Chapter 5

Structural Equation Models

with Latent Variables
5.1 Introduction to Structural Equation Models.............................................................................. 5-3
5.2 Fitting Structural Equation Models .......................................................................................... 5-11
5.3 Example SEM............................................................................................................................ 5-27
5.4 Equivalent Models ................................................................................................................... 5-43
5-2 | Chapter 5 Structural Equation Models with Latent Variables

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.1 Introduction to Structural Equation Models 5-3

5.1 Introduction to Structural Equation Models

Objectives
 Introduce SEM as a flexible combination of path analysis and CFA
 Review recent SEMs in the literature

Path Analysis Models

 With path analysis (or simultaneous equations) models we could fit
complex structural models for observed variables
1 2

x1 y1 y2

 Here, we have a relatively simple path analysis model where the

effect of x is mediated by y1.
 Although widely used, we have shown that fitting this model will
almost certainly result in biased effect estimates due to failure to
account for measurement error in the observed variables

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Confirmatory Factor Analysis

 In factor analysis, we inferred the presence of underlying, error-
free latent variables from correlated observed variables

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

 Yet here we have not tested any structural model of interest

Here the covariance matrix among the factors is saturated, including non-directional associations among
all latent variables.

Structural Equation Modeling

 In structural equation models, we combine the structural model
of path analysis with the measurement model of CFA
2 3

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

 We are now estimating structural relations among latent

variables that are unbiased by measurement error

Here a causal structure has been applied to the covariance matrix among the factors

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.1 Introduction to Structural Equation Models 5-5

Fixed-x, Exogenous predictors

 We can also include manifest fixed regressors, or exogenous
predictors, in our models, as in the model below

x1 1 2

x2  

y1 y2 y3 y4

1 2 3 4

Fixed-x, Exogenous predictors

 As with the simultaneous equation model, we make no
distributional assumptions about the exogenous predictors
 Exogenous predictors can even be binary
 Gender
 Child of an Alcoholic
 As in simultaneous equation models, however, exogenous
predictors are still assumed to be measured without error. If this
is not true, effect estimates may be biased.
 Preferable to use multiple indicator latent factors as predictors
whenever measurement error is likely

As with path analysis / simultaneous equation models, we can include nominal exogenous predictors via
coding variables.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Models with Manifest Outcomes

 In some models, we may also have manifest variables as outcome
variables, as in the model shown below.

1 y1

2 y2 1

3 y3 y6 3

y7 4
4 y4
2
5 y5

Models with Manifest Outcomes

 Like simultaneous equation models, manifest outcomes are
assumed to be free of measurement error
 Unlike measurement error in predictors, measurement error in
outcomes does not bias the structural paths leading to the outcome.
 It does, however, cause under-estimation of the amount of variance
explained in the outcome, and inflation of the standard error of the
structural paths leading to the outcome (resulting in lower power).
 Thus, given sufficient sample size, still best to use full SEM including
multiple indicator latent outcomes instead of manifest outcomes when
measurement error is likely.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.1 Introduction to Structural Equation Models 5-7

Example Applications
 Hwang & Woods (2009) used a relatively simple SEM to show
that the mental health status of Asian and Latino American
college students is predicted by the acculturation gap with their
parents, as mediated by family conflict.

Hwang W.-C. & Wood, J.J. (2009). Acculturative family distancing: links with self-reported
symptomatology among Asian Americans and Latinos. Child Psychiatry and Human Development, 40,
123–138.

Abstract from manuscript: Objective Our knowledge of how acculturative processes affect families
remains quite limited. This article tests whether acculturative family distancing (AFD), a more proximal
and problem-oriented measure of the acculturation gap, influences the mental health status of Asian
American and Latino college students. AFD occurs along two dimensions: communication difficulties and
cultural value incongruence. Methods Data were collected from 186 Asian American (n = 107) and Latino
(n = 79) undergraduates, who provided self-reports on psychological problems, depressive symptoms,
and family conflict. A new self-report measure of AFD evidencing good psychometric properties was
used to test hypothesized relations among these variables in structural equation models (SEM). Results
For both Asian American and Latinos, results indicated that higher levels of AFD were associated with
higher psychological distress and greater risk for clinical depression, and that family conflict mediated
this relation. Conclusion AFD processes were associated with the mental health of students and the
functioning of their families. These findings highlight potential foci to address in prevention and
intervention programs, such as improving communication and teaching families how to negotiate
cultural value differences.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Example Applications
 Vieno et al. (2010) evaluated the indirect effect of neighborhood
context on children’s antisocial behavior

Vieno, A., Nation, M., Perkins, D.D., Pastore, M. & Santinello, M. (2010). Social capital, safety
concerns, parenting, and early adolescents’ antisocial behavior. Journal of Community Psychology, 38,
314-328.

