Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Published by:
http://www.sagepublications.com
On behalf of:
Additional services and information for Counseling Outcome Research and Evaluation can be found at:
Email Alerts: http://cor.sagepub.com/cgi/alerts
Subscriptions: http://cor.sagepub.com/subscriptions
Reprints: http://www.sagepub.com/journalsReprints.nav
Permissions: http://www.sagepub.com/journalsPermissions.nav
Citations: http://cor.sagepub.com/content/3/1/30.refs.html
A Five-Step Guide to
Conducting SEM Analysis
in Counseling Research
Stephanie A. Crockett1
Abstract
The use of structural equation modeling (SEM), a second-generation multivariate analysis technique
that determines the degree to which a theoretical model is supported by the sample data, is becoming increasingly popular in counseling research. SEM tests models that include both observed and
latent variables, allowing the counseling researcher to confirm the factor structure of a newly developed or existing psychological instruments and to examine the plausibility of complex, theoretical
counseling models. This article provides counseling researchers and practitioners with an overview
of SEM and presents five steps for conducting SEM analysis in counseling research.
Keywords
structural equation modeling, counseling research
Submitted 2 October 2011. Revised 4 December 2011. Accepted 5 December 2011.
Crockett
31
of SEM techniques. This article strives to familiarize counseling researchers and practitioners
with the purpose and uses of SEM, as well as
provide an applied approach to conducting
SEM analysis. In particular, the article begins
with a general overview of SEM, including
key terms and definitions, a brief history of
SEM development, and the advantages and
limitations associated with the approach.
Readers will then learn how to conduct SEM
analysis in counseling research using a series
of five, applicable stages.
Overview of SEM
SEM is a second-generation multivariate analysis technique that is used to determine the
extent to which an a priori theoretical model
is supported by the sample data (Raykov &
Marcoulides, 2000; Schumacker & Lomax,
2010). More specifically, SEM tests models
that specify how groups of variables define a
construct, as well as the relationships among
constructs. For example, consider a counseling
researcher who is interested in the impact of the
therapeutic working alliance, a construct that
cannot be directly measured, on the number
of counseling sessions a client attends. The
researcher could use SEM to determine
whether (a) variables such as agreement on
therapy tasks, agreement on therapy goals, and
the counselorclient emotional bond comprise
the construct therapeutic working alliance, and
(b) the therapeutic working alliance, as a
whole, is predictive of client number of counseling sessions attended.
In essence, SEM uses hypothesis testing to
improve our understanding of the complex relationships that occur among observed variables
and latent constructs. Observed variables (i.e.,
indicator variables) are variables that can be
directly measured using tests, assessments, and
surveys, and are used to define a given latent
construct. Latent constructs cannot be directly
observed or measured and, as a result, must
be inferred from a set of observed variables.
In our example, agreement on therapy tasks,
agreement on therapy goals, and emotional
bond are observed variables that are directly
32
SEM models combine path and factor analytic models allowing for the incorporation of
both observed and latent variables into a model.
SEM procedures ultimately determine the plausibility of a theoretical model by comparing
Pthe
estimated theoretical covariance matrix
to
the observed covariance matrix S (i.e., the
matrix derived from the sample data; Schumacker & Lomax, 2010). Many SEM software
programs are currently available to researchers.
These include LISREL1, AMOS, EQS1,
Mx, Mplus1, Ramona, and SEPATH1. Many
of the SEM software programs allow researchers to statistically analyze raw data and provide
procedures for managing missing data, outliers,
and variable transformations. Programs, such
as AMOS and LISREL1, offer researchers
the option to construct a path diagram that can
be translated by the software program into the
mathematical equations needed for analysis.
