SEM - Basics A Supplement To Multivariate Data Analysis

MultivariateDataAnalysis,PearsonPrenticeHallPublishing Page1
SEMBasics:
ASupplementtoMultivariateDataAnalysis
MultivariateDataAnalysis
PearsonPrenticeHallPublishing


TableofContents
LearningObjectives.....................................................................................................................................................3
Preview.........................................................................................................................................................................3
FundamentalsofStructuralEquationModeling......................................................................................................4
EstimatingRelationshipsUsingPathAnalysis..........................................................................................................4
IdentifyingPaths.................................................................................................................................................5
EstimatingtheRelationship..............................................................................................................................6
UnderstandingDirectandIndirectEffects...............................................................................................................7
IdentificationofCausalversusNonCausaleffects...........................................................................................7
DecomposingEffectsintoCausalversusNoncausal........................................................................................8
CalculatingIndirectEffects................................................................................................................................9
ImpactofModelRespecification......................................................................................................................11
OtherAbsoluteFitIndices.......................................................................................................................................11
SpecificationIssuesinSEMPrograms......................................................................................................................12
TheMultivariateRelationshipsinSEM.....................................................................................................................12
TheMainStructuralEquation...........................................................................................................................12
UsingConstructstoExplainMeasuredVariables:TheMeasurementModel................................................13
CompleteStructuralandMeasurementModelEquations.............................................................................14
SpecifyingAModelinLISRELNotation...................................................................................................................18
SpecificationofaCFAModelwithLISREL.......................................................................................................18
ChangingTheCFASetupinLISRELtoaStructuralModelTest......................................................................19
HBAT:TheCFAModel...............................................................................................................................................21
HBAT:TheStructuralModel...................................................................................................................................23
HowtoFixFactorLoadingstoaSpecificValueinLISREL.......................................................................................25
MeasuredVariableandConstructInterceptTerms................................................................................................27
PathModelSpecificationwithAMOS.....................................................................................................................27
ResultsUsingDifferentSEMPrograms...................................................................................................................28
AdditionalSEMAnalyses...........................................................................................................................................28
TestingforDifferencesinConstructMeans............................................................................................................29
ItemParcelinginCFAandSEM................................................................................................................................29
WhenIsParcelingAppropriate?......................................................................................................................30
HowShouldItemsBeCombinedintoParcels?................................................................................................31
MeasurementBias..................................................................................................................................................31
ContinuousVariableInteractions...........................................................................................................................33


SEMBasics:
ASupplementto
MultivariateDataAnalysis
LEARNINGOBJECTIVES
Inthecourseofcompletingthissupplement,youwillbeintroducedtothefollowing:
Thebasicsofestimatingpathcoefficientsbasedonthespecifiedpathmodel.
Determinationofthedirectandindirecteffectsimpliedinapathmodel,plusdetermination
whethertheycanbecharacterizedascausalornoncausal.
Someadditionalabsolutefitindicesusedincertainsituations.
Specification of the path model as a series of equations for both the structural model and
themeasurementmodel.
Use of LISREL notation to represent these equations and the relationships in the path
model.
Testingformeandifferencesbetweenlatentconstructsindifferentgroups.
Itemparcelingtoreducethenumberofitemsperconstruct.
Assessmentofmeasurementbiasbyintroductionofanadditionallatentconstruct.
Estimationofmoderatingeffectsforcontinuousmultiitemconstructs.

PREVIEW
This supplement to the text Multivariate Data Analysis provides additional coverage of some
basicconceptsthatarethefoundationsforstructuralequationmodeling(SEM).Whilethereis
considerable coverage of the technique in the text, the authors felt that readers may benefit
from further review of certain topics not covered in the text, but issues addressed by many
researchers.Moreover,thereisamorecomprehensivediscussionofthenotationusedinSEM,
particularly those associated with LISREL. There will be some overlap with material in the
chapterssoastofullyintegratetheconcepts.
The supplement is not intended to be a comprehensive primer on all of the SEM

topics not covered in the test, but only those selected issues that may be encountered in the

course of behavioral research. We encourage readers to complement this supplement with
othertreatmentsandtextsontheseconceptsasneeded.
The supplement focuses on three broad areas related to SEM. The first area covers
someofthefundamentalconceptsrelatedthebasicsofpathmodelsestimationofthepath
estimates and determining and interpreting direct and indirect effects. The second area is
specificationoftheSEMmodelinmoreformalterms.Theprimarilyinvolvesdiscussionofwhat
is termed LISREL notation. This involves the notation used in the LISREL software program
whichhasbecomeacommonmethodofdescribingtherelationshipsinboththestructuraland
measurement models. Several examples, including the HBAT CFA and structural models, are
used to illustrate how those models can be expressed in this notation. Included in the
discussionarealsosometechniquestoaccomplishspecializedtasksinLISREL,aswellasabrief
introduction to AMOS, another popular SEM software package. Finally, some more advanced
topics are discussed to provide the user an introduction into some of the more complex, but
oftenused,techniquesavailableinSEManalyses.
FUNDAMENTALSOFSTRUCTURALEQUATIONMODELING
The use of SEM is predicated on a strong theoretical model by which latent constructs are
defined(measurementmodel)andtheseconstructsarerelatedtoeachotherthroughaseries
ofdependencerelationships(structuralmodel).Theemphasisonstrongtheoreticalsupportfor
anyproposedmodelunderliestheconfirmatorynatureofmostSEMapplications.
But many times overlooked is exactly how the proposed structural model is translated
into structural relationships and how their estimation is interrelated. Path analysis is the
processwhereinthestructuralrelationshipsareexpressedasdirectandindirecteffectsinorder
to facilitate estimation. The importance of understanding this process is not so that the
research can understand the estimation process, but instead to understand how model
specification (and respecification) impacts the entire set of structural relationships. We will
first illustrate the process of using path analysis for estimating relationships in SEM analyses.
Thenwewilldiscusstherolethatmodelspecificationhasindefiningdirectandindirecteffects
andclassificationofeffectsascausalversusspurious.Wewillseehowthisdesignationimpacts
theestimationofstructuralmodel.
ESTIMATINGRELATIONSHIPSUSINGPATHANALYSIS
What was the purpose of developing the path diagram? Path diagrams are the basis for path

analysis, the procedure for empirical estimation of the strength of each relationship (path)
depicted in the path diagram. Path analysis calculates the strength of the relationships using
only a correlation or covariance matrix as input. We will describe the basic process in the
following section, using a simple example to illustrate how the estimates are actually
computed.
IdentifyingPaths
The first step is to identify all relationships that connect any two constructs. Path analysis
enablesustodecomposethesimple(bivariate)correlationbetweenanytwovariablesintothe
sum of the compound paths connecting these points. The number and types of compound
paths between any two variables are strictly a function of the model proposed by the
researcher.
Acompoundpathisapathalongthearrowsofapathdiagramthatfollowthreerules:
1. After going forward on an arrow, the path cannot go backward again; but the path can go
backwardasmanytimesasnecessarybeforegoingforward.
2. Thepathcannotgothroughthesamevariablemorethanonce.
3. Thepathcanincludeonlyonecurvedarrow(correlatedvariablepair).
When applying these rules, each path or arrow represents a path. If only one arrow links
two constructs (path analysis can also be conducted with variables), then the relationship
betweenthosetwoisequaltotheparameterestimatebetweenthosetwoconstructs.Fornow,
this relationship can be called a direct relationship. If there are multiple arrows linking one
constructtoanotherasinXYZ,thentheeffectofXonZseemquitecomplicatedbutan
examplemakesiteasytofollow:
Thepathmodelbelowportraysasimplemodelwithtwoexogenousconstructs(X
1
andX
2
)
causallyrelatedtotheendogenousconstruct(Y
1
).ThecorrelationalpathAisX
1
correlatedwith
X
2
,pathBistheeffectofX
1
predictingY
1
,andpathCshowstheeffectofX
2
predictingY
1
.
ThevalueforY
1
canbestatedsimplywitharegressionlikeequation:
1
= b
1
X
1
+ b
2
X
1
C
X
1
Y
1
X
2
A

We can now identify the direct and indirect paths in our model. For ease in referring to the
paths,thecausalpathsarelabeledA,B,andC.
DirectPaths IndirectPaths
A=X
1
toX
2

B=X
1
toY
1
AC=X
1
toY
1
C=X
2
toY
1
AB=X
2
toY
1
EstimatingtheRelationship
Withthedirectandindirectpathsnowdefined,wecanrepresentthecorrelationbetweeneach
constructasthesumofthedirectandindirectpaths.Thethreeuniquecorrelationsamongthe
constructscanbeshowntobecomposedofdirectandindirectpathsasfollows:
Corr
X1 X2
= A
Corr
X1 1
= B +AC
Corr
X2 1
= C +AB
First,thecorrelationofX
1
andX
2
issimplyequaltoA.ThecorrelationofX
1
andY
1
(Corr
X1,Y1
)can
berepresentedastwopaths:BandAC.ThesymbolBrepresentsthedirectpathfromX
1
toY
1
,
and the other path (a compound path) follows the curvedarrow from X
1
to X
2
and then to Y
1
.
Likewise,thecorrelationofX
2
andY
1
canbeshowntobecomposedoftwocausalpaths:Cand
AB.
Once all the correlations are defined in terms of paths, the values of the observed
correlations can be substituted and the equations solved for each separate path. The paths
then represent either the causal relationships between constructs (similar to a regression
coefficient)orcorrelationalestimates.
Assuming that the correlations among the three constructs are as follows: Corr
X1 X2
=
.50, Corr
X1 Y1
= .60 and Corr
X2 Y1
= .70, we can solve the equations for each correlation (see
below) and estimate the causal relationships represented by the coefficients b
1
and b
2
. We
know that A equals .50, so we can substitute this value into the other equations. By solving
these two equations, we get values of B(b
1
) = .33 and C(b
2
) = .53. This approach enables path
analysis to solve for any causal relationship based only on the correlations among the
constructsandthespecifiedcausalmodel.

