Beruflich Dokumente
Kultur Dokumente
Abstract
TheCALISprocedureinSAS/STATisageneralstructuralequation modeling(SEM)tool.
ThisworkshopintroducesthegeneralmethodologyofSEMandtheapplicationsofthe
CALISprocedure.Historicaltopicssuchascasualmodels,pathdiagram,confirmatory
factoranalysis,measurementerrormodel,andlinearstructuralrelations(LISREL)are
reviewed.ApplicationsoftheCALISproceduretoSEMaredemonstratedwithexamples
insocial,educational,behavioral,andmarketingresearch.Specifically,thefollowing
howtotechniquesoftheCALISprocedure(SAS/STAT9.22)arecovered:(1)Specifying
structuralequationmodelswithlatentvariablesbyusingthePATHmodelinglanguage;
(2)Interpretingthemodelfitstatisticsandestimationresults;(3)Testingmodelswith
multiplegroupsandmultiplemodels;(4)Analyzingdirectandindirecteffects;(5)
Modifyingstructuralequationmodels.
Thisworkshopisdesignedforstatisticiansanddataanalystswhowanttooverviewthe
applicationsoftheSEMbytheCALISprocedure.Attendeesshould haveabasic
understandingofregressionanalysisandexperienceusingtheSASlanguage.Previous
exposuretoSEMisuseful,butnotrequired.
Citationofthisworkshop:
Yung,Y.F.(2010).IntroductiontoStructuralEquationModelingUsingtheCALIS
ProcedureinSAS/STAT Software.Computertechnologyworkshoppresentedatthe
JointStatisticalMeetingonAugust4,2010,Vancouver,Canada.
Inthisworkshop,SAS/STAT9.22(TS2M3)orlaterisassumedfortheCALISprocedure.
SomeofthecodemightworkwithPROCTCALIS(anexperimentalprocedure)in
SAS/STAT9.2(TS2M2).However,thereisamajorsyntacticaldifferencebetweenPROC
TCALISandPROCCALIS.InPROCTCALIS,theparameterspecificationforeachpathin
thePATHstatementmustnot beprecededbyanequalsign.Butthisequalsignis
requiredinPROCCALISwhenyouspecifyparameters.Also,PROCTCALISdoesnot
supportthegeneralizedpathspecifications(forvariances,covariances,means,and
intercepts)andmultiplepathspecificationswhenyouusethePATHmodelinglanguage,
whichisthemainfocusoftodaystalk.
Path diagram
b
ey
Thecentralideaofstructuralequationmodelingisthestudyof causalrelationship
betweenvariables.Forexample,youhaveanXandanYvariable. XisthecauseofY,or
doingXresultsinY.Togiveamorerealisticexample:eatingmorevegetables(X)brings
downyourcholesterollevel(Y).However,thiscausalstructureisonlyanidealized
framework.Inmakingcausalinferences,youmusthaveisolatedallotherbackground
variablesandestablishtemporalsequenceofthevariables.Becauseofthecomplicated
philosophicalissuesinvolvedinmakingcausalinferences,ingeneralSEMwouldavoid
claimingcausalinferences.
Apredictoroutcomeframeworkmightbemoreappropriatephilosophically.The
semanticisnow XpredictsY.Mathematicallyandstatistically,thisideaisrepresented
inthesimplelinearregressionanalysis,asshowninthelinear regressionequation:
Y=b*X+e.
Thepathdiagramforthisrepresentationisshownintheslide,wherebiscalledthe
effect,regressioncoefficient,orpathcoefficient.Noticethatanerrortermisaddedto
showthatthepredictionofYfromXisnotperfect.Butessentially,thepredictor
outcomeframeworkisrepresentedbytheYXpathinthepathdiagram.
More variables
More equations
LV
Correlated errors
Group 2
Y
Multiple-group analysis
LV
Z
W
4
WhatarethedifferencesbetweenSEMandregressionanalysis?WhatmorecanSEM
offerthanthelinearregressionanalysis?
YoucanviewSEMasamuchmorecomplicatedsystemformultiplepredictoroutcome
relationships.SEMcanhandlethefollowingsituationswherelinearregressionanalysis
isoflimitedusefulness:
1.Morevariables(notjustXandY,butyoucanalsoaddWandZintothepathdiagram).
2.Moreequationsorfunctionalrelationships(notjustXY,butyoucanalsoanalyze
WZsimultaneously).
3.Correlatederrors,systemofequationscanhavecorrelatederrors.Forexample,the
doubleheadedarrowbetweenYandZ.
4.Directandindirecteffects:XhasadirecteffectonZandanindirecteffectonZviaits
effectonW.Thatis,XZandXWZaredirectandindirecteffects,respectively.
5.Latentvariables.Forexample,LVinthepathdiagramhaseffectsonXandW.
6.Parametricconstraints.Forexample,theconstraintsonthepath coefficientsor
effectslabeledasa intheupperpathdiagrams.
7.Multiplegroupanalysis.Fordifferentgroupsofpopulations,theoverall structureof
themodelarethesame,butthepathconstraintscouldbedifferentwhilethe
constrainedeffectinGroup1isdenotedasa, theconstrainedeffectinGroup2is
denotedasb, whichwillhaveadifferentestimatethanthatfora inGroup1.
SEMhasalotofsynonymsinthefield:Pathanalysis(attributedtoSewallWright),
LISRELmodel(JKWmodel),covariancestructuresanalysis,analysisofmoment
structures,confirmatoryfactoranalysis,causalmodeling,andetc.Intermsofthe
statisticalmethodologyinvolved,allthesenamesaremoreorlessthesame.
PROCCALIS,whichstandsforcovarianceanalysisoflinearstructuralequations,isa
softwarethatwasdesignedtohandlealltheseanalysesundertheumbrellatermSEM.
Hopefully,onedayPROCCALISwouldalsoberememberedasasynonymofSEM.
SEM Software
AMOS, EQS, LISREL, MPLUS,
Why PROC CALIS?
Which is best?
ThereareseveralwellknownsoftwareinthefieldfordoingSEM:AMOS,EQS,LISREL,
andMPLUS,andmaybemore.WhyPROCCALIS?Whichisbest?Althoughtheseare
veryinterestingquestions,asthecurrentdeveloperofPROCCALISprocedureIamnot
atlibertytojudgeotherSEMsoftware.Thisworkshopgivesyouanintroductorytourof
SEMwiththeuseofPROCCALIS.Therefore,youmightcomparePROCCALISwithother
softwareonyourownafterlearningsomeofthefeaturesofPROC CALIS.
LetusstartwithabriefhistoryofPROCCALIS.
Ineighties,WolfgangHartmanndesignedanddevelopedthefirstversionofPROCCALIS.
Thestatisticalandmathematicalmodelwasgreatlyinfluencedby theCOSANmodel
proposedbyR.P.McDonald.Infact,therewasevidencethatCosan,insteadofCalis,
mighthavebeenproposedasthenameoftheprocedure.Themostpopularsyntaxin
PROCCALIS,however,wasundertheinfluenceoftheEQSprogrambyPeterBentler.
TheLINEQSsyntaxinPROCCALISformodelspecificationisbasicallyatwinbrotherof
thesyntaxoftheEQSprogram.
Ipickedupthedevelopmentofthesoftwarearound2000.Iactuallyrewrotethe
mathematicalfoundationsofthesoftware.Ikepttheoptimizationtechniquesandinitial
estimationtechniquessothatthenewCALISiscompatiblewiththeoldCALIS.
In2008,anexperimentalversioncalledTCALISwasreleased.Sincethen,Ihavemodified
thesyntaxalittlemoreandfixedsomemajorbugs.
ThenewCALIS(SAS9.22)hasbeenreleasedthisyear.IfyouhaveusedPROCCALIS
before,youwillnoticeonemajorchange:theemphasisonthePATHmodelinglanguage.
YoucanseeexamplesusingthePATHstatementeverywhereinthePROCCALIS
documentation.Othernoteworthynewfeaturesare:multiplegroupmodeling,
redesignedmeanstructureanalysis,andthenamefreeapproachtoparameter
specifications.Certainly,therearemanymorenewfeaturesthan these,asyouwilllearn
fromthisworkshopandelsewhere.
2.
3.
4.
Multiple-group analysis
2.
3.
4.
Model modifications
ThefirstpartoftheworkshopisaboutthebasicSEMmodelingusingPROCCALIS.Iwill
describetheresearchprocessofSEMbriefly.ThenIwillintroducethePATHmodeling
languageinPROCCALISbyusingasimplelinearregressionexample.Next,Iwillmove
ontomorecomplicatedexamplesthatanalyzeconfirmatoryfactormodels.Iwilluse
PROCCALISintheseexamplestoshowhowyoucanspecifySEMmodelsbythePATH
modelinglanguage,inrelationtothepathdiagramrepresentations.Iwillshowyouhow
tointerprettheresultsgeneratedbyPROCCALIS.Iwillendthe firstpartbyshowingyou
howaLISRELmodelcanbespecifiedbytheLISMODstatementinPROCCALIS.
Thesecondpartoftheworkshopisaboutadvanced modelingrelativelyspeaking.I
willshowhowmultiplegroupanalysiscanbedoneinPROCCALIS.Otherimportant
topicssuchasdirectandindirecteffectanalysis,testingspecifichypotheses,andmodel
modificationsarediscussed.
Therearetwoemphasesofthistalk.
One,IwanttoshowyouanoverallpictureofSEM.ThisaddressesthewhatisSEM?
question.Iwillnotgiveyouatechnicaldefinition,butIwill showyouSEMexamplesso
thatyouwillhaveareal feelingabouttheapplicationsofSEM.
Two,IwanttoshowyouhowtousePROCCALIS.ThisaddressestheHowtodoSEM?
question.Ihopethatintheendoftheworkshop,youwillfindthatPROCCALISisvery
usefulforSEM.
10
10
Start Use
.19
.38
Skill
Positive
Affect
Control
.46
Time Use
-.32
.42
Interactive
Speed
.19
.31
Playfulness
.33
Challenge
.24
.51
.14
Focused
Attention
Future Use
Exploratory
Behavior
Arousal
.70
.51
.19
.45
.47
.25
Importance
.78
.67
.24
Time
Distortion
.13
Stimulation
Level
11
Thisisastructuralequationmodelaboutwebsurfingbehavior.Theresearchers
hypothesizethatthePlayfulness ofawebsitewouldenhancethefutureuse(Future
Use)ofthesamewebsite.However,thetheorydoesnotendthere.Theresearchers
thenhypothesizewhatwouldmakeawebsitetobeperceivedasplayful.Three
additionalconstructsarehypothesizedinthepathdiagram:Control (ofthewebpage),
Arousal (ofinterest),andFocusedAttention aredeterminantsofPlayfulness. In
fact,theresearchershypothesizedevenfurther.Forexample,theyuseStartUse
(whentheusersstartedtousecomputers)andTimeUse (howoftentheyuse
computers)asremotecauses ofalotoflatentconstructsinthepathdiagram.Insum,
thisisarelativelylargeSEMthattheorizescomplicatedrelationshipsamongconstructs.
Inthispathdiagram,theovalshapesrepresentlatentvariables,whicharenotmeasured
butserveasusefulconstructsinthemodel(e.g.,Playfulness).Therectanglesrepresent
measuredorobservedvariables(e.g.,StartUse,TimeUse,FutureUse).Inorderto
analyzethelatentconstructs,somemeasuredvariables(orindicators)areneeded.Inthe
pathdiagram,thosesmallunlabeledrectanglesaremeasuredindicatorsforthelatent
constructs.Inthisresearch,thesemeasuredindicatorsareratingresponsesona
questionnaire.Forexample,Ilosttrackofthetimewhenusingthiswebsite (thisis
notanexactitemfromtheactualresearch)couldbeanitemfor theTimeDistortion
construct.
Giventhispathdiagramforthetheoryaboutwebsurfingbehavior,anSEMsoftwarefits
themodelbasedontheobserveddataandinformsyouthemodelfitandtheestimates
oftheeffects(pathcoefficients)inthepathdiagram.TheSEMsoftwarealsotellsyouthe
significanceoftheseestimates.Ifthemodeldoesnotfitthedatawell,theSEMsoftware
wouldsuggestwaystoimprovethemodel.
11
12
HereisalistofthekeyfeaturesofSEM:
Analyzingcomplicatedrelationshipsamongvariables
Pathdiagramrepresentationsformodels
Abilitytohandlelatentandobservedvariablessimultaneously
Testingthemodelfitandsignificanceoftheparameters
Suggestingwaystoimprovethemodel
12
13
13
y = b x + ey
y: outcome variable
x: predictor variable
ey: error term
b: effect or regression coefficient
14
TointroducethePATHmodelinglanguageinPROCCALIS,asimple linearregression
modelisused.Intheregressionequation,yistheoutcomevariable,xisthepredictor
variable,e_y istheerrortermandbistheeffectorregressioncoefficient.The
regressionmodelwritteninthisformassumesthatxandyarecenteredwithmeans
zero.Butthisassumptionwillnotaffectthegeneralitytouncenteredvariables.
14
15
Onp.34ofFullersbookMeasurementErrorModels,hedescribesadatasetabout
thecountingofhenpheasantsinAprilandAugust.Fifteentrainedobserverscounted
thenumberofbirdsinthetwooccasions.YisthenumberofbirdsinAugustandXisthe
numberofbirdsinApril.Thegoalofthelinearregressionistopredictthenumberof
birdsinAugust(Fall)bythenumberofbirdsinApril(Spring).
15
data hens;
input y x @@;
datalines;
8
6.6
9.8
12.3
10.8
11.9
9.7
11.9
9.3
12
9.2
9.6
6.9
7.5
8.1
10.9
8.7
10.4
8.7
10.2
7.4
7.4
10.1
11
10
11.8
7.3
8.2
;
proc reg data=hens;
model y = x;
run;
16
Toconductalinearregressionanalysis,youcanuseaSASprocedurecalledPROCREG.
Thesyntaxisquitesimple.First,defineyourdataset.Second, callPROCREGwiththe
interesteddatasetspecifiedinthePROCREGstatement.Then,themodelstatement
specifiesthaty=x,whichmeansyispredictedbyx.Noerrortermneedstobespecified,
althoughPROCREGdoesassumethatpredictionisnotperfectsothattheerrordoes
existwithnonzerovarianceintheregression.
16
Parameter Estimates
Parameter
Standard
DF
Estimate
Error
t Value
Intercept
2.14227
0.84513
2.53
0.64941
0.08275
7.85
Variable
Given a base survival of 2.14 birds, every additional bird in Spring predicts a
0.65 bird surviving in August.
17
ThistableshowstheessentialresultsfromPROCREG.Theoutput showsanestimateof
0.65fortheregressioncoefficientb.Theinterceptestimateis2.14.PROCREGalso
showsthestandarderrorestimatesandthetvaluesforjudgingstatisticalsignificance.
Bothestimatesarestatisticallysignificant.
Aninterpretationabouttheseregressionestimatesisthis:Givenabasesurvivalof2.14
birds,everyadditionalbirdinSpringpredictsa0.65birdsurvivinginAugust(Fall).
17
Outcome
Path
Statement
Parameter
<---
ey
Predictor
PATH
b x
Path Relation
b;
Parameter (optional)
18
Asshownpreviously,youcanrepresentthelinearregressionmodelbythepathdiagram,
whichisalsoarepresentationschemeforSEM.Hence,regression modelscouldbe
specifiedasSEMbyusingthepathdiagram.
HereiswhatyoudotospecifyasimplelinearregressionmodelinPROCCALIS.Youuse
thePATHstatementtospecifythepathintheregressionmodel.Inthiscase,itisjustY<
XinthePATHstatement.Optionally,youcandenotethecorrespondingpath
coefficientparameter.Forexample,youcanput=b atthebackofthepathtodenote
theparameterlabelorname.