Crockett
33
Model Specification
Model specification is the first step of SEM
analysis and occurs prior to data collection and
analysis. It is often the most difficult step for
researchers as it involves the development of
a theoretical model using applicable, related
theory and research to determine variables of
interest and the relationships among them
(Cooley, 1978). It is critical that the hypothesized theoretical model be grounded in and
derived from the extant literature. The
researcher must be able to provide plausible
explanations for relationships included in the
model and a rationale for the overall specification of a model. The example theoretical
model attempts to specify the relationship
between supervisor multicultural competence
and supervisee outcomes. The model
34
Model identification
Model estimation
Model testing
Model modification
Description of Step
This step involves the specification of a theoretical model that utilizes applicable,
related theory and research to determine the latent and observed variables of
interest and the relationships among them. In particular, researcher must specify a
measurement and structural model. A path diagram can be constructed to visually
represent the hypothesized relationships among variable in the theoretical model
This step helps the researcher to determine whether the specified model is capable of
producing actual results that can be estimated in SEM analysis. Models must be
indentified and able to generate a unique solution and parameter estimates.
OBriens (1994) criteria can be used to establish whether a measurement model is
identified. To determine whether a structural model is indentified researchers can
use Bollens (1989) recursive rule and the t rule
This step involves the use of an iterative
P procedure (i.e., fitting function) to generate
the theoretical covariance matrix , as wellPas minimize the differences between
the estimated theoretical covariance matrix
and the observed covariance matrix
S. Maximum likelihood (ML) and generalized least squares (GLS) are the most
commonly used fitting functions
This step involves the analysis of both the measurement and structural models in order
to determine (a) the global fit of the entire model, and (b) the fit of individual model
parameters. Multiple indices of fit (i.e., absolute, comparative, and parsimonious)
should be analyzed to determine the degree to which the theoretical model fits the
sample data. The w2 difference test can also be used when working with nested
models to compare the plausibility of the theoretical model to viable alterative
models. It should be noted that the measurement model must yield a good fit to the
data before the structural model can be analyzed
The final step involves using theory trimming or the addition of new parameters to
attempt to improve the theoretical models fit to the data. Researchers should be
advised to model modification is an exploratory procedure and is based on the
sample data instead of the extant literature. Respecified models will need to be
cross-validated with a new sample
Crockett
35
9
Goals = function of supervisory
working alliance error;
Bond = function of supervisory
working alliance error;
The structural equations specify the estimation of three structure coefficients (i.e., elements that comprise
P the estimated theoretical
covariance matrix ). Each equation contains
a prediction error which specifies the degree
of variance in the latent endogenous variable
that is not accounted for by the other variables
in the equation (Schumacker & Lomax, 2010).
Finally, the equations specify the direction of
the predicted relationships.
The hypothesized relationships among
observed and latent variables in a theoretical
model can also be illustrated through a path
diagram (i.e., a graphical representation of the
theoretical model). Such diagrams use a series
of conventional symbols to depict the relationships among model variables (see Figure 1). A
rectangle represents an observed variable,
whereas an oval denotes a latent variable. Unidirectional arrows indicate a hypothesized
relationship in which one variable influences
another. These arrows are often referred to as
model paths. Bidirectional, curved arrows are
used to denote covariance between two independent variables. Finally, the measurement
error for each observed, dependent variable
36
Observed Variable
Latent Variable
Nonrecursive relationship
Model Identification
Model identification is a requirement for producing results that can be estimated in SEM analysis.
This step occurs prior to estimating model parameters (i.e., relationships among variables in the
model) and is concerned with whether a unique
solution to the model can be generated. For a
model to be considered identified, it must be
theoretically possible to establish a unique
estimate for each parameter (Kelloway, 1998;
Schumacker & Lomax, 2010) and is dependent
on the designation of model parameters as free
(i.e., a parameter that is unknown and needs to
be estimated), fixed (i.e., a parameter that is fixed
at a specific value, often a 0 or 1), or constrained
(i.e., a parameter that is unknown, but constrained to equal one or more other parameters).
For example, a theoretical model that exerts x
y 20 has no sole solution; the value of x could
be 10, 15, or 19. In order to find a unique solution
Crockett
37
e
e
Goal
Task
Bond
Supervisory
Working
Alliance
Supervisor
Multicultural
Competence
Counseling
Self-Efficacy
Microskill
Process
Difficult
Culture
Value
Figure 2. This path diagram depicts the hypothesized direct and indirect relationships among supervisor multicultural competence, the supervisory working alliance, and supervisee counseling self-efficacy as specified by the
example theoretical model.
