Structural Equations Modeling

(Introduction)
Psych 203
Quantitative Research in Psychology
Topic #10

Dr. Marshall N. Valencia

mnvalencia@[Link]
 At the end of the session, students should be able to:
• Explain the logic behind structural equation modeling
Learning (SEM).
• Run the different levels of SEM analysis (Confirmatory
Objectives Factor Analysis, Measurement Model Analysis, and
Structural Model Analysis) using the software JASP
and interpret the outputs

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Topics
Structural Equations Modeling
• Components of SEM
• Types of Relationships in a SEM model
• Exogenous vs Endogenous Latent Constructs
• Advantages of SEM (over other multivariate techniques)
• 3 Levels in SEM Analysis
• General process flow of an SEM analysis
• Procedural Issues

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Structural Equations Modeling • A multivariate technique that can simultaneously examine a series of
interrelated dependence relationships among the measured variables
and latent constructs as well as between several latent constructs
(Hair, et al, 2010).

X1 X2

Job X15 X16 X17

Satisfaction

Work Turnover
Engagement
Intent
Org’l
Identity

X12 X13 X14

X3 X4 X5
= latent construct (latent
variable or factors)
= measured variables
(manifest variables or
indicators)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Think of SEM as a “fusion” of Factor Analysis, Linear Regression, and Path Analysis

Factor Analysis Work Load

Sex

Job Engagement
Satisfaction

Linear Regression

Path Analysis

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Model Components ...typically involves two submodels:
1) Measurement model (analyzed using confirmatory factor analysis)
2) Structural model (analyzed using multiple regression

X1 X15
Job
X2 Satisfaction
Work X16
Engagement
X3
Org’l X17
X4 Identity

X5

Measurement Structural Measurement

Model Model Model

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Types of Relationships in
an SEM Model

X1 X15
Job
X2 Satisfaction
Work X16
Engagement
X3
Org’l X17
X4 Identity

X5

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Exogenous vs Endogenous EXOGENOUS constructs
Latent Constructs
• act as an independent variables in
the model
• Determined by factors outside of the
model
• No path (one-headed arrow) towards
them
Job
Satisfaction

Org’l Work Turnover

Identity Engagement Intent

Climate

ENDOGENOUS constructs
• act as dependent variables in the model
• Determined by factors within of the model
• Path (one-headed arrow) towards them are present
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Advantages
• SEM can estimate multiple interrelated dependence
relationships
• SEM can Incorporate latent variables into the analysis
• Theoretical concepts (e.g. satisfaction, quality, organizational
climate) are better represented using multiple measures
• Statistical estimation of the relationships between concepts
are improved by accounting for measurement errors (SEM
deals with this in the analysis of the measurement model)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
3 Levels in SEM Analysis

• Confirmatory Factor Analysis (CFA)

1

• Measurement Model
2

• Structural Model
3

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Purposes of CFA • Confirmatory Factor Analysis (CFA) is applied for the
individual construct

• Test for Model Fit

1

• Convergent Validity
2

• Construct Reliability
3

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 • All latent constructs are simultaneously incorporated
Purposes of a
Measurement Model into one path diagram

• Test for Model Fit

1
• Test for discriminant
2 validity

• Test for normality

3

• Test for outliers

4

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 • Examine and test relationship between exogenous and
Purposes of a endogenous constructs
Structural Model
• Can also be used to test direct and indirect effects

• Test for Model Fit

1

• Test hypotheses on
individual path/regression
2 weights

• Describe coefficient of
determination (R2)
3

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
General Process Flow

Evaluate Modify/
Specify Estimate (how well the
the model the parameters data fit the Respecify the
model) model

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 e1 e1

Specify
Specify the Estimate the Evaluate (how
well the data fit
Modify/
Respecify the
model parameters the model) X1 X2
the model model

Job e15 e16 e17

Satisfaction
X15 X16 X17

e3 X3

e4 X4 Org’l Work Turnover

Identity Engagement
Intent
e5 X5

X12 X13 X14

Climate
e12 e13 e14

X6 X7 X8 X9 X10 X11

e6 e7 e8 e9 e10 e11
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Job
Satisfaction
.28*
.30*

.09 .32*
.35* Org’l Work Turnover
Identity Engagement
Intent

.42*

.45*
Climate

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Specify the Estimate the Evaluate
Evaluate (how
well the data fit the
Modify/
Respecify the
model parameters the fitmodel)
between model
data & the model