18
19
ThisistheentirePROCCALISsyntaxforthesimplelinearregressionmodel.Isntthat
easyandsimple?
19
<---
Parameter
_Parm1
Estimate
Error
t Value
0.64941
0.07974
8.14445
Variance Parameters
Default variance
parameters
Variance
Type
Variable
Parameter
Exogenous
Error
Standard
Estimate
Error
t Value
_Add1
3.62124
1.36870
2.64575
_Add2
0.32233
0.12183
2.64575
20
ThisslideshowstheresultsfromPROCCALIS.
Theestimatedeffectofxony,denotedasy< xintheoutput,is0.65,whichisthe
sameasthatinthePROCREGresults.Becauseyoudidnotnamethisregression
coefficientparameter(butyouspecifythepathnonetheless),PROCCALISgeneratesa
uniqueparameternamecalled_Parm1forit.Thestandarderrorestimateandthet
valuearealittlebitdifferentfromthatofthePROCREGresults.Thisisbecause
differentdegreesoffreedomforcomputingthestandarderrorsareusedinthetwo
approaches.
InPROCCALIS,italsoincludesresultsfortwomoreparametersintheSEM.Thevariance
ofxandtheerrorvarianceofyaretreatedasmodelparameters.Theirestimatesare
alsoshowninthePROCCALISresults.NotethatPROCCALIScreatesdefaultparameter
namesforthesedefaultvarianceseventhoughyoudidnotspecifythem.Inthis
example,thesevarianceparametersarenamed_Add1 and_Add2,respectively.
20
errv_y
y <--- x
= b;
b
y
var_x
pvar
x
= var_x,
= errv_y;
Equivalent representations
run;
b
y
errv_y
var_x
ey
With an explicit error
term
21
YoucouldnamealltheparametersinPROCCALISbyputtingyourpreferrednames.
Inthepathdiagramatthetoprightcorner,theparametersareshowninred.Inthe
regressionmodel,bistheregressioncoefficient,var_x isthevarianceforthepredictor
variablex,anderrv_y istheerrorvarianceofy.Thispathdiagramrepresentationis
equivalenttotheoneshownatthebottomrightcorner,wherean expliciterrortermis
attachedtoY.Theerrortermisrepresentedbyanovalshapebecauseitistreatedasa
latentvariable.Thisrepresentationhasthesamesetofparameters,onlythaterrv_y is
nowattachedtotheerrorvariabledirectly.
YoucanspecifytheseparametersexplicitlyinPROCCALIS.Intheleftpaneloftheslide,
theparameterbisspecifiedafterthey< xpath,separatedbyanequalsign.To
specifythevariancesorerrorvariancesinthemodel,youcanusethePVARstatement.
Forexample,x=var_x meansthatthevarianceofxisaparametercalledvar_x.
Noticethatnamingparametersisentirelyoptional.Forthisexample,naming
parametersappearstoserveonlyasanillustration.Laterinthistalk,youwillfind
situationswheretheuseofparameternamesisnotonlyuseful,butalsonecessary.
21
Parameter
<---
Estimate
Error
t Value
0.64941
0.07974
8.14445
Variance Parameters
Variance
Standard
Type
Variable
Parameter
Estimate
Error
t Value
Exogenous
Error
var_x
3.62124
1.36870
2.64575
errv_y
0.32233
0.12183
2.64575
Asshowninthisslide,thenumericalresultsfromPROCCALISwithexplicitparameter
namesspecifiedarethesameasthosewithoutusingparameternames.Theonly
differenceisthatnowyoucanusetheseparameternamestolocatethecorresponding
resultsdirectly.
22
23
Sofar,Ihaveshownyouthat:
1.ThePATHmodelinglanguageisaseasyasdrawingapathdiagram.
2.YoucanusethePATHstatementtospecifythepathsinpathdiagram,withorwithout
specifytheparameternamesforthepathcoefficients.
3.Youcanalsospecifythevarianceorerrorvarianceparametersexplicitly.Inmost
practicalapplications,variancesanderrorvarianceshavealreadybeensetbydefault
andyoudonotneedtoworryaboutspecifyingthem.TheessentialpartofSEMis
specifiedinthePATHstatement.
4.NamingparametersisoptionalinPROCCALIS.
23
24
LetusmakealittlestepforwardtoshowaspecialSEMfeaturethatlinearregression
cannothandleeasily.
Inthebirdcountingexample,wedidnottakeintoaccountthatbirdcountingcould
involvemeasurementerrors.Inthecurrentcontext,themeasurementerrorinbird
countingcouldbeduetotheenvironmentfactorsintheforest:obstructionfromthe
treebranches,biased anglesfromthebirdobservers,andetc.
Mathematically,youcanhypothesizeavariablecalledfx torepresentthetrue counts
obtainedfromthebirdobserved.Theobservednumberofbirdsxisthesumoffx,the
truescore,andex,anerrorterm.
Whatyougotfromthedataisx,theobservedfalliblescore.However,ideally,you
wouldwanttousefx,thetruescoreinyourregressionanalysis.
24
25
TheprecedingideaisformalizedasthefollowingSEMwithalatentvariablefx.
Inthesocalledstructuralmodel,yispredictedfromfx,thetruescore,inthelinear
regressionmodel.Thissocalledstructuralequationtakestheroleoftheoriginallinear
regressionequationonlynowyouaresupposedtohaveabettermodelbyusingthe
measurementerrorfreefx asthepredictor.
Inthesocalledmeasurementmodel,youhypothesizethattheobservedvariablexis
obtainedasthesumoffx andanmeasurementerrorterme_x.
Canyouestimatebwiththelatentvariablefx inthestructuralequation?
Thisanswerisyes.
Butthetechnicalproblemencounteredhereisthatthemeasurementequation
introducesoneadditionalparametersinvar(ex)errorvarianceofx.Thisproblemwill
maketheSEMunidentified.Inaveryloosesense,thismeansthatyourmodelestimates
moreparametersthanwouldbeallowedbythegiveninformationofthedataset.
Consequently,theparametersinthemodelarenotestimable.Iwilldescribeamethod
todealwiththisidentificationproblemlater.
25
fx
errv_y
var_x
errv_y
var_x
errv_x
One more
parameter
26
Thisslidecomparesthelinearregressionmodelwiththemeasurementerrormodelby
theuseofpathdiagram.Itdemonstrateswhythemeasurementerrormodelhasone
moreparametertoestimate.
Intheleftpanel,thepathdiagramforthesimplelinearregressionanalysisisshown.
Intherightpanelforthemeasurementerrormodel,westillhavexandyasthe
observedvariables.Butnowwehavealatentvariablefx thattakestheroleofthe
predictorofy.Var_x inthismodelnowrepresentsthetruevarianceofthepredictor fx.
Thenewparameterinthemeasurementerrormodeliserrv_x (errorvarianceofx).With
thisadditionalparameter,weneedtomakeadditionalassumption toestimatethe
modelparameters.
26
errv_x = 6* errv_y
27
Fortunately,wehaveareasonableassumptionabouttherelativesizeoftheerror
variancesinthemodel.
ThisassumptionisbasedonthefactthatbirdcountinginFallismoreaccuratethanthat
inspring.ThereasonisthatthefallenleavesinFallmakesthecountingofbirdsless
obstructive.
Theassumptionwearegoingtomakeisbasedonanindependentstudyaboutthe
relativeerrorvariancesinxandiny.InFullersbook,theratioofthesevariancesis
about6.Mathematically,Var(ex)=6*Var(ey).Thatis,errorvarianceforxissixtimesas
muchastheerrorvarianceofy.Or,inthePROCCALISspecification,youwanttostate
thefollowingparametricconstraintinthemodeling:errv_x=6*errv_y.
27
<--- fx
fx ---> x
= b,
= 1;
pvar
y
fx
x
b
y
fx
errv_y
var_x
= errv_y,
= var_x,
= errv_x;
errv_x = 6 * errv_y;
errv_x
errv_x = 6* errv_y
run;
The required constraint is specified
as a SAS programming statement.
28
ItturnsoutthatitisprettystraightforwardtospecifythisparametricconstraintinPROC
CALIS.Youjustsimplyaddonemorelineofcodetorepresentthisrelationship,as
shownintheSAScodeoftheslide.IntheSASliterature,thislineofcodeiscalledaSAS
programmingstatement,whichisusedextensivelyintheDATAstepofSAS.Youcanuse
asmanySASprogrammingstatementsasyouwanttodescribetherelationshipsofthe
parametersinthemodel.
28
Standard
--------Path--------
Parameter
<---
fx
fx
--->
Estimate
Error
t Value
0.75158
0.09228
8.14427
1.00000
A larger estimated effect than the one
estimated without taking the
measurement error into account (0.649)
Variance Parameters
Variance
Standard
Type
Variable
Parameter
Estimate
Error
t Value
Error
errv_y
0.08205
0.03101
2.64575
Exogenous
Error
fx
var_x
3.12893
1.36180
2.29765
errv_x
0.49231
0.18608
2.64575
29
Afteryoutakethemeasurementerrorintoaccount,theregressioncoefficientbisnow
0.75,whichisalargereffectthan0.649,whichyouobtainedfromthelinearregression
modelwithouttakingthemeasurementerrorinxintoaccount.Therefore,theprevious
regressionanalysisunderestimatedthiseffectbecauseitfailed toincorporatethe
measurementerrorintothemodel.However,withSEM,youcaneasilyincorporatethe
measurementerrorsintotheanalysis.
Estimatesofthevariancesanderrorvariancesareshowninthenexttable.Youcansee
thattheconstraintspecifiedinthePROCCALISishonoredintheestimation.Theerror
varianceestimateofxis0.49,whichissixtimesasmuchastheerrorvarianceestimate
ofy,whichis0.08.
29
30
ThisslidesummarizessomefeaturesofthePATHmodelinglanguage.
1.Itisasstraightforwardasdrawingthepaths.
2.Itcandealwithlatentvariableseasily.
3.YoucanusethePVARstatementtospecifyvariancesorerrorvariances(double
headedarrowsattachedtoindividualvariablesinthepathdiagram).
4.YoucanusethePCOVstatementtospecifycovariances orerrorcovariances (double
headedarrowsattachedtopairsofvariablesinthepathdiagram).
5.YoucanspecifyparameterdependencybyusingtheSASprogrammingstatements
directly.Indeed,evenverystrangeandcomplicated(continuous) parametricfunctions
aresupportedinPROCCALIS.
30
31
Wenowmoveontoamorecomplicatedtypeofstructuralequation modelscalled
confirmatoryfactormodels.
31
32
ThisexampleisbasedonanexampleinChapter7ofBollens classictextbook:Structural
EquationModeling.
Inthisexample,twolatentfactorsformeasuringpoliticaldemocracyin75developing
countriesin1960and1965werehypothesized.
Thesetwolatentfactorsarenotobserved,buttheyhavesomerelatedobserved
variablesthatserveasindicators.Ineachyear,youmeasurefourvariablestogaugethe
politicaldemocracy:freedomofpress,freedomofgroupoppositions,fairnessof
elections,andelectivenatureofthelegislativebody.
Thepurposeoftheconfirmatoryfactoranalysisistovalidatethesemeasurement
indicatorsstatistically.
32
Press60
Freop60
Dem60
Fair60
Legis60
Press65
Freop65
Dem65
Fair65
1.
Legis65
33
Thispathdiagramshowsthehypothesizedconfirmatoryfactormodel.
Inthepathdiagram,twolatentfactorsarerepresentedbytwoovals.Dem60isthe
politicaldemocracyin1960andDem65isthepoliticaldemocracy in1965.Theyare
linkedtotherespectivemeasuredvariables,asshowninthepathdiagram.Thesesingle
headedpathsrepresentthetypicalfactorobservedvariablerelationships.
ThedoubleheadedarrowthatconnectsDem60andDem65representsthecovariance
parameterbetweenthetwofactors.Itmeansthatthetwofactorsarecorrelated.
DoubleheadedarrowsthatareattachedtoDem60andDem65individuallyrepresent
thevarianceparametersofthetwofactors.Inthemodel,youfixthesevariancesto1so
thatthescalesofthefactorsareidentified.Thisisconventionallydonebecausethe
scaleoflatentfactorsisarbitrary(youdonotmeasurelatentvariablesdirectlysothat
theycouldbedefinedonanyunitofmeasurement).
Thedoubleheadedarrowsthatareattachedtotheobservedvariablesrepresentthe
errorvariances.Theysignifythefactthatthefactorsinthemodeldonotaccountfor
100%ofthevariancesoftheobservedvariables.Theerrorvariancesaretheunique
partofthevariancesinthevariablesthatarenotduetotheir relationshipswiththe
factorsinthemodel.
33
Press60
Freop60
Dem60
Fair60
Legis60
Press65
Freop65
Dem65
Fair65
1.
Legis65
1.
1.
1.
1.
1.
1.
1.
1.
e1
e2
e3
e4
e5
e6
e7
e8
34
Thisisanalternativepathdiagramrepresentationwiththeuseofexpliciterrorterms.
Noticethatthedoubleheadedarrowsfortheobservedvariablesnowshifttotheerror
terms.Thispathdiagramrepresentationisshownhereonlyforillustrationpurposes.In
thisworkshop,Irelyonthepathdiagramrepresentationthatdoesnotuseexpliciterror
terms.
34
1.
Press60
Freop60
Dem60
Fair60
Legis60
Press65
Freop65
Dem65
Fair65
1.
Legis65
35
SpecifyingtheCFAmodelisnotmuchharderthanthepreviousmeasurementerror
model.Basically,youonlyneedtospecifymorepathsfortheCFAmodel.
InthePATHstatement,youspecifyallthesingleheadedpaths(arrows)inthepath
diagram.
InthePVARstatement,youspecifyalldoubleheadedarrowsthatareattachedto
individualvariables.PVARactuallystandsforpartialvarianceyoucanspecifythe
variancesanderrorvariancesinthisstatement.Dem60=1 meansthatthevarianceof
Dem60isfixedtoone.SimilarlyforDem65=1.Theeightobservedvariableare
specifiedinthePVARstatementtosignifythattheirerrorvariancesarefreeparameters
inthemodel.
InthePCOVstatement,youspecifypairsofvariablesthathavecovariances orerror
covariances asparametersinthemodel.Inthecurrentpathdiagram,Dem60and
Dem65arecorrelated.
35
36
Tomakemodelspecificationmoreefficientanderrorfree,PROCCALISemploysdefault
freeparametersinthemodel.Thesedefaultfreeparametersaresetbecausetheyare
commonlyemployedinpractice.
Forexample,becausepredictionsofoutcomevariablesareusuallynotperfect,theerror
variancesarefreeparametersbydefault.ThismeansthatallthePVARspecificationsfor
theobservedvariablesarenotnecessarybecausePROCCALISwouldhavetreatedthem
asfreeparametersbydefault.
Similarly,thevariancesofDem60andDem65andtheircovariance aredefaultfree
parametersbecausetheyareassumedinmostpracticalapplications.Inthecurrent
example,thismeansthatthePCOVstatementspecificationforthecovariancebetween
Dem60andDem65isnotnecessary.
However,becausethevariancesofDem60andDem65arefixedto1 (foridentification
ofthelatentvariablescales),theymustbespecifiedexplicitlyinthePVARstatement.
Otherwise,thesevarianceswouldhavebeenfreeparametersbydefault.
