following is an example of how the NAMES option is
specified:

NAMES ARE gender ethnic income educatn drink_st agedrink;

Variable names are generated if a list of variables is specified using the

NAMES option. For example,

NAMES ARE y1-y5 x1-x3;

generates the variable names y1 y2 y3 y4 y5 x1 x2 x3.

NAMES ARE itema-itemd;

generates the variable names itema itemb itemc itemd.

SUBSETTING OBSERVATIONS AND VARIABLES

There are options for selecting a subset of observations or variables from
the data set named using the FILE option of the DATA command. The
USEOBSERVATIONS option is used to select a subset of observations
from the data set. The USEVARIABLES option is used to select a
subset of variables from the data set.

537
CHAPTER 15

USEOBSERVATIONS

The USEOBSERVATIONS option is used to select observations for an

analysis from the data set named using the FILE option of the DATA
command. This option is not available for summary data. The
USEOBSERVATIONS option selects only those observations that
satisfy the conditional statement specified after the equal sign. For
example, the following statement selects observations with the variable
ethnic equal to 1 and the variable gender equal to 2:

USEOBSERVATIONS = ethnic EQ 1 AND gender EQ 2;

Only variables from the NAMES statement of the VARIABLE command

can be used in the conditional statement of the USEOBSERVATIONS
option. Logical operators, not arithmetic operators, must be used in the
conditional statement. Following are the logical operators that can be
used in conditional statements to select observations for analysis:

AND logical and

OR logical or
NOT logical not
EQ equal ==
NE not equal /=
GE greater than or equal to >=
LE less than or equal to <=
GT greater than >
LT less than <

As shown above, some of the logical operators can be referred to in two

different ways. For example, equal can be referred to as EQ or ==.

USEVARIABLES

The USEVARIABLES option is used to select variables for an analysis.

It can be used with individual data or summary data. Variables included
on the USEVARIABLES statement can be variables from the NAMES
statement of the VARIABLE command and variables created using the
DEFINE command and the DATA transformation commands. New
variables created using the DEFINE command and the DATA

538
 TITLE, DATA, VARIABLE, And DEFINE Commands

transformation commands must be included on the USEVARIABLES

statement.

The USEVARIABLES option identifies the observed dependent and

independent variables that are used in an analysis. Variables with
special functions such as grouping variables do not need to be included
on the USEVARIABLES statement unless they are new variables
created using the DEFINE command or the DATA transformation
commands. Following is an example of how to specify the
USEVARIABLES option:

USEVARIABLES ARE gender income agefirst;

Variables on the USEVARIABLES statement must follow a particular

order. The order of the variables is important because it determines the
order of variables used with the list function. The set of original
variables from the NAMES statement of the VARIABLE command must
be listed before the set of new variables created using the DEFINE
command or the DATA transformation commands. Within the two sets
of original and new variables, any order is allowed.

If all of the original variables plus some of the new variables are used in
the analysis, the keyword ALL can be used as the first entry in the
USEVARIABLES statement. This indicates that all of the original
variables from the NAMES statement of the VARIABLE command are
used in the analysis. The keyword ALL is followed by the names of the
new variables created using the DEFINE command or the DATA
transformation commands that will be used in the analysis. Following is
an example of how to specify the USEVARIABLES option for this
situation:

USEVARIABLES = ALL hd1 hd2 hd3;

where ALL refers to the total set of original variables and hd1, hd2, and
hd3 are new variables created using the DEFINE command or the DATA
transformation commands.

539
CHAPTER 15

MISSING VALUES
MISSING

The MISSING option is used to identify the values or symbol in the

analysis data set that are treated as missing or invalid. Any numeric
value and the non-numeric symbols of the period, asterisk (*), or blank
can be used as missing value flags. There is no default missing value
flag. Numeric and non-numeric missing value flags cannot be combined.
The blank cannot be used as a missing value flag for data in free format.
When a list of missing value flags contains a negative number, the
entries must be separated by commas.

NON-NUMERIC MISSING VALUE FLAGS

The period (.), the asterisk (*), or the blank can be used as non-numeric
missing value flags. Only one non-numeric missing value flag can be
used for a particular data set. This missing value flag applies to all
variables in the data set. The blank cannot be used with free format data.
With fixed format data, blanks in the data not declared as missing value
flags are treated as zeroes.

The following command indicates that the period is the missing value
flag for all variables in the data set:

MISSING ARE . ;

The blank can be a missing value flag only in fixed format data sets.
The following command indicates that blanks are to be considered as
missing value flags:

MISSING = BLANK;

NUMERIC MISSING VALUE FLAGS

Any number of numeric values can be specified as missing value flags

for the complete data set or for individual variables. The keyword ALL
can be used with the MISSING option if all variables have the same
numeric value(s) as missing value flags.

540
 TITLE, DATA, VARIABLE, And DEFINE Commands

The following statement specifies that the number 9 is the missing value
flag for all variables in the data set:

MISSING ARE ALL (9);

The following example specifies that for the variable ethnic, the
numbers 9 and 99 are missing value flags, while for the variable y1, the
number 1 is the missing value flag:

MISSING ARE ethnic (9 99) y1 (1);

The list function can be used with the MISSING option to specify a list
of missing value flags and/or a set of variables. The order of variables in
the list is determined by the order of variables in the NAMES statement
of the VARIABLE command. Values of 9, 99, 100, 101, and 102 can be
declared as missing value flags for all variables in a data set by the
following specification:

MISSING ARE ALL (9 99-102);

Missing values can be specified for a list of variables as follows:

MISSING ARE gender-income (9 30 98-102);

The above statement specifies that the values of 9, 30, 98, 99, 100, 101,
and 102 are missing value flags for the list of variables beginning with
gender and ending with income.

MEASUREMENT SCALE OF OBSERVED

DEPENDENT VARIABLES
All observed dependent variables are assumed to be measured on a
continuous scale unless the CENSORED, CATEGORICAL, NOMINAL,
and/or COUNT options are used. The specification of the scale of the
dependent variables determines how the variables are treated in the
model and its estimation. Independent variables can be binary or
continuous. The scale of the independent variables has no influence on
the model or its estimation. The distinction between dependent and
independent variables is described in Chapter 17.

541
CHAPTER 15

Variables named using the CENSORED, CATEGORICAL, NOMINAL,

and/or COUNT options can be variables from the NAMES statement of
the VARIABLE command and variables created using the DEFINE
command and the DATA transformation commands.

CENSORED

The CENSORED option is used to specify which dependent variables

are treated as censored variables in the model and its estimation, whether
they are censored from above or below, and whether a censored or
censored-inflated model will be estimated.

The CENSORED option is specified as follows for a censored model:

CENSORED ARE y1 (a) y2 (b) y3 (a) y4 (b);

where y1, y2, y3, y4 are censored dependent variables in the analysis.
The letter a in parentheses following the variable name indicates that the
variable is censored from above. The letter b in parentheses following
the variable name indicates that the variable is censored from below.
The lower and upper censoring limits are determined from the data.

The CENSORED option is specified as follows for a censored-inflated

model:

CENSORED ARE y1 (ai) y2 (bi) y3 (ai) y4 (bi);

where y1, y2, y3, y4 are censored dependent variables in the analysis.
The letters ai in parentheses following the variable name indicate that
the variable is censored from above and that a censored-inflated model
will be estimated. The letters bi in parentheses following the variable
name indicate that the variable is censored from below and that a
censored-inflated model will be estimated. The lower and upper
censoring limits are determined from the data.

With a censored-inflated model, two variables are considered, a

censored variable and an inflation variable. The censored variable takes
on values for individuals who are able to assume values of the censoring
point and beyond. The inflation variable is a binary latent variable in
which the value one denotes that an individual is unable to assume any

542
 TITLE, DATA, VARIABLE, And DEFINE Commands

value except the censoring point. The inflation variable is referred to by

adding to the name of the censored variable the number sign (#)
followed by the number 1. In the example above, the censored variables
available for use in the MODEL command are y1, y2, y3, and y4, and
the inflation variables available for use in the MODEL command are
y1#1, y2#1, y3#1, and y4#1.

CATEGORICAL

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. Ordered categorical dependent variables
cannot have more than 10 categories. The number of categories is
determined from the data. The CATEGORICAL option is specified as
follows:

CATEGORICAL ARE u2 u3 u7-u13;

where u2, u3, u7, u8, u9, u10, u11, u12, and u13 are binary or ordered
categorical dependent variables in the analysis.

For categorical dependent variables, there are as many thresholds as

there are categories minus one. The thresholds are referred to in the
MODEL command by adding to the variable name the dollar sign ($)
followed by a number. The threshold for a binary variable u1 is referred
to as u1$1. The two thresholds for a three-category variable u2 are
referred to as u2$1 and u2$2.

The estimation of the model for binary or ordered categorical dependent

variables uses zero to denote the lowest category, one to denote the
second lowest category, two to denote the third lowest category, etc. If
the variables are not coded this way in the data, they are automatically
recoded as described below. When data are saved for subsequent
analyses, the recoded categories are saved.

Following are examples of situations in which data are recoded by the

program:

543
CHAPTER 15

Categories in Original Data Categories in Recoded Data

1234 0123
2345 0123
2589 0123
01 no recode needed
12 01

In most situations, the default recoding is appropriate. In multiple group

analysis and growth modeling, the default recoding may not be
appropriate because the categories observed in the data for a variable
may not be the same across groups or time. For example, it is sometimes
the case that individuals are observed in lower categories at earlier time
points and higher categories at later time points. Several variations of
the CATEGORICAL option are available for these situations. These are
allowed only for maximum likelihood estimation.

Using the automatic recoding, each variable is recoded using the

categories found in the data for that variable. Following is an example
of how to specify the CATEGORICAL option so that each variable is
recoded using the categories found in the data for a set of variables:

CATEGORICAL u1-u3 (*);

where u1, u2, and u3 are a set of ordered categorical variables and the
asterisk (*) in parentheses indicates that the categories of each variable
are to be recoded using the categories found in the data for the set of
variables not for each variable. Based on the original data shown in the
table below, where the rows represent observations and the columns
represent variables, the set of variables are found to have four possible
categories: 1, 2, 3, and 4. The variable u1 has observed categories 1
and 2; u2 has observed categories 1, 2, and 3; and u3 has observed
categories 2, 3, and 4. The recoded values are shown in the table below.

Categories in the Original Data Set Categories in the Recoded Data Set
u1 u2 u3 u1 u2 u3
1 2 3 0 1 2
1 1 2 0 0 1
2 2 2 1 1 1
2 3 4 1 2 3

544
 TITLE, DATA, VARIABLE, And DEFINE Commands

The CATEGORICAL option can be used to give a set of categories that

are allowed for a variable or set of variables rather than having these
categories determined from the data. Following is an example of how to
specify this:

CATEGORICAL = u1-u3 (1-6);

where the set of variables u1, u2, and u3 can have the categories of 1, 2,
3, 4, 5, and 6. In this example, 1 will be recoded as 0, 2 as 1, 3 as 2, 4 as
3, 5 as 4, and 6 as 5.

Another variation of this is:

CATEGORICAL = u1-u3 (2 4 6);

where the set of variables u1, u2, and u3 can have the categories 2, 4,
and 6. In this example, 2 will be recoded as 0, 4 as 1, and 6 as 2.

The CATEGORICAL option can be used to specify that different sets of

variables have different sets of categories by using the | symbol. For
example,

CATEGORICAL = u1-u3 (*) | u4-u6 (2-5) | u7-u9;

specifies that for the variables u1, u2, and u3, the possible categories are
taken from the data for the set of variables; for the variables u4, u5, and
u6, the possible categories are 2, 3, 4, and 5; and for the variables u7, u8,
and u9, the possible categories are the default, that is, the possible
categories are taken from the data for each variable.

NOMINAL

The NOMINAL option is used to specify which dependent variables are

treated as unordered categorical (nominal) variables in the model and its
estimation. Unordered categorical dependent variables cannot have
more than 10 categories. The number of categories is determined from
the data. The NOMINAL option is specified as follows:

NOMINAL ARE u1 u2 u3 u4;

545
CHAPTER 15

where u1, u2, u3, u4 are unordered categorical dependent variables in

the analysis.

For nominal dependent variables, all categories but the last category can
be referred to. The last category is the reference category. The
categories are referred to in the MODEL command by adding to the
variable name the number sign (#) followed by a number. The three
categories of a four-category nominal variable are referred to as u1#1,
u1#2, and u1#3.

The estimation of the model for unordered categorical dependent

variables uses zero to denote the lowest category, one to denote the
second lowest category, two to denote the third lowest category, etc. If
the variables are not coded this way in the data, they are automatically
recoded as described below. When data are saved for subsequent
analyses, the recoded categories are saved.

Following are examples of situations in which data are recoded by the

program:

Categories in Original Data Categories in Recoded Data

1234 0123
2345 0123
2589 0123
01 no recode needed
12 01

COUNT

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and the type of
model to be estimated. The following models can be estimated for count
variables: Poisson, zero-inflated Poisson, negative binomial, zero-
inflated negative binomial, zero-truncated negative binomial, and
negative binomial hurdle (Long, 1997; Hilbe, 2011). The negative
binomial models use the NB-2 variance representation (Hilbe, 2011, p.
63). Count variables may not have negative or non-integer values.

The COUNT option can be specified in two ways for a Poisson model:

546
 TITLE, DATA, VARIABLE, And DEFINE Commands

COUNT = u1 u2 u3 u4;

or

COUNT = u1 (p) u2 (p) u3 (p) u4 (p);

or using the list function:

COUNT = u1-u4 (p);

The COUNT option can be specified in two ways for a zero-inflated

Poisson model:

COUNT = u1-u4 (i);

or

COUNT = u1-u4 (pi);

where u1, u2, u3, and u4 are count dependent variables in the analysis.
The letter i or pi in parentheses following the variable name indicates
that a zero-inflated Poisson model will be estimated.

With a zero-inflated Poisson model, two variables are considered, a

count variable and an inflation variable. The count variable takes on
values for individuals who are able to assume values of zero and above
following the Poisson model. The inflation variable is a binary latent
variable with one denoting that an individual is unable to assume any
value except zero. The inflation variable is referred to by adding to the
name of the count variable the number sign (#) followed by the number
1.

Following is the specification of the COUNT option for a negative

binomial model:

COUNT = u1 (nb) u2 (nb) u3 (nb) u4 (nb);

or using the list function:

COUNT = u1-u4 (nb);

547
CHAPTER 15

With a negative binomial model, dispersion parameters are estimated.

The dispersion parameters are referred to by using the names of the
count variables.

Following is the specification of the COUNT option for a zero-inflated

negative binomial model:

COUNT = u1- u4 (nbi);

With a zero-inflated negative binomial model, two variables are

considered, a count variable and an inflation variable. The count
variable takes on values for individuals who are able to assume values of
zero and above following the negative binomial model. The inflation
variable is a binary latent variable with one denoting that an individual is
unable to assume any value except zero. The inflation variable is
referred to by adding to the name of the count variable the number sign
(#) followed by the number 1.

Following is the specification of the COUNT option for a zero-truncated

negative binomial model:

COUNT = u1-u4 (nbt);

Count variables for the zero-truncated negative binomial model must

have values greater than zero.

Following is the specification of the COUNT option for a negative

binomial hurdle model:

COUNT = u1-u4 (nbh);

With a negative binomial hurdle model, two variables are considered, a

count variable and a hurdle variable. The count variable takes on values
for individuals who are able to assume values of one and above
following the truncated negative binomial model. The hurdle variable is
a binary latent variable with one denoting that an individual is unable to
assume any value except zero. The hurdle variable is referred to by
adding to the name of the count variable the number sign (#) followed by
the number 1.

548
 TITLE, DATA, VARIABLE, And DEFINE Commands

DSURVIVAL

The DSURVIVAL option is used in conjunction with the PLOT

command to identify the discrete-time survival variables so that survival
curves are generated. The DSURVIVAL option is specified as follows:

DSURVIVAL = u1-u4;

where u1 to u4 are discrete-time survival variables.

VARIABLES WITH SPECIAL FUNCTIONS

There are several options that are used to identify variables that have
special functions. These variables can be variables from the NAMES
statement of the VARIABLE command and variables created using the
DEFINE command and the DATA transformation commands.
Following is a description of these options and their specifications.

GROUPING

The GROUPING option is used to identify the variable in the data set
that contains information on group membership when the data for all
groups are stored in a single data set. Multiple group analysis is
discussed in Chapter 14. A grouping variable must contain only integer
values. Only one grouping variable can be used. If the groups to be
analyzed are a combination of more than one variable, a single grouping
variable can be created using the DEFINE command. Following is an
example of how to specify the GROUPING option:

GROUPING IS gender (1=male 2 = female);

The information in parentheses after the grouping variable name assigns

labels to the values of the grouping variable found in the data set. In the
example above, observations with gender equal to 1 are assigned the
label male, and individuals with gender equal to 2 are assigned the label
female. These labels are used in conjunction with the MODEL
command to specify model statements specific to each group.
Observations that have a value on the grouping variable that is not
specified using the GROUPING option are not included in the analysis.

549
CHAPTER 15

In situations where there are many groups, a shorthand notation can be

used for the GROUPING option. It is specified as follows:

GROUPING = country (101-200 225 350-360);

where country is the grouping variable and 101 through 200, 225, and
350 through 360 are the values of country that will be used as groups.
The values of the variable country are used as labels in group-specific
MODEL commands.

IDVARIABLE

The IDVARIABLE option is used in conjunction with the SAVEDATA

command to provide an identifier for each observation in the data set
that is saved. This is useful for merging the data with another data set.
The IDVARIABLE option is specified as follows:

IDVARIABLE = id;

where id is a variable that contains a unique numerical identifier for each

observation. The length of this variable may not exceed 16.

FREQWEIGHT

The FREQWEIGHT option is used to identify the variable that contains

frequency (case) weight information. Frequency weights are used when
a single record in the data set represents the responses of more than one
individual. Frequency weight values must be integers. Frequency
weights do not have to sum to the total number of observations in the
analysis data set and are not rescaled in any way. Frequency weights are
available for all analysis types except TYPE=COMPLEX,
TYPE=TWOLEVEL, TYPE=THREELEVEL,
TYPE=CROSSCLASSIFIED, and TYPE=EFA. With
TYPE=RANDOM, frequency weights are available only with
ALGORITHM=INTEGRATION. Following is an example of how the
FREQWEIGHT option is specified:

FREQWEIGHT IS casewgt;

550
 TITLE, DATA, VARIABLE, And DEFINE Commands

where casewgt is the variable that contains frequency weight

information.

TSCORES

The TSCORES option is used in conjunction with TYPE=RANDOM to

identify the variables in the data set that contain information about
individually-varying times of observation for the outcome in a growth
model. Variables listed in the TSCORES statement can be used only in
AT statements in the MODEL command to define a growth model. For
TYPE=TWOLEVEL, ALGORITHM=INTEGRATION must be
specified in the ANALYSIS command for this type of analysis. The
TSCORES option is specified as follows:

TSCORES ARE a1 a2 a3 a4;

where a1, a2, a3, and a4 are observed variables in the analysis data set
that contain the individually-varying times of observation for an
outcome at four time points.

AUXILIARY

Auxiliary variables are variables that are not part of the analysis model.
The AUXILIARY option has seven uses. One is to identify a set of
variables that is not used in the analysis but is saved for use in a
subsequent analysis. A second is to identify a set of variables that will
be used as missing data correlates in addition to the analysis variables.
The last five are used with TYPE=MIXTURE. Two are used to identify
a set of variables not used in the analysis that are possible covariates in a
multinomial logistic regression for a categorical latent variable. Three
are used to identify a set of variables not used in the analysis for which
the equality of means across latent classes will be tested. Only one of
these five can be used in an analysis at a time.

In the first use of the AUXILIARY option, variables listed on the

AUXILIARY statement are saved along with the analysis variables if the
SAVEDATA command is used. These variables can be used in
graphical displays if the PLOT command is used. If these variables are
created using the DEFINE command or the DATA transformation
commands, they must be listed on the USEVARIABLES statement of

551
CHAPTER 15

the VARIABLE command in addition to being listed on the

AUXILIARY statement. The AUXILIARY option is specified as
follows:

AUXILIARY = gender race educ;

where gender, race, and educ are variables that are not used in the
analysis but that are saved in conjunction with the SAVEDATA and/or
the PLOT commands.

In the second use, the AUXILIARY option is used in conjunction with

TYPE=GENERAL with continuous dependent variables and maximum
likelihood estimation to identify a set of variables that will be used as
missing data correlates in addition to the analysis variables (Collins,
Schafer, & Kam, 2001; Graham, 2003; Asparouhov & Muthén, 2008b;
Enders, 2010). This use is not available with MODINDICES,
BOOTSTRAP, and models with a set of exploratory factor analysis
(EFA) factors in the MODEL command. The setting M in parentheses is
placed behind the variables on the AUXILIARY statement that will be
used as missing data correlates. Following is an example of how to
specify the M setting:

AUXILIARY = z1-z4 (M);

where z1, z2, z3, and z4 are variables that will be used as missing data
correlates in addition to the analysis variables.

In the third use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE to identify covariates that are important predictors of
latent classes. This is done using pseudo-class draws, that is, the
posterior-probability based multinomial logistic regression of a
categorical latent variable on a set of covariates. The setting R in
parentheses is placed behind the variables on the AUXILIARY
statement that will be used as covariates in the multinomial logistic
regression in a mixture model. Following is an example of how to
specify the R setting:

AUXILIARY = race (R) ses (R) x1-x5 (R);

552
 TITLE, DATA, VARIABLE, And DEFINE Commands

where race, ses, x1, x2, x3, x4, and x5 will be used as covariates in a
multinomial logistic regression in a mixture model.

In the fourth use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
covariates that are important predictors of latent classes using a three-
step approach (Vermunt, 2010; Asparouhov & Muthén, 2012b). It is not
allowed with ALGORITHM=INTEGRATION. The setting R3STEP in
parentheses is placed behind the variables on the AUXILIARY
statement that will be used as covariates in the multinomial logistic
regression in a mixture model. Following is an example of how to
specify the R3STEP setting:

AUXILIARY = race (R3STEP) ses (R3STEP) x1-x5 (R3STEP);

where race, ses, x1, x2, x3, x4, and x5 will be used as covariates in a
multinomial logistic regression in a mixture model.

In the fifth use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE to identify the variables for which the equality of
means across latent classes will be tested using pseudo-class-based
multiple imputations (Asparouhov, 2007). The setting E in parentheses
is placed behind the variables on the AUXILIARY statement for which
equalities of means across latent classes will be tested. Following is an
example of how to specify the E setting:

AUXILIARY = race (E) ses (E) gender (E);

where the equality of means for race, ses, and gender will be tested
across classes.

In the sixth use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
continuous variables for which the equality of means across latent
classes will be tested using a three-step approach (Asparouhov &
Muthén, 2012b). It is not allowed with
ALGORITHM=INTEGRATION. For example, this could be a distal
outcome that is influenced by a categorical latent variable. When more
than one continuous variable is given, each variable is tested separately.
The setting DU3STEP in parentheses is placed behind the variables on

553
CHAPTER 15

the AUXILIARY statement for which equalities of means across latent

classes will be tested. The variances of the variables are allowed to vary
across the classes. Following is an example of how to specify the
DU3STEP setting:

AUXILIARY = drinks (DU3STEP) depress (DU3STEP);

where the equality of means for drinks and depress will be separately
tested across classes.

In the seventh use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
continuous variables for which the equality of means across latent
classes will be tested using a three-step approach (Asparouhov &
Muthén, 2012b). It is not allowed for ALGORITHM=INTEGRATION.
For example, this could be a distal outcome that is influenced by a
categorical latent variable. When more than one continuous variable is
given, each variable is tested separately. The setting DE3STEP in
parentheses is placed behind the variables on the AUXILIARY
statement for which equalities of means across latent classes will be
tested. The variances of the variables are held equal across the classes.
Following is an example of how to specify the DE3STEP setting:

AUXILIARY = drinks (DE3STEP) depress (DE3STEP);

where the equality of means for drinks and depress will be separately
tested across classes.

Following is an example of how to specify more than one setting in the

same AUXILIARY statement:

AUXILIARY = gender race (E) educ ses (E) x1-x5 (E);

where all of the variables on the AUXILIARY statement will be saved if

the SAVEDATA command is used, will be available for plots if the
PLOT command is used, and tests of equality of means across latent
classes will be carried out for the variables race, ses, x1, x2, x3, x4, and
x5.

554
 TITLE, DATA, VARIABLE, And DEFINE Commands

The AUXILIARY option has an alternative specification for the M, R,

R3STEP, E, DU3STEP, and DE3STEP settings that is convenient when
there are several variables that cannot be specified using the list
function. These are AUXILIARY=(M), AUXILIARY=(R),
AUXILIARY=(R3STEP), AUXILIARY(E), AUXILIARY(DU3STEP),
and AUXILIARY(DE3STEP). When M, R, R3STEP, E, DU3STEP, and
DE3STEP in parentheses follows the equal sign, it means that M, R,
R3STEP, E, DU3STEP, or DE3STEP apply to all of the variables that
follow. For example, the following AUXILIARY statement specifies
that the variables x1, x3, x5, x7, and x9 will be used as missing data
correlates in addition to the analysis variables:

AUXILIARY = (M) x1 x3 x5 x7 x9;

CONSTRAINT

The CONSTRAINT option is used to identify the variables that can be

used in the MODEL CONSTRAINT command. These can be not only
variables used in the MODEL command but also other variables. All
variables on the CONSTRAINT list are treated as continuous variables
in the analysis. Only variables used by the following options cannot be
included: GROUPING, PATTERN, COHORT, COPATTERN,
CLUSTER, STRATIFICATION, and AUXILIARY. Variables that are
part of these options can be used in DEFINE to create new variables that
can be used in the CONSTRAINT statement. The CONSTRAINT
option is not available for TYPE=RANDOM, TYPE=TWOLEVEL,
TYPE=THREELEVEL, TYPE=CROSSCLASSIFIED,
TYPE=COMPLEX, and for estimators other than ML, MLR, and MLF.
The CONSTRAINT option is specified as follows:

CONSTRAINT = y1 u1;

where y1 and u1 are variables that can be used in the MODEL

CONSTRAINT command.

PATTERN

The PATTERN option is used when data are missing by design. The
typical use is in situations when, because of the design of the study, all
variables are not measured on all individuals in the analysis. This can

555
CHAPTER 15

occur, for example, when individuals receive different forms of a

measurement instrument that contain different sets of items. Following
is an example of how the PATTERN option is specified:

PATTERN IS design (1= y1 y3 y5 2= y2 y3 y4 3= y1 y4 y5);

where design is a variable in the data set that has integer values of 1, 2,
and 3. The variable names listed after each number and the equal sign
are variables used in the analysis which should have no missing values
for observations with that value on the pattern variable. For example,
observations with the value of one on the variable design should have
information for variables y1, y3, and y5 and have missing values for y2
and y4. Observations with the value of three on the variable design
should have information for variables y1, y4, and y5 and have missing
values for variables y2 and y3. The pattern variable must contain only
integer values. Observations that have a value for the pattern variable
that is not specified using the PATTERN option are not included in the
analysis.

COMPLEX SURVEY DATA

There are several options that are used for complex survey data. These
include options for stratification, clustering, unequal probabilities of
selection (sampling weights), and subpopulation analysis. The variables
used with these options can be variables from the NAMES statement of
the VARIABLE command and variables created using the DEFINE
command and the DATA transformation commands. The exception is
that variables used with the SUBPOPULATION option must be
variables from the NAMES statement of the VARIABLE command.
Following is a description of these options and their specifications.

STRATIFICATION

The STRATIFICATION option is used with TYPE=COMPLEX to

identify the variable in the data set that contains information about the
subpopulations from which independent probability samples are drawn.

556
 TITLE, DATA, VARIABLE, And DEFINE Commands

Following is an example of how the STRATIFICATION option is used:

STRATIFICATION IS region;

where region is the variable that contains stratification information.

CLUSTER

The CLUSTER option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, TYPE=CROSSCLASSIFIED, and
TYPE=COMPLEX to identify the variables in the data set that contain
clustering information. One cluster variable is used for
TYPE=TWOLEVEL and TYPE=COMPLEX. Two cluster variables are
used for TYPE=THREELEVEL, TYPE=CROSSCLASSIFIED, and
TYPE=COMPLEX TWOLEVEL. Three cluster variables are used for
TYPE=COMPLEX THREELEVEL.

Following is an example of how the CLUSTER option is used with

TYPE=TWOLEVEL and TYPE=COMPLEX:

CLUSTER IS school;

where school is the variable that contains clustering information.

Following is an example of how the CLUSTER option is used with

TYPE=THREELEVEL:

CLUSTER IS school class;

where school and class are the variables that contain clustering
information. The cluster variable for the highest level must come first,
that is, classrooms are nested in schools.

Following is an example of how the CLUSTER option is used with

TYPE=CROSSCLASSIFIED:

CLUSTER = neighbor school;

where neighbor and school are the variables that contain clustering
information. Students are nested in schools crossed with neighborhoods.

557
CHAPTER 15

TYPE=COMPLEX TWOLEVEL can be used with either two cluster

variables, one stratification and two cluster variables, or one
stratification and one cluster variable. Following is an example of using
two cluster variables:

CLUSTER = school class;

where school and class are the variables that contain clustering
information. The clusters for TYPE=TWOLEVEL are classroom. The
standard error and chi-square computations for TYPE=COMPLEX are
based on school.

Following is an example with stratification and clustering:

STRATIFICATION = region;
CLUSTER = school;

where the clusters for TYPE=TWOLEVEL are schools and the standard
error and chi-square computations for TYPE=COMPLEX are based on
region.

TYPE=COMPLEX THREELEVEL can be used with either three cluster

variables, one stratification and three cluster variables, or one
stratification and two cluster variables. Following is an example of how
the CLUSTER option is used with TYPE=COMPLEX THREELEVEL:

CLUSTER IS psu school class;

where psu, school, and class are the variables that contain clustering
information. The cluster variable for the highest level must come first,
that is, classrooms are nested in schools and schools are nested in psu’s.
The clusters for TYPE=THREELEVEL are classroom and school. The
standard error and chi-square computations for TYPE=COMPLEX are
based on psu’s.

WEIGHT

The WEIGHT option is used to identify the variable that contains

sampling weight information for non-clustered data and level 1 for
TYPE=TWOLEVEL and TYPE=THREELEVEL. Sampling weights are

558
 TITLE, DATA, VARIABLE, And DEFINE Commands

not available for TYPE=CROSSCLASSIFIED. Sampling weights are

used when data have been collected with unequal probabilities of being
selected. Sampling weight values must be non-negative real numbers. If
the sum of the sampling weights is not equal to the total number of
observations in the analysis data set, the weights are rescaled so that they
sum to the total number of observations. Sampling weights are available
for all analysis types. Sampling weights are available for
ESTIMATOR=MLR, MLM, MLMV, WLS, WLSM, WLSMV, and
ULS and for ESTIMATOR=ML when the BOOTSTRAP option of the
ANALYSIS command is used. There are two exceptions. They are not
available for WLS when all dependent variables are continuous and are
not available for MLM or MLMV for EFA. Following is an example of
how the WEIGHT option is used to identify a sampling weight variable:

WEIGHT IS sampwgt;

where sampwgt is the variable that contains sampling weight

information.

WTSCALE

The WTSCALE option is used with TYPE=TWOLEVEL to adjust the

weight variable named using the WEIGHT option. With
TYPE=TWOLEVEL, the WEIGHT option is used to identify the
variable that contains within-level sampling weight information.

The WTSCALE option has the following settings: UNSCALED,

CLUSTER, and ECLUSTER. CLUSTER is the default.

The UNSCALED setting uses the within weights from the data set with
no adjustment. The CLUSTER setting scales the within weights from
the data set so that they sum to the sample size in each cluster. The
ECLUSTER setting scales the within weights from the data so that they
sum to the effective sample size (Pothoff, Woodbury, & Manton, 1992).

559
CHAPTER 15

The WTSCALE option is specified as follows:

WTSCALE = ECLUSTER;

where scaling the within weights so that they sum to the effective sample
size is chosen.

BWEIGHT

The BWEIGHT option is used with TYPE=TWOLEVEL to identify the

variable that contains between-level sampling weight information.

The BWEIGHT option is specified as follows:

BWEIGHT = bweight;

where bweight is the variable that contains between-level sampling

weight information.

B2WEIGHT

The B2WEIGHT option is used with TYPE=THREELEVEL to identify

the variable that contains level 2 sampling weight information.
Sampling weights are not available for TYPE=CROSSCLASSIFIED.

The B2WEIGHT option is specified as follows:

B2WEIGHT = b2weight;

where b2weight is the variable that contains level 2 sampling weight

information.

B3WEIGHT

The B3WEIGHT option is used with TYPE=THREELEVEL to identify

the variable that contains level 3 sampling weight information.
Sampling weights are not available for TYPE=CROSSCLASSIFIED.

560
 TITLE, DATA, VARIABLE, And DEFINE Commands

The B3WEIGHT option is specified as follows:

B3WEIGHT = b3weight;

where b3weight is the variable that contains level 3 sampling weight

information.

BWTSCALE

The BWTSCALE option is used in with TYPE=TWOLEVEL to adjust

the between-level sampling weight variable named using the BWEIGHT
option. The BWTSCALE option is used in with TYPE=THREELEVEL
to adjust the level 2 and level 3 sampling weight variables named using
the B2WEIGHT and B3WEIGHT options.

The BWTSCALE option has the following settings: UNSCALED and

SAMPLE. SAMPLE is the default.

The UNSCALED setting uses the between weights from the data set
with no adjustment. The SAMPLE option adjusts the between weights
so that the product of the between and the within weights sums to the
total sample size.

The BWTSCALE option is specified as follows:

BWTSCALE = UNSCALED;

where no adjustment is made to the between weights.

REPWEIGHTS

Replicate weights summarize information about a complex sampling

design (Korn & Graubard, 1999; Lohr, 1999). They are used to properly
estimate standard errors of parameter estimates. Replicate weights are
available for TYPE=COMPLEX for continuous variables using
maximum likelihood estimation and for binary, ordered categorical, and
censored variables using weighted least squares estimation. The
SUBPOPULATION option is not available with replicate weights.
Replicate weights can be used or generated. When they are generated,

561
CHAPTER 15

they can be used in the analysis and/or saved (Asparouhov & Muthén,
2009b).

The REPWEIGHTS option is used to identify the replicate weight

variables. The STRATIFICATION and CLUSTER options may not be
used in conjunction with the REPWEIGHTS option. The WEIGHT
option must be used. Following is an example of how to specify the
REPWEIGHTS option:

REPWEIGHTS = rweight1-rweight80;

where rweight1 through rweight80 are replicate weight variables.

USING REPLICATE WEIGHTS

When existing replicate weights are used, the REPWEIGHTS option of

the VARIABLE command is used in conjunction with the WEIGHT
option of the VARIABLE command and the REPSE option of the
ANALYSIS command. The sampling weights are used in the estimation
of parameter estimates. The replicate weights are used in the estimation
of standard errors of parameter estimates. The REPSE option specifies
the resampling method that is used in the computation of the standard
errors.

GENERATING REPLICATE WEIGHTS

When replicate weights are generated, the REPSE option of the

ANALYSIS command and the WEIGHT option of the VARIABLE
command along with the CLUSTER and/or the STRATIFICATION
options of the VARIABLE command are used.

SUBPOPULATION

The SUBPOPULATION option is used with TYPE=COMPLEX to

select observations for an analysis when a subpopulation (domain) is
analyzed. If the SUBPOPULATION option is used, the
USEOBSERVATIONS option cannot be used. When the
SUBPOPULATION option is used, all observations are included in the
analysis although observations not in the subpopulation are assigned

562
 TITLE, DATA, VARIABLE, And DEFINE Commands

weights of zero (see Korn & Graubard, 1999, pp. 207-211). The
SUBPOPULATION option is not available for multiple group analysis.

The SUBPOPULATION option identifies those observations for

analysis that satisfy the conditional statement specified after the equal
sign and assigns them non-zero weights. For example, the following
statement identifies observations with the variable gender equal to 2:

SUBPOPULATION = gender EQ 2;

Only variables from the NAMES statement of the VARIABLE command

can be used in the conditional statement of the SUBPOPULATION
option. Logical operators, not arithmetic operators, must be used in the
conditional statement. Following are the logical operators that can be
used in the conditional statement of the SUBPOPULATION option:

AND logical and

OR logical or
NOT logical not
EQ equal ==
NE not equal /=
GE greater than or equal to >=
LE less than or equal to <=
GT greater than >
LT less than <

As shown above, some of the logical operators can be referred to in two

different ways. For example, equal can be referred to as EQ or ==.

FINITE

For TYPE=COMPLEX, the FINITE option is used to identify the

variable that contains the finite population correction factor, the
sampling fraction, or the population size for each stratum. If the
sampling fraction or the population size for each stratum is provided,
these are used to compute the finite population correction factor. The
finite population correction factor is used to adjust standard errors when
clusters have been sampled without replacement (WOR) from strata in a
finite population. The finite population correction factor is equal to one
minus the sampling fraction. The sampling fraction is equal to the

563
CHAPTER 15

number of sampled clusters in a stratum divided by the number of

clusters in the population for that stratum. The population size for each
stratum is the number of clusters in the population for that stratum. The
FINITE option is not available with replicate weights.

The FINITE option has three settings: FPC, SFRACTION, and

POPULATION. FPC is used when the finite population correction
factor is provided. SFRACTION is used when the sampling fraction is
provided. POPULATION is used when the population size for each
stratum is provided. FPC is the default. Following is an example of
how the FINITE option is used to identify a finite population correction
variable:

FINITE IS sampfrac (SFRACTION);

where sampfrac is the variable that contains the sampling fraction for
each stratum.

MIXTURE MODELS
There are three options that are used specifically for mixture models.
They are CLASSES, KNOWNCLASS, and TRAINING.

CLASSES

The CLASSES option is used to assign names to the categorical latent

variables in the model and to specify the number of latent classes in the
model for each categorical latent variable. This option is required for
TYPE=MIXTURE. Between-level categorical latent variables must be
identified as between-level variables using the BETWEEN option.

Following is an example of how the CLASSES option is used:

CLASSES = c1 (2) c2 (2) c3 (3);

where c1, c2, and c3 are the names of the three categorical latent
variables in the model. The numbers in parentheses specify the number
of classes for each categorical latent variable in the model. The
categorical latent variable c1 has two classes, c2 has two classes, and c3
has three classes.

564
 TITLE, DATA, VARIABLE, And DEFINE Commands

When there is more than one categorical latent variable in the model,
there are rules related to the order of the categorical latent variables.
The order is taken from the order of the categorical latent variables in
the CLASSES statement. Because of the order in the CLASSES
statement above, c1 is not allowed to be regressed on c2 in the model. It
is only possible to regress c2 on c1 and c3 on c2 or c1. This order
restriction does not apply to PARAMETERIZATION=LOGLINEAR.

KNOWNCLASS

The KNOWNCLASS option is used for multiple group analysis with

TYPE=MIXTURE. The KNOWNCLASS option is used to identify the
categorical latent variable for which latent class membership is known
and equal to observed groups in the sample. Only one KNOWNCLASS
variable can be used. Following is an example of how to specify the
KNOWNCLASS option:

KNOWNCLASS = c1 (gender = 0 1);

where c1 is a categorical latent variable named and defined using the

CLASSES option of the VARIABLE command. The information in
parentheses following the categorical latent variable name defines the
known classes using an observed variable. In this example, the observed
variable gender is used to define the known classes. The first class
consists of individuals with the value 0 on the variable gender. The
second class consists of individuals with the value 1 on the variable
gender.

Following is an example with many known classes that uses the list
function:

KNOWNCLASS = c (country = 101-110 112 113-115);

where c is a categorical latent variable named and defined using the

CLASSES option of the VARIABLE command. The information in
parentheses following the categorical latent variable name defines the
known classes using an observed variable. In this example, the observed
variable country is used to define the known classes. The first class
consists of individuals with the value 101 on the variable country. The

565
CHAPTER 15

last class consists of individuals with the value 115 on the variable
country. There are a total of 14 classes.

TRAINING

The TRAINING option is used to identify the variables that contain

information about latent class membership and specify whether the
information is about class membership, probability of class membership,
or priors of class membership. The TRAINING option has three
settings: MEMBERSHIP, PROBABILITIES, and PRIORS.
MEMBERSHIP is the default. Training variables can be variables from
the NAMES statement of the VARIABLE command and variables
created using the DEFINE command and the DATA transformation
commands. This option is available only for models with a single
categorical latent variable.

Following is an example of how the TRAINING option is used for an

example with three latent classes where the training variables contain
information about class membership:

TRAINING = t1 t2 t3;

where t1, t2, and t3 are variables that contain information about latent
class membership. The variable t1 provides information about
membership in class 1, t2 provides information about membership in
class 2, and t3 provides information about membership in class 3. An
individual is allowed to be in any class for which they have a value of
one on a training variable. An individual who is known to be in class 2
would have values of 0, 1, and 0 on t1, t2, and t3, respectively. An
individual with unknown class membership would have the value of 1 on
t1, t2, and t3. An alternative specification is:

TRAINING = t1 t2 t3 (MEMBERSHIP);

Fractional values can be used to provide information about the

probability of class membership when class membership is not estimated
but is fixed at fractional values for each individual. For example, an
individual who has a probability of .9 for being in class 1, .05 for being
in class 2, and .05 for being in class 3 would have t1=.9, t2=.05, and
t3=.05. Fractional training data must sum to one for each individual.

566
 TITLE, DATA, VARIABLE, And DEFINE Commands

Following is an example of how the TRAINING option is used for an

example with three latent classes where the training variables contain
information about probabilities of class membership:

TRAINING = t1 t2 t3 (PROBABILITIES);

where t1, t2, and t3 are variables that contain information about the
probability of latent class membership. The variable t1 provides
information about the probability of membership in class 1, t2 provides
information about the probability of membership in class 2, and t3
provides information about the probability of membership in class 3.

Priors can be used when individual class membership is not known but
when information is available on the probability of an individual being
in a certain class. For example, an individual who has a probability of .9
for being in class 1, .05 for being in class 2, and .05 for being in class 3
would have t1=.9, t2=.05, and t3=.05. Prior values must sum to one for
each individual.

Following is an example of how the PRIORS option is used for an

example with three latent classes where the training variables contain
information about priors of class membership:

TRAINING = t1 t2 t3 (PRIORS);

where t1, t2, and t3 are variables that contain information about the
probability of being in a certain class. The variable t1 provides
information about the probability of membership in class 1, t2 provides
information about the probability of membership in class 2, and t3
provides information about the probability of membership in class 3.

567
CHAPTER 15

MULTILEVEL MODELS
There are two options specific to multilevel models. They are WITHIN
and BETWEEN. Variables identified using the WITHIN and
BETWEEN options can be variables from the NAMES statement of the
VARIABLE command and variables created using the DEFINE
command and the DATA transformation commands.

WITHIN

The WITHIN option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to identify the
variables in the data set that are measured on the individual level and to
specify the levels on which they are modeled. All variables on the
WITHIN list must be measured on the individual level. An individual-
level variable can be modeled on all or some levels.

For TYPE=TWOLEVEL, an individual-level variable can be modeled

on only the within level or on both the within and between levels. If a
variable measured on the individual level is mentioned on the WITHIN
list, it is modeled on only the within level. It has no variance in the
between part of the model. If it is not mentioned on the WITHIN list, it
is modeled on both the within and between levels. The WITHIN option
is specified as follows:

WITHIN = y1 y2 x1;

where y1, y2, and x1 are variables measured on the individual level and
modeled on only the within level.

For TYPE=THREELEVEL, an individual-level variable can be modeled

on only level 1, on levels 1 and 2, levels 1 and 3, or on all levels.
Consider a model where students are nested in classrooms and
classrooms are nested in schools. Level 1 is student; level 2 is
classroom; and level 3 is school. If a variable measured on the
individual level is mentioned on the WITHIN list without a label, it is
modeled on only level 1. It has no variance on levels 2 and 3. If it is
mentioned on the WITHIN list with a level 2 cluster label, it is modeled
on levels 1 and 2. It has no variance on level 3. If it is mentioned on the
WITHIN list with a level 3 cluster label, it is modeled on levels 1 and 3.

568
 TITLE, DATA, VARIABLE, And DEFINE Commands

It has no variance on level 2. If it is not mentioned on the WITHIN list,

it is modeled on all levels.

Following is an example of how to specify the WITHIN option for

TYPE=THREELEVEL:

WITHIN = y1-y3 (class) y4-y6 (school) y7-y9;

In the example, y1, y2, and y3 are variables measured on the individual
level and modeled on only level 1, student. Variables modeled on only
level 1 must precede variables modeled on the other levels. Y4, y5, and
y6 are variables measured on the individual level and modeled on level
1, student, and level 2, class, where class is the level 2 cluster variable.
Y7, y8, and y9 are variables measured on the individual level and
modeled on level 1, student, and level 3, school, where school is the
level 3 cluster variable.

An alternative specification of the WITHIN option above reverses the

order of the level 2 and level 3 variables:

WITHIN = y1-y3 (school) y7-y9 (class) y4-y6;

Variables modeled on only level 1 must precede variables modeled on

the other levels. Another alternative specification is:

WITHIN = y1-y3;
WITHIN = (class) y4-y6;
WITHIN = (school) y7-y9;

In this specification, the WITHIN statement for variables modeled on

only level 1 must precede the other WITHIN statements. The order of
the other WITHIN statements does not matter.

For TYPE=CROSSCLASSIFIED, an individual-level variable can be

modeled on only level 1, on levels 1 and 2a, levels 1 and 2b, or on all
levels. Consider a model where students are nested in schools crossed
with neighborhoods. Level 1 is student; level 2a is school; and level 2b
is neighborhood. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2a and 2b. If it is mentioned on the

569
CHAPTER 15

WITHIN list with a level 2a cluster label, it is modeled on levels 1 and

2a. It has no variance on level 2b. If it is mentioned on the WITHIN list
with a level 2b cluster label, it is modeled on levels 1 and 2b. It has no
variance on level 2a. If it is not mentioned on the WITHIN list, it is
modeled on all levels.

Following is an example of how to specify the WITHIN option for

TYPE=CROSSCLASSIFIED:

WITHIN = y1-y3 (school) y4-y6 (neighbor) y7-y9;

In the example, y1, y2, and y3 are variables measured on the individual
level and modeled on only level 1, student. Variables modeled on only
level 1 must precede variables modeled on the other levels. Y4, y5, and
y6 are variables measured on the individual level and modeled on level
1, student, and level 2a, school, where school is the level 2a cluster
variable. Y7, y8, and y9 are variables measured on the individual level
and modeled on level 1, student, and level 2b, neighborhood, where
neighborhood is the level 2b cluster variable.

BETWEEN

The BETWEEN option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to identify the
variables in the data set that are measured on the cluster level(s) and to
specify the level(s) on which they are modeled. All variables on the
BETWEEN list must be measured on a cluster level. A cluster-level
variable can be modeled on all or some cluster levels.

For TYPE=TWOLEVEL, a cluster-level variable can be modeled on

only the between level. The BETWEEN option is specified as follows:

BETWEEN = z1 z2 x1;

where z1, z2, and x1 are variables measured on the cluster level and
modeled on the between level. The BETWEEN option is also used to
identify between-level categorical latent variables with
TYPE=TWOLEVEL MIXTURE.

570
 TITLE, DATA, VARIABLE, And DEFINE Commands

For TYPE=THREELEVEL, a variable measured on level 2 can be

modeled on only level 2 or on levels 2 and 3. A variable measured on
level 3 can be modeled on only level 3. Consider a model where
students are nested in classrooms and classrooms are nested in schools.
Level 1 is student; level 2 is classroom; and level 3 is school. If a
variable measured on level 2 is mentioned on the BETWEEN list
without a label, it is modeled on levels 2 and 3. If a variable measured
on level 2 is mentioned on the BETWEEN list with a level 2 cluster
label, it is modeled on only level 2. It has no variance on level 3. A
variable measured on level 3 must be mentioned on the BETWEEN list
with a level 3 cluster label. Following is an example of how to specify
the BETWEEN option for TYPE=THREELEVEL:

BETWEEN = y1-y3 (class) y4-y6 (school) y7-y9;

In this example, y1, y2, and y3 are cluster-level variables measured on

level 2, class, and modeled on both levels 2 and 3. Variables modeled
on both levels 2 and 3 must precede variables modeled on only level 2 or
level 3. Y4, y5, and y6 are cluster-level variables measured on level 2,
class, and modeled on level 2. Y7, y8, and y9 are cluster-level variables
measured on level 3 and modeled on level 3.

An alternative specification of the BETWEEN option above reverses the

order of the level 2 and level 3 variables:

BETWEEN = y1-y3 (school) y7-y9 (class) y4-y6;

Variables modeled on both levels 2 and 3 must precede variables

modeled on only level 2 or level 3. Another alternative specification is:

BETWEEN = y1-y3;
BETWEEN = (class) y4-y6;
BETWEEN = (school) y7-y9;

In this specification, the BETWEEN statement for variables modeled on

both levels 2 and 3 must precede the other BETWEEN statements. The
order of the other BETWEEN statements does not matter.

For TYPE=CROSSCLASSIFIED, a variable measured on level 2a must

be mentioned on the BETWEEN list with a level 2a cluster label. It can

571
CHAPTER 15

be modeled on only level 2a. A variable measured on level 2b must be

mentioned on the BETWEEN list with a level 2b cluster label. It can be
modeled on only level 2b. Consider a model where students are nested
in schools crossed with neighborhoods. Level 1 is student; level 2a is
school; and level 2b is neighborhood. Following is an example of how
to specify the BETWEEN option for TYPE=CROSSCLASSIFIED:

BETWEEN = (school) y1-y3 (neighbor) y4-y6;

In this example, y1, y2, and y3 are cluster-level variables measured on

level 2a, school, and modeled on only level 2a. Y4, y5, and y6 are
cluster-level variables measured on level 2b, neighborhood, and modeled
on only level 2b.

CONTINUOUS-TIME SURVIVAL MODELS

There are two options specific to continuous-time survival models. They
are SURVIVAL and TIMECENSORED. Variables identified using the
SURVIVAL and TIMECENSORED options can be variables from the
NAMES statement of the VARIABLE command and variables created
using the DEFINE command and the DATA transformation commands.

SURVIVAL

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option must be used in conjunction with the
TIMECENSORED option. The SURVIVAL option is specified as
follows:

SURVIVAL = t (4*5 10);

where t is the variable that contains time to event information. The

numbers in parentheses specify that four time intervals of length five and
one time interval of length ten are used in the analysis.

572
 TITLE, DATA, VARIABLE, And DEFINE Commands

If no time intervals are specified, a constant baseline hazard is used.

Following is an example of how this is specified:

SURVIVAL = t;

The keyword ALL can be used if the time intervals are taken from the
data. Following is an example of how this is specified:

SURVIVAL = t (ALL);

It is not recommended to use the ALL keyword when the

BASEHAZARD option of the ANALYSIS command is ON because it
results in a large number of baseline hazard parameters.

In continuous-time survival modeling, there are as many baseline hazard

parameters are there are time intervals plus one. When the
BASEHAZARD option of the ANALYSIS command is ON, these
parameters can be referred to in the MODEL command by adding to the
name of the time-to-event variable the number sign (#) followed by a
number. For example, for a time-to-event variable t with 5 time
intervals, the six baseline hazard parameters are referred to as t#1, t#2,
t#3, t#4, t#5, and t#6. With TYPE=MIXTURE and
BASEHAZARD=OFF, the baseline hazard parameters are held equal
across classes as the default in line with Larsen (2004). This equality
constraint can be relaxed using the UNEQUAL setting of the
BASEHAZARD=OFF option. In addition to the baseline hazard
parameters, the time-to-event variable has a mean or an intercept
depending on whether the model is unconditional or conditional. The
mean or intercept is referred to by using a bracket statement, for
example,

[t];

where t is the time-to-event variable.

TIMECENSORED

The TIMECENSORED option is used in conjunction with the

SURVIVAL option to identify the variables that contain information
about right censoring, for example, when an individual leaves a study or

573
CHAPTER 15

when an individual has not experienced the event before the study ends.
There must be the same number and order of variables in the
TIMECENSORED option as there are in the SURVIVAL option. The
variables that contain information about right censoring must be coded
so that zero is not censored and one is right censored. If they are not,
this can be specified as part of the TIMECENSORED option. The
TIMECENSORED option is specified as follows when the variable is
coded zero for not censored and one for right censored:

TIMECENSORED = tc;

The TIMECENSORED option is specified as follows when the variable

is not coded zero for not censored and one for right censored:

TIMECENSORED = tc (1 = NOT 999 = RIGHT);

The value one is automatically recoded to zero and the value 999 is
automatically recoded to one.

THE DEFINE COMMAND

The DEFINE command is used to transform existing variables and to
create new variables. It includes several options for transforming
variables including a do loop for repeating statements. The operations
available in DEFINE can be executed for all observations or selectively
using conditional statements. Transformations of existing variables do
not affect the original data but only the data used for the analysis. If
analysis data are saved using the SAVEDATA command, the
transformed values rather than the original values are saved.

The statements in the DEFINE command are executed one observation

at a time in the order specified except for the CLUSTER_MEAN,
CENTER, and STANDARDIZE options. These options are executed in
the order just mentioned after all transformations specified in the
DEFINE command and the DATA transformation commands are
executed. Any variable listed in the NAMES option of the VARIABLE
command or created in the DEFINE command can be transformed or
used to create new variables. New variables created in the DEFINE

574
 TITLE, DATA, VARIABLE, And DEFINE Commands

command that will be used in an analysis must be listed on the

USEVARIABLES list after the original variables.

Following are examples of the types of transformations available:

DEFINE:

variable = mathematical expression;

IF (conditional statement) THEN transformation

statements;

_MISSING
variable = MEAN (list of variables);
variable = SUM (list of variables);
CUT variable or list of variables (cutpoints);
variable = CLUSTER_MEAN (variable);
CENTER variable or list of variables (GRANDMEAN);
CENTER variable or list of variables (GROUPMEAN);
CENTER variable or list of variables (GROUPMEAN
label);
STANDARDIZE variable or list of variables;
DO (#, #) expression;

DEFINE is not a required command.

LOGICAL OPERATORS, ARITHMETIC

OPERATORS, AND FUNCTIONS
The following logical operators can be used in DEFINE:

AND logical and

OR logical or
NOT logical not
EQ == equal
NE /= not equal
GE >= greater than or equal to
LE <= less than or equal to
GT > greater than
LT < less than

575
CHAPTER 15

As shown above, some of the logical operators can be referred to in two

different ways. For example, equal can be referred to as EQ or = =.

The following arithmetic operations can be used in DEFINE:

+ addition y + x;
- subtraction y - x;
* multiplication y * x;
/ division y / x;
** exponentiation y**2;
% remainder remainder of y/x;

The following functions can be used in DEFINE:

LOG base e log LOG (y);

LOG10 base 10 log LOG10 (y);
EXP exponential EXP (y);
SQRT square root SQRT (y);
ABS absolute value ABS(y);
SIN sine SIN (y);
COS cosine COS (y);
TAN tangent TAN(y);
ASIN arcsine ASIN (y);
ACOS arccosine ACOS (y);
ATAN arctangent ATAN (y);
PHI standard normal PHI (y); PHI (#);
distribution function

NON-CONDITIONAL STATEMENTS
When a non-conditional statement is used to transform existing variables
or create new variables, the variable on the left-hand side of the equal
sign is assigned the value of the expression on the right-hand side of the
equal sign, for example,

y = y/100;

transforms the original variable y by dividing it by 100. Non-conditional

statements can be used to create new variables, for example,

576
 TITLE, DATA, VARIABLE, And DEFINE Commands

abuse = item1 + item2 + item8 + item9;

An individual with a missing value on any variable used on the right-

hand side of the equal sign is assigned a missing value on the variable on
the left-hand side of the equal sign.

CONDITIONAL STATEMENTS
Conditional statements can also be used to transform existing variables
and to create new variables. Conditional statements take the following
form:

IF (gender EQ 1 AND ses EQ 1) THEN group = 1;

IF (gender EQ 1 AND ses EQ 2) THEN group = 2;
IF (gender EQ 1 AND ses EQ 3) THEN group = 3;
IF (gender EQ 2 AND ses EQ 1) THEN group = 4;
IF (gender EQ 2 AND ses EQ 2) THEN group = 5;
IF (gender EQ 2 AND ses EQ 3) THEN group = 6;

An individual with a missing value on any variable used in the

conditional statement is assigned a missing value on the new variable. If
no value is specified for a condition, individuals with that condition are
assigned a missing value.