Abstract from manuscript: This study explores the relations between neighborhood social capital
(neighbor support and social climate), safety concerns (fear of crime and concern for one’s child),
parenting (solicitation and support), and adolescent antisocial behavior in a sample of 952 parents (742
mothers) and 588 boys and 559 girls from five middle schools (sixth through eighth grades) in a midsize
Italian city. In structural equation models, social capital is strongly and inversely related to safety
concerns and positively related to parental support and solicitation. In turn, safety concerns are also
positively related to parental support and solicitation. Social capital and safety concerns have indirect
effects on children’s antisocial behavior through their effects on parenting. Implications are discussed
for parenting and community-based interventions to prevent or reduce youth antisocial behaviors

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.1 Introduction to Structural Equation Models 5-9

Example Applications
 Park, Heppner &
Lee (2010)
evaluated whether
maladaptive coping
and self-esteem
mediate the
relationship
between
perfectionism and
distress

Park, H.-J., Heppner, P.P. & Lee, D.-G. (2010). Maladaptive coping and self-esteem as mediators
between perfectionism and psychological distress. Personality and Individual Differences 48, 469–474.

Abstract from manuscript: This study with 508 Korean college students examined the mediation effects
of maladaptive coping styles and self-esteem on the links of evaluative concerns perfectionism and
psychological distress. Structural equation modeling analyses supported a full mediation effect of
maladaptive coping between evaluative concerns perfectionism and distress. The final model also
revealed a significant path from evaluative concerns perfectionism through maladaptive coping and self-
esteem to distress. Furthermore, a multi-group analysis found that male college students with
evaluative concerns perfectionism tend to use maladaptive coping strategies more compared to their
female counterparts. The findings provided not only external validity for the full mediation effect of
coping but also evidence of more complex relations among the variables.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Summary
 Structural equation models represent the natural combination of
simultaneous equation models with confirmatory factor analysis
 The full model is very flexible, permitting complex structural
models involving both observed and latent predictors and
outcomes
 The presence of the measurement model permits the estimation
of effects without bias or standard error inflation due to
measurement error
 We will now see that fitting SEMs is also a straightforward
extension of procedures we have covered for simultaneous
equation models and CFA

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-11

5.2 Fitting Structural Equation Models

Objectives
 Show specification of SEMs in path diagram and matrix form
 Describe model identification rules
 Describe process of model estimation, evaluation and re-
specification

Fitting SEMs
 As with simultaneous equation models and CFA, model fitting
involves the following steps
 Specification
 Identification
 Estimation
 Evaluation
 Potential re-specification
 Interpretation
 Fortunately, we can draw on the knowledge base we have built
for simultaneous equation models and CFA to develop general
principles for the full structural equation model with latent
variables.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Model Specification
 As with prior models, the initial specification of the model should
be heavily guided by theory.
 Which variables are predictors, mediators, and outcomes?
 Which variables have causal effects, and which are merely associated?
 The model can be specified graphically via a path diagram, or in
equations using an expanded set of model matrices.
 Our model matrices now combine matrices used previously with
simultaneous equation models and matrices used with CFA.
 The full SEM consists of two parts:
 The measurement model (like CFA)
 The structural model (like path analysis)

Model Specification
 The measurement model is used to define the relationships of the
indicator variables to the latent variables:

yi  ν  Ληi  εi , where VAR(εi )  Θ

This is identical to the measurement model of CFA

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-13

Model Specification
 The structural model is used to define the relationships among the
latent variables and between the latent variables and any fixed-x
exogenous predictors:

ηi  α  Βηi  Γxi  ζi , where VAR(ζi )  Ψ

This is identical to the structural model from path analysis, but

latent variables have replaced observed variables as outcomes
 Compare to
Regression: yi    γxi   i
Path Analysis: y i  α  Βy i  Γxi  ζi

Notice that the structural model is the same as path analysis, except that the latent variables defined by
the measurement model have replaced the observed variables.

Model Specification: Example

 Consider specification of the following model

2 3

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Model Specification: Example

 Using scaling indicators to set the metric of the latent variables,
the measurement model is
 y1i   0   1 0 0  1i 
 y   ν     
 2i   2   21 0 0  2i 
 y3i   0   0 1 0   3i 
      1i   
 y4i    ν 4    0 42 0     4i 
 
 y5i   0   0 0 1   2 i   5 i 
         
 y6i   ν 6   0 0 63   3i   6i 
 y  ν   0 0 73   
 7i   7     7i 
 y8i   ν8   0 0 83   8i 
where VAR(εi )  Θ  DIAG 11 ,22 , ,88 

We’ll talk more about scaling latent variables in SEM shortly. For now we just use scaling items.