36
Standard
----------Path----------
Parameter
Estimate
Error
t Value
Dem60
--->
Press60
Dem60
--->
Freop60
_Parm1
2.20567
0.25122
8.77998
_Parm2
3.00132
0.39735
7.55335
Dem60
--->
Dem60
--->
Fair60
_Parm3
2.31033
0.34026
6.78989
Legis60
_Parm4
2.89483
0.31582
Dem65
9.16619
--->
Press65
_Parm5
2.04790
0.25930
7.89771
Dem65
--->
Freop65
_Parm6
2.68003
0.33258
8.05834
Dem65
--->
Fair65
_Parm7
2.70879
0.31804
8.51711
Dem65
--->
Legis65
_Parm8
2.76604
0.30830
8.97190
ThistableshowstheestimatesofpathcoefficientsfromPROCCALIS.
Inthefactoranalysisliterature,thesepathcoefficientsarealsocalledloadings.To
validatetherelationshipsbetweenthedemocracyfactorsandthe observedvariables,
thetvaluesmustbeexaminedforstatisticalsignificance.Usingnormalapproximation,
tvalueswiththeirabsolutevaluesbiggerthan1.96aresignificantlydifferentfromzero.
Inatypicalfactoranalysisstudy,youwouldwantallthesetvaluestobesignificantin
ordertoclaimnonzerofactorvariablerelationships.Aninsignificanttvaluemeansthat
thecorrespondingvariableisnotanindicatorforthepurported factor.Insignificantt
valuesforpathcoefficientswouldchallengethevalidityofyourfactormodel.
37
Variance
Standard
Type
Variable
Exogenous
Dem60
1.00000
Dem65
1.00000
Error
Parameter
Estimate
Error
t Value
Press60
_Add1
2.01359
0.41048
4.90549
Freop60
_Add2
6.57189
1.20964
5.43294
Fair60
_Add3
5.42661
0.96546
5.62076
Legis60
_Add4
2.83887
0.61417
4.62229
Press65
_Add5
2.63180
0.49311
5.33709
Freop65
_Add6
4.19276
0.79422
5.27906
Fair65
_Add7
3.46180
0.68155
5.07928
Legis65
_Add8
2.88292
0.59927
4.81068
38
Estimatesofvariancesanderrorvariancesareshowninthistable.Thevariancesof
Dem60andDem65arefixedto1andthereforetherearenosignificancetestsforthese
variances.Allothererrorvarianceestimatesaresignificantlylargerthanzeros.Thisalso
meansthatthefactorsdonotaccountforallthevariancesoftheobservedvariables.
Thisisnaturalbecausedeterministicrelationshipsbetweenfactorsandobserved
variablesarerare.
38
Standard
Var1
Var2
Parameter
Dem65
Dem60
_Add9
Estimate
Error
t Value
0.97528
0.02656
36.72321
High and significant correlation between the Democracy factors in 1960 and 1965.
39
ThistableshowsthecovariancebetweenDem60andDem65.Thisestimateisalsothe
estimatedcorrelationbetweenthetwolatentfactorsbecausethe variancesofthe
factorsarefixedtoone.Thiscorrelationisextremelyhigh,possiblybecausethepolitical
democracystatusdonotchangemuchduringthose5years.
39
Absolute Index
Parsimony Index
Incremental Index
N Observations
N Variables
N Moments
N Parameters
N Active Constraints
Baseline Model Function Value
Baseline Model Chi-Square
Baseline Model Chi-Square DF
Pr > Baseline Model Chi-Square
Fit Function
Chi-Square
Chi-Square DF
Pr > Chi-Square
Z-Test of Wilson & Hilferty
Hoelter Critical N
Root Mean Square Residual (RMSR)
Standardized RMSR (SRMSR)
Goodness of Fit Index (GFI)
Adjusted GFI (AGFI)
Parsimonious GFI
RMSEA Estimate
RMSEA Lower 90% Confidence Limit
RMSEA Upper 90% Confidence Limit
Probability of Close Fit
ECVI Estimate
ECVI Lower 90% Confidence Limit
ECVI Upper 90% Confidence Limit
Akaike Information Criterion
Bozdogan CAIC
Schwarz Bayesian Criterion
McDonald Centrality
Bentler Comparative Fit Index
Bentler-Bonett NFI
Bentler-Bonett Non-normed Index
Bollen Normed Index Rho1
Bollen Non-normed Index Delta2
James et al. Parsimonious NFI
75
8
36
17
0
6.1482
454.9633
28
<.0001
0.6009
44.4686
19
0.0008
3.1383
51
0.5388
0.0494
0.8658
0.7457
0.5875
0.1346
0.0833
0.1865
0.0062
1.1240
0.9065
1.4608
78.4686
134.8659
117.8659
0.8438
0.9403
0.9023
0.9121
0.8560
0.9416
0.6122
40
Wehavelookedattheestimatesandconcludedthattherelationshipsbetweenthe
factorsandthevariablesarestrongandsignificant.Thoseresultsvalidatedthe
individualfactorvariablerelationships.
Togainsupportfortheoverallconfirmatoryfactormodel,youwouldwanttoexamine
themodelfitstatistics.ThistableshowsvariousfitindicescomputedbyPROCCALIS.In
theSEMfield,alargenumberoffitindiceshavebeenproposed. Thereisnoconsensus
astowhichindicesarebesttoreportintheresearch.Butresearcherstendtoreport
someofthemostpopularonesintheirrespectivefields.
Becausealargenumberofindicesmightbeconfusing,PROCCALIS providesawayto
customizethisfitsummarytable.
40
41
YoucanusetheFITINDEXstatementtocustomizeyourfitsummary table.
UsetheON(ONLY)=optiontoselectyourfavorite fitindices.
UsetheNOINDEXTYPEoptiontosuppresstheprintingofthefitindextypes.
41
Fit Summary
Chi-Square
Chi-Square DF
Pr > Chi-Square
Hoelter Critical N
Standardized RMSR (SRMSR)
Adjusted GFI (AGFI)
RMSEA Estimate
Bentler Comparative Fit Index
44.4686
19
0.0008
51
0.0494
0.7457
0.1346
0.9403
42
ThisisthecustomizedfitsummaryoutputbyusingthepreviousFITINDEXstatement
specifications.
Thistablecontainsthemorepopularfitindicesreportedinresearch(asrecognizedby
theauthor).
Inpractice,themodelfitchisquaremodelstatistic,itsdf,andthecorrespondingp
valueareroutinelyreportedeventhoughveryfewresearchersin thefieldwoulduse
themodelfitchisquarealonetojudgemodelfit.Asshowninthistable,thepvalueis
verysmallsothatstatisticallyitmeansthatthehypothesizedmodelshouldberejected.
However,itisaknownissueinSEMthatevenveryusefulSEMmodelswithminimum
departuresfromthedatawouldberejectedstatistically.Therefore,researchersinthe
SEMfieldtendtofocusmoreonotherfitindicestojudgemodel fit.
TheSRMSR,AGFI,RMSEA,andCFIarefourofthemostpopularfit indicesintheSEM
field.Seetheglossarypageforthedescriptionsofthesefitindices.FortheSRMSRand
RMSEA,thesmallerthevaluesthebetterthefit.Usually,valuesunder0.05indicate
goodmodelfit.Therefore,theSRMSRsaysthatthecurrentmodel isgood,butthe
RMSEAsaysthatthecurrentmodelisbad.FortheAGFIandBentlers CFI,thelargerthe
valuesthebetterthemodelfit.Therefore,theAGFIsaysthatthecurrentmodelisbad,
buttheCFIsaysthatitisgood.Becausetheseindicesdonotconsistentlyindicatea
goodmodelfit,itissafetosaythatthecurrentCFAmodelispromising,butitneeds
furtherconfirmation.
42
43
43
Press60
lam1
Dem60
Freop60
lam2
lam3
Fair60
lam4
Legis60
Press65
Dem65
1.
lam1
lam2
Freop65
lam3
Fair65
lam4
Legis65
44
Inadditiontofittingabasicconfirmatoryfactormodel,PROCCALISenablesyoutoset
upparameterconstraintseasily.Themaintoolistouseparameternamesinthe
specification.
Forthepoliticaldemocracyexample,theresearcherwantstoconstraintthefactor
loadings(pathcoefficients)acrosstime.Thetheoreticalreason isthatbasicallythe
measuredvariablesarethesameinthetwoyears.Inthepathdiagram,youcan
representequalityconstraintsbyputtingthesameparameternamesorlabelstothe
pairsoftherelatedpaths.Forexample,lam1istheloadingofPress60onDem60.Itis
alsotheloadingofPress65onDem65.Similarly,youcansetthe other3setsof
constraintsinthepathdiagram.
44
45
InthePATHmodelinglanguage,theconstraintscouldbehandledsimilarly.Thecode
showninthisslideismodifiedfromthepreviouscodebyadding theparameternames
inthepaths.Thesyntaxistoaddanequalsignandthentheparameternamesafterthe
pathspecificationsinthePATHstatement.Withthesameparameternamesforthe
pairsoftherelatedpaths,theestimateswouldbeexactlythesame.
45
Fit Summary
Chi-Square
Chi-Square DF
Pr > Chi-Square
Hoelter Critical N
Standardized RMSR (SRMSR)
Adjusted GFI (AGFI)
RMSEA Estimate
Bentler Comparative Fit Index
46.8893
23
0.0023
56
0.0714
0.7844
0.1185
0.9440
46
Thistableshowsthefitsummaryofthemodelwiththeloadingconstraints.Becauseof
theconstraints,thismodeldoesnotfitaswellastheprevious model.TheSRMSRis
largerthan0.05.TheAGFIismuchsmallerthan0.9.TheRMSEAismuchlargerthan0.05.
Alltheseshowabadmodelfit.However,Bentlers CFI(0.94)showsagoodmodelfit.
46
PATH List
----------Path----------
Parameter
Dem60
Dem60
Dem60
Dem60
Dem65
Dem65
Dem65
Dem65
lam1
lam2
lam3
lam4
lam1
lam2
lam3
lam4
--->
--->
--->
--->
--->
--->
--->
--->
Press60
Freop60
Fair60
Legis60
Press65
Freop65
Fair65
Legis65
Estimate
Standard
Error
t Value
2.13970
2.80116
2.54987
2.82969
2.13970
2.80116
2.54987
2.82969
0.21716
0.29976
0.27316
0.27285
0.21716
0.29976
0.27316
0.27285
9.85319
9.34473
9.33459
10.37075
9.85319
9.34473
9.33459
10.37075
47
Asrequiredfromthemodel,pathswiththesameloadingparameterhavethesame
estimates.Forexample,bothDem60>Press60andDem65>Press65havealoading
estimateof2.14(lam1).Allloadingestimates,again,arestatisticallysignificant.This
showsthatallthepurportedfactorvariablerelationshipsaresupported.
47
Variable
Dem60
Dem65
Press60
Freop60
Legis60
Press65
Freop65
Legis65
Fair60
Fair65
Parameter
_Add1
_Add2
_Add3
_Add4
_Add5
_Add6
_Add7
_Add8
Estimate
Standard
Error
t Value
1.00000
1.00000
2.01017
6.72037
2.88468
2.61966
4.16958
2.85029
5.40824
3.55382
0.40312
1.21196
0.60956
0.49456
0.79818
0.59556
0.97833
0.67700
4.98647
5.54503
4.73243
5.29699
5.22383
4.78593
5.52803
5.24935
Var1
Var2
Parameter
Dem65
Dem60
_Add9
Estimate
Standard
Error
t Value
0.97480
0.02682
36.34662
48
Alltheerrorvarianceestimatesarealsosignificant.ThecorrelationbetweenDem60and
Dem60isveryhighandsignificant.
48
49
49
Press60
lam1
Dem60
Freop60
lam2
lam3
Fair60
lam4
Legis60
Press65
Dem65
1.
lam1
lam2
Freop65
lam3
Fair65
lam4
Legis65
50
Withtheloadingconstraints,youobservedaworsemodelfit.
Thefourequalityconstraintsontheloadingsyoubasicallyreducethenumberofmodel
parametersby4.Thisnaturallyleadstoaworsemodelfitthanifyouwouldallowallthe
loadingstobefreelyestimated.
Nowyouconsideranoppositedirection.Insteadofreducingparametersbyputtingequality
constraints,youwanttoaddmoreparameterstothemodel.Addingmoreparameterstoyour
modelwouldimprovethemodelfit.Butthedrawbackofaddingmoreparametersisthatit
makesyourmodelmorecomplicated,whichisusuallyjudgedasan undesirablepropertyfora
scientifictheory.Itdoesnotmeanthatyoucannotaddparameters.Itonlymeansthatyou
shouldaddonlythoseparametersthatcouldbejustifiedbytheoreticalorsubstantivereasons.
Inthisexample,ithasbeenarguedthatfreedomofgroupoppositionandtheelectivenatureof
thelegislativebodyhaveapartoftheircorrelationthatisbeyondtheircommonlatentfactors
couldexplain(seeBollen).InSEM,thisextra correlationisconceptualizedasacorrelation(or
covariance)betweentheerrorsofthetwovariables.Inthepath diagram,thiserrorcovarianceis
representedbyadoubleheadedarrowconnectingthetwovariables.Thatis,Freop60and
Legis60areconnectedbyadoubleheadedarrowin1960.Bythesameargument,Freop65and
Legis65arealsoconnectedbyadoubleheadedarrowtorepresenterrorcovariance.
Inaddition,itisarguedinBollen thateachofthevariablepairsthatwereofthesamenaturebut
weremeasuredatdifferenttimeshaveapartofcorrelationthat isbeyondtheircommonlatent
factorscouldexplain.Forexample,Press60andPress65areconnectedbyadoubleheaded
arrowtorepresenttheirerrorcovariance,whichexplainsthepartofthecovariancebetween
thetwovariablesthatisbeyondtheexplanationbythecovariancebetweenDem60andDem65.
Similarly,theFreop,Fair,andLegis pairsareallconnectedbydoubleheadedarrowsto
representerrorcovariances.
50
Press60
lam1
Dem60
Freop60
lam2
lam3
lam4
Fair60
Legis60
Press65
Dem65
1.
lam1
lam2
Freop65
lam3
Fair65
lam4
Legis65
1.
1.
1.
1.
1.
1.
1.
1.
e1
e2
e3
e4
e5
e6
e7
e8
51
Thepathdiagraminthisslideisequivalenttothepreviousrepresentationthatdoesnot
useexpliciterrorvariables.
Inthispathdiagram,errortermsforthemeasuredvariablesare shown.Thedouble
headedarrowsareshiftedtotheerrorterms.Thismakesitobviousthatthosedouble
headedarrowsarecovariances betweentheerrorvariables(butnotaspartial
covariances betweentheobservedvariables,asshowninthepreviousslide).
Therefore,thispathdiagramrepresentationisconceptuallycleareraboutwhatare
reallybeingcorrelatedinthemodel.However,theadditionoftheerrortermsmakes
thepathdiagrammorecluttered.Inthisworkshop,Iwouldstick withthepathdiagram
representationthatdoesnotuseexpliciterrorterms.
51
52
Withthesixadditionalpairsofcorrelatederrors,youhavesix moreerrorcovariance
parametersinthemodel.
InthePATHmodelinglanguage,youcanspecifythesecovarianceparametersinthe
PCOVstatement.Inthisexample,thismeansthatyouenumeratethesixpairsof
measuredvariablesinthePCOVstatement.Forexample,thefirst pairisFreop60and
Legis60,whichrepresentacovarianceparameterbetweentheirerrorterms.
52
Fit Summary
Chi-Square
Chi-Square DF
Pr > Chi-Square
Hoelter Critical N
Standardized RMSR (SRMSR)
Adjusted GFI (AGFI)
RMSEA Estimate
Bentler Comparative Fit Index
15.1946
17
0.5815
135
0.0590
0.9043
0.0000
1.0000
53
Thismodelissupposedtofitbetterbecauseoftheaddedparametersfortheerror
covariances.