The _MISSING keyword can be used in DEFINE to refer to missing

values. The _MISSING keyword can be used to assign a missing value
to a variable, for example,

IF (y EQ 0) THEN u = _MISSING;

It can also be used as part of the condition, for example,

IF (y EQ _MISSING) THEN u = 1;

OPTIONS FOR DATA TRANSFORMATION

The DEFINE command has six options for data transformation. The
first option creates a variable that is the average of a set of variables.
The second option creates a variable that is the sum of a set of variables.

577
CHAPTER 15

The third option categorizes one or several variables using the same set
of cutpoints. The fourth option creates a variable that is the average for
each cluster of an individual-level variable. The fifth option centers a
variable by subtracting the grand mean or group mean from each value.
The sixth option standardizes a variable to have a mean of zero and a
standard deviation of one.

MEAN

The MEAN option is used to create a variable that is the average of a set
of variables. It is specified as follows:

mean = MEAN (y1 y3 y5);

where the variable mean is the average of variables y1, y3, and y5.
Averages are based on the set of variables with no missing values. Any
observation that has a missing value on all of the variables being
averaged is assigned a missing value on the mean variable.

The list function can be used with the MEAN option as follows:

ymean = MEAN (y1-y10);

where the variable ymean is the average of variables y1 through y10.

Variables used with the MEAN option must be original variables from
the NAMES statement of the VARIABLE command or temporary
variables created using the DEFINE command. The order of the
variables for the list function is taken from the NAMES statement not
the USEVARIABLES statement.

SUM

The SUM option is used to create a variable that is the sum of a set of
variables. It is specified as follows:

sum = SUM (y1 y3 y5);

578
 TITLE, DATA, VARIABLE, And DEFINE Commands

where the variable sum is the sum of variables y1, y3, and y5. Any
observation that has a missing value on one or more of the variables
being summed is assigned a missing value on the sum variable.

The list function can be used with the SUM option as follows:

ysum = SUM (y1-y10);

where the variable ysum is the sum of variables y1 through y10.

Variables used with the SUM option must be original variables from the
NAMES statement of the VARIABLE command or temporary variables
created using the DEFINE command. The order of the variables for the
list function is taken from the NAMES statement not the
USEVARIABLES statement.

CUT

The CUT option categorizes a variable or list of variables using the same
set of cutpoints. More than one CUT statement can be included in the
DEFINE command. Following is an example of how the CUT option is
used:

CUT y1 y5-y7 (30 40);

This statement results in the variables y1, y5, y6, and y7 having three
categories: less than or equal to 30, greater than 30 and less than or
equal to 40, and greater than 40, with values of 0, 1, and 2, respectively.
Any observation that has a missing value on a variable that is being cut
is assigned a missing value on the cut variable.

Variables used with the CUT option must be original variables from the
NAMES statement of the VARIABLE command or temporary variables
created using the DEFINE command. The order of the variables for the
list function is taken from the NAMES statement not the
USEVARIABLES statement.

579
CHAPTER 15

CLUSTER_MEAN

The CLUSTER_MEAN option is used with TYPE=TWOLEVEL and

TYPE=COMPLEX along with the CLUSTER option to create a variable
that is the average of the values of an individual-level variable for each
cluster. In multiple group analysis, each group’s means are used for
creating cluster means in that group. It is specified as follows:

clusmean = CLUSTER_MEAN (x);

where the variable clusmean is the average of the values of x for each
cluster. Averages are based on the set of non-missing values for the
observations in each cluster. Any cluster for which all observations have
missing values is assigned a missing value on the cluster mean variable.

Any transformations specified in the DEFINE command or the DATA

transformation commands are done before cluster means are computed.
To be used with the CLUSTER_MEAN option, any new variables
created using the DEFINE command must be placed on the
USEVARIABLES list after the original variables. When a variable on
the CLUSTER_MEAN list is used in other transformations except
CENTER and STANDARDIZE, the original values of the variable are
used. Variables created using the CLUSTER_MEAN option cannot be
used in subsequent DEFINE statements except for the CENTER and
STANDARDIZED options.

CENTER

The CENTER option is used to center continuous observed variables by

subtracting the grand mean or group mean from each variable. In
multiple group analysis, each group’s means are used for centering in
that group. Grand-mean centering is available for all analyses. Group-
mean centering is available with TYPE=TWOLEVEL,
TYPE=THREELEVEL, TYPE=CROSSCLASSIFIED, and
TYPE=COMPLEX in conjunction with the CLUSTER option. The
CENTER option has two settings: GRANDMEAN and GROUPMEAN.

Any transformations specified in the DEFINE command or the DATA

transformation commands are done before centering is done. To be used
with the CENTER option, any new variables created using the DEFINE

580
 TITLE, DATA, VARIABLE, And DEFINE Commands

command must be placed on the USEVARIABLES list after the original

variables. The order of variables for the list function is taken from the
USEVARIABLES list. If there is no USEVARIABLES list, the order is
taken from the NAMES list. When a variable on the CENTER list is
used in other transformations except STANDARDIZE, the original
values of the variable are used. Variables created using the CENTER
option cannot be used in subsequent DEFINE statements except for the
STANDARDIZED option. Any observation that has a missing value on
the variable to be centered is assigned a missing value on the centered
variable.

GRANDMEAN

Grand-mean centering subtracts the overall mean from a variable.

Grand-mean centering is available for all continuous observed variables
used in an analysis. Following is an example of how to specify grand-
mean centering for TYPE=TWOLEVEL and TYPE=COMPLEX:

CENTER x1-x4 (GRANDMEAN);

where x1, x2, x3, and x4 are the variables to be centered using the
overall means for these variables.

For TYPE=THREELEVEL, there are three types of grand-mean

centering. Consider a model where students are nested in classrooms
and classrooms are nested in schools. Level 1 is student; level 2 is
classroom; and level 3 is school. For variables modeled on level 1 and
higher, grand-mean centering subtracts the overall mean from a variable.
For variables modeled on level 2 and higher, grand-mean centering
subtracts the cluster 1, classroom, mean from a variable. For variables
modeled on level 3, grand-mean centering subtracts the cluster 2, school,
mean from a variable.

Following is an example of how to specify grand-mean centering for

TYPE=THREELEVEL:

CENTER x1-x4 (GRANDMEAN);

where x1, x2, x3, and x4 are the variables to be centered using the
overall means for variables modeled on level 1 and higher, the cluster 1

581
CHAPTER 15

means for variables modeled on level 2 and higher, and the cluster 2
means for variables modeled on level 3 and higher.

For TYPE=CROSSCLASSIFIED, there are three types of grand-mean

centering. Consider a model where students are nested in schools
crossed with neighborhoods. Level 1 is student; level 2a is school; and
level 2b is neighborhood. For variables modeled on level 1 and higher,
grand-mean centering subtracts the overall mean from a variable. For
variables modeled on level 2a, grand-mean centering subtracts the
cluster 2a, school, mean from a variable. For variables modeled on level
2b, grand-mean centering subtracts the cluster 2b, neighborhood, mean
from a variable.

Following is an example of how to specify grand-mean centering for

TYPE=CROSSCLASSIFIED:

CENTER x1-x4 (GRANDMEAN);

where x1, x2, x3, and x4 are the variables to be centered using the
overall means for variables modeled on level 1 and higher, the cluster 2a
means for variables modeled on level 2a, and the cluster 2b means for
variables modeled on level 2b.

GROUPMEAN

For TYPE=TWOLEVEL and TYPE=COMPLEX, group-mean centering

subtracts the cluster mean from a variable. For TYPE=COMPLEX,
group-mean centering is available for all continuous observed variables
in an analysis. For TYPE=TWOLEVEL, group-mean centering is
available for all continuous observed variables used in an analysis that
are not modeled at the highest level. Following is an example of how to
specify group-mean centering:

CENTER x1-x4 (GROUPMEAN);

where x1, x2, x3, and x4 are the variables to be centered using the
cluster means for these variables.

For TYPE=THREELEVEL, group-mean centering is available for any

continuous observed variable that is not modeled at the highest level.

582
 TITLE, DATA, VARIABLE, And DEFINE Commands

Consider a model where students are nested in classrooms and

classrooms are nested in schools. Level 1 is student; level 2 is
classroom; and level 3 is school. For variables modeled on levels 1 and
2 and for variables modeled on only level 2, group-mean centering
subtracts the cluster 2, school, mean from a variable where the cluster 2
mean is the average of the cluster 1, class, means for each level 3,
school, cluster. Following is an example of how to specify cluster 2
group-mean centering:

CENTER x1-x4 (GROUPMEAN school);

where school is the cluster 2 variable and x1, x2, x3, and x4 are the
variables to be centered using cluster 2 means for these variables.

For variables modeled on only level 1, both cluster 1, classroom, and

cluster 2, school, group-mean centering are available. Following is an
example of how to specify group-mean centering using cluster 1 and
cluster 2 means:

CENTER x1-x2 (GROUPMEAN class);

CENTER x3 x4 (GROUPMEAN school);

In this example, class is the cluster 1 variable and x1 and x2 are the
variables to be centered using cluster 1 means for these variables.
School is the cluster 2 variable and x3 and x4 are the variables to be
centered using cluster 2 means for these variables. A variable cannot be
centered using both cluster 1 and cluster 2 means.

For TYPE=CROSSCLASSIFIED, group-mean centering is available for

any continuous observed variable that is not modeled at the highest
level. Consider a model where students are nested in schools crossed
with neighborhoods. Level 1 is student; level 2a is school; and level 2b
is neighborhood. For variables modeled on only level 1, group-mean
centering subtracts the cluster 2a, school, mean or the cluster 2b,
neighborhood mean from a variable. Following is an example of how to
specify cluster 2a and 2b group-mean centering:

CENTER x1-x4 (GROUPMEAN school);

CENTER x5-x8 (GROUPMEAN neighbor);

583
CHAPTER 15

In this example, school is the cluster 2a variable and x1, x2, x3, and x4
are the variables to be centered using cluster 2a means for these
variables. Neighbor is the cluster 2b variable and x5, x6, x7, and x8 are
the variables to be centered using the cluster 2b means for these
variables.

STANDARDIZE

The STANDARDIZE option is used to standardize continuous variables

by subtracting the mean from each value and dividing each value by the
standard deviation. In multiple group analysis, each group’s means are
used for standardizing in that group. Following is an example of how
the STANDARDIZE option is used:

STANDARDIZE y1 y5-y10 y14;

where the variables y1, y5 through y10, and y14 will be standardized.
Any observation that has a missing value on the variable to be
standardized is assigned a missing value on the standardized variable.

Any transformations specified in the DEFINE command or the DATA

transformation commands are done before standardization is done. To
be used with the STANDARDIZE option, any new variables created
using the DEFINE command must be placed on the USEVARIABLES
list after the original variables. The order of variables for the list
function is taken from the USEVARIABLES list. If there is no
USEVARIABLES list, the order is taken from the NAMES list. When a
variable on the STANDARDIZE list is used in other transformations, the
original values of the variable are used. Variables created using the
STANDARDIZE option cannot be used in subsequent DEFINE
statements.

DO

The DO option provides a do loop to facilitate specifying the same

transformation for a set of variables. Following is an example of how to
specify a do loop:

DO (1, 5) diff# = y# - x#;

584
 TITLE, DATA, VARIABLE, And DEFINE Commands

where the numbers in parentheses give the range of values the do loop
will use. The number sign (#) is replaced by these values during the
execution of the do loop. Following are the transformations that are
executed based on the DO option specified above:

diff1 = y1 - x1;
diff2 = y2 - x2;
diff3 = y3 - x3;
diff4 = y4 - x4;
diff5 = y5 - x5;

585
CHAPTER 15

586
 ANALYSIS Command

CHAPTER 16
ANALYSIS COMMAND

In this chapter, the ANALYSIS command is discussed. The ANALYSIS

command is used to describe the technical details of the analysis
including the type of analysis, the statistical estimator, the
parameterization of the model, and the specifics of the computational
algorithms.

THE ANALYSIS COMMAND

Following are the options for the ANALYSIS command:

ANALYSIS:

TYPE = GENERAL; GENERAL

BASIC;
RANDOM;
COMPLEX;
MIXTURE;
BASIC;
RANDOM;
COMPLEX;
TWOLEVEL;
BASIC;
RANDOM;
MIXTURE;
COMPLEX;
THREELEVEL;
BASIC;
RANDOM;
COMPLEX;
CROSSCLASSIFIED;
RANDOM;
EFA # #;
BASIC;
MIXTURE;
COMPLEX;
TWOLEVEL;
EFA # # UW* # # UB*;
EFA # # UW # # UB;

587
CHAPTER 16

ESTIMATOR = ML; depends on

MLM; analysis type
MLMV;
MLR;
MLF;
MUML;
WLS;
WLSM;
WLSMV;
ULS;
ULSMV;
GLS;
BAYES;
PARAMETERIZATION = DELTA; depends on
THETA; analysis type
LOGIT;
LOGLINEAR;
PROBABILITY;
LINK = LOGIT; LOGIT
PROBIT;
ROTATION = GEOMIN; GEOMIN
(OBLIQUE value)
GEOMIN (OBLIQUE value);
GEOMIN (ORTHOGONAL value);
QUARTIMIN; OBLIQUE
CF-VARIMAX; OBLIQUE
CF-VARIMAX (OBLIQUE);
CF-VARIMAX (ORTHOGONAL);
CF-QUARTIMAX; OBLIQUE
CF- QUARTIMAX (OBLIQUE);
CF- QUARTIMAX (ORTHOGONAL);
CF-EQUAMAX; OBLIQUE
CF- EQUAMAX (OBLIQUE);
CF- EQUAMAX (ORTHOGONAL);
CF-PARSIMAX; OBLIQUE
CF- PARSIMAX (OBLIQUE);
CF- PARSIMAX (ORTHOGONAL);
CF-FACPARSIM; OBLIQUE
CF- FACPARSIM (OBLIQUE);
CF- FACPARSIM (ORTHOGONAL);
CRAWFER; OBLIQUE 1/p
CRAWFER (OBLIQUE value);
CRAWFER (ORTHOGONAL value);
OBLIMIN; OBLIQUE 0
OBLIMIN (OBLIQUE value);

588
 ANALYSIS Command

OBLIMIN (ORTHOGONAL value);

VARIMAX;
PROMAX;
TARGET;
BI-GEOMIN; OBLIQUE
BI-GEOMIN (OBLIQUE);
BI-GEOMIN (ORTHOGONAL);
BI-CF-QUARTIMAX; OBLIQUE
BI-CF-QUARTIMAX (OBLIQUE);
BI-CF-QUARTIMAX (ORTHOGONAL);
ROWSTANDARDIZATION = CORRELATION; CORRELATION
KAISER;
COVARIANCE;
PARALLEL = number; 0
MODEL = NOMEANSTRUCTURE; means
NOCOVARIANCES; covariances
ALLFREE; equal
REPSE = BOOTSTRAP;
JACKKNIFE;
JACKKNIFE1;
JACKKNIFE2;
BRR;
FAY (#); .3
BASEHAZARD = ON; OFF
OFF;
OFF (EQUAL);
OFF (UNEQUAL);
CHOLESKY = ON; depends on
OFF; analysis type
ALGORITHM = EM; depends on
EMA; analysis type
FS;
ODLL;
INTEGRATION;
INTEGRATION = number of integration points; STANDARD
STANDARD (number of integration points) ; depends on
analysis type
GAUSSHERMITE (number of integration 15
points) ;
MONTECARLO (number of integration points); depends on
analysis type
MCSEED = random seed for Monte Carlo integration; 0
ADAPTIVE = ON; ON
OFF;

589
CHAPTER 16

INFORMATION = OBSERVED; depends on

EXPECTED; analysis type
COMBINATION;
BOOTSTRAP = number of bootstrap draws; STANDARD
number of bootstrap draws (STANDARD);
number of bootstrap draws (RESIDUAL):
LRTBOOTSTRAP = number of bootstrap draws for TECH14; depends on
analysis type
STARTS = number of initial stage starts and number of depends on
final stage optimizations; analysis type
STITERATIONS = number of initial stage iterations; 10
STCONVERGENCE = initial stage convergence criterion; 1
STSCALE = random start scale; 5
STSEED = random seed for generating random starts; 0
OPTSEED = random seed for analysis;
K-1STARTS = number of initial stage starts and number of 20 4
final stage optimizations for the k-1 class model
for TECH14;
LRTSTARTS = number of initial stage starts and number of 0 0 40 8
final stage optimizations for TECH14;
RSTARTS = number of random starts for the rotation depends on
algorithm and number of factor solutions analysis type
printed for exploratory factor analysis;
DIFFTEST = file name;
MULTIPLIER = file name;
COVERAGE = minimum covariance coverage with missing .10
data;
ADDFREQUENCY = value divided by sample size to add to cells .5
with zero frequency;
ITERATIONS = maximum number of iterations for the Quasi- 1000
Newton algorithm for continuous outcomes;
SDITERATIONS = maximum number of steepest descent 20
iterations for the Quasi-Newton algorithm for
continuous outcomes;
H1ITERATIONS = maximum number of iterations for unrestricted 2000
model with missing data;
MITERATIONS = number of iterations for the EM algorithm; 500
MCITERATIONS = number of iterations for the M step of the EM 1
algorithm for categorical latent variables;
MUITERATIONS = number of iterations for the M step of the EM 1
algorithm for censored, categorical, and count
outcomes;
RITERATIONS = maximum number of iterations in the rotation 10000
algorithm for exploratory factor analysis;

590
 ANALYSIS Command

CONVERGENCE = convergence criterion for the Quasi-Newton depends on

algorithm for continuous outcomes; analysis type
H1CONVERGENCE = convergence criterion for unrestricted model .0001
with missing data;
LOGCRITERION = likelihood convergence criterion for the EM depends on
algorithm; analysis type
RLOGCRITERION = relative likelihood convergence criterion for the depends on
EM algorithm; analysis type
MCONVERGENCE = convergence criterion for the EM algorithm; depends on
analysis type
MCCONVERGENCE = convergence criterion for the M step of the EM .000001
algorithm for categorical latent variables;
MUCONVERGENCE = convergence criterion for the M step of the EM .000001
algorithm for censored, categorical, and count
outcomes;
RCONVERGENCE = convergence criterion for the rotation algorithm .00001
for exploratory factor analysis;
MIXC = ITERATIONS; ITERATIONS
CONVERGENCE;
M step iteration termination based on number
of iterations or convergence for categorical
latent variables;
MIXU = ITERATIONS; ITERATIONS
CONVERGENCE;
M step iteration termination based on number
of iterations or convergence for censored,
categorical, and count outcomes;

LOGHIGH = max value for logit thresholds; +15

LOGLOW = min value for logit thresholds; - 15
UCELLSIZE = minimum expected cell size; .01
VARIANCE = minimum variance value; .0001
MATRIX = COVARIANCE; COVARIANCE
CORRELATION;
POINT = MEDIAN; MEDIAN
MEAN;
MODE;
CHAINS = number of MCMC chains; 2
BSEED = seed for MCMC random number generation; 0
STVALUES = UNPERTURBED; UNPERTURBED
PERTURBED;
ML;
MEDIATOR = LATENT; depends on
OBSERVED; analysis type

591
CHAPTER 16

ALGORITHM = GIBBS; GIBBS (PX1)

GIBBS (PX1);
GIBBS (PX2);
GIBBS (PX3);
GIBBS (RW);
MH;
BCONVERGENCE = MCMC convergence criterion using Gelman- .05
Rubin PSR;
BITERATIONS = maximum and minimum number of iterations 50000 0
for each MCMC chain when Gelman-Rubin
PSR is used;
FBITERATIONS = fixed number of iterations for each MCMC
chain when Gelman-Rubin PSR is not used;
THIN = k where every k-th MCMC iteration is saved; 1
DISTRIBUTION = maximum number of iterations used to 10000
compute the Bayes multivariate mode;
KOLMOGOROV = number of draws from the MCMC chains; 100
PRIOR = number of draws from the prior distribution; 1000
INTERACTIVE = file name;
PROCESSORS = # of processors # of threads; 11

The ANALYSIS command is not a required command. Default settings

are shown in the last column. If the default settings are appropriate for
the analysis, it is not necessary to specify the ANALYSIS command.

Note that commands and options can be shortened to four or more

letters. Option settings can be referred to by either the complete word or
the part of the word shown above in bold type.

TYPE
The TYPE option is used to describe the type of analysis. There are six
major analysis types in Mplus: GENERAL, MIXTURE, TWOLEVEL,
THREELEVEL, CROSSCLASSIFIED, and EFA. GENERAL is the
default.

The default is to estimate the model under missing data theory using all
available data; to include means, thresholds, and intercepts in the model;
to compute standard errors; and to compute chi-square when available.
These defaults can be overridden. The LISTWISE option of the DATA
command can be used to delete all observations from the analysis that
have one or more missing values on the set of analysis variables. For

592
 ANALYSIS Command

TYPE=GENERAL, means, thresholds, and intercepts can be excluded

from the analysis model by specifying
MODEL=NOMEANSTRUCTURE in the ANALYSIS command. The
NOSERROR option of the OUTPUT command can be used to suppress
the computation of standard errors. The NOCHISQUARE option of the
OUTPUT command can be used to suppress the computation of chi-
square. In some models, suppressing the computation of standard errors
and chi-square can greatly reduce computational time. Following is a
description of each of the four major analysis types.

GENERAL

Analyses using TYPE=GENERAL include models with relationships

among observed variables, among continuous latent variables, and
among observed variables and continuous latent variables. In these
models, the continuous latent variables represent factors and random
effects. Observed outcome variables can be continuous, censored,
binary, ordered categorical (ordinal), counts, or combinations of these
variable types. In addition, for regression analysis and path analysis for
non-mediating outcomes, observed outcome variables can be unordered
categorical (nominal). Following are models that can be estimated using
TYPE=GENERAL:

• Regression analysis
• Path analysis
• Confirmatory factor analysis
• Structural equation modeling
• Growth modeling
• Discrete-time survival analysis
• Continuous-time survival analysis

Special features available with the above models for all observed
outcome variable types are:

• Multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes

593
CHAPTER 16

• Individually-varying times of observations

• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities

Following is a list of the other TYPE settings that can be used in

conjunction with TYPE=GENERAL along with a brief description of
their functions:

• BASIC computes sample statistics and other descriptive information.

• RANDOM allows models with both random intercepts and random
slopes.
• COMPLEX computes standard errors and a chi-square test of model
fit taking into account stratification, non-independence of
observations, and/or unequal probability of selection.

Following is an example of how to specify the TYPE option for a

regression analysis with a random slope:

TYPE = GENERAL RANDOM;

or simply,

TYPE = RANDOM;

because GENERAL is the default.

MIXTURE

Analyses using TYPE=MIXTURE include models with categorical

latent variables which are also referred to as latent class or finite mixture
models.

For models with only categorical latent variables, observed outcome

variables can be continuous, censored, binary, ordered categorical
(ordinal), unordered categorical (nominal), counts, or combinations of
these variable types. Following are the models that can be estimated
using TYPE=MIXTURE with only categorical latent variables:

594
 ANALYSIS Command

• Regression mixture modeling

• Path analysis mixture modeling
• Latent class analysis
• Latent class analysis with covariates and direct effects
• Confirmatory latent class analysis
• Latent class analysis with multiple categorical latent variables
• Loglinear modeling
• Non-parametric modeling of latent variable distributions
• Finite mixture modeling
• Complier Average Causal Effect (CACE) modeling
• Latent transition analysis and hidden Markov modeling including
mixtures and covariates
• Latent class growth analysis
• Discrete-time survival mixture analysis
• Continuous-time survival mixture analysis

For models that include both continuous and categorical latent variables,
observed outcome variables can be continuous, censored, binary, ordered
categorical (ordinal), counts, or combinations of these variable types. In
addition, for regression analysis and path analysis for non-mediating
outcomes, observed outcome variables can also be unordered categorical
(nominal). Following are models that can be estimated using
TYPE=MIXTURE with both continuous and categorical latent variables:

• Latent class analysis with random effects

• Factor mixture modeling
• SEM mixture modeling
• Growth mixture modeling with latent trajectory classes
• Discrete-time survival mixture analysis
• Continuous-time survival mixture analysis

Special features available with the above models for all observed
outcome variable types are:

• Multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes

595
CHAPTER 16

• Individually-varying times of observations

• Linear and non-linear parameter constraints
• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Bootstrap standard errors and confidence intervals
• Wald chi-square test of parameter equalities
• Analysis with between-level categorical latent variables
• Test of equality of means across latent classes using posterior
probability-based multiple imputations

Following is a list of the other TYPE settings that can be used in

conjunction with TYPE=MIXTURE along with a brief description of
their functions:

• BASIC computes sample statistics and other descriptive information.

• RANDOM allows models with both random intercepts and random
slopes.
• COMPLEX computes standard errors and a chi-square test of model
fit taking into account stratification, non-independence of
observations, and/or unequal probability of selection.

TWOLEVEL

Analyses using TYPE=TWOLEVEL include models with random

intercepts and random slopes that vary across clusters in hierarchical
data. These random effects can be specified for any of the relationships
of the multilevel modeling framework. Observed outcome variables can
be continuous, censored, binary, ordered categorical (ordinal), unordered
categorical (nominal), counts, or combinations of these variable types.

Special features available for two-level models for all observed outcome
variable types are:

• Multiple group analysis

• Missing data
• Complex survey data
• Latent variable interactions and non-linear factor analysis using
maximum likelihood
• Random slopes
• Individually-varying times of observations

596
 ANALYSIS Command

• Linear and non-linear parameter constraints

• Indirect effects including specific paths
• Maximum likelihood estimation for all outcome types
• Wald chi-square test of parameter equalities

Following is a list of the other TYPE settings that can be used in

conjunction with TYPE=TWOLEVEL along with a brief description of
their functions:

• BASIC computes sample statistics and other descriptive information.

• RANDOM allows models with both random intercepts and random
slopes.
• MIXTURE allows models that have both categorical and continuous
latent variables.
• COMPLEX computes standard errors and a chi-square test of model
fit taking into account stratification, non-independence of
observations, and/or unequal probability of selection.

THREELEVEL

Analyses using TYPE=THREELEVEL include models with random

intercepts and random slopes that vary across clusters in hierarchical
data. These random effects can be specified for any of the relationships
of the multilevel modeling framework. Observed outcome variables can
be continuous, binary, or combinations of these variable types.

Special features available for three-level models for all observed

outcome variable types are:

• Multiple group analysis

• Missing data
• Complex survey data
• Random slopes
• Linear and non-linear parameter constraints
• Maximum likelihood estimation for continuous outcomes
• Wald chi-square test of parameter equalities

Following is a list of the other TYPE settings that can be used in

conjunction with TYPE=THREELEVEL along with a brief description
of their functions:

597
CHAPTER 16

• BASIC computes sample statistics and other descriptive information.

• RANDOM allows models with both random intercepts and random
slopes.
• COMPLEX computes standard errors and a chi-square test of model
fit taking into account stratification, non-independence of
observations, and/or unequal probability of selection. This is
available for continuous variables.

CROSSCLASSIFIED

Analyses using TYPE=CROSSCLASSIFIED include models with

random intercepts and random slopes that vary across clusters in
hierarchical data. These random effects can be specified for any of the
relationships of the multilevel modeling framework. Observed outcome
variables can be continuous, binary, ordered categorical (ordinal), or
combinations of these variable types.

Special features available for cross-classified models for all observed

outcome variable types are:

• Missing data
• Random slopes

Following is a list of the other TYPE settings that can be used in

conjunction with TYPE=CROSSCLASSIFIED along with a brief
description of their functions:

• RANDOM allows models with both random intercepts and random

slopes.

EFA

Analyses using TYPE=EFA include exploratory factor analysis of

continuous, censored, binary, ordered categorical (ordinal), counts, or
combinations of these variable types. See the ROTATION option of the
ANALYSIS command for a description of the rotations available for
TYPE=EFA. Modification indices are available for the residual
correlations using the MODINDICES option of the OUTPUT command.

598
 ANALYSIS Command

Special features available for EFA for all observed outcome variable
types are:

• Missing data
• Complex survey data

Following are the other TYPE settings that can be used in conjunction
with TYPE=EFA along with a brief description of their functions:

• BASIC computes sample statistics and other descriptive information.

• MIXTURE allows models that have both categorical and continuous
latent variables.
• COMPLEX computes standard errors and a chi-square test of model
fit taking into account stratification, non-independence of
observations, and/or unequal probability of selection.
• TWOLEVEL models non-independence of observations due to
clustering taking into account stratification and/or unequal
probability of selection.

Following is an example of how to specify the TYPE option for a single-

level exploratory factor analysis:

TYPE = EFA 1 3;

where the two numbers following EFA are the lower and upper limits of
the number of factors to be extracted. In the example above factor
solutions are given for one, two, and three factors.

Following is an example of how to specify the full TYPE option for a

multilevel exploratory factor analysis (Asparouhov & Muthén, 2007):

TYPE = TWOLEVEL EFA 3 4 UW* 1 2 UB*;

where the first two numbers, 3 and 4, are the lower and upper limits of
the number of factors to be extracted on the within level, UW* specifies
that an unrestricted within-level model is estimated, the second two
numbers, 1 and 2, are the lower and upper limits of the number of factors
to be extracted on the between level, and UB* specifies that an
unrestricted between-level model is estimated. The within- and
between-level specifications are crossed. In the example shown above,
the three- and four-factor models and the unrestricted model on the

599
CHAPTER 16

within level are estimated in combination with the one- and two-factor
models and the unrestricted model on between resulting in nine
solutions.

If UW and UB are used instead of UW* and UB*, the unrestricted

models are not estimated but instead the model parameters are fixed at
the sample statistic values. This can speed up the analysis.

For multilevel exploratory factor analysis, the TYPE option can be

specified using only numbers or the UW* and UB* specifications for
each level. For example,

TYPE = TWOLEVEL EFA 3 4 UB*;

specifies that three- and four-factors models on the within level are
estimated in combination with an unrestricted model on the between
level.

TYPE = TWOLEVEL EFA UW* 12;

specifies that an unrestricted model on the within level is estimated in

combination with one- and two-factor model on the between level.

ESTIMATOR
The ESTIMATOR option is used to specify the estimator to be used in
the analysis. The default estimator differs depending on the type of
analysis and the measurement scale of the dependent variable(s). Not all
estimators are available for all models. Following is a table that shows
which estimators are available for specific models and variable types.
The information is broken down by models with all continuous
dependent variables, those with at least one binary or ordered categorical
dependent variable, and those with at least one censored, unordered
categorical, or count dependent variable. All of the estimators require
individual-level data except ML for TYPE=GENERAL and EFA, GLS,
and ULS which can use summary data. The default settings are
indicated by bold type.

The first column of the table shows the combinations of TYPE settings
that are allowed. The second column shows the set of estimators
available for the analysis types in the first column for a model with all

600
 ANALYSIS Command

continuous dependent variables. The third column shows the set of

estimators available for the analysis types in the first column for a model
with at least one binary or ordered categorical dependent variable. The
fourth column shows the set of estimators available for the analysis
types in the first column for a model with at least one censored,
unordered categorical, or count dependent variable.

All continuous At least one At least one

Type of Analysis dependent binary or censored,
variables ordered unordered
TYPE= categorical categorical,
dependent or count
variable dependent
variable
GENERAL ML** WLS WLS****
MLM***** WLSM WLSM****
MLMV***** WLSMV WLSMV****
MLR** ULSMV ML*
MLF** ML* MLR*
GLS***** MLR* MLF*
WLS***** MLF*
BAYES BAYES
GENERAL RANDOM ML** ML* ML*
MLR** MLR* MLR*
MLF** MLF* MLF*
GENERAL RANDOM COMPLEX MLR** MLR* MLR*

GENERAL COMPLEX MLR** WLS WLS****

WLSM WLSM****
WLSMV WLSMV****
ULSMV MLR*
MLR*
MIXTURE ML** ML** ML**
MLR** MLR** MLR**
MLF** MLF** MLF**
BAYES BAYES
MIXTURE RANDOM ML** ML** ML**
MLR** MLR** MLR**
MLF** MLF** MLF**
MIXTURE COMPLEX MLR** MLR** MLR*
MIXTURE COMPLEX RANDOM
TWOLEVEL MUML*** ML* ML*
ML** MLR* MLR*
MLR** MLF* MLF*
MLF** WLS
WLS WLSM
WLSM WLSMV
WLSMV ULSMV
ULSMV BAYES
BAYES
TWOLEVEL RANDOM ML** ML* ML*
MLR** MLR* MLR*
MLF** MLF* MLF*
BAYES BAYES

601
CHAPTER 16

TWOLEVEL MIXTURE ML* ML * ML*

TWOLEVEL RANDOM MIXTURE MLR* MLR* MLR*
MLF* MLF* MLF*
COMPLEX TWOLEVEL MLR** MLR* MLR*
COMPLEX TWOLEVEL RANDOM
COMPLEX TWOLEVEL MIXTURE MLR* MLR* MLR*
COMPLEX TWOLEVEL RANDOM MIXTURE
THREELEVEL ML** BAYES NA
THREELEVEL RANDOM MLR**
MLF**
BAYES
COMPLEX THREELEVEL MLR** NA NA
COMPLEX THREELEVEL RANDOM
CROSSCLASSIFIED BAYES BAYES NA
CROSSCLASSIFIED RANDOM
EFA ML WLS ML*
MLM***** WLSM MLR*
MLMV***** WLSMV MLF*
MLR** ULS
MLF** ULSMV
ULS***** ML*
BAYES MLR*
MLF*
BAYES
EFA MIXTURE ML ML* ML*
MLR** MLR* MLR*
MLF MLF* MLF*
EFA COMPLEX MLR** WLS MLR*
WLSM
WLSMV
ULSMV
MLR*
EFA TWOLEVEL ML WLS NA
MLR** WLSM
MLF WLSMV
WLS ULSMV
WLSM
WLSMV
ULSMV

* Numerical integration is required.

** Numerical integration is an option.
*** Maximum likelihood with balanced data. Limited-information
for unbalanced data. Not available with missing data.
**** Only available for censored outcomes without inflation.
***** Not available with missing data.
NA Not available

602
 ANALYSIS Command

Following is a description of what the above estimator settings represent:

• ML – maximum likelihood parameter estimates with conventional

standard errors and chi-square test statistic
• MLM – maximum likelihood parameter estimates with standard
errors and a mean-adjusted chi-square test statistic that are robust to
non-normality. The MLM chi-square test statistic is also referred to
as the Satorra-Bentler chi-square.
• MLMV – maximum likelihood parameter estimates with standard
errors and a mean- and variance-adjusted chi-square test statistic that
are robust to non-normality
• MLR – maximum likelihood parameter estimates with standard
errors and a chi-square test statistic (when applicable) that are robust
to non-normality and non-independence of observations when used
with TYPE=COMPLEX. The MLR standard errors are computed
using a sandwich estimator. The MLR chi-square test statistic is
asymptotically equivalent to the Yuan-Bentler T2* test statistic.
• MLF – maximum likelihood parameter estimates with standard
errors approximated by first-order derivatives and a conventional
chi-square test statistic
• MUML – Muthén’s limited information parameter estimates,
standard errors, and chi-square test statistic
• WLS – weighted least square parameter estimates with conventional
standard errors and chi-square test statistic that use a full weight
matrix. The WLS chi-square test statistic is also referred to as ADF
when all outcome variables are continuous.
• WLSM – weighted least square parameter estimates using a diagonal
weight matrix with standard errors and mean-adjusted chi-square test
statistic that use a full weight matrix
• WLSMV – weighted least square parameter estimates using a
diagonal weight matrix with standard errors and mean- and variance-
adjusted chi-square test statistic that use a full weight matrix
• ULS – unweighted least squares parameter estimates
• ULSMV – unweighted least squares parameter estimates with
standard errors and a mean- and variance-adjusted chi-square test
statistic that use a full weight matrix
• GLS – generalized least square parameter estimates with
conventional standard errors and chi-square test statistic that use a
normal-theory based weight matrix

603
CHAPTER 16

• Bayes – Bayesian posterior parameter estimates with credibility

intervals and posterior predictive checking

BAYESIAN ESTIMATION

Bayesian estimation differs from frequentist estimation in that

parameters are not considered to be constants but to be variables
(Gelman et al., 2004). The parameters can be given priors
corresponding to theory or previous studies. Together with the
likelihood of the data, this gives rise to posterior distributions for the
parameters. Bayesian estimation uses Markov chain Monte Carlo
(MCMC) algorithms to create approximations to the posterior
distributions by iteratively making random draws in the MCMC chain.
The initial draws in the MCMC chain are referred to as the burnin phase.
In Mplus, the first half of each chain is discarded as being part of the
burnin phase. Convergence is assessed using the Gelman-Rubin
convergence criterion based on the potential scale reduction factor for
each parameter (Gelman & Rubin, 1992; Gelman et al., 2004, pp. 296-
297). With multiple chains, this is a comparison of within- and between-
chain variation. With a single chain, the last half of the iterations is split
into two quarters and the potential scale reduction factor is computed for
these two quarters. Convergence can also be monitored by the trace
plots of the posterior draws in the chains. Auto-correlation plots
describe the degree of non-independence of consecutive draws. These
plots aid in determining the quality of the mixing in the chain. For each
parameter, credibility intervals are obtained from the percentiles of its
posterior distribution. Model comparisons are aided by the Deviance
Information Criterion (DIC). Overall test of model fit is judged by
Posterior Predictive Checks (PPC) where the observed data is compared
to the posterior predictive distribution. In Mplus, PPC p-values are
computed using the likelihood-ratio chi-square statistic for continuous
outcomes and for the continuous latent response variables of categorical
outcomes. Gelman et al. (2004, Chapter 6) and Lee (2007, Chapter 5)
give overviews of model comparison and model checking. For a
technical description of the Bayesian implementation, see Asparouhov
and Muthén (2010b).

Bayesian estimation is available for continuous, binary, ordered

categorical (ordinal) or combinations of these variable types with
TYPE=GENERAL, MIXTURE with only one categorical latent
variable, TWOLEVEL, THREELEVEL, THREELEVEL RANDOM,

604
 ANALYSIS Command

CROSSCLASSIFIED, CROSSCLASSIFIED RANDOM, and EFA. To

obtain Bayesian estimation, specify:

ESTIMATOR=BAYES;

PARAMETERIZATION
The PARAMETERIZATION option is used for two purposes. The first
purpose is to change from the default Delta parameterization to the
alternative Theta parameterization when TYPE=GENERAL is used, at
least one observed dependent variable is categorical, and weighted least
squares estimation is used in the analysis. The second purpose is to
change from the default logit regression parameterization to either the
loglinear or probability parameterization when TYPE=MIXTURE and
more than one categorical latent variable is used in the analysis.

DELTA VERSUS THETA PARAMETERIZATION

There are two model parameterizations available when

TYPE=GENERAL is used, one or more dependent variables are
categorical, and weighted least squares estimation is used in the analysis.
The first parameterization is referred to as DELTA. This is the default
parameterization. In the DELTA parameterization, scale factors for
continuous latent response variables of observed categorical outcome
variables are allowed to be parameters in the model, but residual
variances for continuous latent response variables are not. The second
parameterization is referred to as THETA. In the THETA
parameterization, residual variances for continuous latent response
variables of observed categorical outcome variables are allowed to be
parameters in the model, but scale factors for continuous latent response
variables are not.

The DELTA parameterization is the default because it has been found to

perform better in many situations (Muthén & Asparouhov, 2002). The
THETA parameterization is preferred when hypotheses involving
residual variances are of interest. Such hypotheses may arise with
multiple group analysis and analysis of longitudinal data. In addition,
there are certain models that can be estimated using only the THETA
parameterization because they have been found to impose improper
parameter constraints with the DELTA parameterization. These are
models where a categorical dependent variable is both influenced by and

605
CHAPTER 16

influences either another observed dependent variable or a latent

variable.

To select the THETA parameterization, specify the following:

PARAMETERIZATION = THETA;

LOGIT, LOGLINEAR, AND PROBABILITY

PARAMETERIZATIONS

There are three model parameterizations available when

TYPE=MIXTURE is used and more than one categorical latent variable
is used in the analysis. The first parameterization is referred to as
LOGIT. This is the default parameterization. In the LOGIT
parameterization, logistic regressions are estimated for categorical latent
variables. In the LOGIT parameterization, the ON and WITH options of
the MODEL command can be used to specify the relationships between
the categorical latent variables. The second parameterization is referred
to as LOGLINEAR. In the LOGLINEAR parameterization, loglinear
models are estimated for categorical latent variables allowing two- and
three-way interactions. In the LOGLINEAR parameterization, only the
WITH option of the MODEL command can be used to specify the
relationships between the categorical latent variables. The third
parameterization is referred to as PROBABILITY. In the
PROBABILITY parameterization, categorical latent variable regression
coefficients can be expressed as probabilities rather than logits.

LINK
The LINK option is used with maximum likelihood estimation to select a
logit or probit link for models with categorical outcomes. The default is
a logit link. Following is an example of how to request a probit link:

LINK = PROBIT:

ROTATION
The ROTATION option is used with TYPE=EFA to specify the type of
rotation of the factor loading matrix to be used in exploratory factor
analysis. The default is the GEOMIN oblique rotation (Yates, 1987;

606
 ANALYSIS Command

Browne, 2001). The algorithms used in the rotations are described in

Jennrich and Sampson (1966), Browne (2001), Bernaards and Jennrich
(2005), and Jennrich (2007). For consistency, the names of the rotations
used in the CEFA program (Browne, Cudeck, Tateneni, & Mels, 2004)
are used for rotations that are included in both the CEFA and Mplus
programs. Target rotations (Browne, 2001) and bi-factor rotations
(Jennrich & Bentler, 2011, 2012) are also available.

Standard errors are available as the default for all rotations except
PROMAX and VARIMAX. THE NOSERROR options of the OUTPUT
command can be used to request that standard errors not be computed.
The following rotations are available:

GEOMIN
QUARTIMIN
CF-VARIMAX
CF-QUARTIMAX
CF-EQUAMAX
CF-PARSIMAX
CF-FACPARSIM
CRAWFER
OBLIMIN
PROMAX
VARIMAX
TARGET
BI-GEOMIN
BI-CF-QUARTIMAX

All rotations are available as both oblique and orthogonal except

PROMAX and QUARTIMIN which are oblique and VARIMAX which
is orthogonal. The default for rotations that can be both oblique and
orthogonal is oblique.

The GEOMIN rotation is recommended when factor indicators have

substantial loadings on more than one factor resulting in a variable
complexity greater than one. Geomin performs well on Thurstone’s 26
variable Box Data (Browne, 2001, Table 3, p. 135). The GEOMIN
epsilon (Browne, 2001) default setting varies as a function of the number
of factors. With two factors, it is .0001. With three factors, it is .001.
With four or more factors, it is .01. The default can be overridden using
the GEOMIN option. The epsilon value must be a positive number.

607
CHAPTER 16

The Geomin rotation algorithm often finds several local minima of the
rotation function (Browne, 2001). To find a global minimum, 30
random rotation starts are used as the default. The RSTARTS option of
the ANALYSIS command can be used to change the default.

Following is an example of how to change the GEOMIN epsilon value

for an oblique rotation:

ROTATION = GEOMIN (OBLIQUE .5);

or

ROTATION = GEOMIN (.5);

where .5 is the value of epsilon.

Following is an example of how to specify an orthogonal rotation for the

GEOMIN rotation and to specify an epsilon value different from the
default:

ROTATION = GEOMIN (ORTHOGONAL .5);

The QUARTIMIN rotation uses the direct quartimin rotation of Jennrich

and Sampson (1966). The following rotations are identical to direct
quartimin:

CF-QUARTIMAX (OBLIQUE)
CRAWFER (OBLIQUE 0)
OBLIMIN (OBLIQUE 0)

The rotations that begin with CF are part of the Crawford-Ferguson

family of rotations (Browne, 2001). They are related to the CRAWFER
rotation by the value of the CRAWFER parameter kappa. Following are
the values of kappa for the Crawford-Ferguson family of rotations
(Browne, 2001, Table 1):

CF-VARIMAX 1/p
CF-QUARTIMAX 0
CF-EQUAMAX m/2p
CF-PARSIMAX (m-1)/(p+m-2)
CF-FACPARSIM 1

608
 ANALYSIS Command

where p is the number of variables and m is the number of factors.

The default for these rotations is oblique. Following is an example of

how to specify an orthogonal rotation for the Crawford-Ferguson family
of rotations:

ROTATION = CF-VARIMAX (ORTHOGONAL);

The CRAWFER rotation is a general form of the Crawford-Ferguson

family of rotations where kappa can be specified as a value from 0
through 1. The default value of kappa is 1/p where p is the number of
variables. Following is an example of how to specify an orthogonal
rotation for the CRAWFER rotation and to specify a kappa value
different from 1/p:

ROTATION = CRAWFER (ORTHOGONAL .5);

where .5 is the value of kappa. The kappa value can also be changed for
an oblique rotation as follows:

ROTATION = CRAWFER (OBLIQUE .5);

or

ROTATION = CRAWFER (.5);

The default for the OBLIMIN rotation is oblique with a gamma value of
0. Gamma can take on any value. Following is an example of how to
specify an orthogonal rotation for the OBLIMIN rotation and to specify
a gamma value different than 0:

ROTATION = OBLIMIN (ORTHOGONAL 1);

where 1 is the value of gamma. The gamma value can also be changed
for an oblique rotation as follows:

ROTATION = OBLIMIN (OBLIQUE 1);

or

ROTATION = OBLIMIN (1);

609
CHAPTER 16

The VARIMAX and PROMAX rotations are the same rotations as those
available in earlier versions of Mplus. The VARIMAX rotation is the
same as the CF-VARIMAX orthogonal rotation except that VARIMAX
row standardizes the factor loading matrix before rotation.

The TARGET setting of the ROTATION option (Browne, 2001) is used

with models that have a set of EFA factors in the MODEL command.
This setting allows the specification of target factor loading values to
guide the rotation of the factor loading matrix. Typically these values
are zero. The default for the TARGET rotation is oblique. Following is
an example of how to specify an orthogonal TARGET rotation:

ROTATION = TARGET (ORTHOGONAL);

For TARGET rotation, a minimum number of target values must be

given for purposes of model identification. For the oblique TARGET
rotation, the minimum is m(m-1) where the m is the number of factors.
For the orthogonal TARGET rotation, the minimum is m(m-1)/2. The
target values are given in the MODEL command using the tilde (~)
symbol.

For the bi-factor rotations, BI-GEOMIN and BI-CF-QUARTIMAX, a

general factor is allowed in combination with specific factors. In the
oblique rotation, the specific factors are correlated with the general
factor and are correlated with each other. In the orthogonal rotation, the
specific factors are uncorrelated with the general factor and are
uncorrelated with each other. The default for the BI-GEOMIN and BI-
CF-QUARTIMAX rotations is oblique. Following is an example of how
to specify an orthogonal BI-GEOMIN rotation:

ROTATION = BI-GEOMIN (ORTHOGONAL);

ROWSTANDARDIZATION
The ROWSTANDARDIZATION option is used with exploratory factor
analysis (EFA) and when a set of EFA factors is part of the MODEL
command to request row standardization of the factor loading matrix
before rotation. The ROWSTANDARDIZATION option has three
settings: CORRELATION, KAISER, and COVARIANCE. The
CORRELATION setting rotates a factor loading matrix derived from a
correlation matrix with no row standardization. The KAISER setting

610
 ANALYSIS Command

rotates a factor loading matrix derived from a correlation matrix with

standardization of the factor loadings in each row using the square root
of the sum of the squares of the factor loadings in each row (Browne,
2001). The COVARIANCE setting rotates a factor loading matrix
derived from a covariance matrix with no row standardization. The
default is CORRELATION. The COVARIANCE setting is not allowed
for TYPE=EFA. If factor loading equalities are specified in a model for
EFA factors, the CORRELATION and KAISER settings are not
allowed.

Following is an example of how to specify row standardization using the

Kaiser method:

ROWSTANDARDIZATION = KAISER;

PARALLEL
The PARALLEL option is used with TYPE=EFA to determine the
optimum number of factors in an exploratory factor analysis. It is
available for continuous outcomes using maximum likelihood
estimation. Parallel analysis (see, for example, Fabrigar, Wegener,
MacCallum, & Strahan, 1999; Hayton, Allen, & Scarpello, 2004) is a
method that uses random data with the same number of observations and
variables as the original data. The correlation matrix of the random data
is used to compute eigenvalues. These eigenvalues are compared to the
eigenvalues of the original data. The optimum number of factors is the
number of the original data eigenvalues that are larger than the random
data eigenvalues. TYPE=PLOT2 of the PLOT command gives a plot of
the sample eigenvalues, the parallel analysis eigenvalues, and the
parallel analysis eigenvalues for the 95th percentile. The PARALLEL
option is specified as follows:

PARALLEL = 50;

where 50 is the number of random data sets that is drawn.

MODEL
The MODEL option is used to make changes to the defaults of the
MODEL command. The MODEL option has three settings:
NOMEANSTRUCTURE, NOCOVARIANCES, and ALLFREE. The

611
CHAPTER 16

NOMEANSTRUCTURE setting is used with TYPE=GENERAL to

specify that means, intercepts, and thresholds are not included in the
analysis model. The NOCOVARIANCES setting specifies that the
covariances and residual covariances among all latent and observed
variables in the analysis model are fixed at zero. The WITH option of
the MODEL command can be used to free selected covariances and
residual covariances. Following is an example of how to specify that the
covariances and residual covariances among all latent and observed
variables in the model are fixed at zero:

MODEL = NOCOVARIANCES;

The ALLFREE setting is used with TYPE=MIXTURE, the

KNOWNCLASS option, ESTIMATOR=BAYES, and a special
automatic labeling function to assign zero-mean and small-variance
priors to differences in intercepts, thresholds, and factor loadings across
groups. By specifying MODEL=ALLFREE, factor means, variances,
and covariances are free across groups except for factor means in the last
group which are fixed at zero. In addition, intercepts, thresholds, factor
loadings, and residual variances of the factor indicators are free across
the groups.

Following is an example of how to use the ALLFREE setting and

automatic labeling to assign zero-mean and small-variance priors to
differences in intercepts, thresholds, and factor loadings across the ten
groups.

MODEL: %OVERALL%
f1 BY y1-y3* (lam#_1-lam#_3);
f2 BY y4-y6* (lam#_4-lam#_6);
[y1-y6] (nu#_1-nu#_6);

MODEL PRIORS:
DO(1,6) DIFF(lam1_#-lam10_#)~N(0,0.01);
DO(1,6) DIFF(nu1_#-nu10_#)~N(0,0.01);

In the overall part of the model, labels are assigned to the factor loadings
and the intercepts using automatic labeling for groups. The labels must
include the number sign (#) followed by the underscore (_) symbol
followed by a number. The number sign (#) refers to a group and the
number refers to a parameter. The label lam#_1 is assigned to the factor

612
 ANALYSIS Command

loading for y1; the label lam#_2 is assigned to the factor loading for y2;
and the label lam#_3 is assigned to the factor loading for y3. These
labels are expanded to include group information. For example, the
label for parameter 1 is expanded across the ten groups to give labels
lam1_1, lam2_1 through lam10_1. In MODEL PRIORS, these expanded
labels are used to assign zero-mean and small-variance priors to the
differences across groups of the factor loadings and intercepts using the
DO and DIFF options. They can be used together to simplify the
assignment of priors to a large set of difference parameters for models
with multiple groups and multiple time points. For the DO option, the
numbers in parentheses give the range of values for the do loop. The
number sign (#) is replaced by these values during the execution of the
do loop. The numbers refer to the six factor indicators.

REPSE
The REPSE option is used to specify the resampling method that was
used to create existing replicate weights or will be used to generate
replicate weights (Fay, 1989; Korn & Graubard, 1999; Lohr, 1999;
Asparouhov, 2009). Replicate weights are used in the estimation of
standard errors of parameter estimates. The REPSE option has six
settings: BOOTSTRAP, JACKKNIFE, JACKKNIFE1, JACKKNIFE2,
BRR and FAY. There is no default. The REPSE option must be
specified when replicate weights are used or generated.

With the BOOTSTRAP setting, the BOOTSTRAP option of the

ANALYSIS command is used to specify the number of bootstrap draws
used in the generation of the replicate weights. With the JACKKNIFE
setting, the number of Jackknife draws is equal to the number of PSU’s
in the sample. A multiplier file is required for JACKKNIFE when
replicate weights are used. The size of this file is one column with rows
equal to the number of PSU’s. For each PSU in a stratum, the value in
the file is equal to the number of PSU’s in the stratum minus one divided
by the number of PSU’s in the stratum. All PSU’s in a stratum have the
same value. If replicate weights are generated using JACKKNIFE, a
multiplier file can be saved. JACKKNIFE1 cannot be used when data
are stratified. JACKKNIFE2, balanced repeated replication (BRR), and
FAY are available only when there are two PSU’s in each stratum. The
BRR and FAY resampling methods use Hadamard matrices. With BRR
and FAY, the number of replicate weights is equal to the size of the
Hadamard matrix. The REPSE option is specified as follows:

613
CHAPTER 16

REPSE = BRR;

where BRR specifies that the balanced repeated replication resampling

method is used to generate replicate weights.

For the FAY resampling method, a constant can be given that is used to
modify the sample weights. The constant must range between zero and
one. The default is .3. The REPSE option for the FAY setting is
specified as follows:

REPSE = FAY (.5);

where .5 is the constant used to modify the sample weights.

BASEHAZARD
The BASEHAZARD option is used in continuous-time survival analysis
to specify whether a non-parametric or a parametric baseline hazard
function is used in the estimation of the model. The default is OFF
which uses the non-parametric baseline hazard function. Following is an
example of how to request a parametric baseline hazard function:

BASEHAZARD = ON;

With TYPE=MIXTURE, the OFF setting has two settings, EQUAL and
UNEQUAL. The EQUAL setting is the default. With the EQUAL
setting, the baseline hazard parameters are held equal across classes. To
relax this equality, specify:

BASEHAZARD = OFF (UNEQUAL);

CHOLESKY
The CHOLESKY option is used in conjunction with
ALGORITHM=INTEGRATION to decompose the continuous latent
variable covariance matrix and the observed variable residual covariance
matrix into orthogonal components in order to improve the optimization.
The optimization algorithm starts out with Fisher Scoring used in
combination with EM. The CHOLESKY option has two settings: ON
and OFF. The default when all dependent variables are censored,
categorical, and counts is ON except for categorical dependent variables

614
 ANALYSIS Command

when LINK=PROBIT. Then and in all other cases, it is OFF. To turn

the CHOLESKY option ON, specify:

CHOLESKY = ON;

ALGORITHM
The ALGORITHM option is used in conjunction with
TYPE=MIXTURE, TYPE=RANDOM, and TYPE=TWOLEVEL with
maximum likelihood estimation to indicate the optimization method to
use to obtain maximum likelihood estimates and to specify whether the
computations require numerical integration. The ALGORITHM option
is used with TYPE=TWOLEVEL and weighted least squares estimation
to indicate the optimization method to use to obtain sample statistics for
model estimation. There are four settings related to the optimization
method: EM, EMA, FS, and ODLL. The default depends on the analysis
type.

EM optimizes the complete-data loglikelihood using the expectation

maximization (EM) algorithm (Dempster et al., 1977). EMA is an
accelerated EM procedure that uses Quasi-Newton and Fisher Scoring
optimization steps when needed. FS is Fisher Scoring. ODLL optimizes
the observed-data loglikelihood directly.

To select the EM algorithm, specify the following:

ALGORITHM = EM;

The INTEGRATION setting of the ALGORITHM option is used in

conjunction with numerical integration and the INTEGRATION option
of the ANALYSIS command.

To select INTEGRATION, specify the following:

ALGORITHM = INTEGRATION;

The ALGORITHM option can specify an optimization setting in

addition to the INTEGRATION setting, for example,

ALGORITHM = INTEGRATION EM;

615
CHAPTER 16

OPTIONS RELATED TO NUMERICAL

INTEGRATION

INTEGRATION

The INTEGRATION option is used to specify the type of numerical

integration and the number of integration points to be used in the
computation when ALGORITHM=INTEGRATION is used. The
INTEGRATION option has three settings: STANDARD,
GAUSSHERMITE, and MONTECARLO. The default is STANDARD.
STANDARD uses rectangular (trapezoid) numerical integration. The
default for TYPE=EFA and TYPE=TWOLEVEL with weighted least
squares estimation is 7 integration points per dimension. For all other
analyses, the default is 15 integration points per dimension.
GAUSSHERMITE uses Gauss-Hermite integration with a default of 15
integration point per dimension. MONTECARLO uses randomly
generated integration points. The default number of integration points
varies depending on the analysis type. In most cases, it is 500.

Following is an example of how the INTEGRATION option is used to

change the number of integration points for the default setting of
STANDARD.