Model Specification: Example

 The structural model is specified as

1i  1   0 0 0 1i    1i 

        0 0 2i    2i 
 2i   2   21
3i   3   0 32 0 3i   3i 

where
 11 

VAR(ζi )  Ψ   0  22 

 0 0  33 

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-15

Model Specification: Example 2

 Let us now consider the example model with exogenous x’s:

x1 1 2

x2  

y1 y2 y3 y4

1 2 3 4

Model Specification: Example 2

 Using scaling indicators to set the metric of the latent variables,
the measurement model is

 y1i   0   1 0  1i 
 y   ν      
 2i    2    21 0   1i    2i 
 y3i   0   0 1  2i   3i 
       
 y4i   ν 4   0 42   4i 

where

VAR(εi )  Θ  DIAG 11 ,22 ,33 ,44 

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Model Specification: Example 2

 The structural model is specified as

1i  1   0 0 1i   11  12   x1i   1i 

               x    
   2   21
2 i 0     21
2 i 0   2i   2i 

where

 
VAR(ζi )  Ψ   11 
 0  22 

Model Specification: Example 3

 Recall that our third example model includes two manifest
outcomes:

1 y1

2 y2 1

3 y3 y6 3

y7 4
4 y4
2
5 y5

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-17

Model Specification: Example 3

 Using scaling indicators to set the metric of the latent variables,
the measurement model for the 5 indicator variables is

 y1i   0   1 0  1i 
 y   ν     
 2i   2   21 0  1i   2i 
 y3i    ν 3    31 0      3i 
        
 y4 i   0   0 1   2 i   4 i 
 y5i   ν 5   0 52   5i 

where
VAR(εi )  Θ  DIAG 11 ,22 ,33 ,44 ,55 

The variables y6 and y7 do not appear here as they are not indicators of a latent factor. They will appear
in the structural model instead as, essentially, “manifest factors”.

Model Specification: Example 3

 The structural model is specified as

1i  1   0 0 0 0  1i   1i 

     0 0 0 0  2i   2i 
 2i    2    
 y6i   3   31 32 0 0  y6i   3i 
        
 y7i   4    41  42 0 0  y7i   4i 

where
 11 
 
 21  22 
VAR(ζi )  Ψ 
 0 0  33 
 
 0 0  43  44 

Here we see y6 and y7 alongside the latent variables. The subscript numbering on the regression
coefficients and disturbances is a little unfortunate because it doesn’t match the numbering of the y
variables. Such is life.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Model-Implied Moment Structure

 Recall that the measurement and structural models of the full
SEM are written as

yi  ν  Ληi  εi , VAR(εi )  Θ

ηi  α  Βηi  Γxi  ζi , VAR(ζi )  Ψ

 Together, these imply a specific moment structure. The mean

structure is

μ y (θ)   ν  Λμ η   ν  Λ  I  B  (α  Γμ x ) 
1

μ(θ)     

 μ x   μ x   μx 

As in the previous models we have seen, the mean structure will often be saturated, with as many
parameters estimated as there are observed means.

Model-Implied Moment Structure

 They also imply the covariance structure

 Σyy (θ) 
Σ(θ)   
 Σ xy (θ) Σ xx 
where

Σyy (θ)  ΛΣ ηη Λ  Θ

 Λ I  B
1
 ΓΣxx Γ  Ψ  I  B   Λ  Θ
1

and
Σxy (θ)  Σxx Γ  I  B   Λ
1

Notice that the expression Σyy (θ)  ΛΣηη Λ  Θ is of the same form as the CFA model. The difference
in the SEM is that we also apply a path analytic structure to Σ ηη , the covariance matrix of the factors.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-19

Model-Implied Moment Structure

 Putting all the pieces together, we have the following
model-implied moment structure

 ν  Λ  I  B 1 (α  Γμ x ) 
μ(θ)   
 μx 

 Λ I  B 1 ΓΣ Γ  Ψ I  B 1 Λ  Θ 
    xx   
Σ(θ) 
 1 
 Σ xx Γ  I  B  Λ Σ xx 

There is a lot going on here, but the important thing is just to realize that the model implies a specific
structure for the means and covariances of the data, and one of our key objectives is to see how well
this structure matches the observed means and covariances.