Infact,themodelfitchisquareisnotstatisticallysignificant.Thissupportsthe
hypothesizedmodelinthepopulation.
Allotherfitindicesshowgoodorexcellentfit.TheSRMSRis0.059,whichisonlyslightly
largerthanthe0.05criterion.TheAGFIis0.90,whichisanindicationofgoodmodelfit
byconvention.TheRMSEAisessentiallyzero,whichisthesmallestRMSEAyoucould
everget.TheCFIis1,whichisalsothelargestCFIyoucouldeverget.
53
PATH List
----------Path----------
Parameter
Dem60
Dem60
Dem60
Dem60
Dem65
Dem65
Dem65
Dem65
lam1
lam2
lam3
lam4
lam1
lam2
lam3
lam4
--->
--->
--->
--->
--->
--->
--->
--->
Press60
Freop60
Fair60
Legis60
Press65
Freop65
Fair65
Legis65
Estimate
Standard
Error
t Value
2.16450
2.61630
2.61693
2.75291
2.16450
2.61630
2.61693
2.75291
0.23009
0.32500
0.28700
0.28312
0.23009
0.32500
0.28700
0.28312
9.40738
8.05017
9.11832
9.72356
9.40738
8.05017
9.11832
9.72356
54
Allloading(pathcoefficients)estimatesarestatisticallysignificant,supportingthe
relationshipsbetweenthelatentfactorsandthemeasuredvariables.
54
Variable
Dem60
Dem65
Press60
Freop60
Legis60
Press65
Freop65
Legis65
Fair60
Fair65
Parameter
_Add1
_Add2
_Add3
_Add4
_Add5
_Add6
_Add7
_Add8
Estimate
Standard
Error
t Value
1.00000
1.00000
1.91664
7.65544
3.27028
2.52969
4.87208
3.25392
5.03798
3.32508
0.43982
1.39023
0.73387
0.52882
0.94384
0.73319
0.98299
0.71220
4.35780
5.50661
4.45621
4.78360
5.16199
4.43805
5.12514
4.66875
Allerrorvarianceestimatesaresignificant.
55
Var1
Var2
Parameter
Dem65
Dem60
_Add9
Estimate
Standard
Error
t Value
0.96603
0.02928
32.99044
Error
of
Parameter
Freop60
Freop65
Press60
Freop60
Fair60
Legis60
Legis60
Legis65
Press65
Freop65
Fair65
Legis65
_Parm1
_Parm2
_Parm3
_Parm4
_Parm5
_Parm6
Estimate
Standard
Error
t Value
1.42826
1.26677
0.58548
2.09624
0.74805
0.47686
0.69666
0.59365
0.37178
0.74763
0.62336
0.46214
2.05017
2.13389
1.57478
2.80386
1.20003
1.03186
Thefirsttableshowsthecorrelationbetweenthetwolatentfactors.Again,the
correlationisveryhighandsignificant.
Thesecondtableshowstheestimatesforthenewlyaddedcovariances betweenerrors.
Threeofthesecovariances aresignificant,whiletheothersarenot.Forexample,
Freop60andLegis60,Freop65andLegis65,andFreop60andFreop65arethreeerror
covariances thathavetvalueslargerthan1.96.Theotherthreepairshave insignificant
tvalues.Thismeansthataddingthesethreecovariances mightbesomewhat
undesirablebecausetheirestimatesareactuallynotsignificantlydifferentfromzero,
castingdoubtsabouttheirpresenceinthemodel.
Thelessonhereisthateventhoughaddingerrorcorrelations(orcovariances)might
improvethemodelfit,youshouldnotroutinelyadderrorcovariances onlytoboostthe
modelfit.Addingunjustifiederrorcovariances makesyourmodelmorecomplicatedand
hardertointerpret,especiallywhensomeerrorvarianceestimatesturnouttobe
insignificant.
56
57
57
58
Wecontinuewiththepreviousmodelandaddonemorelatentfactoranditsindicators
intothemodel.
Thisexampleillustratesafullstructuralequationmodel(orafullLISREL)model.
Essentially,thismeansthatourfocusisnotonlyonvalidating therelationshipsbetween
thelatentfactorsandthemeasuredvariables(thatis,themeasurementmodel),but
alsoonvalidatingthefunctionalrelationshipsamonglatentvariables(thatis,the
structuralmodel).
Forexample,youhavealatentfactorcalledindustrialization(Induct)thatissupposedto
bereflectedbythreeobservedvariables:grossnationalproduct percapita(Gnppc60),
energyconsumptionpercapita(enpc60),andpercentoflaborforceinindustrial
occupations(Indlf60).Allthesevariablesweremeasuredin1960.
Theindustrialization(Induct)latentvariableservesasapredictorofthetwodemocracy
factors(Dem60andDem65).Thiskindoffunctionalrelationships betweenlatent
variableshasnotbeenexploredintheconfirmatoryfactormodelsdiscussedpreviously.
58
1.
Freop60
lam2
lam3
Fair60
lam4
Legis60
Indust
Press65
Indlf60
Dem65
1.
lam2
Freop65
lam3
Fair65
lam4
Legis65
59
TheentireSEMmodelisdepictedinthepathdiagramofthecurrentslide.Themost
notableadditionisthepathsfromIndust toDem60andDem65 industrializationin
1960servesasapredictorofdemocracyinboth1960and1965.Threeobserved
variablesserveasindicatorsoftheindustrialization:Gnppc60, Enpc60,andIndlf60.
Therearetwomainmodificationsfromtheprecedingconfirmatory factormodel.
First,insteadofallowingDem60andDem65tofreelycovary intheCFA,thecurrent
modeltreatsDem60asapredictorofDem65.
Second,adifferentmethodforidentifyingthelatentfactorscalesisusedinthecurrent
model.IntheprecedingCFAmodel,variancesofDem60andDem65arefixedtoone.
Butbecausetheybecomeendogenousinthecurrentmodel,youcan nolongerusethis
typeofscaleidentificationmethod.Instead,oneoftheirobservedindicatorvariables
(thatis,Press60andPress65)nowhasafixedpathcoefficientatone.Similarly,thepath
coefficientfromIndust toGnppc60isfixedtooneforscaleidentification.
59
1.
lam2
lam3
lam4,
Dem65
1.
lam2
lam3
lam4,
1.,
---> Dem65;
pcov
Freop60 Legis60, Freop65 Legis65,
Press60 Press65, Freop60 Freop65,
Fair60 Fair65, Legis60 Legis65;
fitindex on(only) = [chisq df probchi rmsea cn srmsr
bentlercfi agfi] noindextype;
run;
60
YoucanusePROCCALIStospecifythisstructuralequationmodel easily.
InthePATHstatement,Iuseamultiplepathspecificationsyntax.Inthefirst
specification,Dem60isapredictorof4outcomevariables:Press60,Freop60,Fair60,
andLegis60.Thisspecifiesfourpathsinasinglepathspecification.Afterusinganequal
sign,Ispecifyfourparametersforthefourpaths.Thefirstoneisafixedconstant1,
whichisappliedtotheDem60>Press60path.Thesecondoneisafreeparameter
lam2,whichisappliedtotheDem60>Freop60path,andsoon.
Inthenext3pathspecifications,Ialsousethemultiplepathspecificationsyntax.The
secondmultiplepathsyntaxspecifiesthatDem65isafactorwithfourindicators.The
pathcoefficients(loadings)arealsospecifiedexplicitly.Thethirdmultiplepathsyntax
specifiesthatIndust isafactorofthreeobservedindicators,withafixedoneforthe
effectofIndust onGnppc60.ThepathcoefficientsforthepathsIndust>Enpc60and
Indust>Indlf60areunnamedfreeparameters(withtheemptyspecifications).The
fourthmultiplepathsyntaxspecifiesthatIndust isapredictorofbothDem60and
Dem65.Thecorrespondingpathcoefficientsare(unnamed)freeparametersinthe
model.
ThelastpathinthePATHstatementspecifiesDem60asapredictorofDem65.Notice
thatnoPVARstatementisusedbecausefixingtheDem60andDem65variancestoone
isnotusedinthecurrentmodel.Thescalesofthelatentfactorsareidentifiedbyfixing
somepathcoefficientsto1.
60
39.6438
38
0.3966
100
0.0558
0.8606
0.0242
0.9975
61
Thefitofthestructuralmodelisacceptable,ifnotexceptionallygood.
Themodelfitchisquareisnotsignificant,supportingthehypothesizedmodel.The
SRMSRiscloseto0.05.TheAGFIis0.86,whichshowsareasonablefit.TheRMSEA
indicatesaverygoodmodelfit,asthevalue(.0242)ismuchlowerthan0.05.TheCFIis
almost1,whichshowsaperfectmodelfit.
61
----------Path----------Dem60
Dem60
Dem60
Dem60
Dem65
Dem65
Dem65
Dem65
Indust
Indust
Indust
Indust
Indust
Dem60
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
Press60
Freop60
Fair60
Legis60
Press65
Freop65
Fair65
Legis65
Gnppc60
Enpc60
Indlf60
Dem60
Dem65
Dem65
Parameter
lam2
lam3
lam4
lam2
lam3
lam4
_Parm01
_Parm02
_Parm03
_Parm04
_Parm05
Estimate
1.00000
1.19079
1.17454
1.25099
1.00000
1.19079
1.17454
1.25099
1.00000
2.17966
1.81821
1.47133
0.60046
0.86504
Standard
Error
t Value
0.14020
0.12121
0.11757
8.49336
9.68988
10.64006
0.14020
0.12121
0.11757
8.49336
9.68988
10.64006
0.13932
0.15290
0.39496
0.22722
0.07538
15.64530
11.89126
3.72529
2.64267
11.47648
62
Allpathcoefficientsaresignificantaprettygoodsign.
62
Variable
Parameter
Indust
Press60
Freop60
Legis60
Press65
Freop65
Legis65
Gnppc60
Enpc60
Indlf60
Fair60
Fair65
Dem60
Dem65
_Add01
_Add02
_Add03
_Add04
_Add05
_Add06
_Add07
_Add08
_Add09
_Add10
_Add11
_Add12
_Add13
_Add14
Estimate
Standard
Error
t Value
0.45466
1.87973
7.68378
3.26801
2.34432
5.03534
3.35236
0.08249
0.12206
0.47297
5.02270
3.60813
3.92767
0.16668
0.08846
0.44229
1.39404
0.73807
0.48851
0.93993
0.71788
0.01986
0.07105
0.09197
0.97587
0.72394
0.88311
0.23158
5.13991
4.25001
5.51189
4.42779
4.79895
5.35716
4.66983
4.15376
1.71776
5.14268
5.14691
4.98402
4.44753
0.71975
63
Someerrorvariancesarenotsignificant:Enpc60andDem65.Enpc60isanindicatorof
theIndustrializationin1960.ThisinsignificanterrorvariancemeansthattheIndust
factorpredictEnpc60perfectly.However,thecorrespondingtvalueis1.71,whichcould
bejudgedasmarginallysignificant.
TheerrorvarianceforDem65isalsonotsignificant,asevidentbythenonsignificantt
valueof0.72.ThismeansthatgivenIndust andDem60,Dem65canbepredictedalmost
perfectly.
63
Error
of
Parameter
Freop60
Freop65
Press60
Freop60
Fair60
Legis60
Legis60
Legis65
Press65
Freop65
Fair65
Legis65
_Parm06
_Parm07
_Parm08
_Parm09
_Parm10
_Parm11
Estimate
Standard
Error
t Value
1.45956
1.39032
0.59042
2.21252
0.72123
0.36769
0.70251
0.58859
0.36307
0.75242
0.62333
0.45324
2.07764
2.36212
1.62619
2.94054
1.15706
0.81125
64
Again,therearesomeinsignificanterrorcovariances.Thisresultchallengestheir
presenceinthemodel.
64
Variable
Enpc60
Fair60
Fair65
Freop60
Freop65
Gnppc60
Indlf60
Legis60
Legis65
Press60
Press65
Dem60
Dem65
Error
Variance
Total
Variance
R-Square
0.12206
5.02270
3.60813
7.68378
5.03534
0.08249
0.47297
3.26801
3.35236
1.87973
2.34432
3.92767
0.16668
2.28211
11.79895
10.09361
14.64875
11.70144
0.53715
1.97602
10.95502
10.70953
6.79166
7.04548
4.91193
4.70116
0.9465
0.5743
0.6425
0.4755
0.5697
0.8464
0.7606
0.7017
0.6870
0.7232
0.6673
0.2004
0.9645
65
Thesquaremultiplecorrelationsareusuallyusedtomeasurethe percentageof
overlappingvariancebetweenthepredictorsandtheoutcomevariables.Inthecurrent
example,Rsquaresrangefrom0.2toextremehighvaluessuchas0.95and0.96.
ThesmallestRsquareistheoneforpredictingDem60,whichis0.2.Thisactuallyisnot
thatsmallanRsquarevalueforsocialsciencedata.
ButtheRsquare(0.96)forDem65isextremelyhigh.ThismeansthatDem65 isalmost
perfectlypredictedfromdemocracyandindustrializationin1960.
65
1.
Measurement
model
Gnppc60
Enpc60
Measurement
model
Dem60
1.
Press60
Freop60
lam2
lam3
Fair60
lam4
Legis60
Indust
Press65
Indlf60
Dem65
Measurement
model
1.
lam2
Freop65
lam3
Fair65
lam4
Legis65
66
ThisfullSEMmodelisalsoagoodillustrationoftheLISRELmodel.
Thepathdiagramfortheprecedingmodelremainsunchangedhere. Inordertocallthis
pathdiagramaLISRELmodel,youhavetoidentifytheLISRELcomponentsinthispath
diagram.ThetwomaincomponentsinLISRELarethemeasurementmodelsandthe
structuralmodel.
First,themeasurementmodelsareidentified.Ameasurementmodelisabouthow
observedvariablesarerelatedtothelatentvariablesorconstructsinthemodel.
Specifically,themeasurementmodelinvolvingindustrializationisthemeasurement
modelofxbecauseIndust servesasanexogenous(independent)variableinthepath
diagram.ThemeasurementmodelinvolvingDem60andDem65isthemeasurement
modelofybecauseDem60andDem65areendogenous(dependent)variablesinthe
pathdiagram.
Second,thestructuralmodelisidentifiedandhighlightedinthecenterofthepath
diagram.Thestructuralmodeldescribesthefunctionalrelationshipsamongthelatent
variables(constructs)inthepathdiagram.
Therefore,alltheessentialcomponentoftheLISRELmodelisidentifiedinthecurrent
pathdiagram.
66
1.
lam2
lam3
lam4,
Dem65
1.
lam2
lam3
lam4,
1.,
---> Dem65;
pcov
Freop60 Legis60, Freop65 Legis65,
Press60 Press65, Freop60 Freop65,
Fair60 Fair65, Legis60 Legis65;
fitindex on(only) = [chisq df probchi rmsea cn srmsr
bentlercfi agfi] noindextype;
Measurement model
run;
67
InthePATHmodelinglanguage,youcanalsoidentifythecodeforthemeasurement
modelsandthestructuralmodel.Theprecedingcodeisrecitedhereforillustrations.
InthePATHstatement,thefirstthreemultiplepathspecificationsareconcernedwith
themeasurementofthelatentconstructs.Inaddition,allspecificationsinthePCOV
statementareforthecovariances ofthemeasurementerrors.
ThelasttwospecificationsinthePATHstatementareforthestructuralmodel.They
describethefunctionalrelationshipsbetweenIndust,Dem60,andDem65.
