Anda di halaman 1dari 70

StructuralEquationModels:

PathAnalysiswithLatentVariables
Introductionto
StructuralEquationModeling
Lecture#12(andlast)
April25,2012
TodaysClass
Puttingitalltogether:
PathAnalysis
Observedvariables
ConfirmatoryFactorAnalysis/MeasurementModels
Latentvariables
Concernsinbuildingstructuralequationmodels
Modelpredictedcovariancematricesforpathanalysiswith
observedandlatentvariables
ExamplesofSEMuses
UNDERLYINGTHEORYOF
STRUCTURALEQUATIONMODELS
StructuralEquationModels
AlthoughthetermSEMcanbeappliedtomanysettings,I
viewthelabelasbeingusedtodescribeanalyseswith
observedandlatentvariables
Astructuralequationmodelconsistsoftwoparts:
Measurementmodel(s)foreachlatentvariable
Pathanalysisbetweenthelatentandobservedvariables
Uptothispoint,wehavecoveredbothinisolation today
weputthemtogethertoshowhowtheprocessworks
Youwillseethisextrastepisprettystraightforward
REVIEWOFPATHANALYSIS
TypesofVariablesintheAnalysis
AnimportantdistinctioninpathanalysisandSEMis
betweenendogenousandexogenousvariables
Endogenousvariable(s): variableswhosevariabilityis
explained byoneormorevariablesinamodel
Inlinearregression,thedependentvariableistheonly
endogenousvariableinananalysis
Exogenousvariable(s):variableswhosevariabilityisnot
explainedbyanyvariablesinamodel
Inlinearregression,theindependentvariable(s)arethe
exogenousvariablesintheanalysis
ERSH8750:Lecture#4 6
DirectandIndirectEffectsofHSLonMSE
ERSH8750:Lecture#4 7
Mathematics
SelfEfficacy
CollegeMath
Experience
HighSchool
Math
Experience
DirectEffect
Residual(Endogenous)
Variance
ExogenousVariances
ExogenousCovariances
4.1S9 (u.46S)
u.696
(u.2S8)
u.S6S
(u.1uS)
SS.797 (2.712)
1uS.128 (8.Su8)
1.74S (u.14u)
NotShownOnPathDiagram:
[
0
CC
= 6.9u4 (1.Su4)
[
0
MSL
= 49.S1S (2.412)
p
HSL
= 4.912 (u.u74)
IndirectEffect
PathAnalysisinMatrixForm
Ourpathmodelsimultaneousequationswere:
CC

= [
0
CC
+ [
HSL
CC
ESI

+ c

CC
HSE

= [
0
MSL
+ [
CC
MSL
CC

+ [
HSL
MSL
ESI

+ c

MSL
p =2 endogenousvariables
q =1 exogenousvariable
Alternatively,wecouldrephrasethisinmatrixform:
y

= u + By

+ Ix

Where:
x

= ESI

(matrixofsizeq x 1 containingobservedexogenousvariables)
y

=
CC

HSE

(matrixofsizep x 1 containingobservedendogenousvariables)
Then:
u =
[
0
CC
[
0
MSL
(matrixofsizep x 1 containinginterceptsforendogenousvariables)
B =
u u
[
CC
MSL
u
(ap x p matrixofcoefficientsrelatingtheendogenousvariablestothemselves)
I =
[
HSL
CC
[
HSL
MSL
(matrixofsizep x q relatingexogenousvariablestoendogenousvariable(s))

=
c

CC
c

MSL
N
2
(, +) (where+isthep x p residualcovariancematrix)
Here,+willbediagonal(nocovariance)aswedonothaveanymoredegreesoffreedom
ERSH8750:Lecture#4 8
PathAnalysisinMatrixForm
Theequationsfromthepreviousslidearecalledthestructural
form ofthepathmodel
Anotherformthatexistsinliteratureisthereducedform,
whereallendogenousvariablesareonthelefthandside


