Structural Equation Modeling

R Demonstration Notes
Daniel J. Bauer & Patrick J. Curran

This material (Version 2019.3) was developed in support of the workshop

Structural Equation Modeling presented on May 13 - 17, 2019, by Dan
Bauer and Patrick Curran in Chapel Hill NC.

This document is copyright Curran-Bauer Analytics and may be used for

any non-commercial purposes in contribution to the public good. There is
no warranty expressed or implied and these materials should be interpreted
and applied at the discretion of the user.

R, RStudio, lavaan, and all R-based packages undergo continual updating

and expansion. These materials were developed using R version 3.5.2
(Windows) and lavaan version 0.6-3. The scripts may require modification
when used with earlier or later versions to function correctly.

Although Curran-Bauer Analytics encourages the use of these materials in

whatever way contributes to a research and teaching endeavor, we are
unable to respond to specific questions that might arise in practice.
However, please share any errors or omissions that you might find at
[email protected] .

www.curranbauer.org
 
 

Citation Information
 

To cite R in publications, use: 

R Core Team (2018). R: A language and environment for statistical computing. R Foundation for 
Statistical Computing, Vienna, Austria. URL https://www.R‐project.org/. 

 
To cite the lavaan package in publications, use: 

Rosseel, Y. (2012). lavaan: An R Package for Structural Equation Modeling. Journal of Statistical 
Software, 48, 1‐36. URL http://www.jstatsoft.org/v48/i02/. 
 

To cite this PDF, use: 

  Bauer, D.J. & Curran, P.J. (2019). Structural equation modeling: R demonstration notes (Version 
2019.3). Curran‐Bauer Analytics, Durham: NC. URL https://curranbauer.org/sem‐r‐notes‐2019‐3/ 

Copyright © 2019 Curran‐Bauer Analytics 

All rights reserved 

No part of these notes may be reproduced, translated, stored in a retrieval system, or transmitted, in 
any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without 
written permission by the copyright holders. 
 
 

Table of Contents
 

 
Preface: Instructions for Downloading R, RStudio, and R Packages ......................................................... iv 
 
Chapter 1: Introduction, Background, and Multiple Regression 
Multiple Regression Analysis of Deviant Peer Affiliations ..................................................................... 1‐3 
Reading in the Data ............................................................................................................................ 1‐3 
Single Predictor Regression ............................................................................................................... 1‐4 
Multiple Regression Model with COA Status, Gender, and Age as Predictors .................................. 1‐8 
Multiple Regression Model with COA Status, Gender, Age, Stress, Emotion, and Negative Affect as 
Predictors ......................................................................................................................................... 1‐11 
 
Chapter 2: Path Analysis: Part I 
Path Analysis with Deviant Peer Affiliation Data ................................................................................... 2‐3 
Path Analysis of Theoretical Peer Affiliation Model .......................................................................... 2‐3 
 
Chapter 3: Path Analysis: Part II 
Path Analysis with Deviant Peer Affiliation Data ................................................................................... 3‐3 
Theoretical Model .............................................................................................................................. 3‐3 
Modification to Peer Affiliation Model: Likelihood Ratio Test .......................................................... 3‐5 
Modification to Peer Affiliation Model: Modification Indices ........................................................... 3‐8 
Tests of Direct, Indirect, and Specific Effects ................................................................................... 3‐16 
 
Chapter 4: Confirmatory Factor Analysis 
Confirmatory Factor analysis of Holzinger‐Swineford Data................................................................... 4‐3 
Preliminary Steps ............................................................................................................................... 4‐3 
Initial Model with Standardized Factors ............................................................................................ 4‐4 
Initial Model with Scaling Items ......................................................................................................... 4‐9 
Model Modification ......................................................................................................................... 4‐13 
 
ii 
 

Chapter 5: Structural Equation Models with Latent Variables 
Structural Equation Modeling of Şenol‐Durak and Ayvaşik’s Posttraumatic Growth Data ................... 5‐3 
Preliminary Steps ............................................................................................................................... 5‐4 
Initial Hypothesized Model ................................................................................................................ 5‐5 
Confirmatory Factor Analysis ............................................................................................................. 5‐9 
Revised Model .................................................................................................................................. 5‐12 
Examining Direct, Indirect and Total Effects .................................................................................... 5‐18 
 
Chapter 6: Multiple Groups Models 
Multiple group analysis of Holzinger‐Swineford Data ........................................................................... 6‐3 
Preliminary Steps ............................................................................................................................... 6‐3 
Evaluating Total Invariance ................................................................................................................ 6‐4 
Evaluating the Level of Invariance ..................................................................................................... 6‐7 
Identifying and Allowing for Partial Strong Invariance .................................................................... 6‐10 
 
Chapter 7: SEM with Categorical Indicators 
Structural Equation Modeling of Schizophrenia Symptoms .................................................................. 7‐3 
Preliminary Steps ............................................................................................................................... 7‐4 
Confirmatory Factor Analysis using Diagonally Weighted Least Squares Estimation ....................... 7‐5 
Revised Model with DWLS Estimation ............................................................................................... 7‐8 
Plotting Item Characteristic Curves ................................................................................................. 7‐13 
 
Chapter 8: Latent Growth Curve Analysis 
Latent Growth Curve Analysis of Antisocial Behavior ............................................................................ 8‐3 
Preliminary Steps ............................................................................................................................... 8‐3 
Intercept‐Only Model ........................................................................................................................ 8‐4 
Intercept and Slope Model ................................................................................................................ 8‐8 
Quadratic Growth Trajectory Model ............................................................................................... 8‐13 
Freed Loading (“Fully Latent”) Growth Model: Additional Demonstration ..................................... 8‐17 
Linear Growth Model Conditional on Gender ................................................................................. 8‐24 
Linear Growth Model Conditional on Cognitive Stimulation in the Home ...................................... 8‐30 
 iii 
 

Appendix A: Generating Path Diagrams in R 
Example 1: Holzinger‐Swineford CFA (Chapter 4) ................................................................................ A‐3 
Example 2: Conditional Latent Curve Model (Chapter 8) ..................................................................... A‐6 
 

 

iv 
 

Preface: Instructions for Downloading R, RStudio, and R Packages

 

Downloading R
One of the nice things about R is that it is free to download and use. Just go to  
https://www.r-project.org/  
and click the “download R” link.  Then run the installer. 

Downloading and Using the RStudio Interface

Another free program, RStudio, provides a convenient interface for using R. The installer for RStudio can 
be found at  
https://www.rstudio.com/products/rstudio/download/
Select the RStudio Desktop / Open Source License for download and then run the installer. 

The default layout for RStudio is shown below 

The top left window shows an R script file.  Saving R script in a file enables you to share code and re‐run 
analyses quickly.  You can highlight individual lines or blocks of lines and click “Run” to submit them to R 
and obtain the results. 

The bottom left window shows the output and also displays a command prompt (the >).  When you 
want to execute a function quickly and don’t care to save it to an R script, you can simply type it at the 
command prompt. 

The top right window shows the objects currently in working memory.  Here we have two data objects, 
afdp (the data file) and fit (containing model results).  We also have a model syntax object called 
model.1 and a vector of variable labels called varnames. 
 v 
 

Finally, the bottom right window is useful for a variety of purposes, including viewing and installing 
packages, displaying plots, etc. 

You can rearrange or resize these windows however you like. 

Installing and Using Lavaan and other R Packages

Although R comes with some built in functionality, much of what you can do with R comes through 
packages contributed by the scientific community. For this workshop, we will be primarily using the 
lavaan (LAtent VAriable ANalaysis) package developed by Yves Rosseel from Ghent University.  Although 
lavaan is still considered to be in beta‐testing (i.e., experimental, meaning there is no guarantee 
everything will work as it should), it is widely used and considered to generate accurate results. 

There are a couple ways you can download this and other packages in R.  Using RStudio, the easiest way 
to install a package is to click the packages tab (in the lower right window under the default 
configuration) and then click the install button, as shown 

This will bring up the Install Packages dialogue box shown below.  Just type “lavaan” on the Packages 
line, as shown, then click “Install”.  

 
 
 
 
 
 
 
Alternatively, you can type the following at the command prompt: 
 

> install.packages("lavaan", dependencies = TRUE)

Regardless of how you install lavaan, to actually use the package, you will need to also run the line 
 

> library(lavaan)
vi 
 

The same steps are used for installing any other package in R (with the exception that some R packages 
are not provided through the CRAN repository accessed with these commands, in which case they must 
be downloaded directly from the developer). 

Additional packages used in various demonstrations presented in these notes include plyr, semTools, 
and psych. 

If you have questions about lavaan, you may wish to consult the discussion group. Go to 
https://groups.google.com/d/forum/lavaan/ and join the group. Then you can email 
questions to [email protected]. 

 
Chapter 1
Introduction, Background,
and Multiple Regression
Multiple Regression Analysis of Deviant Peer Affiliations ....................................................... 1-3
Reading in the Data .............................................................................................................. 1-3
Single Predictor Regression .................................................................................................. 1-4
Multiple Regression Model with COA Status, Gender, and Age as Predictors .................... 1-8
Multiple Regression Model with COA Status, Gender, Age, Stress, Emotion, and Negative
Affect as Predictors............................................................................................................. 1-11
1-2 | Chapter 1 Introduction, Background, and Multiple Regression

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-3

Multiple Regression Analysis of Deviant Peer Affiliations

The data for this demonstration were provided by Dr. Chassin from the Adolescent and Family
Development Project, housed at Arizona State University. Note that these data were
generously provided for strictly pedagogical purposes and should not be used for any other
purposes beyond this workshop. The sample includes 316 adolescents, between 10-16 years of
age. The study was designed to assess the association between parental alcoholism and
adolescent substance use and psychopathology. The data are in the text file afdp.dat and
the accompanying R script is in the ch01.R file. The variables in the data set that we will use
are
peer adolescent report on peer substance use and peer tolerance of use
coa parent report of alcoholism diagnosis where 0=non-alcoholic and 1=alcoholic
gen gender where 0=girl and 1=boy
age age measured in years at assessment
stress self report measure of uncontrollable negative life stressful events
emotion self report measure of temperamental emotional expressiveness
negaff self report measure of depression and anxiety

Reading in the Data

To run our analyses in R, we must first load the data. The first five rows of the afdp.dat data
file (for the first five participants), look like this:
1 1 14 0 2.35 1.37 4.4 0.49
2 1 12 0 0.55 1.46 2.34 0
3 1 14 1 2.37 2.12 2.11 1.73
4 1 15 1 1.14 2.83 2.6 1.86
5 1 12 1 1.37 2.11 2.04 0.36

The data is in free text format, meaning there are not set column positions for each variable.
Instead, variable values are separated only by a space. With this kind of data format, we can
use the read.table function to bring the data into R:
afdp <- read.table("afdp.dat", header=FALSE)
varnames <-c("id","coa","age","gen","stress","emotion","negaff","peer")
names(afdp)<-varnames
View(afdp)

Note that because we have not specified a full file path (e.g., “C:\Users\dbauer\
CBA\SEM\R\afdp.dat”), R will look for afdp.dat in the current working directory. You can see
which directory this is by typing getwd() at the prompt. You can also reset the working
directory by running setwd("<new path>") or, within RStudio, by clicking Session -> Set
Working Directory. We usually choose to set the working directory to the source file
location, meaning the folder within which we have saved our R script file.

Copyright © 2019 Curran-Bauer Analytics, LLC

1-4 | Chapter 1 Introduction, Background, and Multiple Regression

The header argument of the read.table function is set to FALSE here to indicate that there
are no variable names at the top of the data file. If the first row of the data file included
variable names then we would instead set this to header=TRUE. Because there are no names
for the variables, they will initially have the default names V1, V2, V3, etc.
The next two lines of the script assign labels that are more informative to our variables. First,
we define a vector, varnames, to hold the variable labels. Here we use the c function (for
“combine”) to put all the labels into one vector. Then we use the names function to replace the
default variable labels with these new, more useful labels.
Finally, the line View(afdp) requests that the data be presented for viewing in spreadsheet
form, as shown below.

We are now ready to run our analyses.

Single Predictor Regression1

The single-predictor regression model of COA status predicting deviant peer affiliations is
shown in the diagram below:


coa peer i 

The model is of the form peeri     coai   i where peeri is an individual’s report of their
level of association with deviant peers, coai is an adolescent’s parent report of parent
alcoholism status,  represents mean association with deviant peers for children of non-
alcoholics (i.e., where coai  0 ), and  is the expected increase in deviant peer associations for
children of alcoholics (i.e. a one unit increase in coai ). We assume that both coa and peer
contain no measurement error, that  i ~ N (0, ) for all individuals i, and that  i and coai are
uncorrelated.

1
The initial model building procedure presented in Chapter 1 of the lecture notes used OLS estimation in
traditional regression whereas here we use ML as a way of introducing the SEM procedure in R. The results are
virtually identical.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-5

We will fit this regression model using the lavaan package, which we load with the library
function as shown:
library(lavaan)

Under other circumstances, we would use a dedicated regression package to fit such a model,
but our intent here is to gain familiarity with how to specify structural equation models within
lavaan, starting with the very simple single-predictor regression model.
First, we need to create what’s called a model syntax object. The syntax for specifying models
in lavaan is quite straightforward and we will show many of the options as we proceed through
the demonstrations in this book. For this simple model, we only require a single line:
model.1 <- 'peer ~ coa'

Here, we have defined the model syntax object model.1, in which we have specified that peer
is predicted by coa using the regression operator ~. That is, the line peer ~ coa signifies that
peer is regressed on coa. In general, whatever variable is to the left of ~ is the dependent
variable and whatever variable(s) are to the right of ~ are the predictor(s).
We now fit the model using the sem function, placing the results into the data object fit. Note
that we include the meanstructure argument with the sem function so that we will obtain an
estimate of the regression intercept. Last, we use the summary function here to display the
primary output.
fit <- sem(model.1, data=afdp, meanstructure=TRUE)
summary(fit)

We now consider the output produced by the summary function. First, we obtain some basic
information about the model estimation:
lavaan 0.6-2 ended normally after 11 iterations

Optimization method NLMINB

Number of free parameters 3

Number of observations 316

Estimator ML
Model Fit Test Statistic 0.000
Degrees of freedom 0
Minimum Function Value 0.0000000000000

This shows us that we have 316 cases in our sample data and three parameters that we are
estimating. Additionally, we are using ML (maximum likelihood) estimation with the nlminb
optimizer (this is a built-in optimizer in R for finding the minimum of a function).
It is worth pausing here to note that lavaan counts parameters differently than we described in
lecture. For this model, we would normally count five parameters, the mean and variance of
COA, the intercept and slope of the regression of Peer on COA, and the residual variance of
Peer. But lavaan does not count the mean or variance of the predictor, COA, as parameters of

Copyright © 2019 Curran-Bauer Analytics, LLC

1-6 | Chapter 1 Introduction, Background, and Multiple Regression

the model. Nor does it count covariances among predictors (though there are no such
covariances in the current model) as free parameters. Fortunately, lavaan also does not count
these parameters when determining the number of observed moments, so the degrees of
freedom for the test statistic work out regardless (e.g., here it neither counts the mean and
variance of COA as observed moments nor does it count them as estimated moments, leaving a
net difference of zero when calculating the degrees of freedom).
This output also provides some information about model fit. Multiple regression models are
just identified (i.e., every piece of information provided by the sample is ‘used up’ to estimate
model parameters so that no degrees of freedom remain). Therefore, the model fits the data
perfectly and it is not worthwhile to interpret the test statistic (this and other fit indices will be
discussed in later sections).
The estimates for the model are shown next:
Parameter Estimates:

Information Expected
Information saturated (h1) model Structured
Standard Errors Standard

Regressions:
Estimate Std.Err z-value P(>|z|)
peer ~
coa 0.176 0.060 2.956 0.003

Intercepts:
Estimate Std.Err z-value P(>|z|)
.peer 0.298 0.043 6.884 0.000

Variances:
Estimate Std.Err z-value P(>|z|)
.peer 0.280 0.022 12.570 0.000

Recall that our model is

peeri     coai   i
The output is subdivided into three parts, regressions, intercepts, and variances. In the
Regressions section, the peer ~ coa coefficient is the slope parameter estimate ˆ . In the
Intercepts section, the coefficient listed for .peer is the intercept ̂ . Finally, in the Variances
section, the coefficient listed for .peer is the represents represents ˆ , the estimated variance
of  i (i.e., residual variance). The Estimate column lists is the ML point estimate of each
parameter. The Std.Err column gives the standard error for each estimate (where these are
computed using the expected information matrix, as noted above). Next, the z-value column
reports the Wald z-statistic for the null hypothesis test that the parameter is significantly
different from zero in the population, and the P(>|z|) column reports the p-value associated
with this z-statistic.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-7

Here, we see that the average non-COA has a score of .298 on peer and the average COA has a
score that is .176 units higher than non-COAs on peer (.298 + .176 = .474). Both the intercept
and slope are significantly different from zero. Finally, the variance in deviant peer association
that is not explained by COA status is .280. This indicates that, although COA status is a
significant predictor of peer, COA status does not fully account for affiliation with peers.
Because peer is not on an intrinsically meaningful scale, we cannot easily interpret the
differences among the regression parameters or residual variances. To better interpret these
results, we can request standardized estimates. Within lavaan, several methods of
standardization are available within the summary function. The first type, obtained by including
standardized=TRUE, is the typical fully standardized solution. In this case, both the
independent and dependent variables are standardized to have a variance of 1 (note that
lavaan does not, however, also make the means 0, as one might expect). We can see this by
running the following line from the R script:
summary(fit, standardized=TRUE)

The relevant results are shown here:

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
peer ~
coa 0.176 0.060 2.956 0.003 0.176 0.164

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.peer 0.298 0.043 6.884 0.000 0.298 0.554

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.peer 0.280 0.022 12.570 0.000 0.280 0.973

The Std.lv column refers to a solution in which only the latent variables are standardized and
not the observed variables. Since the current model contains no latent variables, this is of little
use here. The Std.all column contains the fully standardized estimates of interest, in which
all variables (observed and latent) are standardized to have a variance of one. Thus, we say that
a one standard deviation increase in COA status is associated with a .164 standard deviation
increase in deviant peer associations.
It may be more useful to retain the original scaling of COA while standardizing peer. If we
include std.nox=TRUE in the summary function, as shown below, we will obtain estimates in
which the latent variables and endogenous observed variables are standardized but the
exogenous observed variables are left in their raw scale. These are commonly known as
partially standardized estimates.
summary(fit, std.nox=TRUE)

Copyright © 2019 Curran-Bauer Analytics, LLC

1-8 | Chapter 1 Introduction, Background, and Multiple Regression

The relevant output is shown here:

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
peer ~
coa 0.176 0.060 2.956 0.003 0.176 0.329

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer 0.298 0.043 6.884 0.000 0.298 0.554

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer 0.280 0.022 12.570 0.000 0.280 0.973

We can interpret these results by saying that the average COA affiliates with deviant peers
about .329 standard deviations more than the average non-COA.
Finally, we can also calculate how much variance is explained in peer by coa by requesting that
lavaan output the r-squared statistic:
summary(fit, rsquare=TRUE)

We now obtain a new section of output, shown here:

R-Square:
Estimate
peer 0.027

R-Square is the estimated proportion of variance in peer that is accounted for by the models
(i.e., COA, the only predictor included in this model). We have explained less than 3% of the
total variance in peer with the COA predictor. Note, however, that measured variable
regression models assume that no measurement error is present. If measurement error is
present, the estimated relationship among the variables in the model may be attenuated and
the R2 value will also be underestimated.

Multiple Regression Model with COA Status, Gender, and Age as Predictors
We now expand the model to include gender and age as additional predictors of deviant peer
relations as shown in the diagram below. All of the predictors are implicitly allowed to covary
with one another.

coa 1

2
gen peer i 
3

age

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-9

The model is of the form peeri     1coai   2 geni   3agei   i . Each  is interpreted as the
effect of the associated predictor on peer, holding the other predictors constant. Unless all
predictors are uncorrelated with one another, these estimates will change depending on which
other predictors are included in the model.  represents the expected value of peer when all
predictors are equal to zero.
We assume that none of the variables in our model contain measurement error, that
 i ~ N (0, ) for all individuals i, that the error variance is constant for all predictors, and that
 i is uncorrelated with all of the predictors in the model.
The R script for fitting this model is shown below:
model.2 <- 'peer ~ coa + gen + age'

fit <- sem(model.2, data=afdp, meanstructure=TRUE)

summary(fit, std.nox=TRUE, rsquare=TRUE)

Here, we have defined a new model syntax object, model.2, in which we have specified that
peer is regression on coa, gen, and age. Note that, again, we use the regression operator ~ to
indicate the regression. Now that our model includes multiple predictors, we use the + operator
in between each predictor.
The results output by the summary function (including both partially standardized estimates and
r-square) are shown here:
lavaan 0.6-2 ended normally after 18 iterations

Optimization method NLMINB

Number of free parameters 5

Number of observations 316

Estimator ML
Model Fit Test Statistic 0.000
Degrees of freedom 0
Minimum Function Value 0.0000000000000

Parameter Estimates:

Information Expected
Information saturated (h1) model Structured
Standard Errors Standard

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
peer ~
coa 0.210 0.054 3.858 0.000 0.210 0.391
gen -0.048 0.055 -0.877 0.381 -0.048 -0.089
age 0.151 0.019 7.993 0.000 0.151 0.281

Copyright © 2019 Curran-Bauer Analytics, LLC

1-10 | Chapter 1 Introduction, Background, and Multiple Regression

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer -1.610 0.250 -6.453 0.000 -1.610 -2.999

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer 0.232 0.018 12.570 0.000 0.232 0.804

R-Square:
Estimate
peer 0.196

We see that gen is not a significant predictor of peer when COA and age are also included in
the model (p = .381). However, age is significantly related with peer such that older
adolescents are more likely to associate with deviant peers. We estimate that a one-year
increase in age is associated with a .151-increase in deviant peer association ratings. COA
remains a significant predictor of peer, and the regression parameter estimate has not
changed much from the single-predictor model to the multiple-predictor model relative to its
standard error (i.e., it has increased from .18 to .21). The stability of this coefficient reflects the
fact that COA is not highly correlated with gen or age.
The regression coefficients can be interpreted as follows. The slope of peer on COA is the
average effect of being a child of an alcoholic, holding age and gender constant. The slope of
age is the average effect of age on peer, holding COA status and gender constant. The
intercept is less informative in this model because age has not been centered. It now
represents the average deviant peer association score for female, non-COAs who are zero years
old. This is obviously outside of the range of our data.
To better interpret these results, and to get a sense for the relative contribution of each
predictor, we requested the partially standardized solution. We prefer this method of
standardization for this example because all three predictors have natural scales. As can be
seen in the Std.nox column of output, after controlling for age and gender, COAs affiliate
with deviant peers about .391 standard deviation units more than non-COAs. Each additional
year of age is associated with a .281 standard deviation increase in self-reported affiliation with
deviant peers.
From the R-square output, we can see that the multiple predictor regression model explains
more of the variance in peer than the single predictor model. Age and gender account for an
additional 17% of the variance in peer, over and above COA status. Still, approximately 80% of
the variance in peer remains unexplained by the variables in our model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-11

Multiple Regression Model with COA Status, Gender, Age, Stress, Emotion, and Negative
Affect as Predictors
To see if we can explain more of the variance in deviant peer associations, we expand the
model to include stress, emotion, and negaff, while retaining COA, gen, and age.
The model is now
peeri     1coai   2 geni   3agei   4emotei   5 stressi   6 negaff i   i
with the same assumptions as before. Although gen was not a significant predictor of deviant
peer relations in the last model, we do not exclude it from this model because it may covary
with the new predictors in such a way that it is important to include it as a control variable in
the model. The full model is shown in the diagram.

coa 1

gen 2

3 
age
peer i 
4
emote
5
stress
6

negaff

The R script for fitting this model is shown below:

model.3 <- 'peer ~ coa + gen + age + emotion + stress + negaff'

fit <- sem(model.3, data=afdp, meanstructure=TRUE)

summary(fit, standardized=TRUE, rsquare=TRUE)
summary(fit, std.nox=TRUE, rsquare=TRUE)

As before, we simply defined a new model syntax object, model.3, which includes the new
predictors of peer. In this case we requested both the fully standardized and partially
standardized solutions in two calls of the summary function. For coa, gen, and age, the
partially standardized coefficients are most interpretable. In contrast, emotion, stress, and

Copyright © 2019 Curran-Bauer Analytics, LLC

1-12 | Chapter 1 Introduction, Background, and Multiple Regression

negaff do not have intrinsically meaningful scales, so the fully standardized estimates are
better. Thus, we consider both. Additionally, by examining the fully standardized estimates
across the full set of predictors we can gauge the relative contributions of the predictors,
despite the fact that they were originally on quite different scales.
Here, we show the output from the first summary function requesting fully standardized
effects:
lavaan 0.6-2 ended normally after 22 iterations

Optimization method NLMINB

Number of free parameters 8

Number of observations 316

Estimator ML
Model Fit Test Statistic 0.000
Degrees of freedom 0

Parameter Estimates:

Information Expected
Information saturated (h1) model Structured
Standard Errors Standard

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
peer ~
coa 0.137 0.055 2.492 0.013 0.137 0.127
gen -0.027 0.052 -0.509 0.611 -0.027 -0.025
age 0.140 0.018 7.660 0.000 0.140 0.378
emotion 0.030 0.058 0.520 0.603 0.030 0.028
stress 0.111 0.044 2.546 0.011 0.111 0.140
negaff 0.109 0.030 3.660 0.000 0.109 0.195

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.peer -1.934 0.267 -7.233 0.000 -1.934 -3.602

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.peer 0.210 0.017 12.570 0.000 0.210 0.728

R-Square:
Estimate
peer 0.272

We see that gen is still not a significant predictor of peer when other predictors are included in
the model (p = .611), but that COA and age remain statistically significant even though the
value of their regression coefficients have changed. Importantly, the effect of COA on peer is
not as strong after controlling for emotion, stress, and negaff, indicating that these

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple Regression Analysis of Deviant Peer Affiliations | 1-13

variables are somewhat related to one another. These new variables could potentially mediate
the relationship between COA and peer; we will later explore this possibility with a path
analysis model. Of the new variables, it appears that stressful life events and negative affect are
significant predictors of association with deviant peers, but self-reported emotional
expressiveness is not significantly related to deviant peer association, after controlling for age,
gen, stress, negaff, and COA.
Examining the standardized estimates in the Std.all column, we see that age is the strongest
relative predictor of peer, followed by negaff, stress, COA, emotion, and then gen.
However, determining the “strongest” predictors is always a tricky endeavor, particularly given
that gen and COA are binary predictors for which fully standardized effects are less useful.
We can examine the results from the second summary call to see the partially standardized
estimates.
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
peer ~
coa 0.137 0.055 2.492 0.013 0.137 0.255
gen -0.027 0.052 -0.509 0.611 -0.027 -0.049
age 0.140 0.018 7.660 0.000 0.140 0.262
emotion 0.030 0.058 0.520 0.603 0.030 0.056
stress 0.111 0.044 2.546 0.011 0.111 0.206
negaff 0.109 0.030 3.660 0.000 0.109 0.204

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer -1.934 0.267 -7.233 0.000 -1.934 -3.602

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.peer 0.210 0.017 12.570 0.000 0.210 0.728

R-Square:
Estimate
peer 0.272

Finally, examining the R-square output, we see that the final model explains about 27% of the
total variance in peer.