INTEGRATION = 10;

where 10 is the number of integration points per dimension to be used in

the computation. An alternative specification is:

INTEGRATION = STANDARD (10);

To select the MONTECARLO setting, specify:

INTEGRATION = MONTECARLO;

The default number of integration points varies depending on the

analysis type. In most cases, 500 integration points are used. Following
is an example of how to specify a specific number of Monte Carlo
integration points:

INTEGRATION = MONTECARLO (1000);

616
 ANALYSIS Command

MCSEED

The MCSEED option is used to specify a random seed when the

MONTECARLO setting of the INTEGRATION option is used. It is
specified as follows:

MCSEED = 23456;

ADAPTIVE

The ADAPTIVE option is used to customize the numerical integration

points for each observation during the computation. The ADAPTIVE
option is available for each of the three settings of the INTEGRATION
option. The ADAPTIVE option has two settings: ON and OFF. The
default is ON. To turn the ADAPTIVE option off, specify:

ADAPTIVE = OFF;

INFORMATION
The INFORMATION option is used to select the estimator of the
information matrix to be used in computing standard errors when the ML
or MLR estimators are used for analysis. The INFORMATION option
has three settings: OBSERVED, EXPECTED, and COMBINATION.
OBSERVED estimates the information matrix using observed second-
order derivatives; EXPECTED estimates the information matrix using
expected second-order derivatives; and COMBINATION estimates the
information matrix using a combination of observed and expected
second-order derivatives. For MLR, OBSERVED, EXPECTED, and
COMBINATION refer to the outside matrices of the sandwich estimator
used to compute standard errors. The INFORMATION option is
specified as follows:

INFORMATION = COMBINATION;

The default is to estimate models under missing data theory using all
available data. In this case, the observed information matrix is used.
For models with all continuous outcomes that are estimated without
numerical integration, the expected information matrix is also available.
For other outcome types and models that are estimated with numerical
integration, the combination information matrix is also available.

617
CHAPTER 16

The defaults for the information matrix when the LISTWISE option of
the DATA command is used are summarized in the tables below. The
information matrix defaults vary depending on the analysis type. The
bolded entry is the default. Only the ML and MLR estimators have
choices beyond the default. Following is the information matrix table
for models with all continuous dependent variables that are estimated
without numerical integration:

Type of Analysis

TYPE=
GENERAL ML Observed
MLR Expected
------------------- -------------------
MLM Expected
MLMV
------------------- -------------------
MLF Observed

GENERAL RANDOM ML Observed

MLR Combination
------------------ -------------------
MLF Observed
GENERAL RANDOM COMPLEX MLR Observed
Combination
GENERAL COMPLEX MLR Observed
Expected
MIXTURE ML Observed
MIXTURE RANDOM MLR Combination
------------------ -------------------
MLF Observed
MIXTURE COMPLEX MLR Observed
MIXTURE COMPLEX RANDOM Combination

TWOLEVEL MUML* Expected

------------------- -------------------
ML Observed
MLR Expected
------------------- -------------------
MLF Observed
TWOLEVEL MIXTURE ML Observed
TWOLEVEL RANDOM MLR Combination
TWOLEVEL RANDOM MIXTURE ------------------- -------------------
MLF Observed
COMPLEX TWOLEVEL MLR Observed
Expected
COMPLEX TWOLEVEL MIXTURE MLR Observed
COMPLEX TWOLEVEL RANDOM Combination

618
 ANALYSIS Command

COMPLEX TWOLEVEL RANDOM MIXTURE

THREELEVEL ML Observed
THREELEVEL RANDOM MLR
COMPLEX THREELEVEL MLR Observed
COMPLEX THREELEVEL RANDOM
EFA ML Observed
Expected
-------------------- --------------------
MLR Observed
Combination
-------------------- --------------------
MLF Observed
-------------------- --------------------
MLM Expected
MLMV

Following is the information matrix table for models with at least one
binary, ordered categorical (ordinal), censored, unordered categorical
(nominal), or count dependent variable and for models estimated using
numerical integration:

Type of Analysis

TYPE=
GENERAL ML Observed
GENERAL RANDOM MLR Combination
MIXTURE ------------------- -------------------
MIXTURE RANDOM MLF Observed
TWOLEVEL
TWOLEVEL RANDOM
TWOLEVEL MIXTURE
TWOLEVEL MIXTURE RANDOM
GENERAL COMPLEX MLR Observed
GENERAL COMPLEX RANDOM Combination
MIXTURE COMPLEX
MIXTURE COMPLEX RANDOM
COMPLEX TWOLEVEL
COMPLEX TWOLEVEL RANDOM
COMPLEX TWOLEVEL MIXTURE
COMPLEX TWOLEVEL MIXTURE RANDOM
EFA COMPLEX
EFA ML Observed
EFA MIXTURE MLR Combination
-------------------- --------------------
MLF Observed

619
CHAPTER 16

BOOTSTRAP
The BOOTSTRAP option is used to request bootstrapping and to specify
the type of bootstrapping and the number of bootstrap draws to be used
in the computation. Two types of bootstrapping are available, standard
and residual (Bollen & Stine, 1992; Efron & Tibshirani, 1993; Enders,
2002). Residual bootstrap is the Bollen-Stine bootstrap. The
BOOTSTRAP option requires individual data.

Standard bootstrapping is available for the ML, WLS, WLSM,

WLSMV, ULS, and GLS estimators. The reason that it is not available
for MLR, MLF, MLM, and MLMV is that parameter estimates and
bootstrap standard errors for these estimators do not differ from those of
ML. Standard bootstrapping is not available for TYPE=RANDOM,
TWOLEVEL, COMPLEX, or EFA; ALGORITHM=INTEGRATION; or
in conjunction with the MONTECARLO command.

Residual bootstrapping is available only for continuous outcomes using

maximum likelihood estimation. In addition to the restrictions for
standard bootstrapping listed above, residual bootstrapping is not
available for TYPE=MIXTURE.

When the BOOTSTRAP option is used alone, bootstrap standard errors

of the model parameter estimates are obtained for standard bootstrapping
and bootstrap standard errors of the model parameter estimates and the
chi-square p-value are obtained for residual bootstrapping. When the
BOOTSTRAP option is used in conjunction with the CINTERVAL
option of the OUTPUT command, bootstrap standard errors of the model
parameter estimates and either symmetric, bootstrap, or bias-corrected
bootstrap confidence intervals for the model parameter estimates can be
obtained. The BOOTSTRAP option can be used in conjunction with the
MODEL INDIRECT command to obtain bootstrap standard errors for
indirect effects. When both MODEL INDIRECT and CINTERVAL are
used, bootstrap standard errors and either symmetric, bootstrap, or bias-
corrected bootstrap confidence intervals are obtained for the indirect
effects.

The BOOTSTRAP option for standard bootstrapping is specified as

follows:

BOOTSTRAP = 500;

620
 ANALYSIS Command

where 500 is the number of bootstrap draws to be used in the

computation. An alternative specification is:

BOOTSTRAP = 500 (STANDARD);

The BOOTSTRAP option for residual bootstrapping is specified as

follows:

BOOTSTRAP = 500 (RESIDUAL);

where 500 is the number of bootstrap draws to be used in the

computation.

LRTBOOTSTRAP
The LRTBOOTSTRAP option is used in conjunction with the TECH14
option of the OUTPUT command to specify the number of bootstrap
draws to be used in estimating the p-value of the parametric
bootstrapped likelihood ratio test (McLachlan & Peel, 2000). The
default number of bootstrap draws is determined by the program using a
sequential method in which the number of draws varies from 2 to 100.
The LRTBOOTSTRAP option is used to override this default.

The LRTBOOTSTRAP option is specified as follows:

LRTBOOTSTRAP = 100;

where 100 is the number of bootstrap draws to be used in estimating the

p-value of the parametric bootstrapped likelihood ratio test.

OPTIONS RELATED TO RANDOM STARTS

For TYPE=MIXTURE, TYPE=TWOLEVEL with categorical outcomes
and weighted least squares estimation, and TYPE=EFA, random sets of
starting values can be generated. Random starts can be turned off or
done more thoroughly using the following set of options.

When TYPE=MIXTURE is used, random sets of starting values are

generated as the default for all parameters in the model except variances
and covariances. These random sets of starting values are random
perturbations of either user-specified starting values or default starting

621
CHAPTER 16

values produced by the program. Maximum likelihood optimization is

done in two stages. In the initial stage, random sets of starting values are
generated. An optimization is carried out for ten iterations using each of
the random sets of starting values. The ending values from the
optimizations with the two highest loglikelihoods are used as the starting
values in the final stage optimizations which are carried out using the
default optimization settings for TYPE=MIXTURE.

When TYPE=TWOLEVEL with categorical outcomes and weighted

least squares estimation or TYPE=EFA is used random sets of starting
values are generated for the factor loading parameters in the model. For
TYPE=TWOLEVEL with categorical outcomes and weighted least
squares estimation, these random sets of starting values are random
perturbations of either user-specified starting values or default starting
values produced by the program. For TYPE=EFA, these random sets of
starting values are random perturbations of default starting values
produced by the program.

STARTS

The STARTS option is used to specify the number of random sets of

starting values to generate in the initial stage and the number of
optimizations to use in the final stage. For TYPE=MIXTURE, the
default is 20 random sets of starting values in the initial stage and 4
optimizations in the final stage. To turn off random starts, the STARTS
option is specified as follows:

STARTS = 0;

Following is an example of how to use the STARTS option for

TYPE=MIXTURE:

STARTS = 100 20;

specifies that 100 random sets of starting values are generated in the
initial stage and 20 optimizations are carried out in the final stage using
the default optimization settings for TYPE=MIXTURE.

Following are recommendations for a more thorough investigation of

multiple solutions:

622
 ANALYSIS Command

STARTS = 400 100;

or

STARTS = 1000 250;

For TYPE=EFA; TYPE=GENERAL; and TYPE=TWOLEVEL

using the WLS, WLSM, WLSMV, and ULSVM estimators, the
STARTS option is specified as follows:

STARTS = 10;

which specifies that 10 random sets of starting values are generated and
ten optimizations are carried out.

STITERATIONS

The STITERATIONS option is used to specify the maximum number of

iterations allowed in the initial stage. The default number of iterations is
10. For a more thorough investigation, 20 iterations can be requested as
follows:

STITERATIONS = 20;

STCONVERGENCE

The STCONVERGENCE option is used to specify the value of the

derivative convergence criterion to be used in the initial stage
optimization. The default is one.

STSCALE

The STSCALE option is used to specify the scale of the random

perturbation. The default is five which represents a medium level scale
of perturbation.

STSEED

The STSEED option is used to specify the random seed for generating
the random starts. The default value is zero.

623
CHAPTER 16

OPTSEED

The OPTSEED option is used to specify the random seed that has been
found to result in the highest loglikelihood in a previous analysis. The
OPTSEED option results in no random starts being used.

K-1STARTS
The K-1STARTS option is used in conjunction with the TECH11 and
TECH14 options of the OUTPUT command to specify the number of
starting values to use in the initial stage and the number of optimizations
to use in the final stage for the k-1 class analysis model. When the
OPTSEED option is used, the default is 20 random sets of starting
values in the initial stage and 4 optimizations in the final stage. When
the OPTSEED option is not used, the default is the same as what is used
for the STARTS option. Following is an example of how to specify the
K-1STARTS option:

K-1STARTS = 80 16;

which specifies that 80 random sets of starting values are generated in

the initial stage and 16 optimizations are carried out in the final stage
using the default optimization settings for TYPE=MIXTURE.

LRTSTARTS
The LRTSTARTS option is used in conjunction with the TECH14
option of the OUTPUT command to specify the number of starting
values to use in the initial stage and the number of optimizations to use
in the final stage for the k-1 and k class models when the data generated
by bootstrap draws are analyzed. The default for the k-1 class model is 0
random sets of starting values in the initial stage and 0 optimizations in
the final stage. One optimization is carried out for the unperturbed set of
starting values. The default for the k class model is 40 random sets of
starting values in the initial stage and 8 optimizations in the final stage.

Following is an example of how to use the LRTSTARTS option:

LRTSTARTS = 2 1 80 16;

624
 ANALYSIS Command

which specifies that for the k-1 class model 2 random sets of starting
values are used in the initial stage and 1 optimization is carried out in the
final stage and for the k class model 80 random sets of starting values
are used in the initial stage and 16 optimizations are carried out in the
final stage.

RSTARTS
The RSTARTS option is used to specify the number of random sets of
starting values to use for the GPA rotation algorithm and the number of
rotated factor solutions with the best unique rotation function values to
print for exploratory factor analysis. The default is 30 random sets of
starting values and printing of the best solution. Following is an
example of how to use the RSTARTS option.

RSTARTS = 10 2;

which specifies that 10 random sets of starting values are used for the
rotations and that the rotated factor solutions with the two best rotation
function values will be printed.

DIFFTEST
The DIFFTEST option is used to obtain a correct chi-square difference
test when the MLMV and the WLSMV estimators are used because the
difference in chi-square values for two nested models using the MLMV
or WLSMV chi-square values is not distributed as chi-square. The chi-
square difference test compares the H0 analysis model to a less
restrictive H1 alternative model in which the H0 model is nested. To
obtain a correct chi-square difference test for MLMV or WLSMV, a two
step procedure is needed. In the first step, the H1 model is estimated. In
the H1 analysis, the DIFFTEST option of the SAVEDATA command is
used to save the derivatives needed for the chi-square difference test. In
the second step, the H0 model is estimated and the chi-square difference
test is computed using the derivatives from the H0 and H1 analyses. The
DIFFTEST option of the ANALYSIS command is used as follows to
specify the name of the data set that contains the derivatives from the H1
analysis:

DIFFTEST = deriv.dat;

625
CHAPTER 16

where deriv.dat is the name of the data set that contains the derivatives
from the H1 analysis that were saved using the DIFFTEST option of the
SAVEDATA command when the H1 model was estimated.

MULTIPLIER
The MULTIPLIER option is used with the JACKKNIFE setting of the
RESPE option when replicate weights are used in the analysis to provide
multiplier values needed for the computation of standard errors. The
MULTIPLIER option is specified as follows:

MULTIPLIER = multiplier.dat;

where multiplier.dat is the name of the data set that contains the
multiplier values needed for the computation of standard errors.

COVERAGE
The COVERAGE option is used with missing data to specify the
minimum acceptable covariance coverage value for the unrestricted H1
model. The default value is .10 which means that if all variables and
pairs of variables have data for at least ten percent of the sample, the
model will be estimated. Following is an example of how to use the
COVERAGE option:

COVERAGE = .05;

where .05 is the minimum acceptable covariance coverage value.

ADDFREQUENCY
The ADDFREQUENCY option is used to specify a value that is divided
by the sample size and added to each cell with zero frequency in the two-
way tables that are used in categorical data analysis. As the default, 0.5
divided by the sample size is added to each cell with zero frequency.
The ADDFREQUENCY option is specified as follows:

ADDFREQUENCY = 0;

626
 ANALYSIS Command

where the value 0 specifies that nothing is added to each cell with zero
frequency. Any non-negative value can be used with this option.

OPTIONS RELATED TO ITERATIONS

ITERATIONS

The ITERATIONS option is used to specify the maximum number of

iterations for the Quasi-Newton algorithm for continuous outcomes. The
default number of iterations is 1,000.

SDITERATIONS

The SDITERATIONS option is used to specify the maximum number of

steepest descent iterations for the Quasi-Newton algorithm for
continuous outcomes. The default number of iterations is 20.

H1ITERATIONS

The H1ITERATIONS option is used to specify the maximum number of

iterations for the EM algorithm for the estimation of the unrestricted H1
model. The default number of iterations is 2000.

MITERATIONS

The MITERATIONS option is used to specify the number of iterations

allowed for the EM algorithm. The default number of iterations is 500.

MCITERATIONS

The MCITERATIONS option is used to specify the number of iterations

for the M step of the EM algorithm for categorical latent variables. The
default number of iterations is 1.

MUITERATIONS

The MUITERATIONS option is used to specify the number of iterations

for the M step of the EM algorithm for censored, categorical, and count
outcomes. The default number of iterations is 1.

627
CHAPTER 16

RITERATIONS

The RITERATIONS option is used to specify the maximum number of

iterations in the GPA rotation algorithm for exploratory factor analysis.
The default number of iterations is 10000.

OPTIONS RELATED TO CONVERGENCE

CONVERGENCE

The CONVERGENCE option is used to specify the value of the

derivative convergence criterion to be used for the Quasi-Newton
algorithm for continuous outcomes. The default convergence criterion
for TYPE=TWOLEVEL, TYPE=MIXTURE, TYPE=RANDOM, and
ALGORITHM=INTEGRATION is .000001. The default convergence
criterion for all other models is .00005.

H1CONVERGENCE

The H1CONVERGENCE option is used to specify the value of the

convergence criterion to be used for the EM algorithm for the estimation
of the unrestricted H1 model. The default convergence criterion for
TYPE=THREELEVEL and TYPE=CROSSCLASSIFIED is .001. The
default convergence criterion for all other models is .0001.

LOGCRITERION

The LOGCRITERION option is used to specify the absolute observed-

data loglikelihood change convergence criterion for the EM algorithm.
The default convergence criterion for TYPE=TWOLEVEL,
TYPE=RANDOM, and ALGORITHM=INTEGRATION is .001. The
default convergence criterion for TYPE=MIXTURE with
PARAMETERIZATION=PROBABILITY is .0001. The default
convergence criterion for all other models is .0000001.

RLOGCRITERION

The RLOGCRITERION option is used to specify the relative observed-

data loglikelihood change convergence criterion for the EM algorithm.
The default convergence criterion for TYPE=TWOLEVEL,

628
 ANALYSIS Command

TYPE=RANDOM, and ALGORITHM=INTEGRATION is .000001.

The default convergence criterion for all other models is .0000001.

MCONVERGENCE

The MCONVERGENCE option is used to specify the observed-data log

likelihood derivative convergence criterion for the EM algorithm. The
default convergence criterion for TYPE=TWOLEVEL,
TYPE=RANDOM, and ALGORITHM=INTEGRATION is .001. The
default for TYPE=MIXTURE with
PARAMETERIZATION=PROBABILITY is .0001. The default
convergence criterion for all other models is .000001.

MCCONVERGENCE

The MCCONVERGENCE option is used to specify the complete-data

log likelihood derivative convergence criterion for the M step of the EM
algorithm for categorical latent variables. The default convergence
criterion is .000001.

MUCONVERGENCE

The MUCONVERGENCE option is used to specify the complete-data

log likelihood derivative convergence criterion for the M step of the EM
algorithm for censored, categorical, and count outcomes. The default
convergence criterion is .000001.

RCONVERGENCE

The RCONVERGENCE option is used to specify the convergence

criterion for the GPA rotation algorithm for exploratory factor analysis.
The default convergence criterion is .00001.

629
CHAPTER 16

MIXC

The MIXC option is used to specify whether to use the number of

iterations or the convergence criterion to terminate the M step iterations
of the EM algorithm for categorical latent variables. Following is an
example of how to select the convergence criterion being fulfilled:

MIXC = CONVERGENCE;

MIXU

The MIXU option is used to specify whether to use the number of

iterations or the convergence criterion to terminate the M step iterations
of the EM algorithm for censored, categorical, and count outcomes.
Following is an example of how to select the convergence criterion
being fulfilled:

MIXU = CONVERGENCE;

LOGHIGH
The LOGHIGH option is used to specify the maximum value allowed for
the logit thresholds of the latent class indicators. The default is +15.

LOGLOW
The LOGLOW option is used to specify the minimum value allowed for
the logit thresholds of the latent class indicators. The default is -15.

UCELLSIZE
The UCELLSIZE option is used to specify the minimum expected cell
size allowed for computing chi-square from the frequency table of the
latent class indicators when the corresponding observed cell size is not
zero. The default value is .01.

630
 ANALYSIS Command

VARIANCE
The VARIANCE option is used in conjunction with TYPE=RANDOM
and TYPE=TWOLEVEL when ESTIMATOR=ML, MLR, or MLF to
specify the minimum value that is allowed in the estimation of the
variance of the random effect variables and the variances of the between-
level outcome variables. The default value is .0001.

MATRIX
The MATRIX option identifies the matrix to be analyzed. The default
for continuous outcomes is to analyze the covariance matrix. The
following statement requests that a correlation matrix be analyzed:

MATRIX = CORRELATION;

The analysis of the correlation matrix is allowed only when all

dependent variables are continuous and there is a single group analysis
with no mean structure. Only the WLS estimator is allowed for this type
of analysis.

For models with all categorical dependent variables, the correlation

matrix is always analyzed. For models with combinations of categorical
and continuous dependent variables, the variances for the continuous
dependent variables are always included.

OPTIONS RELATED TO BAYES ESTIMATION

AND MULTIPLE IMPUTATION
POINT

The POINT option is used to specify the type of Bayes point estimate to
compute. The POINT option has three settings: MEDIAN, MEAN, and
MODE. The default is MEDIAN. With the MODE setting, the mode
reported refers to the multivariate mode of the posterior distribution.
This mode is different from the univariate mode reported in the plot of
the Bayesian posterior parameter distribution. To request that the mean
be computed, specify:

POINT = MEAN;

631
CHAPTER 16

CHAINS

The CHAINS option is used to specify how many independent Markov

chain Monte Carlo (MCMC) chains to use. The default is two. To
request that four chains be used, specify:

CHAINS = 4;

With multiple chains, parallel computing uses one chain per processor.
To benefit from this speed advantage, it is important to specify the
number of processors using the PROCESSORS option.

BSEED

The BSEED option is used to specify the seed to use for random number
generation in the Markov chain Monte Carlo (MCMC) chains. The
default is zero. If one chain is used, the seed is used for this chain. If
more than one chain is used, the seed is used for the first chain and is the
basis for generating seeds for the other chains. The randomly generated
seeds for the other chains can be found in TECH8. If the same seed is
used in a subsequent analysis, the other chains will have the same seeds
as in the previous analysis. To request a seed other than zero be used,
specify:

BSEED = 5437;

STVALUES

The STVALUES option is used to specify starting value information

(Asparouhov & Muthén, 2010b). The STVALUES option has three
settings: UNPERTURBED, PERTURBED, and ML. The default is
UNPERTURBED. If the UNPERTURBED setting is specified, the
default or user-specified starting values are used. If the PERTURBED
setting is used, a BSEED value must be specified. The default or user-
specified starting values are randomly perturbed using the BSEED value.
If the ML setting is used, the model is first estimated using maximum
likelihood estimation and the maximum likelihood parameter estimates
are used as starting values in the Bayesian analysis. To request that
maximum likelihood parameters be used as starting values, specify:

STVALUES = ML;

632
 ANALYSIS Command

MEDIATOR

The MEDIATOR option is used to specify how a categorical mediator

variable is treated when it is an independent variable in a regression.
The MEDIATOR option has two settings: LATENT and OBSERVED.
The default is LATENT where the mediator is treated as a continuous
latent response variable underlying the categorical variable unless data
are imputed using a single-level unrestricted H1 model in which case the
default is OBSERVED. When the OBSERVED setting is specified, the
mediator is treated as a continuous observed variable. To request that
the mediator be treated as observed, specify:

MEDIATOR = OBSERVED;

ALGORITHM

The ALGORITHM option is used to specify the Markov chain Monte

Carlo (MCMC) algorithm to use for generating the posterior distribution
of the parameters (Gelman et al., 2004). The ALGORITHM option has
two settings: GIBBS and MH. The default is GIBBS. The GIBBS
setting uses the Gibbs sampler algorithm which divides the parameters
and the latent variables into groups that are conditionally and
sequentially generated. The GIBBS setting has four choices: PX1, PX2,
PX3, and RW. The default is PX1. PX1, PX2, and PX3 use parameter
extension techniques to generate correlation and covariance matrices.
PX1 is described in Asparouhov and Muthén (2010b). PX2 is described
in Boscardin et al. (2008). PX3 is described in Liu and Daniels (2006).
RW uses a random walk, Metropolis-Hastings algorithm to generate
correlation and covariance matrices (Chib & Greenberg, 1998). This
algorithm can generate a covariance matrix with an arbitrary structure.
Following is an example of how to select an alternative choice for the
GIBBS setting:

ALGORITHM = GIBBS (PX3);

The MH setting uses the Metropolis-Hastings algorithm to generate all

of the parameters simultaneously using the observed-data loglikelihood.
The MH setting uses maximum likelihood starting values. The MH
proposal distribution uses the estimated covariance matrix of the
maximum likelihood parameter estimates. The MH algorithm is not

633
CHAPTER 16

available for TYPE=MIXTURE or TYPE=TWOLEVEL. To request

that the Metropolis-Hastings algorithm be used, specify:

ALGORITHM = MH;

BCONVERGENCE

The BCONVERGENCE option is used to specify the value of the

convergence criterion to use for determining convergence of the
Bayesian estimation using the Gelman-Rubin convergence criterion
(Gelman & Rubin, 1992). The Gelman-Rubin convergence criterion
determines convergence by considering within and between chain
variability of the parameter estimates in terms of the potential scale
reduction (PSR) to determine convergence (Gelman et al., 2004, pp. 296-
298). The default is 0.05. The BCONVERGENCE value is used in the
following formula (Asparouhov & Muthén, 2010b):

a = 1 + BCONVERGENCE* factor,

such that convergence is obtained when PSR < a for each parameter.
The factor value ranges between one and two depending on the number
of parameters. With one parameter, the value of factor is one and the
value of a is 1.05 using the default value of BCONVERGENCE. With a
large number of parameters, the value of factor is 2 and the value of a is
1.1 using the default value of BCONVERGENCE.

With a single chain, PSR is defined using the third and the fourth
quarters of the chain. The first half of the chain is discarded as a burnin
phase. To request a stricter convergence criterion, specify:

BCONVERGENCE = .01;

BITERATIONS

The BITERATIONS option is used to specify the maximum and

minimum numbers of iterations for each Markov chain Monte Carlo
(MCMC) chain when the potential scale reduction (PSR) convergence
criterion (Gelman & Rubin, 1992) is used. The default for the maximum
number of iterations is 50,000. The default for the minimum number of
iterations is zero. To request more Bayes iterations for each chain,
specify:

634
 ANALYSIS Command

BITERATIONS = 60000 (2000);

where 60,000 is the maximum number of iterations and 2,000 is the

minimum number of iterations when the PSR convergence criterion
(Gelman & Rubin, 1992) is used.

Another specification is:

BITERATIONS = (2000);

where the default of 50,000 is the maximum number of iterations and

2,000 is the minimum number of iterations when the PSR convergence
criterion is used.

FBITERATIONS

The FBITERATIONS option is used to specify a fixed number of

iterations for each Markov chain Monte Carlo (MCMC) chain when the
potential scale reduction (PSR) convergence criterion (Gelman & Rubin,
1992) is not used. There is no default. When using this option, it is
important to use other means to determine convergence. To request a
fixed number of iterations for each Markov chain Monte Carlo (MCMC)
chain, specify:

FBITERATIONS = 30000;

THIN

The THIN option is used to specify which iterations from the posterior
distribution to use in the parameter estimation. When a chain is mixing
poorly with high auto-correlations, the estimation can be based on every
k-th iteration rather than every iteration. This is referred to as thinning.
The default is 1 in which case every iteration is used. To request that
every 20th iteration be used, specify:

THIN = 20;

DISTRIBUTION

The distribution option is used with the MODE setting of the POINT
option to specify the maximum number of iterations to use to compute

635
CHAPTER 16

the multivariate mode in Bayes estimation. The default is 10,000. If the

number of iterations used in the estimation exceeds the number of
iterations specified using the DISTRIBUTION option, the number of
iterations specified using the DISTRIBUTION option is used. This
number of iterations is selected from the total iterations using equally-
spaced intervals. To request that more iterations be used to compute the
multivariate mode, specify:

DISTRIBUTION = 15000;

KOLMOGOROV

The KOLMOGOROV option is used to request a Kolmogorov-Smirnov

test of equality of the posterior parameter distributions across the
different chains using draws from the chains. The default is 100. To
request more draws, specify:

KOLMOGOROV = 1000;

PRIOR

The PRIOR option is used to request a plot of the prior distribution for
each parameter that has a proper prior. The plot of the prior
distributions can be viewed by choosing Bayesian prior distributions
from the Plot menu of the Mplus Editor. The default is 1,000 draws
from the prior distribution. To request more draws, specify:

PRIOR = 5000;

INTERACTIVE
The INTERACTIVE option is used to allow changes in technical
specifications during the iterations of an analysis when TECH8 is used.
This is useful in analyses that are computationally demanding. If a
starting value set has computational difficulties, it can be skipped. If too
many random starts have been chosen, the STARTS option can be
changed. If a too strict convergence criterion has been chosen, the
MCONVERGENCE option can be changed. Following is an example of
how to use the INTERACTIVE option:

INTERACTIVE = control.dat;

636
 ANALYSIS Command

where control.dat is the name of the file that contains the technical
specifications that can be changed during an analysis. This file is
created automatically and resides in the same directory as the input file.
The following options of the ANALYSIS command are contained in this
file: STARTS, MITERATIONS, MCONVERGENCE,
LOGCRITERION, and RLOGCRITERION. No other options can be
used in this file except the INTERRUPT statement which is used to skip
the current starting value set and go to the next starting value set. It has
settings of 0 and 1. A setting of 0 specifies that a starting value set is not
skipped. A setting of 1 specifies that the starting value set is skipped.
As the default, the INTERRUPT statement is set to 0 and the other
options are set to either the program default values or the values
specified in the input file.

The following file is automatically created and given the name specified
using the INTERACTIVE option.

INTERRUPT = 0
STARTS = 200 50
MITERATIONS = 500
MCONVERGENCE = 1.0E-06
LOGCRITERION = 1.0E-003
RLOGCRITERION = 1.0E-006

When the file is modified and saved, the new settings go into effect
immediately and are applied at each iteration. Following is an example
of a modified control.dat file where INTERRUPT and STARTS are
changed:

INTERRUPT = 1
STARTS = 150 50
MITERATIONS = 500
MCONVERGENCE = 1.0E-06
LOGCRITERION = 1.0E-003
RLOGCRITERION = 1.0E-006

PROCESSORS
The PROCESSORS option is used to specify the number of processors
to be used for parallel computing to increase computational speed.
When random starts are used, the PROCESSORS option is used in

637
CHAPTER 16

conjunction with the STARTS option to determine the number of threads

to be used for parallel computing. The default is one processor and one
thread. Parallel computing is not available for all analyses. For some
analyses, multiple processors are used alone. In other analyses, multiple
processors are used together with threads.

MULTIPLE PROCESSORS

The use of multiple processors without threads is available for

TYPE=MIXTURE; Bayesian analysis with more than one chain unless
STVALUES=ML; models that require numerical integration; models
with all continuous variables, missing data, and maximum likelihood
estimation; and TYPE=TWOLEVEL with categorical outcomes and
ESTIMATOR= WLSMV. In these cases, the PROCESSORS option is
specified using one number as shown below:

PROCESSORS = 8;

where 8 is the number of processors to be used for parallel computing.

For Bayesian analysis, the PROCESSORS option is specified using one

number as shown below:

PROCESSORS = 2;

where 2 is the number of processors used for parallel computing. The

number of processors used cannot exceed the number of chains. If there
are more processors than chains, only the number of processors equal to
the number of chains is used. If there are more chains than processors,
each processor carries out one chain until it is completed and then the
remaining chains are carried out.

PROCESSORS AND THREADS

When processor and threads are used together, the threads are distributed
across the processors and the memory used is a multiple of the number
of threads. For large models that require a lot of memory, it is important
to have fewer threads than processors because computations are slower
or impossible when the memory used by all processors exceeds the
memory limit.

638
 ANALYSIS Command

The use of multiple processors and multiple threads with random starts
as the default is available for TYPE=MIXTURE; Bayesian analysis with
more than one chain if STVALUES=ML; and models that require
numerical integration. They are also available for TYPE=RANDOM
and TYPE=TWOLEVEL and THREELEVEL with continuous outcomes
using ESTIMATOR=ML, MLR, and MLF without numerical integration
if the STARTS option is used. Without random starts only one
processor is used in these cases.

When the PROCESSORS option is used with random starts, it is used

with two numbers, the number of processors and the number of threads.
The number of threads is the smaller of the number of threads specified
using the PROCESSORS option or the number of final perturbations
specified using the STARTS option. Following is an example:

PROCESSORS = 8 4;
STARTS = 400 40;

where 4 is the number of threads and 40 is the number of final stage

optimizations. Because four is smaller than 40, the number of threads
for this example is four. Two processors are distributed across each of
the four threads. Each of the four threads carries out 100 initial stage
and ten final stage optimizations. When a thread completes, its
processors are distributed across the remaining threads.

If the number of threads is not specified, it is the same as the number of

processors. In this case, the PROCESSORS option is specified as
follows:

PROCESSORS = 4;

where 4 is the number of processors to be used in the analysis for

parallel computing. The number of threads is also 4.

639
CHAPTER 16

640
 MODEL Command

CHAPTER 17
MODEL COMMAND

In this chapter, the MODEL command is discussed. The MODEL

command is used to describe the model to be estimated. The first part of
this chapter describes the general modeling framework used by Mplus
and introduces a set of terms that are used to describe the model to be
estimated. The second part of this chapter explains how a model is
translated into the Mplus language using the options of the MODEL
command. The last part of the chapter describes variations of the
MODEL command. The MODEL command has variations for use with
models with indirect effects, models with linear and non-linear
constraints, models with parameter constraints for the Wald test,
multiple group models, mixture models, multilevel models, and models
for generating data for Monte Carlo simulation studies.

THE Mplus FRAMEWORK

VARIABLES
There are three important distinctions that need to be made about the
variables in an analysis in order to be able to specify a model. The
distinctions are whether variables are observed or latent, whether
variables are dependent or independent, and the scale of the observed
dependent variables.

OBSERVED OR LATENT VARIABLES

Two types of variables can be modeled: observed variables and latent

variables. Observed variables are variables that are directly measured
such as test scores and diagnostic criteria. They are sometimes referred
to as manifest variables, outcomes, or indicators. Latent variables are
variables that are not directly measured such as ability, depression, and
health status. They are measured indirectly by a set of observed
variables. There are two types of latent variables: continuous and
categorical. Continuous latent variables are sometimes referred to as
factors, dimension, constructs, or random effects. Categorical latent
variables are sometimes referred to as latent class variables or mixtures.

641
CHAPTER 17

DEPENDENT OR INDEPENDENT VARIABLES

Observed and latent variables can play the role of a dependent variable
or an independent variable in the model. The distinction between
dependent and independent variables is that of a regression analysis for y
regressed on x where y is a dependent variable and x is an independent
variable. An independent variable is one that is not influenced by any
other variable. Dependent variables are those that are influenced by
other variables. Other terms used for dependent variables are outcome
variable, response variable, indicator variable, y variable, and
endogenous variable. Other terms used for independent variables are
covariate, background variable, explanatory variable, predictor, x
variable, and exogenous variable.

SCALE OF OBSERVED DEPENDENT VARIABLES

The scale of observed dependent variables can be continuous, censored,

binary, ordered categorical (ordinal), unordered categorical (nominal),
counts, or combinations of these variable types.

UNDERLYING GENERAL MODEL

The purpose of modeling data is to describe the structure of a data set in
a simple way so that it is more understandable and interpretable.
Essentially, modeling data amounts to specifying a set of relationships
between variables.

The underlying model of Mplus consists of three parts: the measurement

model for the indicators of the continuous latent variables, the
measurement model for the indicators of the categorical latent variables,
and the structural model involving the continuous and categorical latent
variables and the observed variables that are not indicators of the
continuous or categorical latent variables. A model may consist of only
a measurement model as in confirmatory factor analysis or latent class
analysis, only a structural model as in a path analysis, or both a
measurement model and a structural model as in latent variable
structural equation modeling, longitudinal growth modeling, regression
mixture modeling, or growth mixture modeling.

642
 MODEL Command

THE MODEL COMMAND

The MODEL command is used to describe the model to be estimated. It
has options for defining latent variables, describing relationships among
variables in the model, and specifying details of the model. The
MODEL command has variations for use with models with indirect
effects, models with non-linear constraints, models with parameter
constraints for the Wald test, multiple group models, mixture models,
multilevel models, and models for generating data for Monte Carlo
simulation studies.

Following are the options for the MODEL command:

MODEL:
BY short for measured by -- defines latent variables
example: f1 BY y1-y5;
ON short for regressed on -- defines regression relationships
example: f1 ON x1-x9;
PON short for regressed on -- defines paired regression relationships
example: f2 f3 PON f1 f2;
WITH short for correlated with -- defines correlational relationships
example: f1 WITH f2;
PWITH short for correlated with -- defines paired correlational
relationships
example: f1 f2 f3 PWITH f4 f5 f6;
list of variables; refers to variances and residual variances
example: f1 y1-y9;
[list of variables]; refers to means, intercepts, thresholds
example: [f1, y1-y9];
* frees a parameter at a default value or a specific starting value
example: y1* y2*.5;
@ fixes a parameter at a default value or a specific value
example: y1@ y2@0;
(number) constrains parameters to be equal
example: f1 ON x1 (1);
f2 ON x2 (1);
variable$number label for the threshold of a variable
variable#number label for nominal observed or categorical latent variable
variable#1 label for censored or count inflation variable
variable#number label for baseline hazard parameters
variable#number label for a latent class
(name) label for a parameter
{list of variables}; refers to scale factors
example: {y1-y9};

643
CHAPTER 17

| names and defines random effect variables

example: s | y1 ON x1;
AT short for measured at -- defines random effect variables
example: s | y1-y4 AT t1-t4;
XWITH defines interactions between variables;
MODEL INDIRECT: describes the relationships for which indirect and total effects are
requested
IND describes a specific indirect effect or a set of indirect effects;
VIA describes a set of indirect effects that includes specific
mediators;
MODEL CONSTRAINT: describes linear and non-linear constraints on parameters
NEW assigns labels to parameters not in the analysis model;
DO describes a do loop
PLOT describes y-axis variables
LOOP describes x-axis variables
MODEL TEST: describes restrictions on the analysis model for the Wald test
MODEL PRIORS: specifies the prior distribution for the parameters
COVARIANCE assigns a prior to the covariance between two parameters
DO describes a do loop
DIFFERENCE assigns priors to differences between parameters

Following are variations of the MODEL command:

MODEL: describes the analysis model

MODEL label: describes the group-specific model in multiple group analysis
and the model for each categorical latent variable and
combinations of categorical latent variables in mixture modeling
MODEL:
%OVERALL% describes the overall part of a mixture model
%class label% describes the class-specific part of a mixture model
MODEL:
%WITHIN% describes the individual-level model
%BETWEEN% describes the cluster-level model for a two-level model
%BETWEEN label% describes the cluster-level model for a three-level or cross-
classified model
MODEL POPULATION: describes the data generation model
MODEL POPULATION-label: describes the group-specific data generation model in multiple
group analysis and the data generation model for each
categorical latent variable and combinations of categorical
latent variables in mixture modeling
MODEL POPULATION:
%OVERALL% describes the overall data generation model for a mixture
%class label% model
describes the class-specific data generation model for a mixture
model

644
 MODEL Command

MODEL POPULATION:
%WITHIN% describes the individual-level data generation model for a
multilevel model
%BETWEEN% describes the cluster-level data generation model for a two-level
model
%BETWEEN label% describes the cluster-level data generation model for a three-
level or cross-classified model
MODEL COVERAGE: describes the population parameter values for a Monte Carlo
study
MODEL COVERAGE-label: describes the group-specific population parameter values in
multiple group analysis and the population parameter values for
each categorical latent variable and combinations of categorical
latent variables in mixture modeling for a Monte Carlo study
MODEL COVERAGE:
%OVERALL% describes the overall population parameter values of a mixture
model for a Monte Carlo study
%class label% describes the class-specific population parameter values of a
mixture model
MODEL COVERAGE:
%WITHIN% describes the individual-level population parameter values for
coverage
%BETWEEN% describes the cluster-level population parameter values for a
two-level model for coverage
%BETWEEN label% describes the cluster-level population parameter values for a
three-level or cross-classified model for coverage
MODEL MISSING: describes the missing data generation model for a Monte Carlo
study
MODEL MISSING-label: describes the group-specific missing data generation model for
a Monte Carlo study
MODEL MISSING:
%OVERALL% describes the overall data generation model of a mixture model
describes the class-specific data generation model of a mixture
%class label% model

The MODEL command is required for all analyses except exploratory

factor analysis (EFA), exploratory latent class analysis (LCA), a baseline
model, and TYPE=BASIC.

645
CHAPTER 17

MODEL COMMAND OPTIONS

There are three major options in the MODEL command that are used to
describe the relationships among observed variables and latent variables
in the model. They are:

• BY
• ON
• WITH

BY is used to describe the regression relationships in the measurement

model for the indicators of the continuous latent variables. These
relationships define the continuous latent variables in the model. BY is
short for measured by. ON is used to describe the regression
relationships among the observed and latent variables in the model. It is
short for regressed on. WITH is used to describe correlational
(covariance) relationships in the measurement and structural models. It
is short for correlated with.

The model in the following figure is used to illustrate the use of the BY,
ON, and WITH options. The squares represent observed variables and
the circles represent latent variables. Regression relationships are
represented by arrows from independent variables to dependent
variables. The variables f1 and f2 are continuous latent variables. The
observed dependent variables are y1, y2, y3, y4, y5, y6, y7, y8, and y9.
The measurement part of the model consists of the two continuous latent
variables and their indicators. The continuous latent variable f1 is
measured by y1, y2, y3, y4, and y5. The continuous latent variable f2 is
measured by y6, y7, y8, and y9. The structural part of the model
consists of the regression of the two continuous latent variables on nine
observed independent variables. The observed independent variables are
x1, x2, x3, x4, x5, x6, x7, x8, and x9. Following is the MODEL
command for the figure below:

MODEL: f1 BY y1-y5;
f2 BY y6-y9;
f1 f2 ON x1-x9;

646
 MODEL Command

x1 y1

x2 y2

x3 f1 y3

x4 y4

x5 y5

x6 y6

x7 y7
f2
x8 y8

x9 y9

BY
The BY option is used to name and define the continuous latent
variables in the model. BY is short for measured by. The parameters
that are estimated are sometimes referred to as factor loadings or
lambdas. These are the coefficients for the regressions of the observed
dependent variables on the continuous latent variables. These observed
dependent variables are sometimes referred to as factor indicators. Each
BY statement can be thought of as a set of ON statements that describes
the regressions of a set of observed variables on a continuous latent
variable or factor. However, continuous latent variables in the
measurement model cannot be specified using a set of ON statements

647
CHAPTER 17

because BY statements are used to name the continuous latent variables.

BY statements also provide a set of convenient defaults.

Observed factor indicators for continuous latent variables can be

continuous, censored, binary, ordered categorical (ordinal), or counts.
Factor indicators can also be continuous latent variables or the inflation
part of censored and count variables. Combinations of all factor
indicator types are allowed. With TYPE=TWOLEVEL and
TYPE=TWOLEVEL MIXTURE, factor indicators for continuous latent
variables can be between-level random effects. These factor indicators
can appear only on the BETWEEN level.

CONFIRMATORY FACTOR ANALYSIS MODELING

In this section the use of the BY option for confirmatory factor analysis
(CFA) models is described. Following are the two BY statements that
describe how the continuous latent variables in the figure above are
measured:

f1 BY y1- y5;
f2 BY y6- y9;

The factor loading of any observed variable mentioned on the right-hand

side of the BY statement is free to be estimated with the exception of the
factor loading of the first variable after the BY option. This factor
loading is fixed at one as the default. Fixing a factor loading of an
indicator of a continuous latent variable sets the metric of the continuous
latent variable. Setting the metric can also be accomplished by fixing
the variance of the continuous latent variable to one and freeing the
factor loading of the factor indicator that is fixed at one as the default.
In the example above, the factor loadings of y1 and y6 are fixed at one.
The other factor loadings are estimated using default starting values of
one.

Following is an example of how to set the metric of the continuous latent

variable by fixing the variance of the continuous latent variable to one
and allowing all factor loadings to be free:

f1 BY y1* y2- y5;

f2 BY y6* y7- y9;
f1@1 f2@1;

648
 MODEL Command

where the asterisk (*) after y1 and y6 frees the factor loadings of y1 and
y6, and the @1 after f1 and f2 fixes the variances of f1 and f2 to one.
The use of the asterisk (*); @ symbol; and the specification of means,
thresholds, variances, and covariances are discussed later in the chapter.

Residual variances are estimated as the default when factor indicators

are continuous or censored. Residual covariances among the factor
indicators are fixed at zero as the default. All default settings can be
overridden. How to do so is discussed later in this chapter.

The BY option can also be used to define continuous latent variables

underlying other continuous latent variables that have observed factor
indicators. This is referred to as second-order factor analysis. However,
a continuous latent variable cannot be used on the right-hand side of a
BY statement before it has been defined on the left-hand side of another
BY statement. For example, the following statements are acceptable:

f1 BY y1 y2 y3 y4 y5;
f2 BY y6 y7 y8 y9;
f3 BY f1 f2;

whereas, the following statements are not acceptable:

f3 BY f1 f2;
f1 BY y1 y2 y3 y4 y5;
f2 BY y6 y7 y8 y9;

because f1 and f2 are used on the right-hand side of a BY statement

before they are defined on the left-hand side of a BY statement.

EXPLORATORY STRUCTURAL EQUATION MODELING

In this section the use of the BY option for exploratory structural

equation (ESEM) modeling (Asparouhov & Muthén, 2009a) is
described. One of the differences between CFA and EFA factors is that
CFA factors are not rotated. For a set of EFA factors, the factor loading
matrix is rotated as in conventional EFA using the rotations available
through the ROTATION option of the ANALYSIS command. A set of
EFA factors must have the same factor indicators. A set of EFA factors
can be regressed on the same set of covariates. An observed or latent
variable can be regressed on a set of EFA factors. EFA factors are

649
CHAPTER 17

allowed with TYPE=GENERAL and TYPE=COMPLEX with observed

dependent variables that are continuous, censored, binary, ordered
categorical (ordinal), and combinations of these variable types. EFA
factors are not allowed when summary data are analyzed or when the
MLM, MLMV, or GLS estimators are used.

The BY option has three special features that are used with sets of EFA
factors in the MODEL command. One feature is used to define sets of
EFA factors. The second feature is a special way of specifying factor
loading matrix equality for sets of EFA factors. The third feature is used
in conjunction with the TARGET setting of the ROTATION option of
the ANALYSIS command to provide target factor loading values to
guide the rotation of the factor loading matrix for sets of EFA factors.

DEFINING EFA FACTORS

Following is an example of how to define a set of EFA factors using the

BY option:

f1-f2 BY y1-y5 (*1);

where the asterisk (*) followed by a label specifies that factors f1 and f2
are a set of EFA factors with factor indicators y1 through y5.

Following is an alternative specification:

f1 BY y1-y5 (*1);
f2 BY y1-y5 (*1);

where the label 1 specifies that factors f1 and f2 are part of the same set
of EFA factors. Rotation is carried out on the five by two factor loading
matrix. Labels for EFA factors must follow an asterisk (*). EFA factors
with the same label must have the same factor indicators.

More than one set of EFA factors may appear in the MODEL command.
For example,

f1-f2 BY y1-y5 (*1);

f3-f4 BY y6-y10 (*2);

650
 MODEL Command

specifies that factors f1 and f2 are one set of EFA factors with the label
1 and factors f3 and f4 are another set of EFA factors with the label 2.
The two sets of EFA factors are rotated separately.

Factors in a set of EFA factors can be regressed on covariates but the set
of covariates must be the same, for example,

f1-f2 ON x1-x3;

or

f1 ON x1-x3;
f2 ON x1-x3;

A set of EFA factors can also be used as covariates in a regression, for

example,

y ON f1-f2;

EQUALITIES WITH EFA FACTORS

The BY option has a special convention for specifying equalities of the

factor loading matrices for more than one set of EFA factors. The
equality label is placed after the label that defines the set of EFA factors
and applies to the entire factor loading matrix not to a single parameter.
Following is an example of how to specify that the factor loading
matrices for the set of EFA factors f1 and f2 and the set of EFA factors
f3 and f4 are held equal:

f1-f2 BY y1-y5 (*1 1);

f3-f4 BY y6-y10 (*2 1);

The number 1 following the labels 1 and 2 that define the EFA factors
specifies that the factor loadings matrices for the two sets of EFA factors
are held equal.

TARGET ROTATION WITH EFA FACTORS

The BY option has a special feature that is used with the TARGET
setting of the ROTATION option of the ANALYSIS command to
specify target factor loading values for a set of EFA factors (Browne,

651
CHAPTER 17

2001). The target factor loading values are used to guide the rotation of
the factor loading matrix. Typically these values are zero. For the
TARGET rotation, a minimum number of target values must be given for
purposes of model identification. For the default oblique TARGET
rotation, the minimum is m(m-1) where the m is the number of factors.
For the orthogonal TARGET rotation, the minimum is m(m-1)/2. The
target values are given in the MODEL command using the tilde (~)
symbol. The target values are specified in a BY statement using the tilde
(~) symbol as follows:

f1 BY y1-y5 y1~0 (*1);

f2 BY y1-y5 y5~0 (*1);

where the target factor loading values for the factor indicator y1 for
factor f1 and y5 for factor f2 are zero.

ON
The ON option is used to describe the regression relationships in the
model and is short for regressed on. The general form of the ON
statement is:

y ON x;

where y is a dependent variable and x is an independent variable.

Dependent and independent variables can be observed or latent
variables.

In the previous figure, the structural relationships are the regressions of

the continuous latent variables f1 and f2 on the nine independent
variables x1 through x9. The ON statements shown below are used to
specify these regressions:

f1 ON x1-x9;
f2 ON x1-x9;

These statements specify that regression coefficients are free to be

estimated for f1 and f2 regressed on the independent variables x1
through x9 with default starting values of zero.

652
 MODEL Command

For continuous latent variables, the residual variances are estimated as

the default. The residuals of the latent variables are correlated as the
default because residuals are correlated for latent variables that do not
influence any other variable in the model except their own indicators.
These defaults can be overridden. Means, variances, and covariances of
the independent variables in the model should not be mentioned in the
MODEL command because the model is estimated conditioned on the
covariates.

An ON statement can be used to describe the regression relationship

between an observed dependent variable and an observed independent
variable. Following is an example of how to specify the regression of an
observed dependent variable y9 on the observed independent variable
x9:

y9 ON x9;

The general form of the ON statement is used to describe regression

relationships for continuous latent variables and observed variables that
are continuous, censored, binary, ordered categorical (ordinal), counts,
censored inflated, and count inflated. The ON option has special
features for categorical latent variables and unordered categorical
(nominal) observed variables which are described below.

CATEGORICAL LATENT VARIABLES AND

UNORDERED CATEGORICAL (NOMINAL) OBSERVED
VARIABLES

For categorical latent variables and unordered categorical (nominal)

observed variables, the ON option is used to describe the multinomial
logistic regression of the categorical latent variable or the unordered
categorical (nominal) observed variable on one or more independent
variables.

For a categorical latent variable, an ON statement is specified for each

latent class except the last class which is the reference class. A class
label is used to refer to each class. Class labels use the convention of
adding to a variable name the number symbol (#) followed by a number.
For a categorical latent variable c with three classes,

c#1 c#2 ON x1-x3;

653
CHAPTER 17

specifies that regression coefficients are free to be estimated for classes

1 and 2 of the categorical latent variable c regressed on the independent
variables x1, x2, and x3. The intercepts in the regression of the
categorical latent variable on the independent variables are free to be
estimated as the default.

The statement above can be simplified to the following:

c ON x1-x3;

The multinomial logistic regression of one categorical latent variable on

another categorical latent variable where c2 has four classes and c1 has
three classes is specified as follows:

c2 ON c1;

or

c2#1 c2#2 c2#3 ON c1#1 c1#2;

For an unordered categorical (nominal) observed variable, an ON

statement is specified for each category except the last category which is
the reference category. A category label is used to refer to each
category. Category labels use the convention of adding to a variable
name the number symbol (#) followed by a number. For a three-
category variable u,

u#1 u#2 ON x1-x3;

specifies that regression coefficients are free to be estimated for

categories 1 and 2 of the unordered categorical (nominal) observed
variable u regressed on the independent variables x1, x2, and x3. The
thresholds in the regression of the unordered categorical (nominal)
observed variable on the independent variables are free to be estimated
as the default.

The statement above can be simplified to the following:

u ON x1-x3;

654
 MODEL Command

Following is a table that describes how the relationships between

dependent variables and observed mediating variables or latent variables
are specified. Relationships with the designation of NA are not allowed.
Relationships not specified using the ON option are specified by listing
for each class the intercepts or thresholds in square brackets and the
residual variances with no brackets. Not shown in the table is that all
dependent variables can be regressed on independent variables that are
not mediating variables using the ON option.

Scale of Scale of Observed Scale of Latent Variable

Dependent Mediating Variable
Variable Continuous Censored, Nominal Continuous Categorical Inflation
Categorical, Part of
and Count Censored
and Count
Continuous ON ON NA ON Mean and NA
variance vary
across classes
Censored, ON ON NA ON Mean/ NA
Categorical, threshold
and Count and variance
vary across
classes
Nominal ON ON NA ON Means vary NA
across classes
Continuous ON ON NA ON Mean and NA
Latent variance vary
across classes
Categorical ON ON NA ON ON NA
Latent
Inflation ON ON NA ON Mean varies NA
Part of across classes
Censored
and Count

PON
A second form of the ON option is PON. PON is used to describe the
paired regression relationships in the model and is short for regressed
on. PON pairs the variables on the left-hand side of the PON statement
with the variables on the right-hand side of the PON statement. For

655
CHAPTER 17

PON, the number of variables on the left-hand side of the PON statement
must equal the number of variables on the right-hand side of the PON
statement. For example,

y2 y3 y4 PON y1 y2 y3;

implies

y2 ON y1;
y3 ON y2;
y4 ON y3;

The PON option cannot be used with the simplified language for
categorical latent variables or unordered categorical (nominal) observed
variables.

WITH
The WITH option is used to describe correlational relationships in a
model and is short for correlated with. Correlational relationships
include covariances among continuous observed variables and
continuous latent variables and among categorical latent variables. With
the weighted least squares estimator, correlational relationships are also
allowed for binary, ordered categorical, and censored observed
variables. For all other variable types, the WITH option cannot be used
to specify correlational relationships. Special modeling needs to be used
in these situations, for example, using a latent variable that influences
both variables.

The NOCOVARIANCES setting of the MODEL option of the

ANALYSIS command specifies that the covariances and residual
covariances among all latent and observed variables in the analysis
model are fixed at zero. The WITH option is used to free selected
covariances and residual covariances.

Following is an example of how to specify the WITH option:

f1 WITH f2;

This statement frees the covariance parameter for the continuous latent
variables f1 and f2.

656
 MODEL Command

Several variables can be included on both sides of the WITH statement.

In this situation, the variables on the left-hand side of the WITH
statement are crossed with the variables on the right-hand side of the
WITH statement resulting in all possible combinations of left- and right-
hand side variables.

The association between two categorical latent variables c1 and c2

where c1 has three classes and c2 has four classes is specified as
follows:

c1#1 c1#2 WITH c2#1 c2#2 c2#3;

The statement above can be simplified to:

c1 WITH c2;

The association coefficient for the last class of each categorical latent
variable is fixed at zero as the default as in loglinear modeling.

PWITH
A second form of the WITH option is PWITH. PWITH pairs the
variables on the left-hand side of the PWITH statement with those on the
right-hand side of the PWITH statement. For PWITH, the number of
variables on the left-hand side of the PWITH statement must equal the
number of variables on the right-hand side of the PWITH statement. For
example,

y1 y2 y3 PWITH y4 y5 y6;

implies

y1 WITH y4;
y2 WITH y5;
y3 WITH y6;

whereas,

y1 y2 y3 WITH y4 y5 y6;

implies

657
CHAPTER 17

y1 WITH y4;
y1 WITH y5;
y1 WITH y6;
y2 WITH y4;
y2 WITH y5;
y2 WITH y6;
y3 WITH y4;
y3 WITH y5;
y3 WITH y6;

The PWITH option cannot be used with the simplified language for
categorical latent variables.

VARIANCES/RESIDUAL VARIANCES
For convenience, no distinction is made in how variances and residual
variances are referred to in the MODEL command. The model defines
whether the parameter to be estimated is a variance or a residual
variance. Variances are estimated for independent variables and residual
variances are estimated for dependent variables. Variances of
continuous and censored observed variables and continuous latent
variables are free to be estimated as the default. Variances of categorical
observed variables are not estimated. When the Theta parameterization
is used in either a growth model or a multiple group model, variances for
continuous latent response variables for the categorical observed
variables are estimated. Unordered categorical (nominal) observed
variables, observed count variables, and categorical latent variables have
no variance parameters.