Job
Satisfaction
.28*
.30*

.09 .32*
.35* Org’l Work Turnover
Identity Engagement
Intent

.42*

.45*
Climate Χ2 =207.31, df=84 p<.001
NFI=.90, NNFI =.89
CFI=.92, RMSEA=.053, CI=.050, .056
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Specify the Estimate the Evaluate (how
well the data fit the
Modify
Modify/
Respecify the
model parameters Respecify
model) model the
model

Job
Satisfaction
.28*
.30*

.35* Org’l .32*

Work Turnover
Identity Engagement
Intent

.42*
.45* .21*

Climate
Χ2 =207.31, df=84 p<.001
NFI=.90, NNFI =.89
CFI=.92, RMSEA=.053, CI=.050, .056
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Kinds of Research Questions
• Adequacy of the model
• Testing theory
• Amount of variance in the variables accounted for by
the factors
• Reliability of the indicators
• Parameter estimates
• Mediation
• Group differences
• Longitudinal differences
• Multilevel Modeling

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Procedural Issues in SEM
Step 1: Specify the model Step 3: Evaluate (how well the data fit the
• Steps model)
• Issues • Goodness of fit (GOF)
• Unidimensionality of measures • Fit Indexes
• Number of items per construct Step 4: Modify / respecify the model
• Identification • Basic methods
Step 2: Estimate the parameters • Chi-square difference test
• Basic logic • LM Test
• Estimation methods • Wald’s Test
• Issues
• Type of input data
• Missing data
• Data assumptions: multivariate
normality, linearity,
multicollinearity and singularity
• Residuals
• Sample size

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Step 1: Specify the model

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Specify
Specify the
model
Estimate the
parameters
Evaluate (how
well the data fit
Modify/
Respecify the
the model)
the model model

X1 X2

X15 X16 X17

Steps: JobSat
1. Specify a series of
relationships (identify Engagement Turnover
which constructs are
exogenous or endogenous) OrgID
2. Create a path diagram X12 X13 X14
• Define the individual X3 X4 X5
constructs
• Develop and specify the
measurement model Exogenous Constructs Endogenous Constructs
Job Satisfaction Work Engagement
Organizational Identity Turnover Intent

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 JOBS1 JOBS2 JOBS3
Specifying the model in Jamovi

TURN1 TURN2
Lavaan Basics: JobSat
• Spacing does not matter
• New lines should be included for each new
“equation” Engagement Turnover
• =~ creates latent variables (is measured by)
• Latent =~ variable1 + variable2…
• You name it =~ names from the dataset OrgID
• Prediction is that latent variable predicts
Eng_Vig Eng_Ded Eng_Abs
the manifest variables
• <~ creates latent variables (is predicted by)
• Latent <~variable1 + variable2 IDORG2 IDORG3 IDORG4 IDORG5
• Prediction is that latent variable is
predicted by the manifest variables # measurement model
• ~indicates regression
• Y~X Jobsat =~ JOBS1 + JOBS2 + JOBS3
• Names need to be defined in dataset or OrgID =~ IDORG2 + IDORG3 + IDORG4 + IDORG5
previously defined latents
• ~~indicates covariance/correlation Engagement =~ Eng_Vig + Eng_Ded + Eng_Abs
• Does not matter which order you write it in Turnover =~ TURN1 + TURN2
• Names need to be dfined in dataset or
previously defined latents
• You can set specific variances by using: # regressions
• variable ~~ NUMBER*variable Engagement ~ Jobsat + OrgID
• ~1 indicates intercept
• Variable ~ 1 Turnover ~ Engagement
Jobsat ~~ OrgID
Buchanan, E. M. (2020, March 15). Structural
EquationModeling. Psych 203 (Quantitative Research in Psychology)
[Link] Marshall N. Valencia (mnvalencia@[Link])
 cross-loading of error
Specify
Specify the
model
Estimate the
parameters
Evaluate (how
well the data fit
Modify/
Respecify the covariance
the model)
the model model

VIG1 E1
Issues / Considerations VIG2 E2
1. Unidimensionality of measures Vigor
VIG3 E3
• Indicators (measured variables
VIG4 E4
should relate/load only to a
single construct
• Cross loadings are indications of Dedi-
DED2 E5

a lack of construct validity cation DED3 E6

• Suggestion: check cross-loadings DED4 E7
beforehand by running an EFA;
exclude problematic indicators
from the measurement model ABS2 E8
Absor
• Correlated error covariance is p-tion ABS3 E9
also a problem
ABS4 E10

