Вы находитесь на странице: 1из 12

Structural Equation Modeling:

A Primer for Health Behavior Researchers


Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

Eric R. Buhi, MPH, PhD; Patricia Goodson, PhD; Torsten B. Neilands, PhD

Objective: To introduce the state that can be tested with empirical


of the art of structural equation data. Although use of SEM alone
modeling (SEM). Method: This is not a magic solution, new soft-
primer is organized in a manner ware developments provide users
allowing readers to review any with unparalleled flexibility for
one of 5 freestanding sections. improving research. Conclusion:
Copyright (c) PNG Publications. All rights reserved.

Results: SEM maintains several SEM must be thrust into the daily
advantages over regression and vocabulary and routine practice
other multivariate techniques. of health behavior researchers.
Through a 2-step modeling pro- Key words: health behavior re-
cess, SEM strengthens research search, structural equation models,
by allowing for the specification SEM, multivariate statistics, primer
of complex, theory-driven models Am J Health Behav. 2007;31(1):74-85

S
tructural equation modeling (SEM) 7 reports in the American Journal of Health
is a powerful multivariate statisti- Behavior, 5 in the American Journal of
cal method being used in the social Health Promotion, and 7 in Health Educa-
sciences with increasing frequency.1,2 In tion and Behavior which utilized SEM (Table
the psychological literature, SEM cita- 1). This amounted to approximately 1 in
tions have risen since 1979 (Figure 1); 53 data-based journal articles using SEM.
SEM now rivals analysis of variance Furthermore, only a fraction of the jour-
(ANOVA) in statistical method popular- nals’ reports used some other multivari-
ity.3 In health behavior research, how- ate technique, such as multivariate
ever, SEM has yet to reach such popular- analysis of variance (MANOVA) or canoni-
ity. cal correlation analysis (CCA; Table 1).
In an electronic search of articles pub- According to the American Academy of
lished between 1996 and 2004 in 3 health Health Behavior Work Group on Doctoral
behavior research journals, we found only Research Training,4 “a working knowl-
edge of multivariate statistical procedures
is crucial for generating high quality re-
Eric R. Buhi, Assistant Professor, Department search and answering complex questions”
of Community and Family Health, University of (p. 554). Why, then, are health behavior
South Florida, Tampa, FL. Patricia Goodson, As- and health promotion researchers not
sociate Professor, Department of Health and Ki- going beyond univariate/bivariate proce-
nesiology, Texas A&M University, College Sta- dures, such as ANOVA and regression, in
tion, TX. Torsten B. Neilands, Adjunct Assistant
Professor, Center for AIDS Prevention Studies,
their research efforts? Although it is be-
University of California San Francisco, San Fran- yond the scope of the current paper to fully
cisco, CA. answer this question, we speculate there
Address correspondence to Dr Buhi, Depart- may be 3 reasons.
ment of Community and Family Health, 13201 First, some researchers may not know
Bruce B. Downs Boulevard, MDC 56, University why SEM is particularly useful or valu-
of South Florida, Tampa, FL 33612. E-mail: able. Because SEM is a relatively new
ebuhi@health.usf.edu tool, classes and trainings addressing

74
Buhi et al

Figure 1
Citation Frequencies of SEM and (M)ANOVA in the PsychINFO
Database Between 1979 and 2002
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37
Copyright (c) PNG Publications. All rights reserved.

Note.
The numbers are standardized with respect to the total number of records per year. Figure
reproduced with permission from MPR Online.

SEM’s value have only recently been de- introduce—or update, depending upon the
veloped. Thus, researchers’ knowledge of reader’s familiarity—the state of the art
the assets afforded by SEM, such as the in SEM as a multivariate analytic tech-
ability to account for measurement error nique in health behavior research. This
in the modeling process, may be limited. primer is organized in a manner allowing
Second, some researchers may not know readers, both new and experienced SEM
how to conduct SEM. SEM analyses in- users, to review any one of 5 freestanding
volve a distinctive nomenclature that can sections most pertinent to their needs or
be intimidating for novices. Further, the knowledge level. In section 1, we define
complexities of models available within the purpose of SEM; and in section 2, we
SEM analyses are mirrored in complex present SEM’s strengths as a multivari-
SEM software packages, which can also ate analytic tool. These sections are most
be intimidating for the uninitiated ana- useful for individuals questioning why
lyst. Third, some researchers may not be SEM might be valuable. For those unfa-
familiar with resource materials and com- miliar with “how to do SEM,” sections 3
puter programs available to aid in SEM and 4 are particularly helpful. In section
analyses. The past 3 decades have been 3, we provide an overview of the basic
filled with rapid developments in SEM steps involved in conducting SEM analy-
theory and software. For instance, the ses. In the fourth section, we present
ability to model dichotomous dependent cautionary notes related to using SEM.
variables (eg, ever had sexual intercourse Finally, in section 5, we review a host of
= yes/no) in SEM software packages such available resources and provide a com-
as Mplus became widely available only parative treatment of SEM software
within the last few years. Even if an packages. This concluding section may
individual took a graduate-level SEM prove useful for beginners and seasoned
course as recently as 5 years ago, he or “SEMers” alike. It will be most helpful
she might not know about these develop- for researchers seeking analytic mate-
ments. rials or programs that can aid in SEM
The purpose of this paper, then, is to analyses.