A list of observed or latent variables refers to the variances or residual

variances of those variables. For example,

y1 y2 y3;

refers to the variances of y1, y2, and y3 if they are independent variables
and refers to the residual variances of y1, y2, and y3 if they are
dependent variables. The statement means that the variances or residual
variances are free parameters to be estimated using default starting
values.

658
 MODEL Command

MEANS/INTERCEPTS/THRESHOLDS
Means, intercepts, and thresholds are included in the analysis model as
the default. The NOMEANSTRUCTURE setting of the MODEL option
of the ANALYSIS command is used with TYPE=GENERAL to specify
that means, intercepts, and thresholds are not included in the analysis
model.

For convenience, no distinction is made in how means and intercepts are

referred to in the MODEL command. The model defines whether the
parameter to be estimated is a mean or an intercept. Means are
estimated for independent observed variables and observed variables that
are neither independent nor dependent variables in the model. Means for
nominal variables are logit coefficients corresponding to probabilities
for each category except the last category. Means for count variables are
log rates. Means for time-to-event variables in continuous-time survival
analysis are log rates. Means are also estimated for independent
continuous latent variables and independent categorical latent variables.
For an independent categorical latent variable, the means are logit
coefficients corresponding to probabilities for each class except the last
class.

Intercepts are estimated for continuous observed dependent variables,

censored observed dependent variables, unordered categorical (nominal)
observed dependent variable, count observed dependent variables,
baseline hazard parameters for continuous-time survival analysis,
continuous latent dependent variables, and categorical latent dependent
variables.

Thresholds are estimated for binary and ordered categorical observed

variables. The sign of a threshold is the opposite of the sign of a mean
or intercept for the same variable. For example, with a binary dependent
variable, a threshold of -0.5 is the same as an intercept of .5.

A list of observed or latent variables enclosed in brackets refers to

means, intercepts, or thresholds.

For example,

[y1 y2 y3];

659
CHAPTER 17

refers to the means of variables y1, y2, and y3 if they are independent
variables and refers to the intercepts if they are continuous dependent
variables. This statement indicates that the means or intercepts are free
parameters to be estimated using the default starting values.

If the variables are categorical, the thresholds are referred to as follows,

[y1$1 y1$2 y1$3 y2$1 y2$2];

where y1 is a four category variable with three thresholds and y2 is a

three category variable with two thresholds. y1$1 refers to the first and
lowest threshold of variable y1; y1$2 refers to the next threshold; and
y1$3 refers to the highest threshold. This statement means that the
thresholds are free parameters to be estimated using the default starting
values.

For models with a mean structure, all means, intercepts, and thresholds
of observed variables are free to be estimated at the default starting
values. The means and intercepts of continuous latent variables are
fixed at zero in a single group analysis. In a multiple group analysis, the
means and intercepts of the continuous latent variables are fixed at zero
in the first group and are free to be estimated in the other groups. In a
mixture model, the means and intercepts of the continuous latent
variables are fixed at zero in the last class and are free to be estimated in
the other classes. The means and intercepts of categorical latent
variables are fixed at zero in the last class and are free to be estimated in
the other classes.

CONVENIENCE FEATURES FOR THE MODEL

COMMAND
There are several features that make it easier for users to specify the
model to be estimated. One feature is the list function. A user can use a
hyphen to specify a list of variables, a list of equality constraints, and a
list of parameter labels.

When using the list function, it is important to know the order of

observed and latent variables that the program expects. The order of
observed variables is determined by the order of variables in the
NAMES or USEVARIABLES options of the VARIABLE command. If

660
 MODEL Command

all of the variables in NAMES statement are used in the analysis, then
the order is taken from there. If the variables for the analysis are a
subset of the variables in the NAMES statement, the order is taken from
the USEVARIABLES statement.

The order of continuous latent variables is determined by the order of

the BY and | statements in the MODEL command. Factors defined using
the BY option come first in the order that they occur in the MODEL
command followed by the random effects defined using the | symbol in
the order that they occur in the MODEL command.

The list function can be used on the left- and right-hand sides of ON and
WITH statements and on the right-hand side of BY statements. A list on
the left-hand side implies multiple statements. A list on the right-hand
side implies a list of variables.

Following is an example of the use of the list function on the right-hand

side of a BY statement. It assumes the variables are in the order: y1, y2,
y3, y4, y5, y6, y7, y8, y9, y10, y11, and y12.

f1 BY y1-y4;
f2 BY y5-y9;
f3 BY y10-y12;

The program would interpret these BY statements as:

f1 BY y1 y2 y3 y4;
f2 BY y5 y6 y7 y8 y9;
f3 BY y10 y11 y12;

To use the list function with latent variables the order of latent variables
would be f1, f2, f3 because of the order of the BY statements in the
MODEL command.

Following is an example of using the list function on both the left- and
the right-hand sides of the ON statement:

f1-f3 ON x1-x3;

This implies the multiple statements:

661
CHAPTER 17

f1 ON x1 x2 x3;
f2 ON x1 x2 x3;
f3 ON x1 x2 x3;

The list function can also be used with the WITH option,

y1-y3 WITH y4-y6;

This implies

y1 WITH y4;
y1 WITH y5;
y1 WITH y6;
y2 WITH y4;
y2 WITH y5;
y2 WITH y6;
y3 WITH y4;
y3 WITH y5;
y3 WITH y6;

FREEING PARAMETERS AND ASSIGNING

STARTING VALUES
The asterisk (*) is used to free a parameter and/or assign a starting value
for the estimation of that parameter. It is placed after a parameter with a
number following it. For example:

y1*.5;

is interpreted as freeing the variance/residual variance of y1 to be

estimated with a starting value of 0.5.

Consider the BY statements from the previous section:

f1 BY y1 y2 y3 y4 y5;
f2 BY y6 y7 y8 y9;

As mentioned previously, the above statements result in the factor

loadings for y1 and y6 being fixed at one in order to set the metric of the
latent variables f1 and f2. All of the other parameters mentioned are free

662
 MODEL Command

to be estimated with starting values of one. Consider the following BY

statements:

f1 BY y1* y2*0.5 y3 y4 y5;

f2 BY y6 y7-y9*0.9;

By putting an asterisk (*) after y1, the y1 parameter is freed at the

default starting value of one instead of being fixed at one by default. By
placing an asterisk (*) followed by 0.5 behind y2, the parameter starting
value is changed from the default starting value of 1 to 0.5. The
variables y3, y4, and y5 are free to be estimated at the default starting
value of one. In the BY statement for f2 the variables y7, y8, and y9 are
specified using the list function, y7-y9, followed by an asterisk (*) and
the value 0.9. This changes the starting values for y7, y8, and y9 from
the default starting value of 1 to 0.9.

These same features can be used with the ON and WITH options and for
assigning starting values to variances, means, thresholds, and scales.

Following are examples of assigning starting values to a variety of

parameters:

f1 ON x1-x3*1.5;
f1 WITH f2*.8;
y1-y12*.75;
[f1-f3*.5];
{y1-y12*5.0};

FIXING PARAMETER VALUES

In some cases, it is necessary to fix a parameter at a specific value. The
@ symbol is used to fix the values of parameters. Consider the
following example based on the measurement model in the earlier figure.
Following are the specifications needed to free the value of the first
indicator of each latent variable at starting values of one and to fix the
value of the second indicator of each latent variable to one in order to set
the metric of each latent variable:

f1 BY y1* y2@1 y3 y4 y5;

f2 BY y6* y7@1 y8 y9;

663
CHAPTER 17

By placing an asterisk (*) after y1, the factor loading for y1 is estimated
using the starting value of one. By placing @1 after y2, the factor
loading for y2 is fixed at one. Likewise, by placing an asterisk (*) after
y6, the factor loading for y6 is estimated using the starting value of one.
By placing @1 after y7, the factor loading for y7 is fixed at one.

The @ symbol can be used to fix any parameter in a model. The

following example fixes the covariance between f1 and f2 at zero:

f1 WITH f2@0;

CONSTRAINING PARAMETER VALUES TO BE

EQUAL
Parameters can be constrained to be equal by placing the same number in
parentheses following the parameters that are to be held equal. This
convention can be used for all parameters. Following is an example in
which regression coefficients, residual variances, and residual
covariances are held equal:

y1 ON x1 (1) ;
y2 ON x2 (1) ;
y3 ON x3 (1) ;
y1 y2 y3 (2);
y1 WITH y2-y3 (3);

In the above example, the regression coefficients for the three

regressions are constrained to be equal, the three residual variances are
constrained to be equal, and the two residual covariances are constrained
to be equal.

There can be only one number in parentheses on each line. If a

statement continues on more than one line, the number in parentheses
must be stated at the end of each line.

For example,

f1 BY y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12 (1)

y13 y14 y15 (1);

664
 MODEL Command

specifies that the factor loadings of y2 through y15 are constrained to be

equal. The factor loading of y1 is fixed at one as the default.

The following statement,

f1 BY y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 y11 y12

y13 y14 y15 (1);

specifies that the factor loadings of y13, y14, and y15 are constrained to
be equal because (1) refers to only the information on the line on which
it is located.

The following statement,

f1 BY y1 y2 y3 y4 y5 (1)
y6 y7 y8 y9 y10 (2)
y11 y12 y13 y14 y15 (3);

specifies that the factor loading of y1 is fixed at one and that the factor
loadings of y2, y3, y4, and y5 are held equal, that the factor loadings of
y6, y7, y8, y9, and y10 are held equal, and that the factor loadings of
y11, y12, y13, y14, and y15 are held equal.

Following are examples of how to constrain the parameters of means,

intercepts and/or thresholds to be equal.

[y1 y2 y3] (1);

indicates that the means/intercepts of variables y1, y2, and y3 are

constrained to be equal. The statements

[u1$1 u2$1 u3$1] (2);

[u1$2 u2$2 u3$2] (3);
[u1$3 u2$3 u3$3] (4);

indicate that the first threshold for variables u1, u2, and u3 are
constrained to be equal; that the second threshold for variables u1, u2,
and u3 are constrained to be equal; and that the third threshold for
variables u1, u2, and u3 are constrained to be equal. Out of nine
possible thresholds, three parameters are estimated. Only one set of
parentheses can be included on each line of the input file.

665
CHAPTER 17

USING THE LIST FUNCTION FOR ASSIGNING

STARTING VALUES, FIXING VALUES,
CONSTRAINING VALUES TO BE EQUAL, AND
ASSIGNING LABELS TO PARAMETERS
The list function is convenient for assigning starting values to
parameters, fixing parameters values, constraining parameter values to
be equal, and assigning labels to parameters.

ASSIGNING STARTING VALUES TO PARAMETERS

Following is an example of how to use the list function to assign starting

values to parameters:

f1 BY y1 y2-y4*0;
f2 BY y5 y6-y9*.5;
f3 BY y10 y11-y12*.75;

The program interprets these BY statements as:

f1 BY y1 y2*0 y3*0 y4*0;

f2 BY y5 y6*.5 y7*.5 y8*.5 y9*.5;
f3 BY y10 y11*.75 y12*.75;

where the starting value of 0 is assigned to the factor loadings for y2, y3,
and y4; the starting value of .5 is assigned to the factor loadings for y6,
y7, y8, and y9; and the starting value of .75 is assigned to the factor
loadings for y11 and y12. The factor loading for the first factor
indicator of each factor is fixed at one as the default to set the metric of
the factor.

FIXING PARAMETER VALUES

Following is an example of using the list function to fix parameter

values:

f1-f3@1;

The statement above fixes the variances/residual variances of f1, f2, and
f3 at one.

666
 MODEL Command

CONSTRAINING PARAMETER VALUES TO BE EQUAL

Following is an example of using the list function to constrain parameter

values to be equal:

f1 BY y1-y5 (1)
y6-y10 (2);

The statement above specifies that the factor loadings of y2, y3, y4, and
y5 are held equal and that the factor loadings of y6, y7, y8, y9, and y10
are held equal. The factor loading of y1 is fixed at one as the default to
set the metric of the factor.

The list function can be used to assign equalities to a list of parameters

using a list of equality constraints. A list of equality constraints cannot
be used with a set of individual parameters. Following is an example of
how to use the list function with a list of parameters on the right-hand
side of the BY option:

f1 BY y1
y2-y5 (2-5);
f2 BY y6
y7-y10 (2-5);

The statements above specify that the factor loadings for y2 and y7 are
held equal, the factor loadings for y3 and y8 are held equal, the factor
loadings for y4 and y9 are held equal, and the factor loadings for y5 and
y10 are held equal. This can also be specified as shown below for
convenience.

f1 BY y1-y5 (1-5);
f2 BY y6-y10 (1-5);

No equality constraint is assigned to y1 and y6 even though they are part

of the list of variables because they are fixed at one to set the metric of
the factors. The number of equalities in the list must equal the number
of variables on the right-hand side of the BY option.

A list of equality constraints cannot be used with a list of parameters on

the left-hand side of an option. Following is an example of how equality

667
CHAPTER 17

constraints are specified for a list of parameters on the left-hand side of

the ON option:

y1-y3 ON x (1 2 3);
y4-y6 ON x (1 2 3);

where the regression coefficients in the regression of y1 on x and y4 on

x are constrained to be equal; the regression coefficients in the
regression of y2 on x and y5 on x are constrained to be equal; and the
regression coefficients in the regression of y3 on x and y6 on x are
constrained to be equal.

Following is an example of how equalities are specified when a list of

parameters appears on both the left- and right-hand sides of an option:

y1-y3 ON x1-x2 (1-2 3-4 5-6);

y4-y6 ON x1-x2 (1-2 3-4 5-6);

Each variable on the left-hand side of the ON option must have a list of
equalities for use with the variables on the right-hand side of the ON
option. Because there are three variables on the left-hand side of the ON
statement and two variables on the right-hand side of the ON statement,
three lists of two equalities are needed. A single list cannot be used.

Following is what this specifies:

y1 ON x1 (1);
y1 ON x2 (2);
y2 ON x1 (3);
y2 ON x2 (4);
y3 ON x1 (5);
y3 ON x2 (6);
y4 ON x1 (1);
y4 ON x2 (2);
y5 ON x1 (3);
y5 ON x2 (4);
y6 ON x1 (5);
y6 ON x2 (6);

The list function can be used with the simplified language for categorical
latent variables and unordered categorical (nominal) observed variables.

668
 MODEL Command

The multinomial logistic regression of one categorical latent variable on

another categorical latent variable where c2 has four classes and c1 has
three classes is specified as follows:

c2 ON c1;

Following is an example of how equalities are specified when a list of

parameters appears on both the left- and right-hand sides of the ON
option using the simplified language:

c2 ON c1 (1-2 1-2 1-2);

or

c2 ON c1 (1-2);

Following is what this specifies:

c2#1 ON c1#1 (1);

c2#2 ON c1#2 (2);
c2#3 ON c1#1 (1);
c2#1 ON c1#2 (2);
c2#2 ON c1#1 (1);
c2#3 ON c1#3 (2);

ASSIGNING LABELS TO PARAMETERS

The list function can be used to assign labels to parameters in the

MODEL command. Following is an example of how to use the list
function in this way:

[y1-y5] (p1-p5);

The statement above assigns the parameter label p1 to y1, p2 to y2, p3 to

y3, p4 to y4, and p5 to y5.

The list function can be used to assign labels to a list of parameters using
a list of labels. A list of labels cannot be used with a set of individual
parameters. Following is an example of how to use the list function with
a list of parameters on the right-hand side of the BY option:

669
CHAPTER 17

f1 BY y1
y2-y5 (p2-p5);

The statement above assigns the label p2 to the factor loading for y2, p3
to the factor loading for y3, p4 to the factor loading for y4, and p5 to the
factor loading for y5.

A list of labels cannot be used with a list of parameters on the left-hand

side of an option. Following is an example of how labels are specified
for a list of parameters on the left-hand side of the ON option:

y1-y3 ON x (p1 p2 p3);

where the regression coefficient in the regression of y1 on x is assigned

the label p1; the regression coefficient in the regression of y2 on x is
assigned the label p; and the regression coefficient in the regression of
y3 on x is assigned the label p3.

Following is an example of how labels are specified when a list of

parameters appears on both the left- and right-hand sides of an option:

y1-y3 ON x1-x2 (p1-p2 p3-p4 p5-p6);

Each variable on the left-hand side of the ON option must have a list of
labels for use with the variables on the right-hand side of the ON option.
Because there are three variables on the left-hand side of the ON
statement and two variables on the right-hand side of the ON statement,
three lists of two equalities are needed. A single list cannot be used.

Because there are three variables on the left-hand side of the ON

statement and two variables on the right-hand side of the ON statement,
three lists of two equalities are needed.

Following is what this specifies:

y1 ON x1 (p1);
y1 ON x2 (p2);
y2 ON x1 (p3);
y2 ON x2 (p4);
y3 ON x1 (p5);
y3 ON x2 (p6);

670
 MODEL Command

The list function can be used with the simplified language for categorical
latent variables and unordered categorical (nominal) observed variables.

The multinomial logistic regression of one categorical latent variable on

another categorical latent variable where c2 has four classes and c1 has
three classes is specified as follows:

c2 ON c1;

Following is an example of how labels are assigned when a list of

parameters appears on both the left- and right-hand sides of the ON
option using the simplified language:

c2 ON c1 (p1-p2 p3-p4 p5-p6);

Following is what this specifies:

c2#1 ON c1#1 (p1);

c2#2 ON c1#2 (p2);
c2#3 ON c1#1 (p3);
c2#1 ON c1#2 (p4);
c2#2 ON c1#1 (p5);
c2#3 ON c1#3 (p6);

SPECIAL LIST FUNCTION FEATURE

The list function has a special feature that can make model specification
easier. This feature allows a parameter to be mentioned in the MODEL
command more than once. The last specification is used in the analysis.
For example,

f1 BY y1-y6*0 y5*.5;

is interpreted by the program as

f1 BY y1*0 y2*0 y3*0 y4*0 y5*.5 y6*0;

Although y5 is assigned a starting value of 0.0 in the beginning of the

BY statement using the list function, y5 is assigned a starting value of
0.5 later in the statement. The program uses the last specification.

671
CHAPTER 17

If a variable is mentioned more than once on the right-hand side of a BY,

ON, or WITH statement or in a list of variances, means, or scale factors,
the program uses the last value it reads. This makes it convenient when
a user wants all of the starting values in a list to be the same except for a
few. The same feature can be used when fixing values. For example,

f1-f4@1 f3@2;

fixes the variances/residual variances of f1, f2, and f4 at one and fixes
the variance/residual variance of f3 at 2.

This feature can also be used with equalities, however, the variable from
the list that is not to be constrained to be equal must appear on a separate
line in the input file. In a line with an equality constraint, anything after
the equality constraint is ignored. For example,

f1 BY y1-y5 (1)
y4
y6-y10 (2);

indicates that the factor loadings for y2, y3, and y5 are held equal, the
factor loading for y4 is free and not equal to any other factor loading,
and the factor loadings for y6, y7, y8, y9, and y10 are held equal. The
factor loading for y1 is fixed at one as the default.

LABELING THRESHOLDS
For binary and ordered categorical dependent variables, thresholds are
referred to by using the convention of adding to a variable name a dollar
sign ($) followed by a number. The number of thresholds is equal to the
number of categories minus one. For example, if u1 is an ordered
categorical variable with four categories it has three thresholds. These
thresholds are referred to as u1$1, u1$2, and u1$3.

LABELING CATEGORICAL LATENT

VARIABLES AND UNORDERED CATEGORIAL
(NOMINAL) OBSERVED VARIABLES
The classes of categorical latent variables and the categories of
unordered categorical (nominal) observed variables are referred to by

672
 MODEL Command

using the convention of adding to a variable name a number sign (#)

followed by the category/class number. For example, if c is a
categorical latent variable with three classes, the first two classes are
referred to as c#1 and c#2. The third class has all parameters fixed at
zero as a reference category. If u1 is a nominal variable with three
categories, the first two categories are referred to as u1#1 and u1#2. The
third category has all parameters fixed at zero as a reference category.
With the ON option categorical latent variables and unordered
categorical (nominal) observed variables can be referred to by their
variable name. With the WITH option, categorical latent variables can
be referred to by their variable name.

LABELING INFLATION VARIABLES

Censored and count inflation variables are referred to by using the
convention of adding to a variable name a number sign (#) followed by
the number one. For example, if y1 is a censored variable, the inflation
part of y1 is referred to as y1#1. If u1 is a count variable, the inflation
part of u1 is referred to as u1#1.

LABELING BASELINE HAZARD PARAMETERS

In continuous-time survival modeling, there are as many baseline hazard
parameters are there are time intervals plus one. When the
BASEHAZARD option of the ANALYSIS command is ON, these
parameters can be referred to by using the convention of adding to the
name of the time-to-event variable the number sign (#) followed by a
number. For example, for a time-to-event variable t with 5 time
intervals, the six baseline hazard parameters are referred to as t#1, t#2,
t#3, t#4, t#5, and t#6.

LABELING CLASSES OF A CATEGORICAL

LATENT VARIABLE
In the MODEL command, categorical latent variable classes are referred
to using labels. These labels are constructed by using the convention of
adding to the name of the categorical latent variable a number sign (#)
followed by a number. For example, if c is a categorical latent variable
with four classes, the labels for the four classes are c#1, c#2, and c#3.
The last class is the reference class.

673
CHAPTER 17

LABELING PARAMETERS
Labels can be assigned to parameters by placing a name in parentheses
following the parameter in the MODEL command. These labels are
used in three ways. First, they are used in conjunction with the MODEL
CONSTRAINT command to define linear and non-linear constraints on
the parameters in the model. Second, they are used with the MODEL
TEST command to test linear restrictions on the model defined in the
MODEL and MODEL CONSTRAINT commands. Third, they are used
with ESTIMATOR=BAYES and the MODEL PRIORS command to
specify the prior distribution for parameters in the MODEL command.

The parameter labels follow the same rules as variable names. They can
be up to 8 characters in length; must begin with a letter; can contain only
letters, numbers, and the underscore symbol; and are not case sensitive.
Only one label can appear on a line. Following is an example of how to
label parameters:

MODEL: y ON x1 (p1)
x2 (p2)
x3 (p3);

where p1 is the label assigned to the regression slope for y on x1, p2 is

the label assigned to the regression slope for y on the x2, and p3 is the
label assigned to the regression slope for y on x3.

The list function can be used to assign labels. Following is an example

of how to use the list function to label parameters:

MODEL: [y1-y10] (q1-q10);

y1-y10 (p1-p10);
f BY y1-y10 (z2-z10);

where the labels q1 through q10 are assigned to the intercepts of y1

through y10, the labels p1 through p10 are assigned to the residual
variances of y1 through y10, and the labels z2 through z10 are assigned
to the factors loadings for y2 through y10. The factor loading of y1 is
fixed at one as the default to set the metric of the factor.

If a list of labels is used, for example, in MODEL PRIORS, the order of

the labels is alphabetical not the order of the labels in the MODEL

674
 MODEL Command

command. For example, the list p4-q2 includes p4, p5, p6, p7, p8, p9,
p10, q1, and q2.

The list function can be used with the ON and WITH options when there
are lists of variable names on both the right- and left-hand sides of these
options. Following is an example of how to use the list function to
assign labels when there are lists of variables on both the right- and left-
hand sides of ON:

y1-y3 ON x1-x2 (p1-p6);

The first variable on the left-hand side of ON is paired with all variables
on the right-hand side. Then the second variable on the left-hand side of
ON is paired with all variables on the right-hand side etc. The label p1
is assigned to the regression slope for y1 on x1. The label p2 is assigned
to the regression slope for y1 on x2. The label p3 is assigned to the
regression slope for y2 on y1. The label p6 is assigned to the regression
slope for y3 on x2.

Following is an example of how to use the list function to assign labels

when there are lists of variables on both the right- and left-hand sides of
WITH:

y1-y3 WITH y1-y3 ( p1-p3);

The labels are assigned to the upper triangle of a symmetric matrix read
row-wise. The label p1 is assigned to the covariance between y1 and y2.
The label p2 is assigned to the covariance between y1 and y3. The label
p3 is assigned to the covariance between y2 and y3.

SCALE FACTORS
In models that use TYPE=GENERAL, it may be useful to multiply each
observed variable or latent response variable by a scale factor that can be
estimated. For example, with categorical observed variables, a scale
factor refers to the underlying latent response variables and facilitates
growth modeling and multiple group analysis because the latent response
variables are not restricted to have across-time or across-group equalities
of variances. With continuous observed variables, using scale factors
containing standard deviations makes it possible to analyze a sample
covariance matrix by a correlation structure model.

675
CHAPTER 17

A list of observed variables in curly brackets refers to scale factors. For

example,

{u1 u2 u3};

refers to scale factors for variables u1, u2, and u3. This statement means
that the scale factors are free parameters to be estimated using the
default starting values of one.

The | SYMBOL
The | symbol is used to name and define random effect variables in the
model. It can be used with all analysis types to specify growth models.
It can be used with TYPE=RANDOM to name and define random effect
variables that are slopes, to specify growth models with individually-
varying times of observation, and to specify latent variable interactions.

GROWTH MODELS

Following is a description of the language specific to growth models.

The | symbol can be used with all analysis types to specify growth
models. The names on the left-hand side of the | symbol name the
random effect variables, also referred to as growth factors. The
statement on the right-hand side of the | symbol names the outcome and
specifies the time scores for the growth model.

Following is an example of the MODEL command for a quadratic

growth model for a continuous outcome specified without using the |
symbol:

MODEL: i BY y1-y4@1;
s BY y1@0 y2@1 y3@2 y4@3;
q BY y1@0 y2@1 y3@4 y4@9;
[y1-y4@0 i s q];

If the | symbol is used to specify the same growth model for a continuous
outcome, the MODEL command is:

MODEL: i s q | y1@0 y2@1 y3@2 y4@3;

676
 MODEL Command

All of the other specifications shown above are done as the default. The
defaults can be overridden by mentioning the parameters in the MODEL
command after the | statement. For example,

MODEL: i s q | y1@0 y2@1 y3@2 y4@3;

[y1-y4] (1);
[i@0 s q];

changes the parameterization of the growth model from one with the
intercepts of the outcome variable fixed at zero and the growth factor
means free to be estimated to a parameterization with the intercepts of
the outcome variable held equal, the intercept growth factor mean fixed
at zero, and the slope growth factor means free to be estimated.

Many other types of growth models can be specified using the | symbol.
Following is a table that shows how to specify some of these growth
models using the | symbol and also how to specify the same growth
models using the BY option and other options. All examples are for
continuous outcomes unless specified otherwise.

Growth Language Alternative

Intercept only MODEL: MODEL:
i | y1-y4; i BY y1-y4@1;
[y1-y4@0 i];
Linear MODEL: MODEL:
i s | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
s BY y1@0 y2@1 y3@2 y4@3;
[y1-y4@0 i s];
Linear with free MODEL: MODEL:
time scores i s | y1@0 y2@1 y3 y4; i BY y1-y4@1;
s BY y1@0 y2@1 y3 y4;
[y1-y4@0 i s];

677
CHAPTER 17

Quadratic MODEL: MODEL:

i s q | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
s BY y1@0 y2@1 y3@2 y4@3;
q BY y1@0 y2@1 y3@4 y4@9;
[y1-y4@0 i s q];
Piecewise MODEL: MODEL:
i s1 | y1@0 y2@1 y3@2 y4@2 y5@2; i BY y1-y4@1;
i s2 | y1@0 y2@0 y3@0 y4@1 y5@2; s1 BY y1@0 y2@1 y3@2 y4@2 y5@2;
s2 BY y1@0 y2@0 y3@0 y4@1 y5@2;
[y1-y4@0 i s1 s2];
Linear for a MODEL: MODEL:
censored i s | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
outcome s BY y1@0 y2@1 y3@2 y4@3;
[y1-y4@0 i s];
Linear for a MODEL: MODEL:
censored i s | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
outcome and the ii si |y1#1@0 y2#1@1 y3#1@2 s BY y1@0 y2@1 y3@2 y4@3;
inflation part of y4#1@3; ii BY y1#1-y4#1@1;
a censored si BY y1#1@0 y2#1@1 y3#1@2
outcome y4#1@3;
[y1-y4@0 i s];
[y1#1-y4#1] (1);
[ii@0 si];
Linear for a MODEL: MODEL:
binary outcome i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
with the Delta s BY u1@0 u2@1 u3@2 u4@3;
parameterization [u1$1-u4$1] (1);
[i@0 s];
{u1@1 u2-u4};
Linear for a MODEL: MODEL:
binary outcome i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
with the Theta s BY u1@0 u2@1 u3@2 u4@3;
parameterization [u1$1-u4$1] (1);
[i@0 s];
u1@1 u2-u4;
Linear for a MODEL: MODEL:
binary outcome i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
with the logistic s BY u1@0 u2@1 u3@2 u4@3;
model [u1$1-u4$1] (1);
[i@0 s];

678
 MODEL Command

Linear for a MODEL: MODEL:

count outcome i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
s BY u1@0 u2@1 u3@2 u4@3;
[u1-u4@0 i s];
Linear for a MODEL: MODEL:
count outcome i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
and the inflation ii si | u1#1@0 u2#1@1 u3#1@2 s BY u1@0 u2@1 u3@2 u4@3;
part of a count u4#1@3; ii BY u1#1-u4#1@1;
outcome si BY u1#1@0 u2#1@1 u3#1@2
u4#1@3;
[u1-u4@0 i s];
[u1#1-u4#1] (1);
[ii@0 si];
Multiple group MODEL: MODEL:
i s | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
s BY y1@0 y2@1 y3@2 y4@3;
[y1-y4@0 i s];
MODEL g1:
[i s];
Multiple group MODEL: MODEL:
for a binary i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
outcome with s BY u1@0 u2@1 u3@2 u4@3;
the Delta [u1$1-u4$1] (1);
parameterization {u1@1};
MODEL g1:
[s];
{u2-u4};
Multiple group MODEL: MODEL:
for a three- i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
category s BY u1@0 u2@1 u3@2 u4@3;
outcome with [u1$1-u4$1] (1);
the Delta [u1$2-u4$2] (2);
parameterization MODEL g1:
[s];
{u2-u4};

679
CHAPTER 17

Multiple group MODEL: MODEL:

for a binary i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
outcome with s BY u1@0 u2@1 u3@2 u4@3;
the Theta [u1$1-u4$1] (1);
parameterization u1@1;
MODEL g1:
[s];
u2-u4;
Multiple group MODEL: MODEL:
for a three- i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
category s BY u1@0 u2@1 u3@2 u4@3;
outcome with [u1$1-u4$1] (1);
the Theta [u1$2-u4$2] (2);
parameterization MODEL g1:
[s];
u2-u4;
Mixture MODEL: MODEL:
%OVERALL% %OVERALL%
i s | y1@0 y2@1 y3@2 y4@3; i BY y1-y4@1;
s BY y1@0 y2@1 y3@2 y4@3;
[y1-y4@0 i s];
%c#1%
[i s];
Mixture for a MODEL: MODEL:
binary outcome %OVERALL% %OVERALL%
i s | u1@0 u2@1 u3@2 u4@3; i BY u1-u4@1;
s BY u1@0 u2@1 u3@2 u4@3;
[u1$1-u4$1] (1);
[ i s];
%c#1%
[i s];
%c#2%
[i@0 ];

680
 MODEL Command

Multilevel MODEL: MODEL:

%WITHIN% %WITHIN%
iw sw | y1@0 y2@1 y3@2 y4@3; iw BY y1-y4@1;
%BETWEEN% sw BY y1@0 y2@1 y3@2 y4@3;
ib sb | y1@0 y2@1 y3@2 y4@3; %BETWEEN%
ib BY y1-y4@1;
sb BY y1@0 y2@1 y3@2 y4@3;
[y1-y4@0 ib sb];
Multiple MODEL: MODEL:
indicator f1 BY y11 f1 BY y11
y21 (1); y21 (1);
f2 BY y12 f2 BY y12
y22 (1); y22 (1);
f3 BY y13 f3 BY y13
y23 (1); y23 (1);
f4 BY y14 f4 BY y14
y24 (1); y24 (1);
[y11 y12 y13 y14] (2); [y11 y12 y13 y14] (2);
[y21 y22 y23 y24] (3); [y21 y22 y23 y24] (3);
i s | f1@0 f2@1 f3@2 f4@3; i BY f1-f4@1;
s BY f1@0 f2@1 f3@2 f4@3;
[f1-f4@0 i@0 s];
Multiple MODEL: MODEL:
indicator for a f1 BY u11 f1 BY u11
binary outcome u21 (1); u21 (1);
with the Delta f2 BY u12 f2 BY u12
parameterization u22 (1); u22 (1);
f3 BY u13 f3 BY u13
u23 (1); u23 (1);
f4 BY u14 f4 BY u14
u24 (1); u24 (1);
[u11$1 u12$1 u13$1 u14$1] (2); [u11$1 u12$1 u13$1 u14$1] (2);
[u21$1 u22$1 u23$1 u24$1] (3); [u21$1 u22$1 u23$1 u24$1] (3);
{u11-u21@1 u12-u24}; {u11-u21@1 u12-u24};
i s | f1@0 f2@1 f3@2 f4@3; i BY f1-f4@1;
s BY f1@0 f2@1 f3@2 f4@3;
[f1-f4@0 i@0 s];

681
CHAPTER 17

Multiple MODEL: MODEL:

indicator for a f1 BY u11 f1 BY u11
binary outcome u21 (1); u21 (1);
with the Theta f2 BY u12 f2 BY u12
parameterization u22 (1); u22 (1);
f3 BY u13 f3 BY u13
u23 (1); u23 (1);
f4 BY u14 f4 BY u14
u24 (1); u24 (1);
[u11$1 u12$1 u13$1 u14$1] (2); [u11$1 u12$1 u13$1 u14$1] (2);
[u21$1 u22$1 u23$1 u24$1] (3); [u21$1 u22$1 u23$1 u24$1] (3);
u11-u21@1 u12-u24; u11-u21@1 u12-u24;
i s | f1@0 f2@1 f3@2 f4@3; i BY f1-f4@1;
s BY f1@0 f2@1 f3@2 f4@3;
[f1-f4@0 i@0 s];

The defaults for the means/intercepts of the growth factors vary

depending on the scale of the outcome variable as described below. The
variances/residual variances and covariances/residual covariances of
growth factors are free to be estimated for all outcomes as the default.

For continuous, censored, and count outcomes, the means/intercepts of

the growth factors are free to be estimated. For a binary outcome, an
ordered categorical (ordinal) outcome, the inflation part of a censored
outcome, the inflation part of a count outcome, and a multiple indicator
growth model, the mean/intercept of the intercept growth factor is fixed
at zero. The means/intercepts of the slopes growth factors are free to be
estimated.

In multiple group analysis for continuous, censored, and count outcomes,

the means/intercepts of the growth factors are free to be estimated in all
groups. In multiple group analysis for a binary outcome, an ordered
categorical (ordinal) outcome, the inflation part of a censored outcome,
the inflation part of a count outcome, and a multiple indicator growth
model, the mean/intercept of the intercept growth factor is fixed at zero
in the first group and is free to be estimated in the other groups. The
means/intercepts of the slopes growth factors are free to be estimated in
all groups.

682
 MODEL Command

In mixture models for continuous, censored, and count outcomes, the

means/intercepts of the growth factors are free to be estimated in all
classes. In mixture models for a binary outcome, an ordered categorical
(ordinal) outcome, the inflation part of a censored outcome, the inflation
part of a count outcome, and a multiple indicator growth model, the
mean/intercept of the intercept growth factor is fixed at zero in the last
class and is free to be estimated in the other classes. The
means/intercepts of the slopes growth factors are free to be estimated in
all classes.

The residual variances of continuous and censored outcome variables are

free as the default. The inflated part of censored outcomes, binary
outcomes, ordered categorical (ordinal) outcomes, count outcomes, and
the inflated part of count outcomes have no variance parameters. An
exception is the Theta parameterization used for binary and ordered
categorical (ordinal) outcomes. In the Theta parameterization, residual
variances are fixed at one at the first time point and are free at the other
time points.

RANDOM SLOPES

The | symbol is used in conjunction with TYPE=RANDOM to name and

define the random slope variables in the model. The name on the left-
hand side of the | symbol names the random slope variable. The
statement on the right-hand side of the | symbol defines the random slope
variable. Random slopes are defined using the ON option. ON
statements used on the right-hand side of the | symbol may not use the
asterisk (*) or @ symbols. Otherwise, the regular rules regarding ON
apply. The means and the variances of the random slope variables are
free as the default. Covariances among random slope variables are fixed
at zero as the default. Covariances between random slope variables and
growth factors, latent variables defined using BY statements, and
observed variables are fixed at zero as the default.

With TYPE=TWOLEVEL RANDOM and TYPE=CROSSCLASSIFIED

RANDOM, the random slope variables are named and defined in the
within part of the MODEL command and used in the between part of the
MODEL command. For TYPE=THREELEVEL RANDOM, the random
slope variables are named and defined in the within and between level 2
parts of the MODEL command. Random slope variables defined in
within part of the MODEL command can be used in both the between

683
CHAPTER 17

level 2 and between level 3 parts of the MODEL command. Random

slope variables defined in the between level 2 part of the MODEL
command can be used only in the between level 3 part of the MODEL
command.

For TYPE=THREELEVEL and TYPE=CROSSCLASSIFIED, random

slope variables are between-level variables. For TYPE=TWOLEVEL,
random slope variables are between-level variables unless specifically
designated as having variation on both levels. This is done by placing an
asterisk (*) after the name on the left-hand side of the | symbol.
Following is an example of how to specify a random slope using the |
symbol:

s | y ON x;

where s is a random slope in the regression of y on x where y is a

continuous dependent variable and x is an independent variable. Both
dependent and independent variables can be latent or observed variables.

For TYPE=TWOLEVEL, the asterisk (*) convention is specified as

follows:

s* | y ON x;

where the asterisk (*) indicates that the random slope variable s has
variation on both the within and between levels.

A random slope variable can refer to more than one slope by being used
on the left-hand side of more than one | statement. In this case, the
random slope variables are the same. For example,

s1 | y1 ON x1;
s1 | y2 ON x2;

defines the random slope, s1, to be the same in the regressions of y1 on

x1 and y2 on x2.

Another example is,

s2 | y1 ON x1 x2;

684
 MODEL Command

which defines the random slope, s2, to be the same in the regressions of
y1 on x1 and y1 on x2.

In the following example,

s3 | y1 y2 ON x1 x2 x3;

the six slopes in the regressions of y1 on x1, x2, and x3 and y2 on x1,
x2, and x3 are the same.

RANDOM FACTOR LOADINGS

For TYPE=TWOLEVEL and TYPE=CROSSCLASSIFIED, the | symbol
is used in conjunction with TYPE=RANDOM to name and define the
random factor loading variables in the model. The names on the left-
hand side of the | symbol name the random factor loading variables. The
statements on the right-hand side of the | symbol define the random
factor loading variables. Random factor loadings are defined using the
BY option. BY statements used on the right-hand side of the | symbol
may not use the asterisk (*) or @ symbol. Otherwise, the regular rules
regarding BY apply. The means and the variances of the random factor
loading variables are free as the default. Covariances between random
factor loading variables and growth factors, latent variables defined
using BY statements, and observed variables are fixed at zero as the
default.

For TYPE=TWOLEVEL, the random factor loading variables are named

and defined in the within part of the MODEL command and used in the
between part of the MODEL command. For
TYPE=CROSSCLASSIFIED, random factor loading variables that are
named and defined in the level 2a part of the MODEL command can be
used in the level 2b part of the MODEL command. Level 2a is the
second cluster variable on the CLUSTER statement. Level 2b is the first
cluster variable on the CLUSTER statement. Random factor loading
variables are between-level variables. Following is an example of how
to specify random factor loadings using the | symbol:

s1-s10 | f BY y1-y10;
f@1;

685
CHAPTER 17

where s1 through s10 are random factor loadings for the factor f. All
factor loadings are free. The metric of the factor is set by fixing the
factor variance to one.

AT
The AT option is used with TYPE=RANDOM to define a growth model
with individually-varying times of observation for the outcome variable.
AT is short for measured at. It is used in conjunction with the | symbol
to name and define the random effect variables in a growth model which
are referred to as growth factors.

Four types of growth models can be defined using AT and the | symbol:
an intercept only model, a model with two growth factors, a model with
three growth factors, and a model with four growth factors. The names
of the random effect variables are specified on the left-hand side of the |
symbol. The number of names determines which of the four models
model will be estimated. One name is needed for an intercept only
model and it refers to the intercept growth factor. Two names are
needed for a model with two growth factors: the first one is for the
intercept growth factor and the second one is for the slope growth factor
that uses the time scores to the power of one. Three names are needed
for a model with three growth factors: the first one is for the intercept
growth factor; the second one is for the slope growth factor that uses the
time scores to the power of one; and the third one is for the slope growth
factor that uses the time scores to the power of two. Four names are
needed for a model with four growth factors: the first one is for the
intercept growth factor; the second one is for the slope growth factor that
uses the time scores to the power of one; the third one is for the slope
growth factor that uses the time scores to the power of two; and the
fourth one is for the slope growth factor that uses the time scores to the
power of three. Following are examples of how to specify these growth
models:

intercpt | y1 y2 y3 y4 AT t1 t2 t3 t4;
intercpt slope1 | y1 y2 y3 y4 AT t1 t2 t3 t4;
intercpt slope1 slope2 | y1 y2 y3 y4 AT t1 t2 t3 t4;
intercpt slope1 slope2 slope3 | y1 y2 y3 y4 AT t1 t2 t3 t4;

where intercpt, slope1, slope2, and slope3 are the names of the intercept
and slope growth factors; y1, y2, y3, and y4 are the outcome variables in

686
 MODEL Command

the growth model; and t1, t2, t3, and t4 are observed variables in the data
set that contain information on times of measurement. The TSCORES
option of the VARIABLE command is used to identify the variables that
contain information about individually-varying times of observation for
the outcome in a growth model. The variables on the left-hand side of
AT are paired with the variables on the right-hand side of AT.

The intercepts of the outcome variables are fixed at zero as the default.
The residual variances of the outcome variables are free to be estimated
as the default. The residual covariances of the outcome variables are
fixed at zero as the default. The means, variances, and covariances of
the intercept and slope growth factors are free as the default.

XWITH
The XWITH option is used with TYPE=RANDOM to define
interactions between continuous latent variables or between a continuous
latent variable and an observed variable. XWITH is short for multiplied
with. It is used in conjunction with the | symbol to name and define
interaction variables in a model. Following is an example of how to use
XWITH and the | symbol to name and define an interaction:

int | f1 XWITH f2;

where int is the name of the interaction between f1 and f2. Interaction
variables can be used only on the right-hand side of ON statements.

The XWITH option can be used to create an interaction variable that is

the square of a latent variable. Following is an example of how this is
specified:

fsquare | f XWITH f;

Latent variable interactions are estimated using maximum likelihood for

all regular models as well as mixture models and multilevel models.
Interactions are allowed between continuous latent variables and
between a continuous latent variable and an observed variable. Factor
indicators for the continuous latent variables can be continuous,
censored, binary, ordered categorical (ordinal), counts, or combinations
of these variable types. Observed variables in the interaction can be
independent or mediating variables and the mediating variables can be

687
CHAPTER 17

censored, binary, ordered categorical (ordinal), counts, or combinations

of these variable types. Dependent observed variables in the interaction
can have missing data. In many cases, numerical integration is required
in the maximum likelihood estimation of latent variable interactions.
Numerical integration becomes increasingly more computationally
demanding as the number of factors and the sample size increase.

Interactions between observed variables are handled using the DEFINE

command where an interaction variable is created via multiplication. If
an observed dependent variable has missing data, a latent variable can be
created for this observed variable and the procedure described above can
be used.

Interactions between categorical latent variables and between a

categorical latent variable and an observed or continuous latent variable
are handled using mixture modeling.

Following is a table that summarizes ways of obtaining interactions for

different variable types.

Types of Variables Interaction Options

observed continuous with DEFINE
observed continuous
observed categorical with DEFINE
observed continuous Multiple Group
observed continuous with XWITH
continuous latent
observed categorical with XWITH
continuous latent Multiple Group
observed continuous with MIXTURE
categorical latent
observed categorical with MIXTURE
categorical latent KNOWNCLASS
continuous latent with XWITH
continuous latent
continuous latent with MIXTURE
categorical latent
categorical latent with MIXTURE
categorical latent

688
 MODEL Command

THE MODEL INDIRECT COMMAND

The MODEL INDIRECT command is used to request indirect effects

and their standard errors. Indirect effects are available only for
continuous, binary, and ordered categorical (ordinal) outcomes. The
MODEL INDIRECT command is not available for TYPE=RANDOM,
ALGORITHM=INTEGRATION, the CONSTRAINT option of the
VARIABLE command, and TYPE=EFA.

Delta method standard errors for the indirect effects are computed as the
default. Bootstrap standard errors for the indirect effects can be
obtained by using the MODEL INDIRECT command in conjunction
with the BOOTSTRAP option of the ANALYSIS command.

The MODEL INDIRECT command can be used in conjunction with the

STANDARDIZED option of the OUTPUT command to obtain
standardized indirect effects. Standard errors are available for the
standardized indirect effects when the BOOTSTRAP option is used.

MODEL INDIRECT can also be used in conjunction with the

CINTERVAL option of the OUTPUT command to obtain confidence
intervals for the indirect effects and the standardized indirect effects.
Three types of 95% and 99% confidence intervals can be obtained:
symmetric, bootstrap, or bias-corrected bootstrap confidence intervals
(MacKinnon, Lockwood, & Williams, 2004; MacKinnon, 2008). The
bootstrapped distribution of each parameter estimate is used to
determine the bootstrap and bias-corrected bootstrap confidence
intervals. These intervals take non-normality of the parameter estimate
distribution into account. As a result, they are not necessarily symmetric
around the parameter estimate.

Total, total indirect, specific indirect, and direct effects are obtained
using the IND and VIA options of the MODEL INDIRECT command.
The IND option is used to request a specific indirect effect or a set of
indirect effects. The VIA option is used to request a set of indirect
effects that includes specific mediators.

689
CHAPTER 17

IND
The variable on the left-hand side of IND is the dependent variable in the
indirect effect. The last variable on the right-hand side of IND is the
independent variable in the indirect effect. Other variables on the right-
hand side of IND are mediating variables. If there are no mediating
variables included in the IND option, all indirect effects between the
independent variable and dependent variable are computed. The total
indirect effect is the sum of all indirect effects. The total effect is the
sum of all indirect effects and the direct effect.

VIA
The variable on the left-hand side of VIA is the dependent variable in
the indirect effect. The last variable on the right-hand side of VIA is the
independent variable in the indirect effect. Other variables on the right-
hand side of VIA are mediating variables. All indirect effects that go
from the independent variable to the dependent variable and include the
mediating variables are computed. The total indirect effect is the sum of
all indirect effects.

x1 y1

y3

x2 y2

Following is an example of the model shown in the picture above for

which IND and VIA options will be specified.

MODEL: y3 ON y1 y2;
y2 ON y1 x1 x2;
y1 ON x1 x2;

Following is an example of how MODEL INDIRECT can be used to

request indirect effects using the IND option:

690
 MODEL Command

MODEL INDIRECT:
y3 IND y1 x1;
y3 IND y2 x1;
y3 IND x2;

The first IND statement requests the specific indirect effect from x1 to
y1 to y3. The second IND statement requests the specific indirect effect
from x1 to y2 to y3. The third IND statement requests all indirect effects
from x2 to y3. These include x2 to y1 to y3, x2 to y2 to y3, and x2 to y1
to y2 to y3.

Following is an example of how MODEL INDIRECT can be used to

request indirect effects using the VIA option:

MODEL INDIRECT:
y3 VIA y1 x1 ;

The VIA statement requests all indirect effects from x1 to y3 that are
mediated by y1. These include x1 to y1 to y3 and x1 to y1 to y2 to y3.

Following is an example of how MODEL INDIRECT can be used to

request total, total indirect, specific indirect, and direct effects:

MODEL INDIRECT:
y3 IND x1;

The IND statement requests all indirect effects from x1 to y3. These
include x1 to y1 to y3, x1 to y2 to y3, and x1 to y1 to y2 to y3, the total
effect, and the total indirect effect.

THE MODEL CONSTRAINT COMMAND

The MODEL CONSTRAINT command is used to define linear and non-
linear constraints on the parameters in the model. These constraints can
be implicit or explicit. The default setting for the INFORMATION
option of the ANALYSIS command for MODEL CONSTRAINT is
OBSERVED. The MODEL CONSTRAINT command is not available
for TYPE=EFA.

691
CHAPTER 17

LABELING THE PARAMETERS AND

SELECTING VARIABLES
The MODEL CONSTRAINT command specifies parameter constraints
using labels defined for parameters in the MODEL command, labels
defined for parameters not in the MODEL command using the NEW
option of the MODEL CONSTRAINT command, and names of observed
variables that are identified using the CONSTRAINT option of the
VARIABLE command. See Labeling Parameters to see how parameters
are labeled using the MODEL command.

NEW

The NEW option is used to assign labels and starting values to

parameters not in the analysis model. These parameters are used to
constrain the parameters in the analysis model. The default starting
value for these parameters is 0.5. Following is an example of how the
NEW option is specified:

MODEL:
[y1-y3] (p1-p3);

MODEL CONSTRAINT:
NEW (c*.6);
p2 = p1 + c;
p3 = p1 + 2*c;

where c is a parameter that constrains the means to change linearly

across the three variables. The value .6 following the asterisk (*)
specifies that the value .6 will be used as a starting value for model
estimation. If the analysis is a Monte Carlo simulation study, the value
will also be used as a coverage value.

CONSTRAINT

The CONSTRAINT option of the VARIABLE command is used to

identify the variables that can be used in the MODEL CONSTRAINT
command. These can be not only variables used in the MODEL
command but also other variables. All variables on the CONSTRAINT
list are treated as continuous variables in the analysis. Only variables

692
 MODEL Command

used by the following options cannot be included: GROUPING,

PATTERN, COHORT, COPATTERN, CLUSTER, STRATIFICATION,
and AUXILIARY. Variables that are part of these options can be used
in DEFINE to create new variables that can be used in the
CONSTRAINT statement. The CONSTRAINT option is not available
for TYPE=RANDOM, TYPE=TWOLEVEL, TYPE=THREELEVEL,
TYPE=CROSSCLASSIFIED, TYPE=COMPLEX, and for estimators
other than ML, MLR, and MLF. The CONSTRAINT option is specified
as follows:

CONSTRAINT = y1 u1;

where y1 and u1 are variables that can be used in the MODEL

CONSTRAINT command.

DEFINING LINEAR AND NON-LINEAR

CONSTRAINTS
Linear and non-linear constraints can be defined using the equal sign (=),
the greater than sign (>), the less than sign (<), and all arithmetic
operators and functions that are available in the DEFINE command with
the exception of the absolute value function.

In the MODEL CONSTRAINT command, labels from the MODEL

command and the NEW option of the MODEL CONSTRAINT
command can be used on both the left-hand and right-hand sides of one
or more parameter constraint statements. Variables listed on the
CONSTRAINT option of the VARIABLE command can appear on the
right-hand side of one or more parameter constraint statements.
Following is an example of how to define an explicit constraint:

MODEL:
[y1-y3] (p1-p3);

MODEL CONSTRAINT:
p1 = p2**2 + p3**2;

where the parameter p1 is constrained to be equal to the sum of the

squares of the p2 and p3 parameters.

693
CHAPTER 17

Following is an example of how to define an implicit constraint where a

function of parameters is constrained to be zero:

MODEL:
[y1-y5] (m1-m5);

MODEL CONSTRAINT:
0 = - m4 + m1*m3 - m2;
0 = exp(m3) - 1 - m2;
0 = m4 - m5;

Following is an example of how to define an implicit constraint where a

parameter appears in a set of parameter constraints:

MODEL:
[y1-y4] (p1-p4);

MODEL CONSTRAINT:
p1 = p2**2 + p3**2;
p2 = p4;

DO
The DO option provides a do loop to facilitate specifying the same
expression for a set of parameters. Following is an example of how to
specify a do loop:

MODEL:
y1 ON x1 (p1);
y1 ON x2 (p2);
y1 ON x3 (p3);
y2 ON x1 (q1);
y2 ON x2 (q2);
y2 ON x3 (q3);

MODEL CONSTRAINT:
NEW (ratio1-ratio3);
DO (1, 3) ratio# = p#/q#;

where the numbers in parentheses give the range of values for the do
loop. The number sign (#) is replaced by these values during the

694
 MODEL Command

execution of the do loop. Following are the calculations that are

executed based on the DO option specified above:

ratio1 = p1/q1;
ratio2 = p2/q2;
ratio3 = p2/q3;

The DO option can be used to place constraints on a set of parameters.

Following is an example of how to use a do loop in this way:

MODEL:
f BY y1-y4;
y1-y4 (p1-p4);
MODEL CONSTRAINT:
DO (1, 4) p# > 0;

where the numbers in parentheses give the range of values for the do
loop. The number sign (#) is replaced by these values during the
execution of the do loop. In this example the residual variances of y1
through y4 are constrained to be greater than zero.

PLOT
The PLOT option is used to name the variables that will be plotted on
the y-axis in the plots created using the LOOP option. Following is an
example of how to specify the PLOT option:

PLOT (ind1 ind2);

where ind1 and ind2 are the variables that will be plotted on the y-axis in
the plots created using the LOOP option.

LOOP
The LOOP option is used in conjunction with the PLOT option to create
plots of variables. For example, it is useful for plotting indirect effects
with moderation and mediation (Preacher, Rucker, & Hayes, 2007),
cross-level interactions in multilevel regression (Bauer & Curran, 2005),
and sensitivity graphs for causal effect mediation modeling (Imai, Keele,
& Tingley, 2010; Muthén, 2011). Following is an example of how to
specify the LOOP option:

695
CHAPTER 17

LOOP = (mod, -1, 1, 0.01);

where mod is a variable that will be used on the x-axis, the numbers -1
and 1 are the lower and upper values of mod, and 0.01 is the incremental
value of mod to use in the computations. When mod appears in a
MODEL CONSTRAINT statement involving a new parameter, that
statement is evaluated for each value of mod specified by the LOOP
option. For example, the first value of mod is -1; the second value of
mod is -1 plus 0.01 or -0.99; the third value of mod is -0.99 plus 0.01 or -
0.98; the last value of mod is 1. Plots are created with mod on the x-axis
and the names in the PLOT option on the y-axis.

Following is an example of how to use the PLOT and LOOP options

with the MODEL and MODEL CONSTRAINT commands:

MODEL:
y ON x (p1);

MODEL CONSTRAINT:
PLOT (ypred);
LOOP (age, 10, 50, 1);
ypred = p1*age;

Using TYPE=PLOT2 in the PLOT command, the plot of ypred and age
can be viewed by choosing Loop plots from the Plot menu of the Mplus
Editor. The plot presents the computed values along with a 95%
confidence interval. For frequentist estimation, the default confidence
interval uses plus and minus 1.96 times the standard error. The
CINTERVAL option of the OUTPUT command can be used in
conjunction with the BOOTSTRAP option of the ANALYSIS command
to obtain bootstapped or bias-corrected bootstrap confidence intervals.
For Bayesian estimation, the default is credibility intervals of the
posterior distribution with equal tail percentages. The CINTERVAL
option of the OUTPUT command can be used to obtain credibility
intervals of the posterior distribution that give the highest posterior
density.

696
 MODEL Command

THE MODEL TEST COMMAND

The MODEL TEST command is used to test linear restrictions on the
parameters in the MODEL and MODEL CONSTRAINT commands
using the Wald chi-square test. These restrictions are defined using
labels from the MODEL command and labels from the NEW option of
the MODEL CONSTRAINT command. See Labeling Parameters to see
how parameters are labeled using the MODEL command. See the NEW
option to see how parameters are labeled using this option. Variables
listed on the CONSTRAINT statement of the VARIABLE command
cannot be used in MODEL TEST. Model restrictions can be defined
using the equal sign (=) and all arithmetic operators and functions that
are available in the DEFINE command with the exception of the
absolute value function.

Following is an example of how to place restrictions on the parameters

in the MODEL and MODEL CONSTRAINT commands using MODEL
TEST:

MODEL:
f1 BY y1
y2-y6 (p2-p6);

MODEL CONSTRAINT:
p4 = 2*p2;

MODEL TEST:
0 = p3;
p6 = .5*p5;

where in the MODEL command, p2 represents the factor loading for y2,
p3 represents the factor loading for y3, p4 represents the factor loading
for y4, p5 represents the factor loading for y5, and p6 represents the
factor loading for y6. In the MODEL CONSTRAINT command, the
factor loading for y4 is constrained to be two times the factor loading for
y2. In the MODEL TEST command, the factor loading for y3 is fixed at
zero and the factor loading for y6 is equal to one half the factor loading
for y5. The model described in the MODEL and MODEL
CONSTRAINT commands is tested against the same model except for
the restrictions described in MODEL TEST. A Wald chi-square test
with two degrees of freedom is computed.