SolvingfortheStructuralCoefficients
.50=A
.60=B+AC
.70=C+AB
SubstitutingA=.50
.60=B+.50C
.70=C+.50B
SolvingforBandC
B=.33
C=.53
As you can see from this simple example, if we change the path model in some way, the
causal relationships will change as well. Such a change provides the basis for modifying the
modeltoachievebetterfit,iftheoreticallyjustified.
With these simple rules, the larger model can now be modeled simultaneously, using
correlationsorcovariancesastheinputdata.Weshouldnotethatwhenusedinalargermodel,
we can solve for any number of interrelated equations. Thus, dependent variables in one
relationship can easily be independent variables in another relationship. No matter how large
thepathdiagramgetsorhowmanyrelationshipsareincluded,pathanalysisprovidesawayto
analyzethesetofrelationships.
UNDERSTANDINGDIRECTANDINDIRECTEFFECTS
Whilepathanalysisplaysakeyroleinestimatingtheeffectsrepresentedinastructuralmodel,
it also provides additional insight into not only the direct effects of one construct versus
another,butallofthemyriadsetofindirecteffectsbetweenanytwoconstructs.Whiledirect
effects can always be considered causal if a dependence relationship is specified, indirect
effects require further examination to determine if they are causal (directly attributable to a
dependence relationship) or noncausal (meaning that they represent relationship between
constructs,butitcannotbeattributedtoaspecificcausalprocess).
IdentificationofCausalversusNonCausaleffects
The prior section discussed the process of identifying all of the direct and indirect effects
betweenanytwoconstructsbyaseriesofrulesforcompoundpaths.Herewewilldiscusshow
tocategorizethemintocausalversusnoncausalandthenillustratetheiruseinunderstanding
theimplicationsofmodelspecification.
An important question is: Why is the distinction important? The parameter estimates
are made without any distinction as described above. But the estimated parameters in the
Multiva

structura
theindire
another
distinguis
inferany
A
types: a
director
of the re
canarise
andCisc
Decompo
We will
following

We can
correlatio
theeffec
Thetotal
C
identifyin
direct ef
estimate
ariateDataA
al model ref
ecteffects,w
construct d
shbetween
ydependenc
Assume we a
direct caus
indirectcom
elationships
efromthree
correlatedw
osingEffects
use a simp
gpathdiagra
see that th
onal relation
cts(directan
lsetofeffec
1C2.Let
ng compoun
ffect of C
1
o
forP
2,1
repr
Analysis,Pea
flect only th
whichcanbe
does not m
thesemyria
cerelationsh
are identifyi
al effect ( A
mpoundpat
is never rev
conditions:
withB)andre
sintoCausa
le example
am.
here are on
nships amon
ndindirect)w
ctsforeachr
tsstartwith
nd paths ea
on C
2
. The
resentsthet
arsonPrent
e direct effe
esubstantia
ean that th
adtypesofi
hip(causal)a
ing the pos
A B) or an
th,onlydepe
versed. Non
commonef
eciprocaleff
alversusNo
of four con
nly direct d
ng exogenou
wouldbeca
relationship
hthesimple
rlier, we ca
re are no p
totaleffects
ticeHallPub
ect of one c
al?Justbeca
here is no
indirecteffe
attributable
sible effects
n indirect ca
endencerela
causal (som
ffects(C
fects(Aa
ncausal
nstructs all
ependence
us construct
usalaswell.
isshownin
relationship
n see that
possible ind
ofC
1
onC
2
.
blishing
construct on
auseaconst
impact. Th
ectsandbea
tothemeve
s of A B
ausal effect (
ationshipsa
metimes refe
AandCB
andBA).
related to e
relationship
ts). So we m
.Butaswe
thetablebe
pbetweenC
there is on
irect or non
n another. W
tructisnotd
hus, we nee
abletounde
enthoughth
. Causal ef
( A C
arepresenta
erred to as
B),correlate
each other

ps in this e
might presu
willsee,tha
elow.
C
1
andC
2
.U
ly one poss
ncausal effe
Pa
What about
directlyrelat
ed a metho
erstandifwe
heyareindire
ffects are of
B). In eithe
andthedire
spurious eff
deffects(C
as shown in
example (i.e
ppose that
atisnotthe
singtherule
ible effect
ects, so the
age8
all of
edto
od to
ecan
ect.
f two
er the
ection
fects)
CA
n the
e., no
all of
case.
esfor
the
path

C1 C3. The next relationship is C
1
with C
3
. Here we can see two effects: the direct
effect(P
3,1
)andtheindirecteffect(P
3,2
xP
2,1
).Sincethedirectionofthepathsneverreversesin
the indirect effect, it can be categorized as causal. So the direct and indirect effects are both
causaleffects.
C2C3.Thisrelationshipintroducesthefirstnoncausaleffectswehaveseen.Thereis
the direct effect of B
3,2
, but there is also the noncausal effect (due to common cause) seen in
B
3,1
x B
2,1
. Here we see the result of two causal effects creating a noncausal effect since they
bothoriginatefromacommonconstruct(C
1
C
2
andC
1
C
3
).
C1 C4. In this relationship we will see the potential for numerous indirect causal
effectsinadditiontodirecteffects.Inadditiontothedirecteffect(B4,1),weseethreeother
indirecteffectsthatarealsocausal:B
4,2
xB
2,1
;B
4,3
xB
3,1
;andB
4,3
xB
3,2
xB
2,1
.
C3C4.Thisfinalrelationshipwewillexaminehasonlyonecausaleffects(B4,2),but
four different noncausal effects, all a result of C
1
or C
2
acting as common causes. The two
noncausal effects associated with C
1
are B
4,1
x B
3,1
and B
4,1
x B
2,1
x B
3,2
. The two other
noncausaleffectsareassociatedwithC
2
(B
4,2
xB
3,2
andB
4,2
xB
2,1
xB
3,1
).
TheremainingrelationshipisC
2
C
4
.Seeifyoucanidentifythecausalandnoncausal
effects. Hint: There are all three types of effects direct and indirect causal effects and
noncausaleffectsaswell.
Relationship
Effects
Direct(Causal) Indirect(Causal) Indirect(Noncausal)
C
1
C
2
P
2,1
None None
C
1
C
3
P
3,1
P
3,2
xP
2,1
None
C
2
C
3
P
3,2
None P
3,1
xP
2,1
C
1
C
4
P
4,1
P
4,2
xP
2,1
P
4,3
xP
3,1
P
4,3
xP
3,2
xP
2,1
None
C
3
C
4
P
4,3
None P
4,1
xP
3,1
P
4,1
xP
2,1
xP
3,2
P
4,2
xP
3,2
P
4,2
xP
2,1
xP
3,1
CalculatingIndirectEffects
In the previous section we discussed the identification and categorization of both direct and
indirecteffectsforanypairofconstructs.Thenextstepistocalculatetheamountoftheeffect
basedonthepathestimatesofthemodel.Assumethispathmodelwithestimatesasfollows:
Multiva

For the r
effectsa
Howare
T
software
Buttose
compoun
Fo
calculate
forC1
P
P
P
T
Sointhis
= .55. It
directeff
ariateDataA
relationship
reshowndir
theycalcula
he size of a
etypicallypr
eethesizeo
ndpath.
or instance,
etheamoun
C4canbec
C2,C1
P
C4,C2
C3,C1
P
C4,C3
C2,C1
P
C3,C2
otalIndirect
sexample,th
t is interesti
fects.
Analysis,Pea
between C
rectlybythe
ated?
an indirect e
roducesata
feacheffec
, one of the
tofthiseffe
calculatedas
=.50.40
=.40.30
P
C4,C3
=.50
tEffects=.2
hetotaleffe
ing to note
arsonPrent
1 C4, the
epathestim
effect is a fu
bleshowing
tyoucanco
e indirect ef
ectas.50
sfollows:
=.20
=.12
0.20.30
0+.12+.03
ectofC1C
that in this
ticeHallPub
ere are both
ateP
C4,C1
=.
unction of t
gthetotalof
omputethem
ffects for C1
.40=.20.T
=.03
3=.35
C4equalsth
s example th
blishing
h direct and
20.Butwha
the direct ef
ftheindirec
mbymultipl
1 C4 is P
Thus,thecom
hedirectand
he indirect e
indirect eff
ataboutthe
ffects that m
teffectsimp
lyingthedir
P
C2,C1
P
C4,C
mpleteseto
dindirecteff
effects are g
Pag
fects. The d
eindirecteff
make it up.
pliedbyam
ecteffectsi
C2
. We can
ofindirectef
fects,or.20
greater than
ge10
direct
fects?
SEM
model.
nthe
then
ffects
+.35
n the