-1 -1

-1

Where

p

Thereducedformisnotasfrequentlyusedinpractice,but
doesariseinsomeresearchareasandinidentification
ERSH8750:Lecture#4 9
PathAnalysiswithMatrices
Althoughnotexplainedbyourmodel,wecouldstatethat
themeanvectorofexogenousvariableswas:
x HSL
Likewise,wecanstatethatthecovariancematrixofthe
exogenousvariablesis
HSL
2
Wewillusethesetermsinourmatrixversionofthe
modelpredictedmeanandcovariancematrix
ERSH8750:Lecture#4 10
ModelPredictedMeanVectorand
CovarianceMatrix
Theunconditionalmeanoftheendogenousvariablesis:

-1 -1
x
Thecovariancematrixoftheexogenousandendogenous
variablesisthen:
,x
-1 1 1
-1
-1
1 1
-1
Thepoint:thatmodelspecificationshavedirect
implicationsfortheparametersofthemultivariate
normaldistribution
ERSH8750:Lecture#4 11
MatchingMatriceswithResults
Tomorespecificallylinkourresultstothematricesfrom
thepreviouspage:
ERSH8750:Lecture#4 12
Name Matrix ModelEstimates
ResidualCovariance Matrix +
SS.797 u
u 1uS.128
Regression WeightsofExogenousonto
Endogenous
I
u.696
4.1S9
CovarianceMatrixofExogenous Variables 4 1.74S
MeanVectorof ExogenousVariables
x
4.912
VectorofEndogenous VariableIntercepts u
6.9u4
49.S1S
MatrixofEndogenousRegression Weights B
u u
u.S6S u
Inverse matrixusedincalculations I - B
-1
1 u
-u.S6S 1
ModelPredictedMeanVectorand
CovarianceMatrix
Theestimatedconditionalmeanoftheendogenousvariablesis:
Thesevaluescorrespondexactly(saturatedmodel)
Theestimatedcovariancematrixoftheexogenousand
endogenousvariablesis:
Thesearemostlyexact smalldifferences
ERSH8750:Lecture#4 13
REVIEWOFCONFIRMATORY
FACTORANALYSIS
OneFactorModelofFiveGRIItems
TheCFAmodelforthefiveGRIitems:

s1
= p
1
+ z
11
F
s1
+ c
s1

s2
= p
2
+ z
21
F
s1
+ c
s2

s3
= p
3
+ z
31
F
s1
+ c
s3

s4
= p
4
+ z
41
F
s1
+ c
s4

s5
= p
5
+ z
51
F
s1
+ c
s5
Here:

s
responseofsubjects onitemi
p

interceptofitemi (listedasameanasthisistypicallywhatitbecomes)
z
1
factorloadingofitemi onfactor1(onlyonefactortoday)
F
s1
latentfactorscoreforsubjects (sameforallitems)tofactor1(onlyonetoday)
c
s
regressionlikeresidualforsubjects onitemi
Weassumec
s
N u,

2
;

2
iscalledtheuniquevariance ofitemi
Wealsoassumec
s
andF
s1
areindependent
Also,wewillassumeF
s1
N p
P
1
, o
P
1
2
Typicallyp
P
1
= u (butnotalways)
Factorvariancecanbeestimatedorfixed(moreonbothinidentification)
ERSH8750:Lecture#5 15
OurCFAModelPathDiagram
MeasurementModel:
s=factorloadings
es=errorvariances
s=itemintercepts
StructuralModel:
o
F
1
2
=factorvariance

F1
=factormean
(Someofthesevalueswillhavetobe
restrictedforthemodeltobeidentified)
ERSH8750:Lecture#5
e
1
e
2
e
3
e
4
e
5
Y
4
Y
5
Y
3
Y
2
Y
1
F
1

11

21

31

41

51
o
F
1
2

F1
1
16
ModelPredictedMeanVector
Combiningacrossallitems,themeanvectorfortheitems
isgivenby:
I F

S
I
1
I
2
I
3
I
4
I
S
11
21
31
41
51
P
1
I
1
11 P
1
I
2
21 P
1
I
3
31 P
1
I
4
41 P
1
I
S
51 P
1
ERSH8750:Lecture#5 17
ModelImpliedCovarianceMatrix
Combiningacrossallitems,thecovariancematrixfortheitemsis
givenby:
L