697
CHAPTER 17

THE MODEL PRIORS COMMAND

The MODEL PRIORS command is used with ESTIMATOR=BAYES to
specify the prior distribution for each parameter. The default is to use
diffuse (non-informative) priors. Following is a table that shows the
distributions available and the default priors for different types of
parameters:

Type of Parameter Distributions Available Default Priors

Observed continuous normal normal (0, infinity)
dependent variable
means/intercepts (nu)
Observed continuous inverse Gamma inverse Gamma (-1, 0)
dependent variable
variances/residual
variances (theta)
Observed categorical normal normal (0, infinity)
dependent variable uniform
thresholds (tau)
Factor loadings (lambda) normal normal (0, infinity)*
normal (0, 5)**
Regression coefficients (beta) normal normal (0, infinity)*
normal (0, 5)**
Continuous latent variable normal normal (0, infinity)
means/intercepts (alpha)
Continuous latent variable One latent variable inverse Gamma (-1, 0)
variances/residual variances inverse Gamma
(psi) Gamma
uniform
lognormal
normal

More than one latent variable inverse Wishart (0, -p-1)

inverse Wishart
Categorical latent variable Dirichlet Dirichlet (10, 10)
parameters (varies)

* Continuous variables
** Categorical variables

For the normal distribution default, infinity is ten to the power of ten.
For the inverse Gamma default, the settings imply a uniform prior
ranging from minus infinity to plus infinity. For the inverse Wishart
default, p is the dimension of the multivariate block of latent variables.

698
 MODEL Command

For the Dirichlet default, the first number gives the number of
observations to add to the class referred to and the second number gives
the number of observations to add to the last class. For a discussion of
priors, see Gelman et al. (2004), Browne and Draper (2006), and Gelman
(2006).

ASSIGNING PRIORS TO PARAMETERS

Priors are assigned to the parameters using the tilde (~) symbol, which
means distributed as, using the following distribution settings:

Normal – N
Lognormal – LN
Uniform – U
Inverse Gamma – IG
Gamma – G
Inverse Wishart – IW
Dirichlet – D

Each setting has two numbers in parentheses following the setting. For
the normal and lognormal distributions, the first number is the mean and
the second number is the variance. For the uniform distribution, the first
number is the lower limit and the second number is upper limit. For the
inverse Gamma distribution, the first number is the shape parameter and
the second number is the scale parameter. For the Gamma distribution,
the first number is the shape parameter and the second number is the
inverse scale parameter. For the inverse Wishart distribution, the first
number is used to form a covariance matrix and the second number is the
degrees of freedom. For the Dirichlet distribution, the first number gives
the number of observations to add to the class referred to and the second
number gives the number of observations to add to the last class. For a
technical description of the implementation of priors, see Asparouhov
and Muthén (2010b).

In the MODEL PRIORS command, labels from the MODEL command

are used to represent parameters for which prior distributions are
specified. See Labeling Parameters to see how parameters are labeled
using the MODEL command. Following is an example of how to assign
priors to the factor loading parameters:

699
CHAPTER 17

MODEL:
f BY y1-y10* (p1-p10);
f@1;

MODEL PRIORS:
p1-p10 ~ N (1, 0.5);

where parameters p1 through p10 have normal priors with mean one and
variance 0.5.

COVARIANCE
The COVARIANCE option is used to assign a prior to the covariance
between two parameters. Only normal priors are available. Covariance
priors can be assigned to only factor loadings, regression coefficients,
intercepts, and thresholds for binary variables. Following is an example
of how to specify the COVARIANCE option:

MODEL:
y ON x1 (p1)
x2 (p2);

MODEL PRIORS:
p1 ~ N (10, 4);
p2 ~ N (6, 1);
COVARIANCE (p1, p2) = 0.5;

where the bivariate prior distribution of p1 and p2 has a covariance of

0.5 which corresponds to a correlation of 0.25 computed as 0.5 / (sqrt (4)
* sqrt (1)).

DO
The DO option provides a do loop to facilitate specifying the same
expression for a set of parameters. With MODEL PRIORS it can be
used with the DIFFERENCE option to assign priors to differences
among a set of parameters.

700
 MODEL Command

DIFFERENCE
The DIFFERENCE option is used to assign priors to the difference
between two parameters. Only normal priors are available. Difference
priors can be assigned to only factor loadings, regression coefficients,
intercepts, and thresholds for binary variables. Following is an example
of how to specify the DIFFERENCE option:

MODEL:
y ON x1 (p1)
x2 (p2);

MODEL PRIORS:
DIFFERENCE (p1, p2) ~ N (0, 0.01);

where the difference between p1 and p2 has a normal prior with mean
zero and variance 0.01.

The DO option provides a do loop to facilitate specifying the same

expression for a set of parameters. The DO and DIFFERENCE options
can be used together to simplify the assignment of priors to a large set of
difference parameters for models with multiple groups and multiple time
points. In Bayesian estimation, multiple group analysis is obtained using
TYPE=MIXTURE and the KNOWNCLASS option. Following is an
example of how to use the DO and DIFFERENCE options to assign
priors to differences across groups among a set of parameters:

MODEL:
%OVERALL%
f BY y1-y5;
%c#1%
f BY y2-y5 (p12-p15);
%c#2%
f BY y2-y5 (p22-p25);
%c#3%
f BY y2-y5 (p32-p35);

MODEL PRIORS:
DO (2, 5) DIFFERENCE (p1#-p3#) ~ N (0, 0.01);

701
CHAPTER 17

where the numbers in parentheses give the range of values for the do
loop. The number sign (#) is replaced by these values during the
execution of the do loop. Following are the differences that were
assigned normal priors with mean zero and variance 0.01:

p12 - p22
p12 - p32
p22 - p32
p13 - p23
p13 - p33
p23 - p33
p14 - p24
p14 - p34
p24 - p34
p15 - p25
p15 - p35
p25 - p35

MODEL COMMAND VARIATIONS

The MODEL command is used to describe the analysis model using the
options described in the previous sections. This section discusses
variations of the MODEL command for use with multiple group models,
mixture models, multilevel models, and models for generating data for
Monte Carlo simulations studies.

MODEL:
The MODEL command is used to describe the analysis model for a
single group analysis and the overall analysis model for multiple group
analysis.

MODEL label:
MODEL followed by a label is used to describe the group-specific
analysis models in multiple group analysis and the analysis model for
each categorical latent variable in mixture modeling when there are more
than one categorical latent variable in the analysis.

702
 MODEL Command

In multiple group analysis, MODEL followed by a label is used to

describe differences between the overall analysis model described in the
MODEL command and the analysis model for each group. Labels are
defined using the GROUPING option of the VARIABLE command for
raw data in a single file, by the FILE option of the DATA command for
raw data in separate files, and by the program for summary data.
MODEL followed by a label is used in conjunction with the
%WITHIN% and %BETWEEN% specifications in multiple group
multilevel analysis.

In mixture modeling, MODEL followed by a label is used to describe the

analysis model for each categorical latent variable when there are more
than one categorical latent variable in the analysis and for combinations
of categorical latent variables when there are more than two categorical
latent variables in the analysis. Labels are defined by using the names of
the categorical latent variables.

When there are more than one categorical latent variable in the model,
the class-specific parts of the model for each categorical latent variable
must be specified within a MODEL command for that categorical latent
variable. The %OVERALL% specification is not included in the
MODEL commands for each categorical latent variable. Following is an
example of how to specify the MODEL command when there are more
than one categorical latent variable in the model:

MODEL c1:
%c1#1%
%c1#2%

where the class-specific parts of the model for c1 is specified after

MODEL c1.

When there are more than two categorical latent variables in the model,
MODEL commands for pairs of categorical latent variables are allowed.
These are used to specify parameters that are specific to the
combinations of classes for those two categorical latent variables.
Categorical latent variables can be combined in sets involving all but one
categorical latent variable. For example, with three categorical latent
variables c1, c2, and c3, combinations of up to two categorical latent
variables are allowed. Following is an example of how this is specified:

703
CHAPTER 17

MODEL c1.c2:
%c1#1.c2#1%

where %c1#1.c2#1% refers to a combination of class 1 for c1 and class 1

for c2.

MODEL:
%OVERALL%
%class label%
The MODEL command used in conjunction with %OVERALL% and
%class label% is used to describe the overall and class-specific models
for mixture models. Statements following %OVERALL% refer to the
model common to all latent classes. Statements following %class
label% refer to class-specific model statements.

Class labels are created by adding to the name of the categorical latent
variable a number sign (#) followed by the class number. For example,
if c is a categorical latent variable with two latent classes, the class
labels are c#1 and c#2.

MODEL:
%WITHIN%
%BETWEEN%
%BETWEEN label%
The MODEL command used in conjunction with %WITHIN%,
%BETWEEN%, and %BETWEEN label% is used to describe the
individual-level and cluster-level models for multilevel modeling. For
TYPE=TWOLEVEL, the statements following %WITHIN% describe
the individual-level model and the statements following %BETWEEN%
describe the cluster-level model. With multilevel mixture models, the
%OVERALL% and %class label% specifications are used with the
%WITHIN% and %BETWEEN% specifications to describe the mixture
part of the model.

For TYPE=THREELEVEL and TYPE=CROSSCLASSIFIED, the

statements following %WITHIN% describe the individual-level model

704
 MODEL Command

and the statements following %BETWEEN label% describe the cluster-

level model. The label is a cluster variable from the CLUSTER option.

THE MODEL POPULATION COMMAND

The MODEL POPULATION command is used to provide the
population parameter values to be used in data generation using the
options of the MODEL command described earlier. The MODEL
POPULATION command has variations for use with multiple group
models, mixture models, and multilevel models. These are described
below.

In the MODEL POPULATION command, each parameter in the model

must be specified followed by the @ symbol or the asterisk (*) and the
population parameter value. Any model parameter not specified will
have the value of zero as the population parameter value.

Parameter estimates can be saved from a real data analysis using the
ESTIMATES option of the SAVEDATA command and used in a
subsequent Monte Carlo analysis as population parameter values. This
is done by using the POPULATION option of the MONTECARLO
command.

MODEL POPULATION:
The MODEL POPULATION command is used to provide the
population parameter values to be used in data generation for single
group analysis and the overall analysis model for multiple group
analysis.

MODEL POPULATION-label:
MODEL POPULATION followed by a dash and a label is used to
provide parameter values to be used in the generation of data for the
group-specific analysis models in multiple group analysis and the
analysis model for each categorical latent variable in mixture modeling
when there are more than one categorical latent variable in the analysis.

In multiple group analysis, the label following the dash refers to the
group. The first group is referred to by g1, the second group by g2, and

705
CHAPTER 17

so on. In mixture modeling, the label following the dash is the name of
each categorical latent variable when there are more than one categorical
latent variables in the generation of the data.

In addition, the NGROUPS option of the MONTECARLO command is

used for the generation of data for multiple group Monte Carlo
simulation studies.

MODEL POPULATION:
%OVERALL%
%class label%
MODEL POPULATION used in conjunction with %OVERALL% and
%class label% is used to provide the population parameter values to be
used in the generation of data for mixture models. Statements following
%OVERALL% refer to the model common to all latent classes.
Statements following %class label% refer to class-specific model
statements. In addition, the GENCLASSES option of the
MONTECARLO command is used for the generation of data for mixture
models.

The statements in the overall part of the model include information

about the means, variances, and covariances of the background variables
and the specification of the overall model including information about
parameters that do not vary across the classes. The class-specific parts
of the model describe the parameters that vary across classes.

MODEL POPULATION:
%WITHIN%
%BETWEEN%
%BETWEEN label%
MODEL POPULATION used in conjunction with %WITHIN%,
%BETWEEN%, and %BETWEEN label% is used to provide the
population parameter values to be used in the generation of clustered
data. For TYPE=TWOLEVEL, %WITHIN% is used to provide
population parameter values for the individual-level model parameters.
%BETWEEN% is used to provide population parameter values for the
cluster-level model parameters. With multilevel mixture models, the

706
 MODEL Command

%OVERALL% and %class label% specifications are used with the

%WITHIN% and %BETWEEN% specifications to provide the
population parameter values to be used in the generation of data for
mixture models.

For TYPE=THREELEVEL and TYPE=CROSSCLASSIFIED,

%WITHIN% is used to provide population parameter values for the
individual-level model parameters. %BETWEEN label% is used to
provide population parameter values for the cluster-level model
parameters. For TYPE=THREELEVEL, the labels level2 and level3 are
used to provide the level 2 and level 3 population parameter values. For
TYPE=CROSSCLASSIFIED, the labels level2a and level2b are used to
describe the level 2a and level 2b population parameter values.

In addition, the NCSIZES, CSIZES, BETWEEN, and WITHIN options

of the MONTECARLO command are used for the generation of
clustered data.

THE MODEL COVERAGE COMMAND

In Monte Carlo simulation studies, the MODEL command describes not
only the analysis model but also provides values for each parameter that
are used both as the population parameter values for computing coverage
and as starting values in the estimation of the model. If the MODEL
COVERAGE command is used, coverage is computed using the
population parameter values specified in the MODEL COVERAGE
command, and the values specified in the MODEL command are used
only as starting values. The MODEL COVERAGE command has
special options for multiple group models, mixture models, and
multilevel models.

In MODEL COVERAGE, each parameter in the model must be

specified followed by the @ symbol or the asterisk (*) and the
population parameter value. Any model parameter not specified will
have the value of zero as the population parameter value.

Parameter estimates can be saved from a real data analysis using the
ESTIMATES option of the SAVEDATA command and used in a
subsequent Monte Carlo analysis as population parameter values. This
is done by using the COVERAGE option of the MONTECARLO
command.

707
CHAPTER 17

MODEL COVERAGE:
The MODEL COVERAGE command is used to provide the population
parameter values to be used for computing coverage for single group
analysis and the overall analysis model for multiple group analysis.

MODEL COVERAGE-label:
MODEL COVERAGE followed by a dash and a label is used in multiple
group analysis to provide group-specific parameter values to be used in
computing coverage. The label following the dash refers to the group.
The first group is referred to by g1, the second group by g2, and so on.
In mixture modeling, the label following the dash is the name of each
categorical latent variable when there are more than one categorical
latent variable.

MODEL COVERAGE:
%OVERALL%
%class label%
MODEL COVERAGE used in conjunction with %OVERALL% and
%class label% is used to provide the population parameter values to be
used in computing coverage. Statements following %OVERALL% refer
to the model common to all latent classes. Statements following %class
label% refer to class-specific model statements.

Class labels are created by adding to the name of the categorical latent
variable a number sign (#) followed by the class number. For example,
if c1 is a categorical latent variable with two latent classes, the class
labels are c1#1 and c1#2.

The statements in the overall part of the model include information

about the means, variances, and covariances of the background variables
and the specification of the overall model including information about
latent class parameters that do not vary across the classes. The class-
specific parts of the model describe the latent class parameters that do
vary across classes.

708
 MODEL Command

MODEL COVERAGE:
%WITHIN%
%BETWEEN%
%BETWEEN label%
MODEL COVERAGE used in conjunction with %WITHIN% and
%BETWEEN% is used to provide the population parameter values to be
used in computing coverage. For TYPE=TWOLEVEL, %WITHIN% is
used to provide the population parameter values for the individual-level
model parameters. %BETWEEN% is used to provide the population
parameter values for the cluster-level model parameters.

For TYPE=THREELEVEL and TYPE=CROSSCLASSIFIED,

%WITHIN% is used to provide population parameter values for the
individual-level model parameters. %BETWEEN label% is used to
provide population parameter values for the cluster-level model
parameters. For TYPE=THREELEVEL, the labels level2 and level3 are
used to provide the level 2 and level 3 population parameter values. For
TYPE=CROSSCLASSIFIED, the labels level2a and level2b are used to
describe the level 2a and level 2b population parameter values.

THE MODEL MISSING COMMAND

The MODEL MISSING command is used to provide information about
the population parameter values for the missing data model to be used in
the generation of data. The MODEL MISSING command has special
options for multiple group models and for mixture models. The
MISSING option of the MONTECARLO command is required for
MODEL MISSING.

Each parameter in the model must be specified followed by the @

symbol or the asterisk (*) and the population parameter value. Any
model parameter not specified will have the value of zero as the
population parameter value.

Information about each variable listed in the MISSING statement of the

MONTECARLO command must be included as part of the MODEL
MISSING command. These variables must be dependent variables in
the MODEL command. The probability of having missing data or not on
these dependent variables is described by logistic regressions in the

709
CHAPTER 17

MODEL MISSING command. In the MODEL MISSING command, the

missing data indicators use the same names as the dependent variables in
the MODEL command. For each dependent variable, the intercept and
slopes for all covariates must be specified for the logistic regression.
The covariates in these logistic regressions can be both independent and
dependent variables in the MODEL command. When a dependent
variable in the MODEL command is used as a dependent variable in the
logistic regression, it is a missing value indicator. When it is used as a
covariate in the logistic regression, it is the original variable in the
MODEL command. In the following example, in the first ON statement
y1 is a missing value indicator variable. In the second ON statement, y1
is treated as the original variable.

MODEL MISSING:
y1 ON x;
y2 ON y1 x;

A dependent variable that is censored, categorical, or count is treated as

a continuous covariate in the logistic regressions.

MODEL MISSING:
The MODEL MISSING command is used to provide information about
the population parameter values for the missing data model to be used in
the generation of data for single group analysis and the overall analysis
model for multiple group analysis.

MODEL MISSING-label:
MODEL MISSING followed by a dash and a label is used in multiple
group analysis to provide group-specific population parameter values for
the missing data model to be used in the generation of data. The label
following the dash refers to the group. The first group is referred to by
g1, the second group by g2, and so on.

710
 MODEL Command

MODEL MISSING:
%OVERALL%
%class label%
MODEL MISSING used in conjunction with %OVERALL% and %class
label% is used to provide the population parameter values for the
missing data model to be used in the generation of data for mixture
models. Statements following %OVERALL% refer to the model
common to all latent classes. Statements following %class label% refer
to class-specific model statements.

Class labels are created by adding to the name of the categorical latent
variable a number sign (#) followed by the class number. For example,
if c1 is a categorical latent variable with two latent classes, the class
labels are c1#1 and c1#2.

711
CHAPTER 17

712
 OUTPUT, SAVEDATA, And PLOT Commands

CHAPTER 18
OUTPUT, SAVEDATA, AND PLOT
COMMANDS

In this chapter, the OUTPUT, SAVEDATA, and PLOT commands are

discussed. The OUTPUT command is used to request additional output
beyond that included as the default. The SAVEDATA command is used
to save the analysis data and/or a variety of model results in an ASCII
file for future use. The PLOT command is used to request graphical
displays of observed data and analysis results.

THE OUTPUT COMMAND

The OUTPUT command is used to request additional output not
included as the default.

Following are the option settings for the OUTPUT command:

OUTPUT:

SAMPSTAT;
CROSSTABS; ALL
CROSSTABS (ALL);
CROSSTABS (COUNT);
CROSSTABS (%ROW);
CROSSTABS (%COLUMN);
CROSSTABS (%TOTAL);
STANDARDIZED; ALL
STANDARDIZED (ALL);
STANDARDIZED (STDYX);
STANDARDIZED (STDY);
STANDARDIZED (STD);
RESIDUAL;
MODINDICES (minimum chi-square); 10
MODINDICES (ALL);
MODINDICES (ALL minimum chi-square); 10

713
CHAPTER 18

CINTERVAL; SYMMETRIC
CINTERVAL (SYMMETRIC);
CINTERVAL (BOOTSTRAP);
CINTERVAL (BCBOOTSTRAP);
CINTERVAL (EQTAIL); EQTAIL
CINTERVAL (HPD);
SVALUES;
NOCHISQUARE;
NOSERROR;
H1SE;
H1TECH3;
PATTERNS;
FSCOEFFICIENT;
FSDETERMINACY;
BASEHAZARD;
LOGRANK;
D(#); 1
TECH1;
TECH2;
TECH3;
TECH4;
TECH5;
TECH6;
TECH7;
TECH8;
TECH9;
TECH10;
TECH11;
TECH12;
TECH13;
TECH14;
TECH15;
TECH16;

The OUTPUT command is not a required command. Note that

commands can be shortened to four or more letters. Option settings can
be referred to by either the complete word or the part of the word shown
above in bold type.

The default output for all analyses includes a listing of the input setup, a
summary of the analysis specifications, and a summary of the analysis
results. Analysis results include a set of fit statistics, parameter
estimates, standard errors of the parameter estimates, the ratio of each

714
 OUTPUT, SAVEDATA, And PLOT Commands

parameter estimate to its standard error, and a two-tailed p-value for the
ratio. Analysis results for TYPE=EFA include eigenvalues for the
sample correlation matrix, a set of fit statistics, estimated rotated factor
loadings and correlations and their standard errors, estimated residual
variances and their standard errors, the factor structure matrix, and factor
determinacies. Output for TYPE=BASIC includes sample statistics for
the analysis data set and other descriptive information appropriate for
the particular analysis.

Mplus OUTPUT
Following is a description of the information that is provided in the
output as the default. Information about optional output is described in
the next section.

INPUT SETUP

The first information printed in the Mplus output is a restatement of the

input file. The restatement of the input instructions is useful as a record
of which input produced the results provided in the output. Following is
the input file that produced the output that will be used in this chapter to
illustrate most of the output features:
TITLE: example for the output chapter
DATA: FILE = output.dat;
VARIABLE: NAMES = y1-y4 x;
MODEL: f BY y1-y4;
f ON x;
OUTPUT: SAMPSTAT MODINDICES (0) STANDARDIZED
RESIDUAL TECH1 TECH2 TECH3 TECH4
TECH5 FSCOEF FSDET CINTERVAL PATTERNS;
SAVEDATA: FILE IS output.sav;
SAVE IS FSCORES;

SUMMARY OF ANALYSIS SPECIFICATIONS

A summary of the analysis specifications is printed in the output after

the restatement of the input instructions. This is useful because it shows
how the program has interpreted the input instructions and read the data.
It is important to check that the number of observations is as expected.
It is also important to read any warnings and error messages that have
been generated by the program. These contain useful information for
understanding and modifying the analysis.

715
CHAPTER 18

Following is the summary of the analysis for the example output:

SUMMARY OF ANALYSIS

Number of groups 1
Number of observations 500

Number of dependent variables 4

Number of independent variables 1
Number of continuous latent variables 1

Observed dependent variables

Continuous
Y1 Y2 Y3 Y4

Observed independent variables

X

Continuous latent variables

F

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

Input data file(s)

output.dat

Input data format FREE

SUMMARY OF ANALYSIS RESULTS

The third part of the output consists of a summary of the analysis results.
Fit statistics, parameter estimates, and standard errors can be saved in an
external data set by using the RESULTS option of the SAVEDATA
command. Following is a description of what is included in the output.

Tests of model fit are printed first. For most analyses, these consist of
the chi-square test statistic, degrees of freedom, and p-value for the
analysis model; the chi-square test statistic, degrees of freedom, and p-
value for the baseline model of uncorrelated dependent variables; CFI
and TLI; the loglikelihood for the analysis model; the loglikelihood for
the unrestricted model; the number of free parameters in the estimated
model; AIC, BIC, and sample-size adjusted BIC; RMSEA; and SRMR.

716
 OUTPUT, SAVEDATA, And PLOT Commands

MODEL FIT INFORMATION

Number of Free Parameters 13

Loglikelihood

H0 Value -3329.929
H1 Value -3326.522

Information Criteria

Akaike (AIC) 6685.858

Bayesian (BIC) 6740.648
Sample-Size Adjusted BIC 6699.385
(n* = (n + 2) / 24)

Chi-Square Test of Model Fit

Value 6.815
Degrees of Freedom 5
P-Value 0.2348

RMSEA (Root Mean Square Error Of Approximation)

Estimate 0.027
90 Percent C.I. 0.000 0.072
Probability RMSEA <= .05 0.755

CFI/TLI

CFI 0.999
TLI 0.997

Chi-Square Test of Model Fit for the Baseline Model

Value 1236.962
Degrees of Freedom 10
P-Value 0.0000

SRMR (Standardized Root Mean Square Residual)

Value 0.012

The results of the model estimation are printed after the tests of model
fit. The first column of the output labeled Estimates contains the model
estimated value for each parameter. The parameters are identified using
the conventions of the MODEL command. For example, factor loadings
are found in the BY statements. Other regression coefficients are found
in the ON statements. Covariances and residual covariances are found in

717
CHAPTER 18

the WITH statements. Variances, residual variances, means, intercepts,

and thresholds are found under these headings. The scale factors used in
the estimation of models with categorical outcomes are found under the
heading Scales.

The type of regression coefficient produced during model estimation is

determined by the scale of the dependent variable and the estimator
being used in the analysis. For continuous observed dependent variables
and for continuous latent dependent variables, the regression coefficients
produced for BY and ON statements for all estimators are linear
regression coefficients. For censored observed dependent variables, the
regression coefficients produced for BY and ON statements for all
estimators are censored-normal regression coefficients. For the inflation
part of censored observed dependent variables, the regression
coefficients produced for BY and ON statements are logistic regression
coefficients. For binary and ordered categorical observed dependent
variables, the regression coefficients produced for BY and ON
statements using a weighted least squares estimator such as WLSMV are
probit regression coefficients. For binary and ordered categorical
observed dependent variables, the regression coefficients produced for
BY and ON statements using a maximum likelihood estimator are
logistic regression coefficients using the default LINK=LOGIT and
probit regression coefficients using LINK=PROBIT. Logistic regression
for ordered categorical outcomes uses the proportional odds
specification. For categorical latent dependent variables and unordered
categorical observed dependent variables, the regression coefficients
produced for ON statements are multinomial logistic regression
coefficients. For count observed dependent variables and time-to-event
variables in continuous-time survival analysis, the regression
coefficients produced for BY and ON statements are loglinear regression
coefficients. For the inflation part of count observed dependent
variables, the regression coefficients produced for BY and ON
statements are logistic regression coefficients.

718
 OUTPUT, SAVEDATA, And PLOT Commands

MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value
F BY
Y1 1.000 0.000 999.000 999.000
Y2 0.907 0.046 19.908 0.000
Y3 0.921 0.045 20.509 0.000
Y4 0.949 0.046 20.480 0.000

F ON
X 0.606 0.049 12.445 0.000

Intercepts
Y1 0.132 0.051 2.608 0.009
Y2 0.118 0.049 2.393 0.017
Y3 0.061 0.048 1.268 0.205
Y4 0.076 0.050 1.529 0.126

Residual Variances
Y1 0.479 0.043 11.061 0.000
Y2 0.558 0.045 12.538 0.000
Y3 0.492 0.041 11.923 0.000
Y4 0.534 0.044 12.034 0.000
F 0.794 0.073 10.837 0.000

The second column of the output labeled S.E. contains the standard
errors of the parameter estimates. The type of standard errors produced
during model estimation is determined by the estimator that is used. The
estimator being used is printed in the summary of the analysis. Each
analysis type has a default estimator. For several analysis types, the
default estimator can be changed using the ESTIMATOR option of the
ANALYSIS command. A table of estimators that are available for each
analysis type can be found in Chapter 16.

The third column of the output labeled Est./S.E. contains the value of the
parameter estimate divided by the standard error (column 1 divided by
column 2). This statistical test is an approximately normally distributed
quantity (z-score) in large samples. The critical value for a two-tailed
test at the .05 level is an absolute value greater than 1.96. The fourth
column of the output labeled Two-Tailed P-Value gives the p-value for
the z-score in the third column.

The value of 999 is printed when a value cannot be computed. This

happens most often when there are negative variances or residual
variances. A series of asterisks (*) is printed when the value to be

719
CHAPTER 18

printed is too large to fit in the space provided. This happens when
variables are measured on a large scale. To reduce the risk of
computational difficulties, it is recommended to keep variables on a
scale such that their variances do not deviate too far from the range of
one to ten. Variables can be rescaled using the DEFINE command.

OUTPUT OPTIONS
SAMPSTAT

The SAMPSTAT option is used to request sample statistics for the data
being analyzed. For continuous variables, these include sample means,
sample variances, sample covariances, and sample correlations. For
binary and ordered categorical (ordinal) variables using weighted least
squares estimation, these include sample thresholds; first- and second-
order sample proportions if the model has all binary dependent variables;
sample tetrachoric, polychoric and polyserial correlations for models
without covariates; and sample probit regression coefficients and sample
probit residual correlations for models with covariates. The
SAMPSTAT option is not available for censored variables using
maximum likelihood estimation, unordered categorical (nominal)
variables, count variables, binary and ordered categorical (ordinal)
variables using maximum likelihood estimation, and time-to-event
variables. The sample correlation and covariance matrices can be saved
in an ASCII file using the SAMPLE option of the SAVEDATA
command.

Following is an example of the sample statistics for the continuous

outcome example introduced at the beginning of this chapter.

SAMPLE STATISTICS

Means
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
0.104 0.092 0.035 0.050 -0.046

Covariances
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 1.608
Y2 1.028 1.487
Y3 1.027 0.957 1.451

720
 OUTPUT, SAVEDATA, And PLOT Commands

Y4 1.065 0.974 0.992 1.551

X 0.593 0.453 0.506 0.524 0.912

Correlations
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 1.000
Y2 0.665 1.000
Y3 0.672 0.652 1.000
Y4 0.675 0.641 0.661 1.000
X 0.489 0.389 0.440 0.441 1.000

CROSSTABS

The CROSSTABS option is used to request bivariate frequency tables

for pairs of binary, ordered categorical (ordinal), and/or unordered
categorical (nominal) variables. Row, column, and total counts are
given along with row, column, and total percentages for each category of
the variable and the total counts. The CROSSTAB option has the
following settings: ALL, COUNT, %ROW, %COLUMN, and
%TOTAL. The default is ALL. These settings can be used to request
specific information in the bivariate frequency table. For example,

CROSSTABS (COUNT %ROW);

provides a bivariate frequency table with count and row percentages.

STANDARDIZED

The STANDARDIZED option is used to request standardized parameter

estimates and their standard errors. Standard errors are computed using
the Delta method. Only symmetric confidence intervals are available
using the CINTERVAL option of the OUTPUT command.

Three types of standardizations are provided as the default. The first

type of standardization is shown under the heading StdYX in the output.
StdYX uses the variances of the continuous latent variables as well as
the variances of the background and outcome variables for
standardization. The StdYX standardization is the one used in the linear
regression of y on x,

bStdYX = b*SD(x)/SD(y),

721
CHAPTER 18

where b is the unstandardized linear regression coefficient, SD(x) is the

sample standard deviation of x, and SD(y) is the model estimated
standard deviation of y. The standardized coefficient bStdYX is
interpreted as the change in y in y standard deviation units for a standard
deviation change in x.

The second type of standardization is shown under the heading StdY in

the output. StdY uses the variances of the continuous latent variables as
well as the variances of the outcome variables for standardization. The
StdY standardization for the linear regression of y on x is

bStdY = b/SD(y).

StdY should be used for binary covariates because a standard deviation

change of a binary variable is not meaningful. The standardized
coefficient bStdY is interpreted as the change in y in y standard deviation
units when x changes from zero to one.

The third type of standardization is shown under the heading Std in the
output. Std uses the variances of the continuous latent variables for
standardization.

Covariances are standardized using variances. Residual covariances are

standardized using residual variances. This is the case for both latent
and observed variables.

Options are available to request one or two of the standardizations.

They are STANDARDIZED (STDYX), STANDARDIZED (STDY), and
STANDARDIZED (STD). To request only the standardization that uses
the variances of the continuous latent variables as well as the variances
of the background and outcome variables, specify:

STANDARDIZED (STDYX);

To request both the standardization that uses the variances of the

continuous latent variables as well as the variances of the background
and outcome variables and the standardization that uses the variances of
the continuous latent variables and the variances of the outcome
variables, specify:

STANDARDIZED (STDYX STDY);

722
 OUTPUT, SAVEDATA, And PLOT Commands

The STANDARDIZED option is not available for models with random

slopes defined using the | symbol in conjunction with the ON, AT and
XWITH options or for parameters in MODEL CONSTRAINT involving
variables named on the CONSTRAINT statement of the VARIABLE
command. For weighted least squares estimation when the model has
covariates, STDY and standard errors for standardized estimates are not
available. For the MUML estimator, STDY and standard errors for
standardized estimates are not available.

Following is the output obtained when requesting STDYX:

STDYX Standardization
Two-Tailed
Estimates S.E. Est./S.E. P-Value

F BY
Y1 0.838 0.018 47.679 0.000
Y2 0.790 0.020 38.807 0.000
Y3 0.813 0.019 42.691 0.000
Y4 0.810 0.019 42.142 0.000

F ON
X 0.545 0.034 15.873 0.000

Intercepts
Y1 0.104 0.040 2.605 0.009
Y2 0.097 0.040 2.391 0.017
Y3 0.051 0.040 1.269 0.205
Y4 0.061 0.040 1.529 0.126

Residual Variances
Y1 0.298 0.029 10.106 0.000
Y2 0.375 0.032 11.657 0.000
Y3 0.339 0.031 10.964 0.000
Y4 0.345 0.031 11.078 0.000
F 0.703 0.037 18.822 0.000

where the first column of the output labeled Estimates contains the
parameter estimate that has been standardized using the variances of the
continuous latent variables as well as the variances of the background
and outcome variables for standardization, the second column of the
output labeled S.E. contains the standard error of the standardized
parameter estimate, the third column of the output labeled Est./S.E.
contains the value of the parameter estimate divided by the standard
error (column 1 divided by column 2), and the fourth column of the
output labeled Two-Tailed P-Value gives the p-value for the z-score in

723
CHAPTER 18

the third column. When standardized parameter estimates and standard

errors are requested, an R-square value and its standard error are given
for each observed and latent dependent variable in the model.

RESIDUAL

The RESIDUAL option is used to request residuals for the observed

variables in the analysis. Residuals are computed for the model
estimated means/intercepts/thresholds and the model estimated
covariances/correlations/residual correlations. Residuals are computed
as the difference between the value of the observed sample statistic and
its model estimated value. With missing data, the observed sample
statistics are replaced by the estimated unrestricted model for the
means/intercepts/thresholds and the covariances/correlations/residual
correlations. Standardized and normalized residuals are available for
continuous outcomes with TYPE=GENERAL and maximum likelihood
estimation. Standardized residuals are computed as the difference
between the value of the observed sample statistic and its model
estimated value divided by the standard deviation of the difference
between the value of the observed sample statistic and its model
estimated value. Standardized residuals are approximate z-scores.
Normalized residuals are computed as the difference between the value
of the observed sample statistic and its model estimated value divided by
the standard deviation of the value of the observed sample statistic. The
RESIDUAL option is not available for TYPE=RANDOM and the
CONSTRAINT option of the VARIABLE command. Following is an
example of the residual output for a covariance matrix:
RESIDUAL OUTPUT

ESTIMATED MODEL AND RESIDUALS (OBSERVED - ESTIMATED)

Model Estimated Covariances/Correlations/Residual Correlations

Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 1.608
Y2 1.024 1.487
Y3 1.041 0.944 1.451
Y4 1.072 0.972 0.987 1.551
X 0.553 0.501 0.509 0.524 0.912

724
 OUTPUT, SAVEDATA, And PLOT Commands

Residuals for Covariances/Correlations/Residual Correlations

Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 0.000
Y2 0.004 0.000
Y3 -0.014 0.013 0.000
Y4 -0.006 0.002 0.005 0.000
X 0.040 -0.048 -0.003 0.000 0.000

Standardized Residuals (z-scores) for Covariances/Correlations/

Residual Correlations
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 0.000
Y2 0.252 0.000
Y3 -1.241 0.819 0.000
Y4 -0.505 0.143 0.336 0.000
X 1.906 -2.284 -0.132 -0.003 0.000

Normalized Residuals for Covariances/Correlations/Residual

Correlations
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 0.000
Y2 0.043 0.000
Y3 -0.165 0.171 0.000
Y4 -0.073 0.029 0.061 0.00
X 0.664 -0.864 -0.049 -0.001 0.000

MODINDICES

The MODINDICES option is used to request the following indices:

modification indices, expected parameter change indices, and two types
of standardized expected parameter change indices for all parameters in
the model that are fixed or constrained to be equal to other parameters.
Model modification indices are available for most models when
observed dependent variables are continuous, binary, and ordered
categorical (ordinal). The MODINDICES option is used with EFA to
request modification indices and expected parameter change indices for
the residual correlations. The MODINDICES option is not available for
the MODEL CONSTRAINT command,
ALGORITHM=INTEGRATION, TYPE=TWOLEVEL using the
MUML estimator, the BOOTSTRAP option of the ANALYSIS
command, and for models with more than one categorical latent variable.

725
CHAPTER 18

When model modification indices are requested, they are provided as the
default when the modification index for a parameter is greater than or
equal to 10. The following statement requests modification indices
greater than zero:

MODINDICES (0);

Model modification indices are provided for the matrices that are opened
as part of the analysis. To request modification indices for all matrices,
specify:

MODINDICES (ALL);

or

MODINDICES (ALL 0);

The first column of the output labeled M.I. contains the modification
index for each parameter that is fixed or constrained to be equal to
another parameter. A modification index gives the expected drop in
chi-square if the parameter in question is freely estimated. The
parameters are labeled using the conventions of the MODEL command.
For example, factor loadings are found in the BY statements. Other
regression coefficients are found in the ON statements. Covariances and
residual covariances are found in the WITH statements. Variances,
residual variances, means, intercepts, and thresholds are found under
these headings. The scale factors used in the estimation of models with
categorical outcomes are found under the heading Scales.

MODEL MODIFICATION INDICES

Minimum M.I. value for printing the modification index 0.000

M.I. E.P.C. Std E.P.C. StdYX E.P.C.

WITH Statements

Y2 WITH Y1 0.066 0.010 0.010 0.019

Y3 WITH Y1 1.209 -0.042 -0.042 -0.086
Y3 WITH Y2 0.754 0.031 0.031 0.059
Y4 WITH Y1 0.226 -0.019 -0.019 -0.037
Y4 WITH Y2 0.021 0.005 0.005 0.010
Y4 WITH Y3 0.116 0.013 0.013 0.024

726
 OUTPUT, SAVEDATA, And PLOT Commands

The second column of the output labeled E.P.C. contains the expected
parameter change index for each parameter that is fixed or constrained to
be equal to another parameter. An E.P.C. index provides the expected
value of the parameter in question if it is freely estimated. The third and
fourth columns of the output labeled Std E.P.C. and StdYX E.P.C.
contain the two standardized expected parameter change indices. These
indices are useful because the standardized values provide relative
comparisons. The Std E.P.C. indices are standardized using the
variances of the continuous latent variables. The StdYX E.P.C. indices
are standardized using the variances of the continuous latent variables as
well as the variances of the background and/or outcome variables.

CINTERVAL

The CINTERVAL option is used to request confidence intervals for

frequentist model parameter estimates and credibility intervals for
Bayesian model parameter estimates. Confidence intervals are also
available for indirect effects and standardized indirect effects. The
CINTERVAL option has three settings for frequentist estimation and
two settings for Bayesian estimation.

The frequentist settings are SYMMETRIC, BOOTSTRAP, and

BCBOOTSTRAP. SYMMETRIC is the default for frequentist
estimation. SYMMETRIC produces 90%, 95% and 99% symmetric
confidence intervals. BOOTSTRAP produces 90%, 95%, and 99%
bootstrap confidence intervals. BCBOOTSTRAP produces 90%, 95%,
and 99% bias-corrected bootstrap confidence intervals. The
bootstrapped distribution of each parameter estimate is used to
determine the bootstrap and bias-corrected bootstrap confidence
intervals. These intervals take non-normality of the parameter estimate
distribution into account. As a result, they are not necessarily symmetric
around the parameter estimate.

The Bayesian settings are EQTAIL and HPD. EQTAIL is the default for
Bayesian estimation. EQTAIL produces 90%, 95%, and 99% credibility
intervals of the posterior distribution with equal tail percentages. HPD
produces 90%, 95%, and 99% credibility intervals of the posterior
distribution that give the highest posterior density (Gelman et al., 2004).

With frequentist estimation, only SYMMETRIC confidence intervals are

available for standardized parameter estimates. With Bayesian

727
CHAPTER 18

estimation, both EQTAIL and HPD confidence intervals are available

for standardized parameter estimates.

The following statement shows how to request bootstrap confidence

intervals:

CINTERVAL (BOOTSTRAP);

In the output, the parameters are labeled using the conventions of the
MODEL command. For example, factor loadings are found in the BY
statements. Other regression coefficients are found in the ON
statements. Covariances and residual covariances are found in the
WITH statements. Variances, residual variances, means, intercepts, and
thresholds will be found under these headings. The scale factors used in
the estimation of models with categorical outcomes are found under the
heading Scales. The CINTERVAL option is not available for
TYPE=EFA.

The outputs for frequentist confidence intervals and Bayesian credibility

intervals have the same format. Following is output showing symmetric
frequentist confidence intervals:
CONFIDENCE INTERVALS OF MODEL RESULTS

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

F BY
Y1 1.000 1.000 1.000 1.000 1.000 1.000 1.000
Y2 0.790 0.818 0.832 0.907 0.982 0.996 1.024
Y3 0.806 0.833 0.847 0.921 0.995 1.009 1.037
Y4 0.829 0.858 0.872 0.949 1.025 1.039 1.068

F ON
X 0.481 0.511 0.526 0.606 0.686 0.702 0.732

Intercepts
Y1 0.002 0.033 0.049 0.132 0.215 0.231 0.262
Y2 -0.009 0.021 0.037 0.118 0.199 0.214 0.245
Y3 -0.063 -0.033 -0.018 0.061 0.141 0.156 0.186
Y4 -0.052 -0.022 -0.006 0.077 0.159 0.175 0.205

Residual Variances
Y1 0.367 0.394 0.408 0.479 0.550 0.564 0.590
Y2 0.443 0.471 0.485 0.558 0.631 0.645 0.673
Y3 0.386 0.411 0.424 0.492 0.560 0.573 0.599
Y4 0.420 0.447 0.461 0.534 0.607 0.621 0.649
F 0.606 0.651 0.674 0.794 0.915 0.938 0.983

The fourth column of the output labeled Estimate contains the parameter
estimates. The third and fifth columns of the output labeled Lower 5%

728
 OUTPUT, SAVEDATA, And PLOT Commands

and Upper 5%, respectively, contain the lower and upper bounds of the
90% confidence interval. The second and sixth columns of the output
labeled Lower 2.5% and Upper 2.5%, respectively, contain the lower
and upper bounds of the 95% confidence interval. The first and seventh
columns of the output labeled Lower .5% and Upper .5%, respectively,
contain the lower and upper bounds of the 99% confidence interval.

SVALUES

The SVALUES option is used to create input statements that contain

parameter estimates from the analysis. These values are used as starting
values in the input statements. The input statements can be used in a
subsequent analysis in the MODEL or MODEL POPULATION
commands. Not all input statements are reported, for example, input
statements with the | symbol followed by ON, AT, or XWITH. For
MODEL CONSTRAINT, input statements are created for only the
parameters of the NEW option. Input statements are created as the
default when a model does not converge. To request that these input
statements be created, specify the following:

SVALUES;

NOCHISQUARE

The NOCHISQUARE option is used to request that the chi-square fit

statistic not be computed. This reduces computational time when the
model contains many observed variables. The chi-square fit statistic is
computed as the default when available. To request that the chi-square
fit statistic not be computed, specify the following:

NOCHISQUARE;

This option is not available for the MONTECARLO command unless

missing data are generated.

NOSERROR

The NOSERROR option is used to request that standard errors not be

computed. This reduces computational time when the model contains
many observed variables. To request that standard errors not be
computed, specify the following:

729
CHAPTER 18

NOSERROR;

This option is not available for the MONTECARLO command.

H1SE

The H1SE option is used with the ML, MLR, and MLF estimators to
request standard errors for the unrestricted H1 model. It must be used in
conjunction with TYPE=BASIC or the SAMPSTAT option of the
OUTPUT command. It is not available for any other analysis type and it
cannot be used in conjunction with the BOOTSTRAP option of the
ANALYSIS command.

H1TECH3

The H1TECH3 option is used to request estimated covariance and

correlation matrices for the parameter estimates of the unrestricted H1
model. It is not available for any other analysis types, and it cannot be
used in conjunction with the BOOTSTRAP option of the ANALYSIS
command.

PATTERNS

The PATTERNS option is used to request a summary of missing data

patterns. The first part of the output shows the missing data patterns that
occur in the data. In the example below, there are 13 patterns of
missingness. In pattern 1, individuals are observed on y1, y2, y3, and
y4. In pattern 7, individuals are observed on y1 and y4.

SUMMARY OF MISSING DATA PATTERNS

MISSING DATA PATTERNS (x = not missing)

1 2 3 4 5 6 7 8 9 10 11 12 13
Y1 x x x x x x x x
Y2 x x x x x x x
Y3 x x x x x x x
Y4 x x x x x x

730
 OUTPUT, SAVEDATA, And PLOT Commands

MISSING DATA PATTERN FREQUENCIES

Pattern Frequency Pattern Frequency Pattern Frequency

1 984 6 12 11 1
2 127 7 14 12 1
3 56 8 87 13 1
4 139 9 9
5 48 10 3

The second part of the output shows the frequency with which each
pattern is observed in the data. For example, 984 individuals have
pattern 1 whereas 14 have pattern 7.

FSCOEFFICIENT

The FSCOEFFICIENT option is used to request factor score coefficients

and a factor score posterior covariance matrix. It is available only for
TYPE=GENERAL and TYPE=COMPLEX with all continuous
dependent variables. The factor score posterior covariance matrix is the
variance/covariance matrix of the factor scores. Following is the
information produced by the FSCOEFFICIENT option:

FACTOR SCORE INFORMATION (COMPLETE DATA)

FACTOR SCORE COEFFICIENTS

Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
F 0.254 0.197 0.227 0.216 0.093

FACTOR SCORE POSTERIOR COVARIANCE MATRIX

F
________
F 0.122

FSDETERMINACY

The FSDETERMINACY option is used to request a factor score

determinacy value for each factor in the model. It is available only for
TYPE=EFA, TYPE=GENERAL, and TYPE=COMPLEX with all
continuous dependent variables. The factor score determinacy is the
correlation between the estimated and true factor scores. It ranges from
zero to one and describes how well the factor is measured with one being

731
CHAPTER 18

the best value. Following is the information produced by the

FSDETERMINACY option:

FACTOR DETERMINACIES

F 0.945

BASEHAZARD

The BASEHAZARD option is used to request baseline hazard values for

each time interval used in a continuous-time survival analysis. This
option is available only with the SURVIVAL option. The baseline
hazard values can be saved using the BASEHAZARD option of the
SAVEDATA command.

LOGRANK

With TYPE=MIXTURE, the LOGRANK option is used to request the

logrank test also known as the Mantel-Cox test (Mantel, 1966). This test
compares the survival distributions between pairs of classes for both
continuous-time and discrete-time survival models. It is a nonparametric
test for right-censored data. For discrete-time survival models, the
DSURVIVAL option of the VARIABLE command must be used to
identify the discrete-time survival variables.

The D option is used with one-factor models with all categorical factor
indicators to provide a number to be used in the translation from the
factor model parameterization to the IRT parameterization. The default
is one. Any number can be provided. The number 1.7 is commonly
used in this translation. The D option is specified as follows:

D(1.7);

where 1.7 is the number to be used in the translation from the factor
model parameterization to the IRT parameterization.

732
 OUTPUT, SAVEDATA, And PLOT Commands

TECH1

The TECH1 option is used to request the arrays containing parameter

specifications and starting values for all free parameters in the model.
The number assigned to the parameter in the parameter specification
matrices is the number used to refer to the parameter in error messages
regarding non-identification and other issues. When saving analysis
results, the parameters are saved in the order used in the parameter
specification matrices. The starting values are shown in the starting
value matrices. The TECH1 option is not available for TYPE=EFA.
TECHNICAL 1 OUTPUT

PARAMETER SPECIFICATION

NU
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
1 2 3 4 0

LAMBDA
F X
________ ________
Y1 0 0
Y2 5 0
Y3 6 0
Y4 7 0
X 0 0

THETA
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 8
Y2 0 9
Y3 0 0 10
Y4 0 0 0 11
X 0 0 0 0 0

ALPHA
F X
________ ________
0 0

BETA
F X
________ ________
F 0 12
X 0 0

733
CHAPTER 18

PSI
F X
________ ________
F 13
X 0 0

STARTING VALUES

NU
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
0.104 0.092 0.035 0.050 0.000

LAMBDA
F X
________ ________
Y1 1.000 0.000
Y2 1.000 0.000
Y3 1.000 0.000
Y4 1.000 0.000
X 0.000 1.000

THETA
Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 0.806
Y2 0.000 0.745
Y3 0.000 0.000 0.727
Y4 0.000 0.000 0.000 0.777
X 0.000 0.000 0.000 0.000 0.000

ALPHA
F X
________ ________
0.000 -0.046

BETA
F X
________ ________
F 0.000 0.000
X 0.000 0.000

PSI
F X
________ ________
F 0.050
X 0.000 0.912

734
 OUTPUT, SAVEDATA, And PLOT Commands

TECH2

The TECH2 option is used to request parameter derivatives. The

TECH2 option is not available for TYPE=EFA and the CONSTRAINT
option of the VARIABLE command unless TYPE=MIXTURE is used.

TECHNICAL 2 OUTPUT

DERIVATIVES

Derivatives With Respect to NU

Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
0.000 0.000 0.000 0.000 0.000

Derivatives With Respect to LAMBDA

F X
________ ________
Y1 0.000 -0.084
Y2 0.000 0.086
Y3 0.000 0.006
Y4 0.000 0.000
X 0.000 -0.014

Derivatives With Respect to THETA

Y1 Y2 Y3 Y4 X
________ ________ ________ ________ ________
Y1 0.000
Y2 -0.014 0.000
Y3 0.058 -0.049 0.000
Y4 0.024 -0.008 -0.019 0.000
X 0.000 0.000 0.000 0.000 0.000

Derivatives With Respect to ALPHA

F X
________ ________
0.000 0.000

Derivatives With Respect to BETA

F X
________ ________
F 0.000 0.000
X 0.000 0.000

Derivatives With Respect to PSI

F X
________ ________
F 0.000
X 0.000 0.000

735
CHAPTER 18

TECH3

The TECH3 option is used to request estimated covariance and

correlation matrices for the parameter estimates. The parameters are
referred to using the numbers assigned to them in TECH1. The TECH3
covariance matrix can be saved using the TECH3 option of the
SAVEDATA command. The TECH3 option is not available for
ESTIMATOR=ULS, the BOOTSTRAP option of the ANALYSIS
command, and TYPE=EFA.

TECH4

The TECH4 option is used to request estimated means, covariances, and

correlations for the latent variables in the model. The TECH4 means
and covariance matrix can be saved using the TECH4 option of the
SAVEDATA command. The TECH4 option is not available for
TYPE=RANDOM, the CONSTRAINT option of the VARIABLE
command, and TYPE=EFA.

TECHNICAL 4 OUTPUT

ESTIMATES DERIVED FROM THE MODEL

ESTIMATED MEANS FOR THE LATENT VARIABLES

F X
________ ________
1 -0.028 -0.046

ESTIMATED COVARIANCE MATRIX FOR THE LATENT VARIABLES

F X
________ ________
F 1.130
X 0.553 0.912

ESTIMATED CORRELATION MATRIX FOR THE LATENT VARIABLES

F X
________ ________
F 1.000
X 0.545 1.000

736
 OUTPUT, SAVEDATA, And PLOT Commands

TECH5

The TECH5 option is used to request the optimization history in

estimating the model. The TECH5 option is not available for
TYPE=EFA.

TECH6

The TECH6 option is used to request the optimization history in

estimating sample statistics for categorical observed dependent
variables. TECH6 is produced when at least one outcome variable is
categorical but not when all outcomes are binary unless there is an
independent variable in the model.

TECH7

The TECH7 option is used in conjunction with TYPE=MIXTURE to

request sample statistics for each class using raw data weighted by the
estimated posterior probabilities for each class.

TECH8

The TECH8 option is used to request that the optimization history in

estimating the model be printed in the output. TECH8 is printed to the
screen during the computations as the default. TECH8 screen printing is
useful for determining how long the analysis takes. TECH8 is available
for TYPE=RANDOM, MIXTURE, TWOLEVEL and analyses where
numerical integration is used.

TECH9

The TECH9 option is used in conjunction with the MONTECARLO

command, the MONTECARLO and IMPUTATION options of the
DATA command, and the BOOTSTRAP option of the ANALYSIS
command to request error messages related to convergence for each
replication or bootstrap draw. These messages are suppressed if TECH9
is not specified.

737
CHAPTER 18

TECH10

The TECH10 option is used to request univariate, bivariate, and

response pattern model fit information for the categorical dependent
variables in the model. This includes observed and estimated (expected)
frequencies and standardized residuals. TECH10 is available for
TYPE=MIXTURE and categorical and count variables with maximum
likelihood estimation.

TECH11

The TECH11 option is used in conjunction with TYPE=MIXTURE to

request the Lo-Mendell-Rubin likelihood ratio test of model fit (Lo,
Mendell, & Rubin, 2001) that compares the estimated model with a
model with one less class than the estimated model. The Lo-Mendell-
Rubin approach has been criticized (Jeffries, 2003) although it is unclear
to which extent the critique affects its use in practice. The p-value
obtained represents the probability that the data have been generated by
the model with one less class. A low p-value indicates that the model
with one less class is rejected in favor of the estimated model. An
adjustment to the test according to Lo-Mendell-Rubin is also given. The
model with one less class is obtained by deleting the first class in the
estimated model. Because of this, it is recommended when using
starting values that they be chosen so that the last class is the largest
class. In addition, it is recommended that model identifying restrictions
not be included in the first class. TECH11 is available only for
ESTIMATOR=MLR. The TECH11 option is not available for the
MODEL CONSTRAINT command, the BOOTSTRAP option of the
ANALYSIS command, training data, and for models with more than one
categorical latent variable.

TECH12

The TECH12 option is used in conjunction with TYPE=MIXTURE to

request residuals for observed versus model estimated means, variances,
covariances, univariate skewness, and univariate kurtosis. The observed
values come from the total sample. The estimated values are computed
as a mixture across the latent classes. The TECH12 option is not
available for TYPE=RANDOM, the MONTECARLO command, the
CONSTRAINT option of the VARIABLE command, and when there are
no continuous dependent variables.

738
 OUTPUT, SAVEDATA, And PLOT Commands

TECH13

The TECH13 option is used in conjunction with TYPE=MIXTURE to

request two-sided tests of model fit for univariate, bivariate, and
multivariate skew and kurtosis (Mardia’s measure of multivariate
kurtosis). Observed sample values are compared to model estimated
values generated over 200 replications. Each p-value obtained
represents the probability that the estimated model has generated the
data. A high p-value indicates that the estimated model fits the data.
TECH13 is available only when the LISTWISE option of the DATA
command is set to ON. TECH13 is not available for
TYPE=TWOLEVEL MIXTURE, ALGORITHM=INTEGRATION, the
BOOTSTRAP option of the ANALYSIS command, the CONSTRAINT
option of the VARIABLE command, and when there are no continuous
dependent variables.

TECH14

The TECH14 option is used in conjunction with TYPE=MIXTURE to

request a parametric bootstrapped likelihood ratio test (McLachlan &
Peel, 2000) that compares the estimated model to a model with one less
class than the estimated model. The p-value obtained represents an
approximation to the probability that the data have been generated by the
model with one less class. A low p-value indicates that the model with
one less class is rejected in favor of the estimated model. The model
with one less class is obtained by deleting the first class in the estimated
model. Because of this, it is recommended that model identifying
restrictions not be included in the first class. In addition, it is
recommended when using starting values that they be chosen so that the
last class is the largest class. TECH14 is not available for
TYPE=RANDOM unless ALGORITHM=INTEGRATION is used, the
BOOTSTRAP option of the ANALYSIS command, training data, the
CONSTRAINT option of the VARIABLE command, with sampling
weights, and for models with more than one categorical latent variable.

Following is a description of the bootstrap method that is used in

TECH14. Models are estimated for both the number of classes in the
analysis model (k) and the number of classes in the analysis model
minus one (k–1). The loglikelihood values from the k and k-1 class
analyses are used to compute a likelihood ratio test statistic (-2 times the
loglikelihood difference). Several data sets, referred to as bootstrap

739
CHAPTER 18

draws, are then generated using the parameter estimates from the k-1
class model. These data are analyzed for both the k and k-1 class models
to obtain loglikelihood values which are used to compute a likelihood
ratio test statistic for each bootstrap draw. The likelihood ratio test
statistic from the initial analysis is compared to the distribution of
likelihood ratio test statistics obtained from the bootstrap draws to
compute a p-value which is used to decide if the k-1 class model fits the
data as well as the k class model.

