See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/279269157

Structural equation modeling-A second-

generation multivariate analysis

Conference Paper April 2015

DOI: 10.13140/RG.2.1.4837.4885

CITATIONS READS

0 113

2 authors:

Elangovan N. Raju Rajendran

NIFT Sri Ramakrishna Institute of Technology
23 PUBLICATIONS 23 CITATIONS 24 PUBLICATIONS 44 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Small Business Strategic Alignment View project

All content following this page was uploaded by Elangovan N. on 29 June 2015.

The user has requested enhancement of the downloaded file.

IBM 2015 Proceedings - SRIT

Structural equation modeling-A second-generation multivariate analysis

Dr Elangovan. N
Christ University, Bengaluru-560029

Dr Rajendran. R
Sri Ramakrishna Institute of Technology, Coimbatore-641010

Abstract
The application of Structural Equation Modeling (SEM) in the business research is
growing. The second-generation multivariate data analysis technique, SEM is easy to use and
provides a high quality statistical analysis. Many visual SEM software programs help in a quick
design of the theoretical model and to modify them graphically using simple drawing tools.
Further it can estimate the models fit, make any modifications and arrive at a final valid model.
It is essential to understand how to design the research process appropriate to the SEM analysis.
This article describes the steps in SEM analysis, the conventions used in presenting a model, the
elements of the SEM technique, the interpretation of the SEM results. The rule of thumb that
evaluates the results and the other issues in SEM reporting are discussed.
Keywords: Structural equation modeling, Components, Characteristics, Steps, Reporting.
1. Introduction
The first generation regression models such as linear regression, LOGIT, ANOVA, and
MANOVA, can only analyze one layer of linkages between independent and dependent variables
at a time. The modern research calls for the examination of the interaction of many independent
variables and the dependent variable, whose influence is either direct or sometimes indirect. The
first generation techniques need to run the analysis many times with different combination to
arrive at a valid model. Contrast to this method structural equation modeling (SEM) enables
researchers to answer a set of interrelated research questions in a single, systematic, and
comprehensive analysis by modeling the relationships among multiple independent and
dependent constructs simultaneously [Gerbing and Anderson, 1988]. SEM permits complicated
variable relationships to be expressed through hierarchical or non-hierarchical, recursive or non-
recursive structural equations, to present a more complete picture of the entire model [Bullock et
al., 1994, Hanushek and Jackson, 1977].
SEM in a single analysis can assesses the assumed causation among a set of dependent
and independent constructs i.e. validation of the structural model and the loadings of observed
items (measurements) on their expected latent variables (constructs) i.e. validation of the
measurement model. The combined analysis of the measurement and the structural model
enables the measurement errors of the observed variables to be analyzed as an integral part of the
model, and factor analysis combined in one operation with the hypotheses testing.
The result is a more rigorous analysis of the proposed research model and, very often, a
better methodological assessment tool [Bollen, 1989, Bullock et al., 1994, Jreskog and Srbom,
1989]. Thus, in SEM, factor analysis and hypotheses are tested in the same analysis. SEM

ISBN : 9-38423-409-X 33
IBM 2015 Proceedings - SRIT
techniques also provide full information about the extent to which the research model is
supported by the data than in regression techniques.
The benefit of any statistical tools can only be obtained by proper application of the tool.
The power of SEM depends on the data characteristics, reliability and validity. Missteps
compromise results validity and hamper interpretation and researchers ability to develop
knowledge that will be useful to practicing managers. (Shook, Hult, Kacmar, & Ketchen, 2004).
Such as other statistical procedure, researchers can easily misuse SEM. The respecification of
model helps researchers to change the hypothesis and pose as though the proposed model is
found fit. Rather testing a model formed out of sound theory and research design, the researchers
will collect data using one model and use it to depict a new model based entirely on statistical
criteria. (Weston & Gore, 2006). This article addresses this need to make researchers to
understand the SEM better and to use it appropriately in the analysis that can produce a relevant
knowledge for practice.
This paper explains the steps in using the structural equation modeling (SEM) and the
overall research design required for the SEM analysis. In the sections that follow, we introduce
to SEM technique, and then look at the types of analysis done. We then discuss the steps in
designing a SEM research and the analysis. The conventions of developing a model for the
SEM analysis are described along with the elements of the model. Finally, the issues that are to
be reported in a SEM research are listed and the conclusion is drawn.
2. The SEM technique
Research interest has moved from just correlation and descriptive statistics to more
complete picture from multiple regression and factor analysis. Factor analysis is an exploratory
tool that finds and groups variables that load to a set of underlying factors. Multiple regression
measures the variance explained by a set of independent variables on a dependent variable
(Tabachnick & Fidell, 2001). Further to factor analysis and multiple regression structural
equation modeling (SEM) further enriches the analysis by testing the models for expected
relationships between a set of variables with their direction of influence and the factors upon
which they are expected to load. (Abramson, Rahman, & Buckley, 2005).
Distinguishing characteristics of SEM is its ability to estimate multiple and interrelated
dependence relationships simultaneously. If a dependent variable becomes independent variable
in subsequent relationships, it gives rise to interdependent nature of the structural model. Many
such variables affect each of the dependent variables with varying effects which can be
represented in a structural model. The relationships in a structural model are translated into series
of structural equations similar to regression equations (Hair, Anderson, Tatham and Black,
1998).
The illustration below (Fig. 1) depicts a structural model with dependent variables being
independent variable of another dependent variable.