Model Identification
 Like CFA, one requirement for model identification is that the
scale of the latent variables be defined through specific
restrictions in the model.
 As before, one way to set the scale of the latent factors is by
fixing n and  to zero and one for a scaling indicator, respectively
 This is what we did when specifying the preceding models
 Alternatively, we can set the scale of the latent variables by fixing
 and  to zero and one, respectively.
 Unlike CFA, this is not always equivalent to putting the factor on a
standard normal scale.
 If the factor is predicted by other variables in the model, then we are
actually standardizing the factor disturbances, and the total variance of
the factor will exceed one

One can put the factors on a standard normal scale post-estimation by requesting the standardized
solution. We will see this in our example later.

Copyright © 2014 Curran-Bauer Analytics, LLC

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

Model Identification
 Aside from assigning a metric to the latent variables, other
restrictions are also necessary to identify the model.
 e.g., common to assume simple structure and local independence of
items (but not required)
 Model identification can be determined by
 Matrix algebra (hard)
 t-rule (necessary but not sufficient)
 Two-step rule (sufficient but not necessary)

t-Rule
 As before, the t-rule simply stipulates that there should be more
observed means and unique (co)variances than estimated
parameters.
 More formally, if t is the number of freely estimated parameters in
 then a necessary (but not sufficient) condition for identification
is
t   p  q  p  q  1 2   p  q 
where p and q are the number of y and x variables, respectively
 Because  x and Σ xx are always saturated, some software programs (e.g.,
Mplus) do not count the elements in these matrices as parameters to be
estimated
 In assessing the t-rule, however, one should include unique elements in
 x and Σ xx when counting the number of parameters t.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-21

Two-Step Rule
 Step 1: Respecify the model as a confirmatory factor analysis
with all possible associations between factors. Determine the
identification of the measurement model as you would do for a
CFA.
 Step 2: Treat all latent variables within the structural model as if
they were observed variables. Determine the identification of
the structural model as you would do for a simultaneous
equation model for observed variables.
 Passing the two-step rule is sufficient to ensure model
identification, but not necessary (i.e., some identified models fail
the two-step rule).

Example of Two-Step Rule

 Consider the following model

2 3

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

Copyright © 2014 Curran-Bauer Analytics, LLC

5-22 | Chapter 5 Structural Equation Models with Latent Variables

Example of Two-Step Rule

 The two-step rule involves first respecifying the model as a CFA
with all possible associations between factors

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

 This model is identified by the two-indicator rule.

Example of Two-Step Rule

 Second, we evaluate the structural model as we would a
simultaneous equation model for observed variables
1 2

x y y
(1)  

 The structural model is identified by the recursive rule (the B

matrix can be written as a lower triangular matrix and the Ψ
matrix is diagonal)

 0 0  
B  Ψ   11 
  21 0   0  22 

Note that the first latent variable is now treated as if it were a fixed-x exogenous predictor, so the B and
 matrices are only 2 x 2 for the simultaneous equation model (whereas they are 3 x 3 for the SEM as
actually fit to the data).

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-23

Example of Two-Step Rule

 Because this model satisfies the two-step rule, that is sufficient to
say that the model is identified
2 3

  

y1 y2 y3 y4 y5 y6 y7 y8

1 2 3 4 5 6 7 8

 Empirical under-identification is still possible (e.g., if certain

structural paths or factor loadings approach values of zero)

Model Estimation
 Given identification of model, we can now turn to estimation
 As with simultaneous equation models and CFA, multiple
estimators exist for SEMs
 Maximum likelihood
 Ordinary least squares
 Generalized least squares
 Weighted least squares
 As before, we will concentrate on maximum likelihood
 The only significance difference from path analysis and CFA is in the
computation of the model-implied means and covariances

Copyright © 2014 Curran-Bauer Analytics, LLC

5-24 | Chapter 5 Structural Equation Models with Latent Variables

Model Estimation
 The normal-theory ML fitting function is again

2  θ    ni ln  2   ln | Σi (θ) |   z i  μi (θ)  Σi1 (θ)  z i  μi (θ)  

N

i 1
 

 where zi   yi xi  , i.e., yi is stacked on xi.

 The i subscripts on n, S, and  are there to allow for missing data.
 As before, an equivalent ML fitting function can be written in
terms of sufficient summary statistics only
FML (θ)  ln | Σ(θ) |  tr SΣ1 (θ)   ln | S |   p  q  

m  μ(θ) Σ1 (θ) m  μ(θ)

 Assumes complete data

ni is the number of observed values in zi for case i.