38
Model Estimation
Model Testing
Model estimation, the third step of SEM analysis, involves estimating the parameters of the
theoretical model in such a way that the theoretical parameter values yield a covariance
matrix as close as possible to the observed
covariance matrix S. SEM analysis programs
use an iterative procedure, often referred to as
a fitting function, to minimize the differences
betweenPthe estimated theoretical covariance
matrix
and the observed covariance matrix
S. Specifically, the iterative procedure attempts
to improve the preliminary parameter estimates
with subsequent calculation cycles. The final
parameter estimates represent the best fit to
observed covariance matrix S.
Several fitting functions are available to
researchers (e.g., ordinary least squares [OLS],
generalized least squares [GLS], maximum likelihood [ML]). ML is the most widely used type
of estimation, followed by GLS (Kelloway,
1998). Although ML and GLS are comparable
to OLS estimation used in multiple regression,
they are slightly different from and several
advantages over OLS estimation. In particular,
ML and GLS are (a) not scale-dependent, (b)
allow dichotomous exogenous variables (Skosireva, 2010), and (c) consistent and asymptotically efficient in large samples (Bollen, 1989;
Kelloway, 1998; Schumacker & Lomax,
2010). ML and GLS assume multivariate normality of dependent variables and, unlike OLS,
are full information techniques, meaning that
they estimate all model parameters simultaneously to produce a full estimation model. When
the assumption of multivariate normality is violated, researchers may use an asymptotically
distribution-free (ADF) estimator. ADF is not
dependent on the underlying distribution of the
data, but it does require a large sample size as the
estimator yields inaccurate chi-square (w2) statistics for smaller sample sizes (Mueller, 1996). For
more information on ADF, please see Raykov
and Widaman (1995). In the example model,
ML was employed by LISERL1 during the
SEM analysis to minimize the differences
betweenPthe estimated theoretical covariance
matrix
and the observed covariance matrix S.
As mentioned earlier SEM allows for the simultaneous analysis of direct and indirect relationships among latent and observed variables;
however, many researchers (e.g., Anderson &
Gerbing, 1988; James, Mulaik, & Brett, 1982)
recommend a two-step approach to model testing. In particular, James, Mulaik, and Brett
(1982) argued that model testing involved the
analysis of two conceptually distinct models: the
measurement model and the structural model.
The researcher must first determine whether the
proposed measurement model holds, ensuring
that the chosen observed indicators for a latent
construct actually measure the construct. If the
chosen indicators for a construct do not accurately measure the construct, then the structural
model is meaningless (Joreskog & Sorbom,
1993). Accordingly, it is recommended that
researchers conduct a CFA of the measurement
model to determine whether the factor indicators
loaded on the latent variables in the direction
expected prior to testing the structural model.
A CFA of the example measurement model
was run prior to estimating the structural model
to ensure that all factors loaded on the latent
variables in the direction expected. Results
indicated an adequate fit of the CFA model,
w2(19) 44.72, p < .05; root mean square
error of approximation (RMSEA) .07; comparative fit index (CFI) .97; Parsimonious
normed fit index (PNFI) .65, to the data (see
information on model fit). The standardized
parameter estimates were significant at the
p < .05 level and consistent with the specified
hypotheses, loading in the appropriate direction. The individual parameters comprising the
model were also analyzed. As predicted, the
latent variable supervisory working alliance
was significantly positively correlated with its
factor indicators: WAI-SF bond subscale (r
.83, p < .05), WAI-SF task subscale (r .92,
p < .05), and WAI-SF goal subscale (r .82,
p < .05). The latent variable supervisor CSE
was also significantly positively correlated
with its factor indicators: COSE microskills
subscale (r .83, p < .05), COSE counseling
process subscale (r .79, p < .05), COSE
Crockett
39
40
Crockett
41
Microskill
Process
Supervisor
Multicultural
Competence
Counseling
Self-Efficacy
Difficult
Culture
Value
Figure 3. This path diagram depicts the hypothesized relationship between supervisor multicultural competence and supervisee counseling self-efficacy as specified by the direct, alternative model.
42
Task
Goal
Bond
Microskill
Process
Supervisor
Multicultural
Competence
Supervisory
Working
Alliance
Counseling
Self-Efficacy
Difficult
Culture
Value
Figure 4. This path diagram depicts the hypothesized mediated relationship between supervisor multicultural
competence and supervisee counseling self-efficacy through the supervisory working alliance as specified by the
indirect, alternative model.