Item DED4 is cross-loading with

the Absorption construct
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Specify
Specify the
model
Estimate the
parameters
Evaluate (how
well the data fit
Modify/
Respecify the
the model)
the model model

Issues / Considerations
2. Items per construct
– Good practice: minimum of 3 items per factor, preferably 4 (Hair,
et al 2010)
– Single item constructs are problematic (construct validity issues
and “identification” issues)
– Some concepts may be very simple that they can be captured by
a single item (e.g. directly observable behaviours such as job
hopping, purchase, and consumption); when in doubt, use
multiple items
– If you have several items (for example you have 9 items), you can
reduce them to 3 “parcels” (get the average of 3 items to
represent one indicator)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Specify
Specify the
model
Estimate the
parameters
Evaluate (how
well the data fit
Modify/
Respecify the
Issues / Considerations
the model)
the model model
3. Identification
– Is there enough information to identify a solution to a set of
structural equations

Underidentified Just Identified Overidentified

Work Work Work

Engagement Engagement Engagement

X1 X2 X3 X1 X2 X3 X4
X1 X2

E1 E2 E3 E1 E2 E3 E4
E1 E2
6 parameters to estimate 8 parameters to estimate
4 parameters to estimate X1 X2 X3 X4
X1 X2 X1 X2 X3 X1 Var1
X1 Var1 X1 Var1 X2 Cov1,2 Var2
X2 Cov 1,2 Var2 X2 Cov1,2 Var2 X3 Cov1,3 Cov2,3 Var 3
X3 Cov1,2 Cov2,3 Var 3 X4 Cov1,4 Cov2,4 Cov3,4 Var4
3 unique terms
6 unique terms 10 unique terms
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Dealing with Identification
problems
To compute the number of
• 3-indicator rule • Underidentified models will not unique variances/ covariances:
• “set the scale”of each construct generate a unique solution ½ [p(p+1)]
(in AMOS, one indicator is
always set to 1.0) • Just-identified models will have a 0 where p is the number of indicators
degrees of freedom thus it will have a (or measured items)
perfect fit; it’s a “saturated” model
with a chi square of 0.
• Overidentified models have more
unique covariance and variance terms df (degrees of freedom)
than parameters to be estimated; a …amount of mathematical
unique solution can be generated information available to
with positive df and a corresponding estimate model parameters
χ2 value df=½ [p(p+1)] – k
where k is the no. of estimated
• “error” messages in the EQS may be parameters
due to identification problems

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Step 2: Estimate the Parameters

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

In this step, the strength of relationships (specified in the model) is estimated

Basic logic:
• The parameter values reflect the minimal discrepancy between the observed
(sample covariance matrix) and the expected (population covariance matrix)
• The discrepancy (or the error) is captured by the residual matrix

X1 X2 X3 X1 X2 X3
X1 Var1 X1 Var1
X2 2 Var2 Observed minus Expected X2 1.5 Var2
X3 3 1 Var 3 X3 3.5 1 Var 3

OBSERVED (data) X1 X2 X3 EXPECTED (model)

X1 Var1
X2 .5 Var2
X3 -.5 0 Var 3

ERROR (residual)
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Estimation Methods Optimal Performance Conditions

Generalized Least Square (GLS) Sample < 500
Maximum Likelihood (ML) Sample >500

Elliptical Distribution Theory (EDT) Sample <500; variables need not

to be normally distributed but
kurtosis should be the same
Asymptotically Distribution Free (ADF) Sample > 2500
Yuan-Bentler test statistic (based on N = 60 to 120
adjustment to the ADF)

(Tabachnick & Fidell, 2007)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Jamovi has 10 estimation methods

If all data is continuous, the default

estimator in the lavaan package is
maximum likelihood

Alternative estimators available in lavaan

are:

•"GLS": generalized least squares

•"WLS": weighted least squares
(sometimes called ADF estimation).
•"DWLS": diagonally weighted least
squares
•"ULS": unweighted least squares

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Issues / Considerations
1. Type of input data
• Metric (interval or ordinal) -- covariances can be computed;
some recent versions of softwares allow nonmetric data
(binary, nominal)
• Covariance or correlation matrix?
• When raw data is used, the software takes care of
generating the covariance and correlation matrix
• Correlational matrix can be derived from covariance matrix;
standardized solutions (from correlational matrix) are
easier to interpret
• Suggestion: use covariance matrix whenever possible