Evaluation of Parameter Estimates

 Interpretation of the measurement model is like CFA, and
interpretation of the structural model is like path analysis
 Can again compute standardized parameter estimates to aid
evaluation.
 For factor loadings and causal structural paths this is akin to computing
standardized regression coefficients
 For covariance parameters, rescaling to correlation metric
 For indicator residual variances, rescaled to variance unexplained
 Can compute communality, or h2, for indicators
 Can compute variance explained, or R2, for latent or manifest
outcomes

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.2 Fitting Structural Equation Models 5-25

Model Respecification
 When the model fails to fit, modification indices may suggest
avenues for improvement.
 As always, one must treat these with some skepticism, as use of MIs is
exploratory and model fit can be improved by capitalizing on chance.
 MIs indicate expected improvement in fit for freeing a single parameter
at a time, as opposed to larger scale model misspecifications.

Model Respecification
 Another strategy for locating model misfit is to isolate problems
to the structural or measurement model.
 Useful only if the structural model is not saturated.
 Similar in concept to the two-step rule for identification.
 Overall model c2 can be decomposed into additive components due to
misfit of the measurement model + misfit of the structural model.
 The component for the measurement model can be obtained by fitting a
CFA respecification of the model with a saturated  matrix. The c2
obtained for this model is the misfit due to the measurement model.
 Poor fit indicates a problem with the measurement model.
 The difference in c2 between the CFA and original SEM is the misfit due
to the structural model. Can evaluate by LRT, given these are nested
models (Anderson & Gerbing, 1988).
 A significant c2 difference test indicates a problem with the structural model

Copyright © 2014 Curran-Bauer Analytics, LLC

5-26 | Chapter 5 Structural Equation Models with Latent Variables

Summary
 In sum, the specification, identification, and process of fitting and
interpreting SEMs all follow straightforwardly from procedures
for simultaneous equation models and CFA models.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-27

5.3 Example SEM

Objectives
 Provide a real-data example of the process of fitting, evaluating,
re-specifying, and interpreting an SEM

Copyright © 2014 Curran-Bauer Analytics, LLC

5-28 | Chapter 5 Structural Equation Models with Latent Variables

Example: Posttraumatic Growth

 Senol-Durak & Ayvasik (2010) wished to determine what factors
are associated with posttraumatic growth in spouses of
myocardial infarction patients.
 Sample size of 132.
 Hypotheses were that
 Environmental Resources (ER, i.e., social support) and Individual
Resources (IR) influence PostTraumatic Growth (PTG)
 ER and IR have direct effects on PTG.
 ER and IR also have indirect effects on PTG.
 ER and IR influence Event-Related Factors (ERF)
 ERF influences Cognitive Process Coping (Coping)
 Cope influences PTG

It is worth noting that the sample size is a bit low, especially for the complexity of the fitted model, but
this is not uncommon for published applications of SEM.

The full reference is:

Senol-Durak, E. & Ayvasik, H.B. (2010). Factors associated with posttraumatic growth among the
spouses of myocardial infarction patients. Journal of Health Psychology, 15, 5–95.

Abstract from manuscript: To clarify the rationale behind Posttraumatic Growth (PTG), a model by
Schaefer and Moos describes the relative contribution of environmental resources, individual resources,
event related factors, cognitive processing and coping (CPC) on PTG. In the present study, this model
was tested with the spouses of myocardial infarction patients with data from various hospitals in Turkey.
A structural equation model revealed that neither individual nor environmental resources had indirect
effects on PTG through the effect of event-related factors and CPC, while they showed direct effects on
PTG. The findings were discussed in the context of the theoretical model.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-29

Example: Multiple Indicator Latent Factors

 Environmental Resources
 Three indicators: Social support from family (fa), friends (fr), significant
others (si)
 Individual Resources
 Five indicators: psychological hardiness indicators: commitment (comt),
control (con), challenge (cha); self esteem scale scores (es), and locus of
control scale scores (lo)
 Event-Related Factors
 Three indicators: subjective evaluations of prognosis (pro), threat to
future health (th), and time since diagnosis (time)

One could probably make a case that the indicators for Event-Related Factors should be causal
indicators rather than effect indicators. The authors used a reflective measurement model for all
factors, however, and we remain consistent with their analysis.

Example: Multiple Indicator Latent Factors

 Cognitive process coping
 Six indicators: emotion focused coping (em), indirect coping (in),
rumination (ru), avoidance (av), hypervigilence (hy), and religious belief
(rb)
 Posttraumatic growth
 Five indicators, subscale scores from postraumatic growth inventory:
improved relationship (ptgi1), new possibilities for one’s life (ptgi2),
greater appreciation of life (ptgi3), greater sense of personal strength
(ptgi4), spiritual development (ptgi5).