ImpactofModelRespecification
Theimpactofmodelrespecificationonboththeparameterestimatesandthecausal/noncausal
effects can be seen in our example as well. Look back at the C
3
C
4
relationship. What
happensifweeliminatetheC1C4relationship?DoesitimpacttheC3C4relationshipin
any way? If we look back at the indirect effects, we can see that two of the four noncausal
effects would be eliminated (B
4,1
x B
3,1
and B
4,1
x B
2,1
x B
3,2
). How would this impact the
model?IftheseeffectsweresubstantialbuteliminatedwhentheC
1
C
4
pathwaseliminated,
then most likely the C
3
C
4
relationship would be underestimated, resulting in a larger
residual for this covariance and overall poorer model fit. Plus, a number of other effects that
used this path would be eliminated as well. This illustrates how the removal or addition of a
path in the structural model can impact not only that direct relationship (e.g., C
1
C
4
), but
manyotherrelationshipsaswell.
OTHERABSOLUTEFITINDICES
Most SEM programs today provide the user with many different fit indices. In the text we
focusedmorecloselyonthosethataremostwidelyused.Inthissection,webrieflytouchona
fewotherabsolutefitindicesthataresometimesreported:
The expected crossvalidation index (ECVI) is an approximation of the goodnessoffit the
estimated model would achieve in another sample of the same size. Based on the sample
covariancematrix,ittakesintoaccounttheactualsamplesizeandthedifferencethatcould
be expected in anothersample. The ECVI also takes into account the number ofestimated
parametersforagivenmodel.Itismostusefulincomparingtheperformanceofonemodel
toanother.
The actual crossvalidation index (CVI) can be formed by using the computed covariance
matrixderivedfromamodelinonesampletopredicttheobservedcovariancematrixtaken
from a validation sample. Given a sufficiently large sample (i.e., N > 500 for most
applications), the researcher can create a validation sample by splitting the original
observationsrandomlyintotwogroups.
Gamma Hat also attempts to correct for both the sample size and model complexity by
includingeachinitscomputation.TypicalGammaHatvaluesrangebetween.9and1.0.Its
primaryadvantageisthatithasaknowndistribution[10].


SPECIFICATIONISSUESINSEMPROGRAMS
InthissectionweprovideanoverviewofspecificationissuesinSEMfortwosoftwarepackages.
We will first discuss the notation used in LISREL, a popular SEM program. This notation has
becomeastandardlanguageofSEMinreferringtobothmeasurementandstructuralmodel
relationships.Wewillthenexaminehowtheformulationofthepathmodelistranslatedinto
programcommandswhileconformingtotheLISRELnotation.Thefirstexamplewillbeasimple
path model to illustrate the basic issues involved. The discussion then shifts to the HBAT
analysisfromthetextbookforboththeCFAandstructuralmodels.Wethenreviewtheseissues
brieflyforAMOSaswell.Inthefinalsectionseveralmorecomplexissuesinmodelspecification
arediscussed.
THEMULTIVARIATERELATIONSHIPSINSEM
As we discussed in the text, SEM models are defined by two submodels the measurement
model and the structural model. Each submodel can be expressed is a set of multivariate
equations.Itisntcalledstructuralequationsmodelingfornothing!Eventhoughitispossible
to learn how to run a SEM model without a full and complete understanding of its equations,
knowingthebasicequationscanbehelpfulinunderstandingthedistinctionbetweenmeasured
variables and constructs and between exogenous and endogenous constructs. Moreover, the
equations introduce the notation used in LISREL, which we will discuss in more detail in the
followingsection.Finally,theequationsalsohelpshowhowSEMissimilartoothertechniques.
TheMainStructuralEquation
In regression, our goal was to build a model that predicted a single dependent variable. Here,
we are trying to predict and explain a set of endogenous constructs. Therefore, we need
equationsthatexplainendogenousconstructs()inadditiontothoseexplainingthemeasured
items (individual x and y variables used as indicators). Not surprisingly, we find that these
equationsaresimilartothemultipleregressionequationthatexplainsthedependentvariable
(y) with multiple independent variables (i.e., x
1
and x
2
). This fundamental equation for the
structuralmodelisasfollows(refertotheabbreviationguideintheAppendixofthisdocument
foranyneededhelpwithpronunciationsordefinitions):
p = p + +
Therepresentstheendogenousconstructsinamodel.Wewillhaveaseparateequation
for each endogenous construct. The appears on both sides of the equation because
endogenousconstructscanbedependentononeanother(i.e.,oneendogenousconstructcan

beapredictorofanotherendogenousconstruct).TheBrepresentstheparametercoefficients
thatlinkendogenousconstructswithotherendogenousconstructs.TheBisamatrixconsisting
of as many rows and columns as there are endogenous constructs. So is there are two
endogenous constructs in the model, B would be a 2 2 matrix, with 2 rows and 2 columns.
The individual elements of B are designated by a . The is the corresponding matrix of
parametercoefficientslinkingtheexogenousconstructs()withtheendogenousconstructs().
It also is a matrix that has as many rows as there are exogenous constructs and as many
columnsasthereareendogenousconstructs.Iftherearethreeexogenousconstructsandtwo
endogenousconstructs,therewouldbea32matrix.Itsindividualelementsaredesignated
by as shown in the figure. Finally, represents the error in the prediction of . It can be
thoughtofastheresidualorconverseoftheR
2
conceptfromregression(i.e.,1R
2
).
Another way to think of the structural equation is as a multiple regression equation
predicting (a construct) instead of y, with the other values and the values as predictors.
The B (
1,1
, ) and (
1,1
, ) provide structural parameter estimates. In the regression
equation, the predictor values were represented by x and the standardized parameter
estimates by the regression coefficients. In both cases, the parameter estimate depicts the
linear relationship between a predictor and an outcome. Thus, clear similarities exist between
SEMandregressionanalysis.
UsingConstructstoExplainMeasuredVariables:TheMeasurementModel
Oncevaluesforareknown,wecanalsopredicttheyvariablesusinganequationoftheform:
= z
,]
p
]
+ e

Here, eachmeasured variable y is predicted by its loadings on the endogenous constructs.
Typicallyameasuredvariableonlyhasaloadingononeconstruct,butthatcanvaryincertain
situations. Predicted values for each x also can be computed in the same manner using the
followingequation:
x
= z
,]
]
+ o
]
Thepredictedvaluesforeachobservedvariable,(whetherapredictedxory)canbeusedto
computecovarianceestimatesthatcouldbecomparedtotheactualobservedcovarianceterms
inassessingmodelfit.Inotherwords,wecanusetheparameterestimatestomodeltheactual
observedvariables.Theestimatedcovariancematrixobtainedbycomputingcovariationamong
predicted values for the measured items is
k
. Recall that the difference between the actual
covariance matrix for observed items (S) and the estimated covariance matrix is an important
partofanalyzingthevalidityofanySEMmodel.
Rarely is it necessary in most applications to actually list predicted values based on the
values of the other variables or constructs. Although it is useful to understand how predicted

values can be obtained because it helps demonstrate the way SEM works, the focus in social
sciencetypicallyisonexplainingrelationships.
CompleteStructuralandMeasurementModelEquations
As noted earlier, LISREL notation has become in some sense the language of SEM. A
researcher, therefore, must have a basic understanding of the notation no matter what
software program is being used. The example below illustrates the complete equations for a
model consisting of three exogenous constructs, two endogenous constructs and four
indicatorseachforthesetsofendogenousandexogenousconstructs.
StructuralModelEquations
Endogenous
Construct
Exogenous
Construct
Endogenous
Construct
Error
1
=
11
1
+
12
2
+
13
3
+
11
1
+
12
3
+
2
=
21
1
+
22
2
+
23
3
+
21
2
+
22
2
+
2
MeasurementModelEquations
ExogenousIndicator ExogenousConstruct Error
X
1
=
x
11
1
+
x
12
2
+
x
13
3
+
1
X
2
=
x
21
1
+
x
22
2
+
x
23
3
+
2
X
3
=
x
31
1
+
x
32
2
+
x
33
3
+
3
X
4
=
x
41
1
+
x
42
2
+
x
43
3
+
4
EndogenousIndicator EndogenousConstructs Error
1
=
y
11
1
+
y
12
2
+
1
2
=
y
21
1
+
y
22
2
+
2
3
=
y
31
1
+
y
32
2
+
3
4
=
y
41
1
+
y
42
2
+
4

StructuralEquationCorrelationsAmongConstructs
AmongExogenousConstructs
(Phi)
AmongEndogenousConstructs
(Psi)

1
3

1
1

1

21

2

21
31
32

CorrelationsAmongIndicators
AmongExogenousIndicators
(Thetadelta
)
AmongEndogenousIndicators
(Thetaepsilon)
X
1
X
2
X
3
X
4

1

2

3

4
X
1

1

X
2

21

2

21

X
3

31

32

3

31

32

X
4

41

42

43

4

41

42

43

The path model not only represents the structural relationships between constructs, but also
provides a means of depicting the direct and indirect effects implied in the structural
relationships.Aworkingknowledgeofthedirectandindirecteffectsofanypathmodelgives
the researcher not only the basis for understanding the foundations of model estimation, but
also insight into the total effects of one construct upon another. Moreover, the indirect
effects can be further subdivided into casual and noncausal/spurious to provide greater
specificity into the types of effects involved. Finally, an understanding of the indirect effects
allows for greater understanding of the implications of model respecification, either through
additionordeletionofadirectrelationship.
The following table provides an overview of the notation used for matrices, constructs
and indicators commonly used in SEM. SEM terminology often is abbreviated with a
combination of Greek characters and roman characters to help distinguish different parts of a
SEMmodel.Itisfollowedbyaguidetoaidinthepronunciationandunderstandingofcommon
SEMabbreviations.