Copyright © 2019 Curran-Bauer Analytics, LLC

1-14 | Chapter 1 Introduction, Background, and Multiple Regression

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 2
Path Analysis: Part I
Path Analysis with Deviant Peer Affiliation Data ..................................................................... 2-3
2-2 | Chapter 2 Path Analysis: Part I

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 2-3

Path Analysis with Deviant Peer Affiliation Data

The data for this demonstration were provided by Dr. Laurie Chassin from the Adolescent and
Family Development Project, housed at Arizona State University. Note that these data were
generously provided for strictly pedagogical purposes and should not be used for any other
purposes beyond this workshop. The sample includes 316 adolescents, between 10-16 years of
age. The study was designed to assess the association between parental alcoholism and
adolescent substance use and psychopathology. The data are in the text file afdp.dat and the
accompanying R script is in the ch02.R file. The variables in the data set that we will use are
peer adolescent report on peer substance use and peer tolerance of use
coa parent report of alcoholism diagnosis where 0=non-alcoholic and 1=alcoholic
gen gender where 0=girl and 1=boy
age age measured in years at assessment
stress self report measure of uncontrollable negative life stressful events
emotion self report measure of temperamental emotional expressiveness
negaff self report measure of depression and anxiety

Path Analysis of Theoretical Peer Affiliation Model

Theory dictates that alcoholic parents increase the number of stressful life events that their
children experience, leading to an increase in child negative affect. Further, children of
alcoholics are hypothesized to have higher levels of emotionality, leading to more negative
affect. Negative affect is thought to be related to higher rates of affiliation with deviant peers.
Stressful life events and emotionality should covary, but we hypothesize no directional relation
among these variables. We allow age and gen to predict stress and emotion, and we allow
all exogenous characteristics (coa, gen, and age) to covary.

s
coa
n p
stress
gen negaff peer
emote
age e

Copyright © 2019 Curran-Bauer Analytics, LLC

2-4 | Chapter 2 Path Analysis: Part I

The model is of the form

 ystressi      12  13   0 0 0 0  stressi    1i 
   1i   11   coai   0    
 yemotei    2i    21  22  23     0 0 0  emotei    2i 
y       0 0 0  
geni 
  32 0

0  negaffi    3i 
where
 negaffi   3i    agei    31
   
 y peer    4i   0 0 0   0 0  43 0  peeri    4i 
 i 
 11 
 
  22
COV  ζi    21 
 0 0  33 
 
 0 0 0  44 

The R script that fits this model with the lavaan package is shown below (note that this assumes
the data file is in the same directory as the R script file, i.e., the source directory):
#reading in data
afdp <- read.table("afdp.dat", header=FALSE)
varnames <-c("id","coa","age","gen","stress","emotion","negaff","peer")
names(afdp)<-varnames

#loading lavaan package

library(lavaan)

#fitting path analysis model

pathmodel.1 <- '
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion
peer ~ negaff

#covariances
stress ~~ emotion
'
fit <- sem(pathmodel.1, data=afdp, meanstructure=TRUE)
summary(fit, standardized=TRUE, rsquare=TRUE)
summary(fit, std.nox=TRUE, rsquare=TRUE)

fitted(fit)
resid(fit, type="raw")
resid(fit, type="normalized")

We have seen most of this code in Chapter 1, so here we will focus only on what’s new in the
definition of the model syntax object. Note that some lines in this script, those that are
preceded with a # character, are comments and are not part of the actual syntax.
In Chapter 1, peer was the only dependent variable and was regressed on all of the other
variables in the model. Here, we have multiple endogenous variables and each has its own

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 2-5

regression equation, as specified in the section of the syntax beginning with the #regressions
comment. Thus, we indicate that there are direct effects of coa, gen, and age on stress with
the line stress ~ coa + gen + age. In each subsequent line, we indicate the direct effects
of the predictors for each endogenous variable in the model, ending with peer being regressed
on negaff.
As a default, lavaan allows all exogenous variables to covary but it fixes the residual covariances
among endogenous variables to zero. The residual terms of stress and emotion are freed to
covary within the section of the code marked #covariances. To designate a covariance (or
variance) we use the ~~ operator. Thus, stress ~~ emotion tells lavaan to include a residual
covariance for stress and emotion.
In the two summary function calls we requested standardized estimates and partially
standardized estimates, respectively, to aid interpretation of the results. The final three lines
request the model-implied covariance matrix (fitted(fit)) and the raw and standardized
residual matrices (resid(fit, type="raw") and resid(fit, type="normalized"),
respectively). These residuals represent the differences between the observed covariance
matrix and the matrix that is implied by the structure of the model.
Typically, we would evaluate model fit prior to interpreting parameter estimates. For
pedagogical purposes, however, we will put aside a discussion of model fit until Chapter 3 and
move directly to parameter estimates for our model. Here we see the raw and fully
standardized estimates produced by the first summary function call:
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
stress ~
coa 0.451 0.072 6.223 0.000 0.451 0.331
gen -0.016 0.073 -0.215 0.830 -0.016 -0.011
age 0.002 0.025 0.078 0.938 0.002 0.004
emotion ~
coa 0.110 0.056 1.963 0.050 0.110 0.110
gen -0.048 0.056 -0.843 0.399 -0.048 -0.047
age -0.027 0.019 -1.374 0.170 -0.027 -0.077
negaff ~
stress 0.246 0.078 3.134 0.002 0.246 0.175
emotion 0.553 0.106 5.206 0.000 0.553 0.290
peer ~
negaff 0.176 0.030 5.892 0.000 0.176 0.315
Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress ~~
.emotion 0.112 0.019 5.896 0.000 0.112 0.352
Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress 0.687 0.333 2.066 0.039 0.687 1.010
.emotion 2.341 0.258 9.088 0.000 2.341 4.663
.negaff 1.527 0.207 7.373 0.000 1.527 1.594
.peer -0.118 0.091 -1.298 0.194 -0.118 -0.220

Copyright © 2019 Curran-Bauer Analytics, LLC

2-6 | Chapter 2 Path Analysis: Part I

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress 0.412 0.033 12.570 0.000 0.412 0.890
.emotion 0.247 0.020 12.570 0.000 0.247 0.979
.negaff 0.778 0.062 12.570 0.000 0.778 0.848
.peer 0.260 0.021 12.570 0.000 0.260 0.901

R-Square:
Estimate
stress 0.110
emotion 0.021
negaff 0.152
peer 0.099

The second summary function call reproduces these same raw parameter estimates but now
computes the partially standardized estimates:
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
stress ~
coa 0.451 0.072 6.223 0.000 0.451 0.663
gen -0.016 0.073 -0.215 0.830 -0.016 -0.023
age 0.002 0.025 0.078 0.938 0.002 0.003
emotion ~
coa 0.110 0.056 1.963 0.050 0.110 0.219
gen -0.048 0.056 -0.843 0.399 -0.048 -0.095
age -0.027 0.019 -1.374 0.170 -0.027 -0.053
negaff ~
stress 0.246 0.078 3.134 0.002 0.246 0.175
emotion 0.553 0.106 5.206 0.000 0.553 0.290
peer ~
negaff 0.176 0.030 5.892 0.000 0.176 0.315

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.stress ~~
.emotion 0.112 0.019 5.896 0.000 0.112 0.352

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.stress 0.687 0.333 2.066 0.039 0.687 1.010
.emotion 2.341 0.258 9.088 0.000 2.341 4.663
.negaff 1.527 0.207 7.373 0.000 1.527 1.594
.peer -0.118 0.091 -1.298 0.194 -0.118 -0.220

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
.stress 0.412 0.033 12.570 0.000 0.412 0.890
.emotion 0.247 0.020 12.570 0.000 0.247 0.979
.negaff 0.778 0.062 12.570 0.000 0.778 0.848
.peer 0.260 0.021 12.570 0.000 0.260 0.901

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 2-7

R-Square:
Estimate
stress 0.110
emotion 0.021
negaff 0.152
peer 0.099

We see that the hypothesized pathways to deviant peer affiliation do contain statistically
significant components. To aid in interpretation, standardized values are included in the model
path diagram, shown below. We report fully standardized effects (from Std.all column
generated by the first summary function call) except for coa, gen, and age, for which we report
partially standardized effects (from Std.nox column generated by the second summary function
call). Recall that only the outcome variables are standardized in computing the partially
standardized effects, which makes them particularly useful for examining the effects of coding
variables (e.g., coa and gen) or predictors with natural metrics (e.g., age).
These partially standardized parameter estimates represent the expected change in standard
deviation units in y given a one unit increase in x. By comparison, the fully standardized
estimates are computed by standardizing both the predictors and the outcome variables so that
parameter estimates represent the expected change in standard deviation units in y given a one
standard deviation increase in x. These are most useful for interpreting effects when neither
predictors nor outcomes have natural scales and when gauging relative effect sizes across
predictors on different scales. For covariance parameters, fully standardized estimates can be
interpreted as correlations.
From the R-Square section of output, we can also see that only a modest amount of the total
variance in any of these variables has been explained by the model.
Finally, the model-implied covariance matrix and the raw and standardized residuals between
the observed and model-implied covariance matrices are:
> fitted(fit)
$`cov`
stress emotin negaff peer coa gen age
stress 0.463
emotion 0.125 0.252
negaff 0.183 0.170 0.917
peer 0.032 0.030 0.162 0.288
coa 0.112 0.029 0.044 0.008 0.249
gen -0.003 -0.010 -0.006 -0.001 0.002 0.249
age -0.019 -0.059 -0.037 -0.007 -0.055 -0.070 2.095

$mean
stress emotion negaff peer coa gen age
0.941 2.034 2.883 0.390 0.525 0.538 12.718

Copyright © 2019 Curran-Bauer Analytics, LLC

2-8 | Chapter 2 Path Analysis: Part I

> resid(fit, type="raw")

$`type`
[1] "raw"

$cov
stress emotin negaff peer coa gen age
stress 0.000
emotion 0.000 0.000
negaff 0.000 0.000 0.000
peer 0.055 0.006 0.000 0.000
coa 0.000 0.000 -0.004 0.036 0.000
gen 0.000 0.000 -0.042 -0.021 0.000 0.000
age 0.000 0.000 0.247 0.314 0.000 0.000 0.000

$mean
stress emotion negaff peer coa gen age
0 0 0 0 0 0 0

> resid(fit, type="normalized")

$`type`
[1] "normalized"

$cov
stress emotin negaff peer coa gen age
stress 0.000
emotion 0.000 0.000
negaff 0.000 0.000 0.000
peer 2.626 0.394 0.000 0.000
coa 0.000 0.000 -0.159 2.375 0.000
gen 0.000 0.000 -1.541 -1.383 0.000 0.000
age 0.000 0.000 3.130 6.681 0.000 0.000 0.000

$mean
stress emotion negaff peer coa gen age
0 0 0 0 0 0 0

The normalized residuals are rescaled to follow a standard normal distribution and values
exceeding plus or minus two are often taken as potentially meaningful in magnitude. The
normalized residuals for this model suggest that the hypothesized structure is doing a poor job
in reproducing the observed covariances between several variables, most notably between age
and peer, between age and negaff, and between stress and peer. We will explore methods for
more formally assessing overall model fit in the next chapter.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 2-9

Finally, below is a path diagram denoting all estimated parameters. Significant effects are
denoted with solid lines (p<.05) and non-significant with dashed lines. Partially standardized
effects are presented for COA, gender, and age and fully standardized are presented for all
other predictors.

 s .89
coa .66 .85 .90
.22 n p
.35
stress .18
-.02
.32
gen .29
negaff peer
-.10
.003
emote
-.05
age  e .98

Copyright © 2019 Curran-Bauer Analytics, LLC

2-10 | Chapter 2 Path Analysis: Part I

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 3
Path Analysis: Part II
Path Analysis with Deviant Peer Affiliation Data ..................................................................... 3-3
Theoretical Model ................................................................................................................ 3-3
Modification to Peer Affiliation Model: Likelihood Ratio Test ............................................. 3-5
Modification to Peer Affiliation Model: Modification Indices ............................................. 3-8
Tests of Direct, Indirect, and Specific Effects ..................................................................... 3-16
3-2 | Chapter 3 Path Analysis: Part II

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-3

Path Analysis with Deviant Peer Affiliation Data

The data for this demonstration were provided by Dr. Laurie Chassin from the Adolescent and
Family Development Project, housed at Arizona State University. Note that these data were
generously provided for strictly pedagogical purposes and should not be used for any other
purposes beyond this workshop. The sample includes 316 adolescents, between 10-16 years of
age. The study was designed to assess the association between parental alcoholism and
adolescent substance use and psychopathology. The data are in the text file afdp.dat and the
accompanying R script is in the ch03.R file. The variables in the data set that we will use are
peer adolescent report on peer substance use and peer tolerance of use
coa parent report of alcoholism diagnosis where 0=non-alcoholic and 1=alcoholic
gen gender where 0=girl and 1=boy
age age measured in years at assessment
stress self report measure of uncontrollable negative life stressful events
emotion self report measure of temperamental emotional expressiveness
negaff self report measure of depression and anxiety

Theoretical Model
We tested our hypothesized theoretical model of pathways to deviant peer affiliation for COAs
and non-COAs. The original model is illustrated below with standardized parameter estimates
overlaid on the path diagram (regression paths from coa, gen, age are partially standardized;
all other estimates are fully standardized):

 s .89
coa .66 .85 .90
.22 n p
.35
stress .18
-.02
.32
gen .29
negaff peer
-.10
.003
emote
-.05
age  e .98

Copyright © 2019 Curran-Bauer Analytics, LLC

3-4 | Chapter 3 Path Analysis: Part II

The R script for fitting this model was shown in Chapter 2 and is included here again for
reference.
pathmodel.1 <-'
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion
peer ~ negaff

#covariances
stress ~~ emotion
'
fit.1 <- sem(pathmodel.1, data=afdp, meanstructure=TRUE)

The only modification to the script is that the object previously named fit has been renamed
fit.1, as we will be comparing fit across multiple models in this chapter.

We obtain measures of fit, appropriately enough, by including fit.measures=TRUE as an

argument to the summary function, as shown here:
summary(fit.1, fit.measures=TRUE)

Although the output includes the parameter estimates, we will focus here only on output
related to model fit. Some of this information is already included in the default output header:
lavaan 0.6-2 ended normally after 43 iterations

Optimization method NLMINB

Number of free parameters 18

Number of observations 316

Estimator ML
Model Fit Test Statistic 81.173
Degrees of freedom 8
P-value (Chi-square) 0.000

Lavaan lists the number of free parameters as 18, but recall that it does not count the means,
variances, or covariances of the exogenous predictors (coa, gen, and age) as estimated
parameters. The number of free parameters reported by lavaan for this model thus does not
equal what we computed by the t-rule (27 free parameters, including 3 means, 3 variances, and
3 covariances for coa, gen, and age; 16 + 9 = 27). The t-rule still works out to 8 degrees of
freedom for the chi-square, however, because the observed means, variances, and covariances
of the covariates are also not counted in the number of sample means and variances k.
As we can see, the chi square-distributed likelihood ratio test of model fit rejects the null
hypothesis that the model fits the data. However, to get a full understanding of model fit we
must also consider other fit indices.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-5

The additional output generated by fit.measures=TRUE is shown below

Model test baseline model:

Minimum Function Test Statistic 251.027

Degrees of freedom 18
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.686

Tucker-Lewis Index (TLI) 0.293

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -1158.938

Loglikelihood unrestricted model (H1) -1118.351

Number of free parameters 18

Akaike (AIC) 2353.875
Bayesian (BIC) 2421.479
Sample-size adjusted Bayesian (BIC) 2364.387

Root Mean Square Error of Approximation:

RMSEA 0.170
90 Percent Confidence Interval 0.138 0.205
P-value RMSEA <= 0.05 0.000

Standardized Root Mean Square Residual:

SRMR 0.085

In addition to poor model chi-square, the CFI and the TLI are far lower than .9, the standard
lower bound for a good-fitting model. Finally, the 90% confidence interval for the RMSEA does
not even include .10 at the lower bound, indicating terrible fit. As such, we cannot interpret
the obtained parameter estimates with any confidence given the severe misfit of the model.

Modification to Peer Affiliation Model: Likelihood Ratio Test

Since the theoretical model of deviant peer affiliation that was presented in Chapter 2 did not
fit the data well, we will consider model modifications to improve our representation of the
data. First, theory might suggest that it is an excessively severe restriction to require that the
influence of parental alcoholism be conveyed entirely by the mediators. Thus we can allow coa
to directly predict negaff and peer, over and above its indirect relationship with these
variables via stress and emotion.

Copyright © 2019 Curran-Bauer Analytics, LLC

3-6 | Chapter 3 Path Analysis: Part II

The new paths are illustrated with dotted lines in the path diagram below.

coa
s

n p
stress
gen negaff peer
emote

e

age

The model is now of the form:

 ystressi         13   0 0 0 0  stressi    1i 
   1i   11 12   coai      
 yemotei    2i    21  22  23     0 0 0 0  emotei    2i 
y        0 0  
geni  32 0

0  negaffi    3i 
 negaffi   3i   31  agei  
31
   
 y peer    4i    41 0 0   0 0  43 0  peeri    4i 
 i 

where
 11 
 
 21  22 
COV  ζi  
 0 0  33 
 
 0 0 0  44 

The R script for this model is shown below:

pathmodel.2 <- '
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion + coa
peer ~ negaff + coa

#covariances
stress ~~ emotion
'
fit.2 <- sem(pathmodel.2, data=afdp, meanstructure=TRUE)
summary(fit.2, fit.measures=TRUE)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-7

The only difference from the pathmodel.1 model syntax object is that coa is included as a
predictor in the lines for peer and negaff.
Let us now turn to the model fit to determine whether freeing these two regression parameters
has led to a significant improvement in model fit.
lavaan 0.6-2 ended normally after 47 iterations

Optimization method NLMINB

Number of free parameters 20

Number of observations 316

Estimator ML
Model Fit Test Statistic 74.321
Degrees of freedom 6
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 251.027

Degrees of freedom 18
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.707

Tucker-Lewis Index (TLI) 0.120

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -1155.511

Loglikelihood unrestricted model (H1) -1118.351

Number of free parameters 20

Akaike (AIC) 2351.023
Bayesian (BIC) 2426.138
Sample-size adjusted Bayesian (BIC) 2362.703

Root Mean Square Error of Approximation:

RMSEA 0.190
90 Percent Confidence Interval 0.153 0.230
P-value RMSEA <= 0.05 0.000

Standardized Root Mean Square Residual:

SRMR 0.081

Copyright © 2019 Curran-Bauer Analytics, LLC

3-8 | Chapter 3 Path Analysis: Part II

The likelihood for the original model (Model A) was TA = 81.17 on 8 degrees of freedom. We
have used two addition model parameters for the modified model (Model B), and the likelihood
test statistic for this model is 74.32. These models are nested, so we can conduct a Likelihood
Ratio Test (LRT) to determine whether Model B fits significantly better than Model A. Lavaan
nicely provides a function for conducting the LRT, lavTestLRT, in which you simply provide the
fit objects from Models B and A (in that order, where A is more restricted than B), as shown
here for the present example:
lavTestLRT(fit.2, fit.1)

This generates the output

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.2 6 2351.0 2426.1 74.321
fit.1 8 2353.9 2421.5 81.173 6.8522 2 0.03251 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Thus we have
T  TA  TB  81.17  74.32  6.85
df   df A  df B  8  6  2
T ~  2 (df  )  p  .033
Given that p < .05, we can reject the null hypothesis that there is no difference in model fit
between Model A and Model B. There is thus support for including the two additional paths.
However, given that the sample size is rather large and the effect size (gain in model
improvement) is rather small, the LRT suggests only a trivial gain in model fit associated with
this modification.
Examining relative tests of model fit for this new model, however, we see that the fit is still
terrible. Our attempt to modify the model based upon a priori hypotheses has failed to
produce an acceptable model.

Modification to Peer Affiliation Model: Modification Indices

Returning to the original hypothesized model, we will take a different approach to model
modification. We can request that lavaan suggest changes to our model based on the expected
change in model chi-square if a fixed parameter were freed; these are the modification indices
(MIs). We can obtain MIs with the modindices function, as shown here:
modindices(fit.1, sort.=TRUE, minimum.value=10)

The first argument is the fit/results object generated by lavaan for the model of interest, here
fit.1. The sort. argument is set to TRUE so that lavaan will present the largest MIs first.
Additionally, the minimum.value argument is set to suppress printing of any MIs less than 10.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-9

The minimum.value argument can be changed to a smaller or larger number, as desired, or the
omitted entirely to look at all possible MIs.
Model modification indices are shown below:
lhs op rhs mi epc sepc.lv sepc.all sepc.nox
63 age ~ peer 53.942 1.114 1.114 0.413 0.413
47 peer ~ age 42.540 0.129 0.129 0.348 0.241
42 negaff ~ age 11.889 0.119 0.119 0.179 0.124
62 age ~ negaff 10.877 0.303 0.303 0.201 0.201

Here, lhs stands for “on the left-hand side of the operator” and rhs stands for “on the right-
hand side of the operator.” The operator, denoted op, for these parameters is ~, indicating
these are regression slopes. mi denotes the expected improvement in the model chi square
test statistic if the modification is accepted. epc denotes “expected parameter change (EPC)”
and gives the expected parameter value if the modification is implemented (since it is changing
from zero). sepc.lv re-scales the EPC by standardizing any latent variables in the model,
sepc.all re-scales the EPC by standardizing all variables (observed and latent), and sepc.nox
re-scales the EPC by standardizing all variables except exogenous predictors. Here, sepc.lv is
equivalent to the EPC because there are no latent variables in the model. It can be helpful to
rely on standardized EPC values in order to get a sense of the relative magnitude of each
potential modification.
Note the overlap in suggested modification indices. This tells us that our original model does
not allow for a significant relation between age and negaff or between age and peer. These
suggestions are not independent from one another; allowing negaff to relate directly with
age would imply an increased correlation between negaff and peer.
It is more theoretically justifiable to regress peer on age than to regress negaff on age; thus,
we will proceed with this model modification, as shown in the code below.
pathmodel.3 <-'
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion
peer ~ negaff + age

#covariances
stress ~~ emotion
'
fit.3 <- sem(pathmodel.3, data=afdp, meanstructure=TRUE)
summary(fit.3, fit.measures=TRUE)

Relative to the model syntax object for our first model, all we have done is add age as a
predictor to the right of the ~ statement for peer.

Copyright © 2019 Curran-Bauer Analytics, LLC

3-10 | Chapter 3 Path Analysis: Part II

We obtain the following model fit:

lavaan 0.6-2 ended normally after 49 iterations

Optimization method NLMINB

Number of free parameters 19

Number of observations 316

Estimator ML
Model Fit Test Statistic 34.368
Degrees of freedom 7
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 251.027

Degrees of freedom 18
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.883

Tucker-Lewis Index (TLI) 0.698

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -1135.535

Loglikelihood unrestricted model (H1) -1118.351

Number of free parameters 19

Akaike (AIC) 2309.070
Bayesian (BIC) 2380.429
Sample-size adjusted Bayesian (BIC) 2320.166

Root Mean Square Error of Approximation:

RMSEA 0.111
90 Percent Confidence Interval 0.076 0.150
P-value RMSEA <= 0.05 0.003

Standardized Root Mean Square Residual:

SRMR 0.057

We can statistically compare the fit of this model with that of the original model using a LRT:
lavTestLRT(fit.3, fit.1)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-11

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.3 7 2309.1 2380.4 34.368
fit.1 8 2353.9 2421.5 81.173 46.805 1 7.841e-12 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

In other words, we have

T  TA  TB  81.17  34.37  46.80
df   df A  df B  8  7  1
T ~  2 (df  )  p  .001
Thus, including age as a predictor of deviant peer affiliation has significantly improved the fit of
the model. However, the relative model fit is still poor. We can again examine the modification
indices to determine whether a justifiable model modification would lead to further
improvements in model fit.
modindices(fit.3, sort.=TRUE, minimum.value=10)

lhs op rhs mi epc sepc.lv sepc.all sepc.nox

44 peer ~ stress 13.381 0.149 0.149 0.192 0.192
63 age ~ peer 12.858 1.977 1.977 0.721 0.721
43 negaff ~ age 11.889 0.119 0.119 0.179 0.124
46 peer ~ coa 11.853 0.185 0.185 0.175 0.350
62 age ~ negaff 10.859 0.303 0.303 0.200 0.200

The MIs continue to suggest that negaff be directly regressed on age, even after we have
regressed peer on age. However, the MIs also suggest regressing peer on coa, a modification
that we consider to be the most theoretically defensible of the suggested modifications.
To make this modification, we run the following script:
pathmodel.4 <-'
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion
peer ~ negaff + age + coa

#covariances
stress ~~ emotion
'
fit.4 <- sem(pathmodel.4, data=afdp, meanstructure=TRUE)
summary(fit.4, fit.measures=TRUE)

Copyright © 2019 Curran-Bauer Analytics, LLC

3-12 | Chapter 3 Path Analysis: Part II

We obtain the following model fit:

lavaan 0.6-2 ended normally after 50 iterations

Optimization method NLMINB

Number of free parameters 20

Number of observations 316

Estimator ML
Model Fit Test Statistic 22.274
Degrees of freedom 6
P-value (Chi-square) 0.001

Model test baseline model:

Minimum Function Test Statistic 251.027

Degrees of freedom 18
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.930

Tucker-Lewis Index (TLI) 0.790

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -1129.488

Loglikelihood unrestricted model (H1) -1118.351

Number of free parameters 20

Akaike (AIC) 2298.977
Bayesian (BIC) 2374.092
Sample-size adjusted Bayesian (BIC) 2310.657

Root Mean Square Error of Approximation:

RMSEA 0.093
90 Percent Confidence Interval 0.054 0.135
P-value RMSEA <= 0.05 0.038

Standardized Root Mean Square Residual:

SRMR 0.044

A LRT would show that the inclusion of a direct path from coa to peer results in a significant
improvement to the model fit; however, we are still not satisfied with the relative model fit.
Once again, we turn to the modification indices.
modindices(fit.4, sort.=TRUE, minimum.value=10)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-13

lhs op rhs mi epc sepc.lv sepc.all sepc.nox

44 negaff ~ age 11.889 0.119 0.119 0.179 0.124
62 age ~ negaff 10.875 0.303 0.303 0.201 0.201

Modification indices persist in suggesting that age is directly related with negaff. This is the
only remaining modification that has an expected change in model fit that is greater than 10.
Because this is not an unreasonable modification, we will make this change as well:
pathmodel.5 <-'
#regressions
stress ~ coa + gen + age
emotion ~ coa + gen + age
negaff ~ stress + emotion + age
peer ~ negaff + age + coa

#covariances
stress ~~ emotion
'
fit.5 <- sem(pathmodel.5, data=afdp, meanstructure=TRUE)
summary(fit.5, fit.measures=TRUE)

And we obtain the following fit to the data:

lavaan 0.6-2 ended normally after 55 iterations

Optimization method NLMINB

Number of free parameters 21

Number of observations 316

Estimator ML
Model Fit Test Statistic 10.156
Degrees of freedom 5
P-value (Chi-square) 0.071

Model test baseline model:

Minimum Function Test Statistic 251.027

Degrees of freedom 18
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.978

Tucker-Lewis Index (TLI) 0.920

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -1123.429

Loglikelihood unrestricted model (H1) -1118.351

Number of free parameters 21

Copyright © 2019 Curran-Bauer Analytics, LLC

3-14 | Chapter 3 Path Analysis: Part II

Akaike (AIC) 2288.858

Bayesian (BIC) 2367.728
Sample-size adjusted Bayesian (BIC) 2301.122

Root Mean Square Error of Approximation:

RMSEA 0.057
90 Percent Confidence Interval 0.000 0.108
P-value RMSEA <= 0.05 0.345

Standardized Root Mean Square Residual:

SRMR 0.027

By including three data-driven but theoretically acceptable modifications to our original model,
we have obtained good model fit. The CFI and the TLI are both above .9 and the 90%
confidence interval for the RMSEA includes 0. Note, however, that the confidence interval also
includes values greater than .10, so the model fit is not outstanding.
We turn now to the raw and standardized parameter estimates associated with our final model,
again computing both fully and partially standardized estimates.
summary(fit.5, standardized=TRUE, rsquare=TRUE)
summary(fit.5, std.nox=TRUE, rsquare=TRUE)

Here we see the raw and fully standardized estimates:

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
stress ~
coa 0.451 0.072 6.223 0.000 0.451 0.331
gen -0.016 0.073 -0.215 0.830 -0.016 -0.011
age 0.002 0.025 0.078 0.938 0.002 0.004
emotion ~
coa 0.110 0.056 1.963 0.050 0.110 0.110
gen -0.048 0.056 -0.843 0.399 -0.048 -0.047
age -0.027 0.019 -1.374 0.170 -0.027 -0.077
negaff ~
stress 0.243 0.077 3.157 0.002 0.243 0.173
emotion 0.582 0.105 5.568 0.000 0.582 0.305
age 0.119 0.034 3.515 0.000 0.119 0.179
peer ~
negaff 0.137 0.028 4.942 0.000 0.137 0.244
age 0.138 0.018 7.525 0.000 0.138 0.372
coa 0.185 0.053 3.512 0.000 0.185 0.172

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress ~~
.emotion 0.112 0.019 5.896 0.000 0.112 0.352

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-15

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress 0.687 0.333 2.066 0.039 0.687 1.010
.emotion 2.341 0.258 9.088 0.000 2.341 4.663
.negaff -0.038 0.489 -0.078 0.938 -0.038 -0.040
.peer -1.856 0.239 -7.771 0.000 -1.856 -3.457

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.stress 0.412 0.033 12.570 0.000 0.412 0.890
.emotion 0.247 0.020 12.570 0.000 0.247 0.979
.negaff 0.749 0.060 12.570 0.000 0.749 0.816
.peer 0.216 0.017 12.570 0.000 0.216 0.748

And here we see the raw and partially standardized estimates (just shown for the regression
slopes):
Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.nox
stress ~
coa 0.451 0.072 6.223 0.000 0.451 0.663
gen -0.016 0.073 -0.215 0.830 -0.016 -0.023
age 0.002 0.025 0.078 0.938 0.002 0.003
emotion ~
coa 0.110 0.056 1.963 0.050 0.110 0.219
gen -0.048 0.056 -0.843 0.399 -0.048 -0.095
age -0.027 0.019 -1.374 0.170 -0.027 -0.053
negaff ~
stress 0.243 0.077 3.157 0.002 0.243 0.173
emotion 0.582 0.105 5.568 0.000 0.582 0.305
age 0.119 0.034 3.515 0.000 0.119 0.124
peer ~
negaff 0.137 0.028 4.942 0.000 0.137 0.244
age 0.138 0.018 7.525 0.000 0.138 0.257
coa 0.185 0.053 3.512 0.000 0.185 0.345

Finally, the r-square values for the model are given here:
R-Square:
Estimate
stress 0.110
emotion 0.021
negaff 0.184
peer 0.252

Copyright © 2019 Curran-Bauer Analytics, LLC

3-16 | Chapter 3 Path Analysis: Part II

The final path diagram with standardized estimates overlaid is shown below. Note that, as
before, partially standardized estimates (from the std.nox output) are reported for regression
paths emanating from the exogenous variables, COA, gen, and age, as all of these predictors
have meaningful metrics. All other values are fully standardized.

coa .35
 s .89
.66 .82
.75
.22 n
p
stress .17 .35

.24
gen .31 negaff peer
emote
.12
 e .98
.26
age

We originally hypothesized that coa would lead to an increase in uncontrolled stressful life
events, and this was supported by the model. COA status leads to a moderately high increase
in stress. We hypothesized that being a COA would lead to an increase in emotionality, and
this was supported: coa leads to a small-to-moderate increase in emotion. We hypothesized
that stress and emotion would increase negaff, and model results suggest a small, positive
effect of stress on negaff and a moderate effect of emotion on negaff. Finally, we
hypothesized that negative affect would increase affiliation with deviant peers, and we
estimated a moderately positive direct relation between negaff and peer.
Additionally, we found support for a direct effect of coa on peer, suggesting that stress,
emotion, and negaff do not fully account for the total relation between COA status and
affiliation with deviant peers. Furthermore, age appears to account for a significant amount of
the variation in negative affect and affiliation with deviant peers, but age does not appear to be
related to uncontrolled stressful life events or emotionality.