The parametric bootstrapped likelihood ratio test can be obtained in two

ways. The default method is a sequential method that saves
computational time by using a minimum number of bootstrap draws to
decide whether the p-value is less than or greater than 0.05. The number
of draws varies from 2 to 100. This method gives an approximation to
the p-value. A more precise estimate of the p-value is obtained by using
a full set of bootstrap draws using the LRTBOOTSTRAP option of the
ANALYSIS command. A common value suggested in the literature is
100 bootstrap draws (McLachlan & Peel, 2000). For more information
about TECH14 see Nylund et al. (2007) and Asparouhov and Muthén
(2012c).

In the TECH14 output, the H0 loglikelihood value given is for the k-1
class model. It is important to check that the H0 loglikelihood value in
the TECH14 output is the same as the loglikelihood value for the H0
model obtained in a previous k-1 class analysis. If it is not the same, the
K-1STARTS option of the ANALYSIS command can be used to
increase the number of random starts for the estimation of the k-1 class
model for TECH14.

TECH14 computations are time consuming because for each bootstrap

draw random starts are needed for the k class model. The default for the
k class model is to generate 20 random sets of starting values in the
initial stage followed by 5 optimizations in the final stage. The default
values can be changed using the LRTSTARTS option of the ANALYSIS
command. The following steps are recommended to save computational
time when using TECH14 (Asparouhov & Muthén, 2012c):

1. Run without TECH14 using the STARTS option of the ANALYSIS

command to find a stable solution if the default starts are not
sufficient.

740
 OUTPUT, SAVEDATA, And PLOT Commands

2. Run with TECH14 using the OPTSEED option of the ANALYSIS

command to specify the seed of the stable solution from Step 1.
3. Run with LRTSTARTS = 0 0 100 20; to check if the results are
sensitive to the number of random starts for the k class model.

TECH15

The TECH15 option is used in conjunction with TYPE=MIXTURE,

PARAMETERIZATION=LOGIT or PROBABILITY, and more than
one categorical latent variable to request marginal and conditional
probabilities, including latent transition probabilities, for the categorical
latent variables in a model. If the model includes binary covariates, the
KNOWNCLASS option is used to represent the categories of the binary
covariates. For example, if the binary covariates of gender and treatment
are used, the two covariates should be combined into one observed
variable with four categories using the DEFINE command. This variable
should be used with the KNOWNCLASS option to create four known
classes. If a continuous covariate is used, the probabilities are evaluated
at the sample mean of the covariate.

TECH16

The TECH16 option is used in conjunction with ESTIMATOR=BAYES

to request test statistics for the Bayes factor approach which is used in
conjunction with MODEL PRIORS to test if variances are greater than
zero (Gelman et al., 2004; Asparouhov & Muthén, 2012a; Verhagen &
Fox, 2012).

Mplus PARAMETER ARRAYS

Following is a description of some parameter arrays that are commonly
used in model estimation. The first nine arrays are for the structural
equation part of the model. The remaining eight arrays are for the
mixture part of the model.

741
CHAPTER 18

ARRAYS FOR THE STRUCTURAL EQUATION PART OF

THE MODEL

TAU

The tau vector contains information regarding thresholds of categorical

observed variables. The elements are in the order of thresholds within
variables.

NU

The nu vector contains information regarding means or intercepts of

continuous observed variables.

LAMBDA

The lambda matrix contains information regarding factor loadings. The

rows of lambda represent the observed dependent variables in the model.
The columns of lambda represent the continuous latent variables in the
model.

THETA

The theta matrix contains the residual variances and covariances of the
observed dependent variables or the latent response variables. The rows
and columns both represent the observed dependent variables.

ALPHA

The alpha vector contains the means and/or intercepts of the continuous
latent variables.

BETA

The beta matrix contains the regression coefficients for the regressions
of continuous latent variables on continuous latent variables. Both the
rows and columns represent continuous latent variables.

742
 OUTPUT, SAVEDATA, And PLOT Commands

GAMMA

The gamma matrix contains the regression coefficients for the

regressions of continuous latent variables on observed independent
variables. The rows represent the continuous latent variables in the
model. The columns represent the observed independent variables in the
model.

PSI

The psi matrix contains the variances and covariances of the continuous
latent variables. Both the rows and columns represent the continuous
latent variables in the model.

DELTA

Delta is a vector that contains scaling information for the observed

dependent variables.

ARRAYS FOR THE MIXTURE PART OF THE MODEL

ALPHA (C)

The alpha (c) vector contains the mean or intercept of the categorical
latent variables.

LAMBDA (U)

The lambda (u) matrix contains the intercepts of the binary observed
variables that are influenced by the categorical latent variables. The
rows of lambda (u) represent the binary observed variables in the model.
The columns of lambda (u) represent the classes of the categorical latent
variables in the model.

TAU (U)

The tau (u) vector contains the thresholds of the categorical observed
variables that are influenced by the categorical latent variables. The
elements are in the order of thresholds within variables.

743
CHAPTER 18

GAMMA (C)

The gamma (c) matrix contains the regression coefficients for the
regressions of the categorical latent variables on observed independent
variables. The rows represent the latent classes. The columns represent
the observed independent variables in the model.

KAPPA (U)

The kappa (u) matrix contains the regression coefficients for the
regressions of the binary observed variables on the observed independent
variables. The rows represent the binary observed variables. The
columns represent the observed independent variables in the model.

ALPHA (F)

The alpha (f) vector contains the means and/or intercepts of the growth
factors for the categorical observed variables that are influenced by the
categorical latent variables.

LAMBDA (F)

The lambda (f) matrix contains the fixed loadings that describe the
growth of the categorical observed variables that are influenced by the
categorical latent variables. The rows represent the categorical observed
variables. The columns represent the growth factors.

GAMMA (F)

The gamma (f) matrix contains the regression coefficients for the
regressions of the growth factors on the observed independent variables
and the regression coefficients for the regressions of the categorical
observed variables on the observed independent variables.

THE SAVEDATA COMMAND

The SAVEDATA command is used to save the analysis data, auxiliary
variables, and a variety of analysis results. Following is a list of the
types of information that can be saved:

744
 OUTPUT, SAVEDATA, And PLOT Commands

• Analysis data
• Sample correlation or covariance matrix
• Estimated sigma between matrix from TYPE=TWOLEVEL
• Within- and between-level sample statistics for weighted least
squares estimation
• Analysis results
• Parameter estimates for use in the MONTECARLO command
• Derivatives from an H1 model
• TECH3
• TECH4
• Kaplan-Meier survival curve values for continuous-time survival
• Baseline hazard values for continuous-time survival
• Estimated baseline hazard curve values for continuous-time survival
• Factor scores, posterior probabilities, and most likely class
membership for each response pattern
• Factor scores
• Latent response variables
• Posterior probabilities for each class and most likely class
membership
• Outliers
• Bayesian posterior parameter values

Following are the options for the SAVEDATA command:

SAVEDATA:

FILE IS file name;

FORMAT IS format statement; F10.3
FREE;
MISSFLAG = missing value flag; *
RECORDLENGTH IS characters per record; 1000

SAMPLE IS file name;

COVARIANCE IS file name;
SIGBETWEEN IS file name;
SWMATRIX IS file name;
RESULTS ARE file name;
ESTIMATES ARE file name;
DIFFTEST IS file name;
TECH3 IS file name;
TECH4 IS file name;
KAPLANMEIER IS file name;

745
CHAPTER 18

BASEHAZARD IS file name;

ESTBASELINE IS file name;
RESPONSE IS file name;
MULTIPLIER IS file name;
BPARAMETERS IS file name;
TYPE IS COVARIANCE; Varies
CORRELATION;
SAVE = FSCORES;
FSCORES (#);
LRESPONSES (#);
CPROBABILITIES;
REPWEIGHTS;
MAHALANOBIS;
LOGLIKELIHOOD;
INFLUENCE;
COOKS;

FACTORS = names of factors;

LRESPONSES = names of latent response variables;

MFILE = file name;

MNAMES = names of variables in the data set;
MFORMAT = format statement; FREE
FREE;
MMISSING = Variable (#);
*;
.;
MSELECT = names of variables; all variables in
MNAMES

Although SAVEDATA is not a required command, either the FILE,

SAMPLE, SIGBETWEEN, SWMATRIX, RESULTS, ESTIMATES,
DIFFTEST, TECH3, TECH4, KAPLANMEIER, BASEHAZARD,
ESTBASELINE, RESPONSE or MULTIPLIER option is required if the
SAVEDATA command is used.

Note that commands and options can be shortened to four or more

letters. Option settings can be referred to by either the complete word or
the part of the word shown above in bold type.

746
 OUTPUT, SAVEDATA, And PLOT Commands

FILE
The FILE option is used to specify the name of the ASCII file in which
the individual-level data used in the analysis will be saved. Following is
an example of how to specify the FILE option:

FILE IS newdata.dat;

where newdata.dat is the name of the file in which the individual-level

data used in the analysis will be saved. If the working directory contains
a file of the same name, it will be overwritten. The data are saved in a
fixed format unless the FORMAT option is used. Any original and/or
transformed variables used in the analysis will be saved. Missing values
are saved as an asterisk (*). If categorical variables have been recoded
by the program, the recoded values are saved. If the weight variable has
been rescaled by the program, the rescaled values are saved. The order
in which the variables are saved is given at the end of the output under
SAVEDATA INFORMATION.

The AUXILIARY option of the VARIABLE command can be used in

conjunction with the SAVEDATA command to save variables that are
not used in the analysis along with the analysis variables.

FORMAT
The FORMAT option is used to specify the format in which the analysis
data will be saved. This option cannot be used for saving other types of
data. All dependent and independent variables used in the analysis are
saved. In addition, all other variables that are used in conjunction with
the analysis are saved. The name of the data set along with the names of
the variables saved and the format are printed in the output. The default
is to save the analysis variables using a fixed format.

Following is an example of how to specify the FORMAT option to save

individual data in a free format:

FORMAT IS FREE;

Individual data can also be saved in a fixed format specified by the user.
The user has the choice of which F or E format the analysis variables are

747
CHAPTER 18

saved in with the format of other saved variables determined by the

program. This option is specified as:

FORMAT IS F2.0;

which indicates that all analysis variables will be saved with an F2.0
format.

MISSFLAG
The MISSFLAG option is used to specify the missing value flag to use
in the data set named in the FILE option of the SAVEDATA command.
The default is the asterisk (*). The period (.) and any number can be
used instead. All variables must have the same missing value flag.

RECORDLENGTH
The RECORDLENGTH option is used to specify the number of
characters per record in the file to which the analysis data are saved. It
cannot be used for saving other types of data. The default and maximum
record length is 5000. Following is an example of how the
RECORDLENGTH option is specified:

RECORDLENGTH = 220;

SAMPLE
The SAMPLE option is used to specify the name of the ASCII file in
which the sample statistics such as the correlation or covariance matrix
will be saved. Following is an example of how to specify the SAMPLE
option:

SAMPLE IS sample.dat;

where sample.dat is the name of the file in which the sample statistics
will be saved. If the working directory contains a file of the same name,
it will be overwritten. The data are saved using free format delimited by
a space.

748
 OUTPUT, SAVEDATA, And PLOT Commands

For continuous outcomes, the default is the covariance matrix. For

categorical outcomes, the default is the correlation matrix. For
combinations of continuous and categorical outcomes, the default is the
correlation matrix. The TYPE option can be used in conjunction with
the SAMPLE option to obtain a matrix other than the default matrix.

For TYPE=TWOLEVEL and maximum likelihood estimation, the

sample correlation and covariance matrices are the maximum likelihood
estimated sigma within covariance and correlation matrices. For
TYPE=TWOLEVEL and weighted least squares estimation, the sample
correlation and covariance matrices are the pairwise maximum
likelihood estimated sigma within covariance and correlation matrices.
For ESTIMATOR=MUML, the sample correlation and covariance
matrices are the sample pooled-within correlation and covariance
matrices.

COVARIANCE
The COVARIANCE option is used to specify the name of the ASCII file
in which the model estimated covariance matrix for continuous analysis
variables is saved. Following is an example of how this option is
specified:

COVARIANCE = cov.dat;

where cov.dat is the name of the file in which the covariance matrix for
continuous analysis variables will be saved. If the working directory
contains a file of the same name, it will be overwritten. The data are
saved using free format delimited by a space.

SIGBETWEEN
The SIGBETWEEN option is used to specify the name of the ASCII file
in which the estimated sigma between covariance matrix or the estimated
sigma between correlation matrix will be saved. For maximum
likelihood estimation, it is the consistent maximum likelihood estimate
of sigma between. For weighted least squares estimation, it is the
pairwise maximum likelihood estimated sigma between covariance and
correlation matrices. For ESTIMATOR=MUML, it is the unbiased
estimate of sigma between. Following is an example of how to specify
the SIGB option:

749
CHAPTER 18

SIGBETWEEN IS sigma.dat;

where sigma.dat is the name of the file in which the estimated sigma
between matrix will be saved. If the working directory contains a file of
the same name, it will be overwritten. The data are saved using free
format delimited by a space.

The default is to save the estimated sigma between covariance matrix.

The TYPE option can be used in conjunction with the SIGB option to
obtain the estimated sigma between correlation matrix.

SWMATRIX
The SWMATRIX option is used with TYPE=TWOLEVEL and
weighted least squares estimation to specify the name of the ASCII file
in which the within- and between-level sample statistics and their
corresponding estimated asymptotic covariance matrix will be saved.
The univariate and bivariate sample statistics are estimated using one-
and two-dimensional numerical integration with a default of 7
integration points. The INTEGRATION option of the ANALYSIS
command can be used to change the default. It is recommended to save
this information and use it in subsequent analyses along with the raw
data to reduce computational time during model estimation. Analyses
using this information must have the same set of observed dependent and
independent variables, the same DEFINE command, the same
USEOBSERVATIONS statement, and the same USEVARIABLES
statement as the analysis which was used to save the information.

Following is an example of how to specify the SWMATRIX option:

SWMATRIX IS swmatrix.dat;

where swmatrix.dat is the name of the file in which the analysis results
will be saved. If the working directory contains a file of the same name,
it will be overwritten.

For the DATA IMPUTATION command and the IMPUTATION option

of the DATA command, the SWMATRIX option is specified as follows:

SWMATRIX IS sw*.dat;

750
 OUTPUT, SAVEDATA, And PLOT Commands

where the asterisk (*) is replaced by the number of the imputed data set.
A file is also produced that contains the names of all of the imputed data
sets. To name this file, the asterisk (*) is replaced by the word list. The
file, in this case swlist.dat, contains the names of the imputed data sets.
This file is used with the SWMATRIX of the DATA command in
subsequent analyses.

RESULTS
The RESULTS option is used to specify the name of the ASCII file in
which the results of an analysis will be saved. The results saved include
parameter estimates, standard errors of the parameter estimates, and fit
statistics. If the STANDARDIZED option of the OUTPUT command is
used, standardized parameters estimates and their standard errors will
also be saved. Following is an example of how to specify the RESULTS
option:

RESULTS ARE results.dat;

where results.dat is the name of the file in which the analysis results will
be saved. If the working directory contains a file of the same name, it
will be overwritten. The data are saved using free format delimited by a
space.

ESTIMATES
The ESTIMATES option is used to specify the name of the ASCII file in
which the parameter estimates of an analysis will be saved. The saved
parameter estimates can be used in a subsequent Monte Carlo simulation
study as population values for data generation and/or coverage values
using the POPULATION and/or COVERAGE options of the
MONTECARLO command. The SVALUES option is an alternative to
the ESTIMATES option. The SVALUES option creates input
statements that contain parameter estimates from the analysis as starting
values.

Following is an example of how to specify the ESTIMATES option:

ESTIMATES ARE estimate.dat;

751
CHAPTER 18

where estimate.dat is the name of the file in which the parameter

estimates will be saved. If the working directory contains a file of the
same name, it will be overwritten. The data are saved using free format
delimited by a space.

DIFFTEST
The DIFFTEST option is used in conjunction with the MLMV and
WLSMV estimators to specify the name of the ASCII file in which the
derivatives from an H1 model will be saved. These derivatives are used
in the subsequent estimation of an H0 model to compute a chi-square
difference test using the DIFFTEST option of the ANALYSIS
command. The H1 model is the less restrictive model. The H0 model is
the more restrictive model nested within H1. Following is an example of
how to specify the DIFFTEST option:

DIFFTEST IS deriv.dat;

where deriv.dat is the name of the file in which the derivatives from the
H1 model will be saved. If the working directory contains a file of the
same name, it will be overwritten. The data are saved using free format
delimited by a space.

TECH3
The TECH3 option is used to specify the name of the ASCII file in
which the covariance matrix of parameter estimates will be saved.
Following is an example of how to specify the TECH3 option:

TECH3 IS tech3.dat;

where tech3.dat is the name of the file in which the covariance matrix of
parameter estimates will be saved. If the working directory contains a
file of the same name, it will be overwritten. The data are saved using
free format delimited by a space.

TECH4
The TECH4 option is used to specify the name of the ASCII file in
which the estimated means and covariance matrix for the latent variables

752
 OUTPUT, SAVEDATA, And PLOT Commands

in the analysis will be saved. Following is an example of how to specify

the TECH4 option:

TECH4 IS tech4.dat;

where tech4.dat is the name of the file in which the estimated means and
covariance matrix for the latent variables will be saved. If the working
directory contains a file of the same name, it will be overwritten. The
data are saved using free format delimited by a space.

KAPLANMEIER
The KAPLANMEIER option is used to specify the name of the ASCII
file in which the y- and x-axis values for the Kaplan-Meier survival
curve for continuous-time survival analysis will be saved. This option is
available only with the SURVIVAL option. Following is an example of
how this option is specified:

KAPLANMEIER IS kapmeier.dat;

where kapmeier.dat is the name of the file in which the survival curve
values will be saved. If the working directory contains a file of the same
name, it will be overwritten. The data are saved using free format
delimited by a space.

BASEHAZARD
The BASEHAZARD option is used to specify the name of the ASCII file
in which the estimated baseline hazard values for continuous-time
survival analysis will be saved. This option is available only with the
SURVIVAL option. Following is an example of how this option is
specified:

BASEHAZARD IS base.dat;

where base.dat is the name of the file in which the estimated baseline
hazard values will be saved. If the working directory contains a file of
the same name, it will be overwritten. The data are saved using free
format delimited by a space.

753
CHAPTER 18

ESTBASELINE
The ESTBASELINE option is used to specify the name of the ASCII file
in which the y- and x-axis values for the estimated baseline hazard curve
of the continuous-time survival analysis will be saved. This option is
available only with the SURVIVAL option. Following is an example of
how this option is specified:

ESTBASELINE IS estbase.dat;

where estbase.dat is the name of the file in which the estimated baseline
hazard curve values will be saved. If the working directory contains a
file of the same name, it will be overwritten. The data are saved using
free format delimited by a space.

RESPONSE
The RESPONSE option is used with single-level models and the ML,
MLR, and MLF estimators when all dependent variables are categorical
to specify the name of the ASCII file in which information about each
response pattern is saved. It is not available for models with covariates.
It is available for TYPE=EFA and TYPE=MIXTURE EFA when the
lower and upper limits of the number of factors to be extracted is the
same. If the model has continuous latent variables, factor scores and the
standard errors of the factor scores are saved. For TYPE=MIXTURE,
the factor scores based on most likely class membership are saved in
addition to posterior probabilities for each class and most likely class
membership for each response pattern. The RESPONSE option is not
available for the KNOWNCLASS and TRAINING options of the
VARIABLE command. Following is an example of how to specify the
RESPONSE option:

RESPONSE IS response.dat;

where response.dat is the name of the file in which information about

each response pattern is saved. If the working directory contains a file
of the same name, it will be overwritten. The data are saved using free
format delimited by a space. Response pattern frequencies, factor
scores, standard errors of the factor scores, and posterior probabilities
are saved as F10.3. Pattern values and most likely class membership are
saved as integers.

754
 OUTPUT, SAVEDATA, And PLOT Commands

MULTIPLIER
The MULTIPLIER option is used with the JACKKNIFE setting of the
REPSE option to specify the name of the ASCII file in which the
multiplier values are saved. Following is an example of how to specify
the MULTIPLIER option:

MULTIPLIER IS multiplier.dat;

where multiplier.dat is the name of the file in which the multiplier values
are saved. If the working directory contains a file of the same name, it
will be overwritten. The values are saved as E15.8.

BPARAMETERS
The BPARAMETERS option is used in Bayesian analysis to specify the
name of the ASCII file in which the Bayesian posterior parameter values
for all iterations are saved. Following is an example of how this option
is specified:

BPARAMETERS = bayes.dat;

where bayes.dat is the name of the file in which the Bayesian posterior
parameter values for all iterations will be saved. If the working directory
contains a file of the same name, it will be overwritten. The data are
saved using free format delimited by a space.

TYPE
The TYPE option is used to specify the type of matrix to be saved. It
can be used in conjunction with the SAMPLE and SIGB options to
override the default. The default matrix for the SAMPLE option is the
covariance matrix for continuous outcomes, the correlation matrix for
categorical outcomes, and the correlation matrix for combinations of
continuous and categorical outcomes. The default matrix for the SIGB
option is the covariance matrix. If the default matrix is the covariance
matrix, a correlation matrix can be requested by the following statement:

TYPE = CORRELATION;

755
CHAPTER 18

SAVE
The SAVE option is used to save factor scores, posterior probabilities
for each class, and outliers along with the analysis and/or auxiliary
variables.

FSCORES

When SAVE=FSCORES is used with frequentist estimation, factor

scores are saved along with the other analysis variables. Following is an
example of how this option is specified:

SAVE = FSCORES;

Factor scores are available when observed dependent variables are

continuous, censored, binary, ordered categorical (ordinal), count or a
combination of these variable types. Factor scores are not available for
TYPE=BASIC, TYPE=EFA, or TYPE=TWOLEVEL with weighted
least squares estimation. For censored and count dependent variables,
factor scores are available only for maximum likelihood estimators using
numerical integration.

When SAVE=FSCORES is used with ESTIMATOR=BAYES, a

distribution of factor scores, called plausible values, is obtained for each
observation. The following summaries are saved along with the other
analysis variables: mean, median, standard deviation, lower 2.5% limit,
and upper 97.5% limit. Following is an example of how this option is
specified:

SAVE = FSCORES (50);

where 50 is the number of imputations or draws that are used from the
Bayesian posterior distribution to compute the plausible value
distribution for each observation. The number of imputations or draws
must be specified. There is no default.

The FACTORS option is used to specify the names of the factors for
which the plausible value distributions will be saved. Following is an
example of how this option is specified:

FACTORS = f1 f2 f3;

756
 OUTPUT, SAVEDATA, And PLOT Commands

where f1, f2, and f3 are the factors for which the plausible value
distributions will be saved. If the PLOT command is used, these
plausible values will be saved for plotting.

LRESPONSES

When SAVE=LRESPONSES is used with ESTIMATOR=BAYES, a

distribution of latent response variable scores is obtained for each
observation. The following summaries are saved along with the other
analysis variables: mean, median, standard deviation, lower 2.5% limit,
and upper 97.5% limit. Following is an example of how this option is
specified:

SAVE = LRESPONSES (50);

where 50 is the number of imputations or draws that are used from the
Bayesian posterior distribution to compute the latent response variable
distribution for each observation. The number of imputations or draws
must be specified. There is no default.

The LRESPONSES option is used to specify the names of the latent

response variables underlying categorical outcomes for which the latent
response variable distributions will be saved. Following is an example
of how this option is specified:

LRESPONSES = u1 u2 u3;

where u1, u2, and u3 are the latent response variables underlying
categorical outcomes for which the latent response variable distributions
will be saved.

CPROBABILITIES

When SAVE=CPROBABILITIES is used in conjunction with

TYPE=MIXTURE in the ANALYSIS command, individual posterior
probabilities for each class are saved along with the other analysis
variables. In addition, a variable is saved that contains the most likely
class membership, that is, the class with the highest posterior probability

757
CHAPTER 18

for each individual. Following is an example of how this option is

specified:

SAVE = CPROBABILITIES;

REPWEIGHTS

When SAVE=REPWEIGHTS is used in conjunction with the REPSE

option of the ANALYSIS command, the replicate weights generated are
saved along with the other analysis variables. Following is an example
of how this option is specified:

SAVE = REPWEIGHTS;

MAHALANOBIS

When SAVE=MAHALANOBIS is used, the Mahalanobis distance and

its p-value (Rousseeuw & Van Zomeren, 1990) are saved for each
observation along with the other analysis variables. The
MAHALANOBIS option is available only for continuous outcomes. It
is not available for TYPE=MIXTURE, TWOLEVEL, RANDOM, EFA,
and BASIC; for ESTIMATOR=WLS, WLSM, WLSMV, and ULS; for
the MONTECARLO command; and for the BOOTSTRAP option of the
ANALYSIS command. Following is an example of how this option is
specified:

SAVE = MAHALANOBIS;

LOGLIKELIHOOD

When SAVE=LOGLIKELIHOOD is used, the loglikelihood

contribution from each observation is saved along with the other analysis
variables. The LOGLIKELIHOOD option is available only for the
maximum likelihood estimators. It is not available for TYPE=EFA and
BASIC, the MONTECARLO command, and the BOOTSTRAP option
of the ANALYSIS command. Following is an example of how this
option is specified:

SAVE = LOGLIKELIHOOD;

758
 OUTPUT, SAVEDATA, And PLOT Commands

INFLUENCE

When SAVE=INFLUENCE is used, the loglikelihood distance influence

measure (Cook & Weisberg, 1982) is saved for each observation along
with the other analysis variables. This measure is an overall influence
statistic that computes the influence of an observation on the function
being optimized. This measure is also referred to as likelihood
displacement for maximum likelihood estimators. An analogous fit
function displacement is available for the weighted least squares
estimators. The INFLUENCE option is not available for TYPE=EFA
and BASIC, the MONTECARLO command, and the BOOTSTRAP
option of the ANALYSIS command. The INFLUENCE option can be
computationally demanding because the model is re-estimated as many
times as there are observations. Following is an example of how this
option is specified:

SAVE = INFLUENCE;

COOKS

When SAVE=COOKS is used, Cook’s D (Cook, 1977) is saved for each

observation along with the other analysis variables. This measure is a
statistic that computes the influence of an observation on the parameter
estimates. The COOKS option is not available for TYPE=EFA and
BASIC, the MONTECARLO command, and the BOOTSTRAP option
of the ANALYSIS command. The COOKS option can be
computationally demanding because the model is re-estimated as many
times as there are observations. Following is an example of how this
option is specified:

SAVE = COOKS;

FACTORS
The FACTORS option is used in conjunction with
ESTIMATOR=BAYES to specify the names of the factors for which the
distribution of factor scores, called plausible values, will be saved.
Following is an example of how to specify the FACTORS option:

FACTORS = f1 f2 f3;

759
CHAPTER 18

where f1, f2, and f3 are the factors for which the plausible value
distributions will be saved.

LRESPONSES
The LRESPONSES option is used in conjunction with ESTIMATOR =
BAYES to specify the names of the latent response variables underlying
the categorical outcomes for which the plausible value distributions will
be saved. Following is an example of how to specify the LRESPONSES
option:

LRESPONSES = u1 u2 u3;

where u1, u2, and u3 are the latent responses variables underlying the
categorical outcomes for which the latent response variable distributions
will be saved.

MERGING DATA SETS

The following options are used in conjunction with the FILE option of
the DATA command and the FILE option of the SAVEDATA command
to merge the analysis data set with the data set named using the MFILE
option described below. Only individual data sets can be merged. Both
data sets must contain an ID variable which is used for merging.

MFILE

The MFILE option is used to specify the name and location of the ASCII
file that is merged with the file named in the FILE option of the DATA
command. It is specified as

MFILE IS c:\merge\merge.dat;

where merge.dat is the name of the ASCII file containing the data to be
merged with the data set named using the FILE option of the DATA
command. In this example, the file merge.dat is located in the directory
c:\merge. If the full path name of the data set contains any blanks, the
full path name must have quotes around it.

760
 OUTPUT, SAVEDATA, And PLOT Commands

If the name of the data set is specified with a path, the directory
specified by the path is checked. If the name of the data set is specified
without a path, the local directory is checked. If the data set is not found
in the local directory, the directory where the input file is located is
checked.

MNAMES

The MNAMES option is used to assign names to the variables in the

data set named using the MFILE option of the SAVEDATA command.
The variable names can be separated by blanks or commas and can be up
to 8 characters in length. Variable names must begin with a letter. They
can contain only letters, numbers, and the underscore symbol. The
program makes no distinction between upper and lower case letters.
Following is an example of how the MNAMES option is specified:

MNAMES ARE id gender ethnic income educatn drink_st agedrink;

The ID variable from the IDVARIABLE option of the VARIABLE

command must be one of the variables listed in the MNAMES statement,

Variable names are generated if a list of variables is specified using the

MNAMES option. For example,

MNAMES ARE y1-y5 x1-x3;

generates the variable names y1 y2 y3 y4 y5 x1 x2 x3.

MNAMES ARE itema-itemd;

generates the variable names itema itemb itemc itemd.

MFORMAT

The MFORMAT option is used to describe the format of the data set to
be merged with the analysis data set. Individual data can be in fixed or
free format. Free format is the default. Fixed format is recommended
for large data sets because it is faster to read data using a fixed format.

For data in free format, each entry on a record must be delimited by a

comma, space, or tab. When data are in free format, the use of blanks is

761
CHAPTER 18

not allowed. The number of variables in the data set is determined from
information provided in the MNAMES option of the SAVEDATA
command. Data are read until the number of pieces of information equal
to the number of variables is found. The program then goes to the next
record to begin reading information for the next observation.

For data in fixed format, each observation must have the same number of
records. Information for a given variable must occupy the same position
on the same record for each observation. A FORTRAN-like format
statement describing the position of the variables in the data set is
required. See the FORMAT option of the DATA command for a
description of how to specify a format statement.

MMISSING

The MMISSING option is used to identify the values or symbol in the

data set to be merged with the analysis data set that are treated as
missing or invalid. Any numeric value and the non-numeric symbols of
the period, asterisk (*), or blank can be used as missing value flags.
There is no default missing value flag. Numeric and non-numeric
missing value flags cannot be combined. The blank cannot be used as a
missing value flag for data in free format. When a list of missing value
flags contains a negative number, the entries must be separated by
commas. See the MISSING option of the VARIABLE command for
further information about missing value flags.

MSELECT

The MSELECT option is used to select the variables from the data set to
be merged with the analysis data set. Variables included on the
MSELECT list must come from the MNAMES statement. The
MSELECT option is specified as follows:

MSELECT ARE gender income agefirst;

THE PLOT COMMAND

The PLOT command is used to request graphical displays of observed
data and analysis results. These graphical displays can be viewed after
the analysis is completed using a post-processing graphics module.

762
 OUTPUT, SAVEDATA, And PLOT Commands

Following are the options for the PLOT command:

PLOT:

TYPE IS PLOT1;
PLOT2;
PLOT3;
SERIES IS list of variables in a series plus x-axis
values;
FACTORS ARE names of factors (#);
LRESPONSES ARE names of latent response variables (#);
OUTLIERS ARE MAHALANOBIS;
LOGLIKELIHOOD;
INFLUENCE;
COOKS;
MONITOR IS ON; OFF
OFF;

The PLOT command is not a required command. Note that commands

can be shortened to four or more letters. Option settings can be referred
to by either the complete word or the part of the word shown above in
bold type.

The AUXILIARY option of the VARIABLE command can be used in

conjunction with the PLOT command to save variables that are not used
in the analysis for subsequent use in graphical displays.

TYPE
The TYPE option is used to specify the types of plots that are requested.
The TYPE option has three settings: PLOT1, PLOT2, and PLOT3.

Following is a list of the plots obtained with TYPE=PLOT1:

• Histograms of sample values

• Scatterplots (sample values)
• Sample means
• Sample proportions
• Observed individual values

Following is a list of the plots obtained with TYPE=PLOT2 in addition

to the plots listed above for PLOT1:

763
CHAPTER 18

• Estimated means
• Sample proportions and estimated probabilities
• Plot estimated probabilities only
• Plot sample proportions only
• Plot estimated probabilities and sample proportions
• Plot estimated probabilities, conditional on a set of covariates
• Plot conditional estimated probabilities as a function of one
covariate
• Sample and estimated means
• Adjusted estimated means
• IRT plots
• Item characteristic curves
• Information curves
• Eigenvalues for EFA
• Mixture distributions
• Survival curves
• Kaplan-Meier curve
• Sample log cumulative hazard curve
• Estimated baseline hazard curve
• Estimated baseline survival curve
• Estimated log cumulative baseline curve
• Kaplan-Meier curve with estimated baseline survival curve
• Sample log cumulative hazard curve with estimated log
cumulative baseline curve
• Estimated survival curve
• Estimated log cumulative curve
• Missing data plots
• Dropout means
• Sample means
• Bayesian plots
• Posterior parameter distributions
• Posterior parameter trace plots
• Autocorrelation plots
• Prior parameter distributions
• Posterior predictive checking scatterplots
• Posterior predictive checking distribution plots
• Loop plots

764
 OUTPUT, SAVEDATA, And PLOT Commands

Following is a list of the plots obtained with TYPE=PLOT3 in addition

to the plots listed above for PLOT1 and PLOT2:

• Histograms of outliers
• Histograms of estimated values
• Scatterplots (sample values, estimated factor scores, outliers,
estimated values)
• Observed individual values
• Estimated individual values
• Estimated individual probability values
• Estimated means and observed individual values
• Estimated means and estimated individual values
• Adjusted estimated means and observed individual values
• Adjusted estimated means and estimated individual values
• Estimated probabilities for a categorical latent variable as a function
of its covariates
• Latent variable distribution plots

Plots can be generated for the total sample, by group, by class, and
adjusted for covariates.

SERIES
The SERIES option is used to list the names of the set of variables to be
used in plots where the values are connected by a line. The x-axis values
for each variable must also be given. For growth models, the set of
variables is the repeated measures of the outcome over time, and the x-
axis values are the time scores in the growth model. For other models,
the set of variables reflects an ordering of the observed variables in the
plot. Non-series plots such as histograms and scatterplots are available
for all analyses.

Values for the x axis can be given in three ways: by putting the x-axis
values in parentheses following each variable in the series; by using an
asterisk (*) in parentheses to request integer values starting with 0 and
increasing by 1; and for growth models, by putting the name of the slope
growth factor in parentheses following each outcome or a list of the
outcomes to request time score values.

765
CHAPTER 18

Following is an example of putting the x-axis values in parentheses

following each outcome:

SERIES = y1 (0) y2 (1) y3 (2) y4 (3);

where the x-axis value for y1 is 0, for y2 is 1, for y3 is 2, and for y4 is 3.

Following is an example of putting an asterisk (*) in parentheses to

request integer values starting with 0 and increasing by 1:

SERIES = y1 y2 y3 y4 (*);

or

SERIES = y1-y4 (*);

This results in 0 as the first x-axis value and 1, 2, and 3 as subsequent

values.

Following is an example of putting the name of the slope growth factor

in parentheses following each outcome in a growth model:

SERIES = y1 (slope) y2 (slope) y3 (slope) y4 (slope);

where slope is the name of the slope growth factor. The list function can
also be used with the SERIES option. It is specified as follows:

SERIES = y1-y4 (slope);

This results in the time scores for the slope growth factor being used as
the x-axis values.

The SERIES option can be used to give variables and x-axis values for
more than one series. The list of variables for each series is separated by
the | symbol. Following is an example for two growth processes:

SERIES = y1 (0) y2 (1) y3 (2) y4 (3) | y5 (0) y6 (1) y7 (4) y8 (5);

where for the first growth process, the time score for y1 is 0, the time
score for y2 is 1, the time score for y3 is 2, and the time score for y4 is 3;

766
 OUTPUT, SAVEDATA, And PLOT Commands

and for the second growth process, the time score for y5 is 0, the time
score for y6 is 1, the time score for y7 is 4, and the time score for y8 is 5.

Using the list function and the name of the slope growth factor, the
SERIES option is specified as:

SERIES = y1-y4 (s1) | y5-y8 (s2);

where s1 is the name of the slope growth factor for the first growth
process and s2 is the name of the slope growth factor for the second
growth process. The names of the slope growth factors are defined in
the MODEL command.

FACTORS
The FACTORS option is used in conjunction with
ESTIMATOR=BAYES to specify the names of the factors for which the
distributions of factor scores, called plausible values, will be saved for
plotting. Following is an example of how to specify the FACTORS
option:

FACTORS = f1 f2 f3 (50);

where 50 is the number of imputations or draws that are used from the
Bayesian posterior distribution to compute the plausible value
distribution for each observation. F1, f2, and f3 are the factors for which
the plausible value distributions will be saved for plotting.

LRESPONSES
The LRESPONSES option is used in conjunction with
ESTIMATOR=BAYES to specify the names of the latent response
variables underlying the categorical outcomes for which the plausible
value distributions will be saved for plotting. Following is an example
of how to specify the LRESPONSES option:

LRESPONSES = u1 u2 u3 (50);

where 50 is the number of imputations or draws that are used from the
Bayesian posterior distribution to compute the latent response variable
distributions for each observation. U1, u2, and u3 are the latent response

767
CHAPTER 18

variables underlying the categorical outcomes for which the latent

response variable distributions will be saved for plotting.

OUTLIERS
The OUTLIERS option is used to select the outliers that will be saved
for use in graphical displays. The OUTLIERS option has the following
settings:

MAHALANOBIS Mahalanobis distance and its p-value

LOGLIKELIHOOD Loglikelihood contribution
INFLUENCE Loglikelihood distance influence measure
COOKS Cook’s D parameter estimate influence measure

Following is an example of how to specify the OUTLIERS option:

OUTLIERS = MAHALANOBIS COOKS;

With this specification, the Mahalanobis distance and its p-value and
Cook’s D will be saved for use in graphical displays.

The loglikelihood distance influence measure and Cooks D can be

computationally demanding because the model is re-estimated as many
times as there are observations. For further information about the
outliers, see the SAVEDATA command.

For TYPE=TWOLEVEL, INFLUENCE and COOKS are available at the

individual and cluster levels and LOGLIKELIHOOD is available at the
cluster level.

MONITOR
The MONITOR option is used to request that certain plots be shown on
the monitor during model estimation. The default is OFF. To request
that the plots be shown specify:

MONITOR = ON:

For Bayesian analysis, trace plots are shown when one chain is used.
For all models except TYPE=GENERAL and TYPE=EFA,
loglikelihoods are shown.

768
 OUTPUT, SAVEDATA, And PLOT Commands

VIEWING GRAPHICAL OUTPUTS

Mplus includes a dialog-based, post-processing graphics module that can
be accessed using the Plot menu of the Mplus Editor or by clicking on
the V button on the toolbar. Following is a description of some of the
features of the graphics module.

Plots can be viewed by selecting the View plots item under the Plot
menu or by clicking on the V button on the toolbar. A list of plots
available appears in the window as shown below.

After a plot is selected, a window appears showing ways that the plot
can be customized. For example, if observed individual curves is
selected, the following window appears:

769
CHAPTER 18

Individual curves can be viewed in consecutive or random order. The

window above shows that sets of 10 individual curves will be viewed in
consecutive order. Random order can be selected and the number of
curves can be changed. The next set of curves are displayed by either
selecting the Get next sample item under the Individual data submenu of
the Plot menu or by using the arrow button on the toolbar bar.

When viewing a plot, if the mouse is held on a point, information about

the variable values for the individual represented by that point are given
as shown in the window below.

770
 OUTPUT, SAVEDATA, And PLOT Commands

Following is the window that is used to adjust plots of estimated means

for different covariate values. A set of covariates is named by typing a
name in the edit box next to the Name covariate set button and clicking
on the Name covariate set button. The set of covariates for the analysis
then appears in the section under Covariate values. The mean or
particular values of the covariates can be given for the plot.

771
CHAPTER 18

Descriptive statistics can be viewed by using the View descriptive

statistics item of the Plot menu which provides the following
information for each variable.

772
 OUTPUT, SAVEDATA, And PLOT Commands

The plots can be exported as a DIB, EMF, or JPEG file using the Export
plot to item of the Plot menu. In addition, the data for each plot can be
saved in an external file using the Save plot data item of the Plot menu
for subsequent use by another program.

773
CHAPTER 18

774
 MONTECARLO Command

CHAPTER 19
MONTECARLO COMMAND

In this chapter, the MONTECARLO command is discussed. The

MONTECARLO command is used to set up and carry out a Monte Carlo
simulation study.

THE MONTECARLO COMMAND

Following are the options for the MONTECARLO command:

MONTECARLO:

NAMES = names of variables;

NOBSERVATIONS = number of observations;
NGROUPS = number of groups; 1
NREPS = number of replications; 1
SEED = random seed for data generation; 0
GENERATE = scale of dependent variables for data
generation;
CUTPOINTS = thresholds to be used for categorization of
covariates
GENCLASSES = names of categorical latent variables (number
of latent classes used for data generation);
NCSIZES = number of unique cluster sizes for each group
separated by the | symbol;
CSIZES = number (cluster size) for each group
separated by the | symbol;
HAZARDC = specifies the hazard for the censoring
process;
PATMISS = missing data patterns and proportion missing
for each dependent variable;
PATPROBS = proportion for each missing data pattern;
MISSING = names of dependent variables that have
missing data;
CENSORED ARE names and limits of censored-normal
dependent variables;
CATEGORICAL ARE names of ordered categorical dependent
variables;
NOMINAL ARE names of unordered categorical dependent
variables;
COUNT ARE names of count variables;

775
CHAPTER 19

CLASSES = names of categorical latent variables (number

of latent classes used for model estimation);
AUXILIARY = names of auxiliary variables (R);
names of auxiliary variables (R3STEP);
names of auxiliary variables (E);
names of auxiliary variables (DU3STEP);
names of auxiliary variables (DE3STEP);
SURVIVAL = names and time intervals for time-to-event
variables;
TSCORES = names, means, and standard deviations of
observed variables with information on
individually-varying times of observation;
WITHIN = names of individual-level observed variables;
BETWEEN = names of cluster-level observed variables;
POPULATION = name of file containing population parameter
values for data generation;
COVERAGE = name of file containing population parameter
values for computing parameter coverage;
STARTING = name of file containing parameter values for
use as starting values for the analysis;
REPSAVE = numbers of the replications to save data from
or ALL;
SAVE = name of file in which generated data are
stored;
RESULTS = name of file in which analysis results are
stored;
BPARAMETERS = name of file in which Bayesian posterior
parameter values are stored;

The MONTECARLO command is not a required command. When the

MONTECARLO command is used, however, the NAMES and
NOBSERVATIONS options are required. Default settings are shown in
the last column. If the default settings are appropriate for the analysis,
nothing besides the required options needs to be specified. Following is
a description of the MONTECARLO command.

GENERAL SPECIFICATIONS
The NAMES, NOBSERVATIONS, NGROUPS, NREPS, and SEED
options are used to give the basic specifications for a Monte Carlo
simulation study. These options are described below.

776
 MONTECARLO Command

NAMES

The NAMES option is used to assign names to the variables in the

generated data sets. These names are used in the MODEL
POPULATION and MODEL commands to specify the data generation
and analysis models. As in regular analysis, the list feature can be used
to generate variable names. Consider the following specification of the
NAMES option:

NAMES = y1-y10 x1-x5;

which is the same as specifying:

NAMES = y1 y2 y3 y4 y5 y6 y7 y8 y9 y10 x1 x2 x3 x4 x5;

NOBSERVATIONS

The NOBSERVATIONS option is used to specify the sample size to be

used for data generation and in the analysis. The NOBSERVATIONS
option is specified as follows:

NOBSERVATIONS = 500;

where 500 is the sample size to be used for data generation and in the
analysis.

If the data being generated are for a multiple group analysis, a sample
size must be specified for each group. In multiple group analysis, the
NOBSERVATIONS option is specified as follows:

NOBSERVATIONS = 500 1000;

where a sample size of 500 is used for data generation and in the
analysis in the first group and a sample size of 1000 is used for data
generation and in the analysis for the second group.

NGROUPS

The NGROUPS option is used to specify the number of groups to be

used for data generation and in the analysis. The NGROUPS option is
specified as follows:

777
CHAPTER 19

NGROUPS = 3;

where 3 is the number of groups to be used for data generation and in the
analysis. The default for the NGROUPS option is 1. The NGROUPS
option is not available for TYPE=MIXTURE.

For Monte Carlo studies, the program automatically assigns the label g1
to the first group, g2 to the second group, etc. These labels are used with
the MODEL POPULATION and MODEL commands to describe the
data generation and analysis models for each group.

NREPS

The NREPS option is used to specify the number of replications for a

Monte Carlo study, that is, the number of samples that are drawn from
the specified population and the number of analyses that are carried out.
The NREPS option is specified as follows:

NREPS = 100;

where 100 is the number of samples that are drawn and the number of
analyses that are carried out. The default for the NREPS option is 1.

SEED

The SEED option is used to specify the seed to be used for the random
draws. The SEED option is specified as follows:

SEED = 23458256;

where 23458256 is the random seed to be used for the random draws.
The default for the SEED option is zero.

DATA GENERATION
The GENERATE, CUTPOINTS, GENCLASSES, NCSIZES, CSIZES,
and HAZARDC options are used in conjunction with the MODEL
POPULATION command to specify how data are to be generated for a
Monte Carlo simulation study. These options are described below.

778
 MONTECARLO Command

GENERATE

The GENERATE option is used to specify the scales of the dependent

variables for data generation. Variables not mentioned using the
GENERATE option are generated as continuous variables. In addition
to generating continuous variables which is the default, dependent
variables can be generated as censored, binary, ordered categorical
(ordinal), unordered categorical (nominal), count variables, and time-to-
event variables.

Censored variables can be generated with censoring from above or from

below and can be generated with or without inflation. The letters ca
followed by a censoring limit in parentheses following a variable name
indicate that the variable is censored from above. The letters cb
followed by a censoring limit in parentheses following a variable name
indicate that the variable is censored from below. The letters cai
followed by a censoring limit in parentheses following a variable name
indicate that the variable is censored from above with inflation. The
letters cbi followed by a censoring limit in parentheses following a
variable name indicate that the variable is censored from below with
inflation.

For binary and ordered categorical (ordinal) variables using maximum

likelihood estimation, the number of thresholds followed by the letter l
for a logistic model or the letter p for a probit model is put in
parentheses following the variable name. If no letter is specified, a
logistic model is used. For binary and ordered categorical (ordinal)
variables using weighted least squares estimation, only the probit model
is allowed. If p is not specified, the probit model is used. The number
of thresholds is equal to the number of categories minus one.

For unordered categorical (nominal) variables, the letter n followed by

the number of intercepts is put in parentheses following the variable
name. The number of intercepts is equal to the number of categories
minus one because the intercepts are fixed to zero in the last category
which is the reference category.

Count variables can be generated for the following six models: Poisson,
zero-inflated Poisson, negative binomial, zero-inflated negative
binomial, zero-truncated negative binomial, and negative binomial
hurdle (Long, 1997; Hilbe, 2011). The letter c or p in parentheses

779
CHAPTER 19

following the variable name indicates that the variable is generated using
a Poisson model. The letters ci or pi in parentheses following the
variable name indicate that the variable is generated using a zero-inflated
Poisson model. The letters nb in parentheses following the variable
name indicate that the variable is generated using a negative binomial
model. The letters nbi in parentheses following the variable name
indicate that the variable is generated using a zero-inflated negative
binomial model. The letters nbt in parentheses following the variable
name indicate that the variable is generated using a zero-truncated
negative binomial model. The letters nbh in parentheses following the
variable name indicate that the variable is generated using a negative
binomial hurdle model.

For time-to-event variables in continuous-time survival analysis, the

letter s and the number and length of time intervals of the baseline
hazard function is put in parentheses following the variable name. When
only s is in parentheses, the number of intervals is equal to the number
of observations.

The GENERATE option is specified as follows:

GENERATE = u1-u2 (1) u3 (1 p) u4 (1 l) u5 u6 (2 p) y1 (ca 1) y2 (cbi 0)

u7 (n 2) u8 (ci) t1 (s 5*1);

where the information in parentheses following the variable name or list

of variable names defines the scale of the dependent variables for data
generation. In this example, the variables u1, u2, u3, and u4 are binary
variables with one threshold. Variables u1, u2, and u4 are generated
using the logistic model. This is specified by placing nothing or the
letter l after the number of thresholds. Variable u3 is generated using
the probit model. This is specified by placing the letter p after the
number of thresholds. Variables u5 and u6 are three-category ordered
categorical (ordinal) variables with two thresholds. The p in parentheses
specifies that they are generated using the probit model. Note that if a
variable has nothing in parentheses after it, the specification in the next
set of parentheses is applied. This means that both u5 and u6 are
ordered categorical (ordinal) variables with two thresholds. Variable y1
is a censored variable that is censored from above with a censoring limit
of one. Variable y2 is a censored variable with an inflation part that is
censored from below with a censoring limit of zero. Variable u7 is a
three-category unordered categorical (nominal) variable with two

780
 MONTECARLO Command

intercepts. Variable u8 is a count variable with an inflation part.

Variable t1 is a time-to-event variable. The numbers in parentheses
specify that five time intervals of length one will be used for data
generation.

In MODEL POPULATION, the inflation part of a censored or count

variable is referred to by adding to the name of the censored or count
variable the number sign (#) followed by the number 1. The baseline
hazard parameters in continuous-time survival analysis are referred to by
adding to the name of the time-to-event variable the number sign (#)
followed by a number. There are as many baseline hazard parameters as
there are time intervals plus one.

CUTPOINTS

The CUTPOINTS option is used to create binary independent variables

from the multivariate normal independent variables generated by the
program. The CUTPOINTS option specifies the value of the cutpoint to
be used in categorizing an independent variable. Following is an
example of how the CUTPOINTS option is specified:

CUTPOINTS = x1 (0) x2 ( 1);

where x1 has a cutpoint of 0 and x2 has a cutpoint of 1. For x1,

observations having a value less than or equal to 0 are assigned the value
of zero and observations having values greater than 0 are assigned the
value of one. Any independent variable not mentioned using the
CUTPOINTS option is assumed to be continuous.

In multiple group analysis, the CUTPOINT option is specified as follows

where the cutpoints for the groups are separated using the | symbol:

CUTPOINTS = x1 (0) x2 ( 1) | x1 (1) x2 (0);

where the cutpoints before the | symbol are the cutpoints for group 1 and
the cutpoints after the | symbol are the cutpoints for group 2.

GENCLASSES

The GENCLASSES option is used to assign names to the categorical

latent variables in the data generation model and to specify the number

781
CHAPTER 19

of latent classes to be used for data generation. This option is used in

conjunction with TYPE=MIXTURE. The GENCLASSES option is
specified as follows:

GENCLASSES = c1 (3) c2 (2) c3 (3);

where c1, c2, and c3 are the names of the three categorical latent
variables in the data generation model. The numbers in parentheses are
the number of classes that will be used for each categorical latent
variable for data generation. Three classes will be used for data
generation for c1, two classes for c2, and three classes for c3.

The letter b following the number of classes specifies that the categorical
latent variable is a between-level variable. Following is an example of
how to specify that a categorical latent variable being generated is a
between-level variable:

GENCLASSES = cb (2 b);

Categorical latent variables that are to be treated as between-level

variables in the analysis must be specified as between-level variables
using the BETWEEN option.

NCSIZES

The NCSIZES option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to specify the
number of unique cluster sizes to be used for data generation. If the data
being generated are for a multiple group analysis, the number of unique
cluster sizes must be specified for each group.

For TYPE=TWOLEVEL, the NCSIZES option is specified as follows:

NCSIZES = 3;

where 3 is the number of unique cluster sizes to be used for data

generation.

In multiple group analysis, the NCSIZES option is specified as follows

where the number of unique cluster sizes for the groups are separated
using the | symbol:

782
 MONTECARLO Command

NCSIZES = 3 | 2;

where 3 is the number of unique cluster sizes to be used for data

generation for group 1 and 2 is the number of unique cluster sizes for
group 2.

For TYPE=THREELEVEL, consider a model where students are nested

in classrooms and classrooms are nested in schools. Level 1 is student;
level 2 is classroom; and level 3 is school. For TYPE=THREELEVEL,
the NCSIZES option is specified as follows:

NCSIZES = 3 [2];

where the numbers 3 and 2 are the unique cluster sizes to be used for
data generation. In this example, 3 is the number of unique cluster sizes
for level 3, school, and 2 is the number of unique cluster sizes for level
2, classroom.

In multiple group analysis, the NCSIZES option is specified as follows

where the number of unique cluster sizes for the groups are separated
using the | symbol:

NCSIZES = 3 [2] | 4 [3];

where the numbers 3 and 2 are the unique cluster sizes to be used for
data generation in group 1 and the numbers 4 and 3 are the unique
cluster sizes to be used for data generation in group 2. In this example, 3
is the number of unique cluster sizes for level 3, school, and 2 is the
number of unique cluster sizes for level 2, classroom, for group 1 and 4
is the number of unique cluster sizes for level 3, school, and 3 is the
number of unique clusters sizes for level 2, classroom, for group 2.

For TYPE=CROSSCLASSIFIED, consider a model where students are

nested in schools crossed with neighborhoods. Level 1 is student; level
2a is school; and level 2b is neighborhood. For
TYPE=CROSSCLASSIFIED, the NCSIZES option is specified as
follows:

NCSIZES = 3 [2];

783
CHAPTER 19

where the numbers 3 and 2 are the unique cluster sizes to be used for
data generation. In this example, 3 is the number of unique cluster sizes
for level 2a, school, and 2 is the number of unique cluster sizes for level
2b, neighborhood.

CSIZES

The CSIZES option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to specify the
number of clusters and the sizes of the clusters to be used for data
generation.

For TYPE=TWOLEVEL, the CSIZES option is specified as follows:

CSIZES = 100 (10) 30 (5) 15 (1);

where 100 clusters of size 10, 30 clusters of size 5, and 15 clusters of

size 1 will be used for data generation.

In multiple group analysis, the CSIZES option is specified as follows

where the number of clusters and the sizes of the clusters to be used for
the groups are separated by the | symbol:

CSIZES = 100 (10) 30 (5) 15 (1) | 80 (10) 20 (5);

where 100 clusters of size 10, 30 clusters of size 5, and 15 clusters of

size 1 will be used for data generation for group 1 and 80 clusters of size
10 and 20 clusters of size 5 will be used for data generation for group 2.

For TYPE=THREELEVEL, consider a model where students are nested

in classrooms and classrooms are nested in schools. Level 1 is student;
level 2 is classroom; and level 3 is school. For TYPE=THREELEVEL,
the CSIZES option is specified as follows:

CSIZES = 40 [15(2) 10(5)] 30 [6(8)] 7 [20(2)];

where the numbers 40, 30, and 7 are the number of level 3, school,
clusters. There are a total of 77 level 3, school, clusters. The 40 level 3,
school, clusters are made up of 15 level 2, class, clusters of size two and
10 level 2, class, clusters of size 5 for a total of 3200 observations. The
30 level 3, school, clusters are made up of 6 level 2, class, clusters of

784
 MONTECARLO Command

size 8 for a total of 1440 observations. The 7 level 3, school, clusters are
made up of 20 level 2, class, clusters of size 2 for a total of 280
observations. The total sample size for data generation is 4920.

In multiple group analysis, the CSIZES option is specified as follows

where the number of clusters and the sizes of the clusters to be used for
the groups are separated by the | symbol:

CSIZES = 30 [6(8)] 7 [20(2)] | 40 [5(6)] 20 [4(2)] ;

where the numbers 30 and 7 are the number of level 3, school, clusters
for group 1 and the numbers 40 and 20 are the number of level 3, school,
clusters for group 2. There are a total of 37 level 3, school, clusters for
group 1 and 60 level 3, school, clusters for group 2. For group 1, the 30
level 3, school, clusters are made up of 6 level 2, class, clusters of size 8
for a total of 1440 observations. The 7 level 3, school, clusters are made
up of 20 level 2, class, clusters of size 2 for a total of 280 observations.
For group 1, the total sample size for data generation is 1720. For group
2, the 40 level 3, school, clusters are made up of 5 level 2, class, clusters
of size 6 for a total of 1200 observations. The 20 level 3, school,
clusters are made up of 4 level 2, class, clusters of size 2 for a total of
160 observations. For group 2, the total sample size for data generation
is 1360.