= A4A
1
++
Getusedtoseeingthis althoughyoualreadyhave(seetheregression
slides)
o

1
2
o

1
,
2
o

1
,
3
o

1
,
4
o

1
,
S
o

1
,
2
o

2
2
o

2
,
3
o

2
,
4
o

2
,
S
o

1
,
3
o

2
,
3
o

3
2
o

3
,
4
o

3
,
S
o

1
,
4
o

2
,
4
o

3
,
4
o

4
2
o

4
,
S
o

1
,
S
o

2
,
S
o

3
,
S
o

4
,
S
o

S
2
=
z
11
z
21
z
31
z
41
z
51
o
P
1
2
z
11
z
21
z
31
z
41
z
51
+

1
2
u u u u
u
2
2
u u u
u u
3
2
u u
u u u
4
2
u
u u u u
5
2
=
ERSH8750:Lecture#5 18
PUTTINGITTOGETHER:PATH
ANALYSISWITHLATENTVARIABLES
ASmallSEMExample
TodemonstratehowSEMworks,wewilluseaverysmall
example:
Measurementmodel:threeGRIitemsformingonelatent
construct(gambling)
Note:withthreeitems,themeasurementmodelisjustidentified
(meaningperfectfit)
Pathmodel:ThepredictionofgamblingbytheSOGscore
Note:herewetreatSOGSscoreasbeingobservedwithouterror
GRI1
GRI3
GRI5
Gambling SOGS
Step#1:BuildingtheMeasurementModel
Thefirststepinastructuralequationmodelistobuildthe
measurementmodel
Here,themeasurementmodelissimplifiedsoastoshowhow
SEMworks
MeasurementModelFitAssessment
Ourthreeitemmeasurementmodelfitsperfectly
MeasurementModelParameterEstimates
MeasurementModel:
ImpliedCovarianceMatrix
Themeasurementmodelimpliedcovariancematrixis:

1

=
1.uuu
u.726
u.996
u.4u7
1.uuu u.726 u.996
+
u.648 u u
u u.SSS u
u u u.S46
1.uSS u.29S u.4uS
u.29S u.749 u.294
u.4uS u.297 u.9Su
Step2:Estimatingthe
StructuralEquationModel
Oncethemeasurementmodelisfoundtofit,thenext
stepistoestimatethefullstructuralequationmodel
SOGSsum istreatedasanexogenousvariable
Alsocalledanindependentvariable
GAMBLING(andtheitemsmeasuringit)aretreatedas
endogenousvariables
Alsocalleddependentvariables
SEM:ModelIdentification
AsSEMintegratesbothmeasurementandpathmodels,the
identificationrulesforSEMborrowfromboth
Themeasurementmodel(foralllatentvariables)mustbelocallyidentified
Includingrulesforsettingscaleoflatentfactor(s)
Thepathmodelmustbeidentified
Anecessarybutnotsufficientwayofensuringidentificationisthet
rule(countingrule)
Thenumberofparametersmustbelessthanthetotalnumberofmeans+
variances/covariancesofall observedvariablesintheanalysis
Numberofobservedvariablesinouranalysis:4
Numberofvariances/covariances:4*(4+1)/2=10
Numberofmeans:4
Total:14
Numberofparametersinouranalysis
2factorloadings+1factorvariance+3uniquevariances+1directeffect+
3itemintercepts+1exogenousvariance=12
SEM:MplusSyntax
TheMplussyntaxisacombinationofpathand
measurementmodels
SEM:ModelFitAssessment
Becausewehavefewerparametersthanthetotal
possible,wemustnowassessourmodelfit
SEM:ModelParameterOutput
Note:ourmeasurementmodelparametershavechanged
slightly(moreonwhyinamoment)
SEM:StandardizedModelParameters
Here,weseethattheSOGShasacorrelationof.577with
theGAMBLINGlatentvariable
SEM:ModelImpliedCovarianceMatrices
Noticeourmodelimpliedcovariancematrix:
Andtheresidualsfromthesaturatedmodel:
NEWWRINKLES:
MEASUREMENTMODELDOESNOTFITSATURATEDMODEL
PERFECTLY
OFFDIAGONALOFBLOCKCANCAUSEMODELMISFIT
EquationFormof
OverallStructuralEquationModel
Ourstructuralequationmodelsimultaneousequations
were,forthemeasurementportion:
s I
1
11 s1 s1
s I
3
31 s1 s3
s I
S
51 s1 s5
And,forthestructuralportion:
s1 0
GAMBLING
1
S0uS
s gs
p = S endogenousvariables
q = 1 exogenousvariable
MatrixFormofStructuralEquationModel
Inmatrices,themeasurementportionofthemodelisgiven
by:
s I s s s
with
s
Further,thestructuralportionofthemodelisgivenby:
s s s s
with
s
Intermsofourmodel:
Therearenodirectexogenouspredictorsofourendogenous
measurementmodelparameters(soK = )
Therearenodirectpredictorsofourendogenouslatentvariablesby
otherlatentvariables(soB = )
Westandardizedourfactormeantozero(sou = )
PuttingValuesintoMatrices:
MeasurementModel