Tests of Direct, Indirect, and Specific Effects

Significant links in a mediational pathway are not sufficient to infer mediation. In order to
formally test the full mediation effect, we need an inferential test of the entire specific indirect
effect in question. Here, we want to know whether the specific mediational pathway of COA
status on deviant peer affiliation via stressful events and negative affect is statistically
significant, whether the specific mediational effect of coa on peer via emotionality and
negative affect is significant, and whether the overall indirect effect of coa on peer is

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-17

significant. Finally, we would like to have an estimate of the total effect of coa on peer
considering both direct and indirect pathways.
The first thing we need to do is define the effects of interest in a new model syntax object, as
shown below:
effects.5 <-'
#regressions
stress ~ c_s*coa + gen + age
emotion ~ c_e*coa + gen + age
negaff ~ s_n*stress + e_n*emotion + age
peer ~ n_p*negaff + age + c_p*coa

#covariances
stress ~~ emotion

#direct effect
dir := c_p

#specific indirect effects

ind_s := c_s*s_n*n_p
ind_e := c_e*e_n*n_p

#total indirect effect

tot_ind := c_s*s_n*n_p + c_e*e_n*n_p

#total effect
tot := c_s*s_n*n_p + c_e*e_n*n_p + c_p

This is precisely the same model as was defined in the pathmodel.5 syntax object. What’s new
here is that we’ve added parameter labels for the paths involved in the direct and indirect
effects of coa on peer. For instance, in the line stress ~ c_s*coa + gen + age, we have
supplied the label c_s for the effect of coa on stress. Likewise, the other paths involved in
computing the effects of interest have also been labeled. To help keep track of things, our
labels consist of the first letter of the predictor, an underscore, and first letter of outcome, but
any labels are fine. We can now refer to these labels when defining direct, indirect, and total
effects.
Following the core model specification, we have new lines to define the direct, specific indirect,
total indirect, and total effects. This is done using the “define” operator :=. For instance, ind_s
:= c_s*s_n*n_p requests that lavvan compute the quantity ind_s as the product of the three
paths previously labeled c_s, s_n, and n_p.
If we were to fit this model using the default options, lavaan would generate point estimates
for all of the newly defined effects as well as delta-method standard errors. As we noted in
lecture, however, the current best practices is to conduct inferential tests of indirect effects
using bootstrapped confidence intervals. Fortunately, it is straightforward to compute these
using lavaan.

Copyright © 2019 Curran-Bauer Analytics, LLC

3-18 | Chapter 3 Path Analysis: Part II

The script for computing bootrapped confidence intervals is shown below:

set.seed(62973)
fit.5b <- sem(effects.5, data=afdp, meanstructure=TRUE, se="bootstrap",
bootstrap=1000)
parameterEstimates(fit.5b, boot.ci.type = "perc")
parameterEstimates(fit.5b, boot.ci.type = "bca.simple")

Note that we have used the set.seed function to set the random number seed for selecting
the bootstrapped samples so that we can reproduce exactly these same results each time we
run the script.
Then, in the sem function, we specified se="bootstrap", bootstrap=1000 to request that
lavaan compute bootstrapped standard errors with 1000 bootstrapped samples. We actually
don’t care much about the standard errors. What we really want are bootstrapped confidence
intervals, which are generated in the parameterEstimates function with the boot.ci.type
argument. The parameterEstimates function produces a simple list of the parameter
estimates from the model. By including the boot.ci.type argument, we request that this
output include 95% boostrapped confidence intervals.
Note that there exist multiple methods for calculating bootrapped CIs. The percentile method
is probably easiest to understand – the 1000 boostrapped estimates are ordered by magnitude
and the 25th estimate (2.5th percentile) and 975th estimate (97.5th percentile) are taken as the
confidence limits, yielding a 95% confidence interval. This method, which is widely available in
many software programs, is obtained via boot.ci.type = "perc". That is the method we
used in the lecture notes. A slightly preferable approach, however, is to use bias-corrected
bootstrapped confidence intervals, and these are readily available in lavaan. To obtain the bias-
corrected confidence intervals, we simply use boot.ci.type = "bca.simple". The bias-
corrected intervals are shown in the output below. In terms of null hypothesis testing, the
same conclusions are generated with the bias-corrected CIs as with the percentile CIs presented
in lecture, but there are some slight differences in the specific values of the confidence limits.
lhs op rhs label est se z pvalue ci.lower ci.upper
1 stress ~ coa c_s 0.451 0.071 6.385 0.000 0.319 0.597
2 stress ~ gen -0.016 0.075 -0.209 0.835 -0.163 0.126
3 stress ~ age 0.002 0.026 0.075 0.940 -0.050 0.055
4 emotion ~ coa c_e 0.110 0.058 1.908 0.056 -0.004 0.226
5 emotion ~ gen -0.048 0.056 -0.843 0.399 -0.167 0.062
6 emotion ~ age -0.027 0.022 -1.215 0.225 -0.072 0.018
7 negaff ~ stress s_n 0.243 0.091 2.674 0.007 0.080 0.430
8 negaff ~ emotion e_n 0.582 0.109 5.324 0.000 0.354 0.787
9 negaff ~ age 0.119 0.030 3.891 0.000 0.062 0.180
10 peer ~ negaff n_p 0.137 0.030 4.526 0.000 0.077 0.196
11 peer ~ age 0.138 0.016 8.636 0.000 0.108 0.169
12 peer ~ coa c_p 0.185 0.052 3.529 0.000 0.069 0.280
13 stress ~~ emotion 0.112 0.019 6.006 0.000 0.076 0.150
14 stress ~~ stress 0.412 0.041 10.132 0.000 0.343 0.507
15 emotion ~~ emotion 0.247 0.017 14.356 0.000 0.217 0.286
16 negaff ~~ negaff 0.749 0.066 11.352 0.000 0.637 0.896
17 peer ~~ peer 0.216 0.029 7.426 0.000 0.162 0.277
18 coa ~~ coa 0.249 0.000 NA NA 0.249 0.249
19 coa ~~ gen 0.002 0.000 NA NA 0.002 0.002
20 coa ~~ age -0.055 0.000 NA NA -0.055 -0.055
21 gen ~~ gen 0.249 0.000 NA NA 0.249 0.249

Copyright © 2019 Curran-Bauer Analytics, LLC

 Path Analysis with Deviant Peer Affiliation Data | 3-19

22 gen ~~ age -0.070 0.000 NA NA -0.070 -0.070

23 age ~~ age 2.095 0.000 NA NA 2.095 2.095
24 stress ~1 0.687 0.339 2.028 0.043 0.016 1.346
25 emotion ~1 2.341 0.291 8.033 0.000 1.777 2.939
26 negaff ~1 -0.038 0.449 -0.085 0.932 -0.950 0.880
27 peer ~1 -1.856 0.195 -9.514 0.000 -2.236 -1.474
28 coa ~1 0.525 0.000 NA NA 0.525 0.525
29 gen ~1 0.538 0.000 NA NA 0.538 0.538
30 age ~1 12.718 0.000 NA NA 12.718 12.718
31 dir := c_p dir 0.185 0.052 3.527 0.000 0.069 0.280
32 ind_s := c_s*s_n*n_p ind_s 0.015 0.007 2.024 0.043 0.005 0.036
33 ind_e := c_e*e_n*n_p ind_e 0.009 0.005 1.647 0.100 0.000 0.022
34 tot_ind := c_s*s_n*n_p+c_e*e_n*n_p tot_ind 0.024 0.010 2.477 0.013 0.009 0.047
35 tot := c_s*s_n*n_p+c_e*e_n*n_p+c_p tot 0.209 0.054 3.900 0.000 0.094 0.306

Here, we are principally concerned with the estimates and bias-adjusted bootstrapped
confidence intervals given in lines 31-35.
The total effect of coa on peer is equal to .209 and represents a combination of the direct
effect (.185) and the total indirect effect (.024). The 95% CI is equal to .094 and .306; because
this does not contain zero, the total effect is deemed to be significant.
Examining the mediational pathways, we see that the specific indirect effect
coa→stress→negaff→peer, labeled ind_s to indicate it is the indirect effect through
stress, is equal to .015 (95% CI=.005, .036) and is significant (does not include zero). The
biological pathway from coa→emotion→negaff→peer, labeled ind_e to indicate it is the
indirect effect through emotion, is equal to .009 (95% CI=0, .022) and thus does not reach
statistical significance (because the lower CI is equal to zero).
In sum, examining the specific indirect mediational pathways has provided a more nuanced
understanding of how parental alcoholism is related to children’s affiliation with deviant peers.

Copyright © 2019 Curran-Bauer Analytics, LLC

3-20 | Chapter 3 Path Analysis: Part II

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 4
Confirmatory Factor
Analysis
Confirmatory Factor analysis of Holzinger-Swineford Data..................................................... 4-3
Preliminary Steps .................................................................................................................. 4-3
Initial Model with Standardized Factors............................................................................... 4-4
Initial Model with Scaling Items ........................................................................................... 4-9
Model Modification ............................................................................................................ 4-13
4-2 | Chapter 4 Confirmatory Factor Analysis

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-3

Confirmatory Factor analysis of Holzinger-Swineford Data

The data for this demonstration were provided by Holzinger & Swineford in their 1939
monograph A Study in Factor Analysis: The Stability of a Bi-Factor Solution. The sample includes
301 7th and 8th grade students, between 11-16 years of age, drawn from two schools. The
data is in the text file hs.dat and the accompanying R-script file is ch04.R. The variables in
the data set that we will use are
visperc visual perception test in which participants select the next image in a series
cubes visual perception test in which participants must mentally rotate a cube
lozenges visual perception test involving mental “flipping” of a parallelogram (“lozenge”)
parcomp paragraph comprehension test
sencomp sentence completion task in which participants select most appropriate word to
put at the end of a sentence
wordmean verbal ability test in which participants must select a word most similar in
meaning to a word used in a sentence.
addition participants have 2 minutes to complete as many 2-number addition problems
as they can
countdot participants have 4 minutes to count the number of dots in each of a series of
dot pictures
sccaps participants have 3 minutes to indicate whether capital letters are composed
entirely of straight lines or include curved lines.
Other variables in the data not included in the models fit here are school, female, age, and
month.
Preliminary Steps
We begin by reading in the data and assigning variable names, as follows (where the working
directory has been set to the source file directory as described in Chapter 1):
hs <- read.table("hs.dat", header=FALSE)
names(hs) <- c("school", "female", "age", "month",
"visperc", "cubes", "lozenges",
"parcomp", "sencomp", "wordmean",
"addition", "countdot", "sccaps")

Next, we rescaled the variables addition, countdot, and sccaps by dividing by four:
hs$addition <- hs$addition/4
hs$countdot <- hs$countdot/4
hs$sccaps <- hs$sccaps/4

This was done to facilitate model estimation, as numerical instability problems can arise when
using variables on widely differing scales. Dividing these variables by four brings their standard
deviations closer to the standard deviations of the other observed variables.

Copyright © 2019 Curran-Bauer Analytics, LLC

4-4 | Chapter 4 Confirmatory Factor Analysis

Finally, we load the lavaan package that we will use to fit the CFA models:
library(lavaan)

Initial Model with Standardized Factors

The hypothesized model for the data includes three factors and is shown in the diagram below:

Visual Verbal Speed

visper cubes lozenges parcomp sencomp wordmean addition countdot sccaps

1 2 3 4 5 6 7 8 9

The model is of the form

 visperi    1   11 0 0   1i 
       
 cubesi   2   21 0 0    2i 
 lozengesi   3   31 0 0    3i 
       
 parcompi   4   0 42 0   visuali    4i 
 sencompi    5    0 52 0   verbali     5i 
         

 wordmeani   6   0 62 0  speed i    6i 
 addition     0 0 73   
 i
  7    7i 
 countdoti   8   0 0 83    8i 
 sccaps     0 93   
 i   9  0  9i 
where
COV  εi   Θ  DIAG 11 ,22 ,33 ,44 ,55 ,66 ,77 ,88 ,99 

 1   11 
   
E  ηi   α    2  ; COV  ηi   Ψ   21  22 
     
  3  31 32 33 

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-5

Recall that to identify the model we must set the scale of the latent variables. Two options for
doing so are to (1) standardize the latent variables or (2) set the intercept and factor loading of
one item per factor to zero and one, respectively. We shall begin with the standardized scaling
by setting these parameters as fixed:
0  1 
   
E  ηi   α   0  ; COV  ηi   Ψ   21 1 
0  1 
   31  32
The R script to specify and fit this model is shown below:
cfa.1a <-'#factor loadings
visual =~ visperc + cubes + lozenges
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps
'

fit.1a <- cfa(cfa.1a, data=hs, meanstructure=TRUE, std.lv=TRUE)

summary(fit.1a, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

We have seen most of this syntax before, so here we will highlight only portions of the syntax,
especially new syntax that involves the use of latent variables.
First, we will use the cfa function in lavaan to fit the model. This function imposes a number of
defaults consistent with how CFA models are typically specified and enables us to specify
models with compact model syntax objects. The default scaling for the latent variables is a
hybrid of the approaches we described in lecture. By default, the latent variables will have
means set to zero but freely estimated variance. Additionally, the factor loading of the first
indicator for each latent variable will be set to one but the intercept for this indicator will be
freely estimated. We will override these defaults to implement the scaling options described in
lecture.
Second, in defining the model syntax object, we implement the =~ operator, which is used to
define latent variables (left hand side) and to specify the variables that load on them (right
hand side). Thus, visual =~ visperc + cubes + lozenges defines a new latent variable,
visual, and indicates that visperc, cubes, and lozenges are indicator variables.
Third, in fitting the model, we specify std.lv=TRUE within the cfa function call. This overrides
the default hybrid scaling and instead asks lavaan to (1) standardize the factors and (2) freely
estimate all loadings.
Let us now turn to the output, considering first the fit indices for the model:

Copyright © 2019 Curran-Bauer Analytics, LLC

4-6 | Chapter 4 Confirmatory Factor Analysis

lavaan 0.6-2 ended normally after 28 iterations

Optimization method NLMINB

Number of free parameters 30
Number of observations 301
Estimator ML
Model Fit Test Statistic 85.306
Degrees of freedom 24
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 918.852

Degrees of freedom 36
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.931

Tucker-Lewis Index (TLI) 0.896

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8326.241

Loglikelihood unrestricted model (H1) -8283.589

Number of free parameters 30

Akaike (AIC) 16712.483
Bayesian (BIC) 16823.696
Sample-size adjusted Bayesian (BIC) 16728.553

Root Mean Square Error of Approximation:

RMSEA 0.092
90 Percent Confidence Interval 0.071 0.114
P-value RMSEA <= 0.05 0.001

Standardized Root Mean Square Residual:

SRMR 0.060

We see here that the fit indices indicate that the model does not fit the data particularly well.
The estimates (both raw and fully standardized) for the model are shown next. These values
must be interpreted cautiously, given the lack of fit of the model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-7

Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc 5.398 0.485 11.128 0.000 5.398 0.772
cubes 1.992 0.310 6.429 0.000 1.992 0.424
lozenges 5.249 0.595 8.817 0.000 5.249 0.581
verbal =~
parcomp 2.969 0.170 17.474 0.000 2.969 0.852
sencomp 4.406 0.251 17.576 0.000 4.406 0.855
wordmean 6.416 0.376 17.082 0.000 6.416 0.838
speed =~
addition 3.562 0.400 8.903 0.000 3.562 0.570
countdot 3.655 0.330 11.090 0.000 3.655 0.723
sccaps 6.030 0.585 10.305 0.000 6.030 0.665

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual ~~
verbal 0.459 0.064 7.189 0.000 0.459 0.459
speed 0.471 0.073 6.461 0.000 0.471 0.471
verbal ~~
speed 0.283 0.069 4.117 0.000 0.283 0.283

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 29.615 0.403 73.473 0.000 29.615 4.235
.cubes 24.352 0.271 89.855 0.000 24.352 5.179
.lozenges 18.003 0.521 34.579 0.000 18.003 1.993
.parcomp 9.183 0.201 45.694 0.000 9.183 2.634
.sencomp 17.362 0.297 58.452 0.000 17.362 3.369
.wordmean 15.299 0.441 34.667 0.000 15.299 1.998
.addition 24.069 0.360 66.766 0.000 24.069 3.848
.countdot 27.635 0.291 94.854 0.000 27.635 5.467
.sccaps 48.367 0.523 92.546 0.000 48.367 5.334
visual 0.000 0.000 0.000
verbal 0.000 0.000 0.000
speed 0.000 0.000 0.000

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 19.766 4.090 4.833 0.000 19.766 0.404
.cubes 18.141 1.628 11.146 0.000 18.141 0.821
.lozenges 54.037 5.800 9.317 0.000 54.037 0.662
.parcomp 3.341 0.429 7.779 0.000 3.341 0.275
.sencomp 7.140 0.934 7.642 0.000 7.140 0.269
.wordmean 17.454 2.109 8.277 0.000 17.454 0.298
.addition 26.430 2.691 9.823 0.000 26.430 0.676
.countdot 12.192 1.855 6.573 0.000 12.192 0.477
.sccaps 45.857 5.730 8.003 0.000 45.857 0.558

Copyright © 2019 Curran-Bauer Analytics, LLC

4-8 | Chapter 4 Confirmatory Factor Analysis

visual 1.000 1.000 1.000

verbal 1.000 1.000 1.000
speed 1.000 1.000 1.000

Note that the standard error, z-statistic, and p-value for the factor means and variances are all
missing. This simply indicates that these parameters were not estimated, and hence no
inferential tests were conducted on their values.
Because the items are on different scales, we cannot easily interpret the differences among the
raw estimates for the factor loadings or residual variances. To better interpret these results,
we requested the standardized solution. Note that because we have already standardized the
latent variables, the column labeled Std.lv is redundant with the raw estimates. We are more
interested in the fully-standardized Std.all values. The fully-standardized factor loadings can
be compared directly (e.g., visperc is a better indicator than cubes for the visual factor).
One odd thing to note is that the indicator intercepts are not zero in the standardized output,
as you would expect if the observed variables had been standardized in the usual way of
deviating the mean and dividing by the standard deviation. In lavaan, the standardized solution
merely rescales the observed variables to have standard deviations of one, and does not center
the variables to have means of zero. This is of no consequence here, as the mean structure is
saturated and of little interest.
Last, the standardized residual variances of the indicators are interpretable as the proportion of
variance unexplained by the latent factors (or 1 – communality). The communalities
themselves were obtained by setting the rsquare argument of the summary function to TRUE:
R-Square:
Estimate
visperc 0.596
cubes 0.179
lozenges 0.338
parcomp 0.725
sencomp 0.731
wordmean 0.702
addition 0.324
countdot 0.523
sccaps 0.442

Items with high communality are generally regarded as better items. Given the poor model fit,
we may wish to examine modification indices to get an idea of where the model may be
misspecified.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-9

The modification indices for the model are obtained from the modindices function as follows:
modindices(fit.1a, sort.=TRUE, minimum.value=10)

lhs op rhs mi epc sepc.lv sepc.all sepc.nox

42 visual =~ sccaps 36.411 4.672 4.672 0.515 0.515
88 addition ~~ countdot 34.145 15.423 15.423 0.859 0.859
40 visual =~ addition 18.631 -2.182 -2.182 -0.349 -0.349
90 countdot ~~ sccaps 14.946 -19.039 -19.039 -0.805 -0.805

Here we see the largest modification indices are associated with a cross-loading for sccaps on
visual, and a correlated uniqueness for countdot with addition. It is important to keep in
mind that both modification indices may be related to the same misspecification.
Before proceeding to respecify the model, let us also consider how the model would be input
into lavaan if we chose to scale the latent variables using scaling items rather than
standardizing.

Initial Model with Scaling Items

We shall choose the first indicator for each factor to be the scaling item. The intercept and
factor loading for each scaling item is set to zero and one, respectively. This scaling option
permits the means and variances of the latent factors to be estimated. The model is thus now
specified as
 visperi   0   1 0 0   1i 
       
 cubesi   2   21 0 0    2i 
 lozengesi   3   31 0 0    3i 
       
 parcompi   0   0 1 0   visuali    4i 
 sencompi    5    0 52 0   verbali     5i 
         

 wordmeani   6   0 62 0   speedi    6i 
 addition   0   0 0 1   
 i
      7i 
 countdoti   8   0 0 83    8i 
 sccaps     0 93   
 i   9  0  9i 
where
COV  εi   Θ  DIAG 11 ,22 ,33 ,44 ,55 ,66 ,77 ,88 ,99 

 1   11 
   
E  ηi   α    2  ; COV  ηi   Ψ   21  22 
     
 3  31 32 33 

and all elements in  and  are now estimated.

Copyright © 2019 Curran-Bauer Analytics, LLC

4-10 | Chapter 4 Confirmatory Factor Analysis

The corresponding R script for fitting the model and displaying the results is given here:
cfa.1b <-'#factor loadings
visual =~ visperc + cubes + lozenges
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps

#means/intercepts
visual ~ 1
verbal ~ 1
speed ~ 1
visperc ~ 0*1
parcomp ~ 0*1
addition ~ 0*1
'

fit.1b <- cfa(cfa.1b, data=hs, meanstructure=TRUE)

summary(fit.1b, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

In this case, we will not invoke the std.lv option when fitting the model as we do not wish to
standardize the factors. Here, we wish to use a scaling indicator instead. We can rely on the
default that the factor loading for the first indicator per factor will be fixed to one (this need
not be specified explicitly). We must, however, assign a numeric value of zero to the intercept
for each of the anchor items, e.g., via visperc ~ 0*1, to override the default of freely
estimating intercepts for all observed variables.
Similarly, we can rely on the default that the factor variances will be freely estimated but we
need to indicate that lavvan should estimate the factor means, e.g., via visual ~ 1. (Since
there are no predictors of the factors, their intercepts are interpretable as means).
The model fit output is shown here:
lavaan 0.6-2 ended normally after 180 iterations

Optimization method NLMINB

Number of free parameters 30

Number of observations 301

Estimator ML
Model Fit Test Statistic 85.306
Degrees of freedom 24
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 918.852

Degrees of freedom 36
P-value 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-11

User model versus baseline model:

Comparative Fit Index (CFI) 0.931

Tucker-Lewis Index (TLI) 0.896

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8326.241

Loglikelihood unrestricted model (H1) -8283.589

Number of free parameters 30

Akaike (AIC) 16712.483
Bayesian (BIC) 16823.696
Sample-size adjusted Bayesian (BIC) 16728.553

Root Mean Square Error of Approximation:

RMSEA 0.092
90 Percent Confidence Interval 0.071 0.114
P-value RMSEA <= 0.05 0.001

Standardized Root Mean Square Residual:

SRMR 0.060

Note that the tests of model fit are identical to the standardized factor model presented
previously. The two models are equivalent, and merely scaled differently.
The difference in scales is apparent when considering the model estimates:
Parameter Estimates:

Information Expected
Information saturated (h1) model Structured
Standard Errors Standard

Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc 1.000 5.398 0.772
cubes 0.369 0.066 5.554 0.000 1.992 0.424
lozenges 0.972 0.145 6.685 0.000 5.249 0.581
verbal =~
parcomp 1.000 2.969 0.852
sencomp 1.484 0.087 17.014 0.000 4.406 0.855
wordmean 2.161 0.129 16.703 0.000 6.416 0.838
speed =~
addition 1.000 3.562 0.570
countdot 1.026 0.143 7.152 0.000 3.655 0.723
sccaps 1.693 0.237 7.155 0.000 6.030 0.665

Copyright © 2019 Curran-Bauer Analytics, LLC

4-12 | Chapter 4 Confirmatory Factor Analysis

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual ~~
verbal 7.348 1.323 5.552 0.000 0.459 0.459
speed 9.047 1.942 4.660 0.000 0.471 0.471
verbal ~~
speed 2.993 0.851 3.518 0.000 0.283 0.283

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual 29.615 0.403 73.473 0.000 5.487 5.487
verbal 9.183 0.201 45.694 0.000 3.093 3.093
speed 24.069 0.360 66.766 0.000 6.757 6.757
.visperc 0.000 0.000 0.000
.parcomp 0.000 0.000 0.000
.addition 0.000 0.000 0.000
.cubes 13.424 1.985 6.762 0.000 13.424 2.855
.lozenges -10.797 4.336 -2.490 0.013 -10.797 -1.195
.sencomp 3.734 0.831 4.496 0.000 3.734 0.725
.wordmean -4.545 1.233 -3.685 0.000 -4.545 -0.594
.countdot 2.940 3.472 0.847 0.397 2.940 0.582
.sccaps 7.622 5.730 1.330 0.183 7.622 0.841

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 19.766 4.090 4.833 0.000 19.766 0.404
.cubes 18.141 1.628 11.146 0.000 18.141 0.821
.lozenges 54.037 5.800 9.317 0.000 54.037 0.662
.parcomp 3.341 0.429 7.779 0.000 3.341 0.275
.sencomp 7.140 0.934 7.642 0.000 7.140 0.269
.wordmean 17.454 2.109 8.277 0.000 17.454 0.298
.addition 26.430 2.691 9.823 0.000 26.430 0.676
.countdot 12.192 1.855 6.573 0.000 12.192 0.477
.sccaps 45.857 5.730 8.003 0.000 45.857 0.558
visual 29.135 5.237 5.564 0.000 1.000 1.000
verbal 8.815 1.009 8.737 0.000 1.000 1.000
speed 12.688 2.850 4.451 0.000 1.000 1.000

R-Square:
Estimate
visperc 0.596
cubes 0.179
lozenges 0.338
parcomp 0.725
sencomp 0.731
wordmean 0.702
addition 0.324
countdot 0.523
sccaps 0.442

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-13

The raw estimates for the factor loadings and intercepts are now interpreted in a relative scale,
with reference to the scaling item. Further, because the factor variances are no longer set to
one, the covariances among the factors can no longer be interpreted as correlations.
Again, the difference in metric across the tests makes interpretation of differences in intercepts
and loadings somewhat difficult. The standardized solution can again be considered to aid in
interpretation. Note that it is identical to the standardized solution shown previously. Similarly,
modification indices are no different with this scaling option and are not repeated here.
We now consider respecification of the model, returning to the standardized scaling option to
set the metric of the latent variables.