ISBN : 9-38423-409-X 34
IBM 2015 Proceedings - SRIT

X1 C
A Y1
H
X2 D
G Y3

X3 E
Y2 I
B

X4 F

Fig 1. Path diagram

The two independent variables X1 and X2 are correlated and predict the dependent
variable Y1. C and D are causal path which may calculated from the correlation or covariance
and expressed as b1 and b2. The regression equation for this path can be expressed as:

Y1=b1X1+b2 X2

For example using the correlation between X1, X2 and Y1 shown in table 1 to estimate
the causal effect of path C(b1) and D(b2). The correlation of X1 and Y1 can be represented by
direct path (C) from X1 to Y1 and a compound path from X1 to X2 and then to Y1(AD).
Likewise the correlation of X2 and Y1 can be shown by direct path (D) and compound path
(AC).
Table 1. Correlation matrix

X1 X2 Y1
X1 1
X2 .673 1
Y1 .535 .552 1

Correlation between X1 and X2 = rX1X2 = A = 0.673

Correlation between X1 and Y1 = rX1Y1 = B+AC

Correlation between X2 and Y2 = rX2Y2 = C+AB ---- Solving these equation we get the
values C (b1) =0.299 and D (b2) = 0.350

Similarly the regression equation for different path can be formed and the larger model
can be modeled simultaneously, using correlations or covariances as the input data.

ISBN : 9-38423-409-X 35
IBM 2015 Proceedings - SRIT

Y2=b3X3+b4X4

Y2=b5Y1+ b3X3+b4X4

Y3=b5Y1+b6Y2

Thus, dependent variable in one equation can be calculated as independent variable in

another relationship. This is done any number of relationships in a path analysis
SEM is capable of analyzing the model by easily incorporating moderators and
mediators as necessary and run several multiple regression equations simultaneously (Byrne,
2001; Kline, 1998). Another strength of SEM is its ability to distinguishes latent and manifest
variables, as well as between endogenous and exogenous variables.
In SEM, the relationship between variables can be easily changed and a modified model
can be tested for theoretical sense. The initial model is usually theory-driven or based on past
research that is further developed by the researchers experience and insight by a disciplined
imagination. The goal of SEM is to test the goodness of fit of the new model to the data and if it
is parsimonious (Kline, 1998; Arbuckle & Wothke, 1999; Ullman, 2001).
Structural equation models are either recursive or non-recursive. We can represent
unidirectional and bidirectional causal relationship easily in a SEM model. SEM can also
examine structural models, measurement models, as well as full structural models in a
single analysis.
There are two distinctive approaches in SEM analysis (Fig 2); they are
1) Covariance analysis and
2). Partial least squares (PLS).

Structural Equation Modeling

Covariance Structure Analysis Component-based

Partial Least Square

Amos PLS
CALIS procedure of SAS/STAT, PLS-Graph
EQS, PLS-PM
LISREL, GSCA
Mplus,
XLSTAT-PLSPM
Mx Graph,
RAMONA module of SYSTAT, Smart PLS
SEPATH module of STATISTICA. Visual PLS

Fig 2. Types of SEM approches

ISBN : 9-38423-409-X 36
 IBM 2015 Proceedings - SRIT
The objective of analysis and the statistical assumptions are different for these two
approaches. The Covariance analysis as employed in LISREL, EQS and AMOS is to analyse the
plausibility of the given model using the data. It analyses the entire path in the specified model,
tests the goodness of fit, and corroborate that the theory is supported by the data collected.
Covariance-based SEM is thought to provide better coefficient estimates and more accurate
model analyses [Bollen, 1989] unlike linear regression and PLS which depends on the F-
statistics and R2 values.
PLS as employed in PLS and PLS-Graph (Chin, 1998b, Thompson et al., 1995) is similar
to liner regression in presenting the high R2 and significant t-values, thus rejecting the null
hypothesis of no-effect.
SEM technique using covariate analysis is more robust that it eliminates the parallel
correlation of a variable with different constructs and assesses the unidimensinality of the
variable (Segars, 1997). We present the comparison of the three methods as discussed by Gefen,
Straub, & Boudreau (2000) in Table 2.
Table 2. Comparison between SEM approaches and Linear Regression
Issue Covariance Approach PLS approach Linear regression
Objective of Overall Path specific null Path specific null Path specific null
Analysis hypotheses are hypotheses are hypotheses is rejected
rejected. As well as the rejected.
null hypothesis of the
entire proposed model
is plausible.
Objective of Overall model fit, Variance explanation Variance explanation
Variance such as insignificant (high R-square) (high R-square)
Analysis 2 or high AGFI.
Required Theory Requires sound theory Does not necessarily Does not necessarily
Base base. Supports require sound theory require sound theory
confirmatory research. base. Supports both base. Supports both
exploratory and exploratory and
confirmatory confirmatory research.
research.
Assumed Multivariate normal, if Relatively robust to Relatively robust to
Distribution estimation is through deviations from a deviations from a
ML. Deviations from multivariate multivariate distribution,
multivariate normal distribution. with established
are supported with methods of handling
other estimation non- multivariate
techniques. distributions.
Minimum sample At least 100-150 At least 10 times the Supports smaller
size required cases. number of items in sample sizes, although
the most complex a sample of at least 30
constructs. is required.
Adopted from Gefen, Straub, & Boudreau (2000)

ISBN : 9-38423-409-X 37
IBM 2015 Proceedings - SRIT
3. Elements and characteristics of the SEM model
The key elements of essentially all structural equation models are their parameters (often referred
to as model parameters or unknown parameters) The elements and characteristics of SEM are
presented in the Table 3.
Table 3. The elements and characteristics of SEM