™ 2007;31(1):74-85
Am J Health Behav.™ 75
Structural Equation Modeling

Table 1
Utilization of Structural Equation Modeling and Other
Multivariate Analytic Techniques in Published Articles
From 3 Health Behavior Research Journals, 1996-2004
Journal Name # of Articles # of Articles Using Total # of Articles # of Data-based Articles Published,
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

Using SEMa Other Multivariate Using Multivariate Including in Supplements and Special
Techniquesb Techniques Issues

American Journal of Health Behavior 7 24 31 416


American Journal of Health Promotion 5 15 20 325
Health Education and Behavior 7 14 21 275c

Note.
a Boolean search in MEDLINE/PsycINFO: (structural equation modeling) or (SEM) or (LISREL) or
(AMOS) or (EQS) or (mplus) AND (American Journal of Health Behavior) AND (American Journal
of Health Promotion) AND (Health Education and Behavior)
b Boolean search in MEDLINE/PsycINFO: (MANOVA) or (multivariate analysis of variance) or
(cluster analysis) or (factor analysis) or (multidimensional scaling) or (CCA) or (canonical
correlation analysis) or (discriminant analysis) AND (American Journal of Health Behavior) AND
Copyright (c) PNG Publications. All rights reserved.

(American Journal of Health Promotion) AND (Health Education and Behavior)


c Issues from 1997 (volume 24, issue 5) to 2004. Prior to this volume/issue, journal was Health
Education Quarterly, and we do not have access to these issues

1. What Is the Purpose of SEM? ANOVA, and regression) as well as com-


Structural equation modeling includes plex ones, such as multilevel modeling13,14
a wide range of multivariate methods (eg, examining youth-within-classrooms-
aimed at examining the underlying rela- within-schools) and latent curve model-
tionships, or structure, among variables ing (which examines change as a con-
in a model. SEM was created to test and tinuous process over time).14
refine theoretical models attempting to
explain or predict social or behavioral 2. Why Use SEM?
phenomena.5-7 Understanding these phe- Although SEM’s increased application
nomena allows us to appreciate “why in the social sciences partially stems
people engage in health-risk or health- from improvements in software pack-
compromising behavior and why (as well ages,15 we argue such increasing use has
as how) they adopt health protective been driven by 4 factors. First, multivari-
behavior.” 8,p.1 These theoretical models ate methods such as SEM best honor the
inform the development and improvement reality to which investigators are attempt-
of health-related interventions. Moreover, ing to generalize.16 In health behavior
SEM is a useful tool in estimating these research, most outcomes (ie, behaviors)
interventions’ effects.9 have multiple causes (ie, predictors), and
Often referred to as causal, path, latent most causes have multiple outcomes, all
variable, or covariance structure models, interacting dynamically. Health behavior
SEM is similar to regression (and other researchers investigate multivariate, not
correlational methods) because it belongs univariate/bivariate or isolated, phenom-
to the general linear model (GLM) family. ena with only one or 2 determinants. It is
For instance, SEM and regression analy- impossible to assess how multiple vari-
ses both rely on a linear combination of ables behave in each other’s company
variables, use weights (eg, β weights) to when a researcher limits an analysis to a
optimize the explained variance and mini- univariate/bivariate examination. In-
mize model error variance, focus on la- stead, SEM allows all variables—multiple
tent (or not directly observed) variables, independent and dependent variables—
and yield variance-accounted-for effect to be examined simultaneously.
sizes (eg, R2, eta2).10 In short, SEM sub- Second, multivariate methods such as
sumes a range of other analytic meth- SEM control for inflation of experimentwise
ods11,12 and may be utilized to conduct both (EW or Type I) error. Type I error is defined
simple analyses (including t-tests, by alpha (α), usually set at .05, and is “the

76
Buhi et al

probability of getting a result … that leads researcher basically “dumps in” all the
to an incorrect decision to reject the null variables, SEM allows researchers to test
hypothesis.”17,p.38 Inflated EW error may theories and assumptions directly by
occur when a researcher conducts mul- specifying which variables are related to
tiple univariate/bivariate tests (ie, with other variables. That is, the researcher
a single dependent variable or hypoth- can test some paths (or relationships) but
esis, such as in ANOVA) with a single not others in the analysis. Some SEM
sample’s data. These analyses can lead a programs even allow researchers to draw
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

researcher to falsely reject too many of these hypothesized relationships visu-


the hypotheses being tested.18 For ex- ally and fit the drawn model to the under-
ample, assume a researcher conducts c lying data, an intuitive and user-
number of tests in a single study (eg, 20 friendly process. Finally, SEM allows re-
statistical significance tests), each at α = searchers to examine relationships
.05. Using the following formula, the risk among latent variables with multiple ob-
of making a Type I error across the entire served measures. The relationships
set of tests is 64%.17 among latent variables, thus, are purged
of measurement error, leading to more
α
EW = 1 – (1 - α)c accurate and often stronger relationships
between latent variables than what would
α
EW = 1 – (1 - .05)20 = .64 be observed using multivariate methods
Copyright (c) PNG Publications. All rights reserved.