itemintercepts
factorloadingsforendogenousvariables


uniquevariancesof
endogenousvariables
PuttingValuesintoMatrices:
StructuralModel
directregressioncoefficientofexogenous
variableontoendogenousfactor
residualvarianceofendogenousfactor
variance/covariancematrixforthe
exogenousvariables
ModelPredictedMeanVectorand
CovarianceMatrix
Thecovariancematrixoftheexogenousandendogenous
variablesisthen:
,x

1

1

1
Thepoint:thestructuralequationmodelcanhave
significantmodelmisfitduetoboththemeasurement
modelandthestructuralmodel
ERSH8750:Lecture#4 36
IssuesinBuildingStructuralEquationModels
BecauseofthemultiplewaysSEMscanexhibitmodelmisfit,the
processofbuildingSEMscanbedifficult
Ingeneral,currentpracticestatesthatmeasurementmodelsshould
bebuiltfirst thenthefullSEM
Someresearchersofferquestionableadvice:
Useonlyjustidentifiedmeasurementmodels
Why:fewerdegreesoffreedomwheremisfitcanhappen
Badidea:poorreliabilityforlatentconstructs
BuildmeasurementmodelswithSEMssimultaneously
Why:fullcalibrationcanleadtobetteroverallmodelfit
Badidea:measurementshouldhappeninabsenceofexogenousvariables
UsetwostageanalysesforSEMs
Why:measurementmodelthencannotchange
Badidea:propagationofmeasurementerrorforsomefactorscoremethods
SEMINPRACTICE:EXAMPLES
FROMREALWORLDANALYSES
SEMinPractice
Todemonstratethepracticalsideofbuildingstructural
equationmodels,Iwillgooveracoupleexamplesfrom
realdataanalyses
Intheseexamples,themodelbuildingprocesswillbe
discussed,alongwithvaryingmethodsforanalysis
Thedatafortheseexamplesisnotavailable butthe
practiceshouldshowhowdecisionsaremadeabouthow
SEMsareconstructedandinterpreted
Example#1:EvaluationofAcademicProgress
Thisexamplecomesfromdatafromalargesoutheastern
university
Datainclude:
PRE:scoresonapretestofmathematicsability,administered
tostudentswhentheyarriveattheuniversity
Scoresarefromtotalnumbercorrect alphareliabilityof.81
POST:scoresonaposttestofmathematicsability(usingthe
sameitems),administeredtostudentsaftertwoyearsatthe
university
Scoresarefromtotalnumbercorrect alphareliabilityof.81
CourseEnrollments:
Ifastudenthadenrolledinoneof29coursesrelatedtomathand
scienceeducationattheuniversity
Dataarebinary 0=didnotenroll;1=enrolled
Example#1:ResearchQuestions
Theevaluationsoughttoanswerthefollowingquestions:
Didscoresimproveontheposttestwhencomparedwiththe
pretest?
Didcourseworksignificantlyaffecttheposttestscores?
Didthescoreonthepretestpredictthecourseworkstudents
took?
Didcourseworkmediatetherelationshipbetweenpretestand
posttest?
BuildingtheSEM:ModelingIssues
Becauseofthenatureofthedata,severalmodelingissuesmustbe
consideredwhenusingSEMtoanswertheresearchquestions
Becausepretestandposttestaresumscores(withaknown
reliability),eachcanbeusedasasingleindicator
Inthiscase,theposttestsingleindicatorwillbeproblematicbecauseofthe
residualvariance(afterprediction)islessthantheoverallvariance