For TYPE=CROSSCLASSIFIED, consider a model where students are

nested in schools crossed with neighborhoods. Level 1 is student; level
2a is school; and level 2b is neighborhood. For
TYPE=CROSSCLASSIFIED, the CSIZES option is specified as follows:

CSIZES = 40 [15(2) 10(5)] 30 [6(8)] 7 [20(2)];

where the numbers 40, 30, and 7 are the number of level 2b,
neighborhood, clusters. There are a total of 77 level 2b, neighborhood,
clusters. The numbers 15, 10, 6, and 20 are the number of level 2a,
school, clusters. There are a total of 51 level 2a, school, clusters. The
40 level 2b, neighborhood, clusters are crossed with the 15 level 2a,
school, clusters and the 10 level 2a, school, clusters. Each cell of the 40
by 15 cross-classification contains 2 students for a total of 1200
observations. Each cell of the 40 by 10 cross-classification contains 5
students for a total of 2000 observations. The 30 level 2b,
neighborhood, clusters are crossed with 6 level 2a, school, clusters.

785
CHAPTER 19

Each cell of the 30 by 6 cross-classification contains 8 students for a

total of 1440 observations. The 7 level 2b, neighborhood, clusters are
crossed with the 20 level 2a, school, clusters. Each cell of the 7 by 20
cross-classification contains 2 students for a total of 280 observations.
The total sample size for data generation is 4920.

HAZARDC

The HAZARDC option is used to specify the hazard for the censoring
process in continuous-time survival analysis when time-to-event
variables are generated. This information is used to create a censoring
indicator variable where zero is not censored and one is right censored.
The HAZARDC option is specified as follows:

HAZARDC = t1 (.5);

where t1 is the name of the time-to-event variable that is generated and

.5 is the hazard for censoring.

MISSING DATA GENERATION

The PATMISS, PATPROBS, and MISSING options and the MODEL
MISSING command are used to specify how missing data will be
generated for a Monte Carlo simulation study. These options are
described below. Missing data can be generated using two approaches.
In the first approach, the PATMISS and PATPROBS options are used
together to generate missing data. In the second approach, the MISSING
option is used in conjunction with the MODEL MISSING command to
generate missing data. The approaches cannot be used in combination.
When generated data are saved, the missing value flag is 999. The
PATMISS and PATPROBS options are not available for multiple group
analysis. For multiple group analysis, missing data are generated using
the MISSING option in conjunction with the MODEL MISSING
command. These options are described below.

PATMISS

The PATMISS option is used to specify the missing data patterns and
the proportion of data that are missing to be used in missing data
generation for each dependent variable in the model. Any variable in the
NAMES statement that is not listed in a missing data pattern is assumed

786
 MONTECARLO Command

to have no missing data for all individuals in that pattern. The

PATMISS option is used in conjunction with the PATPROBS option.
The PATMISS option is specified as follows:

PATMISS = y1 (.2) y2 (.3) y3 (.1) |

y2 (.2) y3 (.1) y4 (.3) |
y3 (.1) y4 (.3);

The statement above specifies that there are three missing data patterns
which are separated by the | symbol. The number in parentheses
following each variable is the probability of missingness to be used for
that variable in data generation. In the first pattern, y1, y2, y3 are
observed with missingness probabilities of .2, .3, and .1, respectively. In
the second pattern, y2, y3, y4 are observed with missingness
probabilities of .2, .1, and .3, respectively. In the third pattern, y3 and y4
are observed with missingness probabilities of .1 and .3, respectively.
Assuming that the NAMES statement includes variables y1, y2, y3, and
y4, individuals in the first pattern have no missing data on variable y4;
individuals in the second pattern have no missing data on variable y1;
and individuals in the third pattern have no missing data on variables y1
and y2.

PATPROBS

The PATPROBS option is used to specify the proportion of individuals

for each missing data pattern to be used in the missing data generation.
The PATPROBS option is used in conjunction with the PATMISS
option. The proportions are listed in the order of the missing data
patterns in the PATMISS option and are separated by the | symbol. The
PATPROBS option is specified as follows:

PATPROBS = .4 | .3 | .3;

where missing data pattern one has probability .40 of being observed in
the data being generated, missing data pattern two has probability .30 of
being observed in the data being generated, and missing data pattern
three has probability .30 of being observed in the data being generated.
The missing data pattern probabilities must sum to one.

787
CHAPTER 19

MISSING

The MISSING option is used to identify the dependent variables for

which missing data are be generated. This option is used in conjunction
with the MODEL MISSING command. Missing data are not allowed on
the observed independent variables. The MISSING option is specified
as follows:

MISSING = y1 y2 u1;

which indicates that missing data will be generated for variables y1, y2,
and u1. The probabilities of missingness are described using the
MODEL MISSING command which is described in Chapter 17.

SCALE OF DEPENDENT VARIABLES FOR

ANALYSIS
The CENSORED, CATEGORICAL, NOMINAL, and COUNT options
are used to specify the scale of the dependent variables for analysis.
These options are described below.

All observed dependent variables are assumed to be measured on a

continuous scale for the analysis unless the CENSORED,
CATEGORICAL, NOMINAL, and/or COUNT options are specified.
The specification of the scale of the dependent variables determines how
the variables are treated in the model and its estimation. Independent
variables can be binary or continuous. The scales of the independent
variables have no impact on the model or its estimation. The distinction
between dependent and independent variables is described in the
discussion of the MODEL command.

CENSORED

The CENSORED option is used to specify which dependent variables

are treated as censored variables in the model and its estimation, whether
they are censored from above or below, and whether a censored or
censored-inflated model will be estimated.

The CENSORED option is specified as follows for a censored model:

788
 MONTECARLO Command

CENSORED ARE y1 (a) y2 (b) y3 (a) y4 (b);

where y1, y2, y3, y4 are censored dependent variables in the analysis.
The letter a in parentheses following the variable name indicates that the
variable is censored from above. The letter b in parentheses following
the variable name indicates that the variable is censored from below.
The lower and upper censoring limits are determined from the data
generation.

The CENSORED option is specified as follows for a censored-inflated

model:

CENSORED ARE y1 (ai) y2 (bi) y3 (ai) y4 (bi);

where y1, y2, y3, y4 are censored dependent variables in the analysis.
The letters ai in parentheses following the variable name indicates that
the variable is censored from above and that a censored-inflated model
will be estimated. The letter bi in parentheses following the variable
name indicates that the variable is censored from below and that a
censored-inflated model will be estimated. The lower and upper
censoring limits are determined from the data generation.

With a censored-inflated model, two variables are considered, a

censored variable and an inflation variable. The censored variable takes
on values for individuals who are able to assume values of the censoring
point and beyond. The inflation variable is a binary latent variable with
one denoting that an individual is unable to assume any value except the
censoring point. The inflation variable is referred to by adding to the
name of the censored variable the number sign (#) followed by the
number 1. In the example above, the censored variables available for
use in the MODEL command are y1, y2, y3, and y4, and the inflation
variables available for use in the MODEL command are y1#1, y2#1,
y3#1, and y4#1.

CATEGORICAL

The CATEGORICAL option is used to specify which dependent

variables are treated as binary or ordered categorical (ordinal) variables
in the model and its estimation. Ordered categorical dependent variables
cannot have more than 10 categories. The number of categories is

789
CHAPTER 19

determined from the data generation. The CATEGORICAL option is

specified as follows:

CATEGORICAL ARE u2 u3 u7-u13;

where u2, u3, u7, u8, u9, u10, u11, u12, and u13 are binary or ordered
categorical dependent variables in the analysis.

For categorical dependent variables, there are as many thresholds as

there are categories minus one. The thresholds are referred to in the
MODEL command by adding to the variable name the dollar sign ($)
followed by a number. The threshold for a binary variable u1 is referred
to as u1$1. The two thresholds for a three-category variable u2 are
referred to as u2$1 and u2$2.

NOMINAL

The NOMINAL option is used to specify which dependent variables are

treated as unordered categorical (nominal) variables in the model and its
estimation. Unordered categorical dependent variables cannot have
more than 10 categories. The number of categories is determined from
the data generation. The NOMINAL option is specified as follows:

NOMINAL ARE u1 u2 u3 u4;

where u1, u2, u3, u4 are unordered categorical dependent variables in

the analysis.

For nominal dependent variables, all categories but the last category can
be referred to. The last category is the reference category. The
categories are referred to in the MODEL command by adding to the
variable name the number sign (#) followed by a number. The three
categories of a four-category nominal variable are referred to as u1#1,
u1#2, and u1#3.

COUNT

The COUNT option is used to specify which dependent variables are

treated as count variables in the model and its estimation and the type of
model to be estimated. The following models can be estimated for count
variables: Poisson, zero-inflated Poisson, negative binomial, zero-

790
 MONTECARLO Command

inflated negative binomial, zero-truncated negative binomial, and

negative binomial hurdle (Long, 1997; Hilbe, 2011). The negative
binomial models use the NB-2 variance representation (Hilbe, 2011, p.
63). Count variables may not have negative or non-integer values.

The COUNT option can be specified in two ways for a Poisson model:

COUNT = u1 u2 u3 u4;

or

COUNT = u1 (p) u2 (p) u3 (p) u4 (p);

or using the list function:

COUNT = u1-u4 (p);

The COUNT option can be specified in two ways for a zero-inflated

Poisson model:

COUNT = u1-u4 (i);

or

COUNT = u1-u4 (pi);

where u1, u2, u3, and u4 are count dependent variables in the analysis.
The letter i or pi in parentheses following the variable name indicates
that a zero-inflated Poisson model will be estimated.

With a zero-inflated Poisson model, two variables are considered, a

count variable and an inflation variable. The count variable takes on
values for individuals who are able to assume values of zero and above
following the Poisson model. The inflation variable is a binary latent
variable with one denoting that an individual is unable to assume any
value except zero. The inflation variable is referred to by adding to the
name of the count variable the number sign (#) followed by the number
1.

Following is the specification of the COUNT option for a negative

binomial model:

791
CHAPTER 19

COUNT = u1 (nb) u2 (nb) u3 (nb) u4 (nb);

or using the list function:

COUNT = u1-u4 (nb);

Following is the specification of the COUNT option for a zero-inflated

negative binomial model:

COUNT = u1- u4 (nbi);

With a zero-inflated negative binomial model, two variables are

considered, a count variable and an inflation variable. The count
variable takes on values for individuals who are able to assume values of
zero and above following the negative binomial model. The inflation
variable is a binary latent variable with one denoting that an individual is
unable to assume any value except zero. The inflation variable is
referred to by adding to the name of the count variable the number sign
(#) followed by the number 1.

Following is the specification of the COUNT option for a zero-truncated

negative binomial model:

COUNT = u1-u4 (nbt);

Count variables for the zero-truncated negative binomial model must

have values greater than zero.

Following is the specification of the COUNT option for a negative

binomial hurdle model:

COUNT = u1-u4 (nbh);

With a negative binomial hurdle model, two variables are considered, a

count variable and a hurdle variable. The count variable takes on values
for individuals who are able to assume values of one and above
following the truncated negative binomial model. The hurdle variable is
a binary latent variable with one denoting that an individual is unable to
assume any value except zero. The hurdle variable is referred to by
adding to the name of the count variable the number sign (#) followed by
the number 1.

792
 MONTECARLO Command

OPTIONS FOR DATA ANALYSIS

The CLASSES, AUXILIARY, and SURVIVAL options are used only in
the analysis. These options are described below.

CLASSES

The CLASSES option is used to assign names to the categorical latent

variables in the model and to specify the number of latent classes in the
model for each categorical latent variable. This option is required for
TYPE=MIXTURE. Between-level categorical latent variables must be
identified as between-level variables using the BETWEEN option. The
CLASSES option is specified as follows:

CLASSES = c1 (2) c2 (2) c3 (3);

where c1, c2, and c3 are the names of the three categorical latent
variables in the analysis model. The numbers in parentheses specify the
number of classes that will be used for each categorical latent variable in
the analysis. The categorical latent variable c1 has two classes, c2 has
two classes, and c3 has three classes.

AUXILIARY

Auxiliary variables are variables that are not part of the analysis model.
The AUXILIARY option has five uses with TYPE=MIXTURE. The
first two uses are to identify a set of variables not used in the analysis
that are possible covariates in a multinomial logistic regression for a
categorical latent variable. The last three uses are to identify a set of
variables not used in the analysis for which the equality of means across
latent classes will be tested. Only one of these five can be used in an
analysis at a time.

In the first use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE to identify covariates that are important predictors of
latent classes. This is done using pseudo-class draws, that is, the
posterior-probability based multinomial logistic regression of a
categorical latent variable on a set of covariates. The setting R in
parentheses is placed behind the variables on the AUXILIARY
statement that will be used as covariates in the multinomial logistic

793
CHAPTER 19

regression in a mixture model. Following is an example of how to

specify the R setting:

AUXILIARY = race (R) ses (R) x1-x5 (R);

where race, ses, x1, x2, x3, x4, and x5 will be used as covariates in a
multinomial logistic regression in a mixture model.

In the second use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
covariates that are important predictors of latent classes using a 3-step
approach (Vermunt, 2010; Asparouhov & Muthén, 2012b). It is not
allowed with ALGORITHM=INTEGRATION. The setting R3STEP in
parentheses is placed behind the variables on the AUXILIARY
statement that will be used as covariates in the multinomial logistic
regression in a mixture model. Following is an example of how to
specify the R3STEP setting:

AUXILIARY = race (R3STEP) ses (R3STEP) x1-x5 (R3STEP);

where race, ses, x1, x2, x3, x4, and x5 will be used as covariates in a
multinomial logistic regression in a mixture model.

In the third use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE to identify the variables for which the equality of
means across latent classes will be tested using pseudo-class-based
multiple imputations (Asparouhov, 2007). The setting E in parentheses
is placed behind the variables on the AUXILIARY statement for which
equalities of means across latent classes will be tested. Following is an
example of how to specify the E setting:

AUXILIARY = race (E) ses (E) gender (E);

where the equality of means for race, ses, and gender will be tested
across classes.

In the fourth use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
continuous variables for which the equality of means across latent
classes will be tested using a 3-step approach (Asparouhov & Muthén,
2012b). It is not allowed with ALGORITHM=INTEGRATION. For

794
 MONTECARLO Command

example, this could be a distal outcome that is influenced by a

categorical latent variable. When more than one continuous variable is
given, each variable is tested separately. The setting DU3STEP in
parentheses is placed behind the variables on the AUXILIARY
statement for which equalities of means across latent classes will be
tested. The variances of the variables are allowed to vary across the
classes. Following is an example of how to specify the DU3STEP
setting:

AUXILIARY = drinks (DU3STEP) depress (DU3STEP);

where the equality of means for drinks and depress will be separately
tested across classes.

In the fifth use, the AUXILIARY option is used in conjunction with

TYPE=MIXTURE with one categorical latent variable to identify
continuous variables for which the equality of means across latent
classes will be tested using a 3-step approach (Asparouhov & Muthén,
2012b). It is not allowed for ALGORITHM=INTEGRATION. For
example, this could be a distal outcome that is influenced by a
categorical latent variable. When more than one continuous variable is
given, each variable is tested separately. The setting DE3STEP in
parentheses is placed behind the variables on the AUXILIARY
statement for which equalities of means across latent classes will be
tested. The variances of the variables are held equal across the classes.
Following is an example of how to specify the DE3STEP setting:

AUXILIARY = drinks (DE3STEP) depress (DE3STEP);

where the equality of means for drinks and depress will be separately
tested across classes.

The AUXILIARY option has an alternative specification for the R,

R3STEP, E, DU3STEP, and DE3STEP settings that is convenient when
there are several variables that cannot be specified using the list
function. These are AUXILIARY=(R), AUXILIARY=(R3STEP),
AUXILIARY(E), AUXILIARY(DU3STEP), and
AUXILIARY(DE3STEP). When R, R3STEP, E, DU3STEP, and
DE3STEP in parentheses follows the equal sign, it means that R,
R3STEP, E, DU3STEP, or DE3STEP apply to all of the variables that
follow. For example, the following AUXILIARY statement specifies

795
CHAPTER 19

that the equality of means for drinks and depress will be separately
tested across classes:

AUXILIARY = (DE3STEP) drinks depress;

SURVIVAL

The SURVIVAL option is used to identify the variables that contain

information about time to event and to provide information about the
time intervals in the baseline hazard function to be used in the analysis.
The SURVIVAL option is specified as follows:

SURVIVAL = t (4*5 10);

where t is the variable that contains time to event information. The

numbers in parentheses specify that four time intervals of length five and
one time interval of length ten are used in the analysis.

If no time intervals are specified, a constant baseline hazard is used.

Following is an example of how this is specified:

SURVIVAL = t;

The keyword ALL can be used if the time intervals are taken from the
data. Following is an example of how this is specified:

SURVIVAL = t (ALL);

It is not recommended to use the ALL keyword when the

BASEHAZARD option of the ANALYSIS command is ON because it
results in a large number of baseline hazard parameters.

In continuous-time survival modeling, there are as many baseline hazard

parameters are there are time intervals plus one. When the
BASEHAZARD option of the ANALYSIS command is ON, these
parameters can be referred to in the MODEL command by adding to the
name of the time-to-event variable the number sign (#) followed by a
number. For example, for a time-to-event variable t with 5 time
intervals, the six baseline hazard parameters are referred to as t#1, t#2,
t#3, t#4, t#5, and t#6. In addition to the baseline hazard parameters, the
time-to-event variable has a mean or an intercept.

796
 MONTECARLO Command

VARIABLES WITH SPECIAL FUNCTIONS FOR

DATA GENERATION AND ANALYSIS
The TSCORES, WITHIN, and BETWEEN options are used for both
data generation and in the analysis. These options are described below.

TSCORES

The TSCORES option is used in conjunction with TYPE=RANDOM to

name and define the variables to be generated that contain information
about individually-varying times of observation for the outcome in a
longitudinal study. Variables listed in the TSCORES statement can be
used only in AT statements in the MODEL and MODEL POPULATION
commands to define a growth model. They cannot be used with other
statements in the MODEL command. The TSCORES option is specified
as follows:

TSCORES ARE a1 (0 0) a2 (1 .1) a3 (2 .2) a4 (3 .3);

where a1, a2, a3, and a4 are variables to be generated that contain the
individually-varying times of observation for an outcome at four time
points. The first number in parentheses is the mean of the variable. The
second number in parentheses is the standard deviation of the variable.
Each variable is generated using a univariate normal distribution using
the mean and standard deviation specified in the TSCORES statement.

WITHIN

The WITHIN option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to identify the
variables in the data set that are measured on the individual level and to
specify the levels on which they are modeled. All variables on the
WITHIN list must be measured on the individual level. An individual-
level variable can be modeled on all or some levels.

For TYPE=TWOLEVEL, an individual-level variable can be modeled

on only the within level or on both the within and between levels. If a
variable measured on the individual level is mentioned on the WITHIN
list, it is modeled on only the within level. It has no variance in the
between part of the model. If it is not mentioned on the WITHIN list, it

797
CHAPTER 19

is modeled on both the within and between levels. The WITHIN option
is specified as follows:

WITHIN = y1 y2 x1;

where y1, y2, and x1 are variables measured on the individual level and
modeled on only the within level.

For TYPE=THREELEVEL, an individual-level variable can be modeled

on only level 1, on levels 1 and 2, levels 1 and 3, or on all levels.
Consider a model where students are nested in classrooms and
classrooms are nested in schools. Level 1 is student; level 2 is
classroom; and level 3 is school. If a variable measured on the
individual level is mentioned on the WITHIN list without a label, it is
modeled on only level 1. It has no variance on levels 2 and 3. If it is
mentioned on the WITHIN list with a level 2 cluster label, it is modeled
on levels 1 and 2. It has no variance on level 3. If it is mentioned on the
WITHIN list with a level 3 cluster label, it is modeled on levels 1 and 3.
It has no variance on level 2. If it is not mentioned on the WITHIN list,
it is modeled on all levels.

Following is an example of how to specify the WITHIN option for

TYPE=THREELEVEL:

WITHIN = y1-y3 (level2) y4-y6 (level3) y7-y9;

In the example, y1, y2, and y3 are variables measured on the individual
level and modeled on only level 1. Variables modeled on only level 1
must precede variables modeled on the other levels. Y4, y5, and y6 are
variables measured on the individual level and modeled on levels 1 and
2. Y7, y8, and y9 are variables measured on the individual level and
modeled on levels 1 and 3.

An alternative specification of the WITHIN option above reverses the

order of the level 2 and level 3 variables:

WITHIN = y1-y3 (level3) y7-y9 (level2) y4-y6;

Variables modeled on only level 1 must precede variables modeled on

the other levels. Another alternative specification is:

798
 MONTECARLO Command

WITHIN = y1-y3;
WITHIN = (level2) y4-y6;
WITHIN = (level3) y7-y9;

In this specification, the WITHIN statement for variables modeled on

only level 1 must precede the other WITHIN statements. The order of
the other WITHIN statements does not matter.

For TYPE=CROSSCLASSIFIED, an individual-level variable can be

modeled on only level 1, on levels 1 and 2a, levels 1 and 2b, or on all
levels. Consider a model where students are nested in schools crossed
with neighborhoods. Level 1 is student; level 2a is school; and level 2b
is neighborhood. If a variable measured on the individual level is
mentioned on the WITHIN list without a label, it is modeled on only
level 1. It has no variance on levels 2a and 2b. If it is mentioned on the
WITHIN list with a level 2a cluster label, it is modeled on levels 1 and
2a. It has no variance on level 2b. If it is mentioned on the WITHIN list
with a level 2b cluster label, it is modeled on levels 1 and 2b. It has no
variance on level 2a. If it is not mentioned on the WITHIN list, it is
modeled on all levels.

Following is an example of how to specify the WITHIN option for

TYPE=CROSSCLASSIFIED:

WITHIN = y1-y3 (level2a) y4-y6 (level2b) y7-y9;

In the example, y1, y2, and y3 are variables measured on the individual
level and modeled on only level 1. Variables modeled on only level 1
must precede variables modeled on the other levels. Y4, y5, and y6 are
variables measured on the individual level and modeled on levels 1 and
2a. Y7, y8, and y9 are variables measured on the individual level and
modeled on levels 1 and 2b.

BETWEEN

The BETWEEN option is used with TYPE=TWOLEVEL,

TYPE=THREELEVEL, and TYPE=CROSSCLASSIFIED to identify the
variables in the data set that are measured on the cluster level(s) and to
specify the level(s) on which they are modeled. All variables on the
BETWEEN list must be measured on a cluster level. A cluster-level
variable can be modeled on all or some cluster levels.

799
CHAPTER 19

For TYPE=TWOLEVEL, a cluster-level variable can be modeled on

only the between level. The BETWEEN option is specified as follows:

BETWEEN = z1 z2 x1;

where z1, z2, and x1 are variables measured on the cluster level and
modeled on the between level. The BETWEEN option is also used to
identify between-level categorical latent variables with
TYPE=TWOLEVEL MIXTURE.

For TYPE=THREELEVEL, a variable measured on level 2 can be

modeled on only level 2 or on levels 2 and 3. A variable measured on
level 3 can be modeled on only level 3. Consider a model where
students are nested in classrooms and classrooms are nested in schools.
Level 1 is student; level 2 is classroom; and level 3 is school. If a
variable measured on level 2 is mentioned on the BETWEEN list
without a label, it is modeled on levels 2 and 3. If a variable measured
on level 2 is mentioned on the BETWEEN list with a level 2 cluster
label, it is modeled on only level 2. It has no variance on level 3. A
variable measured on level 3 must be mentioned on the BETWEEN list
with a level 3 cluster label. Following is an example of how to specify
the BETWEEN option for TYPE=THREELEVEL:

BETWEEN = y1-y3 (level2) y4-y6 (level3) y7-y9;

In this example, y1, y2, and y3 are cluster-level variables measured on

level 2 and modeled on both levels 2 and 3. Variables modeled on both
levels 2 and 3 must precede variables modeled on only level 2 or level 3.
Y4, y5, and y6 are cluster-level variables measured on level 2 and
modeled on level 2. Y7, y8, and y9 are cluster-level variables measured
on level 3 and modeled on level 3.

An alternative specification of the BETWEEN option above reverses the

order of the level 2 and level 3 variables:

BETWEEN = y1-y3 (level3) y7-y9 (level2) y4-y6;

Variables modeled on both levels 2 and 3 must precede variables

modeled on only level 2 or level 3. Another alternative specification is:

BETWEEN = y1-y3;

800
 MONTECARLO Command

BETWEEN = (level2) y4-y6;

BETWEEN = (level3) y7-y9;

In this specification, the BETWEEN statement for variables modeled on

both levels 2 and 3 must precede the other BETWEEN statements. The
order of the other BETWEEN statements does not matter.

For TYPE=CROSSCLASSIFIED, a variable measured on level 2a must

be mentioned on the BETWEEN list with a level 2a cluster label. It can
be modeled on only level 2a. A variable measured on level 2b must be
mentioned on the BETWEEN list with a level 2b cluster label. It can be
modeled on only level 2b. Consider a model where students are nested
in schools crossed with neighborhoods. Level 1 is student; level 2a is
school; and level 2b is neighborhood. Following is an example of how
to specify the BETWEEN option for TYPE=CROSSCLASSIFIED:

BETWEEN = (school) y1-y3 (neighbor) y4-y6;

In this example, y1, y2, and y3 are cluster-level variables measured on

level 2a, school, and modeled on only level 2a. Y4, y5, and y6 are
cluster-level variables measured on level 2b, neighborhood, and modeled
on only level 2b.

POPULATION, COVERAGE, AND STARTING

VALUES
The POPULATION, COVERAGE, and STARTING options are used as
population parameter values for data generation; population parameter
values for computing parameter coverage that are printed in the first
column of the output labeled Population; and as starting values for the
analysis. These values are the parameter estimates obtained from a
previous analysis where the parameter estimates are saved using the
ESTIMATES option of the SAVEDATA command. These options are
described below.

POPULATION

The POPULATION option is used to name the data set that contains the
population parameter values to be used in data generation. Following is
an example of how the POPULATION option is specified:

801
CHAPTER 19

POPULATION = estimates.dat;

where estimates.dat is a file that contains the parameter estimates from a

previous analysis of the model that is specified in the MODEL
POPULATION command.

COVERAGE

The COVERAGE option is used to name the data set that contains the
population parameter values to be used for computing parameter
coverage in the Monte Carlo summary. They are printed in the first
column of the output labeled Population. Following is an example of
how the COVERAGE option is specified:

COVERAGE = estimates.dat;

where estimates.dat is a file that contains the parameter estimates from a

previous analysis of the model that is specified in the MODEL
command.

STARTING

The STARTING option is used to name the data set that contains the
values to be used as starting values for the analysis. Following is an
example of how the STARTING option is specified:

STARTING = estimates.dat;

where estimates.dat is a file that contains the parameter estimates from a

previous analysis of the model that is specified in the MODEL
command.

SAVING DATA AND RESULTS

The REPSAVE, SAVE, RESULTS, and BPARAMETERS options are
used to save data and results. These options are described below.

REPSAVE

The REPSAVE option is used in conjunction with the SAVE option to

save some or all of the data sets generated in a Monte Carlo study. The

802
 MONTECARLO Command

REPSAVE option specifies the numbers of the replications for which the
data are saved. The keyword ALL can be used to save the data from all
of the replications. The list function is also available with REPSAVE.
To save the data from specific replications, REPSAVE is specified as
follows:

REPSAVE = 1 10-15 100;

which results in the data from replications 1, 10, 11, 12, 13, 14, 15, and
100 being saved. To save the data from all replications, REPSAVE is
specified as follows:

REPSAVE = ALL;

SAVE

The SAVE option is used to save data from the first replication for
future analysis. It is specified as follows:

SAVE = rep1.dat;

where rep1.dat is the name of the file in which data from the first
replication is saved. The data are saved using a free format.

The SAVE option can be used in conjunction with the REPSAVE option
to save data from any or all replications. When the SAVE option is used
with the REPSAVE option, it is specified as follows:

SAVE = rep*.dat;

where the asterisk (*) is replaced by the replication number. For

example, if replications 10 and 30 are saved, the data are stored in the
files rep10.dat and rep30.dat. A file is also produced that contains the
names of all of the data sets. To name this file, the asterisk (*) is
replaced by the word list. The file, in this case replist.dat, contains the
names of the generated data sets. The variables are not always saved in
the order that they appear in the NAMES statement.

803
CHAPTER 19

RESULTS

The RESULTS option is used to save the analysis results for each
replication of the Monte Carlo study in an ASCII file. The results saved
include the replication number, parameter estimates, standard errors, and
a set of fit statistics. The parameter estimates and standard errors are
saved in the order shown in the TECH1 output in free format delimited
by a space. The values are saved as E15.8. The RESULTS option is
specified as follows:

RESULTS = results.sav;

where results.sav is the name of the file in which the analysis results for
each replication will be saved.

BPARAMETERS

The BPARAMETERS option is used in Bayesian analysis to specify the

name of the ASCII file in which the Bayesian posterior parameter values
for all iterations are saved. Following is an example of how this option
is specified:

BPARAMETERS = bayes.dat;

where bayes.dat is the name of the file in which the Bayesian posterior
parameter values for all iterations will be saved.

804
 A Summary Of The Mplus Language

CHAPTER 20
A SUMMARY OF THE Mplus
LANGUAGE
This chapter contains a summary of the commands, options, and settings
of the Mplus language. For each command, default settings are found in
the last column. Commands and options can be shortened to four or
more letters. Option settings can be referred to by either the complete
word or the part of the word shown in bold type.

THE TITLE COMMAND

TITLE: title for the analysis

THE DATA COMMAND

DATA:
FILE IS file name;
FORMAT IS format statement; FREE
FREE;
TYPE IS INDIVIDUAL; INDIVIDUAL
COVARIANCE;
CORRELATION;
FULLCOV;
FULLCORR;
MEANS;
STDEVIATIONS;
MONTECARLO;
IMPUTATION;
NOBSERVATIONS ARE number of observations;
NGROUPS = number of groups; 1
LISTWISE = ON; OFF
OFF;
SWMATRIX = file name;
VARIANCES = CHECK; CHECK
NOCHECK;

805
CHAPTER 20

DATA IMPUTATION:
IMPUTE = names of variables for which missing values
will be imputed;
NDATASETS = number of imputed data sets; 5
SAVE = names of files in which imputed data sets
are stored;
MODEL = COVARIANCE; COVARIANCE
SEQUENTIAL;
REGRESSION;
VALUES = values imputed data can take; no restrictions
ROUNDING = number of decimals for imputed continuous 3
variables;
THIN = k where every k-th imputation is saved; 100
DATA WIDETOLONG:
WIDE = names of old wide format variables;
LONG = names of new long format variables;
IDVARIABLE = name of variable with ID information;
REPETITION = name of variable with repetition information;
DATA LONGTOWIDE:
LONG = names of old long format variables;
WIDE = names of new wide format variables;
IDVARIABLE = name of variable with ID information;
REPETITION = name of variable with repetition information
(values); 0, 1, 2, etc.
DATA TWOPART:
NAMES = names of variables used to create a set of
binary and continuous variables;
CUTPOINT = value used to divide the original variables 0
into a set of binary and continuous
BINARY = variables;
CONTINUOUS = names of new binary variables;
TRANSFORM = names of new continuous variables; LOG
function to use to transform new continuous
variables;
DATA MISSING:
NAMES = names of variables used to create a set of
binary variables;
BINARY = names of new binary variables;
TYPE = MISSING;
SDROPOUT;
DDROPOUT;
DESCRIPTIVE = sets of variables for additional descriptive
statistics separated by the | symbol;

806
 A Summary Of The Mplus Language

DATA SURVIVAL:
NAMES = names of variables used to create a set of
binary event-history variables;
CUTPOINT = value used to create a set of binary event-
history variables from a set of original
variables;
BINARY = names of new binary variables;
DATA COHORT:
COHORT IS name of cohort variable (values);
COPATTERN IS name of cohort/pattern variable (patterns);
COHRECODE = (old value = new value);
TIMEMEASURES = list of sets of variables separated by the |
symbol;
TNAMES = list of root names for the sets of variables in
TIMEMEASURES separated by the |
symbol;

THE VARIABLE COMMAND

VARIABLE:

NAMES ARE names of variables in the data set;

USEOBSERVATIONS ARE conditional statement to select observations; all observations
in data set
USEVARIABLES ARE names of analysis variables; all variables in
NAMES
MISSING ARE variable (#);
.;
*;
BLANK;
CENSORED ARE names, censoring type, and inflation status for
censored dependent variables;
CATEGORICAL ARE names of binary and ordered categorical
(ordinal) dependent variables;
NOMINAL ARE names of unordered categorical (nominal)
dependent variables;
COUNT ARE names of count variables (model);
DSURVIVAL ARE names of discrete-time survival variables;
GROUPING IS name of grouping variable (labels);
IDVARIABLE IS name of ID variable;
FREQWEIGHT IS name of frequency (case) weight variable;

807
CHAPTER 20

TSCORES ARE names of observed variables with information

on individually-varying times of observation;
AUXILIARY = names of auxiliary variables (function);
CONSTRAINT = names of observed variables that can be used
in the MODEL CONSTRAINT command;
PATTERN IS name of pattern variable (patterns);
STRATIFICATION IS name of stratification variable;
CLUSTER IS name of cluster variables;
WEIGHT IS name of sampling weight variable;
WTSCALE IS UNSCALED; CLUSTER
CLUSTER;
ECLUSTER;
BWEIGHT IS name of between-level sampling weight
variable;
B2WEIGHT IS name of the level 2 sampling weight variable;
B3WEIGHT IS name of the level 3 sampling weight variable;
BWTSCALE IS UNSCALED; SAMPLE
SAMPLE;
REPWEIGHTS ARE names of replicate weight variables;
SUBPOPULATION IS conditional statement to select subpopulation; all observations
in data set
FINITE = name of variable; FPC
name of variable (FPC);
name of variable (SFRACTION);
name of variable (POPULATION);
CLASSES = names of categorical latent variables (number
of latent classes);
KNOWNCLASS = name of categorical latent variable with known
class membership (labels);
TRAINING = names of training variables; MEMBERSHIP
names of variables (MEMBERSHIP);
names of variables (PROBABILITIES);
names of variables (PRIORS);
WITHIN ARE names of individual-level observed variables;
WITHIN ARE (label) names of individual-level observed variables;
BETWEEN ARE names of cluster-level observed variables;
BETWEEN ARE (label) names of cluster-level observed variables;
SURVIVAL ARE names and time intervals for time-to-event
variables;
TIMECENSORED ARE names and values of variables that contain (0 = NOT
right censoring information; 1 = RIGHT)

808
 A Summary Of The Mplus Language

THE DEFINE COMMAND

DEFINE:

variable = mathematical expression;

IF (conditional statement) THEN transformation

statements;

_MISSING
variable = MEAN (list of variables);
variable = SUM (list of variables);
CUT variable or list of variables (cutpoints);
variable = CLUSTER_MEAN (variable);
CENTER variable or list of variables (GRANDMEAN);
CENTER variable or list of variables (GROUPMEAN);
CENTER variable or list of variables (GROUPMEAN
label);
STANDARDIZE variable or list of variables;
DO (#, #) expression;

THE ANALYSIS COMMAND

ANALYSIS:

TYPE = GENERAL; GENERAL

BASIC;
RANDOM;
COMPLEX;
MIXTURE;
BASIC;
RANDOM;
COMPLEX;
TWOLEVEL;
BASIC;
RANDOM;
MIXTURE;
COMPLEX;
THREELEVEL;
BASIC;
RANDOM;
COMPLEX;

809
CHAPTER 20

CROSSCLASSIFIED;
RANDOM;
EFA # #;
BASIC;
MIXTURE;
COMPLEX;
TWOLEVEL;
EFA # # UW* # # UB*;
EFA # # UW # # UB;
ESTIMATOR = ML; depends on
MLM; analysis type
MLMV;
MLR;
MLF;
MUML;
WLS;
WLSM;
WLSMV;
ULS;
ULSMV;
GLS;
BAYES;
PARAMETERIZATION = DELTA; depends on
THETA; analysis type
LOGIT;
LOGLINEAR;
PROBABILITY;
LINK = LOGIT; LOGIT
PROBIT;
ROTATION = GEOMIN; GEOMIN
(OBLIQUE value)
GEOMIN (OBLIQUE value);
GEOMIN (ORTHOGONAL value);
QUARTIMIN; OBLIQUE
CF-VARIMAX; OBLIQUE
CF-VARIMAX (OBLIQUE);
CF-VARIMAX (ORTHOGONAL);
CF-QUARTIMAX; OBLIQUE
CF- QUARTIMAX (OBLIQUE);
CF- QUARTIMAX (ORTHOGONAL);
CF-EQUAMAX; OBLIQUE
CF- EQUAMAX (OBLIQUE);
CF- EQUAMAX (ORTHOGONAL);

810
 A Summary Of The Mplus Language

CF-PARSIMAX; OBLIQUE
CF- PARSIMAX (OBLIQUE);
CF- PARSIMAX (ORTHOGONAL);
CF-FACPARSIM; OBLIQUE
CF- FACPARSIM (OBLIQUE);
CF- FACPARSIM (ORTHOGONAL);
CRAWFER; OBLIQUE 1/p
CRAWFER (OBLIQUE value);
CRAWFER (ORTHOGONAL value);
OBLIMIN; OBLIQUE 0
OBLIMIN (OBLIQUE value);
OBLIMIN (ORTHOGONAL value);
VARIMAX;
PROMAX;
TARGET;
BI-GEOMIN;
BI-GEOMIN (OBLIQUE);
BI-GEOMIN (ORTHOGONAL);
BI-CF-QUARTIMAX;
BI-CF-QUARTIMAX (OBLIQUE);
BI-CF-QUARTIMAX (ORTHOGONAL);
ROWSTANDARDIZATION = CORRELATION; CORRELATION
KAISER;
COVARIANCE;
PARALLEL = number; 0
MODEL = NOMEANSTRUCTURE; means
NOCOVARIANCES; covariances
ALLFREE; equa;
REPSE = BOOTSTRAP;
JACKKNIFE;
JACKKNIFE1;
JACKKNIFE2;
BRR;
FAY (#); .3
BASEHAZARD = ON; OFF
OFF;
OFF (EQUAL);
OFF (UNEQUAL);
CHOLESKY = ON; depends on
OFF; analysis type
ALGORITHM = EM; depends on
EMA; analysis type
FS;
ODLL;
INTEGRATION;

811
CHAPTER 20

INTEGRATION = number of integration points; STANDARD

STANDARD (number of integration points) ; depends on
analysis type
GAUSSHERMITE (number of integration 15
points) ; depends on
MONTECARLO (number of integration points); analysis type
MCSEED = random seed for Monte Carlo integration; 0
ADAPTIVE = ON; ON
OFF;
INFORMATION = OBSERVED; depends on
EXPECTED; analysis type
COMBINATION;
BOOTSTRAP = number of bootstrap draws; STANDARD
number of bootstrap draws (STANDARD);
number of bootstrap draws (RESIDUAL):
LRTBOOTSTRAP = number of bootstrap draws for TECH14; depends on
analysis type
STARTS = number of initial stage starts and number of depends on
final stage optimizations; analysis type
STITERATIONS = number of initial stage iterations; 10
STCONVERGENCE = initial stage convergence criterion; 1
STSCALE = random start scale; 5
STSEED = random seed for generating random starts; 0
OPTSEED = random seed for analysis;
K-1STARTS = number of initial stage starts and number of 20 4
final stage optimizations for the k-1 class
model;
LRTSTARTS = number of initial stage starts and number of 0 0 40 8
final stage optimizations for TECH14;
RSTARTS = number of random starts for the rotation depends on
algorithm and number of factor solutions analysis type
printed for exploratory factor analysis;
DIFFTEST = file name;
MULTIPLIER = file name;
COVERAGE = minimum covariance coverage with missing .10
data;
ADDFREQUENCY = value divided by sample size to add to cells .5
with zero frequency;
ITERATIONS = maximum number of iterations for the Quasi- 1000
Newton algorithm for continuous outcomes;
SDITERATIONS = maximum number of steepest descent 20
iterations for the Quasi-Newton algorithm for
continuous outcomes;
H1ITERATIONS = maximum number of iterations for unrestricted 2000
model with missing data;

812
 A Summary Of The Mplus Language

MITERATIONS = number of iterations for the EM algorithm; 500

MCITERATIONS = number of iterations for the M step of the EM 1
algorithm for categorical latent variables;
MUITERATIONS = number of iterations for the M step of the EM 1
algorithm for censored, categorical, and count
outcomes;
RITERATIONS = maximum number of iterations in the rotation 10000
algorithm for exploratory factor analysis;
CONVERGENCE = convergence criterion for the Quasi-Newton depends on
algorithm for continuous outcomes; analysis type
H1CONVERGENCE = convergence criterion for unrestricted model .0001
with missing data;
LOGCRITERION = likelihood convergence criterion for the EM depends on
algorithm; analysis type
RLOGCRITERION = relative likelihood convergence criterion for the depends on
EM algorithm; analysis type
MCONVERGENCE = convergence criterion for the EM algorithm; depends on
analysis type
MCCONVERGENCE = convergence criterion for the M step of the EM .000001
algorithm for categorical latent variables;
MUCONVERGENCE = convergence criterion for the M step of the EM .000001
algorithm for censored, categorical, and count
outcomes;
RCONVERGENCE = convergence criterion for the rotation algorithm .00001
for exploratory factor analysis;
MIXC = ITERATIONS; ITERATIONS
CONVERGENCE;
M step iteration termination based on number
of iterations or convergence for categorical
latent variables;
MIXU = ITERATIONS; ITERATIONS
CONVERGENCE;
M step iteration termination based on number
of iterations or convergence for censored,
categorical, and count outcomes;

LOGHIGH = max value for logit thresholds; +15

LOGLOW = min value for logit thresholds; - 15
UCELLSIZE = minimum expected cell size; .01
VARIANCE = minimum variance value; .0001
MATRIX = COVARIANCE; COVARIANCE
CORRELATION;
POINT = MEDIAN; MEDIAN
MEAN;
MODE;

813
CHAPTER 20

CHAINS = number of MCMC chains; 2

BSEED = seed for MCMC random number generation; 0
STVALUES = UNPERTURBED; UNPERTURBED
PERTURBED;
ML;
MEDIATOR = LATENT; depends on
OBSERVED; analysis type
ALGORITHM = GIBBS; GIBBS (PX1)
GIBBS (PX1);
GIBBS (PX2);
GIBBS (PX3);
GIBBS (RW);
MH;
BCONVERGENCE = MCMC convergence criterion using Gelman- .05
Rubin;
BITERATIONS = maximum and minimum number of iterations 50,000 0
for each MCMC chain when Gelman-Rubin
PSR is used;
FBITERATIONS = fixed number of iterations for each MCMC
chain when Gelman-Rubin PSR is not used;
THIN = k where every k-th MCMC iteration is saved; 1
DISTRIBUTION = maximum number of iterations used to 10,000
compute the Bayes multivariate mode;
KOLMOGOROV = number of draws from the MCMC chains; 100
PRIOR = number of draws from the prior distribution; 1000
INTERACTIVE = file name;
PROCESSORS = number of processors; 1
number of processors (STARTS);

THE MODEL COMMAND

MODEL:
BY short for measured by -- defines latent variables
example: f1 BY y1-y5;
ON short for regressed on -- defines regression relationships
example: f1 ON x1-x9;
PON short for regressed on -- defines paired regression relationships
example: f2 f3 PON f1 f2;
WITH short for correlated with -- defines correlational relationships
example: f1 WITH f2;
PWITH short for correlated with -- defines paired correlational
relationships
example: f1 f2 f3 PWITH f4 f5 f6;
list of variables; refers to variances and residual variances
example: f1 y1-y9;

814
 A Summary Of The Mplus Language

[list of variables]; refers to means, intercepts, thresholds

example: [f1, y1-y9];
* frees a parameter at a default value or a specific starting value
example: y1* y2*.5;
@ fixes a parameter at a default value or a specific value
example: y1@ y2@0;
(number) constrains parameters to be equal
example: f1 ON x1 (1);
f2 ON x2 (1);
variable$number label for the threshold of a variable
variable#number label for nominal observed or categorical latent variable
variable#1 label for censored or count inflation variable
variable#number label for baseline hazard parameters
variable#number label for a latent class
(name) label for a parameter
{list of variables}; refers to scale factors
example: {y1-y9};
| names and defines random effect variables
example: s | y1 ON x1;
AT short for measured at -- defines random effect variables
example: s | y1-y4 AT t1-t4;
XWITH defines interactions between variables;
MODEL INDIRECT: describes the relationships for which indirect and total effects
are requested
IND describes a specific indirect effect or a set of indirect effects;
VIA describes a set of indirect effects that includes specific
mediators;
MODEL CONSTRAINT: describes linear and non-linear constraints on parameters
NEW assigns labels to parameters not in the analysis model;
DO describes a do loop
PLOT describes y-axis variables
LOOP describes x-axis variables
MODEL TEST: describes restrictions on the analysis model for the Wald test
MODEL PRIORS: specifies the prior distribution for the parameters
COVARIANCE assigns a prior to the covariance between two parameters
DO describes a do loop
DIFFERENCE assigns priors to differences between parameters
MODEL: describes the analysis model
MODEL label: describes the group-specific model in multiple group analysis
and the model for each categorical latent variable and
combinations of categorical latent variables in mixture modeling
MODEL:
%OVERALL% describes the overall part of a mixture model
%class label% describes the class-specific part of a mixture model

815
CHAPTER 20

MODEL:
%WITHIN% describes the individual-level model
%BETWEEN% describes the cluster-level model for a two-level model
%BETWEEN label% describes the cluster-level model for a three-level or cross-
classified model
MODEL POPULATION: describes the data generation model for a Monte Carlo study
MODEL POPULATION-label: describes the group-specific data generation model in multiple
group analysis and the data generation model for each
categorical latent variable and combinations of categorical
latent variables in mixture modeling for a Monte Carlo study
MODEL POPULATION:
%OVERALL% describes the overall data generation model of a mixture
%class label% model
describes the class-specific data generation model of a mixture
model
MODEL POPULATION:
%WITHIN% describes the individual-level data generation model for a
multilevel model
%BETWEEN% describes the cluster-level data generation model for a two-
%BETWEEN label% level model
describes the cluster-level data generation model for a three-
level or cross-classified model
MODEL COVERAGE: describes the population parameter values for a Monte Carlo
study
MODEL COVERAGE-label: describes the group-specific population parameter values in
multiple group analysis and the population parameter values for
each categorical latent variable and combinations of categorical
latent variables in mixture modeling for a Monte Carlo study
MODEL COVERAGE:
%OVERALL% describes the overall population parameter values of a mixture
model for a Monte Carlo study
%class label% describes the class-specific population parameter values of a
mixture model

816
 A Summary Of The Mplus Language

MODEL COVERAGE:
%WITHIN% describes the individual-level population parameter values for
coverage
%BETWEEN% describes the cluster-level population parameter values for a
%BETWEEN label% two-level model for coverage
describes the cluster-level population parameter values for a
three-level or cross-classified model for coverage
MODEL MISSING: describes the missing data generation model for a Monte Carlo
study
MODEL MISSING-label: describes the group-specific missing data generation model for
a Monte Carlo study
MODEL MISSING:
%OVERALL% describes the overall data generation model of a mixture model
describes the class-specific data generation model of a mixture
%class label% model

THE OUTPUT COMMAND

OUTPUT:

SAMPSTAT;
CROSSTABS; ALL
CROSSTABS (ALL);
CROSSTABS (COUNT);
CROSSTABS (%ROW);
CROSSTABS (%COLUMN);
CROSSTABS (%TOTAL);
STANDARDIZED; ALL
STANDARDIZED (ALL);
STANDARDIZED (STDYX);
STANDARDIZED (STDY);
STANDARDIZED (STD);
RESIDUAL;
MODINDICES (minimum chi-square); 10
MODINDICES (ALL);
MODINDICES (ALL minimum chi-square); 10
CINTERVAL; SYMMETRIC
CINTERVAL (SYMMETRIC);
CINTERVAL (BOOTSTRAP);
CINTERVAL (BCBOOTSTRAP);
CINTERVAL (EQTAIL); EQTAIL
CINTERVAL (HPD);
SVALUES;
NOCHISQUARE;

817
CHAPTER 20

NOSERROR;
H1SE;
H1TECH3;
PATTERNS;
FSCOEFFICIENT;
FSDETERMINACY;
BASEHAZARD;
LOGRANK;
D(#);
TECH1;
TECH2;
TECH3;
TECH4;
TECH5;
TECH6;
TECH7;
TECH8;
TECH9;
TECH10;
TECH11;
TECH12;
TECH13;
TECH14;
TECH15;
TECH16;

THE SAVEDATA COMMAND

SAVEDATA:

FILE IS file name;

FORMAT IS format statement; F10.3
FREE;
MISSFLAG = missing value flag; *
RECORDLENGTH IS characters per record; 1000

SAMPLE IS file name;

COVARIANCE IS file name;
SIGBETWEEN IS file name;
SWMATRIX IS file name;
RESULTS ARE file name;
ESTIMATES ARE file name;
DIFFTEST IS file name;
TECH3 IS file name;
TECH4 IS file name;

818
 A Summary Of The Mplus Language

KAPLANMEIER IS file name;

BASEHAZARD IS file name;
ESTBASELINE IS file name;
RESPONSE IS file name;
MULTIPLIER IS file name;
BPARAMETERS IS file name;
TYPE IS COVARIANCE; varies
CORRELATION;
SAVE = FSCORES;
FSCORES (#);
LRESPONSES (#);
CPROBABILITIES;
REPWEIGHTS;
MAHALANOBIS;
LOGLIKELIHOOD;
INFLUENCE;
COOKS;

FACTORS = names of factors;

LRESPONSES = names of latent response variables;

MFILE = file name;

MNAMES = names of variables in the data set;
MFORMAT = format statement; FREE
FREE;
MMISSING = Variable (#);
*;
.;
MSELECT = names of variables; all variables in
MNAMES

THE PLOT COMMAND

PLOT:

TYPE IS PLOT1;
PLOT2;
PLOT3;
SERIES IS list of variables in a series plus x-axis
values;
FACTORS ARE names of factors (#);
LRESPONSES ARE names of latent response variables (#);

819
CHAPTER 20

OUTLIERS ARE MAHALANOBIS;

LOGLIKELIHOOD;
INFLUENCE;
COOKS;
MONITOR IS ON; OFF
OFF;

THE MONTECARLO COMMAND

MONTECARLO:

NAMES = names of variables;

NOBSERVATIONS = number of observations;
NGROUPS = number of groups; 1
NREPS = number of replications; 1
SEED = random seed for data generation; 0
GENERATE = scale of dependent variables for data
generation;
CUTPOINTS = thresholds to be used for categorization of
covariates
GENCLASSES = names of categorical latent variables (number
of latent classes used for data generation);
NCSIZES = number of unique cluster sizes for each group
separated by the | symbol;
CSIZES = number (cluster size) for each group
separated by the | symbol;
HAZARDC = specifies the hazard for the censoring
process;
PATMISS = missing data patterns and proportion missing
for each dependent variable;
PATPROBS = proportion for each missing data pattern;
MISSING = names of dependent variables that have
missing data;
CENSORED ARE names and limits of censored-normal
dependent variables;
CATEGORICAL ARE names of ordered categorical dependent
variables;
NOMINAL ARE names of unordered categorical dependent
variables;
COUNT ARE names of count variables;
CLASSES = names of categorical latent variables (number
of latent classes used for model estimation);

820
 A Summary Of The Mplus Language

AUXILIARY = names of auxiliary variables (R);

names of auxiliary variables (R3STEP);
names of auxiliary variables (E);
names of auxiliary variables (DU3STEP);
names of auxiliary variables (DE3STEP);
SURVIVAL = names and time intervals for time-to-event
variables;
TSCORES = names, means, and standard deviations of
observed variables with information on
individually-varying times of observation;
WITHIN = names of individual-level observed variables;
BETWEEN = names of cluster-level observed variables;
POPULATION = name of file containing population parameter
values for data generation;
COVERAGE = name of file containing population parameter
values for computing parameter coverage;
STARTING = name of file containing parameter values for
use as starting values for the analysis;
REPSAVE = numbers of the replications to save data from
or ALL;
SAVE = name of file in which generated data are
stored;
RESULTS = name of file in which analysis results are
stored;

821
CHAPTER 20

822
 REFERENCES
Agresti, A. (1996). An introduction to categorical data analysis. New York: John Wiley & Sons.

Agresti, A. (2002). Categorical data analysis. Second Edition. New York: John Wiley & Sons.

Aitkin, M. (1999). A general maximum likelihood analysis of variance components in

generalized linear models. Biometrics, 55, 117-128.

Asparouhov, T. (2005). Sampling weights in latent variable modeling. Structural Equation

Modeling, 12, 411-434.

Asparouhov, T. (2006). General multi-level modeling with sampling weights. Communications in

Statistics: Theory and Methods, 35, 439-460.

Asparouhov, T. (2007). Wald test of mean equality for potential latent class predictors in mixture
modeling. Technical appendix. Los Angeles: Muthén & Muthén.

Asparouhov, T. & Muthén, B. (2005). Multivariate statistical modeling with survey data.
Proceedings of the FCMS 2005 Research Conference.

Asparouhov, T. & Muthén, B. (2006a). Multilevel modeling of complex survey data.

Proceedings of the Joint Statistical Meeting in Seattle, August 2006. ASA Section on Survey
Research Methods, 2718-2726.

Asparouhov, T. & Muthén, B. (2006b). Constructing covariates in multilevel regression. Mplus

Web Notes: No. 11. www.statmodel.com.

Asparouhov, T. & Muthén, B. (2007). Computationally efficient estimation of multilevel high-

dimensional latent variable models. Proceedings of the Joint Statistical Meeting in Salt Lake
City, August 2007. ASA section on Biometrics.

Asparouhov, T. & Muthén, B. (2008a). Multilevel mixture models. In G.R. Hancock, & K.M.
Samuelson (eds.), Advances in latent variable mixture models. Charlotte, NC: Information Age
Publishing, Inc.

Asparouhov, T. & Muthén, B. (2008b). Auxiliary variables predicting missing data. Technical
appendix. Los Angeles: Muthén & Muthén.

Asparouhov, T. & Muthén, B. (2008c). Chi-square statistics with multiple imputation. Technical
appendix. Los Angeles: Muthén & Muthén.

823
Asparouhov, T. & Muthén, B. (2009a). Exploratory structural equation modeling. Structural
Equation Modeling, 16, 397-438.

Asparouhov, T. & Muthén, B. (2009b). Resampling methods in Mplus for complex survey data.
Technical appendix. Los Angeles: Muthén & Muthén.

Asparouhov, T. & Muthén, B. (2010a). Weighted least squares estimation with missing data.
Technical appendix. Los Angeles: Muthén & Muthén.

Asparouhov, T. & Muthén, B. (2010b). Bayesian analysis using Mplus: Technical

implementation. Technical Report. Los Angeles: Muthén & Muthén.

Asparouhov, T. & Muthén, B. (2012a). General random effect latent variable modeling:
Random subjects, items, contexts, and parameters. Technical Report. Los Angeles: Muthén &
Muthén.

Asparouhov, T. & Muthén, B. (2012b). Auxiliary variables in mixture modeling: A 3-step

approach using Mplus. Mplus Web Notes: No. 15. www.statmodel.com.

Asparouhov, T. & Muthén, B. (2012c). Using Mplus TECH11 and TECH14 to test the number
of latent classes. Mplus Web Notes: No. 14. www.statmodel.com.

Asparouhov, T., Masyn, K. & Muthén, B. (2006). Continuous time survival in latent variable
models. Proceedings of the Joint Statistical Meeting in Seattle, August 2006. ASA section on
Biometrics, 180-187.

Baker, F.B. & Kim, S. (2004). Item response theory. Parameter estimation techniques. Second
edition. New York: Marcel Dekker, Inc.

Bauer, D.J. & Curran, P.J. (2005). Probing interactions in fixed and multilevel regression:
Inferential and graphical techniques. Multivariate Behavioral Research, 40, 373-400.

Bauer, D.J., Preacher, K.J., & Gil, K.M. (2006). Conceptualizing and testing random indirect
effects and moderated mediation in multilevel models: New procedures and recommendations.
Psychological Methods, 11, 142-163.

Bernaards, C.A. & Jennrich, R.I. (2005). Gradient projection algorithms and software for
arbitrary rotation criteria in factor analysis. Educational and Psychological Measurement, 65,
676-696.

824
Bijmolt, T.H.A., Paas, L.J., & Vermunt, J.K. (2004). Country and consumer segmentation:
Multi-level latent class analysis of financial product ownership. International Journal of Research
in Marketing, 21, 323-340.

Bollen, K.A. (1989). Structural equations with latent variables. New York: John Wiley & Sons.

Bollen, K.A. & Stein, R.A. (1992). Bootstrapping goodness-of-fit measures in structural equation
models. Sociological Methods & Research, 21, 205-229.