that consider observed variables only (eg,


Further, out of 100 statistical tests MANOVA or even regression). In short,
conducted, the researcher could be re- although these older techniques assume
jecting the null hypothesis, incorrectly, zero measurement error in sample data
64 times. In these cases, the probability (which is never the case), SEM controls for
of making one or more Type I errors can measurement error.
be very serious (see Fish19 for examples), Lastly, SEM is useful because it en-
and the implications for health behavior ables the advanced treatment of incom-
research can be sobering. If a researcher plete data. Missing data in health behav-
conducts multiple univariate/bivariate ior research can represent an important
statistical tests to investigate determi- problem during analyses. SEM software
nants of an individual’s involvement in a developers have dealt well with the prob-
health-risk behavior, the researcher may lem of missing data by incorporating so-
erroneously (due to Type I error) conclude phisticated missing data techniques—
that a predictor is associated with the such as optimal full information maxi-
outcome when, in reality, it is not. Em- mum likelihood (FIML)20,21—ahead of the
ploying multivariate methods such as general purpose software vendors (eg,
SEM, however, can correct “up front” for SPSS, SAS, and Stata). Thus, ANOVA,
this analytic limitation by avoiding the regression, MANOVA, and ANCOVA can
use of multiple univariate/bivariate tests be conducted using SEM programs with or
and, instead, testing hypotheses/research without incomplete data, and a researcher
questions across several variables at once. can thereby capitalize on the more so-
Third, researchers have begun to real- phisticated missing-data-handling capa-
ize the utility of SEM over other multi- bilities (Table 2 lists SEM software pro-
variate analytic methods. SEM gives grams capable of conducting FIML or
health behavior researchers unparalleled multiple imputation). These capabilities
flexibility in specifying theory-driven allow researchers to proceed with SEM or
models that can be tested with empirical other analyses as if there were no miss-
data. SEM goes further than older multi- ing data, if certain missing data assump-
variate techniques, such as MANOVA and tions are met.22,23
CCA, by allowing users to automatically
and efficiently compute indirect, direct, 3. What Are the Basic Steps of SEM?
and total effects in complex models, in- In SEM, the researcher utilizes the
cluding models that evaluate statistical theoretical literature to specify a health
mediation in which an exogenous predic- behavior model for testing. The researcher
tor variable X impacts an intermediary subsequently determines how to mea-
variable Y that in turn exerts influence sure the variables pertinent to the theory
on a distal outcome Z. Additionally, un- and collects data, for instance, using a
like these older techniques in which the survey instrument. Next, he or she passes

™ 2007;31(1):74-85
Am J Health Behav.™ 77
Structural Equation Modeling

Table 2
A Comparison of Various SEM Features by Program Package
Program Name
Amos 5 EQS 6 Mplus 3 LISREL 8 PROC CALIS Mx
Key Features www.spss.com/amos/ www.mvsoft.com/ www.statmodel.com/ www.ssicentral.com/ www.sas.com/ http://www.vcu.edu/mx/
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

Has the ability to handle U A U U B U


missing data via FIML
or MI

Offers accessible/online U U U U U U
technical support

Allows for graphical U U U U


forming of models

Can conduct analyses with U U U U


categorical outcome variables

Has the ability to fit multilevel U U C U U


or hierarchical SEM models

Has the ability to model U U U U U U


Copyright (c) PNG Publications. All rights reserved.

non-normal continuous data

Has the ability to model non- U U U


normal continuous incomplete data

Has the ability to model U U


categorical incomplete data

Computes direct, indirect, D E U U E U


and total effects and associated
asymmetric confidence intervals

Offers tests for multivariate U U F U U


normality

PricingG 599(Academic), 595(Academic), 595(Academic), 495(Academic Proc Calis is This program is


in dollars 549(Government), 695(Government/ 695(Commercial/ and included with freely available
999(Commercial)G Corporate)G Government/ Commercial)G SAS. Contact local through
Non-profit)G sales office through the Mx website.
SAS website.

Comments * Easy to use, * Provides models * Very flexible overall * Program syntax * Can manage * Does many of
quick learning in a graphical * Excellent online is sufficiently missing values the usual SEM
curve form, using a technical support complex to with PROC MIB analyses (and
* Provides models drawing tool (Linda Muthén, require an * Non-graphical then some) and
and model results * Comes in one of the software unusually care- format, requires is free to try
in a graphical Windows or authors, handles ful attention to knowledge of * Offers model
form, using a Macintosh all online queries) detail on the syntax fit indexes found
drawing tool, which versions * Can only read part of the in the major
may be useful for data from a text analyst commercial
more “visual” (.dat) file. * Detractors refer programs, such
analysts to its user inter- as LISREL, EQS,
* Reads a wide variety face and the and Amos
of data file formats, level of * Can only read
including SPSS and sophistication. It data from a text
Excel is more compli- (.dat) file
cated than other
programs

Note.
A Implements ML-EM procedure, which is similar to Arbuckle’s FIML.20
B Multiple imputation available through SAS PROC MI (available in version 9.x)
C LISREL 8.72 for Windows includes a multilevel SEM module which allows general 2-level structural equation models.
D Can compute as long as there are no missing data.
E Can compute direct, indirect, and total effects, but not asymmetric CIs.
F Available for mixture models only.
G Pricing is current as of summer 2005, and is listed for the basic-level package only. Add-ons, user’s guides, and technical support may not be included.

data to an SEM software package, which analyses entail essentially a 2-step mod-
fits the data to the specified model and eling process25 of building and testing (a) a
produces results, including model fit sta- measurement model and (b) a structural
tistics and parameter estimates.24 SEM model. Although these 2 are the funda-

78
Buhi et al

mental analytic steps, there is an addi- ric sense, this step is imperative in as-
tional “up front” step in the analytic pro- certaining the validity of the constructs.
cess: examining the critical assumption In Figure 2, for example, 2 sample mea-
of multivariate normality of the data. surement models (within the narrow dot-
ted boxes) have been formulated using a
Examining Multivariate Normality drawing tool found in one SEM package,
Prior to any analyses, the researcher Amos. This measurement model’s pur-
should test a critical assumption under- pose is to take into account measure-
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

lying both SEM and other multivariate ment error in all variables which are not
techniques: the assumption regarding directly observable (eg, the constructs of
normality of the distribution of multivari- behavioral intentions, perceived norms,
ate data. Before testing this assumption, attitudes, or motivation).7 In short, the
however, analysts must first assess shared variance derived from the correla-
univariate normality because “normality tions/covariances among multiple ob-
on each of the variables [in a model] served variables, such as survey items
separately is a necessary, but not suffi- (the boxes in Figure 2), is used to infer the
cient, condition for univariate normality presence of a common latent factor (the
to hold.”26,p.262 Testing multivariate nor- ovals in the model). The software package
mality can be accomplished (in both SPSS converts survey data from Excel, SPSS, or
and SAS) graphically—by examining nor- raw data (in ASCII text format) into cova-
Copyright (c) PNG Publications. All rights reserved.