Somustputsingleindicatormodelinlast
Eachofthecoursesisbinary(dichotomous),soincludingthemin
themodeldirectlyisnotanoption
Modelwouldtreatthemasnormallydistributedifnotcategorical
Softwarewontallowcategoricalmediators
Couldusethemas:
Countsforspecificcategories(thentreatcountasapproximatelynormal)
Whatwedid
Indicatorsofacourseworkfactor
Hardtoenvision
ModelingStrategy
Courses:
Createcountsofeachcoursecategory(3categoriestotal)
Treatcountsasapproximatelynormal(anduseMLR)
Useallvariablesinapathmodelwhere:
Pretestpredictscoursecountsandposttestscore
Coursecountspredictposttestscore
Treatpretestandposttestassingleindicatorswherevariance
ofeachisweightedbythe.81reliabilityofeach
Finalstepintheanalysis
InitialSyntax:ForDescriptiveStatistics
InitialOutput:DescriptiveStatistics
Model#1:PathModelw/o
PosttestSingleIndicator
TheMplussyntax:
Modelfit:
Model#1:RelevantOutput
Forbuildingasingleindicatoroutofposttest:
Model#2:Pre/PostSingleIndicators
MplusSyntax:
Model#2:ModelFitAssessment
MplusOutput:
Normalizedresiduals:
Needforresidualcovariancesbetweencourseworksums
Model#3:SingleIndicatorswith
ResidualCovariances
Mplussyntax:
Note:thismodelhasnodegreesoffreedomleft itis
justidentified
Thereforemodelfitisperfect
Model#3:Results
Model#3Results
Model#3Results
Model#3Results
Model#3Results
Example#1:ResearchQuestionsAnswered
Theevaluationsoughttoanswerthefollowingquestions:
Didscoresimproveontheposttestwhencomparedwiththe
pretest?
Yes,posttestscoresimprovedby.654SDforeveryoneSDincreaseinthe
pretestscore(p<.001),holdingcourseworkconstant
Didcourseworksignificantlyaffecttheposttestscores?
No,nocourseworkwassignificantlyrelatedtotheposttest
Didthescoreonthepretestpredictthecourseworkstudentstook?
TheG1courseworkwassignificantlyreduced,with.101SDinnumberof
coursestakenforeverySDincreaseinthepretestscore(p=.030)
Didcourseworkmediatetherelationshipbetweenpretest
andposttest?
No,therewasnoindirecteffectofpretestonposttestasmediatedby
coursework(p=.687)
Example#2:DivorceeAdjustment
Thesecondexamplecomesfromarecentdissertation
titled:
RISKSANDBENEFITSOFSEEKINGANDRECEIVING
EMOTIONALSUPPORTDURINGTHEDIVORCEPROCESS:AN
EXAMINATIONOFDIVORCEEINDIVIDUALADJUSTMENT,
CLOSENESS,ANDRELATIONALSATISFACTIONWITH
MULTIPLEPARTNERSFROMASOCIALNETWORK
Thedissertationusedastructuralequationmodelto
investigatethepsychometricpropertiesofascaleand
theninvestigatedtheroleofthefactorsinpredicting
outcomesrelevanttotheresearch
Example#2:ResearchQuestions
Thefollowinghypotheseswereexamined:
H1:Initiatorstatusoffilingforthelegaldivorceisrelatedto
higheradjustmenttodivorce.
H2:Femaleshavehigheradjustmenttodivorce.
H3:Havingchildrenwiththeexspouseisrelatedtolowerlevels
ofadjustment.
H4:Thereisapositiverelationshipbetweenlengthoftimesince
thedivorceandadjustmenttodivorce.
H5:Thereisanegativerelationshipbetweenageandadjustment