X Y Observed, Measured, Manifest, Indicator,

Reference variable

Unobserved, Latent Variable, Factor or

Construct, Error or distrubance

Direct effect, Factor loading, Path loading,

Correlation between variables, Parameter, Path

Non directional, Bidirectional association,

Correlation and covariance, Parameter, Path

e e
X Y Measurement error on observed variable

e e
Disturbance or error on latent variable

2 exogenous, 1 endogenous Exogenous Latent variable

Exogenous variables are those
Exogenous variables used to explain relationships
Latent Variable within the model. Exogenous variables
are equated to the independent variables
Endogenous
in a multiple regression equation.
Latent Variable

Exogenous
Latent Variable

ISBN : 9-38423-409-X 38
IBM 2015 Proceedings - SRIT

1 exogenous, 2 endogenous

Exogenous
Latent Variable Endogenous Latent Variable
Endogenous variables are equated to
Endogenous
dependent variables as in a multiple
Latent Variable
regression equation.
Endogenous
Latent Variable

Manifest Effect 1 Reflective measurement:

Manifest Variable reflects the latent
Latent variable. Latent Variable effects
Manifest Effect 2
Variable

Manifest Effect n

Formative measurement:
Manifest Cause 1 Latent Variable caused by the Manifest
Variable
Manifest Cause 2 Latent
Variable

Manifest Cause n

X1 Y1 E1 Recursive model
Recursive models have unidirectional
causal relationships and independent
error terms
X2 Y2 E2

X1 Nonrecursive
In case that a model contemplates two
Y1 E1 or more LVs with reciprocal causation
the model is said to be non-recursive,
X2 i.e. there are some paths going forward
and backward.

X3 Y2 E

ISBN : 9-38423-409-X 39
IBM 2015 Proceedings - SRIT
Moderators are those that interact with
Predictor the relationship of one variables
Variable impact on anothers
Outcome Mediators are those variables that
Variable affect the relationship between two
other variables.
Moderator Mediators come between two variables
Variable such that the first variable has an
indirect effect on the second variable,
through its direct effect on the mediator
Mediator
Indirect Effect (Baron & Kenny, 1986).
Variable

Predictor Outcome
Variable Direct Effect Variable

E11 IND 11 IND = Indicator, E= Error

IND 12 FACTOR 1
E12

E13 IND 13

E21 IND 21

IND 22 FACTOR 2
E22

Measurement Model is mapping of

E23 IND 23
measures onto theoretical constructs.
The constituent parts of the
E31 IND 41
measurement model are factor loading
of the measure on the variable, error
IND 42 FACTOR 3
E32 variances and the error covariances
(correlated errors)
E33 IND 43

E41 IND 41

IND 42 FACTOR 4
E42

E43 IND43
Measurement Model

ISBN : 9-38423-409-X 40
IBM 2015 Proceedings - SRIT

FACTOR

Structural Model presents the causal

FACTOR 3 FACTOR and correlational links between
theoretical variables. The constituent
parts of a structural model are paths,
FACTOR variances of the exogenous variables,
Structural model covariances between exogenous
variables, variances of the disturbances
of endogenous variables, covariances
between disturbances, covariances
between disturbances and exogenous
variables (usually set to zero)

E11 IND 11
E31 E32 E33

IND 12 FACTOR 1 IND 31 IND 32 IND 33

E12 IND 41 E41

E13 IND 13 FACTOR 4

FACTOR 3 IND 42 E42

E21 IND 21
E3 IND 43 E43
E2
FACTOR 2
E22 IND 22
IND = Indicator, E= Error
E1
E23 IND 23

Full Model

4. Alternate Approaches in SEM

SEM is usually viewed as a confirmatory rather than exploratory procedure. There are
three approaches in using the SEM for analysis (Fig. 3): strictly confirmatory, exploration of
alternative models, and model-generating models (Joreskog, 1993).

SEM analysis

Strictly confirmatory Alternative and Explorative Model generation

Fig 3. Alternate approaches of SEM

ISBN : 9-38423-409-X 41
IBM 2015 Proceedings - SRIT

Strictly confirmatory approach: Strictly confirmatory approaches involve testing a

theoretically-developed model, without modification, against a dataset. A model developed
out of sound theoretical base is tested using SEM goodness-of-fit tests to determine if the pattern
of variances and covariances in the data is consistent with a structural (path) model specified in
the model. An unexamined model may fit the data better; however an accepted model is only that
is not a disconfirmed model.
Alternative models approach: In alternative models approach several priori, alternate,
plausible models are tested with the collected data to obtain the model that best fits theory and
data. This approach is a kind of explorative method of theory development process in which the
prior theory does not exist or is in a rudimentary form. However, this kind of finding has to be
replicated across other samples from the same population to make the theory a trustworthy one.
Model development approach: An alternative model is developed from the modification index
produced in the SEM analysis. Whatever may be the SEM approach; the causal arrows are
specified based on the theoretical insight and experience of the researcher. Since this kind of
model development is obtained post-hoc from the available data may not be stable and requires
cross validation by independent sample.
In the alternative model approach, the model selected is done so based on best fit to
theory and data, along with the goal of parsimony (Byrne, 2001; Joreskog, 1993; Kline, 1998).
However as the model-generating approach is exploratory, the generated model should be
tested on a second dataset. The other way of generating model from data is to randomly split a
dataset into two subsample (Breckler, 1990).