mal probability plots or Q-Q Plots—or riances and means which are then used
nongraphically—by assessing skewness in subsequent analyses.24 In the Figure 2
and kurtosis coefficients, or through sta- sample model, the boxes—items E1-E6,
tistical testing with the Shapiro-Wilk test. N1-N5, and ABSTINENCE—are observed
Once the univariate normality assump- measurement items, or indicators (E1-E6
tion has been evaluated and met, the and N1-N5 are scales). The ovals in the
multivariate normality assumption can model—INTENTION and NORMS—are the
be assessed. Because ordinary least latent factors, or constructs being repre-
squares, generalized least squares, and sented by their respective scales.
maximum likelihood statistical estima- The measurement model step is
tion theories all presume a multivariate equivalent to performing a confirmatory
normal distribution, not meeting this factor analysis. In this approach, the num-
assumption can be problematic, particu- bers of hypothesized underlying con-
larly when assessing statistical signifi- structs (ie, factors) are specified by the
cance.27 In SEM, nonnormality can result literature (the theory under examina-
in an underestimation of overall model tion) and researcher a priori, and the
fit, downwardly biased parameter esti- model is fitted to sample data to assess its
mates, and underestimated standard er- convergent and discriminant validity
rors.28 Many SEM software packages such (which together provide evidence for con-
as EQS and Amos offer multivariate nor- struct validity). Ideally, after an accept-
mality tests, such as Mardia’s measure of able fit is achieved, the measurement
multivariate kurtosis,29,30 which can be model is then cross-validated using a
carried out with a single mouse click. second set of sample data. In other words,
Recent advances in many packages even in this step the researcher assesses how
make it possible for researchers to ana- well the scales measure the latent con-
lyze nonnormal continuous and categori- structs which will be included in the
cal data (Table 2). structural model.
Once the measurement model has been
Using the 2-Step Modeling Approach formulated and tested, a structural model
Once assumptions are evaluated and is estimated as the second step. In a
met, the researcher may commence with structural model, the goal is to examine
the 2-step modeling process of building the underlying relationship, or structure,
and testing a (a) measurement model and between the latent constructs tested in
(b) structural model. First, the researcher the measurement model and other (ob-
begins by building and testing the mea- served) variables proposed by the theory
surement model within the SEM software (see Figure 2, within the wide-dotted box).
package. The purpose of the measure- This structure accounts for the direct,
ment model step is to test indicator/ indirect, and total effects among factors.31
construct relationships. In a psychomet- A direct effect is the directional relation-

™ 2007;31(1):74-85
Am J Health Behav.™ 79
Structural Equation Modeling

Figure 2
Sample Measurement and Structural Model

Measurement Model A
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

e1 X E1 - Have oral sex before marriage X


Structural Model
e2 X E2 - Have vaginal sex during next year X

E3 - Touch someone else's sexual organs


e3 X W
during the next year β = 0.50
Intention to Have Sexually Abstinent
Sex before Marriage W X Behavior
(INTENTION) (ABSTINENCE)
e4 X E4 - Have oral sex during the next year W
W
S
e5 X E5 - Have vaginal sex before marriage X
Copyright (c) PNG Publications. All rights reserved.

e6 E6 - Touch someone else's sexual X


X organs before marriage β = 0.25
β = 0.35

Subjective Norms
about Abstinence
(Norms)

X X W W
N1 - Most people N3 - Most people N4 - My friends N5 - Most of my
my age think they my age think it's think that absti- friends intend to
should wait until OK to have sex with nence until mar- wait to have sex
marriage to have a serious boy- riage is the best until they get mar-
sex. friend or girlfriend. choice. ried.
S S S S

e7 e8 e9 e10

Measurement Model B

ship between 2 variables and is the type of effect on ABSTINENCE


relationship usually examined through (NORMS→INTENTION*INTENTION→ABSTINENCE
ANOVA and regression. An indirect effect = 0.175). To compute the total effect of
is an independent variable’s influence on NORMS on ABSTINENCE, one would take
a dependent variable, through a single or the sum of the NORMS→ABSTINENCE
possibly multiple mediating variables.32 direct effect and the
The standardized indirect effect is the prod- NORMS→INTENTION*INTENTION→ABSTINENCE
uct of the standardized direct effects. Us- indirect effect (0.25+0.175=0.425).
ing the model in Figure 2 as an example, To test the fit of sample data to the
if NORMS has a direct effect on INTEN- structural model, SEM software exam-
TION (β=0.35), and INTENTION has a di- ines covariances rather than individual
rect effect on ABSTINENCE (β=0.50), then cases (as happens in regression tech-
NORMS can be said to have an indirect niques). To examine relationships in SEM,