Matrices,Construct/IndicatorsandModelEquationNotationoftheLISRELModel
LISRELModel
Element
Description Notation
Matrix Element
Matrices
StructuralModel
Beta Relationshipsofendogenoustoendogenous
constructs

nn
Gamma Relationshipsofexogenoustoendogenous
constructs

nm
Phi Correlationamongexogenousconstructs
mm
Psi Correlationofstructuralequationsor
endogenousconstructs

n
MeasurementModel
LambdaX Correspondence(loadings)ofexogenous
indicators

x

x
pm
LambdaY Correspondence(loadings)ofendogenous
indicators

y

y
qn
Thetadelta Matrixofpredictionerrorforexogenous
constructindicators

pp
Thetaepsilon Matrixofpredictionerrorforendogenous
constructindicators

qq
Construct/Indicators
Construct
Exogenous Exogenousconstruct
Endogenous Endogenousconstruct
Indicator
Exogenous Exogenousindicator X
Endogenous Endogenousindicator Y
StructuralandMeasurementModelEquations
StructuralModel Relationshipsbetweenexogenousand
endogenousconstructs

=++
MeasurementModel
Exogenous Specificationofindicatorsforexogenous
constructs

X=
x
+
Endogenous Specificationofindicatorsforendogenous
constructs

Y=
y
+


PronunciationGuidetoSEMNotation
Symbol
Pronunciation Meaning
xi(KSIorKZI) An exogenous construct associated with measured X

variables
eta(eightta) An endogenous construct associated with measured Y

variables
lambdax A path representing the factor loading between an

exogenouslatentconstructandameasuredxvariable
lambday A path representing the factor loading between an

endogenouslatentconstructandameasuredyvariable
capitallambda The set of loading estimates represented in a matrix where

rows represent measured variables and columns represent
latentconstructs
phi(fi) Anarcedtwoheadedarrowdenotingthecovariationoftwo
exogenous()constructs
capitalphi A way of referring to the covariance or correlation matrix

betweenasetofexogenous()constructs
gamma Apathrepresentingacausalrelationshipfromanexogenous
construct()toanendogenousconstruct()
capitalgamma A way of referring to the entire set of relationships for a

givenmodel
beta(bayta) A path representing a causal relationship from one

endogenous()constructtoanotherconstruct
capitalbeta A way of referring to the entire set of relationships for a

givenmodel
delta The error term associated with an estimated, measured x

variable
theta(theyta)
delta
Theresidualvariancesandcovariancesassociatedwiththex
estimates;theerrorvarianceitemsarethediagonal
epsilon The error term associated with an estimated, measured y

variable
thetaepsilon Theresidualvariancesandcovariancesassociatedwiththey
estimates;theerrorvarianceitemsarethediagonal
zeta(zayta) Thecovariationbetweenconstructerrors
tau(likenow) Theintercepttermsforameasuredvariable
kappa Theintercepttermsforalatentconstruct
chi(ki)squared Thelikelihoodratio


SPECIFYINGAMODELINLISRELNOTATION
For LISREL and AMOS, the user can either use the dropdown menus to generate the syntax
thatmatchesthemeasurementmodel,drawthemeasurementmodelusingapathdiagram,or
write the appropriate program commands into a syntax window. If either of the first two
alternativesisdonecorrectly,theprogramsgeneratetheprogramsyntaxautomatically.Wewill
discuss this third approach for LISREL since this best illustrates how to use LISREL notation in
specifyingthemodel.
SpecificationofaCFAModelwithLISREL
SpecificationisquitedifferentusingCFAcomparedtoEFA.Thecommandsbelowillustratehow
the simple CFA model shown below is communicated using LISREL program statements. Note
that here we only provide the commands needed to define the model. The complete set of
programcommandsaregiveninourHBATexampleinalatersection. Also,linenumbershave
beenaddedtothecommandsforreference,buttheyarenotneededasinputtoLISREL.
In our example we have four constructs, each with four indicators. See the CFA path
modelbelow:
TheLISRELcommandsforthisCFAareasfollows:
01 MO NX=16 NK=4 PH=SY,FR
02 VA 1.0 LX 1 1 LX 5 2 LX 9 3 LX 13 4
03 FR LX 2 1 LX 3 1 LX 4 1 LX 6 2 LX 7 2 LX 8 2
04 FR LX 10 3 LX 11 3 LX 12 3 LX 14 4 LX 15 4 LX 16 4
We begin with the Model command (MO) indicating the numbers of measured and latent

variables and descriptions of the key matrices of parameters. In a CFA model we only have
exogenousconstructsandthusofxvariables.NXstandsfornumberofxvariables,inthiscase
16.NKstandsforthenumberofexogenous()constructs,inthiscase4.PHindicatesthatthe
matrix of covariances between the 4 constructs () will be symmetric (SY) and free (FR). In
otherwords,theconstructvariances(thediagonalof)andthecovariancebetweeneachpair
ofconstructswillbeestimated.
Line2isavaluestatement(VA)whereweassignavaluetoafixedparameter.Inthiscase,
each of the parameters listed on this line is fixed to 1.0 to set the scale for the constructs.
Oneitemisfixedto1.0oneachconstruct.LX1,1representstheparameterforthefirstloading
onthefirstconstruct(
x1,1
).TheLstandsforlambda,theXisanxvariableand11standforthe
measured variable number and construct number, respectively. Thus, LX2,1 stands for the
parameter representing the factor loading of the second measured variable (x
2
) on the first
latentconstruct(
1
),or
x2,1
.
Lines3and4designatethefreeloadingestimates(FR).The12loadingsreferredtoonthese
lines will be estimated and shown as factor results in the output (in
x
). Thus, this model
estimatesatotalof16loadings,oneforeachindicator(actually12areestimatedandfourfixed
toavalueof1.0)asshowninthepathdiagram.ThiscomparestoEFA,wheretherewouldbea
totalof64loadings(oneforeachindicatoroneachconstruct).
ChangingTheCFASetupinLISRELtoaStructuralModelTest
As discussed in the text, the CFA model forms the foundation from which the structural
model is formulated. In making the conversion from a CFA to a structural model, the
research must make two fundamental decisions: distinguish between exogenous and
endogenousconstructsandspecifythestructuralrelationshipsbetweenconstructs.Notethat
inmostinstances,themeasurementmodelwillbespecifiedandanalyzedintheCFAstage.
Toillustratetheprocess,weutilizetheCFAexamplediscussedinthesectionabove.As
canbeseenfromthepathmodelbelow,twooftheconstructsaredefinedasendogenouswith
relationshipstothetworemainingexogenousconstructs.

ShownbelowarethemodifiedLISRELsyntaxchangescorrespondingtothestructuralmodel.
Asbefore,linenumbers(notrequiredintheactualLISRELsyntax)areincludedforreference
andonlythecommandsrelatingtothemodelspecificationareshown.
MO NY=8 NE=2 NX=8 NK=2 PH=SY,FR PS=DI,FR GA=FU,FI BE=FU,FI

VA 1.0 LX 1 1 LX 5 2 LY 1 1 LY 5 2
FR LX 2 1 LX 3 1 LX 4 1 LX 6 2 LX 7 2 LX 8 2
FR LY 2 1 LY 3 1 LY 4 1 LY 6 2 LY 7 2 LY 8 2
FR GA 1 1 GA 1 2
FR BE 2 1
Thestructuralmodelcommandshaveseveralchanges:
1. TheMOstatementnowprovidesvaluesfor:
a. Thenumberofindicatorsofendogenousconstructs(NY=8)
b. Thenumberofendogenousconstructs(NE=2)
c. Thenewnumberofindicatorsofexogenousconstructs(NX=8)
d. Thenewnumberofexogenousconstructs(NK=2)
2. The MO statement now provides the parameter matrices for the structural parameter
estimates:
a. GA stands for the relationships between exogenous and endogenous constructs (, or
gamma).Itisspecifiedasfull(FU)andfixed(FI).Theconventionistospecifyindividual
freeelementsbelow.