5 Research steps for SEM

The basic building block of all SEM analyses follows a logical sequence. When the steps
of SEM analysis is strictly adopted, the researcher can ensure that the results are valid and
worthy for the practicing managers. The six steps to analyzing a dataset using SEM as described
by Kline (2005) are presented in the fig 4. Kline (2005) has developed from his earlier five step
method into a six step method by adding the reporting as an important issue in using SEM
analysis. He also supplements the six-step method with addition steps of replicating the model
and applying the results that reflect the cyclic process of the research. In using the steps in SEM
analysis, the researchers should understand and modify it for the different approaches explained
in the previous chapter. Even as the three approaches to SEM are qualitatively different, the
strictly confirmatory approaches end at the fourth step and other approaches can continue upto
the final step. The next section describes the steps to undertaking SEM. They are actually iterative
because problems at a later step may require a return to an earlier one.

ISBN : 9-38423-409-X 42
IBM 2015 Proceedings - SRIT

Literature review, Theoretical basis,

Model Specification Interviews with Practitioners and Model
Development

Specifying the constraints to obtain a

Model Identifiability
unique estimate of every model parameter

Measure selection,
Data collection,
Cleaning and preparation
Exploratory and Confirmatory Study
Fit indicies & Unidimensionality
Model Analysis and Evaluation assessment
Discriminant validity
Construct reliability
Model Respecification Test structural model

Reporting

Replicate the results

Apply the results

Fig 4. Steps of SEM analysis by Kline (2005) and equivalent steps by Lu, Lai, & Cheng (2007)
Model Specification
The first step of SEM analysis is the development of the model. Going by the various
theory and model development methods researchers hypotheses are expressed in the form of a
structural equation model. Holton and Lowe (2007), operationalized the process of building
theory on the Dubins (1978) theory development method. This process serves well for the
model development. The researcher first understands the phenomena of his research area by
conducting an initial review of literature and preliminary interview with users, experts and
academicians. The researcher arrives at a refined area within the broader perspective of a
phenomenon and makes a review of literature further in the narrow area.
The researcher is now ready with variables or constructs of the study and understands the
existing relationship between them. At this point, the researcher can introduce a new construct or
a relationship out of insight and experience and propose a conceptual theory. The main challenge
in the model specification is to care that all necessary variables are included and their
relationships are indicated. The endogenous and exogenous variables including moderators and
mediators both from literature and conceptualized by the researcher that probably influence the
endogenous variable. Errors of omission and/or inclusion of any variable or parameter are known

ISBN : 9-38423-409-X 43
IBM 2015 Proceedings - SRIT
as specification error.
This conceptual theoretical framework will constitute and will define the units, the laws
of interaction of theory, the boundaries of theory, the system states of theory and the propositions
of a theory. The theory is represented pictorially as model, which is helpful to show causal
relationships in a logical ordering so that readers can understand the chain of causation or how a
third variable intervene in or moderates in a relationship. Researchers specify by drawing a
diagram of a model using a set of more-or-less standard symbols (as in section 3).

5.2. Model Identifiability:

A model that is specified as in the first step is checked for identifiability. A model is said to be
identifiable if there is a unique solution possible for it and each of its parameters. An unidentified
model will have infinite number of possible solution and makes it difficult for the analysis to
converge on a single solution. For example, consider an equation X + Y = 5, the two variables X
and Y in the equation can have more number of combination and will not be called to have a
unique value. A deficit observation in which the number of observation is less than the number
of parameters can lead to this kind of model and we call it an under-identified model. In an
model, if number of observations are equal to number of parameters we call it just-identified
model and when the number of observation are more than the number of parameters we call it
over-indentified model. Software like AMOS will indicate when a path is unidentified and helps
the researcher to specify constraints to make the model just-identified.
If a model fails to meet the relevant identification requirements, attempts to estimate it may
be unsuccessful; such models need to be respecified to be identifiable by (Kline, 1998; Ullman,
2001). Recursive models are usually identifiable, but result in an unidentifiable model (Kline,
1998).
Though an over-identified model can obtain an unique solution it may not produce a perfect
observation. An under-identified model is respecified by imposing more constraints to the latent
variable by assigning a scale

5.3 Measure selection, Data collection, Cleaning and preparation:

Structural equation modeling being a correlation and covariance based research method
care should be taken on the data characteristics. A valid data will provide a better analysis and
interpretation for the benefit of the managers.
Joreskog (1977) recommends that the latent construct be measured by at least two
manifest variables. Ullman, (2001) recommends that the best scale of the variable that is used in
SEM is of an interval scale If an ordinal or nominal scale is used, a fewer response options will
be better. The selected measures should be ensured to have good psychometric properties of
reliability and validity; otherwise, the statistical estimates become biased. A measure is
considered reliable when it gives consistent, or repeatable, results. It is considered valid when it
measures what it says it measures.
During data collection, a sufficiently large sample is required to be collected to analyse
the model better. As a rule of thumb for covariance approach, at least 100-150 cases are collected
and for PLS approach at least, 10 times the number of items in the most complex construct. Data
cleaning is also an important step that is required to ensure the data integrity to avoid Garbage In,
Garbage Out. Simple techniques such as descriptive statistics, histograms and scatter plot can
highlight the outliers in the data and can be corrected.