80
Buhi et al

matrix algebra is used to account for behavior research studies, ratios between
variances of each variable and covari- 2 and 5 have often been employed.
ances of each pair of variables.26 This Hu and Bentler33 further recommend
covariation makes SEM a more appli- assessing and reporting results from sev-
cable and generalizable technique than eral approximate fit indexes because one
regression, allowing for the simultaneous or more are insensitive to sample size
examination of multiple independent and and/or impervious to estimation meth-
dependent variables. It is important to ods (eg, maximum likelihood or general-
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

bear in mind that these variables are ized least squares). The Comparative Fit
known in SEM, respectively, as exogenous Index (CFI) and Tucker-Lewis Index (TLI),
and endogenous variables. The name ex- for example, indicate how much better
ogenous variable means that the cause of the model fits the sample data than a null
the variable is determined outside of the model, which stipulates that there are no
specified model. Endogenous variables, on common factors. Unlike the CFI, how-
the other hand, are determined within ever, the TLI moderately corrects for model
the model (that is, endogenous variables parsimony. The Root Mean Square Error
are hypothesized to be predicted by other of Approximation (RMSEA) is similar to
variables in the model). the chi-square test in that it is more or
less a “badness of fit” test. Possible values
Assessing Model Fit for these fit indexes range between 0 and
Copyright (c) PNG Publications. All rights reserved.

A strength of SEM is that the analyst 1, although TLI can exceed 1. With re-
obtains both a global assessment of model gards to what a good fit actually means,
fit and tests of individual parameters. there is little consensus. Cutoff values of
The researcher begins by evaluating glo- 0.95 for TLI/CFI (the higher the better)
bal model fit. Quantifying the correspon- and .06 for RMSEA (the lower the better)
dence between the predicted covariances have been suggested.34 However, Browne
and the observed covariances (which is and Cudeck35 suggested that fair fitting
the analytic focus of SEM) generates a models have RMSEA values in the range
goodness-of-fit value or index. In classical of 0.05 to 0.08, even as Marsh et al36
statistics, effect sizes characterize the fit cautioned against broadly applying these
of a model to data (eg, R² for a regression “golden rules of fit” without first consider-
model). Similarly, in SEM, fit indexes may ing their limitations. In brief, then, val-
be thought of as effect sizes. Consulting ues further away from the recommended
these indexes and checking for model fit cutoff points indicate potential inconsis-
can lead to important model improve- tency between the model and sample data,
ments. For instance, specific paths can whereas values near the recommenda-
be re-drawn to hypothesize new relation- tions suggest that the model might be
ships or the entire model can be useful. Once global fit is met, then the
respecified to exclude factors with weak researcher examines individual param-
explanatory power. eter estimates and confidence intervals
Although there are a number of fit to learn which paths denote the strongest
indexes available, unfortunately there is relationships or explain the greatest
not one index appropriate for all analytic amount of model variance.
conditions. There are, nevertheless, gen- In sum, the theory the researcher is
eral fit index “rules of thumb” to consider examining guides the measurement and
with recommended cutoff values. Hu and structural modeling process by specifying
Bentler,33 for example, suggest that re- (a) the number of underlying constructs
searchers always examine and report chi- in the model and (b) the interrelation-
square (χ2) for exact fit, which tests whether ships of these constructs. Results pro-
there is a statistically significant differ- duced by the software package, such as
ence between the model and the sample model fit statistics and parameter esti-
data and degrees of freedom (df) for each mates, are used to test and improve over-
model estimated. Because χ 2 can be all model fit. For good examples of this
heavily influenced by sample size, how- modeling process, see Blue et al,37 Park et
ever, the χ2/df ratio may be reported. al,38 and Saunders et al.39
According to Bollen,31 there is little consen-
sus, for the χ2/df ratio, on what represents 4. What Cautions Should SEM Users
a “good fit,” with recommendations as high Exercise?
as 5 and as low as 2 (or less). In health Although SEM is a sophisticated ana-

™ 2007;31(1):74-85
Am J Health Behav.™ 81
Structural Equation Modeling

lytic tool for testing theoretical models Lastly, a single variable Y must be isolated
with multiple endogenous/exogenous from all influences, with exception of a
variables, its application alone does not second variable X. If a change in X accom-
resolve (or even address) the limitations panies a change in Y, then it can be said
of behavioral and social science research. that Y causes X. In reality (eg, in health
When considering SEM, researchers must behavior research), however, the isola-
exercise the following 4 cautions con- tion of Y from other variables of influence
cerning its utilization. is virtually impossible. Thus, according to
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

Bollen,31 all models must be looked upon


SEM Does Not Compensate for Poorly as estimations of reality. So, SEM may
Conceived Ideas or Weak Theoretical give an indication of causal relations but,
Grounding by itself, SEM cannot ensure that asso-
Statistics is an important tool in data ciation, temporal priority, and isolation
analysis, but it represents only one of its have been met.41 Rather, rigorous meth-
components. Logical reasoning is another odological planning and implementation
vital data analysis component. Critical as in research must accompany use of SEM
it is, the use of statistical techniques is to increase confidence that causality is
becoming separated from the sound ma- being observed (or established).
nipulation of ideas,40 in part due to the
rapid development of powerful computer Model Parsimony Should Be a Top
Copyright (c) PNG Publications. All rights reserved.