AfteridentifyingtheLISRELcomponentsinthepathdiagramandintheSAScode,you
nowhaveacluetospecifytheLISRELmodelinPROCCALIS.Essentially,itshouldbe
clearthatthesamepathdiagramisbeingusedbythePATHmodelinglanguageandthe
LISRELmodel.TheonlytaskistotranscribethecodeinthePATHmodelinglanguageto
thelanguagefortheLISRELmodel.
67
[ 1 to 2, 1];
[2,1];
68
PROCCALISsupportsthesocalledLISMODmodelinglanguage.Inordertofullyunderstandthe
PROCCALIScodefortheLISRELmodel,knowledgeaboutmatrixalgebraisneeded.ButIwillonly
describethecodeinaconceptualway.
IntheLISMODstatement,youfirstclassifyyourvariablesintooneofthefourcategories:
1.xvariables:observedindicatorsfortheexogenous(independent)latentfactorsinthemodel.
2.yvariables:observedindicatorsfortheendogenous(dependent)latentfactorsinthemodel.
3.xivariables:exogenous(independent)latentfactorsinthemodel.
4.etavariables:endogenous(dependent)latentfactorsinthemodel.
TheLISRELmodelorLISRELprogramhadbeendevelopedasamatrixbasedlanguage.
Parametersinthemodelsarespecifiedasmatrixelementsinsomespecificmodelmatriceswith
Greeknames.PROCCALISsupportsthematrixinputoftheseLISRELmodelmatrices.For
example,inthemeasurementmodelfory,_LambdaY_isthematrixthatrelatestheyvariables
totheetavariables.InsteadofspecifyingthepathsasinthePATHstatement,theMATRIX
statementfor_LambdaY_servesthesamepurposeintheLISMODmodelinglanguage.The
MATRIXstatementfor_ThetaY_specifiestheerrorvariancesandcovariances oftheyvariables,
muchlikethespecificationsofthePCOVstatementinthePATHmodelinglanguage.Inother
words,thePATHmodelspecificationsaretranscribedintotheLISMODmodelspecificationsfor
theyvariables.
Similarly,theMATRIXstatementfor_LAMBDAX_specifiestheparametersinthemeasurement
modelforthexvariables.
Finally,thestructuralrelationshipsorthepathrelationshipsamongthelatentfactorsare
specifiedintheMATRIXstatementsforthe_GAMMA_and_BETA_matrices.
Tosimplifytheoutput,IusedtwooptionsinPROCCALISstatement.TheNOSEoption
suppressestheprintingofstandarderrorsandtheNOPARMNAMEoptionsuppressesthe
printingoftheparameternames.
68
Dem60
Dem65
Press60
1.0000
Freop60
1.1908
Fair60
1.1745
Legis60
1.2510
Press65
1.0000
Freop65
1.1908
Fair65
1.1745
Legis65
1.2510
ThefollowingfewslidesshowtheoutputfromPROCCALISforthe LISRELmodel.Allthe
resultsarematrixoriented.DetailsfortheseresultshavebeendiscussedforthePATH
modeloutputandwillnotberepeatedhere.Ingeneral,youcanfindcorrespondence
betweentheLISMODandthePATHresults.
Thisslideshowsthemeasurementmodelfortheyvariables.
69
Press60
Freop60
Fair60
Legis60
Press65
Freop65
Fair65
Legis65
Press60
1.8797
0.5904
Freop60
7.6838
1.4596
2.2125
Fair60
5.0227
0.7212
Legis60
1.4596
3.2680
0.3677
Press65
0.5904
2.3443
Freop65
2.2125
5.0353
1.3903
Fair65
0.7212
3.6081
Legis65
0.3677
1.3903
3.3524
70
Thisslideshowsthemeasurementerrorvariancesandcovariances fortheyvariables.
70
Indust
Gnppc60
1.0000
Enpc60
2.1797
Indlf60
1.8182
_THETAX_ Matrix
Gnppc60
Gnppc60
Enpc60
Indlf60
0.0825
Enpc60
0.1221
Indlf60
0.4730
Thisslideshowstheresultsofthemeasurementmodelforthexvariables,includingthe
pathcoefficientsandtheerrorvariances.
71
Dem60
Dem65
Dem60
Dem65
0.8650
_GAMMA_ Matrix
Indust
Dem60
1.4713
Dem65
0.6005
72
Thisslideshowsthefunctionalrelationshipsbetweenlatentconstructs.
72
Dem60
Dem65
Dem60
3.9277
Dem65
0.1667
_PHI_ Matrix
Indust
Indust
0.4547
Note: All variances (diagonal elements in _PSI_ and _PHI_) were set by default.
73
Thisslideshowstheerrorvariancesoftheetavariableandthevarianceofthexi
variable.
73
Insum,theLISMODmodelinglanguageinPROCCALISsupportstheLISRELmodelbyproviding
syntaxtospecifytheessentialcomponentsoftheLISRELmodel.However,LISMODitselfdoesnot
interpretaLISRELprogram.
TheLISMODmodelinglanguageinPROCCALISalsosupportsthemeanstructureanalysis.Thisis
donebyprovidingadditionalMATRIXstatementinputforthemean modelmatricesintheLISREL
model.
IfyouunderstandtheLISRELmodel,herearethreethingsyouinputbyusingtheLISMODlanguage:
1.Theorderedlistsofx,y,,and variables
2.MATRIXstatementstodefinefreeandfixedparameters
3.Namesforparameters(notrequiredforfreeparameters)
(Note:Thisslidewasupdatedaftertheprintingofthehandout)
TheDefaultcovariancestructureparametersintheLISMODlanguageare:
1.Diagonalelementsofallcovariancematrices(allvariances)
2.Lowertriangularelementsofthe_PHI_matrix(covariances ofthe variables)
Specifyingthedefaultparametersexplicitlyiscertainlyallowed,especiallywhenyouneedtoset
constraintsontheseparameters.
Inaddition,whenthemeanstructuresaremodeled,theinterceptsofthex andy variablesare
defaultfreeparameters,whiletheinterceptsofthe variablesandthemeansofthe variables
arefixedzerosbydefault.
74
The Generalized
PATH Modeling Language
75
75
1.
Freop60
lam2
lam3
Fair60
lam4
Legis60
Indust
Press65
Indlf60
Dem65
1.
lam2
Freop65
lam3
Fair65
lam4
Legis65
76
TogeneralizethePATHmodelinglanguage,errorcovariances inthepathdiagramcould
alsobespecifiedaspaths inPROCCALIS.Infact,covariances inthepathdiagramare
alreadyrepresentedasdoubleheadedarrows,asshowninthepoliticaldemocracyand
industrializationexample.
76
Legis60
Legis65
=
=
=
1. lam2
1. lam2
1.,
lam3
lam3
lam4,
lam4,
=
=
=
1. lam2
1. lam2
1.,
Legis65,
Freop65,
Legis65;
lam3
lam3
lam4,
lam4,
ThetoppanelshowstheuseofPCOVstatementtospecifythecovariances.Thebottom
panelshowsthatthesecovariances arespecifiedasdoubleheadedpaths,which
resembletheirrepresentationsinthepathdiagram.
ThetwoPROCCALISspecificationsshownaboveareequivalent.
77
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
--->
<-->
<-->
<-->
<-->
<-->
<-->
Press60
Freop60
Fair60
Legis60
Press65
Freop65
Fair65
Legis65
Gnppc60
Enpc60
Indlf60
Dem60
Dem65
Dem65
Legis60
Legis65
Press65
Freop65
Fair65
Legis65
Parameter
lam2
lam3
lam4
lam2
lam3
lam4
_Parm01
_Parm02
_Parm03
_Parm04
_Parm05
_Parm06
_Parm07
_Parm08
_Parm09
_Parm10
_Parm11
Estimate
1.00000
1.19079
1.17454
1.25099
1.00000
1.19079
1.17454
1.25099
1.00000
2.17966
1.81821
1.47133
0.60046
0.86504
1.45956
1.39032
0.59042
2.21252
0.72123
0.36769
Standard
Error
t Value
0.14020
0.12121
0.11757
8.49336
9.68988
10.64006
0.14020
0.12121
0.11757
8.49336
9.68988
10.64006
0.13932
0.15290
0.39496
0.22722
0.07538
0.70251
0.58859
0.36307
0.75242
0.62333
0.45324
15.64530
11.89126
3.72529
2.64267
11.47648
2.07764
2.36212
1.62619
2.94054
1.15706
0.81125
78
TheresultsobtainedfromPROCCALISnowshowsthecovarianceestimatesaspaths
inthePATHlist.Thistablecouldbeuseddirectlyinresearchpaperforreportingthe
SEMestimationresults.
78
Y <--> Y
Covariance-path: X <--> Y
Mean or intercept (one-path): 1 ---> Y
79
Insum,thegeneralizedpathmodelinglanguageenablesyoutospecifyalltypesof
arrowsinthepathdiagramaspaths, includingthevariance,covariance,intercept,and
meanparameters.
VarianceofYisapathlikeY<>Y.
CovariancebetweenXandYisapathlikeX<>Y
MeanorinterceptforYisonepathlike1>Y.
79
The main purpose of setting default parameters is to enable you to specify only the
functional relationships among variables in most practical applications.
80
KnowingthedefaultfreeandfixedparametersinPROCCALISareusefulbecauseit
enhancesthecodingefficiencyandaccuracy.Hereisalistofdefaultfreeparameters
andfixedzerosusedinPROCCALIS:
(Thisslidehasbeenchangedslightlyaftertheprintingofthehandout)
Defaultfreeparameters
o Variancesofandcovariances amongallexogenous (independent)variables
(observedorlatent,exceptforerrorterms)
o Errorvariancesofallendogenous (dependent)variables
o Meansorinterceptsofallobserved variables
Defaultfixedzeros
o Unspecifiedpathsanderrorcovariances
o Meansorinterceptsofalllatent variables
Atthefirstglance,itmightseemtobetediousanddemandingthatmodelersmust
rememberallthesedefaultparameterrulestospecifyanSEMaccurately.However,the
defaultparameterizationusedinPROCCALISmatchesthatofregressionanalysisandit
isdesignedwiththefollowingmainpurposeinmind:Inmostpracticalapplications,you
wouldonlyneedtospecifythefunctionalrelationshipsamongvariables(thatis,the
singleheadedpathsinthepathdiagram)andthefixedvariancesofthe latentvariables.
80
Multiple-Group Analysis
81
MultiplegroupanalysisrepresentsanimportantclassofSEMapplications.
81
82
TheskeletonofthemultiplegroupsyntaxofPROCCALISisshowninthisslide.
Inthisspecification,youhavetwoindependentgroups,whicharestoredinseparate
datasetsds1 andds2. Youdefinetwomodels:Model1andModel2,respectively,for
fittingtothetwogroups.
UndereachMODELstatement,youspecifythemodelbyoneofthemodelinglanguages
supportedbyPROCCALIS.Forexample,aPATHstatementoraLISMODstatement.
82
Note: REFMODEL does not constrain default free parameters in the two models.
It constrains only those parameters that are specified explicitly.
83
Thisexampleshowsthreegroupsofdata.Model1isfittedtoGroups1and2bythe
PATHmodelspecifiedimmediatelyafter.Model2isfittedtoGroup3.TodefineModel
2,aREFMODELstatementisused.REFMODEL1meansthatthespecificationsinModel
1isreferencedhere.Thismeansthatallexplicit specificationsinModel1arecopied
intothecurrentmodel.Forexample,allthePATHstatementspecificationsinModel1
arecopiedintoModel2.However,REFMODELallowsmodificationsfromthereference
model.Inthisexample,thex3>x5pathisrespecifiedwithanewpathcoefficient
calledeffect2. Thiseffect2 inModel2isadifferentparameterthaneffect1 inModel
1,eventhoughbothmodelhavethisx3>x5path.Exceptforthisx3>x5path,Models
1and2sharealltheremainingspecifiedpathswiththesamesetofpathcoefficients.
NoticethatbyusingtheREMODELstatementtoconnectmodels,all(andonly)the
explicit specificationsinthemodelsareconstrained(unlessbeingmodifiedorre
specified).Theconstraints,however,willnotapplytodefaultparameters(assumingthat
theyarenotexplicitlyspecified).Toconstrainthedefaultparameters,theywillhaveto
bespecifiedexplicitlyintheMODELdefinitions.
83
Females
v1
1.
Spatial
a2
Visual
ev1
Cubes
ev2
v1
1.
Spatial
a3
a2
Visual
ev1
Cubes
ev2
Lozenges
ev3
Paragraph
ev4
Sentence
ev5
WordMean
ev6
a3
Lozenges
ev3
cv
cv
1.
Verbal
a5
Paragraph
ev4
Sentence
ev5
1.
a6
a5
a6
WordMean
v2
Verbal
ev6
v2
Data source: Holzinger and Swineford1939 (an example used in Arbuckle 2008)
84
Letususeanexampletoillustratethemultiplegroupanalysis.TheHolzinger and
Swineford dataaresued.ThisdatasetisalsousedinArbucklesAMOSmanual.
Inthisresearch,visualandverbaltestscoreswereobserved.Visual,Cubes,and
Lozengesarespatialtests.Paragraph,Sentence,andWordMean areverbaltests.CFA
modelswerehypothesizedforthetwogroups:onegroupisformalesandtheotherfor
females.Thebasicfactorstructuresforthegroupsarethesame.Twolatentfactors
Spatial andVerbal areassumedforthemeasuredvariables.Allparameters(including
thosedefaultparametersinPROCCALIS)inthetwomodelsforthegroupsarelabeledin
red.
Thisisacompletelyconstrainedmultiplegroupmodelbecausethetwopathdiagrams
formalesandfemalesareexactlythesame(thatis,bothhavethesamepathstructures
andthesamesetofparameters).
84
85
InPROCCALIS,youcanuseoneofthefollowingthreewaystospecifythepreceding
completelyconstrainedmodel:
1.Maleandfemalegroupsfittedbytwomodels,butwithallparameters(includingall
defaultparameters)beingconstrainedinthemodels
2.Maleandfemalegroupsfittedbyasinglemodeldefinition
3.Maleandfemalegroupsfittedbytwomodelsthatareconstrained throughthe
REFMODELspecification
Iwilldescribeeachofthesemethods.
85
Thefirstmethodistodefinetwomodelsforthetwogroups.The modelspecifications
underthetwoMODELstatementsmustbeexactlythesame.
Thismethodisintuitive,butalittleclumsybecauseyouneedtospecifyallparameters
withmatchingnamesinthemodels(althoughyoucancutandpastethemodel
specificationstoensureanexactcopy).Youalsoneedtospecifyeachparameterinthe
model,includingthedefaultparameters,whichyoumightsometimesmiss.
86
proc calis;
group 1 / label='Males' data=males;
group 2 / label='Females' data=females;
Groups 1 and 2 are fitted
model 1 / group = 1, 2;
by the same model.
path
Spatial ---> Visual Cubes Lozenges
= 1. ,
Verbal
---> Paragraph Sentence Wordmean = 1. ;
fitindex on(only) = [chisq df probchi rmsea srmsr aic
bentlercfi agfi] noindextype;
run;
87
Thesecondmethodisverysimpleandintuitive.Youspecifyonemodelandfitthis
modeltothetwogendergroups.Thisensuresthegroupsarefittedexactlybythesame
model.
Theadvantageofthismethodisthatitissimple,intuitive,andnoparameternamesare
necessaryforconstrainingmodels.Also,youdonotneedtospecifyanyofthedefault
parametersexplicitlyforsettingupconstraints.
Thisisanidealspecificationmethodifthismodelisallwhatyouwanttofit.However,if
youaregoingtofitasequenceofmultiplegroupmodelstothegroups,youmightwant
toconsiderthenextmethod.