Model Modification
We first introduce a cross loading of sccaps on visual, as shown in the diagram below:

Visual Verbal Speed

visper cubes lozenges parcomp sencomp wordmean addition countdot sccaps

1 2 3 4 5 6 7 8 9

The model is now of the form

 visperi    1   11 0 0   1i 
       
 cubesi   2   21 0 0    2i 
 lozengesi   3   31 0 0    3i 
       
 parcompi   4   0 42 0   visuali    4i 
 sencompi    5    0 52 0   verbali     5i 
         

 wordmeani   6   0 62 0  speed i    6i 
 addition     0 0 73   
 i
  7    7i 
 countdoti   8   0 0 83    8i 
 sccaps      
93   
 i   9   91 0  9i 

Copyright © 2019 Curran-Bauer Analytics, LLC

4-14 | Chapter 4 Confirmatory Factor Analysis

where
COV  εi   Θ  DIAG 11 ,22 ,33 ,44 ,55 ,66 ,77 ,88 ,99 

0  1 
   
E  ηi   α   0  ; COV  ηi   Ψ   21 1 
0  1 
   31  32
The difference between the modified model and the original CFA is that the element in 9th row
(corresponding to sccaps) and 1st column (corresponding to visual) of the factor loading
matrix has been freed from zero to 91 .
The new cross-loading can be estimated by defining and fitting a slightly modified model syntax
object, cfa.2, as follows:
cfa.2 <- '#factor loadings
visual =~ visperc + cubes + lozenges + sccaps
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps
'
fit.2 <- cfa(cfa.2, data=hs, meanstructure=TRUE, std.lv=TRUE)
summary(fit.2, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

Notice that sccaps now appears twice: once on speed and once on visual. Factor loadings
for sccaps are freely estimated for both factors.
The model fit information is shown here:
lavaan 0.6-2 ended normally after 29 iterations

Optimization method NLMINB

Number of free parameters 31

Number of observations 301

Estimator ML
Model Fit Test Statistic 52.382
Degrees of freedom 23
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 918.852

Degrees of freedom 36
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.967

Tucker-Lewis Index (TLI) 0.948

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-15

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8309.780

Loglikelihood unrestricted model (H1) -8283.589

Number of free parameters 31

Akaike (AIC) 16681.560
Bayesian (BIC) 16796.480
Sample-size adjusted Bayesian (BIC) 16698.166

Root Mean Square Error of Approximation:

RMSEA 0.065
90 Percent Confidence Interval 0.042 0.089
P-value RMSEA <= 0.05 0.133

Standardized Root Mean Square Residual:

SRMR 0.041

Freeing the cross-loading significantly improved the model fit according to a likelihood ratio test
of the original CFA and the model estimated above:
lavTestLRT(fit.2, fit.1a)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.2 23 16682 16797 52.382
fit.1a 24 16713 16824 85.305 32.923 1 9.586e-09 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Thus we have:
 2  dfOriginal  dfCrossLoad    2Original   2CrossLoad
 2  24  23  85.305  52.382
 2 1  32.923, p  .001
Other fit indices suggest that the modified model may have satisfactory fit to the data.
The estimates for the model are shown next. To the extent that we are justified in including the
new cross loading, we can have more faith in these estimates due to improved model fit.
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc 5.308 0.454 11.685 0.000 5.308 0.759
cubes 2.046 0.305 6.698 0.000 2.046 0.435
lozenges 5.333 0.577 9.241 0.000 5.333 0.590
sccaps 3.480 0.577 6.034 0.000 3.480 0.384

Copyright © 2019 Curran-Bauer Analytics, LLC

4-16 | Chapter 4 Confirmatory Factor Analysis

verbal =~
parcomp 2.967 0.170 17.455 0.000 2.967 0.851
sencomp 4.411 0.251 17.602 0.000 4.411 0.856
wordmean 6.414 0.376 17.073 0.000 6.414 0.838
speed =~
addition 3.830 0.419 9.134 0.000 3.830 0.612
countdot 4.021 0.368 10.934 0.000 4.021 0.795
sccaps 4.049 0.593 6.824 0.000 4.049 0.447

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual ~~
verbal 0.453 0.062 7.242 0.000 0.453 0.453
speed 0.301 0.080 3.763 0.000 0.301 0.301
verbal ~~
speed 0.206 0.070 2.937 0.003 0.206 0.206

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 29.615 0.403 73.473 0.000 29.615 4.235
.cubes 24.352 0.271 89.855 0.000 24.352 5.179
.lozenges 18.003 0.521 34.579 0.000 18.003 1.993
.sccaps 48.367 0.523 92.546 0.000 48.367 5.334
.parcomp 9.183 0.201 45.694 0.000 9.183 2.634
.sencomp 17.362 0.297 58.452 0.000 17.362 3.369
.wordmean 15.299 0.441 34.667 0.000 15.299 1.998
.addition 24.069 0.360 66.766 0.000 24.069 3.848
.countdot 27.635 0.291 94.854 0.000 27.635 5.467
visual 0.000 0.000 0.000
verbal 0.000 0.000 0.000
speed 0.000 0.000 0.000

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 20.726 3.616 5.731 0.000 20.726 0.424
.cubes 17.924 1.607 11.153 0.000 17.924 0.811
.lozenges 53.148 5.586 9.515 0.000 53.148 0.651
.sccaps 45.234 4.845 9.336 0.000 45.234 0.550
.parcomp 3.353 0.430 7.800 0.000 3.353 0.276
.sencomp 7.098 0.934 7.602 0.000 7.098 0.267
.wordmean 17.483 2.110 8.285 0.000 17.483 0.298
.addition 24.450 2.845 8.595 0.000 24.450 0.625
.countdot 9.383 2.362 3.973 0.000 9.383 0.367
visual 1.000 1.000 1.000
verbal 1.000 1.000 1.000
speed 1.000 1.000 1.000

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-17

R-Square:
Estimate
visperc 0.576
cubes 0.189
lozenges 0.349
sccaps 0.450
parcomp 0.724
sencomp 0.733
wordmean 0.702
addition 0.375
countdot 0.633

In the results, note that sccaps loads significantly on visual. The residual variance of
sccaps remains fairly high and the correlation between visual and speed is still statistically
significant. The new modification indices exceeding 10 are reported below:
modindices(fit.2, sort.=TRUE, minimum.value=10)

[1] lhs op rhs mi epc sepc.lv sepc.all sepc.nox

<0 rows> (or 0-length row.names)

Note that no further changes are suggested for the model.

Next, we illustrate how implementing a different model modification might affect conclusions
drawn from the model. We allow the residual errors for addition and countdot to
correlate, as shown in the diagram below:

Visual Verbal Speed

visper cubes lozenges parcomp sencomp wordmean addition countdot sccaps

1 2 3 4 5 6 7 8 9

Copyright © 2019 Curran-Bauer Analytics, LLC

4-18 | Chapter 4 Confirmatory Factor Analysis

The model is now of the form

 visperi    1   11 0 0   1i 
       
 cubesi   2   21 0 0    2i 
 lozengesi   3   31 0 0    3i 
       
 parcompi   4   0 42 0   visuali    4i 
 sencompi    5    0 52 0   verbali     5i 
         

 wordmeani   6   0 62 0  speed i    6i 
 addition     0 0 73   
 i
  7    7i 
 countdoti   8   0 0 83    8i 
 sccaps     0 
93   
 i   9  0  9i 
where
 11 
 
 0  22 
 0 0 33 
 
 0 0 0  44 
COV  εi   Θ   0 0 0 0 55 
 
 0 0 0 0 0  66 
 0 0 0 0 0 0  77 
 
 0 0 0 0 0 0 87 88 
 0 0 99 
 0 0 0 0 0 0

0  1 
   
E  ηi   α   0  ; COV  ηi   Ψ   21 1 
0  1 
   31  32
The difference between the modified model and the original CFA is that Θ is no longer
diagonal; it contains a covariance between the residual terms for addition and countdot.
This change can be included in the model by defining and fitting a new model syntax object,
cfa.3, as follows:
cfa.3 <- '#factor loadings
visual =~ visperc + cubes + lozenges
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps

#covariances
addition ~~ countdot
'
fit.3 <- cfa(cfa.3, data=hs, meanstructure=TRUE, std.lv=TRUE)
summary(fit.3, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-19

The difference between this input and the original input is the final statement: addition ~~
countdot. In lavaan, when observed variable names are joined by the ~~ operator, it indicates
that the residuals associated with those variables should covary (i.e., be correlated).
We first consider the model fit:
lavaan 0.6-2 ended normally after 51 iterations

Optimization method NLMINB

Number of free parameters 31

Number of observations 301

Estimator ML
Model Fit Test Statistic 53.272
Degrees of freedom 23
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 918.852

Degrees of freedom 36
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.966

Tucker-Lewis Index (TLI) 0.946

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8310.225

Loglikelihood unrestricted model (H1) -8283.589

Number of free parameters 31

Akaike (AIC) 16682.450
Bayesian (BIC) 16797.370
Sample-size adjusted Bayesian (BIC) 16699.056

Root Mean Square Error of Approximation:

RMSEA 0.066
90 Percent Confidence Interval 0.043 0.090
P-value RMSEA <= 0.05 0.118

Standardized Root Mean Square Residual:

SRMR 0.043

Copyright © 2019 Curran-Bauer Analytics, LLC

4-20 | Chapter 4 Confirmatory Factor Analysis

Allowing a single residual correlation among these items resulted in a large improvement in
model fit, which we can evaluate via a likelihood ratio test (i.e., chi-square difference test):
lavTestLRT(fit.3, fit.1a)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.3 23 16682 16797 53.272
fit.1a 24 16713 16824 85.305 32.033 1 1.516e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

The improvement in model fit is highly significant:

 2  dfOriginal  df ResCorr    2Original   2 ResCorr

 2  24  23  85.305  53.272
 2 1  32.033, p  .001
Other fit indices suggest that the modified model may also have satisfactory overall fit to the
data.
The estimates for the model are shown next. To the extent that we are justified in including the
unhypothesized cross loading, we can have more faith in these estimates due to improved
model fit.
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc 5.308 0.457 11.620 0.000 5.308 0.759
cubes 2.037 0.305 6.673 0.000 2.037 0.433
lozenges 5.323 0.576 9.238 0.000 5.323 0.589
verbal =~
parcomp 2.967 0.170 17.458 0.000 2.967 0.851
sencomp 4.411 0.251 17.601 0.000 4.411 0.856
wordmean 6.413 0.376 17.068 0.000 6.413 0.838
speed =~
addition 2.202 0.421 5.229 0.000 2.202 0.352
countdot 2.383 0.366 6.505 0.000 2.383 0.471
sccaps 8.667 0.958 9.051 0.000 8.667 0.956

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.addition ~~
.countdot 10.140 1.936 5.239 0.000 10.140 0.389
visual ~~
verbal 0.457 0.064 7.142 0.000 0.457 0.457
speed 0.544 0.078 6.965 0.000 0.544 0.544
verbal ~~
speed 0.270 0.065 4.141 0.000 0.270 0.270

Copyright © 2019 Curran-Bauer Analytics, LLC

 Confirmatory Factor analysis of Holzinger-Swineford Data | 4-21

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 29.615 0.403 73.473 0.000 29.615 4.235
.cubes 24.352 0.271 89.855 0.000 24.352 5.179
.lozenges 18.003 0.521 34.579 0.000 18.003 1.993
.parcomp 9.183 0.201 45.694 0.000 9.183 2.634
.sencomp 17.362 0.297 58.452 0.000 17.362 3.369
.wordmean 15.299 0.441 34.667 0.000 15.299 1.998
.addition 24.069 0.360 66.766 0.000 24.069 3.848
.countdot 27.635 0.291 94.854 0.000 27.635 5.467
.sccaps 48.367 0.523 92.546 0.000 48.367 5.334
visual 0.000 0.000 0.000
verbal 0.000 0.000 0.000
speed 0.000 0.000 0.000

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.visperc 20.730 3.651 5.678 0.000 20.730 0.424
.cubes 17.960 1.608 11.171 0.000 17.960 0.812
.lozenges 53.255 5.576 9.552 0.000 53.255 0.653
.parcomp 3.350 0.430 7.791 0.000 3.350 0.276
.sencomp 7.099 0.934 7.600 0.000 7.099 0.267
.wordmean 17.496 2.111 8.287 0.000 17.496 0.298
.addition 34.268 2.980 11.501 0.000 34.268 0.876
.countdot 19.870 1.989 9.988 0.000 19.870 0.778
.sccaps 7.090 15.210 0.466 0.641 7.090 0.086
visual 1.000 1.000 1.000
verbal 1.000 1.000 1.000
speed 1.000 1.000 1.000

R-Square:
Estimate
visperc 0.576
cubes 0.188
lozenges 0.347
parcomp 0.724
sencomp 0.733
wordmean 0.702
addition 0.124
countdot 0.222
sccaps 0.914

Note that the newly freed parameter, addition ~~ countdot is statistically significantly
different from zero. The standardized estimate indicates that addition and countdots are
moderately correlated above and beyond the correlation implied by the common speed factor
(r = .389). Note further the substantial reduction in residual variance estimates for items
loading on speed. Specifically, even though sccaps is not directly involved in the residual
correlation between addition and countdot, its residual variance was substantially reduced.
By accounting for the local dependence between addition and countdot with the residual
covariance, speed was able to account for more of the variance in sccaps. This is further

Copyright © 2019 Curran-Bauer Analytics, LLC

4-22 | Chapter 4 Confirmatory Factor Analysis

evidenced by the fact that sccaps is now the highest loading item on the factor. It now
appears that sccaps is almost definitional of the speed factor, with a communality of .914,
whereas the communalities of countdots and addition are much lower.
As with the previous example, introducing this single new parameter resulted in no further
sizeable modification indices.
modindices(fit.3, sort.=TRUE, minimum.value=10)

[1] lhs op rhs mi epc sepc.lv sepc.all sepc.nox

<0 rows> (or 0-length row.names)

This model is not nested with the alternative modified model; however, the two models
provide nearly equivalent fit (RMSEA=.065 versus .066, TLI=.948 versus .946, etc.). Thus, model
selection should be based on which modification is most plausible, and upon the
interpretability of parameter estimates.

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 5
Structural Equation Models
with Latent Variables
Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data ..... 5-3
Preliminary Steps .................................................................................................................. 5-4
Initial Hypothesized Model................................................................................................... 5-5
Confirmatory Factor Analysis ............................................................................................... 5-9
Revised Model .................................................................................................................... 5-12
Examining Direct, Indirect and Total Effects ...................................................................... 5-18
5-2 | Chapter 5 Structural Equation Models with Latent Variables

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-3

Structural Equation Modeling of Şenol-Durak and Ayvaşik’s

Posttraumatic Growth Data
The data for this demonstration were provided by Şenol-Durak & Ayvaşik in their 2010 Journal
of Health Psychology manuscript, “Factors associated with posttraumatic growth among the
spouses of myocardial infraction patients.” The sample includes 132 spouses of myocardial
infarction patients. The correlation matrix is in the text file mip.dat. We will enter the means
and standard deviations into R directly. The R script for this chapter is contained in the file
ch05.R.
The variables in the data set that we will use are
fa social support from family
fr social support from friends Environmental Resources
si social support from significant others
lo Locus of Control Scale score
comt Commitment score
con Control score Individual Resources

cha Challenge score

es Rosenberg Self Esteem Scale score
pro subjective evaluation of prognosis
th threat to future health Event Related Factors

time time since diagnosis

em emotion focused coping
ind indirect coping (labeled in in diagrams)
ru rumination
Cognitive Process Coping
av avoidance
hy hypervigilance
rb religious beliefs
ptgi1 improved relationships
ptgi2 new possibilities for one’s life
ptgi3 greater appreciation of life Posttraumatic Growth
ptgi4 greater sense of personal strength
ptgi5 spiritual development

Refer to the article for definitions of variables not included in the model.

Copyright © 2019 Curran-Bauer Analytics, LLC

5-4 | Chapter 5 Structural Equation Models with Latent Variables

Preliminary Steps
This example illustrates how summary statistics can be used to fit structural equation models in
lavaan. To do so, we require the covariance matrix of the observed variables. For generality,
we will also bring in the mean vector, although it plays no part in the fit of this particular model
(because the model is saturated in the means). In this case, we have the correlation matrix and
thus must construct the covariance matrix from the correlations and standard deviations.
To begin, we load the lavaan package
library(lavaan)

We then define the variable names (this includes variables not used in the models fit here):
names <- c("ptgi","ptgi1", "ptgi2", "ptgi3", "ptgi4", "ptgi5",
"marital", "fa", "fr", "si", "child", "child18",
"age", "gender", "depres", "comt", "con", "cha",
"es", "lo", "pro", "th", "diord", "time", "problem",
"em", "ind", "ru", "av", "hy", "relipart", "rb")

As a side note, here we used the name ind for indirect coping rather than in because the
latter has a special meaning and thus cannot be used as a variable name.
Next, we enter the means and standard deviations reported in the manuscript (in the same
order as the variable names):
mip.mns <- c(59.18, 19.84, 10.53, 12.65, 9.80, 6.35,
9.53, 24.96, 19.89, 18.24, 2.68, .45,
52.04, .11, 9.37, 10.36, 10.35, 9.93,
30.33, 80.47, 2.67, 1.78, .47, 3.91, 70.58,
38.87, 24.34, 12.26, 11.80, 9.17, .81, 2.74)

mip.sds <- c(24.24, 9.00, 6.89, 5.22, 3.73, 3.05,

2.59, 3.87, 6.07, 7.48, 1.07, .81,
11.04, .318, 6.57, 2.65, 3.27, 2.45,
5.92, 18.12, .74, 1.14, .50, 8.05, 11.94,
11.41, 6.83, 7.51, 5.51, 5.94, 1.24, .92)

Finally, we read in the correlations from the file mip.dat using the read.table function and
convert the correlation matrix to a covariance matrix with the following commands:
mip.cor <- read.table("mip.dat", header=FALSE, row.names=names, col.names=names)
mip.cor <- data.matrix(mip.cor)
mip.cov <- cor2cov(mip.cor,sds=mip.sds)

The mip.dat file is a space-delimited text file that contains a full correlation matrix (diagonal
of 1’s as well as all elements above and below the diagonal). The read.table function creates
a data frame from this file called mip.cor with rows and columns labeled by the variable
names we input previously. We then convert mip.cor to a numeric matrix using the
data.matrix function. Finally, we use the cor2cov function of lavaan to convert the
correlation matrix into a covariance matrix, utilizing the information we provided on the
standard deviations in the vector mip.sds.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-5

Note that it is also possible to read in a lower triangular correlation matrix and convert it to a
full covariance matrix using the getCov function in lavaan. Here, however, the data was
originally entered as a full correlation matrix.
We now have the sufficient statistics, the covariance matrix mip.cov and the mean vector
mip.mns, ready for use with lavaan. We can thus proceed to specify and fit the hypothesized
structural equation model.

Initial Hypothesized Model

The hypothesized model for the data predicts that both individual and environmental resources
directly lead to increased posttraumatic growth, but also indirectly lead to somewhat
decreased posttraumatic growth by reducing event-related hardship, thus decreasing the need
for coping and reducing opportunities for posttraumatic growth. The hypothesized model also
predicts that neither environmental nor individual resources have a direct impact on cognitive
coping. Further, the effect of event-related factors on posttraumatic growth is hypothesized to
be purely mediated by coping.

9 10 11
1 2 3
pro th time
fa fr si

 ERF
Event-Related ptgi1 18
Factors  PTG
Environmental
Resources ptgi2 19

Posttraumatic
ptgi3 20
Growth
Individual
ptgi4 21
Resources
Coping  Cope
ptgi5 22

lo comt con cha es

em in ru av hy rb
4 5 6 7 8

12 13 14 15 16 17

We can also express the model using matrix algebra, as shown on the next page.

Copyright © 2019 Curran-Bauer Analytics, LLC

5-6 | Chapter 5 Structural Equation Models with Latent Variables

The measurement model is:

 fai    1i   11 0 0 0 0   1i 
 fr      0 0 0 0    
 i   2i   21  2i 
 sii    3i   31 0 0 0 0    3i 
       
 loi    4i   0 42 0 0 0    4i 
 comti    5i   0 52 0 0 0    5i 
       
 coni    6i   0 62 0 0 0    6i 
 cha      0 72 0 0 0   
 i
  7i     7i 
 esi    8i   0 82 0 0 0    8i 
 pro      0 93   
   ERi    9i 
0 0 0
 i   9i  
 thi   10i   0 0 10,3 0 0   10i 
        IRi   
 timei    11i    0 0 11,3 0 0  
  ERFi    11i 
 emi   12i   0 0 0 12,4 0    12i 
      COPEi   
 ini   13i   0 0 0 13,4 0 
   13i 
 ru     0 0 0 14,4 0   PTGi   
 i   14i     14i 
 avi   15i   0 0 0 15,4 0   15i 
 hy     0 0 0 16,4 0   
 i   16i     16i 
 rbi   17 i   0 0 0 17,4 0  17 i 
     18,5   
 ptgi1i   18i   0 0 0 0  18i 
 ptgi2i   19i   0 0 0 0 19,5  19i 
       
 ptgi3i   20i   0 0 0 0 20,5   20i 
 ptgi4i   21i   0 0 0 0 21,5   21i 
       
 ptgi5i   22i   0 0 0 0 22,5   22i 

where Θ  DIAG 11 ,22 , ,22,22 

The latent variable model is:

  ERi  0  0 0 0 0 0    ERi    ERi 
        
 IRi  0  0 0 0 0 0    IRi    IRi 
  ERFi   0   31 32 0 0 0    ERFi     ERFi 
        
COPEi  0  0 0  43 0 0  COPEi   COPEi 
     0   PTGi    PTGi 
 PTGi  0  51 52 0 54

 1 
 1 
 21 
where Ψ   0 0 1 
 
 0 0 0 1 
 0 0 0 0 1

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-7

Note that the means/intercepts and (residual) variances of the factors have been fixed to 0 and
1, respectively to scale the latent variables. In class, we used the t-rule and the two-step rule to
verify that the model is identified. We can thus go on to specify the model in lavaan.
Here we specify the model syntax object for the hypothesized SEM and fit the model.
mod.1 <-'#specifying measurement model portion
ER =~ fa + fr + si
IR =~ comt + con + cha + es + lo
ERF =~ pro + th + time
CPP =~ em + ind + ru + av + hy + rb
PTG =~ ptgi1 + ptgi2 + ptgi3 + ptgi4 + ptgi5

#specifying structural model portion

ERF ~ ER + IR
CPP ~ ERF
PTG ~ ER + IR + CPP
'

fit.1 <- sem(mod.1, sample.cov=mip.cov, sample.mean=mip.mns, sample.nobs=132,

meanstructure=TRUE, std.lv=TRUE)
summary(fit.1, fit.measures=TRUE, estimates=FALSE)

We have seen most of these commands before, so here we will highlight only portions of the
code.
Lavaan accepts either raw data or summary data (i.e., means, standard deviations, and
correlations among variables). As a default, lavaan assumes that data are in raw format. Since
the data from this example are in summary form, we used the sample.cov, sample.mean, and
sample.nobs arguments of the sem function, indicating the covariance matrix, mean vector,
and number of observations in the data, respectively.
As in Chapter 4, the measurement model for the latent variables is specified using the =~
operator. As in Chapters 1-3, regression equations are specified via the ~ operator, the only
difference here being that we are referencing the latent variables.
To set the scale of the latent factors, we have standardized the factor (residual) variances by
using std.lv=TRUE in the sem function call.
Within the summary function, we requested only the fit measures. We will examine the model
estimates only after obtaining a model that we deem to fit the data reasonably well.
Let us now turn to the fit indices for the model:
lavaan 0.6-2 ended normally after 53 iterations

Optimization method NLMINB

Number of free parameters 73

Number of observations 132

Estimator ML

Copyright © 2019 Curran-Bauer Analytics, LLC

5-8 | Chapter 5 Structural Equation Models with Latent Variables

Model Fit Test Statistic 350.000

Degrees of freedom 202
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 1205.231

Degrees of freedom 231
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.848

Tucker-Lewis Index (TLI) 0.826

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8012.004

Loglikelihood unrestricted model (H1) -7837.004

Number of free parameters 73

Akaike (AIC) 16170.009
Bayesian (BIC) 16380.453
Sample-size adjusted Bayesian (BIC) 16149.553

Root Mean Square Error of Approximation:

RMSEA 0.075
90 Percent Confidence Interval 0.061 0.087
P-value RMSEA <= 0.05 0.002

Standardized Root Mean Square Residual:

SRMR 0.101

The fit indices indicate that the model does not fit the data well. Rather than examining the
parameter estimates, we will next look at the modification indices to get a sense for what might
be causing the model to fit poorly, keeping in mind that any model modification must be
theoretically justifiable.
modindices(fit.1, sort.=TRUE, minimum.value=10)

lhs op rhs mi epc sepc.lv sepc.all sepc.nox

368 ru ~~ hy 32.238 34.381 34.381 4.461 4.461
233 comt ~~ cha 27.844 2.684 2.684 0.546 0.546
348 em ~~ ind 21.340 -26.655 -26.655 -0.410 -0.410
183 fa ~~ ind 14.250 -7.574 -7.574 -0.333 -0.333
250 con ~~ cha 12.343 -2.215 -2.215 -0.381 -0.381
119 IR =~ ptgi5 11.877 -0.730 -0.730 -0.241 -0.241
256 con ~~ em 10.695 -8.717 -8.717 -0.304 -0.304
160 PTG =~ cha 10.321 0.590 0.688 0.282 0.282

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-9

Modification indices suggest that the largest improvement to the model chi-square could be
achieved by allowing hypervigilance to correlate with rumination, over and above the
correlation implied by the coping factor, allowing challenge to correlate with commitment over
and above the individual resources factor, and allowing indirect and emotional coping to
correlate above and beyond the correlation implied by the coping factor. These modifications
reflect misspecification in the measurement model.

Confirmatory Factor Analysis

When building a structural equation model, a useful strategy to isolate misspecification is to
begin by ensuring that the foundation of the overall model, the measurement model, is
correctly specified. Once the measurement model has been properly specified, one can
consider misfit introduced by the structural model. Thus, we turn next to a CFA with saturated
covariances among factors. This strategy will allow us to get measurement right so that
measurement misspecification is not confounded with structural misfit.
The CFA model is shown below.

9 10 11
1 2 3
pro th time
fa fr si

Event-Related ptgi1 18

Factors
Environmental
Resources ptgi2 19

Posttraumatic
ptgi3 20
Growth
Individual
ptgi4 21
Resources
Coping
ptgi5 22

lo comt con cha es

em in ru av hy rb
4 5 6 7 8

12 13 14 15 16 17

Copyright © 2019 Curran-Bauer Analytics, LLC

5-10 | Chapter 5 Structural Equation Models with Latent Variables

The corresponding script to specify and fit this model is

mod.2 <-'#specifying measurement model
ER =~ fa + fr + si
IR =~ comt + con + cha + es + lo
ERF =~ pro + th + time
CPP =~ em + ind + ru + av + hy + rb
PTG =~ ptgi1 + ptgi2 + ptgi3 + ptgi4 + ptgi5
'

fit.2 <- sem(mod.2, sample.cov=mip.cov, sample.mean=mip.mns, sample.nobs=132,

meanstructure=TRUE, std.lv=TRUE)
summary(fit.2, fit.measures=TRUE, estimates=FALSE)

The model fit information is shown below:

lavaan 0.6-2 ended normally after 53 iterations

Optimization method NLMINB

Number of free parameters 76

Number of observations 132

Estimator ML
Model Fit Test Statistic 348.636
Degrees of freedom 199
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 1205.231

Degrees of freedom 231
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.846

Tucker-Lewis Index (TLI) 0.822

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8011.322

Loglikelihood unrestricted model (H1) -7837.004

Number of free parameters 76

Akaike (AIC) 16174.645
Bayesian (BIC) 16393.738
Sample-size adjusted Bayesian (BIC) 16153.348

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-11

Root Mean Square Error of Approximation:

RMSEA 0.075
90 Percent Confidence Interval 0.062 0.088
P-value RMSEA <= 0.05 0.001

Standardized Root Mean Square Residual:

SRMR 0.099

The model still does not fit the data well, confirming our hypothesis that the measurement
model, and not the structural model, is misspecified. Indeed, we can conduct a likelihood ratio
test comparing the CFA model with the originally hypothesized SEM because the hypothesized
model is a constrained version of the CFA with three structural parameters fixed to zero.
lavTestLRT(fit.2, fit.1)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.2 199 16175 16394 348.64
fit.1 202 16170 16380 350.00 1.3639 3 0.714

Thus, there is no significant decrement in fit associated with restricting the relationships among
the latent variables:
 2 (3)  350.00  348.64  1.36, p  .71
Combining this information with the information provided earlier from the modification indices,
we can conclude that the measurement model is the primary source of misfit and requires
respecification. Theoretically, allowing some residuals to correlate (as suggested by the MIs)
makes sense because some factors include multiple subscale scores as indicators. When
combined with other items from different scales, we would expect some degree of local
dependence for subscales from a common measure. Specifically, the IR factor includes three
Psychological Hardiness subscale scores as indicators (comt, con, and cha), but also two
indicators from independent scales (lo and es). The coping factor includes two indicators from
the Ways of Coping Inventory (em and ind), three indicators from the Impact of Event Scale
(ru, av, and hy), and a religious beliefs score from another scale (rb).

Copyright © 2019 Curran-Bauer Analytics, LLC

5-12 | Chapter 5 Structural Equation Models with Latent Variables

Revised Model
We now introduce correlated uniquenesses for comt, con, and cha on the individual resources
factor, between em and ind on the coping factor, and among ru, av, and hy on the coping
factor.