programs. Progress in science, however, Priority


is significantly hampered when research- When building a model, parsimony
ers use, as their guides, implausible theo- should be a top priority for researchers. A
retical reasoning, frameworks, or mod- parsimonious model consists of the few-
els. Although SEM is a more elegant ana- est number of variables, explaining the
lytic technique than univariate/bivari- greatest possible amount of variance in
ate methods, as a statistical tool it will the outcome(s) of interest. In short, the
never compensate for bad logic and poor simplest model is the best model. In the
ideas/models. As Kenny stated in his second section, however, we noted that
seminal contribution, Correlation and Cau- health behavior researchers should stay
sality,6 “Causal modeling provides no cer- true to the examination of the complexi-
tain path to knowledge. In fact, models ties (ie, multivariate nature) of human
are maximally helpful only when good behavior. These complexities force re-
ideas are tested. Good ideas do not come searchers to measure all possible causes
out of computer packages, but from people’s and consequences of the health behavior
heads” (p. 8). In short, SEM should be under study. Thus, the resulting model
used simply as a tool for testing care- may be one of enormous proportion (ie,
fully thought-out ideas that are empiri- nonparsimonious), with a massive num-
cally grounded and/or theoretically gen- ber of variables accounting for these vari-
erated. ous influences and outcomes. A number
of fit statistics, however, penalize the
SEM Analyses Are Correlational researcher for model complexity, or this
SEM is misleadingly called causal mod- lack of parsimony. The challenge for re-
eling, but as part of the GLM family, it is searchers, therefore, is to capture the
still a correlational method. Inferring cau- complexity of human behavior using the
sation requires more than simply employ- fewest number of variables possible. One
ing SEM; instead, a number of conditions solution to the model parsimony issue is
must first be met. The most basic condi- to include large numbers of observed vari-
tion is that an association must exist ables, measured by the smallest possible
between the variables postulated to have number of latent factors with the fewest
the cause-and-effect relationship. Sec- possible number of structural paths. See
ond, temporal priority, temporal ordering, Sivo and Willson42 for a more detailed
or directionality31 of the causal relation- discussion of model parsimony.
ship must be established (ie, the cause
must precede the effect). Rather than Sample-size Issues
being an analytic issue, temporal priority Having an adequate study sample size
is primarily a methodological matter. That can be a major concern in SEM utiliza-
is, to ensure the cause precedes the tion. In general, small samples are more
effect, data must be collected over time. likely to result in unreliable and untrust-

82
Buhi et al

worthy parameter estimates and fit sta- include theoretical, methodological, and
tistics, yielding models which are applied pieces; book reviews; software
nonreplicable. West et al28 noted that de- package reviews; and a teacher’s corner
creasing sample size leads to an increase with instructional modules.
in the probability that analyses will fail to Supplementing the various text and
converge or result in inappropriate solu- journal resources is an electronic mail
tions. What, then, constitutes an adequate network for SEMers called SEMNET. Be-
sample size? Several authors43,44 have gun in 1993, and owned by Dr Carl E.
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

suggested that at least 200 cases are Ferguson Jr, professor of marketing at
necessary for adequate model specifica- the University of Alabama, SEMNET serves
tion. Stevens26 noted that 15 cases per as an open forum for ideas and questions
predictor in standard ordinary least regarding analysis of covariance struc-
squares regression is a good rule of thumb. tures, path analysis, and confirmatory
Because SEM and regression are similar factor analysis. SEMNET archives can be
in many respects, 15 cases per measured searched at http://bama.ua.edu/ar-
variable in SEM is not unreasonable.24 In chives/semnet.html, and additional in-
reality, however, there is no ideal sample formation can be retrieved from the
size for all situations. Adequate sample SEMNET information site (http://
size may depend on the complexity of the www2.gsu.edu/~mkteer/semnet.html).
model being tested and the statistical Finally, numerous software packages
Copyright (c) PNG Publications. All rights reserved.

estimator used. For instance, more com- exist for SEM analyses, including Amos
plex models and those with multiple indi- (Analysis of Moment Structures), EQS
rect effects may need a greater number of (Equations), Mplus, LISREL (Linear Struc-
cases. For confirmatory factor analyses, tural Relationships), CALIS (Covariance
Flora and Curran45 provided evidence that Analysis and Linear Structural Equations,
one can use weighted least-squares esti- in SAS), and Mx. These programs, histori-
mators with sample sizes as small as 100 cally known for their complex code com-
cases. The interested reader should refer mands and large computer memory/space
to other sources34,43 for a more detailed requirements, are now much more user-
discussion of SEM and sample-size is- friendly and accessible. Most programs
sues. See also Muthén and Muthén46 for a still allow users to write code containing
demonstration of how researchers can matrix algebra commands. However, many
use a Monte Carlo (or simulation) study to packages such as Amos and EQS have
decide on sample size and determine graphical interface options, allowing ana-
power for SEMs, using Mplus. lysts to draw their measurement/struc-
tural models on the computer screen and
5. What Resources are Available to tap into a data set to generate output.
SEM Users? Most range in cost from $500-600 for the
Aids and resources abound for both the basic-level package, which may exclude
novice and experienced SEM user, in- add-on features, user’s guides, and tech-
cluding books, journals, e-mail discus- nical support. Mx, however, is freely avail-
sion lists, and statistical software pack- able for download through the WWW, and
ages. For beginners, there are a number the package can do the usual SEM analy-
of textbooks7,32 and book chapters47,48 pro- ses and more. For a fee (usually between
viding useful overviews. Some texts even $900 and 1200), individuals can enroll in
present user-friendly introductions to training courses on conducting SEM with
specific SEM software packages, such as various software packages, which are held
Amos,49 EQS,50 and LISERL.51 For seasoned in major cities and on college campuses.
SEMers, Bollen’s31 encyclopedic reference These trainings are regularly offered
may serve as a key resource. through professional associations (see
In 1994, Structural Equation Modeling: A the APA Advanced Training Institutes
Multidisciplinary Journal began quarterly online at http://www.apa.org/science/
publication. This journal has, since, ati.html), universities (see the Univer-
served as the flagship peer-reviewed pe- sity of Kansas Continuing Education
riodical for researchers utilizing SEM website at http://
analyses in various disciplines, includ- www.continuinged.ku.edu/programs/
ing health/medicine, psychology, educa- rda/index.php), and software manufac-
tion, economics, sociology, business, and turers (see the SPSS/Amos website at
political science. Contents of the journal http://www.spss.com/training/ or the