ISBN : 9-38423-409-X 44
IBM 2015 Proceedings - SRIT
Data preparation is the step during which certain assumptions are checked before
proceeding with data analysis. Maximum likelihood (ML) estimate is the preferred method for
SEM. The maximum likelihood parameter estimation determines the parameters that maximize
the probability (likelihood) of the sample data. The ML estimate can give an unbiased analysis
when data characteristics explained in table 5 is taken care:

Table 5. Assumptions of data characteristics

Assumption Description Identification

Are the variables normally
Multivariate For each value of X, the values of
1 distributed? Check histograms of
normality Y are normally distributed.
all variables in an analysis
The best fitting line through the Check scatterplots and
data is a straight line. correlations between Y and each
2 Linearity
Relationships among variables are of Xs. Check for influence of
linear in nature bivariate outliers
If any two or more variables that
3 Independence
are uncorrelated are independent.
Are the bivariate distributions
reasonably evenly spread about
the line of best fit?
Homoscedasticity For each value of X, the variance Check scatterplots between Y
4
of Y is the same and each of Xs and/or check
scatterplot of the residuals
(ZRESID) and predicted values
(ZPRED)
SEM requires that manifest IVs: MLR involves two or more
Interval scaled variables be measured on an interval/ratio or dichotomous
5 measures interval / continuous scale. variables
Nominal scale are categorical, and DV: One interval/ratio variable
treated as grouping variables
6 Min. 5 cases per predictor (5:1)
Ideally 20 cases per predictor
(20:1), with an overall N of at
least 100; this allows sufficient
Random Random sampling is a statistical power to detect a medium R2
sampling assumption is the underlying for Tabachnick & Fidell (2007)
unbiased probabilistic sample. suggest that N should ideally be
50 + 8(k) for testing a full
regression model or 104 + k when
testing individual predictors
(where k is the number of IVs).

ISBN : 9-38423-409-X 45
IBM 2015 Proceedings - SRIT

Is there multicollinearity
between the IVs? (i.e., are they
Multicollinearity is present when overly correlated e.g., above .7?)
pairs of variables have
Check scatterplots and
correlations of 0.9 or more,
correlations between each of the
Multicollinearity squared multiple correlations
7 Xs and/or check the collinearity
(SMC) approach 0.99, or
statistics in the Coefficients table:
tolerance statistics (the inverse of
The Variance Inflation Factor
SMCs) approach 0.01
(VIF) should be low (< 3) and
(Tabachnick & Fidell, 2001).
Tolerance should be high (> .3)
(Note that TOL=1/VIF).
Singularity is present when a pairs
of variables have correlation of 1
or SMC=1, or tolerance =0.
8 Singularity Singularity happens when
variables are redundant; one
variable is a combination of two
or more other variables.
Missing completely at random Missing Value Analysis (MVA)
Missing data (MCAR).
9
Missing at random (MAR) and
Non-ignorable missing data.
An outlying observation, or Check Mahalanobis' Distance and
Multivariate
outlier, is an observation that is Cook's D
10 Outliers
much different in relation to the
other observations in the sample.

4. Model Analysis and Evaluation.

MLE methods are versatile and provide efficient methods for fitting a model using the
available data. Though the methodology for maximum likelihood estimation is simple, the
calculations are mathematically intensive. Mathematical complexity is not a big obstacle when
computers are utilized for estimation. Many SEM softwares are available that uses the
covariance and PLS approach. Major results and reports are generated in this step.
The types of activities that are done in this step are:
a. Evaluating the model fit. The model is tested to determine whether the specified model
is explained by the data. Many a times the initially specified model does not fit to the data
very well. In such case, the model is respecified as in step-5. When a well-fit model is
arrived after respecification or the initial model, itself is good. The next steps are done.
b. Interpret the parameter estimates. When it is determined that the fit of a structural
equation model to the data is adequate, the estimated parameters of the various parts of
the model are interpreted for their specific effects and significance.
c. Analyse and justify equivalent models. Though a conceptually developed model based
on the literature is well explained by the data, there may be a situation where an
equivalent model can also explains the data well similar to the specified model. There

ISBN : 9-38423-409-X 46
IBM 2015 Proceedings - SRIT
may be possibility of many such models explained well by the same data and the
researcher has to justify why his model should be preferred over the other equivalent
models.
Covariance-based SEM packages generate statistics at three levels: (1) individual
path and construct, (2) overall model fit and (3) individual path modification indexes.
The test that are carried out in the SEM estimates and the critical value of acceptance are
list in the table 6 (Lu, Lai, & Cheng, 2007).

Table 6. SEM Analysis

Statistical Analysis Critical Value
Exploratory Study
Corrected item total Correlation >0.5
Factor analysis within block of loadings
Exploratory factor analysis of entire set
Reliability through Cronbachs alpha >0.7
Confirmatory Study
Convergent Validity
t-value for each loading, significance
Squared correlation
Discriminant validity
Average variance extracted vs squared
correlation between factors
Construct reliability
Composite reliability >0.7
Average variance extracted (AVE) > 0.5
Fit indicies & Unidimensionality assessment
2, Check significance
from table value
2/df <2
GFI, AGFI, CFI >0.9 0(No fit) to 1(Perfect
fit)
RMSR, RMSEA < 0.05
Standard residual <2.58
Modification indices < 4.0
Completely standardized change in Ax <0.4
t-value of structural coefficients for
significance

Of all the analysis listed in the table x the minimal set of fit indexes that should be
reported and interpreted when reporting the results of SEM analyses are: (1) the model chi-
square, (2) the SteigerLind root mean square error of approximation (RMSEA; Steiger, 1990)
with its 90% confidence interval, (3) the Bentler comparative fit index (CFI; Bentler, 1990), and
(4) the standardized root mean square residual (SRMR). The various fit indexes have different
characteristics catering to complexity of the model, sample based model, and Population based
model and the size of the model. The table 7below shows the miscellaneous fit indices.