87
ThethirdmethodconstrainsthemodelsbytheREFMODELstatement.Asdiscussed
previously,theREFMODELcopiesalltheexplicit specificationsfromthereference
modeltothetargetmodel.Inthisslide,allpathcoefficients, varianceparameters,and
covarianceparametersarespecifiedinModel1,whichisfittedtoGroup1(Males).
Model2,whichisfittedtoGroup2(Females),makesreferencetoModel1withoutany
modificationsorrespecifications.
Noticethatinordertocompletelyconstrainthetwomodelsforthetwogroups,all
parameters,includingthosecouldhavebeensetbydefaultbyPROCCALIS(e.g.,
specificationsinthePVARstatementandPCOVstatement),mustbespecifiedexplicitly
inModel1.Thisway,Model2willcopyalltheseparameterspecificationsviathe
REFMODELstatementspecification.
UnlikeMethod1,thismethoddoesnotrequiretheuseofparameternamesfor
constraintsacrossmodels.ConstraintsaredoneviatheREFMODEL statement.
AlthoughnotasintuitiveasMethod2,thismethodwouldbeusefulifyouneedto
considerfittingasequenceofmultiplegroupmodels,whichwillbeillustratedlater.
88
Fit Summary
Chi-Square
Chi-Square DF
Pr > Chi-Square
Standardized RMSR (SRMSR)
Adjusted GFI (AGFI)
RMSEA Estimate
Akaike Information Criterion
Bozdogan CAIC
Schwarz Bayesian Criterion
Bentler Comparative Fit Index
26.0154
29
0.6247
0.0968
0.9235
0.0000
52.0154
103.7130
90.7130
1.0000
89
Thecompletelyconstrainedmodelprovideagoodfitofthedata. Themodelfitchi
squareisnotsignificant.TheRMSEAisperfect,althoughtheSRMSRisnotverygood.
TheAGFIandtheCFIarealsogood.TheAIC,theCAIC,andtheSBCarealsoprinted.
Theseindicescannotbeinterpretedbytheirabsolutevalues,butwillbeusefulwhen
youcomparethefitsofdifferentmultiplegroupmodels.Youwillusetheseindicesto
selectthebest multiplegroupmodelforthedatalater.
89
90
WehavefittedthecompletelyconstrainedmultiplegroupmodelbytheREFMODEL
method.Wecanfurtherfitlessconstrainedmultiplegroupmodelbymodifyingour
PROCCALIScode.
Wecanreleasetheconstraintsonerrorvariances.
Thenwecanreleasetheconstraintsonthestructuralcovariances (amonglatent
variables).
Finally,wecanreleasetheconstraintsonthepathcoefficients (orloadings).
Iamgoingtoshowthesestepbystep.
90
Females
v1
v1
1.
Spatial
a2
Visual
1.
Cubes
Spatial
a3
a2
Visual
Cubes
a3
Lozenges
cv
Lozenges
cv
1.
Verbal
a5
Paragraph
Sentence
a6
1.
Verbal
Sentence
a6
WordMean
v2
a5
Paragraph
WordMean
v2
Thepathdiagramsforthemultiplegroupmodelthatreleasestheconstraintsonthe
errorvariancesareshownabove.
Exceptfortheerrorvarianceparameters,alltheremainingparametersarelabeled.This
meansthatonlytheerrorvarianceparametersarenotinvariantacrossthemodelsfor
thegroups.
91
Comment out the error variance specifications in the PVAR statement, and let PROC CALIS
set two distinct sets of default error variances for the two models.
92
IntermsofPROCCALISspecification,thismeansthatthemodelforfemalesmakes
referencetothemodelformaleswithregardtothoseconstrainedparametersonly.
Thiscouldbedoneveryeasilyfrommodifyingthecompletelyconstrainedmultiple
groupmodel.AllyouneedtodoistocommentoutthePVARstatementspecifications
fortheobservedvariables.
WhenModel2makesreferencetoModel1,onlythoseexplicitspecificationswouldbe
constrainedbetweenthetwomodels.Becausetheerrorvariancesarenotspecifiedin
bothmodels(thatis,commentedoutfromthepreviouscode),PROCCALISwould
generatedifferentsetsofdefaulterrorvarianceparametersfor thetwomodels.In
otherwords,theerrorvarianceconstraintsarereleasedinthis PROCCALISspecification.
92
22.0334
23
0.5182
0.0903
0.9163
0.0000
60.0334
135.5913
116.5913
1.0000
93
Themodelfitchisquarestillisnotsignificant,indicatingagoodmodelfit.TheRMSEA,
theAGFI,andtheCFIareallgood.However,theSRMSRdoesnotindicateagoodmodel
fit.
93
a2
Females
Visual
1.
Cubes
Spatial
a3
a2
Verbal
a5
Cubes
a3
Lozenges
1.
Visual
Lozenges
Paragraph
Sentence
a6
1.
Verbal
a5
Paragraph
Sentence
a6
WordMean
WordMean
Common parameter names for the structural variances and covariance are removed.
94
Howaboutreleasingtheconstraintsonthestructuralcovariances?
Inthepathdiagram,onlythepathcoefficientsareconstrainednow(byusingthesame
setofparameternames).Thismeansthatonlythepatheffectsareinvariantacrossthe
modelsforthegroups.
94
Thisnewmultiplegroupmodelcanbespecifyingbycommentedouttheexplicit
specificationsofthestructuralcovariances (variancesandcovariances amonglatent
variables)inModel1.
WhenModel2makesreferencetoModel1,itcopiestheexplicitspecificationsinthe
PATHstatementofModel1.Errorvariances,structuralvariances andcovariances inthe
twomodelsarenowsetbydefaultandareunconstrainedbetweenthetwomodels.
95
18.2915
20
0.5682
0.0539
0.9179
0.0000
62.2915
149.7796
127.7796
1.0000
96
Themodelfitchisquareisnotsignificant.Now,theSRMSRisacceptable.TheAGFI,the
RMSEA,andtheCFIcontinuetobeverygood.
96
Females
Visual
1.
Cubes
Spatial
Cubes
Lozenges
1.
Verbal
Lozenges
Paragraph
Sentence
Visual
1.
Verbal
WordMean
Paragraph
Sentence
WordMean
97
Finally,forthecompletelyunconstrainedmultiplegroupthepathdiagramsforthetwo
groupsarethesame,butnoparameternames(exceptforfixedvaluesof1)areusedto
denoteconstraints.
97
,
;
,
;
The REFMODEL statement is not used here because the two models are
not constrained with each other.
98
Becausethetwomodelsforthegroupsaretotallyunrelated,you shouldnotneedto
usetheREFMODELstatementanymore.Instead,thetwomodelsare definedexactlyby
thesamePATHstatementspecifications.However,becausenocommonparameter
namesareusedforthepathcoefficients,thetwomodelsarenot constrained.
98
16.4795
16
0.4200
0.0449
0.9077
0.0205
68.4795
171.8746
145.8746
0.9984
The unconstrained SEM model for the groups gives you the best fit,
but it is also the least interesting multiple-group model.
99
Allfitindicesindicateverygoodfitofthecompletelyunconstrainedmultiplegroup
model.
99
Loadings and
Struct. Cov.
Constrained
Loadings
3.892
Structural Covariances
(p=0.32)
Constrained Loadings
Completely Unconstrained
7.724
3.742
(p=0.44)
(p=0.71)
9.536
5.539
1.182
(p=.27)
(p=.41)
(p=.23)
100
Whichmodelisthebestforthedata?
Chisquaredifferencetestsprovideastatisticalmethodtoseeifmodelsaresignificantly
differentfromeachother.Thisslidesshowsthechisquaredifferencetestsfor
comparingthefourmultiplegroupmodels.
Asallpvaluesarebiggerthan0.05,itmeansallthesemultiplegroupmodelsarenot
significantlydifferentfromeachother.
100
Loadings and
Constrained
Completely
Loadings
Unconstrained
16.4795
Struct. Cov.
26.0154
22.0334
18.2915
29
23
20
16
Pr > Chi-Square
0.6247
0.5182
0.5682
0.4200
0.0968
0.0903
0.0539
0.0449
0.9235
0.9163
0.9179
0.9077
Chi-Square DF
RMSEA Estimate
Akaike Information Criterion
Bozdogan CAIC
Schwarz Bayesian Criterion
Bentler Comparative Fit Index
0.0000
0.0000
0.0000
0.0205
52.0154
60.0334
62.2915
68.4795
103.7130
135.5913
149.7796
171.8746
90.7130
116.5913
127.7796
145.8746
1.0000
1.0000
1.0000
0.9984
Wecanalsocomparethefourmodelsbymeansofthefitindexvalues.
Themodelfitchisquarevaluealwaysfavorsthemodelwiththelargestnumberof
parameters.So,accordingtothemodelfitchisquare,thecompletelyunconstrained
modelisthebestmodel.TheSRMSRalsofavorsthecompletelyunconstrainedmodel
simplybecauseitcanbeviewedasamonotonetransformationofthechisquarevalue.
However,youshouldnotselectyourbestmodelbasedontheabsoluteindicessuchas
modelfitchisquarevalueortheSRMSRbecausetheseindicesdonottakemodel
parsimonyintoaccount.Complicatedmodelsmighthaveperfectmodelfitchisquare
andSRMSRvalues(i.e.,0).Butthesecomplexmodelsshouldnotbeselectedasthebest
modelsbecausetheyhaveverylittlescientificvalue.
TheAGFI,theRMSEA,theAIC,theCAIC,andtheSBCalltakesmodelparsimonyinto
account.FortheAGFI,thelargerthebetter.Forotherindices, thesmallerthebetter.All
theseparsimoniousindicespointtothecompletelyconstrainedmodelasthebest
multiplegroupmodelforthedata.
Lastly,theincrementalfitindexBentler CFIfavorsthecompletelyconstrainedmodeltoo.
However,virtuallyallmultiplegroupmodelinthiscomparisonareequallygood
accordingtotheCFI.NoticethatincrementalindicessuchastheCFImeasureshowa
targetmodelmeasuresbetterthanasocalledbaselinemodel.Theydonottakemodel
parsimonyintoaccount.Inaddition,theydependonhowgoodthe baselinemodelis
usedinthecomputingformula.Ifthebaselinemodelisverybad (suchasthecommonly
useduncorrelatedness model),allcompetingmodelswouldhavegoodincrementalfit
onlybecausethebaselinemodelismuchworse.Forthisreason, incrementalfitindices
mightnotserveasgoodcriteriaformodelselection.
101
102
102
1.
Freop60
lam2
lam3
Fair60
lam4
Legis60
Indust
Press65
Indlf60
Dem65
1.
lam2
Freop65
lam3
Fair65
lam4
Legis65
103
AnalyzingdirectandindirecteffectsissomethinguniquetoSEM.
Letuslookatthemodelforthedemocracyandindustrialization data.Onlythe
structuralpartoftheSEMisshowntoillustratetheidea.
103
Dem60
On democracy in 1965
A direct effect:
Indust --->Dem65
An indirect effect:
Indust --->Dem60 --->Dem65
Indust
Dem65
First,letuslookattheeffectofindustrializationonthedemocracymeasurein1960.
ThedirecteffectofIndust onDem60referstothepathIndust >Dem60.Thiseffect
canbeestimateddirectlyfromanySEMsoftware.
Onthedemocracymeasurein1965,industrializationhasadirect andindirecteffect.
ThedirecteffectreferstothepathIndust >Dem65.Theindirecteffectisindicatedby
thetrackIndust > Dem60>Dem65.
Whenyouaddupthedirectandindirecteffect,itgivesyouthe totaleffect.
InSEM,thedirecteffectsareestimatedasthepathcoefficients.Indirecteffectsand
totaleffectsarefunctionsoftheparameterestimates.Fortunately,PROCCALIScan
computethesefunctionsefficientlyanditcanalsoprovidestandarderrorestimatesfor
theseeffects.
104
S2
S3
P1
1.
P2
P3
1.
M1
M2
M3
1.
Social Status
Parental
Encouragement
Family Size
Achievement
Motivation
Mental Ability
1.
A1
A2
A3
105
Thisslideshowsamoreinterestingexampleaboutanalyzingdirect,indirect,andtotaleffects.
TheexampleisinspiredbyamodelofMarjoribanks (1974).Thecurrentmodelisasimplification
andthedataaregenerated.Theresultsheredonotrepresenttheoriginalstudy,butwould
servewellforillustrationpurposes.
Themainideaofthestudyistomodelthementalabilityofstudents.Thementalabilityisa
latentconstruct,whichissupposedtobedetermined(predicted) byparentalencouragement
andachievementmotivation,bothofwhichareformulatedaslatentconstructinthemodel.
Tworemotecauses(predictors),socialstatusandfamilysize,havedirecteffectsonparental
encouragementandachievementmotivation.However,thesetworemotecausesaffectthe
mentalabilityonlyindirectly.Socialstatusisalsoformulated asalatentvariable,whilefamily
sizeisanobservedvariable.Forallthelatentvariables,observedindicatorsareusedandthey
arerepresentedbysmallrectanglesinthepathdiagram.
Therearesomemotivatingquestionsaboutthispathdiagramregardingthedirectandindirect
effects.Forexample,
1.Eventhoughsocialstatusdoesnotaffectthementalability,it doeshaveanindirecteffecton
thementalabilityviaparentalencouragementandachievementmotivation.Onewouldlikethe
SEMsoftwaretocomputethisthisindirecteffectanditssignificance.
2.Parentalencouragementhasadirectandanindirecteffectsonthementalability.Whatisthe
overalltotaleffectofparentalencouragementonthementalability.Onewouldalsolikethe
SEMsoftwaretocomputealltheseeffectsandtheirsignificance.
105
ParentalEncouragement FamilySize
AchievementMotivation,
FamilySize
--->
AchievementMotivation,
ParentalEncouragement
--->
AchievementMotivation MentalAbility,
AchievementMotivation
--->
MentalAbility,
/* Measurement Model */
SocialStatus
---> S1 S2 S3
= 1.,
ParentalEncouragement ---> P1 P2 P3
= 1.,
AchievementMotivation ---> A1 A2 A3
= 1.,
MentalAbility
= 1.;
---> M1 M2 M3
Now,thePATHspecificationforthetargetmodelshouldbeeasyforyou.Youcan
specifythemeasurementmodelandthestructuralmodelbythemultiplepathsyntax.
Youcanlookatthepathdiagramandwritedownthepathsinthe PATHstatement.
Noticethateachpathinthepathdiagramrepresentsadirecteffectofonevariableon
anothervariable.
TheonlynewoptionintroducedhereistheEFFPARToptioninthe PROCCALIS
statement.EFFPARTstandsforeffectpartitioning.Inotherwords,itpartitionsthetotal
effectsofanyvariableonanyothervariableintodirectandindirecteffects.PROCCALIS
willcomputetheseeffectsandthestandardizedversionallwithstandarderror
estimatesprovided.
106
Fit Summary
Fit Summary
0.0936
0.7341
RMSEA Estimate
0.1431
0.8087
107
Themodelfitactuallydoesnotlooktoogoodforthissimulated data.Butthisisnotthe
concernhere.Wewanttostudytheeffectpartitioning,assuming thatwearesatisfied
withthemodelfit.
107
S2
1.
1.20**
S3
P1
1. .58**
.75**
Social Status
-.29*
Family Size
P2
.25**
P3
1.47**
Parental
Encouragement
.22*
1.57**
1.
A1
1.18**
A2
M2
1.
-1.73*
2.11**
M3
2.07**
Mental Ability
1.32**
Achievement
Motivation
-.13**
M1
Negative effect of
parental
encouragement on the
mental ability!