Copyright © 2014 Curran-Bauer Analytics, LLC

5-30 | Chapter 5 Structural Equation Models with Latent Variables

Example: The Model

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

Example: The Model

 Like many SEM’s, the model is rather large and the matrix
representation of the model is unwieldy.
 The path diagram, however, nicely conveys the proposed
structure and is sufficient to infer the matrix representation of
the model.
 We thus forgo showing the matrix representation here – this is provided
in the text

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-31

The measurement model is:

 fai   n 1i   11 0 0 0 0   1i 

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

Copyright © 2014 Curran-Bauer Analytics, LLC

5-32 | Chapter 5 Structural Equation Models with Latent Variables

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

Note that the means/intercepts and (residual) variances of the factors have been fixed to 0 and 1,
respectively to scale the latent variables.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-33

Example: Identification
 We shall scale the latent factors/residuals by setting their means
and variances to 0 and 1, respectively
 Let us now see if the model passes the t-rule.
 The number of estimated parameters is 22 factor loadings + 22
intercepts + 22 residual variances + 1 factor correlation + 6
factor regression slopes = 73
 The number of observed variables is 22, so there are
22  22  1 2  22  275
unique observed first- and second moments from which to fit the
model
 The model thus passes the t-rule and will have 275 – 73 = 202
degrees of freedom.

Example: Identification
 The t-rule is necessary but not sufficient.
 It quickly shows us that our model may be identified.
 Now we can turn to the two-step rule to further evaluate the
identification of the model.
 Recall that the first step is to reparameterize the model as a
confirmatory factor model with all possible associations among
the factors to verify identification of the measurement model.
 The second step is to examine the structural model as if it were
a structural model for observed variables to verify its
identification.

Copyright © 2014 Curran-Bauer Analytics, LLC

5-34 | Chapter 5 Structural Equation Models with Latent Variables

Example: Identification
9 10 11
Here we see the
1 2 3
pro th time
measurement model is
identified by the 3-indicator
fa fr si
rule
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

Example: Identification
 ERF
Event-Related
Factors  PTG
Environmental
Resources
Posttraumatic
Growth
Individual
Resources
Coping  Cope

 The structural model is identified by the recursive rule.

 Note all arrows run left to right; there are no feedback loops
 To verify the recursive rule, we can write out the matrices for the
structural model treating the latent variables as if they were
observed

No causal feedback loops or correlated disturbances are present here.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-35

Example: Identification
 If the latent variables were observed, the simultaneous equation
model would be
 ERFi    ERF   0 0 0  ERFi   11  12    ERFi 
Coping       ER 
 i  COPE     21 0 0 Copingi    0 0   i    COPEi 
0  PTGi   31  32   i    PTGi 
IR
 PTGi    PTG   0 32

with
 ERF 

Ψ   0  COPE 

 0 0  PTG 

 The recursive model is satisfied because the B matrix is lower

triangular and the  matrix is diagonal.

Note that the Environmental and Individual Resources factors are treated as if they were exogenous x
variables in this step.

Example: Fitting the Model

 Since the model is identified, we can proceed to fit it to the data
 Caution: N = 132 is rather low for a model of this complexity.
 Senol-Durak & Ayvasik (2010) provided their correlation matrix,
standard deviations, and means. We can therefore replicate their
analysis via the summary data version of ML estimation.
 We obtain the following fit measures
 c2(202) = 350.00, p < .0001
 CFI = .848; TLI = .826
 RMSEA = .075; CI90 = (.061, .087)
 SRMR = .101
 These fit statistics suggest mediocre fit of the model to the data.

Our fit measures do not perfectly reproduce the results reported in Senol-Durak & Ayvasik (2010) but
this may be due to rounding error in the correlation matrix.

Copyright © 2014 Curran-Bauer Analytics, LLC

5-36 | Chapter 5 Structural Equation Models with Latent Variables

Example: Model Respecification

 The largest MIs for the model are associated with residual
covariances and cross-loadings. This suggests a problem with the
measurement model.
 To more formally locate the source of misfit, we can respecify the
model as a CFA model.
 Poor absolute fit indicative of misfit of measurement model.
 Significant chi-square difference test with original SEM indicative of misfit
of structural model.