ISBN : 9-38423-409-X 47
IBM 2015 Proceedings - SRIT

Table 7. Miscellaneous fit indices.

Model Fit criterion Comment

Incremental fit indexes
NFI Sample-based
1
Normed fit index 0(No fit) to 1(Perfect fit)
NNFI Sample-based, parsimony-adjusted, value can
2
Non-Normed fit index fall outside of range 01.0
PNFI Sample-based, parsimony-adjusted, sensitive
3
Parsimony NFI to model size
RNI
4 Population-based, value can be negative
Relative Noncentrality index
PCFI Population-based, parsimony-adjusted,
5
Parsimony comparative fit index sensitive to model size
Predictive fit indexes
ECVI
6 Population-based, parsimony-adjusted
Expected cross-validation index
CAIC
Population-based, parsimony-adjusted, takes
7 Consistent Akaike information
account of sample size
criterion
Population-based, parsimony-adjusted, takes
BIC
8 account of sample size, greater penalty for
Bayes information criterion
complexity
Absolute fit indexes
GFI Sample-based, analogous to R2, value can
9
Goodness-of-fit index fall outside of range 01.0
Sample-based, parsimony-adjusted,
AGFI
10 analogous to corrected R2, value can fall
Adjusted GFI
outside of range 01.0
Sample-based, parsimony-adjusted,
PGFI analogous to corrected R2, value can fall
11
Parsimony GFI outside of range 01.0, sensitive to model
size

5.5 Model Respecification

This step is carried out if a model that is specified in the initial step and tested as in step-4
does not fit significantly. If the fit indicies is below the acceptable value, we try to understand
what the data really explains by respecifying the model and test it with the same data. Model
respecification is not arrived just by modifying the relationship between constructs or by
dropping the construct, but should be guided by the researchers hypotheses. The respecification can
be either theory-driven or exploratory. When the strictly confirmative approach does not
conclude with the result, the theory-driven alternate models approach or an explorative model-
generating models approach can be used by respecifying the model.
The fit indicies of the model is improved by the model being trimmed or built by

ISBN : 9-38423-409-X 48
IBM 2015 Proceedings - SRIT
removing or adding direct effects (Kline, 1998). Models are also modified by reconfiguring the
direction of direct effects. In SEM method, six model respecification tools can be used to
facilitate the decision to identify which construct or effect to be modified Table 8.
In a theory driven model respecification the Chi-squared goodness of fit index is checked
for improvement by adding or eliminating the construct or their relation that is hypothesized
from the literature. When the Chi-squared goodness of fit is found to be acceptable, we can retain
the model. For explorative respecification five tools are used to help model building or trimming
decisions

Table 8. Tools for Theory Driven and Exploratory Approaches to Model Respecification

Test Statistics Critical Value Interpretation

Theory Driven Model Respecification
Accept model with more
1 Significant
Chi-squared goodness of fit parameters
test Accept model with fewer
Nonsignificant
parameters
Exploratory Approaches to Model Respecification
RMAT resids > |.1| Insert associated variables into
2 Residual matrices
CVAR resids >|2.58| model
Modification indices
3 Largest sig. values Drop parameter from model

4 Wald statistics or z-statistics Nonsignificant Drop parameter from model

Akaike Information Criterion
5 Lower value Accept model with lower value
(AIC)
Consistent Akaike Information
6 Lower value Accept model with lower value
Criterion (CAIC)
(Adapted from Arbuckle & Wothke, 1999; Byrne, 2001; Joreskog, 1977; Kline,
1998)

In a correlation residual matrix, if the absolute value is greater than 0.1 shows that there
is a need for additional direct effect. Similarly, in a covariance residual matrix, the absolute value
greater than 2.58 will require an additional direct effect to modify the model to be acceptable.
(Kline, 1998).
In an explorative model-generating model, separate analysis for testing various
modification of a model is time consuming. Modification index (MI) allows evaluating many
potential modifications in a single analysis. The removal of the direct effect with larger MI value
will contribute greatly to the chi-square statistics. In trying to improve a model we should not
rely exclusively on modification indices. A modification should be considered only if it makes
theoretical or common sense. (Arbuckle, 2007)
The next tool for respecification of the model is the Wald W statistic. Parameters are
dropped without effecting the chi-square statistics when their associated Wald W statistic
is nonsignificant (Kline, 1998; Ullman, 2001). The z-statistic is also used similarly to drop a
relationship or effect in the model when the z-value is insignificant for the path.

ISBN : 9-38423-409-X 49
IBM 2015 Proceedings - SRIT
The two final tools are the Akaike Information Criterion (AIC) and Consistent Akaike
Information Criterion (CAIC). AIC is used to compare alternate models that are equivalent to the
chi-squared goodness of fit test, adjusted for model complexity. The W statistics, z-statistics and
AIC are affected by the sample size. The CAIC adjusts for both sample size and model
complexity. The model is interpreted as best when AIC or CAIC is lower. (Kline, 2005).
5.6. Reporting.
Kline (2005) adds this step in his new edition after the five-step approach. After
obtaining a satisfactory model, a written report should be accurate and complete in describing the
analysis. MacCallum and Austin (2000) reviewed publications using SEM analysis and found
that there were problems with the reporting of the results in many of these reports. They found
that the reports were incomplete in presenting the parameter estimates, if a covariance or
correlation matrix was analyzed, indicators of latent variables not clearly described etc. These
are serious shortcomings, especially considering that there are published guidelines for the
reporting of results in SEM (Boomsma, 2000; Hoyle & Panter, 1995; McDonald & Ho, 2002).
Shook, Hult, Kacmar, & Ketchen (2004) developed a checklist of issues that should be reported
in SEM studies