.60**
A3
* Marginally significant
** Significant
~ Fixed
108
Beforedivingintotheresultsfortheeffectpartitioning,Iwanttolookattheestimates
showninthepathdiagram.Iwanttothrowinonemoremotivationtostudydirectand
indirecteffectsinastructuralequationmodel.
Inthispathdiagram,estimatesareshownwiththeirsignificancemarked.Twoasterisks
afteranestimatemeanstheestimateisstatisticallysignificant.Oneasteriskafteran
estimatemeansthattheestimateismarginallysignificant.
Iwanttofocusontheeffectsofparentalencouragementonmentalability.Thedirect
effectis1.73.Thismeansthatparentalencouragementhasanegativeeffectonmental
ability.Thissoundsalittlestrangeatthefirstglance.Butifwelookatthebiggerpicture
inthepathdiagram,wecanunderstandwhythatisso.Noticethatparental
encouragementhasapositiveeffectonachievementmotivation,whichinturnshasa
positiveeffectonmentalability.Thewholepicturesuggeststhatpurelyparental
encouragementdonotnecessarilyaffectmentalabilityinapositiveway.Sometimes,
themoreencouragementwouldonlyaddmorepressuretotheindividualsmental
performancehencethenegativedirecteffectonmentalabilityobservedinthepath
diagramresult.However,whentheparentalencouragementcanaffectsomethingmore
internaloftheindividualsnamely,theindividualsachievementmotivation,thenitwill
resultinahighermentalabilityscore.Hence,thereisapositiveindirecteffectof
parentalencouragementonthementalability.
Insum,aninterestingquestioninthispathdiagramresultisthatwhatistheoverall
totaleffectofparentalencouragementonmentalability,giventhatithasanegative
directeffectandapositiveindirecteffect?
108
109
Beforeyoucananalyzedirectandindirecteffects,youshouldcheckwhethera
prerequisiteissatisfied.Inordertostudytheeffectpartitioninglegitimately,theso
calledstabilitycoefficientmustbelessthen1.PROCCALISprovidessuchacheck.When
youseethesemessagesfromthePROCCALISoutput,youcouldproceedtoexamine
youreffectpartitioningresults.
109
FamilySize
Achievement
Motivation
Mental
Ability
Parental
Encouragement
SocialStatus
A1
-0.1287
0.0360
-3.5769
0.000348
1.0000
1.5701
0.5176
3.0337
0.002416
0.6523
0.0942
6.9258
<.0001
A2
-0.1522
0.0420
-3.6274
0.000286
.
.
.
.
.
.
.
.
-0.1699
0.0572
-2.9696
0.002982
1.1821
0.1084
10.9012
<.0001
.
.
.
.
.
.
.
.
1.3196
0.4124
3.1995
0.001377
1.8560
0.6051
3.0672
0.002161
.
.
.
.
.
.
.
.
0.3376
0.4045
0.8346
0.4039
0.7710
0.1036
7.4398
<.0001
.
.
.
.
.
.
.
.
0.4244
0.1280
3.3159
0.000914
0.2516
0.0692
3.6356
0.000277
.
.
.
.
.
.
.
.
MentalAbility
ParentalEncouragement
.
.
.
.
.
.
.
.
0
Details
omitted.
110
WiththeEFFPARToption,PROCCALISproducestablesfortotal,direct,andindirect
effectsseparately.Thesetablescouldbelarge.Ijustannotate theseresultshere.Some
resultsarenotshown.
Thistableisabouttheestimatesofthetotaleffects,theirstandarderrors,tvalues,and
significancelevels.
110
FamilySize
Achievement
Motivation
Mental
Ability
Parental
Encouragement
SocialStatus
A1
1.0000
A2
1.1821
0.1084
10.9012
<.0001
.
.
.
.
.
.
.
.
1.3196
0.4124
3.1995
0.001377
.
.
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
-1.7343
0.9047
-1.9169
0.0553
.
.
.
.
.
.
.
.
MentalAbility
ParentalEncouragement
.
.
.
.
.
.
.
.
0
.
.
.
.
.
.
.
.
Details
omitted.
0.2516
0.0692
3.6356
0.000277
111
Thistableisaboutthedirecteffects,theirstandarderrors,tvalues,andsignificance
levels.
111
FamilySize
Achievement
Motivation
Mental
Ability
Parental
Encouragement
SocialStatus
A1
-0.1287
0.0360
-3.5769
0.000348
1.5701
0.5176
3.0337
0.002416
0.6523
0.0942
6.9258
<.0001
A2
-0.1522
0.0420
-3.6274
0.000286
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1.8560
0.6051
3.0672
0.002161
.
.
.
.
.
.
.
.
0.7710
0.1036
7.4398
<.0001
.
.
.
.
.
.
.
.
-0.1699
0.0572
-2.9696
0.002982
2.0719
1.0483
1.9763
0.0481
0.4244
0.1280
3.3159
0.000914
.
.
.
.
.
.
.
.
MentalAbility
ParentalEncouragement
Details
omitted.
112
Thistableisabouttheindirecteffects,theirstandarderrors,tvalues,andsignificance
levels.
112
Whenyouhavelargetableslikethoseshowninpreviousslides,youarelikelytobe
doingexploratoryanalysiswithoutspecificquestionsinyourmind.Theeffecttables
couldgetverylargeandyoumighthaveadifficulttimetolook fortheparticularresults
thatyouareinterestedin.Forexample,thecolumnsoftheeffecttablesconsistoffive
variables,eachofwhichservesasapredictoratleastonceinthepathdiagram.These
fivevariableshavedirectorindirecteffectsontherowvariables.Therowsconsistof
sixteenvariables,eachofwhichservesasanoutcomevariableatleastonce.Inthe
currentpathdiagram,itincludesallvariablesexceptfortheSocialStatusvariable.
However,ifyouhavespecificresearchquestionsinyourmind,youarerecommendedto
docustomizedeffectanalysisthatissupportedbyPROCCALIS.
113
.25**
Social Status
-.29*
Family Size
.22*
-.13**
Parental
Encouragement
1.57**
-1.73*
Mental Ability
1.32**
Achievement
Motivation
114
Inthebeginning,wealreadyhavethesemotivatingquestions.
1.Forthesocialstatusvariable,Whatarethetotal,direct,andindirecteffectsofthis
remotecauseonotherlatentconstructs,especiallyonthementalability?
2.Whatisthetotaleffectofparentalencouragementonthemental ability?
114
--->
ParentalEncouragement
--->
AchievementMotivation MentalAbility,
AchievementMotivation
--->
MentalAbility,
/* Measurement Model */
SocialStatus
---> S1 S2 S3
= 1.,
ParentalEncouragement ---> P1 P2 P3
= 1.,
AchievementMotivation ---> A1 A2 A3
= 1.,
MentalAbility
= 1.;
---> M1 M2 M3
The EFFPART
statement defines the
customized effects of
interest.
effpart
SocialStatus
PROCCALISsupportsthecustomizedeffectanalysis.ThiscanbedonebytheEFFPART
statement,asshowninthePROCCALIScodeinthisslide.
First,youwanttostudytheeffectpartitioningofsocialstatusonthesethreevariables:
parentalencouragement,achievementmotivation,andmentalability.Hence,youusethe
followingcodeintheEFFPARTstatement:
SocialStatus -> ParentalEncouragement AchievementMotivation
MentalAbility,
Second,youwanttostudytheeffectpartitioningofparentalencouragementonmental
ability.Hence,youusethefollowingcodeintheEFFPARTstatement:
ParentalEncouragement -> MentalAbility;
115
Total
Direct
Indirect
ParentalEncouragement
0.2516
0.2516
Direct effect
only
0.0692
0.0692
3.6356
3.6356
0.000277
0.000277
AchievementMotivation
0.6523
0.2193
0.4330
0.0942
0.1147
0.1203
6.9258
1.9125
3.5985
<.0001
0.0558
0.000320
0.4244
0.4244
MentalAbility
Indirect effect
only
0.1280
0.1280
3.3159
3.3159
0.000914
0.000914
116
TheeffectpartitioningresultsfromPROCCALISareshowninthisslideandthenextone.
Theeffectsofsocialstatusonthethreespecificlatentvariablesareshowninthistable.
Ontheparentalencouragement,socialstatushasadirecteffect only,whichispositive
andsignificant.
Ontheachievementmotivation,socialstatushasbothadirectandanindirecteffects.
Bothoftheseeffectsaresignificant.Thetotaleffectisthesumofthedirectandindirect
effect.Thetotaleffectisalsosignificant.
Onthementalability,socialstatushasonlyanindirecteffect,whichisalsosignificant.
116
Effects of ParentalEncouragement
Effect / Std Error / t Value / p Value
MentalAbility
Total
Direct
Indirect
0.3376
-1.7343
2.0719
0.4045
0.9047
1.0483
0.8346
-1.9169
1.9763
0.4039
0.0553
0.0481
117
Thisslideshowstheeffectpartitioningofparentalencouragementonmentalability.
Thedirecteffectisnegative,asshownpreviouslyinthepathdiagram.Thisdirecteffect
ismarginallysignificant.
Theindirecteffectispositiveandisstatisticalsignificant.Thisisapieceofcomforting
informationparentalencouragementdoesaffectthementalabilitypositively,butonly
throughitseffectonachievementmotivation.
Thetotaleffect,whichisthesumofdirectandindirecteffect,however,isnotsignificant.
ThisexampleshowsthatSEMeffectanalysiscanshowsomeeffect patternsthatsimply
cannotbeanalyzedbylinearregressionanalysisadequately.The SEMeffectanalysis
providessomethingmoredetailedandrefinedregardingthetotalityofthetheory.In
thisregard,thecustomizedeffectanalysissupportedbyPROCCALISisveryuseful.
117
ParentalEncouragement
AchievementMotivation
MentalAbility
Total
Direct
Indirect
0.6675
0.6675
0.0833
0.0833
8.0141
8.0141
<.0001
<.0001
0.6990
0.2350
0.0597
0.1207
0.1153
11.7105
1.9478
4.0231
<.0001
0.0514
<.0001
0.4960
0.4960
0.4640
0.0850
0.0850
5.8374
5.8374
<.0001
<.0001
118
PROCCALISalsoprovidesthestandardizedresultsforeffectanalysis.Standarderrors,t
values,andpvaluesarealsocomputedforthestandardizedeffectestimates.
118
MentalAbility
Total
Direct
Indirect
0.1487
-0.7639
0.9126
0.1705
0.3096
0.3487
0.8722
-2.4675
2.6171
0.3831
0.0136
0.008869
119
Thisslideshowsthestandardizedeffectsofparentalencouragementonmentalability.
Thepatternisquitesimilartotheunstandardizedversion.
119
Testing Specific
Hypotheses
120
120
v2
S1
v3
S2
1.
g2
S3
P1
g3
P2
P3
1.
M1
M2
1.
Social Status
Parental
Encouragement
Family Size
Achievement
Motivation
r2
M3
r3
Mental Ability
1.
A1
A2
A3
121
Testingspecifichypothesesisaninterestingtopic.Herewelookatsomeexamples.
Inthementalabilitymodel,youhavesomeindicatorvariablesforthelatentvariables.
Twolatentvariablesareselectedtoillustratethetestingofspecifichypotheses.
Forthementalabilityfactor,onemightwanttotestthehypothesisthattheloadings
(pathcoefficients)arethesamefortheM2andM3indicators.Inthepathdiagram,r2
andr3arelabeledasthepathcoefficients.Youwanttotestwhetherr2andr3areequal
withinthemodel.
Forthesocialstatusfactor,younotonlywanttotestthehypothesisthattheloadings
(pathcoefficients)arethesameforthethreeindicators,butyoualsowanttoseeiftheir
correspondingerrorvariancesarethesameinthepopulation.In thepathdiagram,g2,
g3,v1,v2,andv3areparametersofinterest.Youwanttotestsimultaneouslywhether
g2,g3areequalto1andv1,v2,andv3arethesameinthepopulation.
121
Specific Hypotheses
Parallel items for measuring SocialStatus
H1 : g2 = 1
H2: g3 = 1
H3: v1 = v2
H4: v2 = v3
122
Thetestofequalloadingsandequalerrorvariancesforthesocialstatusitemsisatest
ofparallelitems.Thiscouldbestatedmoreformallyasthefollowingfourcomponent
hypothesesH1,H2,H3,andH4,asshownintheslide.Thesefour hypothesesneedtobe
testedsimultaneously.Rejectionofthesimultaneoustestmeanstheitemsarenot
parallel.
Thetestofequalloadingsforthemeasurementindicatorsofthe mentalabilityfactoris
simpler.ItisstatedinH5.RejectionofH5meansthatr2andr3arenotequalinthe
population.
Finally,youcaninventanystrangehypothesisthatcanbeexpressedasacontinuous
functionofthemodelparameters.Forexample,H6statesthattheratioofthesumofr2
andr3tothesumofg2andg3is2.Thishypothesismayormaynotmakesense.Butitis
includedheretodemonstratetheflexibilityofPROCCALIS.
122
123
BeforeIshowyouthePROCCALIScode,itisusefultoreformulatethehypothesesinto
theformsthatmatchthePROCCALISinput.
PROCCALIStestshypothesesoftheformh()=0,whereh()isanycontinuousfunction
ofthemodelparameters(forexample,theerrorvariancesandthepathcoefficientsin
themodel).
Thehypothesesinthepreviousslidecouldallberewritteninthisrequiredform,as
showninthisslide.Withtheseforms,youarereadytospecifythosehypothesesin
PROCCALIS.
123
124
First,youhavetolabelornametheparametersinthecorrectlocationsofthemodel
specification.Forexample,g2andg3arethepathcoefficientsforS2andS3,respectively;andr2
andr3arethepathcoefficientsforM2andM3,respectively.Noticethatyoudidnotnamethese
parametersintheprecedingmodelspecifications.Namingtheseparameterswereoptional
becauseyoudidnotreferencethem.However,becauseyouaregoingtorefertothese
parametersinthehypothesistesting,youmustnameorlabelthemintherespectivelocations
now.Similarly,theerrorvariancesforS1S3arenamedasv1v3,asshowninthePVAR
statement.
ThemaintoolsfortestingspecifichypothesisinPROCCALISare theSIMTESTSandtheTESTFUNC
statements.
TheSIMTESTstatementenablesyoutotestsimultaneoushypothesesliketheparallelhypothesis
withfourcomponenthypotheses.Herewehaveh1,h2,h3,andh4, allofwhicharetreatedjust
asthenamesofthehypothesesthataredefinedlater.
TheTESTFUNCstatementenablesyoutotestindividualhypotheses liketheequalityofloadings
andtheproportionalityhypothesesdescribedpreviously.HereIuselongnamessuchas
h5_equal_load_m2_m3andh6_proportional_sumtoremindmeofthenatureofthetarget
hypotheses.
NowIusethesocalledSASprogrammingstatementstodefinethehypotheses:h1h4,
h5_equal_load_m2_m3,andh6_proportional_sum.TheSASprogrammingstatementsarejust
likecommonmathematicalequations.ThesesixSASprogrammingstatementsdefinethe
parametricfunctionsinthetargethypotheses.PROCCALIStestsallparametricfunctions
equalingzero.
124
Value
Standard
Error
t Value
p Value
0.04147
-0.27995
0.24290
1.02816
0.17072
-0.27228
0.8644
0.7854
125
TheTESTFUNCspecificationproducestheresultsshowninthistable.
YoufailtorejecttheequalityofloadingsforM2andM3becausethepvalueisbigger
than0.05.So,theequalityoftheloadingsissupported.
Youalsofailtorejecttheproportionalsumhypothesis(pvalue=0.79).