™ 2007;31(1):74-85
Am J Health Behav.™ 83
Structural Equation Modeling

Mplus training site at http:// RJ, Crosby RA, Kegler MC, eds. Emerging
www.statmodel.com/courses.html). Table Theories in Health Promotion Practice and
2 summarizes the more commonly used Research: Strategies for Improving Public
SEM applications, including resource and Health. San Francisco: Jossey-Bass; 2002:1-
15.
pricing information as well as specific 9.Short LM, Hennessy M. Using structural equa-
strengths and weaknesses regarding each tions to estimate effects of behavioral inter-
package. ventions. Structural Equation Modeling.
1994;1:68-81.
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

CONCLUSION 10.Thompson B. Five methodology errors in


We have organized this primer to allow educational research: the pantheon of statis-
readers, both new and experienced users tical significance and other faux pas. Invited
of SEM, to review any one of the 5 free- address at the annual meeting of the Ameri-
standing sections most pertinent to their can Educational Research Association, San
Diego, CA. 1998, April. (ERIC Document Re-
needs or knowledge level. It is our hope production Service No. ED 419 023)
that this paper has familiarized health 11.Bagozzi RP, Fornell C, Larcker DF. Canonical
behavior researchers with the purpose of correlation analysis as a special case of a
SEM, why SEM is valuable as an analytic structural relations model. Multivariate Behav
technique, how SEM is conducted, and Res. 1981;16:437-454.
cautionary notes related to using SEM. 12.Fan X. Canonical correlation analysis and
We also hope that we have equipped the structural equation modeling: what do they
Copyright (c) PNG Publications. All rights reserved.

interested reader with the necessary re- have in common? Structural Equation Modeling.
sources and information regarding SEM 1997;4:65-79.
13.Bauer DJ. Estimating multilevel linear mod-
analyses and available software pack- els as structural equation models. Journal of
ages. Finally, we hope that this primer Educational and Behavioral Statistics.
will thrust SEM into the daily vocabu- 2003;28:135-167.
lary and, most importantly, into the 14.Hox J. Multilevel Analysis: Techniques and
routine practice of health behavior and Applications. Mahwah, NJ: Lawrence Erlbaum;
health promotion researchers. The gen- 2002:251-274.
eration of high-quality research de- 15.Mueller RO. Structural equation modeling:
pends upon it. „ back to basics. Structural Equation Modeling.
1997;4:353-369.
REFERENCES 16.Thompson B. Why multivariate methods are
1.Tremblay PF, Gardner RC. On the growth of usually vital in research: some basic con-
structural equation modeling in psychological cepts. Paper presented at the biennial meet-
journals. Structural Equation Modeling. ing of the Southwestern Society for Research
1996;3:93-104. in Human Development, Austin, TX. 1994,
2.Hershberger SL. The growth of structural February. (ERIC Document Reproduction Ser-
equation modeling: 1994-2001. Structural Equa- vice No. ED 367 687)
tion Modeling. 2003;10:35-46. 17.Kline RB. Beyond Significance Testing: Re-
3.Nachtigall C, Kroehne U, Funke F, et al. (Why) forming Data Analysis Methods in Behavioral
should we use SEM? Pros and cons of struc- Research. Washington, DC: American Psy-
tural equation modeling. Methods of Psycho- chological Association; 2004.
logical Research. 2003;8:1-22. 18.Huberty CJ, Morris JD. Multivariate analysis
4. American Academy of Health Behavior Work versus multiple univariate analyses. Psychol
Group on Doctoral Research Training. A vi- Bull. 1989;105:302-308.
sion for doctoral research training in health 19.Fish LJ. Why multivariate methods are usu-
behavior: a position paper from the American ally vital. Measurement and Evaluation in Coun-
Academy of Health Behavior. Am J Health seling and Development. 1988;21:130-137.
Behav. 2005;29:542-556. 20.Arbuckle JL. Full information estimation in
5.Bentler PM. Causal modeling via structural the presence of incomplete data. In:
equation systems. In: Nesselroade JR, Cattell Marcoulides GA, Schumacker RE, eds. Ad-
RB, eds. Handbook of Multivariate Experi- vanced Structural Equation Modeling: Issues
mental Psychology (2nd ed.). New York: Ple- and Techniques. Mahwah, NJ: Lawrence
num Press; 1988:317-335. Erlbaum; 1996:243-277.
6.Kenny DA. Correlation and Causality. New 21.Wothke W. Longitudinal and multigroup mod-
York: John Wiley & Sons; 1979. eling with missing data. In: Little TD, Schnabel
7.Raykov T, Marcoulides GA. A First Course in KU, eds. Modeling Longitudinal and Multi-
Structural Equation Modeling. Mahwah, NJ: level Data: Practical Issues, Applied Ap-
Lawrence Erlbaum; 2000. proaches, and Specific Examples. Mahwah,
8.Crosby RA, Kegler MC, DiClemente RJ. Under- NJ: Lawrence Erlbaum Associates; 2000:219-
standing and applying theory in health pro- 240,269-281.
motion practice and research. In: DiClemente 22.Little RJA, Rubin DB. Statistical Analysis