Fitting as CFA Model

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 © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-37

Fitting as CFA Model

 Fitting the CFA model to test the measurement model, we
obtain:
 c2(199) = 348.64, p < .0001
 CFI = .846; TLI = .822
 RMSEA = .075; CI90 = (.062, .088)
 SRMR = .099
 Conducting the chi-square difference test between the original
SEM and the CFA to test the structural model, we obtain:
 Dc2(3) = 350.00 – 348.64 = 1.36, p = .71.
 In line with the MIs, these results show that the mediocre fit is a
reflection of problems with the measurement model.

Respecifying the Model

 Closer examination of the indicators suggests one possible
respecification of the model.
 For some factors, subsets of indicators are composed of
subscales from a single scale
 Individual Resources factor includes three Psychological Hardiness
subscale scores as indicators (comt, con, cha), but also two indicators
from independent scales (lo and es).
 Coping factor includes two indicators from the Ways of Coping
Inventory (em and in), three indicators from the Impact of Event Scale
(ru, av, & hy), and a religious beliefs score from another scale (rb).

Copyright © 2014 Curran-Bauer Analytics, LLC

5-38 | Chapter 5 Structural Equation Models with Latent Variables

Respecifying the Model

 We will use a correlated uniqueness model to account for
possible scale factors
 Indicators from the same scale likely to be more highly related than
indicators from different scales, above and beyond common factor
 Respecificying the model in this way is theoretically motivated
but is also consistent with the MIs
 Four of the five highest MIs are for residual covariances between
subscale scores from the same scale:
 hy,ru MI = 32.258
 cha,comt MI = 27.845
 in,em MI = 21.340
 in,fa MI = 14.250
 cha,com MI = 12.345

Revised Model
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

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-39

Revised Model
 Fit statistics for the correlated uniqueness model:
 c2(195) = 270.28, p = .0003
 CFI = .923; TLI = .908
 RMSEA = .054; CI90 = (.037, .069)
 SRMR = .090
 CU model fits significantly better than the original model:
 Dc2(7) = 350.00 – 270.28 = 79.72, p < .0001.
 Overall fit is borderline, especially given low power with low N
 Some large MIs still, but suggested paths not consistent with theory
 Especially at low N, suggested paths may reflect only chance
fluctuations in data that would not replicate in another sample
 We will retain the revised, correlated uniqueness model and
proceed to interpret the estimates.

Interpretation
 Interpretation of the measurement model is done similarly to
CFA. We will thus consider just the IR factor here as an example
 Note that locus of control is a
negative indicator of individual Individual
resources (but not clear from Resources
article how this was scored) -.64 .51
.39 .47
 Control subscale of Psychological .71

Hardiness Scale is negatively lo comt con cha es

correlated with commitment and
challenge subscales .59
4
.85
5
.50
6 7
.78
8
.74
-.29 -.39
 unusual for a CU model &
inconsistent with notion that .43
residuals correlate due to a
common “scale” factor

All reported values are standardized estimates. Although we consider only the Individual Resources
factor here, the interpretation of the other factors would proceed similarly.

Copyright © 2014 Curran-Bauer Analytics, LLC

5-40 | Chapter 5 Structural Equation Models with Latent Variables

Interpretation
 The structural is depicted with standardized estimates below:
.60
 ERF
Event-Related .59
-.18 Factors  PTG
Environmental
Resources
.20
.65 -.33 Posttraumatic
.24 Growth
Individual .59
.22
Resources
Coping  Cope
 Findings include: .89

 Environmental resources (social support) and Individual Resources

positively influence posttraumatic growth
 Coping positively influences posttraumatic growth
 Positive event related factors (better prognosis, less threat, more time
from event) predict less coping

Total, Direct, and Indirect Effects

 Note, however, that there is a complex pattern of direct and
indirect effects in the model .60
 ERF
Event-Related .59
-.18 Factors  PTG
Environmental
Resources
.20
.65 -.33 Posttraumatic
.24 Growth
Individual .59
.22
Resources
Coping  Cope
.89

 To fully interpret the model, we need to evaluate total, direct and

indirect effects
 The computation of these effects and their standard errors mimics the
procedures discussed earlier for path analysis

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.3 Example SEM 5-41

Effect Decomposition
 Let’s first consider the effect of Event-Related Factors on Post-
Traumatic Growth
 The effects of Event-Related Factors on Post-Traumatic Growth:
 Total: -.20, p < .05
 Direct : 0
 Indirect : -.20, p < .05
 Interpretation: More positive event-related factors reduce the
need for Coping. Since Coping positively impacts Post-Traumatic
Growth, higher Event-Related Factors lead to less opportunity
for Post-Traumatic Growth
 Note that the direct effect of Event-Related Factors on Post-Traumatic
Growth was constrained to be zero, so the total effect equals the
indirect effect in this case.