Boscardin, J., Zhang, X., & Belin, T. (2008). Modeling a mixture of ordinal and continuous
repeated measures. Journal of Statistical Computation and Simulation, 78, 873-886.

Browne, M.W. (2001). An overview of analytic rotation in exploratory factor analysis.

Multivariate Behavioral Research, 36, 111-150.

Browne, M.W. & Arminger, G. (1995). Specification and estimation of mean- and covariance-
structure models. In G. Arminger, C.C. Clogg & M.E. Sobel (eds.), Handbook of statistical
modeling for the social and behavioral sciences (pp. 311-359). New York: Plenum Press.

Browne, W.J. & Draper, D. (2006). A comparison of Bayesian and likelihood-based methods for
fitting multilevel models. Bayesian Analysis, 3, 473-514.

Browne, M.W., Cudeck, R., Tateneni, K., & Mels, G. (2004). CEFA: Comprehensive
Exploratory Factor Analysis, Version 2.00 [Computer software and manual]. Retrieved from
http://quantrm2.psy.ohio-state.edu/browne/.

Chib, S. & Greenberg, E. (1998). Bayesian analysis of multivariate probit models. Biometrika
85, 347-361.

Collins, L.M. & Lanza, S.T. (2010). Latent class and latent transition analysis. Hoboken, N.J.:
John Wiley & Sons.

Collins, L.M, Schafer, J.L., & Kam, C-H (2001). A comparison of inclusive and restrictive
strategies in modern missing data procedures. Psychological Methods, 6, 330-351.

Cook, R.D. (1977). Detection of influential observations in linear regression. Technometrics, 19,
15-18.

Cook, R.D. & Weisberg, S. (1982). Residuals and influence in regression. New York: Chapman
& Hall.

825
Cudeck, R. & O'Dell, L.L. (1994). Applications of standard error estimates in unrestricted factor
analysis: Significance tests for factor loadings and correlations. Psychological Bulletin, 115,
475-487.

Demirtas, H. & Schafer, J.L. (2003). On the performance of random-coefficient pattern-mixture

models for non-ignorable drop-out. Statistics in Medicine, 22, 2553-2575.

Dempster, A.P., Laird, N.M., & Rubin, D.B. (1977). Maximum likelihood from incomplete data
via the EM algorithm. Journal of the Royal Statistical Society, ser. B, 39, 1-38.

Diggle, P.D. & Kenward, M.G. (1994). Informative drop-out in longitudinal data analysis (with
discussion). Applied Statistics, 43, 49-73.

du Toit, M. (ed.) (2003). IRT from SSI: BILOG-MG MULTILOG PARSCALE TESTFACT.
Lincolnwood, IL: Scientific Software International, Inc.

Efron, B. & Tibshirani, R.J. (1993). An introduction to the bootstrap. New York: Chapman &
Hall.

Enders, C.K. (2002). Applying the Bollen-Stine bootstrap for goodness-of-fit measures to
structural equation models with missing data. Multivariate Behavioral Research, 37, 359-377.

Enders, C.K. (2010). Applied missing data analysis. New York: Guilford Press.

Everitt, B.S. & Hand, D.J. (1981). Finite mixture distributions. London: Chapman & Hall.

Hagenaars, J.A. & McCutcheon, A.L. (2002). Applied latent class analysis. Cambridge, UK:
Cambridge University Press.

Fabrigar, L.R., Wegener, D.T., MacCallum, R.C., & Strahan, E.J. (1999). Evaluating the use of
exploratory factor analysis in psychological research. Psychological Methods, 4, 272-299.

Fay, R.E. (1989). Theoretical application of weighting for variance calculation. Proceedings of
the Section on Survey Research Methods of the American Statistical Association, 212-217.

Fox, J. P. (2010). Bayesian item response modeling. Theory and applications. New York:
Springer.

Gelman, A. (2006). Prior distributions for variance parameters in hierarchical models. Bayesian
Analysis, 3, 515-533.

826
Gelman, A. & Rubin, D.B. (1992). Inference from iterative simulation using multiple sequences
(with discussion). Statistical Science, 7, 457-511.

Gelman, A., Carlin, J.B., Stern, H.S., and Rubin, D.B. (2004). Bayesian data analysis. Second
edition. New York: Chapman & Hall.

Graham, J.W. (2003). Adding missing-data relevant variables to FIML-based structural equation
models. Structural Equation Modeling: A Multidisciplinary Journal, 10, 80-100.

Hayton, J.C., Allen, D.G., & Scarpello, V. (2004). Factor retention decisions in exploratory
factor analysis: A tutorial on parallel analysis. Organizational Research Methods, 7, 191-205.

Hedeker, D. & Gibbons, R.D. (1994). A random-effects ordinal regression model for multilevel
analysis. Biometrics, 50, 933-944.

Hedeker, D. & Gibbons, R.D. (1997). Application of random-effects pattern-mixture models for
missing data in longitudinal studies. Psychological Methods, 2, 64-78.

Hilbe, J.M. (2011). Negative binomial regression. Second edition. New York: Cambridge
University Press.

Hildreth, C. & Houck, J.P. (1968). Some estimates for a linear model with random coefficients.
Journal of the American Statistical Association, 63, 584-595.

Hosmer, D.W. & Lemeshow, S. (2000). Applied logistic regression. Second edition. New York:
John Wiley & Sons.

Hougaard, P. (2000). Analysis of multivariate survival data. New York: Springer.

Imai, K., Keele, L., & Tingley, D. (2010). A general approach to causal mediation analysis.
Psychological Methods, 15, 309-334.

Jedidi, K., Jagpal. H.S. & DeSarbo, W.S. (1997). Finite-mixture structural equation models for
response-based segmentation and unobserved heterogeneity. Marketing Science, 16, 39-59.

Jeffries, N.O. (2003). A note on ‘testing the number of components in a normal mixture’.
Biometrika, 90, 991-994.

Jennrich, R.I. (1973). Standard errors for obliquely rotated factor loadings. Psychometrika, 38,
593-604.

827
Jennrich, R.I. (1974). Simplified formulae for standard errors in maximum-likelihood factor
analysis. The British Journal of Mathematical and Statistical Psychology, 27, 122-131.

Jennrich, R.I. (2007). Rotation methods, algorithms, and standard errors. In R. Cudeck & R.C.
MacCallum (eds.). Factor analysis at 100. Historical developments and future directions (pp.
315-335). Mahwah, New Jersey: Lawrence Erlbaum Associates, Inc.

Jennrich, R.I. & Bentler, P.M. (2011). Exploratory bi-factor analysis. Psychometrika, 76, 537-
549.

Jennrich, R.I. & Bentler, P.M. (2012). Exploratory bi-factor analysis: The oblique case.
Psychometrika, 77, 442-454.

Jennrich, R.I. & Sampson, P.F. (1966). Rotation for simple loadings. Psychometrika, 31, 313-
323.

Johnston, J. (1984). Econometric methods. Third edition. New York: McGraw-Hill.

Joreskog, K.G. & Sorbom, D. (1979). Advances in factor analysis and structural equation
models. Cambridge, MA: Abt Books.

Kaplan, D. (2008). An overview of Markov chain methods for the study of stage-sequential
developmental processes. Developmental Psychology, 44, 457-467.

Kenward, M.G. & Molenberghs, G. (1998). Likelihood based frequentist inference when data are
missing at random. Statistical Science, 13, 236-247.

Klein, A. & Moosbrugger, H. (2000). Maximum likelihood estimation of latent interaction

effects with the LMS method. Psychometrika, 65, 457-474.

Klein J.P. & Moeschberger, M.L. (1997). Survival analysis: Techniques for censored and
truncated data. New York: Springer.

Korn, E.L. & Graubard, B.I. (1999). Analysis of health surveys. New York: John Wiley & Sons.

Kreuter, F. & Muthén, B. (2008). Analyzing criminal trajectory profiles: Bridging multilevel and
group-based approaches using growth mixture modeling. Journal of Quantitative Criminology,
24, 1-31.

Langeheine, R. & van de Pol, F. (2002). Latent Markov chains. In J.A. Hagenaars & A.L.
McCutcheon (eds.), Applied latent class analysis (pp. 304-341). Cambridge, UK: Cambridge
University Press.

828
Larsen, K. (2004). Joint analysis of time-to-event and multiple binary indicators of latent classes.
Biometrics 60, 85-92.

Larsen, K. (2005). The Cox proportional hazards model with a continuous latent variable
measured by multiple binary indicators. Biometrics, 61, 1049-1055.

Lee, S.Y. (2007). Structural equation modeling. A Bayesian approach. New York: John Wiley
& Sons.

Little, R.J. (1995). Modeling the drop-out mechanism in repeated-measures studies. Journal of
the American Statistical Association, 90, 1112-1121.

Little, R.J. & Rubin, D.B. (2002). Statistical analysis with missing data. Second edition. New
York: John Wiley & Sons.

Little, R.J. & Yau, L.H.Y. (1998). Statistical techniques for analyzing data from prevention
trials: Treatment of no-shows using Rubin’s causal model. Psychological Methods, 3, 147-159.

Lo, Y., Mendell, N.R. & Rubin, D.B. (2001). Testing the number of components in a normal
mixture. Biometrika, 88, 767-778.

Lohr, S.L. (1999). Sampling: Design and analysis. Pacific Grove, CA: Brooks/Cole Publishing
Company.

Long, J.S. (1997). Regression models for categorical and limited dependent variables. Thousand
Oaks, CA: Sage Publications, Inc.

Lüdtke, O., Marsh, H.W., Robitzsch, A., Trautwein, U., Asparouhov, T., & Muthén, B. (2008).
The multilevel latent covariate model: A new, more reliable approach to group-level effects in
contextual studies. Psychological Methods, 13, 203-229.

MacKinnon, D.P. (2008). Introduction to statistical mediation analysis. New York: Lawrence
Erlbaum Associates.

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.

MacKinnon, D. P., Lockwood, C. M., Brown, C. H., Wang, W., & Hoffman, J. M. (2007). The
intermediate endpoint effect in logistic and probit regression. Clinical Trials, 4, 499-513.

829
Mantel, N. (1966). Evaluation of survival data and two new rank order statistics arising in its
consideration. Cancer Chemotherapy Reports, 50, 163-170.

Marlow, A.J., Fisher, S.E., Francks, C., MacPhie, I.L., Cherny, S.S., Richardson, A.J., Talcott,
J.B., Stein, J.F., Monaco, A.P., & Cardon, L.R. (2003). Use of multivariate linkage analysis for
dissection of a complex cognitive trait. American Journal of Human Genetics, 72, 561-570.

McCutcheon, A.L. (2002). Basic concepts and procedures in single- and multiple-group latent
class analysis. In J.A. Hagenaars & A.L. McCutcheon (eds.), Applied latent class analysis (pp.
56-85). Cambridge, UK: Cambridge University Press.

McDonald, R.P. (1967). Nonlinear factor analysis. Psychometric Monograph Number 15.
University of Chicago. Richmond, VA: The William Byrd Press.

McLachlan, G. & Peel, D. (2000). Finite mixture models. New York: John Wiley & Sons.

McLachlan, G.J., Do, K.A., & Ambroise, C. (2004). Analyzing microarray gene expression data.
New York: John Wiley & Sons.

Mislevy, R.J., Johnson, E.G., & Muraki, E. (1992). Scaling procedures in NAEP. Journal of
Educational Statistics, 17, 131-154.

Mooijaart, A. (1998). Log-linear and Markov modeling of categorical longitudinal data. In

C.C.J.H. Bijleveld & T. van der Kamp, Longitudinal data analysis: Designs, models, and
methods (pp. 318-370). Newbury Park, CA: Sage Publications.

Muthén, B. (1978). Contributions to factor analysis of dichotomous variables. Psychometrika,

43, 551-560.

Muthén, B. (1984). A general structural equation model with dichotomous, ordered categorical,
and continuous latent variable indicators. Psychometrika, 49, 115-132.

Muthén, B. (1989). Latent variable modeling in heterogeneous populations. Psychometrika, 54,

557-585.

Muthén, B. (1990). Mean and covariance structure analysis of hierarchical data. Paper presented
at the Psychometric Society meeting in Princeton, NJ, June 1990. UCLA Statistics Series 62.

Muthén, B. (1994). Multilevel covariance structure analysis. In J. Hox & I. Kreft (eds.),
Multilevel Modeling, a special issue of Sociological Methods & Research, 22, 376-398.

830
Muthén, B. (1997). Latent variable modeling with longitudinal and multilevel data. In A.
Raftery (ed.), Sociological Methodology (pp. 453-480). Boston: Blackwell Publishers.

Muthén, B. (2002). Beyond SEM: General latent variable modeling. Behaviormetrika, 29, 81-
117.

Muthén, B. (2004). Latent variable analysis: Growth mixture modeling and related techniques for
longitudinal data. In D. Kaplan (ed.), Handbook of quantitative methodology for the social
sciences (pp. 345-368). Newbury Park, CA: Sage Publications.

Muthén, B. (2006). Should substance use disorders be considered as categorical or dimensional?

Addiction, 101 (Suppl. 1), 6-16.

Muthén, B. (2008). Latent variable hybrids: Overview of old and new models. In Hancock, G.
R., & Samuelsen, K. M. (Eds.), Advances in latent variable mixture models, pp. 1-24. Charlotte,
NC: Information Age Publishing, Inc.

Muthén, B. (2011). Applications of causally defined direct and indirect effects in mediation
analysis using SEM in Mplus. Technical Report. Los Angeles: Muthén & Muthén.

Muthén, B. and Asparouhov, T. (2002). Latent variable analysis with categorical outcomes:
Multiple-group and growth modeling in Mplus. Mplus Web Notes: No. 4. www.statmodel.com.

Muthén, B. & Asparouhov, T. (2006). Item response mixture modeling: Application to tobacco
dependence criteria. Addictive Behaviors, 31, 1050-1066.

Muthén, B. & Asparouhov, T. (2007). Non-parametric hierarchical regressions. In preparation.

Muthén, B. & Asparouhov, T. (2009). Growth mixture modeling: Analysis with non-Gaussian
random effects. In Fitzmaurice, G., Davidian, M., Verbeke, G. & Molenberghs, G. (eds.),
Longitudinal Data Analysis, pp. 143-165. Boca Raton: Chapman & Hall/CRC Press.

Muthén, B. & Asparouhov, T. (2011). LTA in Mplus: Transition probabilities influenced by

covariates. Mplus Web Notes: No. 13. www.statmodel.com.

Muthén, B. & Asparouhov, T. (2012). Bayesian SEM: A more flexible representation of

substantive theory. Psychological Methods, 17, 313-335.

Muthén, B. & Christoffersson, A. (1981). Simultaneous factor analysis of dichotomous variables

in several groups. Psychometrika, 46, 407-419.

831
Muthén, B. & Masyn, K. (2005). Discrete-time survival mixture analysis. Journal of Educational
and Behavioral Statistics, 30, 27-28.

Muthén, L.K. & Muthén, B. (2002). How to use a Monte Carlo study to decide on sample size
and determine power. Structural Equation Modeling, 4, 599-620.

Muthén, B. & Satorra, A. (1995). Complex sample data in structural equation modeling. In P.
Marsden (ed.), Sociological Methodology 1995, 216-316.

Muthén, B. & Shedden, K. (1999). Finite mixture modeling with mixture outcomes using the EM
algorithm. Biometrics, 55, 463-469.

Muthén, B., du Toit, S.H.C. & Spisic, D. (1997). Robust inference using weighted least squares
and quadratic estimating equations in latent variable modeling with categorical and continuous
outcomes. Unpublished manuscript.

Muthén, B., Jo., B., & Brown, H. (2003). Comment on the Barnard, Frangakis, Hill, & Rubin
article, Principal stratification approach to broken randomized experiments: A case study of
school choice vouchers in New York City. Journal of the American Statistical Association, 98,
311-314.

Muthén, B., Asparouhov, T. & Rebollo, I. (2006). Advances in behavioral genetics modeling
using Mplus: Applications of factor mixture modeling to twin data. Twin Research and Human
Genetics, 9, 313-324.

Muthén, B., Asparouhov, T., Boye, M.E., Hackshaw, M.D., & Naegeli, A.N. (2009).
Applications of continuous-time survival in latent variable models for the analysis of oncology
randomized clinical trial data using Mplus. Technical Report. www.statmodel.com.

Muthén, B., Asparouhov, T., Hunter, A., & Leuchter, A. (2011). Growth modeling with non-
ignorable dropout: Alternative analyses of the STAR*D antidepressant trial. Psychological
Methods, 16, 17-33.

Muthén, B., Brown, C.H., Masyn, K., Jo, B., Khoo, S.T., Yang, C.C., Wang, C.P., Kellam, S.,
Carlin, J., & Liao, J. (2002). General growth mixture modeling for randomized preventive
interventions. Biostatistics, 3, 459-475.

Nagin, D.S. (1999). Analyzing developmental trajectories: A semi-parametric, group-based

approach. Psychological Methods, 4, 139-157.

Neale, M.C. & Cardon, L.R. (1992). Methodology for genetic studies of twins and families. The
Netherlands: Kluwer Academic Publishers.

832
Nylund, K. (2007). Latent transition analysis: Modeling extensions and an application to peer
victimization. Doctoral dissertation, University of California, Los Angeles.
www.statmodel.com.

Nylund, K.L., 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.

Olsen, M.K. & Schafer, J.L. (2001). A two-part random-effects model for semicontinuous
longitudinal data. Journal of the American Statistical Association, 96, 730-745.

Qu, Y., Tan, M., & Kutner, M.H. (1996). Random effects models in latent class analysis for
evaluating accuracy of diagnostic tests. Biometrics, 52, 797-810.

Posthuma, D., de Geus, E.J.C., Boomsma, D.I., & Neale, M.C. (2004). Combined linkage and
association tests in Mx. Behavior Genetics, 34, 179-196.

Pothoff, R.F., Woodbury, M.A., & Manton, K.G. (1992). “Equivalent sample size” and
“equivalent degrees of freedom” refinements for inference using survey weights under
superpopulation models. Journal of the American Statistical Association, 87, 383-396.

Preacher, K.J., Rucker, D.D., & Hayes, A.F. (2007). Addressing moderated mediation
hypotheses: Theory, methods, and prescriptions. Multivariate Behavioral Research, 42, 185-227.

Prescott, C.A. (2004). Using the Mplus computer program to estimate models for continuous and
categorical data from twins. Behavior Genetics, 34, 17- 40.

Raghunathan, T.E., Lepkowski, J.M., Van Hoewyk, J., & Solenberger, P. (2001). A multivariate
technique for multiply imputing missing values using a sequence of regression models. Survey
Methodology, 27, 85-95.

Raudenbush, S.W. & Bryk, A.S. (2002). Hierarchical linear models: Applications and data
analysis methods. Second edition. Newbury Park, CA: Sage Publications.

Reboussin, B.A., Reboussin, D.M., Liang, K.L., & Anthony, J.C. (1998). Latent transition
modeling of progression of health-risk behavior. Multivariate Behavioral Research, 33, 457-478.

Roeder, K., Lynch, K.G., & Nagin, D.S. (1999). Modeling uncertainty in latent class
membership: A case study in criminology. Journal of the American Statistical Association, 94,
766-776.

833
Rousseeuw P.J. & Van Zomeren B.C. (1990). Unmasking multivariate outliers and leverage
points. Journal of the American Statistical Association. 85, 633-651.

Rubin, D.B. (1987). Multiple imputation for nonresponse in surveys. New York: John Wiley &
Sons.

Schafer, J.L. (1997). Analysis of incomplete multivariate data. London: Chapman & Hall.

Singer, J.D. & Willett, J.B. (2003). Applied longitudinal data analysis: Modeling change and
event occurrence. New York: Oxford University Press.

van Buuren, S. (2007). Multiple imputation of discrete and continuous data by fully conditional
specification. Statistical Methods in Medical Research, 16, 219-242.

Verhagen, J. & Fox, J.P. (2012). Bayesian tests of measurement invariance. British Journal of
Mathematical and Statistical Psychology. Accepted for publication.

Vermunt, J.K. (2003). Multilevel latent class models. In R.M. Stolzenberg (ed.), Sociological
Methodology 2003 (pp. 213-239). Washington, D.C.: ASA.

Vermunt, J.K. (2010). Latent class modeling with covariates: Two improved three-step
approaches. Political Analysis, 18, 450-469.

von Davier, M., Gonzalez, E., & Mislevy, R.J. (2009). What are plausible values and why are
they useful? IERI Monograph Series, 2, 9-36.

Yates, A. (1987). Multivariate exploratory data analysis: A perspective on exploratory factor

analysis. Albany: State University of New York Press.

Yuan, K.H. & Bentler, P.M. (2000). Three likelihood-based methods for mean and covariance
structure analysis with nonnormal missing data. In M.E. Sobel & M.P. Becker (eds.),
Sociological Methodology 2000 (pp. 165-200). Washington, D.C.: ASA.

834
 INDEX
%BETWEEN%, 704–5 AUXILIARY, 551–55, 793–96
%class label%, 704 B2WEIGHT, 560
%COLUMN, 721 B3WEIGHT, 560–61
%OVERALL%, 704 balanced repeated replication, 613–14
%ROW, 721 BASEHAZARD
%TOTAL, 721 ANALYSIS, 614
%WITHIN%, 704–5 OUTPUT, 732
(#), 664–65 SAVEDATA, 753
*, 662–63 baseline hazard function, 149–50, 247–49
@, 663–64 BASIC, 594
[ ], 659–60 Bayes, 604–5, 631–36
_missing, 577 BAYES, 604–5
{ }, 675–76 Bayes factor, 741
| symbol Bayesian estimation, 604, 631–36
growth models, 676–83 Bayesian plots, 763–64
individually-varying times of observation, Bayesian posterior parameter values, 755,
686–87 804
latent variable interactions, 687–88 Bayesian structural equation modeling
random slopes, 683–85 (BSEM)
accelerated cohort, 143–47 bi-factor CFA with two items loading on only
ACE, 83–84, 85–86 the general factor and cross-loadings with
ADAPTIVE, 617 zero-mean and small-variance priors, 105–
ADDFREQUENCY, 626–27 6
ALGORITHM MIMIC model with cross-loadings and direct
Bayesian, 633–34 effects with zero-mean and small-variance
frequentist, 615 priors, 107–8
ALL multiple group model with approximate
CROSSTABS, 721 measurement invariance using zero-mean
MISSING, 540–41 and small-variance priors, 108–10
MODINDICES, 725–27 BCONVERGENCE, 634
REPSAVE, 803 beta, 742
USEVARIABLES, 538–39 BETWEEN
ALLFREE, 108–10, 611–13 Monte Carlo, 799–801
alpha, 742 real data, 570–72
alpha (c), 743 between-level categorical latent variable,
alpha (f), 744 348–51, 352–54, 358–60, 364–66, 367–
ANALYSIS command, 587–639 70, 375–77, 383–85
arithmetic operators, 576 BI-CF-QUARTIMAX, 606–10
AT, 686–87 bi-factor EFA, 52–53
auto-correlated residuals, 142–43 bi-factor rotations, 606–10
automatic labeling, 611–13 BI-GEOMIN, 606–10

835
BINARY chi-square difference test for WLSMV and
DATA MISSING, 527 MLMV, 451–52, 625–26, 752
DATA SURVIVAL, 531 CHOLESKY, 614–15
DATA TWOPART, 525 CINTERVAL, 727–29
birth cohort, 143–47 CINTERVAL (BCBOOTSTRAP), 727
BITERATIONS, 634–35 CINTERVAL (BOOTSTRAP), 727
bivariate frequency tables, 721 CINTERVAL (EQTAIL), 727
BLANK, 540 CINTERVAL (HPD), 727
BOOTSTRAP CINTERVAL (SYMMETRIC), 727
ANALYSIS, 620–21 class probabilities, 757–58
REPSE, 613–14
CLASSES
bootstrap standard errors, 37–38, 620–21 Monte Carlo, 793
Box data, 607–8 real data, 564–65
BPARAMETERS CLUSTER, 557–58
Monte Carlo, 804 cluster size, 784–86
real data, 755
CLUSTER_MEAN, 579–80
BRR, 613–14
cluster-level factor indicators, 280–84
BSEED, 632
cohort, 143–47
BSEM, 105–6, 107–8, 108–10
COHORT, 531–32
burnin, 604–5
COHRECODE, 532–33
BWEIGHT, 560
COMBINATION, 617
BWTSCALE, 561
COMPLEX, 594
BY, 647–52
complex survey data, 556–63
confirmatory factor analysis (CFA), 648–49
exploratory structural equation modeling complier-average causal effect estimation
(ESEM), 649–52 (CACE), 191–93, 193–95
CATEGORICAL conditional independence (relaxed), 180–81,
Monte Carlo, 789 198–99
real data, 543–45 conditional probabilities, 741
categorical latent variables, 4–5, 653–54 confidence intervals, 37–38, 727–29
categorical mediating variables, 491 confirmatory analysis, 60–61, 176–78
CENSORED confirmatory factor analysis (CFA)
Monte Carlo, 788–89 categorical factor indicators, 62
real data, 542–43 censored and count factor indicators, 64–65
CENTER, 580–84 continuous and categorical factor indicators,
CF-EQUAMAX, 606–10 63
CF-FACPARSIM, 606–10 continuous factor indicators, 60–61
mixture, 182
CF-PARSIMAX, 606–10
two-level mixture, 355–57
CF-QUARTIMAX, 606–10
CONSTRAINT, 555
CF-VARIMAX, 606–10
constraints, 31–32, 87–88, 89–90, 90–91,
CHAINS, 632
91–92, 175–76, 200–202, 203–5, 691–96
CHECK, 514–15
contextual effect, 263–67
CONTINUOUS, 525

836
continuous latent variables, 3–4 CROSSCLASSIFIED, 598
continuous-time survival analysis, 148–49, cross-classified modeling
149–50, 151–52, 205–7, 247–49, 308–9, IRT with random binary items, 333–34
437–38 multiple indicator growth model with random
convenience features, 660–62 intercepts and factor loadings, 334–37
CONVERGENCE, 628 path analysis with continuous dependent
variables, 330–32
convergence problems, 467–69
regression for a continuous dependent
COOKS variable, 326–30
PLOT, 768
cross-loadings, 107–8
SAVEDATA, 759
CROSSTABS, 721
Cook's distance, 759
CSIZES, 784–86
COPATTERN, 532
CUT, 579
CORRELATION
ANALYSIS, 631 CUTPOINT
DATA, 510 DATA SURVIVAL, 530–31
ROWSTANDARDIZATION, 610–11 DATA TWOPART, 525
SAVEDATA, 755 CUTPOINTS, 781
COUNT D, 732
CROSSTABS, 721 DATA COHORT, 531–34
Monte Carlo, 790–92 DATA command, 504–34
real data, 546–49 data generation, 779–81
count variable models data imputation, 515–19
negative binomial, 547–48, 791–92 DATA IMPUTATION, 515–19
negative binomial hurdle, 548–49, 792 DATA LONGTOWIDE, 521–23
Poisson, 546–47, 791 DATA MISSING, 526–29
zero-inflated negative binomial, 548, 792 data reading
zero-inflated Poisson, 547, 791 covariance matrix, 444
zero-truncated negative binomial, 548, 792 fixed format, 445
COVARIANCE means and covariance matrix, 444–45
ANALYSIS, 631 DATA SURVIVAL, 529–31
DATA, 510
data transformation, 519–34, 574–85
DATA IMPUTATION, 518
MODEL PRIORS, 700 DATA TWOPART, 523–26
ROWSTANDARDIZATION, 610–11 DATA WIDETOLONG, 519–21
SAVEDATA, 749, 755 DDROPOUT, 526–29
COVERAGE DE3STEP
ANALYSIS, 626 Monte Carlo, 793–96
MONTECARLO, 802 real data, 551–55
Cox regression, 148–49, 247–49, 308–9, decomposition, 258–63
437–38 defaults, 460–64
CPROBABILITIES, 757–58 DEFINE command, 574–85
CRAWFER, 606–10 delta, 743
Crawfer family of rotations, 609 DELTA, 605–6
credibility interval, 727–29 derivatives of parameters, 734–35

837
DESCRIPTIVE, 528–29 estimated correlation matrix, 736
Deviance Information Criterion, 604–5 estimated covariance matrix
Diagrammer, 10–11 OUTPUT, 736
DIC, 604–5 SAVEDATA, 749
DIFFERENCE, 108–10, 700–702 estimated sigma between matrix, 749–50
difference testing, 486–87 estimated time scores, 126
DIFFTEST ESTIMATES, 751–52
ANALYSIS, 625–26 ESTIMATOR, 600–604
SAVEDATA, 752 event history indicators, 433–34
direct effects, 107–8 EXPECTED, 617–19
Dirichlet, 698–702 expected frequencies, 737–38
discrete-time survival analysis, 147–48, exploratory factor analysis (EFA)
433–34 bi-factor with continuous factor indicators,
discrete-time survival mixture analysis, 52–53
245–47 categorical factor indicators, 47
distal outcome, 223–24 continuous factor indicators, 45–46
continuous, censored, categorical, and count
DISTRIBUTION, 635–36
factor indicators, 48–49
DO factor mixture analysis, 49–50
DEFINE, 584–85 parallel analysis, 611
MODEL CONSTRAINT, 694–95 two-level with continuous factor indicators,
MODEL PRIORS, 700 50–51
do loop two-level with individual- and cluster-level
DEFINE, 584–85 factor indicators, 51–52
MODEL CONSTRAINT, 694–95 exploratory structural equation modeling
MODEL PRIORS, 700
(ESEM)
dropout, 392–93 bi-factor EFA, 103–4
DSURVIVAL, 549 bi-factor EFA with two items loading on only
DU3STEP the general factor, 104–5
Monte Carlo, 793–96 EFA at two timepoints, 97–98
real data, 551–55 EFA with covariates (MIMIC), 93–94
E EFA with residual variances constrained to be
Monte Carlo, 793–96 greater than zero, 102–3
real data, 551–55 multiple group EFA with continuous factor
ECLUSTER, 559–60 indicators, 99–101
EFA, 598–600 SEM with EFA and CFA factors, 95–96
eigenvalues, 611 external Monte Carlo simulation, 428–31
EM, 615 factor mixture analysis, 49–50, 198–99
EMA, 615 factor score coefficients, 731
EQUAL, 614 factor score determinacy, 731–32
equalities, 664–65 factor scores, 756–57, 759–60, 767
ESEM, 93–94, 95–96, 97–98, 99–101, 102– FACTORS
3, 103–4, 104–5 PLOT, 767
ESTBASELINE, 754 SAVEDATA, 759–60
FAY, 613–14

838
FBITERATIONS, 635 graded response model, 65–66
FILE GRANDMEAN, 580–84
DATA, 507 graphics module, 768–73
SAVEDATA, 746–48 GROUPING, 549–50
FINITE, 563–64 GROUPMEAN, 580–84
finite population correction factor, 563–64 growth mixture modeling
fixing parameter values, 663–64 between-level categorical latent variable, 375–
FORMAT 77
DATA, 507–9 categorical outcome, 219–20
SAVEDATA, 747–48 censored outcome, 218–19
FPC, 563–64, 563–64 continuous outcome, 213–16
frailty, 1 distal outcome, 223–24
FREE known classes (multiple group analysis), 228–
DATA, 507–8 30
MFORMAT, 761–62 negative binomial model, 223
SAVEDATA, 747–48 sequential process, 225–28
freeing parameters, 662–63 two-level, 371–74
zero-inflated Poisson model, 220–23
frequency tables, 721
growth modeling, 676–83
frequency weights, 550–51
auto-correlated residuals, 142–43
FREQWEIGHT, 550–51 categorical outcome, 121–22
FS, 615 censored outcome, 117–18
FSCOEFFICIENT, 731 censored-inflated outcome, 119–21
FSCORES, 756–57 continuous outcome, 115–17
FSDETERMINACY, 731–32 count outcome, 123
FULLCORR, 510 estimated time scores, 126
FULLCOV, 510 individually-varying times of observation,
functions, 576 130–32
gamma, 742–43 multiple group multiple cohort, 143–47
multiple indicators, 135–37, 137–38
Gamma, 698–702
parallel processes, 133–34
gamma (c), 744 piecewise model, 129–30
gamma (f), 744 quadratic model, 126–27
GAUSSHERMITE, 616 time-invariant covariate, 128–29
Gelman-Rubin convergence, 634 time-varying covariate, 128–29
GENCLASSES, 781–82 two-part (semicontinuous), 138–41
GENERAL, 593–94 using the Theta parameterization, 122
GENERATE, 779–81 with covariates, 128–29
generating data, 779–81 zero-inflated count outcome, 124–25
generating missing data, 786–88 H1CONVERGENCE, 628
GEOMIN, 606–10 H1ITERATIONS, 627
GIBBS, 633–34 H1SE, 730
Gibbs sampler algorithm, 633–34 H1TECH3, 730
GLS, 603–4 HAZARDC, 786
GPA algorithm, 629 heritability, 89–90, 90–91, 200–202, 203–5

839
hidden Markov model, 233–35 ITERATIONS, 627
highest posterior density, 727–29 JACKKNIFE, 613–14
identification, 469–70 JACKKNIFE1, 613–14
identity by descent (IBD), 91–92 JACKKNIFE2, 613–14
IDVARIABLE K-1STARTS, 624
DATA LONGTOWIDE, 522–23 KAISER, 610–11
DATA WIDETOLONG, 520–21 KAPLANMEIER, 753
VARIABLE, 550 kappa (u), 744
imputation, 515–19 known class, 188–89, 228–30
IMPUTATION, 512–13 KNOWNCLASS, 108–10, 565–66
IMPUTE, 516–17 KOLMOGOROV, 636
IND, 690 Kolmogorov-Smirnov test, 636
indirect effect plot, 40–41 labeling
indirect effects, 37–38, 75, 689–91 baseline hazard parameters, 673
INDIVIDUAL, 510 categorical latent variables, 672–73
individually-varying times of observation, classes, 673
130–32, 686–87 inflation variables, 673
INFLUENCE nominal variables, 672–73
PLOT, 768 parameters, 674–75, 692
SAVEDATA, 758–59 thresholds, 672
influential observations, 768 Lagrange multiplier tests. See modification
INFORMATION, 617–19 indices
information curves, 763–64 lambda, 742
INTEGRATION, 616 lambda (f), 744
INTEGRATION setting for ALGORITHM, lambda (u), 743
615 LATENT, 633
interaction between latent variables, 76–77, latent class analysis (LCA)
687–88 binary latent class indicators, 163–65
interactions, 687–88 binary, censored, unordered, and count latent
class indicators, 172–73
INTERACTIVE, 636–37
confirmatory, 175–76, 176–78
intercepts, 659–60 continuous latent class indicators, 170–71
INTERRUPT, 637 three-category latent class indicators, 167–68
inverse Gamma, 698–702 two-level, 361–63
inverse Wishart, 698–702 two-level with a between-level categorical
IRT translation, 732 latent variable, 364–66
item characteristic curves, 763–64 unordered categorical latent class indicators,
item response theory (IRT) models 168–69
factor mixture, 198–99 with a covariate and a direct effect, 174–75
random binary items using cross-classified with a second-order factor (twin analysis),
data, 333–34 183–85
twin, 203–5 with partial conditional independence, 180–81
two-level mixture, 358–60 latent class growth analysis (LCGA)
two-parameter logistic, 65–66 binary outcome, 230–31

840
 three-category outcome, 231–32 LOGRANK, 732
two-level, 378–80 logrank test, 732
zero-inflated count outcome, 232–33 Lo-Mendell-Rubin test, 738
latent response variables, 757, 760, 767–68 LONG
latent transition analysis (LTA) DATA LONGTOWIDE, 522
for two time points with a binary covariate DATA WIDETOLONG, 520
influencing the latent transition LOOP, 40–41, 695–96
probabilities, 235–38 loop plots, 40–41
for two time points with a continuous
LRESPONSES
covariate influencing the latent transition
PLOT, 767–68
probabilities, 238–41
SAVE, 757
mover-stayer for three time points using a
SAVEDATA, 760
probability parameterization, 241–45
two-level, 380–82 LRTBOOTSTRAP, 621
two-level with a between-level categorical LRTSTARTS, 624–25
latent variable, 383–85 LTA calculator, 11, 238–41
latent transition probabilities, 741 M, 551–55
latent variable covariate, 258–63, 263–67 MAHALANOBIS
latent variable interactions, 76–77, 687–88 PLOT, 768
liabilities, 1, 85–86, 90–91, 200–202 SAVEDATA, 758
likelihood ratio bootstrap draws, 739–41 Mantel-Cox test, 732
likelihood ratio test, 738, 739–41 marginal probabilities, 741
linear constraints, 175–76, 691–96 Markov chain Monte Carlo, 631–36
LINK, 606 MATRIX, 631
list function, 666–72 MCCONVERGENCE, 629
LISTWISE, 513–14 MCITERATIONS, 627
listwise deletion, 513–14 MCMC, 604–5
local maxima, 465–67 MCMC chain, 604–5
local solution, 465–67 MCONVERGENCE, 629
log odds, 493–97 MCSEED, 617
LOGCRITERION, 628 MEAN
DEFINE, 578
LOGHIGH, 630
POINT, 631
logical operators, 575–76
mean square error (MSE), 416
logistic regression, 493–97
mean structure, 70–72, 79–80
LOGIT
means, 659–60
LINK, 606
PARAMETERIZATION, 606 MEANS, 510
LOGLIKELIHOOD measurement invariance, 484–87
PLOT, 768 approximate, 108–10
SAVEDATA, 758 MEDIAN, 631
LOGLINEAR, 606 mediation
loglinear analysis, 179–80, 499–500 bootstrap, 37–38
categorical variable, 491
LOGLOW, 630
cluster-level latent variable, 270–71
lognormal, 698–702

841
 continuous variable, 32–33 ML
missing data, 39–40 ESTIMATOR, 603
moderated, 40–41 STVALUES, 632
random slopes, 439–41 MLF, 603
MEDIATOR, 633 MLM, 603
MEMBERSHIP, 566–67 MLMV, 603
merging data sets, 456–57, 760–62 MLR, 603
Metropolis-Hastings algorithm, 633–34 MMISSING, 762
MFILE, 760–61 MNAMES, 761
MFORMAT, 761–62 MODE, 631
MH, 633–34 MODEL
MIMIC ANALYSIS, 611–13
continuous factor indicators, 69–70 DATA IMPUTATION, 518
multiple group analysis, 78–79, 79–80, 80–81 MODEL command, 643–711
MISSFLAG, 748 MODEL command variations, 702–5
MISSING MODEL CONSTRAINT, 691–96
DATA MISSING, 526–29 MODEL COVERAGE, 707–9
MONTECARLO, 787–88 model estimation, 459–73
VARIABLE, 540–41
MODEL INDIRECT, 689–91
missing data, 39–40, 389–91, 392–93, 393–
MODEL label, 702–4
95, 395–96, 397, 398–99, 399–402, 402–
MODEL MISSING, 709–11
4, 417–21, 421–22, 425–27, 487–91
MODEL POPULATION, 705–7
missing data correlate, 389–91, 551–55
MODEL PRIORS, 698–702
missing data generation, 786–88
MODEL TEST, 696–97
missing data patterns, 730–31
modeling framework, 1–6
missing data plots, 763–64
moderated mediation, 40–41, 695–96
missing value flags, 540–41
modification indices, 725–27
non-numeric, 446
numeric, 446–47 MODINDICES, 725–27
MITERATIONS, 627 MONITOR, 768
MIXC, 629–30 Monte Carlo simulation studies
discrete-time survival analysis, 433–34
MIXTURE, 594–96
EFA with continuous outcomes, 427–28
mixture modeling external Monte Carlo, 428–31
confirmatory factor analysis (CFA), 182 GMM for a continuous outcome, 423–25
multivariate normal, 189–91 growth with attrition under MAR, 421–22
randomized trials (CACE), 191–93, 193–95 mediation with random slopes, 439–41
regression analysis, 158–61 MIMIC with patterns of missing data, 417–21
structural equation modeling (SEM), 186–87 missing data, 417–21, 421–22
with known class, 188–89 multiple group EFA with measurement
zero-inflated Poisson regression analysis, invariance, 441–42
162–63 saved parameter estimates, 431–32
zero-inflated Poisson regression as a two-class two-level Cox regression, 437–38
model, 195–96 two-level growth model for a continuous
MIXU, 630 outcomes (three-level analysis), 425–27

842
 two-part (semicontinuous) model, 435–37 three-level MIMIC model with continuous
MONTECARLO factor indicators, two covariates on within,
DATA, 511–12 one covariate on between level 2, and one
INTEGRATION, 616 covariate on between level 3 with random
MONTECARLO command, 775–804 slopes on both within and between level 2,
mover-stayer model, 241–45 318–23
Mplus language, 13–14 three-level path analysis with a continuous
and a categorical dependent variable, 315–
Mplus program
18
base, 17
three-level regression for a continuous
combination add-on, 18
dependent variable, 312–15
mixture add-on, 17
two-level confirmatory factor analysis (CFA)
multilevel add-on, 18
with categorical factor indicators, 277–78
MSE, 416 two-level confirmatory factor analysis (CFA)
MSELECT, 762 with continuous factor indicators, 274–76,
MUCONVERGENCE, 629 278–80
MUITERATIONS, 627 two-level growth for a zero-inflated count
multilevel mixture modeling outcome (three-level analysis), 306–8
two-level confirmatory factor analysis (CFA), two-level growth model for a categorical
355–57 outcome (three-level analysis), 294–95
two-level growth mixture model (GMM), two-level growth model for a continuous
371–74 outcome (three-level analysis), 291–94
two-level growth mixture model (GMM) with two-level MIMIC model with continuous
a between-level categorical latent variable, factor indicators, random factor loadings,
375–77 two covariates on within, and one covariate
two-level growth model with a between-level on between with equal loadings across
categorical latent variable, 367–70 levels, 310–12
two-level item response theory (IRT), 358–60 two-level multiple group confirmatory factor
two-level latent class analysis (LCA), 361–63 analsyis (CFA), 288–90
two-level latent class analysis (LCA) with a two-level multiple indicator growth model,
between-level categorical latent variable, 299–303
364–66 two-level path analysis with a continuous and
two-level latent class growth analysis a categorical dependent variable, 267–70
(LCGA), 378–80 two-level path analysis with a continuous, a
two-level latent transition analysis (LTA), categorical, and a cluster-level observed
380–82 dependent variable, 270–71
two-level latent transition analysis (LTA) with two-level path analysis with random slopes,
a between-level categorical latent variable, 272–74
383–85 two-level regression for a continuous
two-level mixture regression, 342–47, 348– dependent variable with a random
51, 352–54 intercept, 258–63
multilevel modeling two-level regression for a continuous
three-level growth model with a continuous dependent variable with a random slope,
outcome and one covariate on each of the 263–67
three levels, 323–26 two-level structural equation modeling
(SEM), 285–88

843
multinomial logistic regression, 493–97 non-convergence, 467–69
multiple categorical latent variables, 176–78 non-linear constraints, 31–32, 691–96
multiple cohort, 143–47 non-linear factor analysis, 68
multiple group analysis non-parametric, 197
known class, 188–89, 228–30 NOSERROR, 729–30
MIMIC with categorical factor indicators, 80– not missing at random (NMAR)
81 Diggle-Kenward selection model, 393–95
MIMIC with continuous factor indicators, 79– pattern-mixture model, 395–96
80 NREPS, 778
special issues, 473–84 nu, 742
multiple imputation, 453, 512–13, 515–19 numerical integration, 470–73
missing values, 397, 398–99, 402–4
OBLIMIN, 606–10
plausible values, 399–402
OBLIQUE, 606–10
multiple indicators, 135–37, 137–38
OBSERVED
multiple processors, 637–39
INFORMATION, 617–19
multiple solutions, 465–67 MEDIATOR, 633
MULTIPLIER odds, 493–97
ANALYSIS, 626
ODLL, 615
SAVEDATA, 755
OFF
multivariate normal mixture model, 189–91
ADAPTIVE, 617
MUML, 603 BASEHAZARD, 614
NAMES CHOLESKY, 614–15
DATA MISSING, 527 LISTWISE, 513–14
DATA SURVIVAL, 530 MONITOR, 768
DATA TWOPART, 524 ON, 652–55
MONTECARLO, 776–77 ADAPTIVE, 617
VARIABLE, 537 BASEHAZARD, 614
NCSIZES, 782–84 CHOLESKY, 614–15
NDATASETS, 517 LISTWISE, 513–14
negative binomial, 28–29, 546–49 MONITOR, 768
NEW, 692 optimization history, 736–37, 737
NGROUPS OPTSEED, 624
DATA, 513 ORTHOGONAL, 606–10
MONTECARLO, 777–78 outliers, 768
NOBSERVATIONS OUTLIERS, 768
DATA, 513 OUTPUT command, 713–44
MONTECARLO, 777
PARALLEL, 611
NOCHECK, 514–15
parallel analysis, 611
NOCHISQUARE, 729
parallel computing, 637–39
NOCOVARIANCES, 611–13
parallel processes, 133–34
NOMEANSTRUCTURE, 611–13
parameter constraints. See constraints
NOMINAL
parameter derivatives, 736
Monte Carlo, 790
real data, 545–46 parameter extension, 633–34

844
parameterization posterior, 604–5
delta, 80–81 posterior predictive checks, 604–5
logistic, 498–99 potential scale reduction, 604–5, 634
loglinear, 176–78, 179–80, 499–500 PRIOR, 636
probability, 500–501 priors, 698–702
theta, 34, 82, 122
PRIORS, 566–67
PARAMETERIZATION, 605–6
PROBABILITIES, 566–67
parametric bootstrap, 739–41
PROBABILITY, 606
parametric proportional hazards, 149–50,
probability calculations
151–52 logistic regression, 493–97
path analysis multinomial logistic regression, 493–97
categorical dependent variables, 33–34 probit regression, 492–93
combination of censored, categorical, and PROBIT, 606
unordered categorical (nominal) dependent
probit link, 200–202, 606
variables, 36–37
combination of continuous and categorical PROCESSORS, 637–39
dependent variables, 35 profile likelihood, 149–50, 247–49, 308–9
continuous dependent variables, 32–33 PROMAX, 606–10
PATMISS, 786–87 proportional hazards model, 149–50, 151–52
PATPROBS, 787 pseudo-class
PATTERN, 555–56 Monte Carlo, 793–96
PATTERNS, 730–31 real data, 551–55
PERTURBED, 632 psi, 743
piecewise growth model, 129–30 PSR, 634
plausible values, 399–402, 756–57, 759–60, PWITH, 657
767 PX1, 633–34
PLOT, 695 PX2, 633–34
PLOT command, 762–73 PX3, 633–34
PLOT1, 763 quadratic growth model, 126–27
PLOT2, 763–64 quantitative trait locus (QTL), 91–92
PLOT3, 765 QUARTIMIN, 606–10
Plots R
Bayesian, 636, 763–64 Monte Carlo, 793–96
missing data, 763–64 real data, 551–55
survival, 763–64 R3STEP
POINT, 631 Monte Carlo, 793–96
real data, 551–55
Poisson. See zero-inflated Poisson
RANDOM, 594
PON, 655–56
random factor loadings, 310–12, 334–37,
pooled-within covariance matrix. See
685–86
sample covariance matrices
random items, 333–34
POPULATION
FINITE, 563–64 random slopes, 29–31, 130–32, 263–67,
MONTECARLO, 801–2 272–74, 278–80, 285–88, 296–98, 303–5,
population size, 563–64 683–85

845
random starts, 158–61, 167 RW, 633–34
RCONVERGENCE, 629 SAMPLE, 748–49
reading data sample covariance matrices
fixed format, 445 pooled-within, 750–51
RECORDLENGTH, 748 sample, 748–49
REGRESSION, 518 sigma between, 749–50
regression analysis sample statistics, 720–21
censored inflated regression, 24 sampling fraction, 563–64
censored regression, 23–24 sampling weights, 558–59
linear regression, 22–23 SAMPSTAT, 720–21
logistic regression, 25–26 SAVE
multinomial logistic regression, 26–27 DATA IMPUTATION, 517–18
negative binomial regression, 28–29 MONTECARLO, 803
Poisson regression, 27 SAVEDATA, 755–59
probit regression, 25 SAVEDATA command, 744–62
random coefficient regression, 29–31 saving data and results, 744–62
zero-inflated Poisson regression, 28
scale factors, 675–76
REPETITION
SDITERATIONS, 627
DATA LONGTOWIDE, 523
DATA WIDETOLONG, 521 SDROPOUT, 526–29
replicate weights, 457, 458, 561–62 second-order factor analysis, 66–67
REPSAVE, 802–3 SEED, 778
REPSE, 613–14 selection modeling, 393–95
REPWEIGHTS semicontinuous, 138–41, 435–37
SAVEDATA, 758 SEQUENTIAL, 518
VARIABLE, 561–62 sequential cohort, 143–47
RESIDUAL sequential regression, 518
BOOTSTRAP, 621 SERIES, 765–67
OUTPUT, 724–25 SFRACTION, 563–64
residual variances, 658 sibling modeling, 91–92
residuals, 724–25 SIGB, 749–50
RESPONSE, 754 sigma between covariance matrix. See
RESULTS sample covariance matrices
MONTECARLO, 803 STANDARD
SAVEDATA, 751 BOOTSTRAP, 620–21
right censoring, 149–50, 151–52, 308–9 INTEGRATION, 616
RITERATIONS, 628 STANDARDIZE, 584
RLOGCRITERION, 628–29, 628–29 STANDARDIZED, 721–24
robust chi-square, 603 standardized parameter estimates, 721–24
robust standard errors, 603 STARTING, 802
ROTATION, 606–10 starting values
ROUNDING, 518–19 assigning, 662–63
ROWSTANDARDIZATION, 610–11 automatic, 158–61
RSTARTS, 625 saving, 729

846
 user-specified, 165–66, 217, 662–63 TECH3
STARTS, 622–23 OUTPUT, 736
STCONVERGENCE, 623 SAVEDATA, 752
STD, 722 TECH4
STDEVIATIONS, 510 OUTPUT, 736
STDY, 722 SAVEDATA, 752–53
STDYX, 722 TECH5, 736–37
STITERATIONS, 623 TECH6, 737
STRATIFICATION, 556–57 TECH7, 737
structural equation modeling (SEM) TECH8, 737
categorical latent variable regressed on a TECH9, 737
continuous latent variable, 185–86 theta, 742
continuous factor indicators, 73–74 THETA, 605–6
with interaction between latent variables, 76– theta parameterization, 34, 82, 122, 605–6
77 THIN
STSCALE, 623 ANALYSIS, 635
STSEED, 623 DATA IMPUTATION, 519
STVALUES, 632 thinning, 519
SUBPOPULATION, 562–63 threads, 637–39
SUM, 578–79 THREELEVEL, 597–98
summary data, 510–12 three-level analysis, 291–94, 294–95, 367–
SURVIVAL 70
Monte Carlo, 796 three-step mixture analysis
real data, 572–73 Monte Carlo, 793–96
survival analysis. See continuous-time real data, 163–65, 551–55
survival analysis and discrete-time threshold structure, 72–73
survival analysis thresholds, 659–60
survival plots, 763–64 Thurstone's Box data, 607–8
SVALUES, 729 TIMECENSORED, 573–74
SWMATRIX time-invariant covariates, 128–29
DATA, 514 TIMEMEASURES, 533
SAVEDATA, 750–51 time-to-event variable, 148–49, 247–49,
tau, 742 308–9, 437–38
tau (u), 743 time-varying covariates, 128–29
TECH1, 732–34 TITLE command, 503
TECH10, 737–38 TNAMES, 533–34
TECH11, 738 total effect, 689–91
TECH12, 738 TRAINING, 566–67
TECH13, 739 training data, 191–93
TECH14, 739–41 TRANSFORM, 525–26
TECH15, 741 transformation
TECH16, 741 data, 519–34
TECH2, 734–35 variables, 574–85

847
TSCORES dependent, 642
Monte Carlo, 797 independent, 642
real data, 551 latent, 641
twin analysis, 83–84, 85–86, 89–90, 90–91, observed, 641
183–85, 200–202, 203–5 scale of measurement, 642
TWOLEVEL, 596–97 VARIANCE, 630–31
two-part (semicontinuous), 138–41, 435–37 variances, 658
TYPE VARIANCES, 514–15
ANALYSIS, 592–600 VARIMAX, 606–10
DATA, 510–13 VIA, 690
DATA MISSING, 527–28 Wald test, 696–97
PLOT, 763–65 WEIGHT, 558–59
SAVEDATA, 755 WIDE
UB, 599–600 DATA LONGTOWIDE, 522
UB*, 599–600 DATA WIDETOLONG, 520
UCELLSIZE, 630 WITH, 656–57
ULS, 603 WITHIN
ULSMV, 603 Monte Carlo, 797–99
UNEQUAL, 614 real data, 568–70
UNPERTURBED, 632 WLS, 603
UNSCALED WLSM, 603
BWTSCALE, 561 WLSMV, 603
WTSCALE, 559–60 WTSCALE, 559–60
USEOBSERVATIONS, 538 XWITH, 687–88
USEVARIABLES, 538–39 zero cells, 626–27
UW, 599–600 zero-inflated Poisson, 28–29, 124–25, 195–
UW*, 599–600 96, 220–23, 232–33, 546–49
VALUES, 518 zero-mean and small-variance priors, 105–6,
VARIABLE command, 534–74 107–8, 108–10
variables

848
849
 MUTHÉN & MUTHÉN
Mplus SINGLE-USER LICENSE AGREEMENT

Carefully read the following terms and conditions before opening the sealed CD sleeve or downloading the
software. Opening the CD sleeve or downloading the software indicates your acceptance of the terms and
conditions listed below. The Mplus CD and download contains several versions of Mplus. The Mplus
Single-User License allows for the use of only one of these programs. Using more than one is a violation
of the Mplus Single-User License Agreement.

Muthén & Muthén grants you the non-exclusive right to use the copyrighted computer program Mplus and the
accompanying written materials. You assume responsibility for the selection of Mplus to achieve your intended
results, and for the installation, use, and results obtained from Mplus.
1. Copy and Use Restrictions. Mplus and the accompanying written materials are copyrighted. Unauthorized
copying of Mplus and the accompanying written materials is expressly forbidden. One copy of Mplus may be
made for backup purposes, and it may be copied as part of a normal system backup. Mplus may be transferred
from one computer to another but may only be used on one computer at a time.
2. Transfer Restrictions. The Mplus license may be transferred from one individual to another as long as all
copies of the program and documentation are transferred, registered, and the recipient agrees to the terms and
conditions of this agreement.
3. Termination. The license is effective until terminated. You may terminate it at any time by destroying the
written materials and all copies of Mplus, including modified copies, if any. The license will terminate
automatically without notice from Muthén & Muthén if you fail to comply with any provision of this agreement.
Upon termination, you shall destroy the written materials and all copies of Mplus, including modified copies, if
any, and shall notify Muthén & Muthén of same.
4. Limited Warranty. Muthén & Muthén warrants that for ninety (90) days after purchase, Mplus shall
reasonably perform in accordance with the accompanying documentation. Muthén & Muthén specifically does
not warrant that Mplus will operate uninterrupted and error free. If Mplus does not perform in accordance with
the accompanying documentation, you may notify Muthén & Muthén in writing of the non-performance within
ninety (90) days of purchase.
5. Customer Remedies. Muthén & Muthén and its supplier’s entire liability and your exclusive remedy shall be,
at Muthén & Muthén’s option, either return of the price paid, or repair or replacement of the defective copy of
Mplus and/or written materials after they have been returned to Muthén & Muthén with a copy of your receipt.
6. Disclaimer of Other Warranties. Muthén & Muthén and its suppliers disclaim all other warranties, either
express or implied, including, but not limited to, any implied warranties of fitness for a particular purpose or
merchantability. Muthén & Muthén disclaims all other warranties including, but not limited to, those made by
distributors and retailers of Mplus. This license agreement gives you specific legal rights. You may have other
rights that vary from state to state.
7. Disclaimer. In no event shall Muthén & Muthén or its suppliers be liable for any damages, including any lost
profits, lost savings or other incidental or consequential damages arising out of the use or inability to use Mplus
even if Muthén & Muthén or its suppliers have been advised of the possibility of such damages. Some states do
not allow the limitation or exclusion of liability for incidental or consequential damages so the above limitation
or exclusion may not apply to you.
8. Return Policy: All sales are final. Software purchased on-line through our website is considered opened at
the time of purchase. This also applies to hard copy purchases because downloads are made available at the time
of purchase. In rare instances, and only within 30 days of purchase, if due to technical difficulties or platform
incompatibilities, the software will not function, we may, at our discretion, issue a refund. In such instances, an
LOD (Letter Of Destruction) on company letterhead will be required to process the refund.

This agreement is governed by the laws of the State of California.

850