todivorce.
ScalesinAnalysis
Fivescaleswereincludedintheanalysis(N=229)
MLRestimationwasused
Standardizedfactorswereestimated
AdjustmenttoDivorceScale:
10items
RMSEA=.07;CFI=.93
Alphareliability=.88
AnalysisofRiskinSeekingSupportScale
RiskinSeekingSupportScale:
6itemsinitially
Target exspouse:RMSEA=.12;CFI=.85
Target familymember:RMSEA=.06;CFI=.96
Target friend:RMSEA=.09;CFI=.91
Oneitemremoved:
Target exspouse:RMSEA=.13;CFI=.87;Alpha=.69
Target familymember:RMSEA=.00;CFI=1.00;Alpha=.81
Target friend:RMSEA=.08;CFI=.95;Alpha=.75
AnalysisofEmotionalSupportProvision
EffectivenessScale
EmotionalSupportProvisionEffectivenessScale
13items
Target exspouse:RMSEA=.12;CFI=.86
Target familymember:RMSEA=.11;CFI=.87
Target friend:RMSEA=.07;CFI=.92
Removed3reversecodeditems:
10items
Target exspouse:RMSEA=.09;CFI=.92;Alpha=.95
Target familymember:RMSEA=.11;CFI=.91;Alpha=.95
Target friend:RMSEA=.06;CFI=.96;Alpha=.92
AnalysisofClosenessScale
ClosenessScale:
18items
Target exspouse:RMSEA=.11;CFI=.78
Target familymember:RMSEA=.09;CFI=.81
Target friend:RMSEA=.08;CFI=.86
Removed7reversecodeditems
11items
Target exspouse:RMSEA=.09;CFI=.88;Alpha=.89
Target familymember:RMSEA=.06;CFI=.95;Alpha=.90
Target friend:RMSEA=.06;CFI=.95;Alpha=.90
AnalysisofRelationalSatisfactionScale
Relationalsatisfactionscale:
7items
Target exspouse:RMSEA=.07;CFI=.96;Alpha=.83
Target familymember:RMSEA=.05;CFI=.99;Alpha=.92
Target friend:RMSEA=.05;CFI=.98;Alpha=.88
InvarianceTesting
Becausemanyofthescalesaskedaboutspecifictargets,
thesewereusedasmultiplegroupstoinspectfactorial
invariance
Modelspreviouslyreportedrepresentedtheconfigural models
RiskinSeekingSupportScale:
Metricinvarianceheld
Partialscalarinvariance(3itemswithdifferentintercepts)
Partialresidualinvariance(2itemswithdifferentunique
variances)
InvarianceTesting
Emotionalsupportprovisioneffectivenessscale
Partialmetricinvariance(5itemsnotinvariant)
Partialscalarinvariance(5itemsnotinvariant)
Closenessscale
Partialmetricinvariance(1itemnotinvariant)
Partialscalarinvariance(5itemsnotinvariant)
RelationalSatisfactionScale
Partialmetricinvariance(3itemsnotinvariant)
Partialscalarinvariance(3itemsnotinvariant)
SubsequentPathAnalysis
FinalPathAnalysis
CONCLUDINGREMARKS
WrappingUp
Todaywasaboutputtingitalltogether:pathanalysisand
measurementmodels
TheSEMframeworkallowsforpowerfulinferential
analysestobeconductedinastatisticallyrigorousmanner
Butwiththepowercomesalotoffrustration datadonot
alwayscooperate
Youwillfindthatpeopletakegreatlibertieswithhowthey
conductSEManalyses
Ihopethisclasshelpsyouunderstandhowpeopledo
SEMandhowtodoSEMyourself
UpNext
Todayisourlastclassmeeting
Nofinal;Nofinalhomework
Extracreditopportunity:
Filloutcoursesurveyonline
https://portal.coe.uga.edu/apps/authorize/
EnsureyougetanAinthecourseItakethecomments
veryseriously!

Anda mungkin juga menyukai