9 10 11
1 2 3
pro th time
fa fr si

 ERF
Event-Related ptgi1 18
Factors  PTG
Environmental
Resources ptgi2 19

Posttraumatic
ptgi3 20
Growth
Individual
ptgi4 21
Resources
Coping  Cope
ptgi5 22

lo comt con cha es

em in ru av hy rb
4 5 6 7 8
12 13 14 15 16 17

These correlated uniquenesses are included in a new model syntax object:

mod.3 <-'#specifying measurement model portion
ER =~ fa + fr + si
IR =~ comt + con + cha + es + lo
ERF =~ pro + th + time
CPP =~ em + ind + ru + av + hy + rb
PTG =~ ptgi1 + ptgi2 + ptgi3 + ptgi4 + ptgi5

#specifying structural model portion

ERF ~ ER + IR
CPP ~ ERF
PTG ~ ER + IR + CPP

#correlating uniquenesses of same-scale items

comt ~~ con + cha
con ~~ cha
em ~~ ind
ru ~~ av + hy
av ~~ hy
'

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-13

fit.3 <- sem(mod.3, sample.cov=mip.cov, sample.mean=mip.mns, sample.nobs=132,

meanstructure=TRUE, std.lv=TRUE)
summary(fit.3, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

Lavaan does not use separate names for uniquenesses/residuals/disturbances. Instead,

uniquenesses or disturbances are referred to by the referent variable. Thus, the line comt ~~
con + cha thus includes a covariance between the uniquenesses of comt and con and
between comt and cha.
The model fit is shown here:
lavaan 0.6-2 ended normally after 138 iterations

Optimization method NLMINB

Number of free parameters 80

Number of observations 132

Estimator ML
Model Fit Test Statistic 270.278
Degrees of freedom 195
P-value (Chi-square) 0.000

Model test baseline model:

Minimum Function Test Statistic 1205.231

Degrees of freedom 231
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.923

Tucker-Lewis Index (TLI) 0.908

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -7972.143

Loglikelihood unrestricted model (H1) -7837.004

Number of free parameters 80

Akaike (AIC) 16104.287
Bayesian (BIC) 16334.911
Sample-size adjusted Bayesian (BIC) 16081.869

Root Mean Square Error of Approximation:

RMSEA 0.054
90 Percent Confidence Interval 0.037 0.069
P-value RMSEA <= 0.05 0.325

Copyright © 2019 Curran-Bauer Analytics, LLC

5-14 | Chapter 5 Structural Equation Models with Latent Variables

Standardized Root Mean Square Residual:

SRMR 0.090

We can evaluate the improvement in fit associated with adding these 7 new free parameters to
the hypothesized model via the chi-square difference (likelihood ratio) test:
lavTestLRT(fit.3, fit.1)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.3 195 16104 16335 270.28
fit.1 202 16170 16380 350.00 79.722 7 1.569e-14 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

These results show that the correlated uniquenesses significantly improved the fit of the model:
 2 (7)  350.00  270.28  79.72, p  .001 .
Other fit indices suggest that the modified model fits the data reasonably well (though not
wonderfully).
Raw and fully standardized parameter estimates are presented below.
Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
ER =~
fa 1.317 0.362 3.637 0.000 1.317 0.342
fr 5.855 0.767 7.634 0.000 5.855 0.968
si 4.457 0.776 5.740 0.000 4.457 0.598
IR =~
comt 1.028 0.290 3.538 0.000 1.028 0.389
con 2.312 0.325 7.105 0.000 2.312 0.710
cha 1.141 0.264 4.323 0.000 1.141 0.468
es 3.007 0.551 5.453 0.000 3.007 0.510
lo -11.555 1.675 -6.897 0.000 -11.555 -0.640
ERF =~
pro 0.347 0.103 3.365 0.001 0.446 0.606
th -0.342 0.114 -3.012 0.003 -0.440 -0.387
time 1.316 0.743 1.772 0.076 1.693 0.211
CPP =~
em 6.060 1.103 5.495 0.000 6.429 0.566
ind -3.100 0.668 -4.645 0.000 -3.289 -0.483
ru 4.173 0.798 5.229 0.000 4.427 0.592
av 2.958 0.587 5.039 0.000 3.138 0.572
hy 3.725 0.625 5.962 0.000 3.952 0.668
rb 0.221 0.088 2.515 0.012 0.235 0.256

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-15

PTG =~
ptgi1 6.035 0.633 9.538 0.000 7.837 0.874
ptgi2 4.578 0.484 9.457 0.000 5.945 0.866
ptgi3 2.974 0.365 8.146 0.000 3.862 0.743
ptgi4 2.259 0.261 8.649 0.000 2.933 0.790
ptgi5 1.879 0.214 8.794 0.000 2.440 0.803

Regressions:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
ERF ~
ER -0.236 0.181 -1.305 0.192 -0.183 -0.183
IR 0.834 0.293 2.840 0.005 0.648 0.648
CPP ~
ERF -0.276 0.130 -2.120 0.034 -0.334 -0.334
PTG ~
ER 0.259 0.118 2.201 0.028 0.200 0.200
IR 0.284 0.139 2.039 0.041 0.219 0.219
CPP 0.722 0.168 4.299 0.000 0.590 0.590

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.comt ~~
.con -1.601 0.699 -2.292 0.022 -1.601 -0.287
.cha 2.242 0.601 3.729 0.000 2.242 0.427
.con ~~
.cha -1.923 0.648 -2.967 0.003 -1.923 -0.388
.em ~~
.ind -16.748 6.617 -2.531 0.011 -16.748 -0.300
.ru ~~
.av 2.123 3.766 0.564 0.573 2.123 0.078
.hy 19.248 4.875 3.948 0.000 19.248 0.725
.av ~~
.hy -0.383 2.978 -0.129 0.898 -0.383 -0.019
ER ~~
IR 0.244 0.101 2.407 0.016 0.244 0.244

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.fa 24.960 0.336 74.382 0.000 24.960 6.474
.fr 19.890 0.526 37.791 0.000 19.890 3.289
.si 18.240 0.649 28.123 0.000 18.240 2.448
.comt 10.360 0.230 45.087 0.000 10.360 3.924
.con 10.350 0.284 36.503 0.000 10.350 3.177
.cha 9.930 0.212 46.744 0.000 9.930 4.069
.es 30.330 0.513 59.087 0.000 30.330 5.143
.lo 80.470 1.571 51.217 0.000 80.470 4.458
.pro 2.670 0.064 41.612 0.000 2.670 3.622
.th 1.780 0.099 18.008 0.000 1.780 1.567
.time 3.910 0.698 5.602 0.000 3.910 0.488
.em 38.870 0.989 39.289 0.000 38.870 3.420
.ind 24.340 0.592 41.100 0.000 24.340 3.577

Copyright © 2019 Curran-Bauer Analytics, LLC

5-16 | Chapter 5 Structural Equation Models with Latent Variables

.ru 12.260 0.651 18.827 0.000 12.260 1.639

.av 11.800 0.478 24.698 0.000 11.800 2.150
.hy 9.170 0.515 17.804 0.000 9.170 1.550
.rb 2.740 0.080 34.348 0.000 2.740 2.990
.ptgi1 19.840 0.780 25.432 0.000 19.840 2.214
.ptgi2 10.530 0.597 17.632 0.000 10.530 1.535
.ptgi3 12.650 0.453 27.956 0.000 12.650 2.433
.ptgi4 9.800 0.323 30.309 0.000 9.800 2.638
.ptgi5 6.350 0.264 24.018 0.000 6.350 2.091
ER 0.000 0.000 0.000
IR 0.000 0.000 0.000
.ERF 0.000 0.000 0.000
.CPP 0.000 0.000 0.000
.PTG 0.000 0.000 0.000

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
.fa 13.130 1.669 7.867 0.000 13.130 0.883
.fr 2.287 7.782 0.294 0.769 2.287 0.063
.si 35.663 6.303 5.658 0.000 35.663 0.642
.comt 5.913 0.825 7.170 0.000 5.913 0.848
.con 5.269 1.182 4.457 0.000 5.269 0.496
.cha 4.654 0.694 6.704 0.000 4.654 0.781
.es 25.739 3.570 7.209 0.000 25.739 0.740
.lo 192.339 31.799 6.049 0.000 192.339 0.590
.pro 0.344 0.083 4.156 0.000 0.344 0.633
.th 1.096 0.157 6.966 0.000 1.096 0.850
.time 61.446 7.847 7.830 0.000 61.446 0.955
.em 87.866 14.001 6.276 0.000 87.866 0.680
.ind 35.475 5.157 6.879 0.000 35.475 0.766
.ru 36.371 6.685 5.441 0.000 36.371 0.650
.av 20.284 3.598 5.638 0.000 20.284 0.673
.hy 19.397 4.201 4.618 0.000 19.397 0.554
.rb 0.785 0.099 7.895 0.000 0.785 0.934
.ptgi1 18.908 3.311 5.711 0.000 18.908 0.235
.ptgi2 11.741 1.995 5.886 0.000 11.741 0.249
.ptgi3 12.112 1.676 7.227 0.000 12.112 0.448
.ptgi4 5.196 0.752 6.911 0.000 5.196 0.377
.ptgi5 3.275 0.482 6.789 0.000 3.275 0.355
ER 1.000 1.000 1.000
IR 1.000 1.000 1.000
.ERF 1.000 0.604 0.604
.CPP 1.000 0.888 0.888
.PTG 1.000 0.593 0.593

We focus on the structural parameter estimates in this chapter because interpretation of

measurement models has been discussed previously. IR has a significant direct effect on ERF (
  .834 ; S.E. = .293; p = .005) and PTG (   .284 ; S.E. = .139; p = .041). ER has a significant
direct effect on PTG (   .259 ; S.E. = .118; p = .28) and coping (   .276 ; S.E. = .130; p =
.034). Coping is significantly related to PTG (   .722 ; S.E. = .168 p < .001).

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-17

The standardized structural parameter estimates have been drawn on the path diagram below
to enhance interpretation of the model results. The non-significant path from environmental
resources to event-related factors is dashed. All other paths are statistically significant and
shown with solid lines.
.60
 ERF
Event-Related .59
-.18 Factors  PTG
Environmental
Resources
.20
.65 -.33 Posttraumatic
Growth
Individual .59
.22
Resources
Coping  Cope
.89

We can also examine the R-square statistics to get a sense of the magnitude of these effects:
R-Square:
Estimate
fa 0.117
fr 0.937
si 0.358
comt 0.152
con 0.504
cha 0.219
es 0.260
lo 0.410
pro 0.367
th 0.150
time 0.045
em 0.320
ind 0.234
ru 0.350
av 0.327
hy 0.446
rb 0.066
ptgi1 0.765
ptgi2 0.751
ptgi3 0.552
ptgi4 0.623
ptgi5 0.645
ERF 0.396
CPP 0.112
PTG 0.407

Copyright © 2019 Curran-Bauer Analytics, LLC

5-18 | Chapter 5 Structural Equation Models with Latent Variables

Together, these results suggest that environmental and individual resources have a moderate,
direct, positive influence on posttraumatic growth, cognitive coping has a strong, direct,
positive influence on posttraumatic growth, individual resources strongly predict more event-
related factors (shorter time since prognosis, poorer prognosis, and greater threat), and more
positive event-related factors predicts moderately less cognitive coping. Individual resources
and event related factors have a complex relationship with posttraumatic growth. To better
understand these relationships, we must consider direct, indirect, and total effects.
Examining Direct, Indirect and Total Effects
As in Chapter 3, to compute and test indirect and total effects, we need to modify the model
syntax object to include parameter labels for the paths involved in the effects so that we can
define the effects of interest. Here we will focus on the effects on post-traumatic growth.
mod.3b <-'#specifying measurement model portion
ER =~ fa + fr + si
IR =~ comt + con + cha + es + lo
ERF =~ pro + th + time
CPP =~ em + ind + ru + av + hy + rb
PTG =~ ptgi1 + ptgi2 + ptgi3 + ptgi4 + ptgi5

#specifying structural model portion

ERF ~ a1*ER + a2*IR
CPP ~ b*ERF
PTG ~ d1*ER + d2*IR + c*CPP

#correlating uniquenesses of same-scale items

comt ~~ con + cha
con ~~ cha
em ~~ ind
ru ~~ av + hy
av ~~ hy

#effects of ERF
dir_ERF := 0
ind_ERF := b*c
tot_ERF := 0 + b*c

#effects of ER
dir_ER := d1
ind_ER := a1*b*c
tot_ER := d1 + a1*b*c

#effects of IR
dir_IR := d2
ind_IR := a2*b*c
tot_IR := d2 + a2*b*c
'
fit.3b <- sem(mod.3b, sample.cov=mip.cov, sample.mean=mip.mns,
sample.nobs=132, meanstructure=TRUE, std.lv=TRUE)
summary(fit.3b, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Şenol-Durak and Ayvaşik’s Posttraumatic Growth Data | 5-19

We will thus obtain direct, indirect and total effect estimates of ERF, IR, and ER on PTG.
Most of the results obtained by fitting this model are identical to those presented previously, so
only the new information associated with the effect decomposition is shown below:
Defined Parameters:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
dir_ERF 0.000 0.000 0.000
ind_ERF -0.199 0.100 -1.982 0.047 -0.197 -0.197
tot_ERF -0.199 0.100 -1.982 0.047 -0.197 -0.197
dir_ER 0.259 0.118 2.201 0.028 0.200 0.200
ind_ER 0.047 0.041 1.152 0.249 0.036 0.036
tot_ER 0.306 0.123 2.494 0.013 0.236 0.236
dir_IR 0.284 0.139 2.039 0.041 0.219 0.219
ind_IR -0.166 0.091 -1.833 0.067 -0.128 -0.128
tot_IR 0.118 0.132 0.893 0.372 0.091 0.091

For each predictor-to-outcome effect, we have directed lavaan to first compute the direct
effect, then the indirect effect, and finally the total effect (along with standard errors and
significance tests). Standard errors are computed using the delta method (producing what is
known as the Sobel test for the indirect effects). Unlike in Chapter 3, we cannot use non-
parametric bootstrapping to obtain more accurate standard errors because we do not have
access to the raw data.
We will closely examine the effect of IR on PTG. We start by noting that the total effect of IR
on PTG is nonsignificant (p = .372). Upon closer examination, however, it is apparent that IR is
related to PTG both directly and indirectly. These effects are in opposite directions such that
the net, total effect is nearly zero. The direct effect of IR on PTG is positive and significant (p =
.041), whereas the indirect effect is negative and marginally significant (p = .067).

Copyright © 2019 Curran-Bauer Analytics, LLC

5-20 | Chapter 5 Structural Equation Models with Latent Variables

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 6
Multiple Groups Models
Multiple group analysis of Holzinger-Swineford Data ............................................................. 6-3
Preliminary Steps .................................................................................................................. 6-3
Evaluating Total Invariance .................................................................................................. 6-4
Evaluating the Level of Invariance ........................................................................................ 6-7
Identifying and Allowing for Partial Strong Invariance....................................................... 6-10
6-2 | Chapter 6 Multiple Groups Models

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-3

Multiple group analysis of Holzinger-Swineford Data

The data for this demonstration were provided by Holzinger & Swineford in their 1939
monograph A Study in Factor Analysis: The Stability of a Bi-Factor Solution. The sample includes
301 7th and 8th grade students, between 11-16 years of age, drawn from two schools. The
data is in the text file hs.dat and the accompanying R-script file is ch06.R. The variables in
the data set that we will use were described in Chapter 4 except for the grouping variable:
school 0 = Pasteur Elementary (N = 156); 1 =Grant-White Elementary (N = 145)
Please refer to Chapter 4 for a list and description of the other variables used in this
demonstration.

Preliminary Steps
As in previous chapters, we begin by reading in the data and assigning variable names:
hs <- read.table("hs.dat", header=FALSE)
names(hs) <- c("school", "female", "age", "month",
"visperc", "cubes", "lozenges",
"parcomp", "sencomp", "wordmean",
"addition", "countdot", "sccaps")

Then, to put the variables on more similar scales, we do a simple linear transformation of
addition, countdot, and sccaps (see Chapter 4):
hs$addition <- hs$addition/4
hs$countdot <- hs$countdot/4
hs$sccaps <- hs$sccaps/4

Next, to make our results easier to interpret, we will replace the numeric codes for school
with the school names, as follows:
library(plyr)
hs$school <- as.character(hs$school)
hs$school <- revalue(hs$school, c("0"="Pasteur","1"="Grant White"))

Note that this requires installation of the plyr package. The revalue function of the plyr
package provides an easy way to replace the numeric school codes with the school names. We
first convert school from a numeric variable to a character variable using the as.character
function, and then call the revalue function of plyr to conditionally assign the school names.
We are now ready to begin fitting the multiple group models with lavaan. To do so, we will
load the lavaan package. Additionally, invariance testing can be conducted straightforwardly
by also loading the semTools package (which must first be installed).
library(lavaan)
library(semTools)

Copyright © 2019 Curran-Bauer Analytics, LLC

6-4 | Chapter 6 Multiple Groups Models

The semTools package provides the measurementInvariance function, which provides an easy
way to conduct likelihood ratio tests between models imposing configural, weak, strong and
(optionally) strict invariance.

Evaluating Total Invariance

In Chapter 4, we analyzed the pooled data, ignoring the fact that students were grouped within
two schools. We thus implicitly assumed that the same model and set of parameter estimates
apply equally to students from the two groups. The final model that we obtained assuming
homogeneity across schools is illustrated below:

Visual Verbal Speed

visper cubes lozenges parcomp sencomp wordmean addition countdot sccaps

1 2 3 4 5 6 7 8 9

This model fit adequately well:

 2 (23)  52.38, p  .0004
CFI  .97, TLI  .95
RMSEA  .065, CI90  (0.042, 0.089)
SRMR  .041
Because we analyzed pooled data and assessed fit relative to the pooled moments, these fit
statistics do not inform us as to whether our assumption of homogeneity across the two
schools was reasonable. We can explicitly test whether total invariance (an identical model
with identical parameter values) holds across groups by assessing the fit of the model to the
group-specific moments. We do this by fitting a multiple groups CFA, constraining all
parameter estimates to equality.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-5

A model imposing total invariance can be fit in lavaan by first defining a basic CFA model syntax
object as follows:
mg.1 <- '#measurement model
visual =~ visperc + cubes + lozenges + sccaps
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps
'

The model syntax is identical to that described in Chapter 4.

Next, we will use some new options when we fit the model, as shown here:
fit.total <- cfa(mg.1, data=hs, meanstructure=TRUE, std.lv=TRUE,
group="school",
group.equal=c("loadings","intercepts","residuals",
"lv.variances","lv.covariances","means"))
summary(fit.total, fit.measures=TRUE)

We have indicated that we wish to conduct a multiple group analysis by using the group
argument of the fitting function (here cfa, but this is also available with the sem function).
Additionally, we have specified the group.equal argument, which is used to impose equality
constraints across groups on entire sets of parameter estimates. Here, we have indicated that
the factor loadings, intercepts of the indicators, residual variances of the indicators, factor
variances, factor covariances and factor means, respectively, should all be held equal. The
factor variances are also implicitly held equal over groups by requesting that the factors be
standardized with std.lv=TRUE, so specifying that they should be equal using group.equal
too is somewhat redundant. Indeed, upon fitting this model, lavaan outputs the warning:
Warning message:
In lav_partable_flat(FLAT, blocks = "group", meanstructure = meanstructure, :
lavaan WARNING: std.lv = TRUE forces all variances to be unity in all groups,
despite group.equal = "loadings"

In this case, however, this is precisely what we want: We wish to scale the factors to have
variances of one in all groups.
Now we can view model fit statistics to determine how well this highly restrictive model fits the
data. Note that this is the exact model that was estimated at the end of Chapter 4, except that
information about school attended is now included in the likelihood function. Therefore,
parameter estimates for this model will be identical to those obtained in Chapter 4 (and are
thus not shown here), but model fit will be different, as reported below.
lavaan 0.6-2 ended normally after 49 iterations

Optimization method NLMINB

Number of free parameters 62
Number of equality constraints 31

Number of observations per group

Pasteur 156
Grant White 145

Copyright © 2019 Curran-Bauer Analytics, LLC

6-6 | Chapter 6 Multiple Groups Models

Estimator ML
Model Fit Test Statistic 194.023
Degrees of freedom 77
P-value (Chi-square) 0.000

Chi-square for each group:

Pasteur 103.182
Grant White 90.840

Model test baseline model:

Minimum Function Test Statistic 957.769

Degrees of freedom 72
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.868

Tucker-Lewis Index (TLI) 0.876

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8309.780

Loglikelihood unrestricted model (H1) -8212.768

Number of free parameters 31

Akaike (AIC) 16681.560
Bayesian (BIC) 16796.480
Sample-size adjusted Bayesian (BIC) 16698.166

Root Mean Square Error of Approximation:

RMSEA 0.100
90 Percent Confidence Interval 0.083 0.118
P-value RMSEA <= 0.05 0.000

Standardized Root Mean Square Residual:

SRMR 0.111

Including information about school attendance has resulting in a substantial decrement in

model fit. Now that school is included as a grouping variable, we can see that the fit of the
model is inadequate. (Note that the SRMR here deviates slightly from the SRMR reported from
Mplus that was given in the lecture notes but all other fit measures are the same. Evidently the
two programs compute SRMR somewhat differently in the multiple group setting).

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-7

Evaluating the Level of Invariance

When evaluating a latent variable model with multiple groups, it is important to determine the
level of measurement invariance that is present. One way to do this is to start with the least
restrictive model (configural invariance) and move to more restrictive models (weak and then
strong invariance) as allowed by the data. The semTools package provides a function,
measurentInvariance, which nicely automates this process.

Before implementing this function, however, we need to think about how we wish to scale the
latent factors to identify a multiple group factor analysis model. The configural invariance
model imposes the least restrictions so we can consider it to establish minimum conditions for
identification. Recall that, under configural invariance, the form of the model is equivalent
across groups but all of the parameter estimates are allowed to differ except for those that
must be set equal simply to identify the model. There are several ways to set the scale of the
factors for identification purposes. One could, for instance, standardize the factors in both
groups. Another option would be to set the intercept and factor loading for one indicator per
factor to 0 and 1 in both groups, respectively. Our preference is instead to standardize the
factors in a reference group (here Grant-White) and equate the factor loadings and intercepts
for one indicator per factor in each other group (here Pasteur) to the reference group values.
This parameterization provides useful interpretations when moving from configural invariance
to weak and strong invariance models. (All of these options result in the same model fit, and
the chi-square is a simple summation of the chi-squares that would be obtained by fitting the
model separately in each group).
We can define this minimally identified model with the model syntax option shown here:
mg.2 <- '#loading of first indicator per factor set equal over groups
visual =~ NA*visperc + c(a,a)*visperc + cubes + lozenges + sccaps
verbal =~ NA*parcomp + c(b,b)*parcomp + sencomp + wordmean
speed =~ NA*addition + c(c,c)*addition + countdot + sccaps

#intercept of first indicator per factor set equal over groups

visperc ~ c(d,d)*1
parcomp ~ c(e,e)*1
addition ~ c(f,f)*1

#standardizing factor variances in Grant White only

visual ~~ c(NA,1)*visual
verbal ~~ c(NA,1)*verbal
speed ~~ c(NA,1)*speed

#standardizing factor means in Grant White only

visual ~ c(NA,0)*1
verbal ~ c(NA,0)*1
speed ~ c(NA,0)*1
'

Some aspects of this syntax are new, in part because scaling shortcuts (like std.lv) aren’t as
useful in the multiple group context.

Copyright © 2019 Curran-Bauer Analytics, LLC

6-8 | Chapter 6 Multiple Groups Models

First, notice that, in defining each factor, the first indicator is specified twice, first with the NA
modifier and next with the c(label,label) modifier (e.g., visual =~ NA*visperc +
c(a,a)*visperc). The NA modifier overrides the lavaan default to set the fix the first loading
for each factor to one. These loadings will instead be estimated. The c(label,label)
modifier gives the labels for the factor loading in group 1 (Pasteur) and group 2 (Grant-White),
respectively. Since the same label is used for each group’s loading, the estimated value will be
constrained to be equal across groups. When multiple modifiers are specified for the same
coefficient, the modifiers accumulate as lavaan parses the syntax (i.e., they both apply).
Second, we similarly specified that the intercepts of the first item per factor should be held
equal, again using the c(label,label) modifier (e.g., visperc ~ c(d,d)*1).
Third, we standardized the factors in Grant-White only, but estimated factor variances and
means in Pasteur. For the factor variances, this is accomplished by using c(NA,1) as the
modifier (freely estimated for the first group, set to one for the second). In similar fashion, we
use c(NA,0) as the modifier for the factor means (freely estimated for the first group, set to
zero for the second).
By default, all other parameters (intercepts, loadings, residual variances, factor covariances) will
differ between groups.
Now that we have specified a minimally identified model syntax object, we can run the
measurementInvariance function to obtain fit information, likelihood ratio tests, and
parameter estimates for the configural, weak, and strong invariance models simultaneously.
This is all accomplished with the single line:
mi <- measurementInvariance(model=mg.2,data=hs,group="school")

(There is also an option to also include strict invariance but, as we discussed in lecture, this is
not always useful so we don’t illustrate this here.)
We immediately obtain the following output:
Measurement invariance models:

Model 1 : fit.configural
Model 2 : fit.loadings
Model 3 : fit.intercepts
Model 4 : fit.means

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.configural 46 16631 16861 81.547
fit.loadings 53 16625 16829 89.266 7.720 7 0.358
fit.intercepts 59 16654 16836 130.828 41.562 6 2.244e-07 ***
fit.means 62 16688 16859 170.526 39.698 3 1.235e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-9

Fit measures:

cfi rmsea cfi.delta rmsea.delta

fit.configural 0.960 0.072 NA NA
fit.loadings 0.959 0.067 0.001 0.004
fit.intercepts 0.919 0.090 0.040 0.023
fit.means 0.877 0.108 0.041 0.018

Four models were fit to the data, of which the first three are of interest as they correspond to
configural, weak, and strong invariance models. The last model retains strong invariance but
sets the factor means to be equal. While that may be a restriction of interest, rejecting it has
no bearing on the level of measurement invariance so we ignore it for now.
For each model, we obtain the chi-square, AIC, BIC, CFI, and RMSEA. We also obtain chi-square
difference tests and differences in CFI and RMSEA for each model relative to the prior model in
the sequence. Comparing across the models, we can see that the configural invariance model
fits reasonably well and that imposing weak invariance does not significantly degrade model fit
(  2 (7)  7.72, p  .36 ). However, imposing strong invariance results in a noticeable
deterioration of model fit, as reflected both in a significant chi-square difference test relative to
the weak invariance model (  2 (6)  41.56, p  .0001 ) and poorer CFI and RMSEA values.
To see more detailed information on these models, we can use the summary function,
referencing the specific fit object of interest, as shown here (results not shown):
summary(mi$fit.configural)
summary(mi$fit.loadings)
summary(mi$fit.intercepts)

Having established weak invariance, we can safely compare factor variances and covariances
across groups. However, if only weak invariance holds then we cannot also compare factor
means. Although the strong invariance model was rejected, we may still be able to satisfy
partial strong invariance if only a small number of items require different intercepts across the
two groups. We will use modification indices to identify these items.