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Issues / Considerations
2. Missing data
• More than 10% of data items are missing – serious
considerations…
• Remedies:
• Complete case (listwise)
• All-available (pairwise)
• Model-based (ML/EM)
• Full information maximum likelihood (FIML)
• If missing data are random, less than 10%, high factor loadings
(.7 or above), any approach is appropriate; otherwise see Hair et
al (2010) or Enders & Bandalaos (2001)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Issues / Considerations
3. Multivariate normality of data
• Deviations from multivariate normality of data impacts on the
estimates of the parameters
• Screen variables for outliers (univariate and multivariate),
skewness and kurtosis
Remedies:
• Apply data transformations
• exclude cases that contribute to non-normality
• Use estimation procedures that handle non-normal data (e.g. “ADF”
method )

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Issues / Considerations
4. Linearity
• SEM techniques examine only linear relationships
• Remedy: raise the measured variables to powers (e.g. square the
scores)
5. Multicolinearity and Singularity
• Extremely high correlations (>.90) between measured variables
• Inspect the determinant of the covariance matrix – extremely small
determinant may indicate multicollinearity or singularity (usually
EQS will abort -- error)
• Remedy: delete the variable causing the singularity or create
composite variables
6. Residuals
• Residuals should be small and centered around zero. The frequency
distribution of the residual covariances should be symmetrical
• Non symmetrical distributions may signal poor fit
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
 Evaluate (how Modify/
Specify the
model
Estimate
Estimate the
parameters
well the data fit the Respecify the
model) model
the parameters

Issues / Considerations
7. Sample size Minimum Model Complexity Measurement Model Characteristics
• Provides a basis for the Sample Size
estimation of sampling
error 100 5 or fewer constructs More than 3 items (or observed
• How large a sample is variables) per construct
needed to produce High item communalities (.6 or
trustworthy results? higher)
Considerations: 150 7 or fewer constructs Modest communalities (.5) and no
• Multivariate normality underidentified constructs
of data
300 7 or fewer constructs Lower communalities (below .45),
• Estimation technique
and/or multiple underidentified
• Model complexity (fewer than 3 items) constructs
• Amount of missing
data 500 Large number of Some with lower communalities,
constructs and/or having fewer than 3 items
• Average error variance
among the reflective
indicators
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
Step 3: Evaluate (how well the data fit
the model)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Specify the Estimate the Evaluate
Evaluate (how
well the data fit the
Modify/
Respecify the
model parameters the fitmodel)
between model
data & the model

• Goodness-of-fit (GOF) indicates the similarity of the observed and

estimated covariance matrix
• The chi square ( χ2 ) is the fundamental measure of GOF

X1 X2 X3 X1 X2 X3
X1 Var1 X1 Var1
X2 2 Var2 Observed minus Expected X2 1.5 Var2
X3 3 1 Var 3 X3 3.5 1 Var 3

OBSERVED (data) X1 X2 X3 EXPECTED (model)

X1 Var1
X2 .5 Var2
X3 -.5 0 Var 3

ERROR (residual)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Fit Indexes
Fit Index Acceptable Threshold Levels
Absolute Fit Indices
Chi-square χ2 Low χ2 relative to degrees of freedom with an
insignificant p value (p>.05)
Relative χ2 (χ2/df) 2:1 (Tabachnik and Fidell, 2007); 3:1 (Kline, 2005)
McDonald Fit Index (MFI) Values > .89 (Hu & Bentler, 1999)
Goodness-of-fit index (GFI) At least .90 is adequate, values > 0.95 is excellent

Adjusted goodness-of-fit At least .90 is adequate (Bollen, 1990), values > 0.95
index (AGFI) is excellent
Root Mean Square Residual Values less than 0.08, values < 0.05 is excellent
(RMR)
Standardized root mean Values less than 0.08, values < 0.05 is excellent
square residual (SRMR)
Root Mean Square Error of Values .08 to .10 is mediocre fit (Mac Callum et al,
Approximation (RMSEA) 1996) adequate, less than .06 is good (Hu & Bentler,
1999), less than .05 is excellent (Steiger, 2007)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Fit Index Acceptable Threshold Levels
Incremental Fit Indices
Normed fit index (NFI) At least .90 is adequate, values > 0.95 is excellent
Non-normed fit index At least .90 is adequate, values > 0.95 is excellent
(NNFI)
Comparative fit index (CFI) At least .90 is adequate, values > 0.95 is excellent
Incremental fit index (IFI) At least .90 is adequate, values > 0.95 is excellent
(Hu & Bentler, 1999)