b. BEstandsfortherelationshipsbetweenendogenousconstructs().Itisalsospecifiedas
FUandFI.ThefreeelementswillbespecifiedwithaFRstatementbelowtheMOline.
3. The scales are set for the exogenous and endogenous constructs by setting one loading to
beequalto1.0asfollows:
VA 1.0 LX 1 1 LX 5 2 LY 1 1 LY 5 2
4. The factor pattern for both
x
and
y
(the respective loading matrices) is specified in much
the same way as in the CFA model. The only difference is that the item indicators for
constructsCandDarenowreferredtobyLYinsteadofLX.
5. The pattern of hypotheses between latent constructs is represented by freeing the
appropriateelementsofandB,respectively.
FR GA 1 1 GA 1 2
FR BE 2 1
The remainder of the model syntax can be specified as in the CFA model setup with the
exception of the optional label statements (LA). Separate label statements are needed for
exogenous and endogenous constructs. These will be illustrated in the HBAT example in the
followingsection.
HBAT:THECFAMODEL
Theexamplesaboveareforasimplifiedpathmodel,plustheydonotcontainthecomplete
LISRELcommandset.ThefollowingexampleusestheHBATexamplefromthetextbookand
demonstratestheprocessofspecifyingtheCFAmodelinLISRELcommands.Asbefore,line
numbershavebeenaddedforreferencepurposes,butarenotrequiredinLISREL.The
commandsfortheCFAanalysisareasfollows:
01 TI HBAT CFA MODEL
02 DA NI=28 NO=399 MA=CM
03 CM FI=HBAT.COV
04 LA
05 ID JS1 OC1 OC2 EP1 OC3 OC4 EP2 EP3 AC1 EP4 JS2 JS3 AC2 SI1 JS4
SI2 JS5 AC3 SI3 AC4 SI4
06 C1 C2 C3 AGE EXP JP
07 SE
08 JS1 JS2 JS3 JS4 JS5 OC1 OC2 OC3 OC4 SI1 SI2 SI3 SI4 EP1 EP2
EP3 EP4 AC1 AC2 AC3 AC4/
09 MO NX=21 NK=5 PH=SY,FR TD=DI,FR
10 VA 1.00 LX 1 1 LX 6 2 LX 10 3 LX 14 4 LX 18 5
11 FR LX 2 1 LX 3 1 LX 4 1 LX 7 2 LX 8 2 LX 9 2
12 FR LX 11 3 LX 12 3 LX 13 3 LX 15 4 LX 16 4 LX 17 4
LX 19 5 LX 20 5 LX 21 5
13 LK
14 JS OC SI EP CA
15 PD
16 OU RS SC ND=2

Line01issimplyatitlestatement.Theusercanenteranythingonthislinethathelps
identifytheanalysis.Line02isadatastatement.ItmustbeginwithDAandtellstheSEM
programthat28variablesareincludedinthedatasetof399observations.Althoughthedata
setoriginallycontained400observations,oneresponsepointwasdeletedforbeingoutof
rangeandanotherwassimplymissing.Usingpairwisedeletionandthepreviousruleofthumb,
thenumberofobservationswassetattheminimumnumberofobservationsforanycovariance
computation.Inthiscase,atleast399observationsareinvolvedinanysinglecovariance
computation.Thisnumbercanbeverifiedbyexaminingthestatisticaloutputforthecovariance
computations.Iflistwisedeletionhadbeenused,thenNOwouldbesetat398sincebothcases
withamissingresponsewouldbedeletedfromanycomputations.MA=CMdenotesthatthe
inputmatrixisacovariancematrix.Line03indicatesthatacovariancematrix(CM)isstoredina
file(FI)namedHBAT.COV.Line04isalabelsstatementandmustbeginwithLA.Thelabelsare
listedbeginningonthelinebelow.Lines05and06showthelabelsforthe28variables.Users
canchooseanylabelstherespectiveprogramwillallow.Inthiscase,HBATlabeledthevariables
withinitialsfromtheconstructnameslikeJS1,JS2,...,SI4.TheycouldhaveusedX1X28orV1
V28oranyothersimilarabbreviation.Onelabelisnecessaryforeachvariableinthedataset.
Line07isaselectstatementandmustbedenotedwithSE.Itindicatesthatthevariables
listedonthenextline(s)aretheonestobeusedintheanalysis.A/indicatestheendofthe
selectedvariableslist.Theorderisparticularlyimportant.Whateverislistedfirstwillbecome
thefirstobservedvariable.Forexample,thefirstmeasuredvariableintheCFAprogram,
designatedasx
1
(thesmallxwithsubscripthererepresentingthefirstobservedvariable
selectedandcorrespondstotheloadingestimate
x1,1
),willberepresentedbytheinputted
variablelabeledJS1.SI4,the21stvariableontheSEline,willbecomethe21stmeasured
variableorx
21
,andtheloadingestimatesassociatedwiththisvariablewillbefoundinthe21st
rowofthefactorloadingmatrix(
x21,5
of
x
inthiscase).
Onlyinrarecircumstanceswillthevariablesbestoredintheoriginaldatafileintheexact
orderthatwouldmatchtheconfigurationcorrespondingtothetheorybeingtested.Also,the
userseldomincludesallvariablesintheCFAbecausemostdatawillalsocontainsome
classificationvariablesoridentifyingvariablesaswellaspotentialvariablesthatweremeasured
butnotincludedintheCFA.Theselectprocess,whetherthroughastatementoradropdown
menu,isthewaythevariablesinvolvedintheCFAareselected.
Line09isamodelstatementandmustbeginwithMO.Modelstatementsindicatethe
respectivenumbersofmeasuredandlatentvariablesandcanincludedescriptionsofthekey
matricesofparameters.Theabbreviationsshownherearerelativelyeasytofollow.NXstands
fornumberofxvariables,inthiscase21.NKstandsforthenumberofconstructs,inthiscase
5.PHindicatesthatthematrixofcovariancesbetweenthe5constructs()willbesymmetric
(SY)andfree(FR).Inotherwords,theconstructvariances(thediagonalof)andthe
covariancebetweeneachpairofconstructswillbeestimated.TDisthematrixoferror
variancesandcovariances.Itissetasdiagonal(DI)andfree(FR),soonlytheerrorvariancesare
estimated.AnyparametermatrixnotlistedintheMOlineissetattheprogramdefaultvalue.
Thereadercanconsulttheprogramdocumentationforotherpossibleabbreviationsand
defaults.

Line10isavaluestatement(VA).Valuestatementsassignavaluetoafixedparameter.In
thiscase,eachoftheparameterslistedonthislineisfixedto1.0.Thisstatementsetsthescale
fortheconstructssothatoneitemisfixedto1.0oneachconstruct.LX1,1representsthe
parameterforthefirstloadingonthefirstconstruct(
x1,1
).TheLstandsforlambda,theXisan
xvariableand11standforthemeasuredvariablenumberandconstructnumber,respectively.
Thus,LX2,1standsfortheparameterrepresentingthefactorloadingofthesecondmeasured
variable(x
2
)onthefirstlatentconstruct(
1
),or
x2,1
.Factorloadingsinareflectivefactormodel
canbeexpressedequallyascausalpaths.Usingthisterminology,LX21,5standsforthepath
fromconstruct
5
tox
21
(
x21,5
).
Lines11and12startwithFRanddesignatethefreeloadingestimates.The16loadings
referredtoontheselineswillbeestimatedandshownasfactorresultsintheoutput(in
x
).
Withthefiveestimatesfixedat1online10and16loadingsestimated,84elementsremainin
thefactorpatternfixedatzero(21variables5constructs=105potentialloadings;10516
5=84).RecallthatEFAwouldproduceanestimateforall105loadings.Thepatternoffreeand
fixedloadingsmatchesthetheoreticalstructureproposedinthemeasurementmodel.
Consistentwiththecongenericmodelproposed,onlyoneloadingestimateisfreeforeach
measuredindicatorvariable.Inotherwords,eachmeasuredindicatorvariableloadsononly
oneconstruct.
Line13isanotherlabelline.Itiswherethelabelsforthelatentconstructscanbelisted.LK
standsforlabelsforksi().Theactuallabelsappearonthenextlineorlinesifnecessary.Inthis
case,thelabelsmatchtheconstructabbreviationsprovided(JS,OC,SI,EP,andAC).Line15,
withtheabbreviationPD,requeststhatapathdiagrambedrawnbytheprogramdepictingthe
specifiedmodelandpathestimates.TheOUline(16)isrequiredandiswhereanyoneof
numerousoptionscanberequested.Forexample,theSCisrequestingthatcompletely
standardizedestimatesbeincludedintheoutput.RSrequeststhatallmodelresidualsresulting
fromestimatingthemodelbeshown,includingboththestandardizedandnonstandardized
residuals.ND=2meansthatresultswillbeshowntotwosignificantdigits.
AttimesaresearchermaywishtoplaceadditionalconstraintsonaCFAmodel.For
instance,itissometimesusefultosettwoormoreparametersasequal.Itwouldproducea
solutionthatrequiresthevaluesfortheseparametersbethesame.Iftauequivalenceis
assumedforinstance,thisconstraintisneeded.WithLISREL,thistaskcanbedoneusingtheEQ
commandline.Similarly,researcherssometimeswishtosetaspecificparametertoaspecific
valuebyusingtheVAcommandline.Additionalinformationaboutconstraintscanbefoundin
thedocumentationfortheSEMprogramofchoice.
HBAT:THESTRUCTURALMODEL
TheHBATCFAcanbetransformedintotheHBATstructuralmodelaswasdoneearlierinthe
example. The LISREL commands for the structural model are shown below, followed by a
discussion of making the changes from the CFA to the structural model. Again, line numbers
havebeenaddedtothefarlefttoaidindescribingthesyntax.