Copyright © 2019 Curran-Bauer Analytics, LLC

6-10 | Chapter 6 Multiple Groups Models

Identifying and Allowing for Partial Strong Invariance

In prior chapters, we used the modIndices function to obtain modification indices. To obtain
modification indices for equality constraints, however, we must use the lavTestScore function
(score test is another name for modification index), as shown:
lavTestScore(mi$fit.intercepts)

Running this syntax produces the results shown here:

univariate score tests:

lhs op rhs X2 df p.value

1 .p1. == .p38. 0.828 1 0.363
2 .p2. == .p39. 0.272 1 0.602
3 .p3. == .p40. 0.223 1 0.637
4 .p4. == .p41. 1.846 1 0.174
5 .p5. == .p42. 1.050 1 0.306
6 .p6. == .p43. 6.250 1 0.012
7 .p7. == .p44. 2.393 1 0.122
8 .p8. == .p45. 2.947 1 0.086
9 .p9. == .p46. 1.993 1 0.158
10 .p10. == .p47. 0.149 1 0.700
11 .p11. == .p48. 3.117 1 0.077
12 .p12. == .p49. 2.140 1 0.144
13 .p13. == .p50. 16.399 1 0.000
14 .p32. == .p69. 6.197 1 0.013
15 .p33. == .p70. 20.623 1 0.000
16 .p34. == .p71. 1.530 1 0.216

We can see that there are two large MIs, on lines 13 and 15 of the output. These involve
equality constraints for parameter 13 (.p13.) with 50 (.p50.) and parameter 33 (.p33.) with
70 (.p70.). We can clarify which parameters these are by running the following commands:
param<-parTable(mi$fit.intercepts)
param[c(13,50,33,70),]

The first line puts all the parameters from the strong invariance model into the object param,
and then the second command prints out only the 13th, 50th, 33rd, and 70th rows of param, in
that order, corresponding to the parameters associated with the largest MIs.
We obtain the following results:
id lhs op rhs user block group free ustart exo label plabel start est se
13 13 addition ~1 1 1 1 13 NA 0 f .p13. 25.486 23.295 0.472
50 50 addition ~1 1 2 2 50 NA 0 f .p50. 22.545 23.295 0.472
33 33 lozenges ~1 0 1 1 33 NA 0 .p33. .p33. 19.897 17.323 0.636
70 70 lozenges ~1 0 2 2 64 NA 0 .p33. .p70. 15.966 17.323 0.636

Thus we see that the MI of 16.399 for .p13. == .p50. is for the equality constraint on the
intercept of addition. Likewise, the MI of 20.623 for .p33. == .p70. is for the equality
constraint on the intercept of lozenges. These two items appear to be largely responsible for
the lack of fit of the strong invariance model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-11

To fit the model with partial invariance, we first define a new model syntax object, as follows:
mg.3 <- '#loading of first indicator per factor set equal over groups
visual =~ NA*visperc + cubes + lozenges + sccaps
verbal =~ NA*parcomp + sencomp + wordmean
speed =~ NA*addition + countdot + sccaps

#standardizing factor variances in Grant White only

visual ~~ c(NA,1)*visual
verbal ~~ c(NA,1)*verbal
speed ~~ c(NA,1)*speed

#standardizing factor means in Grant White only

visual ~ c(NA,0)*1
verbal ~ c(NA,0)*1
speed ~ c(NA,0)*1
'

It is very similar to the prior model syntax object, mg.2, except that we have removed the
equality constraints on the intercepts and loadings of the first indicator per factor. This is
necessary here because we need to allow the intercepts of addition to differ and this is the
first item on the speed factor. The model will instead be identified through equality
constraints imposed via the group.equal argument of the fitting function, as shown here:
partstr <- cfa(mg.3, data=hs, meanstructure=TRUE, group="school",
group.equal=c("loadings","intercepts"),
group.partial=c("lozenges ~1","addition ~1"))
summary(partstr, fit.measures=TRUE, standardized=TRUE)

We have seen the group.equal argument before (in the total invariance model), which we use
here to set all loadings and intercepts equal over groups. Alone, this would impose full strong
invariance but we can also implement the group.partial argument to remove specific
parameters from these equality constraints, allowing for partial invariance. Specifically, here
we designate group.partial=c("lozenges ~1","addition ~1") to allow the intercepts for
just lozenges and addition to differ over groups.
Model fit is shown below:
lavaan 0.6-2 ended normally after 87 iterations

Optimization method NLMINB

Number of free parameters 68
Number of equality constraints 17

Number of observations per group

Pasteur 156
Grant White 145

Estimator ML
Model Fit Test Statistic 95.064
Degrees of freedom 57
P-value (Chi-square) 0.001

Copyright © 2019 Curran-Bauer Analytics, LLC

6-12 | Chapter 6 Multiple Groups Models

Chi-square for each group:

Pasteur 59.744
Grant White 35.320

Model test baseline model:

Minimum Function Test Statistic 957.769

Degrees of freedom 72
P-value 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.957

Tucker-Lewis Index (TLI) 0.946

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -8260.301

Loglikelihood unrestricted model (H1) -8212.768

Number of free parameters 51

Akaike (AIC) 16622.601
Bayesian (BIC) 16811.664
Sample-size adjusted Bayesian (BIC) 16649.921

Root Mean Square Error of Approximation:

RMSEA 0.067
90 Percent Confidence Interval 0.042 0.090
P-value RMSEA <= 0.05 0.123

Standardized Root Mean Square Residual:

SRMR 0.057

The fit for this model is adequate. Notice also that lavaan has divided the overall model chi-
square into the portion associated with each school (which sum to the total).

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-13

We can also compare the fit of the partial strong invariance model to the less restricted weak
invariance model:
lavTestLRT(partstr, mi$fit.loadings)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

mi$fit.loadings 53 16625 16829 89.266
partstr 57 16623 16812 95.064 5.7981 4 0.2147

We see that the fit for the partial strong invariance model is not significantly worse than the fit
for the weak invariance model (  2 (4)  5.80, p  .21 ).
Additionally, we can evaluate the improvement in fit of the partial strong invariance model
relative to the more restricted full strong invariance model:
lavTestLRT(mi$fit.intercepts, partstr)

Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

partstr 57 16623 16812 95.064
mi$fit.intercepts 59 16654 16836 130.828 35.764 2 1.714e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Thus the partial strong invariance model also fits the data significantly better than the full
strong invariance model (  2 (2)  35.76, p  .0001 ).
Overall, partial strong invariance is supported. With partial strong invariance, we are able to
confidently interpret (co)variance parameters for latent variables. We can also make cautious
inferences related to factor means, but the interpretation of factor mean differences is tenuous
given that these differences are accurately estimated only if we restricted the intercepts of the
truly invariant indicators to equality.
Raw and standardized parameter estimates for the partial strong invariance model are shown
here:
Group 1 [Pasteur]:

Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc (.p1.) 5.132 0.539 9.515 0.000 5.206 0.744
cubes (.p2.) 2.110 0.322 6.552 0.000 2.140 0.432
lozengs (.p3.) 5.499 0.627 8.772 0.000 5.577 0.602
sccaps (.p4.) 3.487 0.600 5.815 0.000 3.537 0.391
verbal =~
parcomp (.p5.) 2.797 0.210 13.320 0.000 2.814 0.814
sencomp (.p6.) 4.183 0.309 13.520 0.000 4.208 0.831
wordmen (.p7.) 6.078 0.472 12.866 0.000 6.115 0.862

Copyright © 2019 Curran-Bauer Analytics, LLC

6-14 | Chapter 6 Multiple Groups Models

speed =~
additin (.p8.) 4.180 0.472 8.862 0.000 3.448 0.558
countdt (.p9.) 4.345 0.427 10.169 0.000 3.584 0.732
sccaps (.10.) 4.464 0.666 6.703 0.000 3.682 0.407

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual ~~
verbal 0.464 0.124 3.739 0.000 0.454 0.454
speed 0.136 0.103 1.319 0.187 0.162 0.162
verbal ~~
speed 0.220 0.093 2.362 0.018 0.265 0.265

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual -0.037 0.147 -0.253 0.801 -0.037 -0.037
verbal -0.618 0.131 -4.733 0.000 -0.614 -0.614
speed 0.109 0.131 0.837 0.403 0.133 0.133
.visperc (.29.) 29.715 0.560 53.063 0.000 29.715 4.246
.cubes (.30.) 24.447 0.309 79.162 0.000 24.447 4.936
.lozengs 20.101 0.896 22.429 0.000 20.101 2.168
.sccaps (.32.) 48.149 0.700 68.812 0.000 48.149 5.320
.parcomp (.33.) 10.063 0.264 38.181 0.000 10.063 2.912
.sencomp (.34.) 18.722 0.390 48.052 0.000 18.722 3.696
.wordmen (.35.) 17.238 0.587 29.353 0.000 17.238 2.429
.additin 25.029 0.626 39.961 0.000 25.029 4.051
.countdt (.37.) 27.397 0.428 63.966 0.000 27.397 5.596

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual 1.029 0.247 4.173 0.000 1.000 1.000
verbal 1.012 0.190 5.335 0.000 1.000 1.000
speed 0.680 0.166 4.105 0.000 1.000 1.000
.visperc 21.879 4.583 4.774 0.000 21.879 0.447
.cubes 19.945 2.455 8.123 0.000 19.945 0.813
.lozenges 54.860 7.878 6.964 0.000 54.860 0.638
.sccaps 51.604 7.096 7.272 0.000 51.604 0.630
.parcomp 4.023 0.625 6.431 0.000 4.023 0.337
.sencomp 7.946 1.306 6.084 0.000 7.946 0.310
.wordmean 12.977 2.460 5.276 0.000 12.977 0.258
.addition 26.285 3.765 6.981 0.000 26.285 0.689
.countdot 11.124 2.665 4.174 0.000 11.124 0.464

Group 2 [Grant White]:

Latent Variables:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual =~
visperc (.p1.) 5.132 0.539 9.515 0.000 5.132 0.734
cubes (.p2.) 2.110 0.322 6.552 0.000 2.110 0.477

Copyright © 2019 Curran-Bauer Analytics, LLC

 Multiple group analysis of Holzinger-Swineford Data | 6-15

lozengs (.p3.) 5.499 0.627 8.772 0.000 5.499 0.663

sccaps (.p4.) 3.487 0.600 5.815 0.000 3.487 0.382
verbal =~
parcomp (.p5.) 2.797 0.210 13.320 0.000 2.797 0.847
sencomp (.p6.) 4.183 0.309 13.520 0.000 4.183 0.862
wordmen (.p7.) 6.078 0.472 12.866 0.000 6.078 0.795
speed =~
additin (.p8.) 4.180 0.472 8.862 0.000 4.180 0.702
countdt (.p9.) 4.345 0.427 10.169 0.000 4.345 0.833
sccaps (.10.) 4.464 0.666 6.703 0.000 4.464 0.489

Covariances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual ~~
verbal 0.545 0.082 6.669 0.000 0.545 0.545
speed 0.393 0.103 3.816 0.000 0.393 0.393
verbal ~~
speed 0.268 0.095 2.834 0.005 0.268 0.268

Intercepts:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual 0.000 0.000 0.000
verbal 0.000 0.000 0.000
speed 0.000 0.000 0.000
.visperc (.29.) 29.715 0.560 53.063 0.000 29.715 4.252
.cubes (.30.) 24.447 0.309 79.162 0.000 24.447 5.526
.lozengs 15.966 0.689 23.175 0.000 15.966 1.925
.sccaps (.32.) 48.149 0.700 68.812 0.000 48.149 5.280
.parcomp (.33.) 10.063 0.264 38.181 0.000 10.063 3.047
.sencomp (.34.) 18.722 0.390 48.052 0.000 18.722 3.860
.wordmen (.35.) 17.238 0.587 29.353 0.000 17.238 2.256
.additin 22.545 0.495 45.576 0.000 22.545 3.785
.countdt (.37.) 27.397 0.428 63.966 0.000 27.397 5.254

Variances:
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
visual 1.000 1.000 1.000
verbal 1.000 1.000 1.000
speed 1.000 1.000 1.000
.visperc 22.501 4.198 5.360 0.000 22.501 0.461
.cubes 15.122 1.947 7.767 0.000 15.122 0.773
.lozenges 38.583 6.016 6.413 0.000 38.583 0.561
.sccaps 38.868 5.875 6.616 0.000 38.868 0.467
.parcomp 3.080 0.557 5.534 0.000 3.080 0.282
.sencomp 6.027 1.174 5.135 0.000 6.027 0.256
.wordmean 21.463 3.266 6.572 0.000 21.463 0.367
.addition 18.011 3.038 5.928 0.000 18.011 0.508
.countdot 8.311 2.519 3.299 0.001 8.311 0.306

Parameter estimates are listed separately by school. Note that in Chapter 6 of the lecture notes
we also present the confidence intervals around the latent factor means. We do not show this

Copyright © 2019 Curran-Bauer Analytics, LLC

6-16 | Chapter 6 Multiple Groups Models

here, but these are easy to obtain. To do this, just include ci=TRUE in the summary function,
i.e., summary(partstr, fit.measures=TRUE, ci=TRUE). For ease of interpretation, the raw
parameter estimates for each school are overlaid on the diagram below:
.39/.14

.55/.46 .27/.22

0/ -.04, 0/ -.62, 0/ .11,

1/1.03 Visual Verbal 1/1.01 Speed 1/.68

3.49

4.18
5.13 2.11 5.50 2.80 4.18 6.01 4.34 4.46

29.7 24.4 16.0/20.1 10.1 18.7 17.2 22.5/25.0 27.4 48.1

visper cubes lozenges parcomp sencomp wordmean addition countdot sccaps

1 2 3 4 5 6 7 8 9
22.5/21.9 15.1/19.9 38.6/54.9 3.1/4.0 6.0/7.9 21.5/13.0 18.0/26.3 8.3/11.1 38.9/51.6

Interpretations for the factor covariances are discussed first because these are the estimates in
which we are most confident. The variances of the Visual and Verbal factors are nearly
identical, but the variance of the Speed factor is unexpectedly smaller in the Pasteur group.
However, we can tell that this difference is not quite significant because the factor mean and
variance estimates for Pasteur are estimated relative to those in the Grant-White group. If a
95% confidence interval constructed for each school contains the fixed value from Grant-White,
then the groups are not significantly different.
As expected, the covariance of the Verbal with Visual is lower in Pasteur than Grant-White,
but the covariances with Speed are harder to interpret given the difference in the variance for
Speed across the two groups. In this case, we can instead consider the correlations among the
factors (reported with the standardized estimates). The Verbal factor is less highly correlated
with Visual in Pasteur, but is equally correlated with Speed across schools. Speed is less
correlated with Visual in Pasteur.
Cautiously moving to group differences in factor means, we see that the mean for the Verbal
factor is significantly lower in the Pasteur group, but the means of the other two factors are
nearly identical. This difference is consistent with our original hypothesis that the performance
of Pasteur students on verbal tasks would be worse than that of Grant-White students.
Finally, we can examine group differences in measurement. Holding the Visual factor
constant, the expected value for lozenges is 4.1 units higher in the Pasteur group than the
Grant White group. Holding the Speed factor constant, the expected value for addition is
2.5 unites higher in the Pasteur group than the Grant White group. Thus, students in Pasteur
find these tests easier than students of equal ability (with respect to Visual and Speed) from
Grant White.

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 7
SEM with Categorical
Indicators
Structural Equation Modeling of Schizophrenia Symptoms .................................................... 7-3
Preliminary Steps .................................................................................................................. 7-4
Confirmatory Factor Analysis using Diagonally Weighted Least Squares Estimation .......... 7-5
Revised Model with DWLS Estimation ................................................................................. 7-8
Plotting Item Characteristic Curves .................................................................................... 7-13
7-2 | Chapter 7 SEM with Categorical Indicators

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-3

Structural Equation Modeling of Schizophrenia Symptoms

The data for this demonstration are from a manuscript written by Mueser, Curran, and McHugo
(1997). The sample includes 437 patients with schizophrenia who were administered the Brief
Psychiatric Rating Scale, which consists of 18 item. Items were initially coded on a 7-point
scale, but were dichotomized due to extreme sparseness in higher categories. Items 17 and 18
were eliminated because exploratory analysis showed that they do not factor well with the
other items. The data file for this example is bprs.dat and the accompanying R-script file is
ch07.R. The binary variables (0=symptom absent, 1=symptom present) are:
Item01 somatic concern
Item02 anxiety
Item03 emotional withdrawal
Item04 conceptual disorganization
Item05 guilt feelings
Item06 tension
Item07 mannerisms/posturing
Item08 grandiosity
Item09 depressed mood
Item10 hostility
Item11 suspiciousness
Item12 hallucinatory behavior
Item13 motor retardation
Item14 uncooperativeness
Item15 unusual thought content
Item16 blunted affect
In lecture we described two general approaches to estimating SEMs with discrete dependent
variables: a limited information approach called diagonally-weighted least squares (DWLS) and
a full information approach called maximum likelihood (ML). There are pros and cons to each,
although when both are available, ML is superior from a statistical standpoint because it is
asymptotically unbiased and maximally efficient. However, ML does not provide omnibus tests
of model fit (e.g., model chi-square or RMSEA) nor does it provide modification indices. In
lecture we demonstrated both DWLS and ML for the BPRS data obtained from Mplus. However,
although lavaan currently implements a version of ML for categorical endogenous variables,
this is still in the early stages of development. We attempted to use this ML estimator for the
full four-factor model in this demonstration and were not able to obtain a converged solution,
although we were able to obtain estimates for a single factor at a time. Because of this, here we
focus solely on the DWLS estimator.

Copyright © 2019 Curran-Bauer Analytics, LLC

7-4 | Chapter 7 SEM with Categorical Indicators

Preliminary Steps
We begin by reading in the data and assigning variable names, as follows:
bprs <- read.table("bprs.dat", header=FALSE)
names <- c(paste("item", 1:9, sep = "0"),
paste("item", 10:18, sep = ""))
names(bprs) <- names

Note that because the variable names all begin with the prefix “item” and then are numbered
consecutively from 01 to 18, we took a syntax shortcut here with the paste function to
generate the labels item01-item18.
Because the data for this example is a bit different from the others we’ve considered so far, it’s
useful to examine the data before proceeding to model specification and fitting. First, we
simply view the first 20 rows of data for the first 8 items to get a sense of the data structure:
bprs[1:20,1:8]

item01 item02 item03 item04 item05 item06 item07 item08

1 1 1 0 0 0 1 0 0
2 1 1 0 1 1 1 1 0
3 0 0 0 1 1 0 0 1
4 0 1 0 0 1 0 0 0
5 0 1 0 0 0 1 0 0
6 1 1 0 1 1 1 1 0
7 1 0 1 1 0 0 0 1
8 1 1 1 1 1 1 0 0
9 1 1 1 1 0 1 1 1
10 1 1 1 1 1 1 0 1
11 0 1 0 0 1 0 0 0
12 0 1 0 0 1 0 1 0
13 1 1 1 0 0 1 1 0
14 1 1 0 0 0 1 0 0
15 0 1 0 0 0 0 0 0
16 1 1 1 1 1 1 0 0
17 1 0 1 1 1 0 0 0
18 0 0 0 0 1 0 0 0
19 1 1 0 1 1 1 1 0
20 0 1 0 0 1 0 0 0

We can clearly see that the data file consists entirely of 0's and 1's. Next, we can use the
colMeans function to see how frequently each item is endorsed (focusing on the 16 items in
our model). The mean of each binary item reflects the proportion of cases endorsing that item.
Because these are psychiatric symptoms, lower endorsement rates reflect more "severe" items.
colMeans(bprs[1:16])

item01 item02 item03 item04 item05 item06 item07 item08

0.5623679 0.7484144 0.5940803 0.5665962 0.3805497 0.5877378 0.3023256 0.3319239
item09 item10 item11 item12 item13 item14 item15 item16
0.5539112 0.4566596 0.6701903 0.5243129 0.4355180 0.3234672 0.7484144 0.8054968

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-5

If we were to use our usual normal-theory maximum likelihood estimator for these data, we
would be violating a number of cardinal assumptions (i.e., linearity, normality,
homoscedasticity). Instead, we will abandon the linear SEM we have considered thus far and
move to a wholly new statistical procedure that is explicitly designed for discretely scaled
dependent variables.

Confirmatory Factor Analysis using Diagonally Weighted Least Squares Estimation

We begin with the following hypothesized confirmatory factor model:

8 11 12 15 3 13 14 16

8 11 12 15 3 13 14 16

Thought Anergia

Affect Disorg

1 2 5 9 10 4 6 7

1 2 5 9 10 4 6 7

Using the underlying latent variable conceptualization, the model follows the following form:
y *i  ν  Ληi  εi
ηi  α  ζ i
where
VAR(εi )  Θ
VAR(ζi )  Ψ

For each item j, yij  1 if yij*   j .

For identification, we assume that each y * is standard normal with a mean of 0 and variance of
1. Thus, we only require an estimate of the correlation matrix of the y * , S* . This is estimated in
two steps: first, item thresholds are estimated using univariate response frequencies; second,
polychoric correlations are estimated using bivariate response frequencies. S* is constructed

Copyright © 2019 Curran-Bauer Analytics, LLC

7-6 | Chapter 7 SEM with Categorical Indicators

using pairwise polychoric correlation estimates. Recall that to calculate the polychorics, there
should be no empty cells in the 2-way bivariate frequency tables, so it is a good idea to check
these before proceeding with this approach.
If we assume that all item intercepts are zero, an assumption that is implicit in assuming that
the y * are standard normal, then it is possible to estimate item thresholds. Likewise, we
implicitly assume the variance of each y * is one, so residual variances in  are not actually
estimated – they are taken as remainders from 1.
As usual, we must identify the scale of the structural latent variables by setting their means and
variances to zero and one or by fixing an item on each factor to have an intercept of zero and a
factor loading of one. We choose to do the former so that covariances among latent variables
are interpretable as correlations and the estimated factor loadings are in a standardized metric.
The lavaan script for specifying and fitting this model is given below:
library(lavaan)

cat.1 <- '#measurement model

Thought =~ item08 + item11 + item12 + item15
Anergia =~ item03 + item13 + item14 + item16
Affect =~ item01 + item02 + item05 + item09 + item10
Disorg =~ item04 + item06 + item07
'
fit.1 <- cfa(cat.1, data=bprs, std.lv=TRUE, ordered=names[1:16])
summary(fit.1, fit.measures=TRUE, estimates=FALSE)
modindices(fit.1, sort.=TRUE, minimum.value=10)

Virtually all of this we have seen before. We have specified the basic measurement model
structure in the cat.1 model syntax object in exactly the same way as was illustrated for
continuous indicators in Chapter 4. Similarly, we’ve requested fit information and modification
indices just as in prior examples. The principal difference is that we use the ordered argument
in the fitting function to tell lavaan that the items are ordered-categorical (in this case, binary).
This automatically triggers a switch from ML to DWLS estimation.
It is also worth remembering that neither model fit statistics nor modification indices would be
available with full maximum likelihood estimation, so obtaining this information is a key
advantage to using DWLS (although at the cost of less efficient estimates for the model
parameters).

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-7

Model fit is shown below:

lavaan 0.6-2 ended normally after 17 iterations

Optimization method NLMINB

Number of free parameters 38

Number of observations 473

Estimator DWLS Robust

Model Fit Test Statistic 305.991 331.217
Degrees of freedom 98 98
P-value (Chi-square) 0.000 0.000
Scaling correction factor 0.978
Shift parameter 18.476
for simple second-order correction (Mplus variant)

Model test baseline model:

Minimum Function Test Statistic 1986.794 1449.631

Degrees of freedom 120 120
P-value 0.000 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.889 0.825

Tucker-Lewis Index (TLI) 0.864 0.785

Robust Comparative Fit Index (CFI) NA

Robust Tucker-Lewis Index (TLI) NA

Root Mean Square Error of Approximation:

RMSEA 0.067 0.071

90 Percent Confidence Interval 0.059 0.076 0.063 0.079
P-value RMSEA <= 0.05 0.001 0.000

Robust RMSEA NA
90 Percent Confidence Interval NA NA

Standardized Root Mean Square Residual:

SRMR 0.107 0.107

By default, when using DWLS, lavaan reports both naïve and robust fit statistics. We will
interpret the robust fit statistics (from the column “Robust”)
The absolute fit of the model, which we are able to assess using this estimator (but not with
ML), is poor for the hypothesized model. (There are a few trivial discrepancies from the
corresponding values reported in the lecture notes which were obtained with Mplus, likely due
to minor variations in how the robust statistics are calculated).

Copyright © 2019 Curran-Bauer Analytics, LLC

7-8 | Chapter 7 SEM with Categorical Indicators

Given the poor fit, we will view modification indices to see how the model might be improved
rather than interpret the current parameter estimates.
lhs op rhs mi epc sepc.lv sepc.all sepc.nox
135 Disorg =~ item13 61.706 -0.460 -0.460 -0.460 -0.460
209 item13 ~~ item16 52.517 1.006 1.006 2.663 2.663
96 Thought =~ item13 50.125 -0.345 -0.345 -0.345 -0.345
124 Affect =~ item13 33.296 -0.301 -0.301 -0.301 -0.301
205 item03 ~~ item04 31.230 0.393 0.393 0.754 0.754
134 Disorg =~ item03 30.432 0.324 0.324 0.324 0.324
216 item13 ~~ item06 30.380 -0.400 -0.400 -0.702 -0.702
116 Anergia =~ item04 29.310 0.394 0.394 0.394 0.394
262 item06 ~~ item07 24.913 0.426 0.426 0.635 0.635
198 item03 ~~ item14 23.859 0.451 0.451 0.732 0.732
147 item08 ~~ item13 22.242 -0.336 -0.336 -0.509 -0.509
260 item04 ~~ item06 21.397 -0.597 -0.597 -1.113 -1.113
95 Thought =~ item03 20.107 0.222 0.222 0.222 0.222
117 Anergia =~ item06 20.046 -0.311 -0.311 -0.311 -0.311
123 Affect =~ item03 18.968 0.230 0.230 0.230 0.230
199 item03 ~~ item16 15.479 -0.615 -0.615 -1.664 -1.664
104 Thought =~ item04 14.749 0.536 0.536 0.536 0.536
155 item08 ~~ item04 14.746 0.274 0.274 0.440 0.440
218 item14 ~~ item16 14.282 -0.487 -0.487 -1.161 -1.161
128 Affect =~ item06 14.194 0.490 0.490 0.490 0.490
106 Thought =~ item07 13.721 -0.376 -0.376 -0.376 -0.376
224 item14 ~~ item04 11.962 0.258 0.258 0.436 0.436
120 Affect =~ item11 11.237 0.400 0.400 0.400 0.400
97 Thought =~ item14 10.223 0.153 0.153 0.153 0.153

Many of the highest modification indices involve item13. Rather than blindly following the
suggestion of the modification indices, we consider why item13 might want to load on all of
the factors in the model and have residual correlations with lots of other items. The content of
item13 concerns motor retardation, a potential side-effect of the psychotropic medications
prescribed for schizophrenia. It is therefore not a “clean” measure of latent anergia, the
construct that we intended to measure with this item. For instance, motor retardation could be
a consequence of pharmacological treatment instigated to reduce thought disorder, resulting in
a high cross-loading on the thought disorder factor. We will thus remove item13 from
subsequent analyses.

Revised Model with DWLS Estimation

We do not present the equations or path diagram for the revised model because they are
virtually identical to the last model, except that item13 has been omitted. Similarly, the new
model syntax object simply omits this item. The script for specifying and fitting the model is
shown next. This time, we requested the parameter estimates and R-square measures for the
model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-9

cat.2 <- '#measurement model

Thought =~ item08 + item11 + item12 + item15
Anergia =~ item03 + item14 + item16
Affect =~ item01 + item02 + item05 + item09 + item10
Disorg =~ item04 + item06 + item07
'
fit.2 <- cfa(cat.2, data=bprs, std.lv=TRUE, ordered=names[1:16])
summary(fit.2, fit.measures=TRUE, rsquare=TRUE)
modindices(fit.2, sort.=TRUE, minimum.value=10)

Model fit for the revised model is presented below:

lavaan 0.6-2 ended normally after 18 iterations

Optimization method NLMINB

Number of free parameters 36

Number of observations 473

Estimator DWLS Robust

Model Fit Test Statistic 156.657 184.102
Degrees of freedom 84 84
P-value (Chi-square) 0.000 0.000
Scaling correction factor 0.916
Shift parameter 13.160
for simple second-order correction (Mplus variant)

Model test baseline model:

Minimum Function Test Statistic 1714.722 1275.104

Degrees of freedom 105 105
P-value 0.000 0.000

User model versus baseline model:

Comparative Fit Index (CFI) 0.955 0.914

Tucker-Lewis Index (TLI) 0.944 0.893

Root Mean Square Error of Approximation:

RMSEA 0.043 0.050

90 Percent Confidence Interval 0.032 0.053 0.040 0.060
P-value RMSEA <= 0.05 0.870 0.468

Robust RMSEA NA
90 Percent Confidence Interval NA NA

Standardized Root Mean Square Residual:

SRMR 0.083 0.083

Model fit is acceptable, and much improved from the original model.