84
Buhi et al

with Missing Data (2nd Ed.). Hoboken, NJ: 38.Park K, Wilson MG, Lee MS. Effects of social
John Wiley & Sons, Inc; 2002. support at work on depression and organiza-
23.Allison PD. Missing Data. Thousand Oaks, tional productivity. Am J Health Behav.
CA: Sage; 2002. 2004;28:444-455.
24.Research Consulting, ITS, UT. Structural 39.Saunders RP, Motl RW, Dowda M, et al.
Equation Modeling using AMOS: An Introduc- Comparison of social variables for under-
tion. Available at: http://ww.utexas.edu/its/ standing physical activity in adolescent girls.
rc/tutorials/stat/amos/. Accessed October 25, Am J Health Behav. 2004;28:426-436.
2005. 40.Aneshensel CS. Theory-Based Data Analysis
Delivered by Publishing Technology to: New Mexico State University IP: 190.129.30.38 on: Mon, 25 Jan 2016 13:26:37

25.Anderson JC, Gerbing DW. Structural equa- for the Social Sciences. Thousand Oaks, CA:
tion modeling in practice: a review and recom- Pine Forge Press; 2002.
mended two-step approach. Psychol Methods. 41.Bullock HE, Harlow LL, Mulaik SA. Causation
1988;103:411-423. issues in structural equation modeling re-
26.Stevens JP. Applied Multivariate Statistics search. Structural Equation Modeling.
for the Social Sciences. Mahwah, NJ: 1994;1:253-267.
Lawrence Erlbaum; 2002:262. 42.Sivo SA, Willson VL. Is parsimony always
27.Henson RK. Multivariate normality: what is it desirable? Identifying the correct model for a
and how is it assessed? Advances in Social longitudinal panel data set. Journal of Experi-
Science Methodology. 1999;5:193-211. mental Education. 1998;66:249-255.
28.West SG, Finch JF, Curran PJ. Structural 43.Fan X, Thompson B, Wang L. Effects of
equation models with nonnormal variables: sample size, estimation methods, and model
problems and remedies. In: Hoyle RH, ed. specification on structural equation modeling
Copyright (c) PNG Publications. All rights reserved.

Structural Equation Modeling: Concepts, Is- fit indexes. Structural Equation Modeling.
sues, and Applications. London: Sage; 1999;6:56-83.
1995:56-75. 44.Chou C, Bentler PM. Estimates and tests in
29.Mardia KV, Kanazawa M. The null distribu- structural equation modeling. In: Hoyle RH,
tion of multivariate kurtosis. Communications ed. Structural Equation Modeling: Concepts,
in Statistics: Simulation and Computation. Issues, and Applications. London: Sage;
1983;12:569-576. 1995:37-55.
30.Yuan K, Lambert PL, Fouladi RT. Mardia’s 45.Flora DB, Curran PJ. An empirical evaluation
multivariate kurtosis with missing data. Mul- of alternative methods of estimation for confir-
tivariate Behav Res. 2004;39:413-437. matory factor analysis with ordinal data.
31.Bollen KA. Structural Equations with Latent Psychol Methods. 2004;9:466-491.
Variables. New York: John Wiley & Sons, Inc; 46.Muthén LK, Muthén BO. How to use a Monte
1989. Carlo study to decide on sample size and
32.Hoyle RH, Panter AT. Writing about structural determine power. Structural Equation Model-
equation models. In: Hoyle RH, ed. Structural ing. 2002;9:599-620.
Equation Modeling: Concepts, Issues, and 47.Klem L. Structural equation modeling. In:
Applications. London: Sage; 1995:158-176. Grimm LG, Yarnold PR, eds. Reading and
33.Hu L, Bentler PM. Fit indices in covariance Understanding MORE Multivariate Statistics.
structure modeling: sensitivity to Washington, DC: American Psychological As-
underparamerterized model misspecification. sociation; 2000:227-260.
Psychol Methods. 1998;3:424-253. 48.Thompson B. Ten commandments of struc-
34.Hu L, Bentler PM. Cutoff criteria for fit tural equation modeling. In: Grimm LG, Yarnold
indexes in covariance structure analysis: con- PR, eds. Reading and Understanding MORE
ventional criteria versus new alternatives. Multivariate Statistics. Washington, DC:
Structural Equation Modeling. 1999;6:1-55. American Psychological Association;
35.Browne MW, Cudeck R. Alternative ways of 2000:261-283.
assessing model fit. In: Bollen KA, Long JS, 49.Byrne BM. Structural Equation Modeling with
eds. Testing Structural Models. Newbury Park, AMOS: Basic Concepts, Applications, and
CA: Sage Publications, Inc; 1993:136-162. Programming. Mahwah, NJ: Lawrence
36.Marsh HW, Hau K, Wen Z. In search of golden Erlbaum Associates; 2001.
rules: comment on hypothesis-testing ap- 50.Byrne SM. Structural Equation Modeling with
proaches to setting cutoff values for fit in- EQS and EQS/Windows: Basic Concepts,
dexes and dangers in overgeneralizing Hu Applications, and Programming. Thousand
and Bentler’s (1999) findings. Structural Equa- Oaks, CA: Sage; 1994.
tion Modeling. 2004;11:320-341. 51.Byrne BM. Structural Equation Modeling with
37.Blue CL, Black DR, Conrad K, et al. Beliefs of LISREL, PRELIS, and SIMPLIS: Basic Con-
blue-collar workers: stage of readiness for cepts, Applications, and Programming.
exercise. Am J Health Behav. 2003;27:408- Mahwah, NJ: Lawrence Erlbaum Associates;
420. 1998.

™ 2007;31(1):74-85
Am J Health Behav.™ 85

Вам также может понравиться