01 TI HBAT EMPLOYEE RETENTION MODEL
02 DA NI=28 NO=399 NG=1 MA=CM
03 CM FI=HBAT.COV
04 LA
05 ID JS1 OC1 OC2 EP1 OC3 OC4 EP2 EP3 AC1 EP4 JS2 JS3 AC2 SI1 JS4
SI2 JS5 AC3 SI3 AC4 SI4
06 C1 C2 C3 AGE EXP JP
07 SE
08 JS1 JS2 JS3 JS4 JS5 OC1 OC2 OC3 OC4 SI1 SI2 SI3 SI4 EP1 EP2
EP3 EP4 AC1 AC2 AC3 AC4/
09 MO NY=13 NE=3 NX=8 NK=2 PH=SY,FR PS=DI,FR BE=FU,FI GA=FU,FI
TD=DI,FR TE=DI,FR
10 VA 1.00 LX 1 1 LX 5 2 LY 1 1 LY 6 2 LY 10 3
11 FR LX 2 1 LX 3 1 LX 4 1 LX 6 2 LX 7 2 LX 8 2
12 FR LY 2 1 LY 3 1 LY 4 1 LY 5 1 LY 7 2 LY 8 2 LY 9 2 LY 11 3 LY
12 3 LY 13 3
13 FR GA 1 1 GA 2 1 GA 1 2 GA 2 2
14 FR BE 2 1 BE 3 1 BE 3 2
15 LK
16 EP AC
17 LE
18 JS OC SI
19 PD
20 OU RS SC MI EF ND=2
The first change from the CFA setup is noted in line 09. The model statement must now
specify a number of variables and constructs for both exogenous and endogenous constructs.
Thus,theMOlinespecifiesNY=13(5itemsforJS,4itemsforOC,4itemsforSI).Eventhough
these are the same items as represented by these constructs in the CFA model, they now
become y variables because they are associated with an endogenous construct. Their loading
parametersarenowchangedtobeconsistentwiththisto
y
(LY).Next,theMOlinespecifiesNE
= 3, indicating three endogenous constructs. This process is repeated for the exogenous
constructs(NX=8andNK=2).PHandTDremainthesame.
Several new matrices are specified. BE = FU,FI means that B, which will list all parameters
linking endogenous constructs with one another (), is set to full and fixed. It means we will
freetheelementscorrespondingtothefollowinghypotheses.GArepresenting,whichwilllist
all parameters linking exogenous constructs with endogenous constructs (), is treated in the
same way. Because we now have endogenous constructs, the error variance terms associated
withthe13yvariablesarenowshownin
,whichisabbreviatedwithTE=DI,FR,meaningitis
adiagonalmatrixandthediagonalelementswillbeestimated.
Line 10 sets the scale for factors just as in the CFA model with the exception that three of
the set values are for y variables (
y
values: LY1,1; LY6,2; LY10,3). Lines 11 and 12 specify the
freevaluesforthemeasureditemsjustasintheCFA.Wearefollowingtheruleofthumbthat
thefreefactorloadingparametersshouldbeestimatedratherthanfixedeventhoughwehave
some idea of their value based on the CFA results. Lines 13 and 14 specify the pattern of free
structuralparameters.Line13specifiesthefreeelementsof.ThesecorrespondwithH
1
H
4

inFigure126(
1,1
islistedasGA1,1).Similarly,line14specifiesthefreeelementsofB.Lines15
and16liststhelabelsfortheconstructs(LK).Lines17and18dothesamefortheconstructs
(LE). Line 19 contains a PD that tells the program to generate a path diagram from the input.
Line20istheoutputlineandisthesameasintheCFAexampleexceptfortheadditionofEF,
whichwillprovideaseparatelistingofalldirectandindirecteffects.
If the user is using a graphical interface (e.g., AMOS or LISREL), the user will need to make
thecorrespondingchangestothepathdiagram.Thesechangeswouldincludemakingsurethe
constructs are properly designated as exogenous or endogenous and that observed variables
each have a corresponding error variance term. Then each of the curved twoheaded arrows
that designated covariance between constructs in CFA will have to be replaced with a single
headed arrow to represent hypothesized relationships. Arrows between constructs for which
no relationship is hypothesized are unnecessary. Therefore, the twoheaded paths between
these constructs in the CFA can be deleted. Once these changes are made, the user can
reestimate the model and the results should now reflect the structural model results. If the
programsyntaxhasbeenchangedasindicated,theprogramwillproducetheappropriatepath
diagramautomatically.
AvisualdiagramcorrespondingtotheSEMcanbeobtainedbyselectingStructuralModel
fromtheviewoptionsandrequestingthatthecompletelystandardizedestimatesbedisplayed
bytheSEMprogram.InLISREL,forexample,thevaluesonthepathdiagramcanberequested
sothateithertheestimatesareshownonthediagram,thetvaluesforeachestimate,orother
keyestimatesareshownincludingthemodificationindices.
HOWTOFIXFACTORLOADINGSTOASPECIFICVALUEINLISREL
IfaresearcherwishedtofixthefactorloadingsofaSEM modeltothevaluesidentifiedinthe
CFA,proceduressuchasthosedescribedherecanbeused.Tospecifythevaluesshowninthe
pathmodelbelow,theresearcherwouldtakethefollowingstepsifusingtheLISRELsoftware.

Thefollowingloadingestimateswouldbefixedandtheirvaluessetasfollows:
FI LX 1 1 LX 2 1 LX 3 1 LX 4 1 LX 5 2 LX 6 2 LX 7 2 LX 8 2
FI LY 1 1 LY 2 1 LY 3 1 LY 4 1 LY 5 2 LY 6 2 LY 7 2 LY 8 2
VA .80 LX 1 1
VA .70 LX 2 1
VA .80 LX 3 1
VA .75 LX 4 1
VA .90 LX 5 2
VA .80 LX 6 2
VA .75 LX 7 2
VA .70 LX 8 2
VA .70 LY 1 1
VA .90 LY 2 1
VA .75 LY 3 1
VA .75 LY 4 1
VA .85 LY 5 2
VA .80 LY 6 2
VA .80 LY 7 2
VA .70 LY 8 2

TheerrorvariancetermsalsocanbefixedtotheirCFAestimatesasshownhere:

FI TD 1 1 TD 2 2 TD 3 3 TD 4 4 TD 5 5 TD 6 6 TD 7 7 TD 8 8
FI TE 1 1 TE 2 2 TE 3 3 TE 4 4 TE 5 5 TE 6 6 TE 7 7 TE 8 8
VA .36 TD 1 1
VA .51 TD 2 2
VA .36 TD 3 3
VA .44 TD 4 4
VA .19 TD 5 5
VA .36 TD 6 6
VA .44 TD 7 7
VA .36 TD 8 8
VA .51 TE 1 1
VA .81 TE 2 2
VA .44 TE 3 3
VA .44 TE 4 4
VA .28 TE 5 5
VA .36 TE 6 6
VA .36 TE 7 7
VA .51 TE 8 8
Theresearchercouldthenproceedtospecifythefreeelementsofthestructuraltheory.
MEASUREDVARIABLEANDCONSTRUCTINTERCEPTTERMS
Itoftenbecomesnecessarytousethemeasuredvariableandlatentvariablemeansindrawing
conclusionsaboutsimilaritiesanddifferencesbetweengroups.Untilnow,noSEMequationhas
shownameanvalue.Now,however,themeansmaybeconsidered.
Onewaythatwecouldthinkaboutthemeanvalueofanymeasuredvariableistothinkofit
asthesumofitszerointerceptterm,plusthefactorloading,timestheaveragevalueofthe
latentconstruct.Inequationform,itwouldlooklikethefollowingexpressedintermsofx
1
:
p
X1
=
X1
+ z
X1
{1
The
1
representsthemeanvalueforthefirstlatentconstruct
1
,the
X1
representsthe
meanofthemeasuredvariablex
1
,andthe
X1
isthezerointerceptforx
1
.Moregenerally,
representsthemeanforanylatentconstruct.Mathematically,itisalsothezerointerceptterm
whensolvingfor.Eventhoughthemathematicsinthiscalculationmaybedifficulttofollow,it
isimportanttoknowthatunlessspecificinstructionsareprovidedtotheSEMprogram,itwill
notconsidernorestimateconstructmeansofanytype.
Thisequationcanberearrangedtosolveforeither
X1
or.Ifanyhypothesesconcern
differencesbetweenconstructmeans,thosedifferencescanbefoundinthevaluesfor.
PATHMODELSPECIFICATIONWITHAMOS
ProgramstatementscanalsobewrittenforAMOSthatwouldformthemodelinthesameway
astheLISRELstatement.However,theassumptionwithAMOSisthattheuserwillworkwitha
pathdiagram.Inessence,thepathdiagramprovidestheframeworkfromwhichtobuildthe