I. Sample issues
(a) General description
(b) Number of observations
(c) Distribution of sample
(d) Statistical power
II. Measurement issues
(a) Reliability of measures
(b) Measures of discriminant validity
(c) Measures of convergent validity
III. Reproduceability issues
(a) Input matrix
(b) Name and version of software package used
(c) Starting values
(d) Computational options used
(e) Analytical anomalies encountered
IV. Equivalent models issues
(a) Potential existence acknowledged as a limitation
V. Respecification issues
(a) Changes cross validated
(b) Respecified models not given status of hypothesized model

Kline (2005) proposes two more steps to the basic ones to make the result of SEM
analysis more contributing to the research community as well as to the society ultimately. The
failure to carry out the steps in SEM analysis makes the outcome unworthy and the following
two steps are to make the SEM analysis a more accepted research method.

5.7. Replicate the results. Replication is relatively rare in SEM. That is, models are seldom
estimated across independent samples either by the same researchers who collected the original
data (e.g., internal replication) or by different researchers who did not (e.g., external replication).

ISBN : 9-38423-409-X 50
IBM 2015 Proceedings - SRIT
The general need for large samples in SEM undoubtedly complicates replication. Nevertheless, it
is critical to eventually replicate a structural equation model if it is ever to represent anything
beyond a mere statistical exercise.

5.8. Apply the results. Kaplan (2000) notes that despite more than 30 years of application of
SEM in the behavioral sciences, it is rare that results from SEM analyses are used for policy or
clinically relevant prediction studies. Neglecting to correctly carry out steps 46 may be part of
the problem.

6 Conclusions and Implications

SEM is a useful, robust and powerful statistical tool. The advantage of SEM using path
model creates a visual understanding of the research model to the researcher. In single step, the
SEM softwares are capable of estimating and evaluating the research model. In comparison with
regression tools, SEMs advantage lies in multiple-equation regression where the dependent
variable of one equation can appear as an independent variable in another equation. In addition,
SEM is capable of analyzing the effect of variables reciprocally in either direction or indirectly
with other variable in between. Another use of SEM lies in finding the effect of unobserved
variable through their observable causes.

SEM enables confirmation of factor structure amongst a set of variables. It can also
accommodate moderator (interaction terms) within the model. Within the constraints of sample
size, SEM can also model mediators. In the absence of missing data, bootstrapping procedures
can also be performed.

All these unique advantages of SEM are not just easily derived. For a robust analysis, the
data need to have certain qualities and satisfy the assumptions. However, SEM can produce
results even when data meets the assumptions, but will lack the practical implications. Another
issue in SEM is the interpretation and reporting of the results. This article help the researchers in
using the SEM effectively while taking care of the requirements of SEM analysis. This article
has attempted to explain the core concepts in SEM and the steps in the analysis. Following these
steps will help the researcher use SEM analysis with ease. This article also explains the standard
practices that are necessary to interpret the estimates and to report sufficient results so that
findings can be reproduced.
References
Abramson, R., Rahman, S., & Buckley, P. (2005). Tricks and Traps in Structural Equation
Modelling: a GEM Australia Example Using AMOS Graphics. Proceedings of ABBSA
Conference (August 5-7 2005), (pp. 558-599). Cairns Australia.
Arbuckle, J. L. (2007). Amos 16.0 User's Guide. PA: Amos Development Corportaion.
Arbuckle, J.L. and Wothke, W. (1999). Amos 4.0 Users Guide. Chicago, IL: Small-Waters
Corporation.
Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social
psychological research: Conceptual, strategic, and statistical considerations. Journal of
Personality and Social Psychology, 51(6), 1173- 1182.

ISBN : 9-38423-409-X 51
IBM 2015 Proceedings - SRIT
Bentler, P.M. (1990). Comparative fit indexes in structural models. PsychologicalBulletin, 107,
238246.
Bollen, K.A. (1989). Structural equations with latent variables. New York: John Wiley.
Boomsma, A. (2000). Reporting analyses of covariance structures. Structural Equation
Modeling, 7, 461483.
Breckler, S. J. (1990). Applications of covariance structure modeling in psychology: Cause for
concern? Psychological Bulletin, 107, 260273.
Bullock, H. E., Harlow, L. L. and Mulaik, S. A. (1994) "Causation Issues in Structural Equation
Modeling Research," Structured Equation Modeling, 1(3), 253-267.
Byrne, B. M. (2001). Structural Equation Modeling with AMOS: Basic Concepts, Applications,
and Programming. Mahway, New Jersey: Lawrence Erlbaum Associates, Publishers.
Chin W. 1998. Issues and opinion on structural equation modeling. MIS Quarterly 22(1), 716.
Dubin, R. (1978). Theory building. New York: Free Press.
Gefen, D., Straub, D. W., & Boudreau, M.C. (2000). Structural Equation Modeling Techniques
and Regression: Guidelines For Research Practice. Communications of AIS Volume , 4
(7), 1-79.
Gerbing, D.W., Anderson, J.C. (1988). An updated paradigm for scale development
incorporating unidimensionality and its assessment. Journal of Marketing Research, 25,
186192.
Hair, Joe F. Anderson, Rolph E. Tatham, Ronald L and Black, William C. (1998). Multivariate
Data Analysis (5th ed.). New Delhi: Pearson Education.
Hanushek, E. A. and. Jackson J. E (1977) Statistical Methods for Social Scientists. Orlando, FL
USA: Academic Press.
Holton, E.F., and Lowe, J.S. (2007). Toward a General Research Process for Using Dubin's
Theory Building Model. Human Resource Development Review, 6, 297-320.
Hoyle, R. H., & Panter, A. T. (1995). Writing about structural equation models. In R. H. Hoyle
(Ed.), Structural equation modeling (pp. 158176). Thousand Oaks, CA: Sage.
Joreskog, K. G. (1977). Structural equation models in the social sciences: Specification,
estimation and testing. Paper presented at the Applications of Statistics: Proceedings of
the Symposium Held at Wright State University, Dayton, Ohio, 14-18 June 1976.
Joreskog, K. G. (1993). Testing structural equation models. In K. A. Bollen & J. S. Long (Eds.),
Testing Structural Equation Models (Vol. 154, pp. 294-316). Newbury Park: SAGE
Publications.
Jreskog, K. G. and Srbom D. (1989) LISREL 7: A Guide to the Program and Applications, 2nd
edition. Chicago: SPSS Inc.
Kaplan, D. (2000) Structural Equation Modeling Foundations and Extensions. Thousand Oaks,
CA: Sage Publications.