Copyright © 2019 Curran-Bauer Analytics, LLC

7-10 | Chapter 7 SEM with Categorical Indicators

Modification indices (shown below) do not suggest any theoretically defensible additional
changes to our model (e.g., item04, conceptual disorganization, does not theoretically belong
on an anergia factor, and we also don’t want to go correlating residuals willy nilly):
lhs op rhs mi epc sepc.lv sepc.all sepc.nox
239 item06 ~~ item07 28.240 0.425 0.425 0.601 0.601
237 item04 ~~ item06 21.782 -0.605 -0.605 -1.160 -1.160
110 Anergia =~ item04 21.353 0.454 0.454 0.454 0.454
121 Affect =~ item06 15.817 0.420 0.420 0.420 0.420
111 Anergia =~ item06 15.103 -0.336 -0.336 -0.336 -0.336
146 item08 ~~ item04 13.379 0.262 0.262 0.451 0.451
114 Affect =~ item11 11.183 0.388 0.388 0.388 0.388
100 Thought =~ item07 10.671 -0.320 -0.320 -0.320 -0.320

Thus, we will retain the current model and examine the parameter estimates provided by the
DWLS estimator with robust standard errors:
Parameter Estimates:

Information Expected
Information saturated (h1) model Unstructured
Standard Errors Robust.sem

Latent Variables:
Estimate Std.Err z-value P(>|z|)
Thought =~
item08 0.477 0.062 7.651 0.000
item11 0.822 0.043 19.318 0.000
item12 0.706 0.044 15.982 0.000
item15 0.915 0.041 22.196 0.000
Anergia =~
item03 0.918 0.106 8.675 0.000
item14 0.635 0.085 7.448 0.000
item16 0.445 0.086 5.172 0.000
Affect =~
item01 0.473 0.068 6.974 0.000
item02 0.693 0.069 9.979 0.000
item05 0.554 0.063 8.734 0.000
item09 0.519 0.064 8.152 0.000
item10 0.576 0.068 8.437 0.000
Disorg =~
item04 0.750 0.076 9.922 0.000
item06 0.615 0.069 8.980 0.000
item07 0.440 0.071 6.227 0.000

Covariances:
Estimate Std.Err z-value P(>|z|)
Thought ~~
Anergia 0.182 0.074 2.452 0.014
Affect 0.590 0.065 9.137 0.000
Disorg 0.518 0.075 6.935 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-11

Anergia ~~
Affect -0.012 0.085 -0.137 0.891
Disorg 0.366 0.085 4.332 0.000
Affect ~~
Disorg 0.488 0.082 5.942 0.000

Intercepts:
Estimate Std.Err z-value P(>|z|)
.item08 0.000
.item11 0.000
.item12 0.000
.item15 0.000
.item03 0.000
.item14 0.000
.item16 0.000
.item01 0.000
.item02 0.000
.item05 0.000
.item09 0.000
.item10 0.000
.item04 0.000
.item06 0.000
.item07 0.000
Thought 0.000
Anergia 0.000
Affect 0.000
Disorg 0.000

Thresholds:
Estimate Std.Err z-value P(>|z|)
item08|t1 0.435 0.060 7.278 0.000
item11|t1 -0.440 0.060 -7.369 0.000
item12|t1 -0.061 0.058 -1.056 0.291
item15|t1 -0.670 0.063 -10.688 0.000
item03|t1 -0.238 0.058 -4.084 0.000
item14|t1 0.458 0.060 7.641 0.000
item16|t1 -0.861 0.066 -13.016 0.000
item01|t1 -0.157 0.058 -2.709 0.007
item02|t1 -0.670 0.063 -10.688 0.000
item05|t1 0.304 0.059 5.182 0.000
item09|t1 -0.136 0.058 -2.342 0.019
item10|t1 0.109 0.058 1.883 0.060
item04|t1 -0.168 0.058 -2.892 0.004
item06|t1 -0.222 0.058 -3.809 0.000
item07|t1 0.518 0.061 8.545 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

7-12 | Chapter 7 SEM with Categorical Indicators

Variances:
Estimate Std.Err z-value P(>|z|)
.item08 0.772
.item11 0.324
.item12 0.501
.item15 0.162
.item03 0.157
.item14 0.597
.item16 0.802
.item01 0.776
.item02 0.519
.item05 0.693
.item09 0.730
.item10 0.668
.item04 0.438
.item06 0.621
.item07 0.806
Thought 1.000
Anergia 1.000
Affect 1.000
Disorg 1.000

<snip>

R-Square:
Estimate
item08 0.228
item11 0.676
item12 0.499
item15 0.838
item03 0.843
item14 0.403
item16 0.198
item01 0.224
item02 0.481
item05 0.307
item09 0.270
item10 0.332
item04 0.562
item06 0.379
item07 0.194

Output for a categorical model is similar to the output that we have seen previously for linear
models, with two differences. First, because DWLS estimation assumes that the latent
distribution underlying each item is standard normal, a threshold has been estimated for each
dichotomous item. The threshold estimates how ‘difficult’ it is for the item to be endorsed.
Second, for the same reason, the residual variances aren’t estimated, but are remainders from
one, and thus have no standard errors or tests.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-13

We also obtain the communalities for the model (in section labeled R-square) As with a linear
factor model, the variance explained in each item is the sum of squared factor loadings for that
item times each factor’s variance. Factor variances are all one in this model and each item only
loads on a single factor, so the communalities are simply equal to the squared factor loadings.
The obtained estimates indicate that each item has a reasonably-sized factor loading and none
of the thresholds are too high or too low (around 3, indicating very low marginal endorsement
rates in one or the other category of the binary item). The model does a fairly good job at
explaining the variance in the fifteen measured variables.
Factor correlations indicate that thought is moderately correlated with affect ( r  .590 ; S.E.
= .065) and disorg ( r  .518 ; S.E. = .075), disorg is moderately correlated with anergia
( r  .366 ; S.E. = .085) and affect ( r  .488 ; S.E. = .082), anergia and thought are mildly
correlated ( r  .182 ; S.E. = .074), and affect and anergia are not significantly related
( r  .012 ; S.E. = .085).

Plotting Item Characteristic Curves

To aid interpretation of the factor-to-item relationships (i.e., the factor loadings and the
thresholds), it is useful to look at plots of the item characteristic curves (ICCs) or tracelines. In
these graphs, an individual’s latent variable score is on the x-axis and the conditional probability
of item endorsement given the latent variable score is on the y-axis. Items with steeper slopes
have higher factor loadings and are more discriminating than items with flatter slopes. Items
that are more to the left of the screen have lower thresholds, and are therefore easier to
endorse.
There are many ways to make plots within R, including a variety of packages. Here we provide
a relatively simple script for generating the ICCs. Because our focus is on fitting SEM models
and not on making graphs, we do not walk through the script in great detail, touching only on
the most important aspects.

As a first step, we need to extract the threshold and loading estimates from the fit object
generated by lavaan for our model. We do that here:
est <- parameterEstimates(fit.2)
threshold <- est[est$op=="|",c("lhs","est")]
loading <- est[est$op=="=~",c("rhs","est")]
names(threshold) <- c("item","est")
names(loading) <- c("item","est")

This produces two matrices, one called threshold that consists of a column of labels for the
items and a column of threshold estimates and another called loading that likewise includes
the item labels in the first column but has a second column consisting of factor loading
estimates.

Copyright © 2019 Curran-Bauer Analytics, LLC

7-14 | Chapter 7 SEM with Categorical Indicators

Next we set up the plot. Here we demonstrate this for the items loading on the thought
disorder factor.
plot(c(-4,4),c(0,1),type="n",xlab="Eta",ylab="Endorsement Probability")
items <- c("item08","item11","item12","item15")
lines <- c(1:length(items))
colors <- rainbow(length(items))

The plot function call here simply sets the frame for the plot, indicating that the x-axis will go
from -4 to 4 and be labeled “Eta” and that the y-axis will go from 0 to 1 and be labeled
“Endorsement probability”. We then put the item labels for the items loading on thought
disorder into the items vector and created the vectors lines and colors to enable us to vary
the line types and colors of the ICCs for the four items.
The actual plotting of the ICCs is done with the following script:
for (i in 1:4){
eta <- seq(-4, 4, by=.1)
int <- -1*threshold[i,"est"]
slp <- loading[i,"est"]
prob <- pnorm(int + slp*eta)
dat <- data.frame(cbind(eta, prob))

lines(dat$eta, dat$prob, type="l", lwd=2, lty=lines[i],

col=colors[i])
}

For each item, we choose factor values from -4 to 4 in increments of .1, pick off the threshold
and loading of the item, and compute the implied probability of endorsement of the item given
these item parameters. The values of the factor and corresponding endorsement probabilities
are put into a data frame called dat, and then the lines function is used to add the ICC for
each item to the plot with different line types and colors across items.
Last, we can give our plot a title and legend as follows:
title("Item Characteristic Curves: Thought Disorder")
legend(-4, 1, items, col=colors, lty=lines, lwd=2)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-15

Copyright © 2019 Curran-Bauer Analytics, LLC

7-16 | Chapter 7 SEM with Categorical Indicators

Copyright © 2019 Curran-Bauer Analytics, LLC

 Structural Equation Modeling of Schizophrenia Symptoms | 7-17

All of the ICCs look reasonable although there is clearly variation in the quality of the items
(with lower quality items having flatter slopes).
For the thought disorder factor there is a great deal of variation in the severity and
discrimination of the items. Item08 “grandiosity” is both the least discriminating item (i.e.,
least strongly related to the factor, with the flattest ICC and lowest factor loading) and the
symptom that is most severe (requires a very high level of thought disorder to have a high
endorsement probability).
Moving to the affect factor we see these items are relatively consistent in their
discrimination but vary somewhat in severity. Item02, “anxiety”, has the lowest threshold
and is thus the least severe symptom of affect. In contrast, item05 “guilt feelings” has the
highest threshold and thus its presence indicates more severe affect disorder.

Copyright © 2019 Curran-Bauer Analytics, LLC

7-18 | Chapter 7 SEM with Categorical Indicators

Copyright © 2019 Curran-Bauer Analytics, LLC

Chapter 8
Latent Growth Curve
Analysis
Latent Growth Curve Analysis of Antisocial Behavior .............................................................. 8-3
Preliminary Steps .................................................................................................................. 8-3
Intercept-Only Model ........................................................................................................... 8-4
Intercept and Slope Model ................................................................................................... 8-8
Quadratic Growth Trajectory Model .................................................................................. 8-13
Freed Loading (“Fully Latent”) Growth Model: Additional Demonstration ....................... 8-17
Linear Growth Model Conditional on Gender .................................................................... 8-24
Linear Growth Model Conditional on Cognitive Stimulation in the Home ........................ 8-30
8-2 | Chapter 8 Latent Growth Curve Analysis

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-3

Latent Growth Curve Analysis of Antisocial Behavior

The data for this demonstration are from the National Longitudinal Study of Youth. The sample
includes 405 children who were aged 6-8 at the first assessment (in 1986). Children were
assessed bi-yearly (until 1992), resulting in up to four measurement occasions per child. Data
are in the file anti.dat and the script is in ch08.R. The variables in the data set that we will
use are described below:
id participant identification; this will not be used in the analysis
gen 0 = female; 1 = male
homecog level of cognitive support in the home (mean-centered); higher values
denote more cognitive support in the home
anti6-anti14 repeated assessments of mother reports of child antisocial behavior
(higher values indicate more antisocial behavior); post-script denotes
child age at time of assessment
We are interested in characterizing the functional form of developmental change in aggressive
behavior for children aged 6-14. We would also like to understand how our time invariant
covariates (gender and cognitive support in the home) contribute to initial levels of, and change
in, aggressive behavior.

Preliminary Steps
As a starting point, we will read in the data and assign variable names:
classes <- rep("numeric",12)
anti <- read.table("anti.dat", header=FALSE, colClasses=classes,
na.strings=".")
y <- paste("anti", 6:14, sep = "")
names(anti) <- c(c("id","gen","homecog"),y)

The classes vector created in the top line is referred to in the read.table function with the
colClasses argument to define the data type for all variables as numeric. Without these
steps, our variables would be assigned the factors data type instead. We also indicate that
missing data is coded as . using the na.strings argument of read.table. The last two lines
assign the variable names. We defined the vector y using the paste function to include the
names of the repeated measures, i.e., anti6, anti7, anti8, …, anti14. Then we used names
function to assign all the variable names.
We are now ready to examine some descriptive statistics. For simple univariate descriptive
statistics, we like to us the describe function of the psych package (which must be installed).
library(psych)
describe(anti[c("gen","homecog",y)])

Copyright © 2019 Curran-Bauer Analytics, LLC

8-4 | Chapter 8 Latent Growth Curve Analysis

The univariate statistics provided by the describe function are shown here:
vars n mean sd median trimmed mad min max range skew kurtosis se
gen 1 405 0.50 0.50 1.00 0.50 0.00 0.00 1.00 1 0.00 -2.00 0.02
homecog 2 405 0.00 2.58 0.11 0.13 2.97 -7.89 5.11 13 -0.45 -0.14 0.13
anti6 3 122 1.57 1.67 1.00 1.34 1.48 0.00 9.00 9 1.33 2.32 0.15
anti7 4 168 1.55 1.55 1.00 1.35 1.48 0.00 7.00 7 1.05 0.82 0.12
anti8 5 146 1.97 1.80 1.00 1.75 1.48 0.00 7.00 7 0.92 0.01 0.15
anti9 6 192 1.89 1.95 1.00 1.58 1.48 0.00 9.00 9 1.23 1.18 0.14
anti10 7 151 2.14 2.14 2.00 1.81 1.48 0.00 10.00 10 1.24 1.22 0.17
anti11 8 174 1.79 1.95 1.00 1.50 1.48 0.00 10.00 10 1.54 3.01 0.15
anti12 9 135 1.84 1.78 2.00 1.62 2.97 0.00 7.00 7 0.75 -0.33 0.15
anti13 10 173 2.24 2.27 2.00 1.93 2.97 0.00 10.00 10 1.09 0.74 0.17
anti14 11 101 1.96 2.09 1.00 1.67 1.48 0.00 10.00 10 1.13 1.11 0.21

We see that there are 405 cases, but there are varying numbers assessed at each specific age
(e.g., 122 of the 405 cases provided data on antisocial at age 6, 168 at age 7, and so on). There
appears to be a slight upward trend in the means over time.

Intercept-Only Model
As we have seen in preceding chapters, it is often useful to start with simple models and then
to increase in model complexity. Adhering to this strategy, we begin with an intercept-only
model. This model evaluates whether individual differences exist with respect to mean levels of
mother-reported aggression, aggregating over age of assessment. That is, we exclude a time
trend from the initial model. The path diagram for this model is:

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-5

The measurement model takes the following form:

 anti6i  1  1i 
 anti7  1  
 i     2i 
 anti8i  1  3i 
     
 anti9i  1  4 i 
 anti10i   1 1i    5i 
     
 anti11i  1  6 i 
 anti12i  1  7 i 
     
 anti13i  1   8i 
 anti14  1  
 i    9i 
where
COV  εi   Θ  DIAG 11,22 , ,99 
and the latent variable model is simply
1i  1   1i

where E 1i   1 and VAR 1i   VAR  1i    .

The repeated measures have the following model-implied mean and covariance structures:
μ(θ) = α
Σ(θ) = 1 1 + Θ
This model implies that the repeated measures of antisocial behavior exhibit no systematic
change over time, and that variation among repeated measurements of antisocial behavior are
solely a function of individual baseline propensities to engage in antisocial behavior and
measurement error (or time-specific true score variance). Variance in the repeated measures
that is not explained by the latent intercept (1 ) is considered to be unsystematic error ( 1... 9 ).
The script for specifying and fitting this model in lavaan is shown here:
library(lavaan)

int <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +

1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
'
fit.int <- growth(int, data=anti, missing="ML")
summary(fit.int, fit.measures=TRUE, estimates=FALSE)

In the model syntax object, we simply list out the repeated measures as indicator variables. To
define I as a latent intercept factor, we assigned each repeated measure a factor loading of
one.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-6 | Chapter 8 Latent Growth Curve Analysis

To fit the model, we use the growth function, which invokes a number of defaults useful for
growth modeling such as including mean structure, setting indicator intercepts to zero, and
estimating means of latent growth factors. Thus, we need not explicitly define these aspects of
the model in the model syntax object. Note that we also indicated missing = “ML” so that
lavaan would use the individual likelihood version of ML to accommodate partially missing data.
Model fit is shown below:
lavaan 0.6-2 ended normally after 40 iterations

Optimization method NLMINB

Number of free parameters 11

Number of observations 405

Number of missing patterns 49

Estimator ML
Model Fit Test Statistic 102.290
Degrees of freedom 43
P-value (Chi-square) 0.000

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2661.782

Loglikelihood unrestricted model (H1) -2610.637

Number of free parameters 11

Akaike (AIC) 5345.564
Bayesian (BIC) 5389.607
Sample-size adjusted Bayesian (BIC) 5354.702

Root Mean Square Error of Approximation:

RMSEA 0.058
90 Percent Confidence Interval 0.044 0.073
P-value RMSEA <= 0.05 0.163

Standardized Root Mean Square Residual:

SRMR 0.191

The model fit statistics reported by lavaan do not match those presented in lecture, which were
generated by Mplus. For instance, lavaan reports a chi-square of 102.29 with 43 degrees of
freedom. In contrast, Mplus reports a chi-square of 101.71 with 39 degrees of freedom.
Additionally, lavaan did not compute CFI or TLI, whereas Mplus did. The reason for these
differences has to do with how the two programs account for missing data.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-7

We can observe the extent of missing data using the lavInspect function of lavaan:
coverage <- lavInspect(fit.int, what = "coverage")
coverage*405

The coverage matrix generated by the what = "coverage" argument to this function indicates
the fraction of data available for each element of the covariance matrix. We multiplied this by
405, our total sample size, to observe the number of cases available with which to estimate
each covariance:
anti6 anti7 anti8 anti9 anti10 anti11 anti12 anti13 anti14
anti6 122
anti7 0 168
anti8 31 0 146
anti9 86 106 0 192
anti10 12 49 98 3 151
anti11 80 82 28 139 1 174
anti12 28 45 81 16 99 14 135
anti13 61 96 22 138 16 145 12 173
anti14 0 35 66 3 85 4 82 11 101

There are four elements for which zero cases provide data. These “zero cells” are due to the
accelerated longitudinal design of the study. For instance, no children were observed at both 6
and 7 years of age given the biennial assessment schedule, thus the covariance between anti6
and anti7 has zero coverage in the data. Mplus automatically detects these zero cells and
removes them from the degrees of freedom for the model chi-square test. Lavaan does not do
so. This is why the degrees of freedom for the chi-square test were lower by four in Mplus
relative to lavaan. Additionally, the precise values of the model chi-squares differ slightly owing
to differences in convergence criteria for the saturated model. These zero-coverage
covariances are technically not identified in the saturated model and thus take on arbitrary
(and irrelevant) values in the estimation process. Lavaan and Mplus will proceed with fitting
the saturated model in any event whereas most other software programs will not (and thus will
not provide any tests of absolute fit).
In this case, regardless of whether we account for the zero-cells or not, we reach the same
conclusion regarding model fit. All measures of fit suggest that the intercept-only model does
not adequately describe our data. The sample means provide a clue for why this is the case;
there appears to be a linear upward trend in the means over time. Thus, we will move to a
more complex model with a latent linear slope in addition to the latent intercept.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-8 | Chapter 8 Latent Growth Curve Analysis

Intercept and Slope Model

We will be able to determine whether including a latent linear slope significantly improves
model fit relative to the intercept-only model because these two models are nested. We want
to fit the following linear growth model:

The model takes the following form:

 anti6i  1 0  1i 
 anti7  1 1   
 i     2i 
 anti8i  1 2  3i 
     
 anti9i  1 3  4i
 1i   

 anti10i   1 4       5i 
      
 anti11i  1 5   2 i   6 i 
 anti12i  1 6  7 i 
     
 anti13i  1 7   8i 
 anti14  1 8   
 i   9i 
where
COV  εi   Θ  DIAG 11,22 , ,99 

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-9

and the latent variable model is now

1i  1   1i 
       
 2i   2   2i 
where
 
E  ηi   α , COV  ηi   COV  ζi   Ψ , and Ψ   11 
 21  22 
The repeated measures have the following model implied mean and covariance structures:
μ(θ) = Λα
Σ(θ) = ΛΨΛ + Θ
Now the repeated measures have an expected value that consists of the model-implied mean
of antisocial behavior at age 6 (  1 ) plus the expected change in antisocial behavior
corresponding to the linear increase in age (  2 ). The covariance matrix of the repeated
measures reflect the variances and covariance of the intercept and slope factors, as transmitted
through the factor loading matrix. Additional variance is added at each time point by the
independent errors of the repeated measures.
An interesting implication of the linear growth model is that the variance contribution from the
linear slope increases as a quadratic function of time because the variance of the slope factor is
multiplied by the square of the factor loading in the calculation of total variance of the
repeated measures.
This model is specified and fit as follows:
lin <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
'
fit.lin <- growth(lin, data=anti, missing="ML")
summary(fit.lin, fit.measures=TRUE, rsquare=TRUE)

lavTestLRT(fit.lin, fit.int)

We have added a new latent variable, S, and specified that each of the repeated measures load
on this latent variable with linearly incrementing fixed factor loadings of 0, 1, 2, etc. This
defines S to be the slope factor of a linear growth curve model. I retains its interpretation as
the intercept of the model, now representing initial levels of antisocial behavior (at age 6).
In addition to requesting fit statistics, parameter estimates, and R-squared statistics, we also
requested a likelihood ratio test to examine the improvement in fit of the linear model relative
to the intercept-only model.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-10 | Chapter 8 Latent Growth Curve Analysis

Model fit statistics for the linear growth model are shown here:
lavaan 0.6-2 ended normally after 53 iterations

Optimization method NLMINB

Number of free parameters 14

Number of observations 405

Number of missing patterns 49

Estimator ML
Model Fit Test Statistic 53.171
Degrees of freedom 40
P-value (Chi-square) 0.079

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2637.223

Loglikelihood unrestricted model (H1) -2610.637

Number of free parameters 14

Akaike (AIC) 5302.445
Bayesian (BIC) 5358.500
Sample-size adjusted Bayesian (BIC) 5314.076

Root Mean Square Error of Approximation:

RMSEA 0.029
90 Percent Confidence Interval 0.000 0.047
P-value RMSEA <= 0.05 0.972

Standardized Root Mean Square Residual:

SRMR 0.123

The chi-square and RMSEA suggest this model fits well. The likelihood ratio test comparing this
model with the intercept-only model is also highly significant (  
 2 3  49.1, p  .001
), indicating
that the linear slope substantially improves the model fit:
Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.lin 40 5302.4 5358.5 53.171
fit.int 43 5345.6 5389.6 102.290 49.119 3 1.231e-10 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-11

We thus now consider the parameter estimates for this model, shown here:
Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
S =~
anti6 0.000
anti7 1.000
anti8 2.000
anti9 3.000
anti10 4.000
anti11 5.000
anti12 6.000
anti13 7.000
anti14 8.000

Covariances:
Estimate Std.Err z-value P(>|z|)
I ~~
S 0.050 0.052 0.962 0.336

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000
.anti13 0.000
.anti14 0.000
I 1.621 0.086 18.870 0.000
S 0.074 0.018 4.128 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

8-12 | Chapter 8 Latent Growth Curve Analysis

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 1.881 0.375 5.021 0.000
.anti7 1.345 0.264 5.087 0.000
.anti8 1.686 0.290 5.818 0.000
.anti9 1.867 0.251 7.430 0.000
.anti10 2.181 0.323 6.748 0.000
.anti11 1.664 0.249 6.691 0.000
.anti12 1.458 0.274 5.329 0.000
.anti13 1.799 0.337 5.335 0.000
.anti14 1.726 0.414 4.174 0.000
I 1.093 0.292 3.743 0.000
S 0.027 0.012 2.167 0.030

R-Square:
Estimate
anti6 0.367
anti7 0.476
anti8 0.454
anti9 0.467
anti10 0.468
anti11 0.576
anti12 0.645
anti13 0.633
anti14 0.676

Notice that the factor loadings and observed variable intercepts are all fixed and hence no
standard errors, z-statistics, or p-values are reported for these values.
The means for the intercept and linear slope components of growth are both significant. On
average, mothers report antisocial behavior of 1.62 at age 6 and these scores increase at a rate
of .07-units per year.
The variances of the intercept and linear slope components of growth are also both significant.
There is significant individual variability in the initial levels of antisocial behavior and in the
rates of increase in antisocial behavior over time.
There is a non-significant covariance between the intercept and the linear slope components of
growth. On average, individual variability in starting point is not related to individual variability
in rate of change over time.
The R-squared statistics show that the growth factors account for an increasing percentage of
the variance in the repeated measures as age advances.
Although this model appears quite reasonable, we have not yet considered functional forms of
growth other than a linear trajectory—we will do this next.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-13

Quadratic Growth Trajectory Model

We now extend the model to include a quadratic growth factor, as shown in the diagram:

The model takes the form:

 anti6i  1 0 0  1i 
 anti7  1 1 1  
 i     2i 
 anti8i  1 2 4  3i 
      
 anti9i  1 3 9   1i   4i 
 anti10i   1 4 16  2i    5i 
     
 anti11i  1 5 25 3i   6i 
 anti12i  1 6 36   7 i 
     
 anti13i  1 7 49   8i 
 anti14  1 64   
 i  8  9 i 
where
COV  εi   Θ  DIAG 11,22 , ,99 

and the latent variable model is now

1i  1    1i 
       
 2i   2   2i 
3i   3   3i 

Copyright © 2019 Curran-Bauer Analytics, LLC

8-14 | Chapter 8 Latent Growth Curve Analysis

where
 11 

COV  ζ i   Ψ   21  22 
 
 31  32  33 

The repeated measures again have the following model implied mean and covariance
structures:
μ(θ) = Λα
Σ(θ) = ΛΨΛ + Θ
The script for specifying and estimating this model is:
quad <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
L =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
Q =~ 0*anti6 + 1*anti7 + 4*anti8 + 9*anti9 +
16*anti10 + 25*anti11 + 36*anti12 + 49*anti13 +
64*anti14
'
fit.quad <- growth(quad, data=anti, missing="ML")
summary(fit.quad, fit.measures=TRUE, rsquare=TRUE)

lavTestLRT(fit.quad, fit.lin)

Now we have specified a third factor called Q to represent quadratic growth by fixing the factor
loadings to the squares of the linear factor loadings. We also requested a likelihood ratio test
of the quadratic model relative to the linear model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-15

Model fit for the quadratic model is given here:

lavaan 0.6-2 ended normally after 86 iterations

Optimization method NLMINB

Number of free parameters 18

Number of observations 405

Number of missing patterns 49

Estimator ML
Model Fit Test Statistic 42.558
Degrees of freedom 36
P-value (Chi-square) 0.210

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2631.916

Loglikelihood unrestricted model (H1) -2610.637

Number of free parameters 18

Akaike (AIC) 5299.832
Bayesian (BIC) 5371.902
Sample-size adjusted Bayesian (BIC) 5314.786

Root Mean Square Error of Approximation:

RMSEA 0.021
90 Percent Confidence Interval 0.000 0.043
P-value RMSEA <= 0.05 0.989

Standardized Root Mean Square Residual:

SRMR 0.130

This model fits the data significantly better than the linear growth model with an LRT of
 2  4  10.6, p  .03
; however, the improvement is modest.
Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.quad 36 5299.8 5371.9 42.558
fit.lin 40 5302.4 5358.5 53.171 10.613 4 0.03127 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Copyright © 2019 Curran-Bauer Analytics, LLC

8-16 | Chapter 8 Latent Growth Curve Analysis

Parameter estimates are shown below:

Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
L =~
anti6 0.000
anti7 1.000
anti8 2.000
anti9 3.000
anti10 4.000
anti11 5.000
anti12 6.000
anti13 7.000
anti14 8.000
Q =~
anti6 0.000
anti7 1.000
anti8 4.000
anti9 9.000
anti10 16.000
anti11 25.000
anti12 36.000
anti13 49.000
anti14 64.000

Covariances:
Estimate Std.Err z-value P(>|z|)
I ~~
L -0.492 0.355 -1.384 0.166
Q 0.049 0.036 1.351 0.177
L ~~
Q -0.036 0.021 -1.710 0.087

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-17

.anti13 0.000
.anti14 0.000
I 1.520 0.107 14.171 0.000
L 0.155 0.061 2.540 0.011
Q -0.011 0.007 -1.433 0.152

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 0.594 0.792 0.750 0.453
.anti7 0.900 0.368 2.446 0.014
.anti8 1.603 0.283 5.662 0.000
.anti9 1.735 0.263 6.605 0.000
.anti10 1.916 0.336 5.709 0.000
.anti11 1.499 0.262 5.713 0.000
.anti12 1.310 0.273 4.801 0.000
.anti13 2.158 0.399 5.411 0.000
.anti14 2.000 0.722 2.770 0.006
I 2.123 0.793 2.677 0.007
L 0.399 0.190 2.097 0.036
Q 0.003 0.003 1.317 0.188

R-Square:
Estimate
anti6 0.781
anti7 0.635
anti8 0.502
anti9 0.530
anti10 0.550
anti11 0.638
anti12 0.678
anti13 0.556
anti14 0.563

ˆ  .011   .003
The mean ( 3 ) and variance ( 33 ) of the quadratic factor were both non-
significant even though the likelihood ratio test was significant. Given the modest LRT and the
non-significant mean and variance of the quadratic factor, we do not regard the quadratic
model as providing a meaningful improvement on the linear model.

Freed Loading (“Fully Latent”) Growth Model: Additional Demonstration

Because the linear LCM fit the data well for the repeated measures of antisocial behavior, we
did not present a freed-loading model in lecture. However, here we demonstrate the
application of this approach in case you would like to apply this to your own data. The fully
latent LCM is an alternative way to capture potentially nonlinear change over time in which we
retain the basic form of the linear growth model but we free some of the loadings associated
with the slope factor. The slope factor is then perhaps better referred to as a “shape” factor, as
estimated loadings that differ from the observed time scores will imply nonlinear trajectories.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-18 | Chapter 8 Latent Growth Curve Analysis

Recall that there are two primary parameterizations of this model, which we refer to in the
course notes as options one and two, respectively. In option one, the first two loadings are
fixed to zero and one to set the scale of the shape factor and the remaining loadings are
estimated to capture potentially nonlinear change over time. In diagram form, this model is
precisely the same as the two-factor linear model except that we estimate a subset of loadings
on the slope factor.
The matrix expression of the model is identical to the linear model, with the sole exception that
the factor loadings for the shape factor are estimated for ages eight through fourteen:

 anti6i  1 0  1i 
 anti7  1 1   
 i     2i 
 anti8i  1 3    3i 
     
 anti9i  1 4   4i
1i   
 anti10i   1 5      5i 
      
 anti11i  1 6   2i   6i 
 anti12i  1 7   7 i 
     
 anti13i  1 8    8i 
 anti14  1 9   
 i   9i 

All else is as defined for the linear model.