model.However,theusermustassignvariablestoeachrectangle,whichrepresentsa
measuredvariable,andassignconstructnamestoeachoval.Likewise,theusermustspecific
namesforeachmeasuredvariableerrorterm.Then,theappropriatearrowsmustbedrawnto
formthemodel.Theusermustbecarefulthatvariablesareassignedcorrectly.Dropdown
windowscanbeusedtoaddconstraintstothemodelandtoperformadvancedapplications
likemultiplegroupanalysis.WhileitispossibletospecifytheSEMmodelthroughcommands,
AMOSisdesignedtobeusedthroughthegraphicalinterfaceandthisistherecommended
routeformostusers.
RESULTSUSINGDIFFERENTSEMPROGRAMS
AlthoughtheinputfordifferentSEMprogramsvaries,theresultsshouldbeessentiallythe
same.Thealgorithmsmayvaryslightly,butamodelthatdisplaysgoodfitusingoneSEM
programalsoshoulddisplaygoodfitinanother.Eachhasitsownidiosyncrasiesthatmay
preventthesamemodelspecificationfrombeingestimated.Forinstance,somemakeitmore
orlessdifficulttouseeachofthemissingvariableoptionsjustmentioned.Eachapproachcan
beeasilyspecifiedwithLISREL,butAMOSusesEMalone.Listwisedeletion,forexample,canbe
performedwithAMOSbyscreeningobservationswithmissingdatapriortobeginningthe
AMOSroutine(e.g.,withSPSS).
Theoverallmodelfitstatistics,includingthe
2
andallfitindices,shouldnotvaryinany
consequentialwaybetweentheprograms.Similarly,theparameterestimatesshouldalsonot
varyinanyconsequentialway.Differencescanbeexpectedintwoareas.
Oneareawheredifferencesinthenumericalestimatesmayvaryisintheresiduals.In
particular,somedifferencesmaybefoundbetweenAMOSandtheotherprograms.Without
gettingintothedetails,AMOSusesadifferentmethodforscalingtheerrortermsofmeasured
variablesthandotheotherprograms.Thisformathastodowithsettingthescalefortheerror
terms,muchaswesetthescaleforthelatentconstructsinaSEMmodel.Thismethodmay
causerelativelysmalldifferencesinthevaluesforresidualsandstandardizedresiduals
computedwithAMOS.However,thedifferencesdonotaffecttherulesofthumbgiveninthe
text.
Anotherareawherenumericalestimatesmayvaryisinthemodificationindices.Again,
AMOStakesadifferentcomputationalapproachthandosomeoftheotherSEMprograms.The
differenceliesinwhetherthechangeinfitisisolatedinoneorseveralparameters.Onceagain,
althoughtheusercomparingresultsbetweenAMOSandotherprogramsmayfindsome
differencesinMI,thedifferencesshouldnotbesolargeastoaffecttheconclusionsinmost
situations.So,onceagain,therulesofthumbfortheMIholdusinganySEMprogram.
ADDITIONALSEMANALYSES

TESTINGFORDIFFERENCESINCONSTRUCTMEANS
Afinaltypeofmultigroupcomparisonisthetestfordifferencesinconstructmeans.Ifatleast
partialscalarinvarianceispresent,wecanoperationalizeavalueforthemeansofthelatent
constructs.Inthisway,wwtelltheSEMprogramthatweareinterestedinanalyzingmeans.An
earlierdiscussionshowedtheequationtorepresentlatentconstructmeans.Inonewayor
anotherthough,theSEMprogrammustbetoldweareinterestedinthemeansofthelatent
constructs.
SEMprogramscomparemeansonlyinarelativesense.Inotherwords,theycantellyou
whetherthemeanishigherorlowerrelativetoanothergroup.Onereasonforthislimitation
hastodowithidentificationgiventhattheintercepttermsarenowbeingestimated.Aresultis
thatthevectoroflatentconstructmeans(containedinthekappamatrix)hastobefixedtozero
inonegrouptoidentifythemodel.Werefertothisgroupasgroup1.Itcanbefreelyestimated
intheothergroup(s)andtheresultingvaluescanbeinterpretedashowmuchhigherorlower
thelatentconstructmeansareinthisgrouprelativetogroup1.
Assumewehaveatwogroupmodelwiththreeconstructsineachgroup.TheSEM
outputwillnowincludeestimatesforthevectoringroup2(i.e.,thecomparisonofgroup2
relativetogroup1).Typically,thisoutputwouldincludeanestimatedvalue,astandarderror,
andatvalueassociatedwitheachvalue.Forinstance,itmaylooklikethis:
KAPPA()
Construct1
Construct2 Construct3
2.6
0.09 3.50
(0.45)
(0.60) (1.55)
5.78
0.10 2.25
Thesevaluessuggestthatthemeanfortheconstruct1is2.6greateringroup2thaningroup1.
Thisdifferenceissignificantasevidencedbythetvalueof5.78(p<.001).Themeanfor
construct2,differingby0.09,isnotsignificantlydifferent(t=0.10).Themeanforconstruct3
ontheotherhand,issignificantlyloweringroup2relativetogroup1(t=2.25;p<.05).

ITEMPARCELINGINCFAANDSEM
Itemparcelingreferstocombiningmeasuredvariablesintosetsofseveralvariablesbyeither
summingoraveragingtheitems.Parcelingprovidesawayofdealingwithanunmanageable
number of measured reflector variables per construct. For example, some psychological

scales may contain more than 100 items to capture only two or three basic personality
dimensions. Thus, even with a few constructs one could end up with far more than 100
measureditems.SEMapplicationsaredifficulttomanagewithsomanymeasuredvariables.
Using item parceling a single latent construct with 40 measured items (x

1
x
40
) could
berepresentedbyeightparcels,eachconsistingof5ofthe40measureditems.Aparcelisa
mathematical combination summarizing multiple variables into one. In the extreme, all
measureditemsforaconstructcanbecombinedintooneaverageorsumofthosevariables.
In Chapter 3, we discussed how to create a summated construct in this fashion. The term
compositeindicatorisgenerallyusedtorefertoparcelingresultinginonlyoneparcelfromall
themeasureditemsforaconstruct.
Numerous issues are associated with item parceling. These issues include the
appropriateness of parceling, which items should be combined into a parcel, and what the
effects of parceling are on evaluating models. Parceling has the potential to improve model
fit simply because it reduces the complexity of the model, and models with fewer variables
have the potential for better fit. Better fit alone, however, is not a sufficient rationale for
combining multiple items into one because the primary goal is creating a model that best
represents the actual data. Further, item parcels can often mask problems with item
measuresandsuggestabetterfitthanactuallyexistsinreality.Parcelingalsocanhideother
latent constructs that exist in the data. So, a covariance matrix that actually contains five
latentconstructsmaybeadequatelybutfalselyrepresentedbythreelatentconstructsusing
parceling.
WhenIsParcelingAppropriate?
Itemparcelingshouldonlybeconsideredwhenaconstructhasalargenumberofmeasured
variable indicators. For instance, applications involving fewer than 15 items do not call for
parceling.Similarly,parcelingisnotusedwithformativemodelsbecauseitisimportantthat
allcausesofaformativefactorbeincluded.Parcelingisappropriatewhenalltheitemsfora
construct are unidimensional. That is, even with a large number of measured items, they
shouldallloadhighlyononlyoneconstructanditshoulddisplayhighreliability(.9orbetter).
Most importantly, parceling is appropriate when information is not lost by using parcels
instead of individual items. Thus, some simple checks prior to parceling would involve
runningaCFAontheindividualfactortocheckforunidimensionalityandtoseewhetherthe
construct reflected by all individual items relates to other constructs in the same way as a