• Incremental (or Comparative) fit indexes measure the proportionate

improvement in fit by comparing a hypothesized model with a more
restricted nested baseline model.
• CFI should be the index of choice (Bentler, 1990)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 m ≤ 12 12< m <30 M ≥ 30
Chi-square χ2 n.s. p-values expected Sig p-values even Sig. p-values
with good fit expected
For N<250 CFI or TLI .97 or better .95 or better Above .92
RNI may not diagnose well .95 or better Above .92
SRMR Biased upward .08 or less (with CFI Less than .09 (with
or .95 or higher) CFI above .92)
RMSEA Values < .08 (with CFI=.97 Values < .08 (with Values < .08 (with
or better) CFI=.95 or better) CFI=.92 or better)

m ≤ 12 12< m <30 M ≥ 30
Chi-square χ2 n.s. p-values even with Sig p-values expected Sig. p-values
good fit expected
For N>250
CFI or TLI .95 or better Above .92 Above .90
RNI .95 or better, not used Above .92, not used Above .90, not used
with N>1,000 with N>1,000 with N>1,000
SRMR Biased upward, use other .08 or less (with CFI .08 or less (with CFI
indices or .92) above .92)
RMSEA Values < .07 (with CFI=.97 Values < .07 (with Values < .07 (with
or higher) CFI=.92 or higher) CFI=.90 or higher)
Psych 203 (Quantitative Research in Psychology)
(Hair, et al, 2010) Marshall N. Valencia (mnvalencia@[Link])
 JOBS1 JOBS2 JOBS3

TURN1 TURN2
JobSat

Engagement Turnover
Chi-square χ2
Relative χ2 (χ2/df) OrgID

GFI (>=.90) Eng_Vig Eng_Ded Eng_Abs

AGFI (>=.90) IDORG2 IDORG3 IDORG4 IDORG5

CFI (>=.90)
IFI (>=.90)
NFI (>=.90)
TLI (>=.90)
RMSEA (<=.08)

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Step 4: Modify / respecify the model

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
Specify the Estimate the Evaluate (how
well the data fit the
Modify
Modify/
Respecify the
model parameters Respecify
model) model the
model

Basic methods for model modification

1. Chi-square Difference Test
• Two nested models are compared (one model is a subset of another)
• the χ2 value of larger model is subtracted from the χ2 value of the
smaller nested model
• The difference, also a χ2 is evaluated with df equal to the difference in
df between the two models
2. La Grange Multiplier Test (LM) – “what to add”
• Requires estimation of only one model
• What parameters should be added to the model to improve the fit?
• Analogous to forward stepwise regression
3. Wald’s Test – “what to drop”
• Which, if any, of the parameters could be deleted to improve the fit?
(which parameters are not necessary in the model?)
• Analogous to backward stepwise regression
Psych 203 (Quantitative Research in Psychology)
Marshall N. Valencia (mnvalencia@[Link])
 JOBS1 JOBS2 JOBS3

TURN1 TURN2
JobSat

Engagement Turnover

OrgID
Eng_Vig Eng_Ded Eng_Abs

IDORG2 IDORG3 IDORG4 IDORG5

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])
 Bentler, P.M. (2006). EQS 6 Structural Equations Program
Manual. Encino, CA: Multivariate Software, Inc.
Bryne, B.M. (2010). Structural Equation Modeling with
AMOS: Basic concepts, application and programing.
London: Lawrence Erlbaum Associates.
Enders, C.K. and Bandalos, D.L. (2001). The relative

References performance of full information maximum likelihood

estimation for missing data in structural equation
models. Structural Equation Modeling, 11, 1-19.
Hair, J.F., Black, W.C., Babin, [Link] Anderson, R.E. (2010).
Multivariate Data Analysis: A Global Perspective, 7th
ed. New Jersey: Pearson Prentice Hall.
Tabachnick, B.G. & Fidell, L.S. (2007). Using Multivariate
Statistics, 5th Ed. MA: Pearson Education, Inc.

Psych 203 (Quantitative Research in Psychology)

Marshall N. Valencia (mnvalencia@[Link])