The corresponding lavaan script is:
free.1 <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + anti8 + anti9 +
anti10 + anti11 + anti12 + anti13 +
anti14
'
fit.free.1 <- growth(free.1, data=anti, missing="ML")
summary(fit.free.1, fit.measures=TRUE, rsquare=TRUE)

lavTestLRT(fit.free.1, fit.lin)

Notice that the factor loadings for anti8 through anti14 on the factor S are no longer
constrained to fixed values and hence will instead be estimated. Additionally, we requested a
likelihood ratio test to compare the fit of this model to the linear growth model.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-19

Resulting model fit is reported here:

Estimator ML
Model Fit Test Statistic 27.651
Degrees of freedom 33
P-value (Chi-square) 0.731

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2624.462

Loglikelihood unrestricted model (H1) -2610.637

Number of free parameters 21

Akaike (AIC) 5290.925
Bayesian (BIC) 5375.006
Sample-size adjusted Bayesian (BIC) 5308.371

Root Mean Square Error of Approximation:

RMSEA 0.000
90 Percent Confidence Interval 0.000 0.028
P-value RMSEA <= 0.05 1.000

Standardized Root Mean Square Residual:

SRMR 0.103

This is a very good fitting model. Indeed, the model may fit too well. Indications that we may be
over-fitting the data include a model chi-square less than the degrees of freedom (the expected
value for a chi-square) and a RMSEA of zero. Comparing this model to the linear growth model,
we find significantly better fit for the freed loadings model (  2 (7)  25.52, p  .001 ):
Chi Square Difference Test

Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)

fit.free.1 33 5290.9 5375.0 27.651
fit.lin 40 5302.4 5358.5 53.171 25.521 7 0.0006132 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

It’s woth noting, however, that the BIC, which favors parsimony, is actually better for the linear
model (5358.5 versus 5375.0).

Copyright © 2019 Curran-Bauer Analytics, LLC

8-20 | Chapter 8 Latent Growth Curve Analysis

Parameter estimates are shown below:

Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
S =~
anti6 0.000
anti7 1.000
anti8 6.478 11.023 0.588 0.557
anti9 7.010 11.719 0.598 0.550
anti10 9.231 16.021 0.576 0.564
anti11 6.703 11.147 0.601 0.548
anti12 7.017 11.907 0.589 0.556
anti13 13.858 24.615 0.563 0.573
anti14 7.688 12.986 0.592 0.554

Covariances:
Estimate Std.Err z-value P(>|z|)
I ~~
S 0.022 0.047 0.472 0.637

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000
.anti13 0.000
.anti14 0.000
I 1.524 0.112 13.563 0.000
S 0.058 0.108 0.537 0.591

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-21

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 1.815 0.389 4.670 0.000
.anti7 1.272 0.326 3.904 0.000
.anti8 1.491 0.260 5.731 0.000
.anti9 1.757 0.262 6.696 0.000
.anti10 1.916 0.360 5.322 0.000
.anti11 1.878 0.255 7.380 0.000
.anti12 1.617 0.277 5.832 0.000
.anti13 1.109 0.536 2.070 0.038
.anti14 2.169 0.395 5.493 0.000
I 1.024 0.335 3.051 0.002
S 0.013 0.048 0.272 0.786

R-Square:
Estimate
anti6 0.361
anti7 0.459
anti8 0.555
anti9 0.529
anti10 0.571
anti11 0.504
anti12 0.550
anti13 0.789
anti14 0.496

The parameter estimates are also somewhat suggestive of over-fit. Notice that the factor
loadings jump around quite a bit and that both the mean and variance of the slope (or shape)
factor are non-significant. Each factor loading can be interpreted as indicating a multiple for the
amount of change occurring between that time and the first time point relative to the change
observed across the first two time points (e.g., the amount of change between age 6 and age 8
is 6.5 times the amount of change between age 6 and age 7).
The alternative parameterization of the freed loading model, or option two, scales the slope
factor in terms of the first and last time points. The diagram is again the same as that for the
linear model except that we freely estimate the interior loadings on the shape factor:
 anti6i  1 0  1i 
 anti7  1 2   
 i     2i 
 anti8i  1 3   3i 
     
 anti9i  1 4   4i
1i   
 anti10i   1 5      5i 
      
 anti11i  1 6   2i   6i 
 anti12i  1 7   7 i 
     
 anti13i  1 8    8i 
 anti14  1 1   
 i   9i 

Copyright © 2019 Curran-Bauer Analytics, LLC

8-22 | Chapter 8 Latent Growth Curve Analysis

The script for fitting this model is shown here:

free.2 <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + anti7 + anti8 + anti9 +
anti10 + anti11 + anti12 + anti13 +
1*anti14
'
fit.free.2 <- growth(free.2, data=anti, missing="ML")
summary(fit.free.2)

Note that the only change to the model syntax object is the placement of the fixed factor
loadings for the shape factor, S. Additionally, because model fit will be identical to the first
parameterization of this model, we did not request this information again in the summary
function. Nor did we request R-squared statistics or conduct the LRT, as these would also all be
identical.
In contrast, parameter estimates will differ given the change in how the shape factor is scaled,
as seen here:
Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
S =~
anti6 0.000
anti7 0.130 0.220 0.592 0.554
anti8 0.843 0.262 3.211 0.001
anti9 0.912 0.266 3.433 0.001
anti10 1.201 0.341 3.525 0.000
anti11 0.872 0.261 3.336 0.001
anti12 0.913 0.287 3.182 0.001
anti13 1.803 0.521 3.463 0.001
anti14 1.000

Covariances:
Estimate Std.Err z-value P(>|z|)
I ~~
S 0.170 0.287 0.594 0.552

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-23

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000
.anti13 0.000
.anti14 0.000
I 1.524 0.112 13.563 0.000
S 0.447 0.158 2.823 0.005

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 1.815 0.389 4.670 0.000
.anti7 1.272 0.326 3.904 0.000
.anti8 1.491 0.260 5.731 0.000
.anti9 1.757 0.262 6.696 0.000
.anti10 1.916 0.360 5.322 0.000
.anti11 1.878 0.255 7.380 0.000
.anti12 1.617 0.277 5.832 0.000
.anti13 1.109 0.536 2.070 0.038
.anti14 2.169 0.395 5.493 0.000
I 1.024 0.335 3.051 0.002
S 0.773 0.514 1.504 0.133

Each estimated factor loading now represents the proportion of total change that has occurred
at each intermediate time point. For instance, the loading for age 8 indicates that 84.3% of the
change observed between age 6 and age 14 has occurred by age 8. This is a particularly
appealing interpretation if the trajectory is monotonically increasing or decreasing. In the
present case, however, some factor loadings exceed one, leading to some awkward
interpretations. For instance, the change between age 6 and age 13 is 180% of the change from
age 6 to age 14. Again, the bouncing factor loadings are suggestions of problems with the
model. For the most part, the observed means are fluctuating around a relatively stable level
and freeing the factor loadings to capture these fluctuations seems to be over-fitting the data.
We thus move forward with the more parsimonious linear growth model.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-24 | Chapter 8 Latent Growth Curve Analysis

Linear Growth Model Conditional on Gender

It is good practice to determine the functional form of growth before introducing exogenous
predictors of growth. Now that we have established a linear growth model, we want to try to
explain the growth with covariates.
First, we will condition the latent intercept and linear slope factors on gender, a binary time
invariant covariate. The path diagram for this model is:

We return to the measurement model for linear growth:

 anti6i  1 0  1i 
 anti7  1 1   
 i     2i 
 anti8i  1 2  3i 
     
 anti9i  1 3  4i
1i   
 anti10i   1 4       5i 
      
 anti11i  1 5   2 i   6 i 
 anti12i  1 6  7 i 
     
 anti13i  1 7   8i 
 anti14  1   
 i  8  9i 
where
COV  εi   Θ  DIAG 11,22 , ,99 
and the latent variable model is now
1i  1   11   1i 
         geni    
 2i   2   21   2i 

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-25

where
 
COV  ζi   Ψ   11 
 21  22 
The repeated measures have the following model implied mean and covariance structures:
μy (θ) = Λ(α + Γμx )

Σyy (θ) = Λ(ΓΣxxΓ + Ψ)Λ + Θ

The script for this model is:
gen <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
I ~ gen
S ~ gen
'
fit.gen <- growth(gen, data=anti, missing="ML")
summary(fit.gen, fit.measures=TRUE, rsquare=TRUE)

All of this is identical to the unconditional linear model from before with the exception that two
lines have been added, I ~ gen and S ~ gen, to indicate that the growth factors are to be
regressed on gender.
Results are given here:

Estimator ML
Model Fit Test Statistic 58.069
Degrees of freedom 47
P-value (Chi-square) 0.129

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2620.972

Loglikelihood unrestricted model (H1) -2591.937

Number of free parameters 16

Akaike (AIC) 5273.943
Bayesian (BIC) 5338.006
Sample-size adjusted Bayesian (BIC) 5287.236

Copyright © 2019 Curran-Bauer Analytics, LLC

8-26 | Chapter 8 Latent Growth Curve Analysis

Root Mean Square Error of Approximation:

RMSEA 0.024
90 Percent Confidence Interval 0.000 0.043
P-value RMSEA <= 0.05 0.992

Standardized Root Mean Square Residual:

SRMR 0.114

Parameter Estimates:

Information Observed
Observed information based on Hessian
Standard Errors Standard

Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
S =~
anti6 0.000
anti7 1.000
anti8 2.000
anti9 3.000
anti10 4.000
anti11 5.000
anti12 6.000
anti13 7.000
anti14 8.000

Regressions:
Estimate Std.Err z-value P(>|z|)
I ~
gen 0.789 0.166 4.756 0.000
S ~
gen 0.014 0.035 0.391 0.696

Covariances:
Estimate Std.Err z-value P(>|z|)
.I ~~
.S 0.054 0.051 1.045 0.296

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-27

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000
.anti13 0.000
.anti14 0.000
.I 1.224 0.119 10.316 0.000
.S 0.066 0.026 2.584 0.010

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 1.900 0.368 5.164 0.000
.anti7 1.376 0.262 5.255 0.000
.anti8 1.706 0.287 5.939 0.000
.anti9 1.859 0.248 7.498 0.000
.anti10 2.182 0.321 6.787 0.000
.anti11 1.637 0.245 6.684 0.000
.anti12 1.467 0.274 5.350 0.000
.anti13 1.837 0.338 5.430 0.000
.anti14 1.740 0.413 4.213 0.000
.I 0.910 0.283 3.214 0.001
.S 0.025 0.012 2.067 0.039

R-Square:
Estimate
anti6 0.359
anti7 0.467
anti8 0.449
anti9 0.467
anti10 0.468
anti11 0.579
anti12 0.643
anti13 0.627
anti14 0.672
I 0.146
S 0.002

There was significant prediction of the intercept as a function of gender ( ˆ11  .789, p  .001 ),
but not of slope ( ˆ21  .014, p  .696 ). On average, boys reported levels of antisocial behavior
.79-units higher than did girls. On average, girls and boys did not differ in the rates of change in
antisocial behavior over time. We have explained 14.6% of the variance in trajectory
intercepts, but virtually no variance in trajectory slopes.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-28 | Chapter 8 Latent Growth Curve Analysis

It is possible to probe the effect of gen by examining the simple intercept and slope for males
and females separately. This can be accomplished in a variety of ways. Here we demonstrate
how to obtain these tests and plots directly within R.
To compute and test simple intercepts and slopes we can define new parameters to represent
these values within the model syntax object (similar to our use of this feature of lavaan in
Chapter 3):
gen <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
I ~ a1*1 + g11*gen
S ~ a2*1 + g21*gen

#simple intercept and slope for girls

int_girl := a1
slp_girl := a2

#simple intercept and slope for boys

int_boy := a1 + g11
slp_boy := a2 + g21
'
fit.gen <- growth(gen, data=anti, missing="ML")
summary(fit.gen)

We have assigned labels a1, g11, a2, and g21 to the intercepts and slopes of the latent
variable regression equations. We then can compute the simple intercepts and slopes for girls
and boys as shown by using the define operator, :=, to specify that these are linear
combinations of the labeled parameters. In the equations for calculating the simple intercpets
and slopes g11 and g21 are, respectively, multiplied by gen; however, for girls gen is equal to
zero so these parameters simply drop out. That is, a1 and a2 already represent the simple
intercept and slope for girls. For boys, gen is equal to one, so g11 and g21 are simply added in
to each equation.
All of the output is identical to what we have seen before except for the following new bit:
Defined Parameters:
Estimate Std.Err z-value P(>|z|)
int_girl 1.224 0.119 10.316 0.000
slp_girl 0.066 0.026 2.584 0.010
int_boy 2.013 0.117 17.219 0.000
slp_boy 0.080 0.025 3.256 0.001

Given the lack of significance of the effect of gen on trajectory slopes, we would not interpret
the difference in simple slopes between boys and girls, as this does not exceed what one might
expect by chance alone at this sample size.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-29

We can also take these values and generate a plot, similarly to how we generated ICC plots in
Chapter 7 based on model estimates:
plot(c(6,14),c(0,3),type="n",xlab="Age",ylab="Antisocial Behavior")

age <- seq(6, 14, by=1)

pred.girl <- 1.224 + .066*(age-6)
pred.boy <- 2.013 + .080*(age-6)
dat <- data.frame(cbind(age,pred.girl,pred.boy))

lines(dat$age, dat$pred.girl, type="l", lwd=2, lty=1)

lines(dat$age, dat$pred.boy, type="l", lwd=2, lty=2)

title("Simple Trajectories Conditional on Sex")

legend(12, .5, c("girl","boy"), lty=c(1,2), lwd=2)

We first created the frame for the plot with the plot function. Then, we generated age scores
of 6, 7, 8, …, 14. Next, we computed predicted values for girls and boys from the simple
intercept and slope estimates, taking care to subtract six so that the first value of age produces
a factor loading of zero for the slope. We put all of these into a data frame called dat and then
added lines to the plot for girls and boys. Last, we added a title and legend to the plot.

From the plot, we see the significant gender difference in intercept for antisocial behavior, but
not in rate of change over time.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-30 | Chapter 8 Latent Growth Curve Analysis

Linear Growth Model Conditional on Cognitive Stimulation in the Home

Next, we will test whether cognitive stimulation in the home influences baseline ratings of
antisocial behavior for children aged 6, and whether cognitive stimulation in the home
influences rate of linear growth of aggressive behavior for children aged 6-14. Typically, both
gen and homecog would be included in the same conditional growth model. However, we
model them separately for demonstration purposes. The path diagram for the model is:

The measurement model is the same as before but the latent variable model is now:

1i  1   11   1i 
        homecogi    
 2i   2   21   2i 
where
 
COV  ζi   Ψ   11 
 21  22 

These equations imply the following mean and covariance structures for the repeated
measures:

μy (θ) = Λ(α + Γμx )

Σyy (θ) = Λ(ΓΣxxΓ + Ψ)Λ + Θ

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-31

This model syntax object is specified almost identically to the prior conditional growth model,
simply replacing gen with homecog as the regressor for the two latent factors:
home <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
I ~ homecog
S ~ homecog
'
fit.home <- growth(home, data=anti, missing="ML")
summary(fit.home, fit.measures=TRUE, rsquare=TRUE) [int slope];

Model fit is shown below:

Estimator ML
Model Fit Test Statistic 68.329
Degrees of freedom 47
P-value (Chi-square) 0.023

User model versus baseline model:

Comparative Fit Index (CFI) NA

Tucker-Lewis Index (TLI) NA

Loglikelihood and Information Criteria:

Loglikelihood user model (H0) -2626.666

Loglikelihood unrestricted model (H1) -2592.501

Number of free parameters 16

Akaike (AIC) 5285.331
Bayesian (BIC) 5349.394
Sample-size adjusted Bayesian (BIC) 5298.624

Root Mean Square Error of Approximation:

RMSEA 0.033
90 Percent Confidence Interval 0.013 0.050
P-value RMSEA <= 0.05 0.951

Standardized Root Mean Square Residual:

SRMR 0.116

The model fits the data adequately well, although not as well as it fit when the growth factors
were conditional on gen. These models are not formally nested, so it is not possible to conduct
a chi-square difference test between the two conditional models.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-32 | Chapter 8 Latent Growth Curve Analysis

Parameter estimates are shown below:

Latent Variables:
Estimate Std.Err z-value P(>|z|)
I =~
anti6 1.000
anti7 1.000
anti8 1.000
anti9 1.000
anti10 1.000
anti11 1.000
anti12 1.000
anti13 1.000
anti14 1.000
S =~
anti6 0.000
anti7 1.000
anti8 2.000
anti9 3.000
anti10 4.000
anti11 5.000
anti12 6.000
anti13 7.000
anti14 8.000

Regressions:
Estimate Std.Err z-value P(>|z|)
I ~
homecog -0.078 0.033 -2.358 0.018
S ~
homecog -0.015 0.007 -2.204 0.028

Covariances:
Estimate Std.Err z-value P(>|z|)
.I ~~
.S 0.044 0.051 0.846 0.398

Intercepts:
Estimate Std.Err z-value P(>|z|)
.anti6 0.000
.anti7 0.000
.anti8 0.000
.anti9 0.000
.anti10 0.000
.anti11 0.000
.anti12 0.000
.anti13 0.000
.anti14 0.000
.I 1.622 0.085 19.041 0.000
.S 0.075 0.018 4.189 0.000

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-33

Variances:
Estimate Std.Err z-value P(>|z|)
.anti6 1.876 0.371 5.061 0.000
.anti7 1.351 0.263 5.126 0.000
.anti8 1.646 0.286 5.761 0.000
.anti9 1.835 0.247 7.436 0.000
.anti10 2.233 0.328 6.810 0.000
.anti11 1.746 0.256 6.825 0.000
.anti12 1.453 0.270 5.380 0.000
.anti13 1.728 0.329 5.244 0.000
.anti14 1.716 0.407 4.219 0.000
.I 1.053 0.289 3.648 0.000
.S 0.025 0.012 2.097 0.036

R-Square:
Estimate
anti6 0.368
anti7 0.475
anti8 0.461
anti9 0.472
anti10 0.464
anti11 0.566
anti12 0.648
anti13 0.644
anti14 0.678
I 0.037
S 0.058

There was significant prediction of the intercept as a function of homecog

( ˆ11  .078, p  .018 ), as well as the slope ( ˆ21  .015, p  .028 ). On average, children
reported initial levels of antisocial behavior .08-units lower per one-unit increase in cognitive
stimulation. On average, children reported trajectories with slopes .02-units less per one-unit
increase in stimulation. The predictor homecog explains 3.7% of the variance in trajectory
intercepts and 5.8% of the variance in trajectory slopes.
We must further probe these effects to understand the precise nature of these relations better.
To do so, we will calculate and plot simple intercepts and slopes at plus and minus one standard
deviation ( sd  2.58 ) from the mean of homecog (where the mean is equal to zero because it
was mean-centered).

Copyright © 2019 Curran-Bauer Analytics, LLC

8-34 | Chapter 8 Latent Growth Curve Analysis

To compute and test the simple intercepts and slopes as a function of homecog we will again
simply modify the model syntax object, labeling the relevant parameters and using the define
operator to specify the quantities of interest:
home <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
I ~ a1*1 + g11*homecog
S ~ a2*1 + g21*homecog

#simple intercept and slope for low homecog

int_lo := a1 - 2.58*g11
slp_lo := a2 - 2.58*g21

#simple intercept and slope for medium homecog

int_md := a1
slp_md := a2

#simple intercept and slope for high homecog

int_hi := a1 + 2.58*g11
slp_hi := a2 + 2.58*g21
'
fit.home <- growth(home, data=anti, missing="ML")
summary(fit.home)

The new output is shown here:

Defined Parameters:
Estimate Std.Err z-value P(>|z|)
int_lo 1.823 0.121 15.020 0.000
slp_lo 0.114 0.026 4.460 0.000
int_md 1.622 0.085 19.041 0.000
slp_md 0.075 0.018 4.189 0.000
int_hi 1.421 0.120 11.863 0.000
slp_hi 0.035 0.025 1.410 0.158

Notice that the simple slope is significantly positive at low and medium levels of home cognitive
support, slp_lo and slp_md, respectively, but is not significantly different from zero when
home cognitive support is high, or slp_hi.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Latent Growth Curve Analysis of Antisocial Behavior | 8-35

We can also plot the simple trajectories based on these intercepts and slopes much as we did
previously for gender.
plot(c(6,14),c(0,3),type="n",xlab="Age",ylab="Antisocial Behavior")

age <- seq(6, 14, by=1)

pred.lo <- 1.823 + .114*(age-6)
pred.md <- 1.622 + .075*(age-6)
pred.hi <- 1.421 + .035*(age-6)
dat <- data.frame(cbind(age,pred.lo,pred.md,pred.hi))

lines(dat$age, dat$pred.lo, type="l", lwd=2, lty=2)

lines(dat$age, dat$pred.md, type="l", lwd=2, lty=1)
lines(dat$age, dat$pred.hi, type="l", lwd=2, lty=3)

title("Simple Trajectories Conditional on Cognitive Support in the Home")

legend(12, .75, c("low","medium", "high"), title="Cognitive Support",
lty=c(2,1,3), lwd=2)

The plot clearly conveys that the slope is relatively flat at high homecog but becomes steeper
as homecog decreases to medium and low levels.
These models could be expanded to include the simultaneous effects of gender and cognitive
stimulation, or even the interaction between the two, but we do not pursue this further here.

Copyright © 2019 Curran-Bauer Analytics, LLC

8-36 | Chapter 8 Latent Growth Curve Analysis

Copyright © 2019 Curran-Bauer Analytics, LLC

Appendix A: Generating
Path Diagrams in R
Example 1: Holzinger-Swineford CFA (Chapter 4).................................................................... A-3
Example 2: Conditional Latent Curve Model (Chapter 8) ........................................................ A-6

This appendix presents two fully worked examples drawn from class demonstrating the
creation of path diagrams from lavaan fit objects using the semPlot package developed by
Sacha Epskamp. There are likely other packages that could be used for the same purpose.
Some software programs also offer the ability to draw a path diagram and infer the model
specification syntax from the diagram. For instance, the free stand-alone program nyx can
produce lavaan script based on a user-specified diagram (see http://onyx.brandmaier.de).
A-2 | Appendix A: Generating Path Diagrams in R

Copyright © 2019 Curran-Bauer Analytics, LLC

 Example 1: Holzinger-Swineford CFA (Chapter 4) | A-3

Example 1: Holzinger-Swineford CFA (Chapter 4)

The data for this demonstration were provided by Holzinger & Swineford in their 1939
monograph A Study in Factor Analysis: The Stability of a Bi-Factor Solution. The sample includes
301 7th and 8th grade students, between 11-16 years of age, drawn from two schools. The
data is in the text file hs.dat and the script file is app_ex1.R. The variables in the data set
that we will use are
visperc visual perception test in which participants select the next image in a series
cubes visual perception test in which participants must mentally rotate a cube
lozenges visual perception test involving mental “flipping” of a parallelogram (“lozenge”)
parcomp paragraph comprehension test
sencomp sentence completion task in which participants select most appropriate word to
put at the end of a sentence
wordmean verbal ability test in which participants must select a word most similar in
meaning to a word used in a sentence.
addition participants have 2 minutes to complete as many 2-number addition problems
as they can
countdot participants have 4 minutes to count the number of dots in each of a series of
dot pictures
sccaps participants have 3 minutes to indicate whether capital letters are composed
entirely of straight lines or include curved lines.
Other variables in the data not included in the models fit here are school, female, age, and
month. We can read in the data with the following:
hs <- read.table("hs.dat", header=FALSE)
names(hs) <- c("school", "female", "age", "month",
"visperc", "cubes", "lozenges",
"parcomp", "sencomp", "wordmean",
"addition", "countdot", "sccaps")
hs$addition <- hs$addition/4
hs$countdot <- hs$countdot/4
hs$sccaps <- hs$sccaps/4

As described in Chapter 4, we rescaled the last three variables to give them a similar range to
the other indicators.
We will also load the two packages we need (install semPlot first if you have not already):
library(lavaan)
library(semPlot)

The initial model we fit to this data posited three factors, visual (with indicators visperc,
cubes and lozenges), verbal (with indicators parcomp, sencomp, and wordmean), and
speed (with indicators addition, countdot, and sccaps). We scaled the latent variables by

Copyright © 2019 Curran-Bauer Analytics, LLC

A-4 | Appendix A: Generating Path Diagrams in R

setting their means to zero and variances to one, permitting all factor loadings and intercepts to
be freely estimated. The corresponding script for specifying and fitting this model in lavaan is
cfa.1a <-'#factor loadings
visual =~ visperc + cubes + lozenges
verbal =~ parcomp + sencomp + wordmean
speed =~ addition + countdot + sccaps
'
fit.1a <- cfa(cfa.1a, data=hs, meanstructure=TRUE, std.lv=TRUE)

We can now use the fit object fit.1a to generate a path diagram.
We first show the default diagram generated by the semPaths function:
semPaths(fit.1a)

Clearly, some aspects of this plot can be improved upon. The triangles used to convey
intercepts create a great deal of clutter and the covariances and variances among the latent
variables are overlapping. Further, we might like to display the model estimates on the diagram
for ease of interpretation.
Fortunately, there are many options for modifying the layout and information provided in the
plot. Fiddling with these a bit, we invoked the following options:
semPaths(fit.1a, whatLabels="est", intercepts="FALSE", curve=1.75)

With these commands we will display the estimates (unstandardized) on the paths, remove the
triangles for the intercepts, and increase the curvature of the double-headed arrows for the
covariances so that they no longer overlap with the variances.

Copyright © 2019 Curran-Bauer Analytics, LLC

 Example 1: Holzinger-Swineford CFA (Chapter 4) | A-5

The resulting plot is shown here:

We can also switch the path labels to standardized estimates using the following options:
semPaths(fit.1a, whatLabels="std", intercepts="FALSE", curve=1.75)

Because the factors were already standardized, the displayed values for the latent variable
variances and covariances are unchanged, but we now see standardized factor loadings and
residuals (equal to one minus the communality for the indicator).
There are many other ways to modify the plot, as described in the documentation for the
package.

Copyright © 2019 Curran-Bauer Analytics, LLC

A-6 | Appendix A: Generating Path Diagrams in R

Example 2: Conditional Latent Curve Model (Chapter 8)

The data for this demonstration are from the National Longitudinal Study of Youth. The sample
includes 405 children who were aged 6-8 at the first assessment (in 1986). Children were
assessed bi-yearly (until 1992), resulting in up to four measurement occasions per child. Data
are in the file anti.dat and the script can be found in app_ex2.R. The variables in the data
set that we will use are described below:
id participant identification; this will not be used in the analysis
gen 0 = female; 1 = male
homecog level of cognitive support in the home (mean-centered); higher values
denote more cognitive support in the home
anti6-anti14 repeated assessments of mother reports of child antisocial behavior
(higher values indicate more antisocial behavior); post-script denotes
child age at time of assessment
We are interested in characterizing the functional form of developmental change in aggressive
behavior for children aged 6-14. We would also like to understand how our time invariant
covariates (gender and cognitive support in the home) contribute to initial levels of, and change
in, aggressive behavior.
To read in the data, we use the following script:
classes <- rep("numeric",12)
anti <- read.table("anti.dat",header=FALSE,colClasses=classes,na.strings=".")
y <- paste("anti", 6:14, sep = "")
names(anti) <- c(c("id","gen","homecog"),y)

We then load the two packages we need:

library(lavaan)
library(semPlot)

For this example, we will fit a linear growth model including both gen and homecog as
predictors (in Chapter 8, we considered these predictors separately, but here we include them
jointly).
The model is specified and fit as follows:
lcm <- 'I =~ 1*anti6 + 1*anti7 + 1*anti8 + 1*anti9 +
1*anti10 + 1*anti11 + 1*anti12 + 1*anti13 +
1*anti14
S =~ 0*anti6 + 1*anti7 + 2*anti8 + 3*anti9 +
4*anti10 + 5*anti11 + 6*anti12 + 7*anti13 +
8*anti14
I ~ gen + homecog
S ~ gen + homecog
'
fit.lcm <- growth(lcm, data=anti, missing="ML")
summary(fit.lcm)

Copyright © 2019 Curran-Bauer Analytics, LLC

 Example 2: Conditional Latent Curve Model (Chapter 8) | A-7

The results of primary interest, regarding the prediction of the growth factors, are shown here:
Regressions:
Estimate Std.Err z-value P(>|z|)
I ~
gen 0.768 0.165 4.664 0.000
homecog -0.073 0.032 -2.263 0.024
S ~
gen 0.015 0.035 0.443 0.658
homecog -0.015 0.007 -2.225 0.026

As in analyses which considered these predictors separately…

 Boys have higher intercepts, indicating higher levels of antisocial behavior at age six.
 There is no sex difference in slopes, such that the greater antisocial behavior of boys is
maintained throughout the age range (through age 14).
 Kids from homes with higher cognitive support have both lower intercepts, i.e., lower
antisocial behavior at age 6, and less steep slopes, i.e., increase in antisocial behavior
less over time.
We can diagram these results as follows:
semPaths(fit.lcm, whatLabels="est", intercepts="FALSE")

Paths that are not estimated as part of the model fitting process are dashed (this includes fixed
paths as well as variance/covariance parameters for exogenous predictors which are not
explicitly estimated).
Again, this plot could be modified in a variety of ways that we do not pursue here.

Copyright © 2019 Curran-Bauer Analytics, LLC

A-8 | Appendix A: Generating Path Diagrams in R

Copyright © 2019 Curran-Bauer Analytics, LLC