ISBN : 9-38423-409-X 52
IBM 2015 Proceedings - SRIT
Kenny, D. A. (1998, November 9). Structural Models with Latent Variables. Retrieved August
11, 2009, from http://davidakenny.net/cm/sem.htm
Kline, R. B. (1998). Principles and Practice of Structural Equation Modeling. New York: The
Guilford Press.
Kline, R. B. (2005). Principles and practice of structural equation modeling (2 ed.). New York,
NY: The Guilford Press.
Lu, C. S., Lai, K. H., & Cheng, T. (2007). Application of structural equation modeling to
evaluate the intention of shippers to use Internet services in liner shipping. European
Journal of Operational Research , 180, 845867.
MacCallum, R. C., & Austin, J. T. (2000). Applications of structural equation modeling in
psychological research. Annual Review of Psychology, 51, 201 236.
McDonald, R. P., & Bolt, D. M. (1998). The determinacy of variables in structural equation
models. Multivariate Behavioral Research, 33(3), 385-401.
Segar, A. (1997). Assessing the unidimensionality of measurement: A paradigm and illustration
within the context of information systems research. Omega, 25 (1), 107121.
Shook, C. L., Hult, G. T., Kacmar, M. K., & Ketchen, J. D. (2004). Research notes and
commentaries an assessment of the use of structural equation modeling in strategic
management research. Strategic Management Journal , 25, 397404.
Steiger, J. H. (1990). Structural model evaluation and modification: An interval estimation
approach. Multivariate Behavioral Research, 25, 173180.
Steiger, J. H. (2001). Driving fast in reverse: The relationship between software development,
theory, and education in structural equation modeling. Journal of the American Statistical
Association, 96, 331338.
Tabachnick, B. G., & Fidell, L. S. (2001). Using Multivariate Statistics (4th ed.). Boston:
Allyn and Bacon.
Thompson, R., D. W. Barclay, and Higgins, C. A. (1995) "The Partial Least Squares Approach
to Causal Modeling: Personal Computer Adoption and Use as an Illustration,"
Technology Studies: Special Issue on Research Methodology, 2 (2) (Fall), pp. 284-324.
Ullman, J. B. (2001). Structural equation modelling. In B. G. Tabachnick & L. S. Fidell (Eds.),
Using Multivariate Statistics (4th ed.). Boston: Allyn and Bacon.
Weston, R., & Gore, P. A. (2006). A brief guide to structural equation modeling. The counseling
psychologist , 34 (5), 719-751.

ISBN : 9-38423-409-X 53
IBM 2015 Proceedings - SRIT
BIOGRAPHY:
Elangovan. N (elangovan.n@christuniversity.in) is an assistant professor and Deanery Research
Coordinator at Christ University Bengaluru. He was awarded Ph.D. by Anna University Chennai.
He also holds a B.E. degree in Mechanical Engineering and MBA Degree with specialization
Marketing. He has also received M.Sc. and M.Phil. degree in Psychology. His primary teaching
interest is in Marketing and Systems Before entering the academics, he spent 15 years managing
family business and his own enterprise producing POP materials. His area of interest and
research is in strategic information systems management. He has published his research findings
in International journal of Enterprise resource planning, Journal of Enterprise Resource Planning
studies and many national and international Journals. He is also a reviewer for conference of
Informing Science Institute and International Journal of Innovation Management.

Rajendran. R. is a professor in the Department of Management at Sri Ramakrishna Institute of

Technology, Coimbatore, India. He holds a Ph.D. and M.Phil. in management from the
Bharathiar University and an M.B.A. and B.E.(Hons.) degree in Mechanical Engineering from
the Madras University. He has prior business and industry experience. He served as a senior
engineer for Bharat Electronics Limited, a public sector undertaking of Government of India
under administrative control of Ministry of Defense, for ten years. Prior to joining academic in
the year 2000, he was managing a small manufacturing partnership firm for six years. His
research focuses on information systems strategic management, electronic commerce, electronic
business and small business management. He has published his findings in journals such as
Digital Business Review, Journal of Electronic Commerce Research, International Journal of E-
Business Research and Organization Management. He is a member of the international editorial
review board of International Journal of E-Business Research.

ISBN : 9-38423-409-X 54
View publication stats