Crockett
43
12
Model Modification
The final step of SEM involves model modification. In this step, researchers employ model
modification methods in an attempt to find a
model that better fits the data. First, researchers
need to perform a specification search that
involves eliminating nonsignificant parameters
from the theoretical model (i.e., theory trimming) and examining the models standardized
residual matrix (i.e., fitted residuals). The most
commonly used procedures for eliminating
parameters include comparing the t statistic for
each parameter to the tabled t value to determine statistical significance (Schumacker &
Lomax, 2010). When examining the standardized residual matrix, researchers should look
to see that all of the values are small in magnitude. Overall large values in the matrix indicate
a misspecification of the general model, while
large values across a single variable point to the
misspecification of that variable only (Bentler,
1989). While the preceding procedures can
improve model fit, they remain highly controversial. Specification searches are exploratory
in nature and are based on the sample data
instead of previous theory and research, as a
result parameters eliminated from the theoretical model may reflect sample characteristics
that do not generalize to the broader population
(Kelloway, 1998). Additionally, model modification may lead to an inflation of Type I error
rates and be misleading (Kelloway, 1998). For
this reason, researchers should strive to balance
the elimination of parameters to the model with
improving the fit of the model.
Conclusions
This article provided an introduction to
SEM techniques and described five applied
steps for conducting SEM analysis: model specification, identification, estimation, testing,
and modification. SEM has several advantages
over first-generation multivariate methods
such as multiple regression and can used to
verify complex relationship among observed
and latent variables in counseling research.
Counseling researchers can use this statistical
44
Crockett
45
Baron, R. M., & Kenny, D. A. (1986). The moderatormediator variable distinction in social psychological research: Conceptual, strategic, and statistical
considerations. Journal of Personality and Social
Psychology, 51, 11731182. doi:10.1037/00223514.51.6.1173
Bentler, P. M. (1989). EQS: A structural equations
program manual. Los Angeles, CA: BMDP Statistical Software.
Bentler, P. M., & Chou, C. (1987). Practical issues
in structural equation modeling. Sociological
Methods and Research, 16, 78117.
Bollen, K. A. (1989). Structural equations with
latent variables. New York, NY: John Wiley.
Bollen, K. A., & Long, J. S. (1993). Testing
structural equation models. Beverley Hills,
CA: Sage.
Bullock-Yowell, E., Peterson, G. W., Reardon, R. C.,
Leirer, S. J., & Reed, C. A. (2011). Relationships
among career and life stress, negative career
thoughts, and career decision state: A cognitive
information processing perspective. Career Development Quarterly, 59, 302314.
Chan, F., Lee, G. K., Lee, E. J., Kubota, C., & Allen,
C. A. (2007). Structural equation modeling in
rehabilitation counseling research. Rehabilitation
Counseling Bulletin, 51, 5366.
Chao, R., Chu-Lien, N., & Sanjay, R. (2011). The
role of ethnic identity, gender roles, and multicultural training in college counselors multicultural counseling competence: A mediation
model. Journal of College Counseling, 14,
5064.
Cochran, D. B., Wang, E. W., Stevenson, S. J.,
Johnson, L. E., & Crews, C. (2010). Adolescent occupational aspirations: Test of Gottfredsons
theory
of
circumscription
and
compromise. Career Development Quarterly,
59, 412427.
Cooley, W. W. (1978). Explanatory observational
studies. Educational Researcher, 7, 915.
Costello, A. B., & Osborne, J. (2005). Best practices
in factor analysis: Four recommendations for getting the most from your analysis. Practical
Assessment Research & Evaluation, 10, 19.
Crockett, S. A. (2011). The role of supervisorsupervisee cultural differences, supervisor multicultural competence, and the supervisory working
alliance in supervision outcomes: A moderated
46
Crockett
47
Bio
Stephanie A. Crockett, PhD, is an assistant professor in the Department of Counseling of the School of
Education and Human Services at Oakland University. Her primary research interests include the use
of SEM in counseling research, testing/assessment
and outcome research in career counseling, and multicultural competence in clinical supervision.