constructreflectedbyasmallernumberofparcels.
HowShouldItemsBeCombinedintoParcels?
Traditionally,littlethoughtwasgiventohowitemsshouldbecombined.However,the
combinationstrategycanaffectthelikelihoodthataCFAisactuallysupportingafalse
measurementtheory.Althoughmanyintricaciesareassociatedwiththecombination
strategies,twosimpleconsiderationsleadtothebestperformancewhenaresearchermustuse
itemparcels.Oneconsiderationisempiricalandtheotheristheoretical.Giventhatthe
individualitemssuggestunidimensionality,thebestparcelsareformedbyitemsthatdisplay
approximatelythesamecovariance,whichshouldleadthemtohaveapproximatelythesame
factorloadingestimates.Further,theparcelsshouldcontaingroupsofitemswiththemost
conceptualsimilarity.Thatis,itemswiththeclosestcontentvalidity.Thus,parcelswithitems
showingapproximatelythesameamountofcovarianceandthatshareaconceptualbasiswill
tendtoperformwellandrepresentthedatamostaccurately.
MEASUREMENTBIAS
Researchers sometimes become concerned that survey responses are biased based on the
way the questions are asked. For instance, it could be argued that the order in which
questions are asked could be responsible for the covariance among items that are grouped
closetogether.Ifso,anuisancefactorbasedonthephysicalproximityofscaleitemsmaybe
explainingsomeoftheinteritemcovariance.
Similarly,researchersoftenarefacedwithresolvingthequestionofconstantmethodsbias.
Constantmethodsbiaswouldimplythatthecovarianceamongmeasureditemsisdrivenbythe
factthatsomeoralloftheresponsesarecollectedwiththesametypeofscale.Aquestionnaire
usingonlysemanticdifferentialscales,forinstance,maybebiasedbecausetheopposingterms
response form becomes responsible for covariance among the items. Thus, the covariance
couldbeexplainedbythewayrespondentsuseacertainscaletypeinadditiontoorinsteadof
thecontentofthescaleitems.Here,asimpleillustrationisprovidedusingtheHBATexample.It
shows how a CFA model can be used to examine the possibility of measurement bias in the
formofanuisancefactor.
The HBAT employee questionnaire consists of several different types of rating scales.
Although it could be argued that respondents prefer a single format on any questionnaire,
severaladvantagescomewithusingasmallnumberofdifferentformats.Oneadvantageisthat
theextenttowhichanyparticularscaletypeisbiasingtheresultscanbeassessedusingCFA.

In this case, HBAT is concerned that the semantic differential items are causing
measurementbias.Theanalystarguesthatrespondentshaveconsistentpatternsofresponses
tosemanticdifferentialscalesnomatterwhatthesubjectoftheitemis.Therefore,asemantic
differential factor may help explain results. A CFA model can be used to test this proposition.
One way to do so is to create an additional construct that is also hypothesized as causing the
semantic differential items. In this case, items EP4, JS2, JS3, AC2, and SI4 are measured with
semantic differential scales. Thus, the model needs to estimate paths between this new
constructandthesemeasureditems.Theadditionofanuisancefactorofthistypeviolatesthe
principlesofgoodmeasurementandsothenewmodelwillnothavecongenericmeasurement
properties.
We will modify the original HBAT CFA model shown in the text. A sixth construct is
introduced (
6
). Next, dependence paths (causal in this case) would be estimated (drawn if
using a path diagram) from
6
to EP4, JS2, JS3, AC2, and SI4. Thus, the factor pattern no
longer exhibits simple structure because each of these measured variables is now
determinedbothbyitsconceptualfactorandbythenewconstruct
6
.
Theanalystteststhismodelandobservesthefollowingfitstatistics.The
2
=232.6with
174 degrees of freedom and the RMSEA, PNFI, and CFI are .028, .80, and .99, respectively.
The added paths have not provided a poor overall fit although the RMSEA has increased
slightly and the PNFI has decreased. However, the
2
= 4.0 (236.6 232.6) with 5 (179
174) degrees of freedom, is insignificant. In addition, none of the estimates associated with
the bias factor (
6
) are significant. The completely standardized estimates of lambda (factor
loadings)andassociatedtvaluesareshownhere:
Parameter
Estimate tValue
x2,6
0.14 1.19
x3,6
0.01 0.08
x17,6
0.16 1.32
x19,6
0.07 0.84
x21,6
0.20 1.48
Also,thevaluesfortheoriginalparameterestimatesremainvirtuallyunchangedas
well.Thus,basedonthemodelfitcomparisons,theinsignificantparameterestimates,and
theparameterstability,noevidencesupportsthepropositionthatresponsestosemantic
differentialitemsarebiasingresults.TheHBATanalystconcludes,therefore,thatthiscaseis
notsubjecttomeasurementbias.Anotherfactorcouldbeaddedtoactasapotential
nuisancecausefortheitemsrepresentinganotherscaletype,suchasallLikertitems.The
testwouldproceedinmuchthesameway.Theendresultofallofthesetestsisthatthe
researchercanproceedtotestmorespecifichypothesesaboutemployeeretentionand

relatedconstructs.
CONTINUOUSVARIABLEINTERACTIONS
Anapproachforhandlingacontinuousmoderatorwhichdoesnotinvolvecreatinggroupsfrom
thecontinuousmoderatoristocreateaninteractionbetweenthemoderatorandthepredictor.
Singlevariable interactions were treated in the text, so we focus here on a moderating
construct that would be measured by multiple indicators. Consider a SEM model with two
exogenous constructs predicting a single endogenous construct. Eachconstruct is indicated by
four measured items. If the first construct (
1
) is hypothesized as the predictor construct and
thesecondconstruct(
2
)ishypothesizedasamoderator,thenaninteractionconstructcanbe
created to represent the moderating effect by multiplying the indicators of the predictor and
moderator constructs together. Using this rationale, the indicators for the third interaction
construct(
3
)canbecomputedasfollows:
x
9
= x
1
x
5
x
10
= x
2
x
6
x
11
= x
3
x
7
x
12
= x
4
x
8
Thesecomputedvariablescanthenbeaddedtotheactualdatacontaining12measured
variablesandthecovariancetermsbetweenthesecomputedvariablesandtheotherscanbe
calculated.Now,thecovariancematrixforthismodelwouldchangefrom1212to1616.
Thiscanbeshowninapathmodelformbythefollowingspecification.

Estimatingthismodeliscomplicatedbyseveralfactors.Thesefactorsincludethefactthat
theassumptionthaterrortermsareuncorrelatedisnolongerfeasiblebecausetheloadingsfor
thethirdconstruct(
3
)areamathematicalfunctionofthoseforconstructs1and2(
1
and
2
).
ThisfactleadstoaquitecomplexSEMmodelsetupthatisrecommendedforadvancedusers
only.Soitisonlybrieflydescribedhere.Thissetuprequiresthattheintercepttermsforthe
measureditems(
x
)beestimatedasdescribedearlier.Theexogenousfactorloadingpattern
cannolongerexhibitsimplestructure.Eventhoughtheloadingestimatesforthethird
constructcanbecomputedbymultiplyingtheloadingestimatescorrespondingtothevariables
thatcreatedeachinteractionindicator,crossconstructloadingsalsoexistfortheinteraction
term.Theyarecomputedbycrossingthematching
x
termswithloadingestimates.Again,this
processisrathercomplextofollow,butasanexample,the12throwof
x
wouldendupas:
4

8
8

4
8


Inaddition,theerrorvariancecovariancematrixforthexvariables(
)mustnowinclude
termsfortheappropriateerrorcovarianceitemsthatexistduetothecomputationalnatureof
theinterceptconstruct.Theseitemsneednotbeestimatedbecausetheyaredetermined
mathematicallyastheintercepttermforthemeasureditemusedtocomputetheinteraction
indicatortimestheerrorvarianceforaconstruct.Thisconceptismoreeasilyillustratedbyan
example.
9
istheerrorvariancetermforx
9
,thefirstindicatorforthemoderatorconstruct.
Becauseitiscomputedasx
1
timesx
5
,anerrorcovariancetermisneededforboth
9,1
and
9,5
.Thevalueswouldbesetas
1
times
1,1
and
5
times
5,5
,respectively.
Afterfinishingasetupfollowingalongtheselines,themodelcanbeestimatedspecifying
onlythestructuralpathbetweentheinteractionconstructandtheoutcome.Ifmoderationis
supported,thecorrespondingestimate,
3,1
inthiscase,wouldbesignificant.Realizethatthe
effectsof
1
and
2
on
1
areofquestionablevalidityinthepresenceofasignificantinteraction.
Therefore,theyshouldbeestimatedandinterpretedonlyifthestructuralinteractionterm(
3,1
)
isinsignificant.
Interactiontermssometimescauseproblemswithmodelconvergenceanddistortionofthe
standarderrors.Therefore,largersamplesareoftenrequiredtominimizethedistortion.An
absoluteminimumsamplesizewouldbe300forthistypeofanalysiswithasamplesizeofmore
than500recommended.

SEM - Basics A Supplement To Multivariate Data Analysis

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

SEM - Basics A Supplement To Multivariate Data Analysis

Hochgeladen von

Copyright:

Verfügbare Formate

MultivariateDataAnalysis,PearsonPrenticeHallPublishing Page1

The supplement is not intended to be a comprehensive primer on all of the SEM

xi(KSIorKZI) An exogenous construct associated with measured X

eta(eightta) An endogenous construct associated with measured Y

lambdax A path representing the factor loading between an

lambday A path representing the factor loading between an

capitallambda The set of loading estimates represented in a matrix where

capitalphi A way of referring to the covariance or correlation matrix

capitalgamma A way of referring to the entire set of relationships for a

beta(bayta) A path representing a causal relationship from one

capitalbeta A way of referring to the entire set of relationships for a

delta The error term associated with an estimated, measured x

epsilon The error term associated with an estimated, measured y

MO NY=8 NE=2 NX=8 NK=2 PH=SY,FR PS=DI,FR GA=FU,FI BE=FU,FI

Using item parceling a single latent construct with 40 measured items (x

Das könnte Ihnen auch gefallen