Effect Decomposition
 Let us now consider the effects of our distal predictors,
Environmental Resources and Individual Resources
 The effects of Environmental Resources on Post-Traumatic
Growth:
 Total: .24, p < .05
 Direct : .20, p < .05
 Indirect : .04, ns
 Environmental Resources do not significantly impact Event-
Related Factors, and this is reflected in the non-significant indirect
effect of Environmental Resources on Post-Traumatic Growth
through Event-Related Factors and Coping.
 Environmental Resources do, however, have a significantly positive
direct effect on Post-Traumatic Growth.

Copyright © 2014 Curran-Bauer Analytics, LLC

5-42 | Chapter 5 Structural Equation Models with Latent Variables

Effect Decomposition
 Last we can consider the effect decomposition for Individual
Resources
 The effects of Individual Resources on Post-Traumatic Growth:
 Total: .09, ns
 Direct : .22, p < .05
 Indirect : -.13, p = .05
 There are countervailing indirect and direct effects that result in
a non-significant total effect.
 Higher Individual Resources predict more positive Event-Related Factors
(which predict less Coping and, in turn, less opportunity for Post-
Traumatic Growth), leading to the negative indirect effect and null total
effect.

Summary
 The procedures we followed in fitting the SEM are very similar to
those used in path analysis and CFA
 Like in path analysis, our models may include complex causal
chains and mediation effects, but now including latent predictors,
mediators and outcomes.
 The computation, testing, and interpretation of total, direct and
indirect effects is similar to path analysis, and can aid in model
interpretation.

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.4 Equivalent Models 5-43

5.4 Equivalent Models

Objectives
 Revisit the issue of equivalent models in the SEM context

Equivalent Models
 When one tests an SEM, generally one considers this a test of the
particular SEM that has been specified.
 But in fact it is a test of the class of all equivalent models to the
specified SEM.
 Some equivalent models, despite having equal fit, may generate
quite different substantive conclusions, so it is a good idea to
consider what alternative model structures cannot be
differentiated from the hypothesized one.

Copyright © 2014 Curran-Bauer Analytics, LLC

5-44 | Chapter 5 Structural Equation Models with Latent Variables

Example: Post-Traumatic Growth

 Consider the structural model that we fit in the Post-Traumatic
Growth example:
 ERF
Event-Related
Factors  PTG
Environmental
Resources
Posttraumatic
Growth
Individual
Resources
Coping  Cope

 In this model, Event-Related Factors constitute a partial mediator

of the effects of Environmental Resources and Individual
Resources, and Coping totally mediates the effect of Event-
Related Factors

Example: Post-Traumatic Growth

 Holding the measurement model constant, this is an equivalent
structural model for the Post-Traumatic Growth example:
Event-Related
Factors  Cope

Coping
 PTG
Environmental
Resources
Posttraumatic
Growth
Individual
Resources

 This model is substantively quite different from the original,

because Event-Related Factors and Coping are no longer
construed as partial mediators of the effects of Environmental
Resources and Individual Resources

Copyright © 2014 Curran-Bauer Analytics, LLC

 5.4 Equivalent Models 5-45

Example: Post-Traumatic Growth

 Examining the standardized estimates, this equivalent model
seems in some ways more sensible
Event-Related
Factors -.33  Cope

-.03
Coping
 PTG
Environmental .59
.60
Resources
.20 Posttraumatic
.24 Growth
Individual
.22
Resources

 More positive Event-Related Factors lead to less need for coping and in
turn less opportunity for post-traumatic growth
 Environmental and Individual Resources promote post-traumatic growth

Summary
 Many other equivalent models may exist that provide precisely
the same fit to the data as the one initially motivated by theory
 Thus the test of theory may not be as strong as initially supposed
(not just testing original model, but also all equivalent models)
 In some cases, some equivalent models may be equally or more
defensible / interpretable as the original model

Copyright © 2014 Curran-Bauer Analytics, LLC

5-46 | Chapter 5 Structural Equation Models with Latent Variables

Chapter Summary
 Structural equation models with latent variables blend together
the core features of path analysis and confirmatory factor analysis
 SEMs are composed of two parts
 Measurement Model (like CFA)
 Structural Model (like path analysis)
 Fitting and interpreting SEMs involves similar procedures to path
analysis and CFA
 Identification, estimation, evaluation, re-specification, interpretation
 Interpretation aided by computation of total, direct, and indirect effects
 Relative to path analysis, SEM with latent variables provides the
key advantage that the coefficients in the structural model are
unbiased by measurement error

Copyright © 2014 Curran-Bauer Analytics, LLC