125
Parametric
Function
Function
Value
DF
Chi-Square
p Value
h1
h2
h3
h4
0.19873
-0.25004
-0.24067
-0.64108
4
1
1
1
1
24.23862
3.57013
8.37521
0.17774
1.49312
<.0001
0.0588
0.0038
0.6733
0.2217
parallel_social_items
Overall parallelism hypothesis is not supported for the SocialStatus items, although
the equality of error variances is supported.
126
TheSIMTESTSstatementspecificationproducestheoutputshowninthistable.
Fortheparallelhypothesis,thesimultaneoustestisrejected(p<.0001).Theparallel
itemhypothesisisnotsupported.
PROCCALISalsoprovidesindividualtestsforthecomponenthypotheses.Thiswouldbe
usefulfordoinganadhocanalysistoprobewhatfailsthesimultaneoushypothesis.For
example,bothh1andh2areatleastmarginallysignificant.But h3andh4arenot
significant.Recallthath1andh2areabouttheequalityofthe loadings(pathcoefficients)
whileh3andh4areabouttheequalityoferrorvariances.Thecurrentresultsshowthat
theitemsmighthavethesameerrorvariancesbutnotthesameloadingsinthe
population.
126
Model Modifications
127
127
196.7455
59
<.0001
0.0936
0.7341
0.1431
0.8087
Thementalabilitymodeldidnotfitwell.TheSRMSRandtheRMSEAarelarge,whilethe
AGFIandtheCFIaresmall.Whenyouencounterabadmodelfit, itwouldjeopardize
yourinterpretationsofthemodelparameters,effectanalysis,hypothesistesting,and
etc.
Modelmodificationisastatisticaltechniquethatsuggestsways toimproveyourmodel
fit.Themostcommonmodelmodificationtechniqueisdonethroughthesocalled
Lagrangemultiplier(LM)tests.Essentially,theLMtestssuggestwhichparametersyou
couldaddtothemodeltosignificantlylowerthemodelfitchisquarevalue.Whenthe
modelfitchisquareislowered,mostotherfitindices(butnotall,especiallythose
parsimoniousindicesthattakemodelcomplexityintoaccount)mightalsoimprove.
128
ParentalEncouragement FamilySize
FamilySize
AchievementMotivation,
AchievementMotivation,
--->
ParentalEncouragement
--->
AchievementMotivation MentalAbility,
AchievementMotivation
--->
MentalAbility,
SocialStatus
---> S1 S2 S3
= 1. ,
ParentalEncouragement
---> P1 P2 P3
= 1.,
AchievementMotivation
---> A1 A2 A3
= 1.,
MentalAbility
---> M1 M2 M3
= 1.;
run;
129
TheoptionyoucanusetodomodelmodificationinPROCCALISis theMODIFICATION
optioninthePROCCALISstatement.YoucansimplyaddthisoptiontothePROCCALIS
statementwhenyourunyourmodel.ThisexampleshowsthattheLMtestsformodel
modificationisrequestedfortheoriginalmentalabilitymodel, whichdoesnothavea
verygoodmodelfit.
129
Parm
To
From
LM Stat
Pr > ChiSq
Change
P2
P1
P1
56.19414
<.0001
-0.73639
P2
56.19349
<.0001
-0.72904
A2
M2
19.17647
<.0001
0.22842
A2
ParentalEncouragement
18.57947
<.0001
-2.31463
A2
MentalAbility
17.20340
<.0001
0.95581
ParentalEncouragement
A1
17.04464
<.0001
0.27042
A1
ParentalEncouragement
15.86099
<.0001
1.88904
FamilySize
A2
14.43548
0.0001
-1.14590
A1
MentalAbility
13.88705
0.0002
-0.75314
A2
P3
12.96818
0.0003
-0.57151
Adding the P2 <--- P1 (or P1 <--- P2) path reduces your model fit chi-square by 56 approximately.
Adding the A2 <--- M2 path reduces your model fit chi-square by 19 approximately.
130
PROCCALISoutputseveraltablesfortheLMtests.Theresultsareshownindifferent
tables,accordingthetypeoftheparameters.ThistableshowstherankingofLM
statisticsforaddingthe(singleheaded)pathsintothementalabilitymodel.Itgivesyou
thetenpathsthatcanimprovethemodelfitchisquarestatisticthemost.
Thetoponeisthep1>p2path.TheLMstatistic56.19meansthatifyouincludethis
pathintothemodel,youcanexpecttoreducethemodelfitchisquarebyabout56.This
isasubstantialimprovementbecauseyoucangetthisbigimprovementbyjustlosing
onedegreeoffreedom.Thesecondoneisthep2>p1path.Essentially,thiswillgive
thesameamountofmodelimprovementasthefirstpath.Thethirdoneisnotthat
dramatic,butstillgiveyouasubstantialimprovement.AddingtheM2>A2path
reducesthemodelfitchisquareby19.
Doyouwanttoaddthesepathsintoyourmodel?Letusdiscussthisafterweexamine
moreresultsabouttheLMtests.
130
Error
Error
of
of
LM Stat
Pr > ChiSq
Change
Parm
P2
P1
56.19312
<.0001
-1.96473
ParentalEncouragement
A1
12.26622
0.0005
0.46050
ParentalEncouragement
A2
12.08031
0.0005
-0.48351
FamilySize
A2
11.22650
0.0008
-1.88205
M2
A2
10.26408
0.0014
1.55895
S2
S1
7.78117
0.0053
1.55314
MentalAbility
A2
7.48800
0.0062
0.78161
AchievementMotivation
A1
6.95709
0.0083
-0.52904
P2
A3
6.54315
0.0105
0.76007
A3
A2
6.21429
0.0127
-0.67173
Adding the error covariance (P2 <---> P1) reduces your model fit chi-square by
56 approximately.
131
ThistableshowstheLMtests(statistics)fortheerrorvariancesandcovariances.Onthe
topofthelististhecovariancebetweentheerrorsofP2andP1.Addingthecovariance
betweentheerrorsofthesetwovariablesreducesthemodelfitchisquarestatisticby
56.ThisisactuallythesameimprovementthatwehaveseenforaddingeithertheP2
>P1or
P1>P2path.Thenextoneinthelisthasamuchlessimprovement. TheLMstatisticis
only12.26.
Forthisparticularlymodel,thesetwotablesareallthatPROCCALISproducesfortheLM
statistics.Thequestionnowiswhichparameterorparametersyouwanttoaddtothe
model.ThiscouldnotbeansweredbyjustlookingattheLMstatistics.Butitmightalso
involvesomejudgmentabouthowreasonabletheaddedparametersare.Dothese
addedparametersrenderyourmodeluninterpretable,orevencontradictorytoyour
theoreticalclaims,despitethefactthattheyimproveyourmodelfitsubstantially?
131
132
Beforegivingananswertothecurrentmodelmodificationanalysis,someimportant
generalpointsabouttheLMstatisticsarediscussed.
First,themodelfitchisquarereductionsasindicatedbytheLMteststatisticsareonly
linearapproximations.Thismeansthatifyouactuallyrefitthe modelbyaddingthe
suggestedparameter,theactualchisquarereductionmightbemoreorless.
Second,thechisquarereductionsassuggestedbytheLMteststatisticsarenot additive.
Thatmeansthatyoucannotaddtwoormoreparametersintothemodelandexpectthe
actualreductioninthenewmodelisexactlythesumofthecorrespondingLMstatistics.
Usually,theactualreductionwouldbesmaller(althoughitcouldbelarger).
Lastbutnotleast,modificationsuggestedbytheLMstatisticsmightnotbe
substantivelymeaningful.
Allthesethreepointsareimportantindecidingwhichparameter youwanttoaddtothe
currentmentalabilitymodelforimprovingthemodelfit.
132
S2
S3
P1
1.
P2
P3
1.
M1
M2
M3
1.
Social Status
Parental
Encouragement
Family Size
Achievement
Motivation
Mental Ability
1.
A1
A2
A3
133
ConsideringthetopsuggestionsfromtheresultsoftheLMteststatistics,Iwouldadd
thecovariancebetweenP1andP2.Theaddedparameterisshowninthepathdiagram.
Basically,allthetopLMsuggestions theP2>P1andP1>P2paths,andtheerror
covariancebetweenP1andP2arejustdifferentmanifestationsofthesamelackoffit
aboutacovarianceelementintheoriginalmodel.Thatis,thecovariancebetweenP1
andP2wasnotwellexplainedbytheoriginalmodel.Addingeitherofthesewillleadto
abetterfittingofthecovariancebetweenP1andP2.Inaddition,addingeitherofthese
willgiveyouanapproximatemodelfitchisquareimprovementof56.But,youwould
notgetthreetimesofthisamountbyaddingallthesethree.In fact,ifyouweretoadd
allthesethreeparameters,itisverylikelythatyourmodelis notidentified,meaning
thatyouwouldnotgetuniqueestimates.
Amongthetopthreechoices,theerrorcovarianceischosenbecausetheinterpretation
ofaddederrorcovarianceisalittlecleaner. P1andP2aremeasurementindicatorsof
thesamefactor(ParentalEncouragement).Theerrorcovarianceinterpretationisthat
thesetwoindicatorshavesomesortofcorrelationthatisunexplainedbytheircommon
factor.Theaddederrorcovariancerepresentsthecovarianceexplainedbysome
unknownsources.However,ifIweretoaddeithertheP1>P2orP2>P1paths,it
wouldcreatesomeconflictswithpurportedcommonfactorstructureforthetwo
indicatorvariables.
Notethatthecurrentconclusionisbasedonaverygeneralargumentthataimsat
preservingtheoriginalfactorvariablestructure.Itisnotauniversalprinciple.Inpractice,
youhavetoalsoconsiderthesubstantivegroundsoftheaddedparameters.
133
---> AchievementMotivation,
ParentalEncouragement
AchievementMotivation
---> MentalAbility,
SocialStatus
---> S1 S2 S3
= 1. ,
ParentalEncouragement
---> P1 P2 P3
= 1.,
AchievementMotivation
---> A1 A2 A3
= 1.,
MentalAbility
---> M1 M2 M3
= 1.;
pcov P1 P2;
run;
134
NowthatIhavedecidedtoaddthecovariancebetweenP1andP2, Irefitthemodelby
addingthePCOVstatementspecificationforthetwovariables,asshownintheSAS
codeinthisslide.IalsousetheMODIFICATIONoptiononemoretimetoseeifthere
couldbeanyfurthersuggestedimprovements.
134
After
Fit Summary
Chi-Square
Chi-Square DF
Fit Summary
196.7455
59
Chi-Square
110.6388
Chi-Square DF
58
Pr > Chi-Square
<.0001
Pr > Chi-Square
<.0001
0.0936
0.0661
0.7341
0.8062
RMSEA Estimate
0.1431
RMSEA Estimate
0.0892
0.8087
0.9269
135
BeforeyouaddtheerrorcovariancebetweenP1andP2,yourmodelfitchisquarewas
about197.Afteraddingthecovariance,themodelfitchisquareisabout111.This
improvementisactuallylargerthanwhattheLMstatisticsuggested,whichwas56.
Otherfitindicesalsoimprove.TheSRMSRandtheRMSEAarenowclosetobe
acceptable.TheAGFIandtheCFIareboostedtohigherlevels.
135
Parm
To
From
LM Stat
Pr > ChiSq
A2
A2
Change
M2
23.93741
<.0001
0.23830
ParentalEncouragement
22.22260
<.0001
-2.30398
A2
MentalAbility
21.70998
<.0001
0.90790
A1
ParentalEncouragement
19.34281
<.0001
1.96513
A1
MentalAbility
18.10545
<.0001
-0.75752
ParentalEncouragement
A1
17.30632
<.0001
0.32504
A1
M2
15.06992
0.0001
-0.17617
A2
FamilySize
15.06786
0.0001
-0.17675
FamilySize
A2
14.29265
0.0002
-0.89658
AchievementMotivation
A1
11.75569
0.0006
-0.35962
Adding the A2 <--- M2 path now reduces your model fit chi-square
by 24 (was 19 before adding the error covariance).
136
ThenewsetofLMtestsforpathssuggeststheadditionoftheM2>A2path.TheLM
statisticisabout24.Ifyoucomparethisresultwiththefirst LMresultsregardingthe
samepath,younoticethattheLMstatisticschangesasthefittedmodelchanges.
Previously,thesamepathhadanLMstatisticof19.Thisillustratesthenonlinearityand
nonadditivityoftheLMstatistics.
136
Error
Error
of
of
LM Stat
Pr > ChiSq
Change
Parm
ParentalEncouragement
A2
14.11717
0.0002
-0.53966
ParentalEncouragement
A1
14.00823
0.0002
0.51242
FamilySize
A2
13.96827
0.0002
-1.98162
M2
A2
11.86015
0.0006
1.65462
AchievementMotivation
A1
11.75570
0.0006
-0.57402
MentalAbility
A2
10.30441
0.0013
0.86542
P1
A1
9.02987
0.0027
0.53357
S2
S1
8.94243
0.0028
1.66876
MentalAbility
FamilySize
8.05365
0.0045
2.49300
MentalAbility
AchievementMotivation
6.94045
0.0084
0.98947
137
ThereisalsoanewsetofLMtestsforaddingerrorcovariances.
Youmightwanttoimproveyourmodelfurtherbyaddingsomeparametersfromthese
twoLMtables,althoughIwillnotattempttodomorehere.
137
Customized LM Tests
Principled modification process
Restrict the set of parameters of interest for
the LM tests
138
ModelmodificationbyusingtheMODIFICATIONoptioniskindofblindsearch
procedurethatyoutrytoimproveyourmodelwithoutanydefinitedirections.As
discussedbefore,theLMteststatisticsmightnotgiveyousuggestionsthatare
substantivelymeaningful.
However,insomeoccasionsyoumightwanttorestrictyourattentiontocertainsetof
potentialpathsorparametersinyourmodel,ratherthanallpossibleparameterspace
searchedbytheMODIFICATIONoption.
Ifyouwanttodosuchaprincipledmodificationprocess,youcanusethecustomizedLM
testssupportedinPROCCALIS.
138
---> AchievementMotivation,
ParentalEncouragement
AchievementMotivation
---> MentalAbility,
SocialStatus
---> S1 S2 S3
= 1. ,
ParentalEncouragement
---> P1 P2 P3
= 1.,
AchievementMotivation
---> A1 A2 A3
= 1.,
MentalAbility
---> M1 M2 M3
= 1.;
ThecustomizedLMtestsdefinesetsofparametersofinterestso thatyourmodel
modificationprocess(orLMstatisticsoutput)wouldbelimitedtothosesetsof
parameters.PROCCALISprovidestheLMTESTSstatementsyntaxtoachievethe
customizedLMtests.
Thementalabilitymodelisusedagain.ThistimeIdefinetwosetsofparametersof
interest.ThefirstsetofLMtestsiscalledcorr_err (itisjustanameyouassign).This
setofparameterscontainstheparameterregionCOVERR,whichis akeywordthat
denotesallerrorcovariancesinthemodel.ThesecondsetofLM testsiscalledpath
anameyouassign.Thissetofparametersdonotexhaustallpathsinthemodel.It
containstheparameterregionsLV>LVandLV>MV,whicharekeywordsthatdenotes
thelatentvariable(LV)tolatentvariable(LV)pathsandthelatentvariable(LV)to
manifestvariable(MV)paths,respectively.Therefore,thiscustomizedsetpath
excludespathsfromobservedvariablestoobservedvariables,or fromobserved
variablestolatentvariablessothatthefactorstructuresofthemodelcouldnotbe
potentiallydestroyedbyaddingthesepaths.TheLMtestsforthesepathsaresimplynot
includedintheresultsforthepath set.
139
Parm
Type
Var1
Var2
LM Stat
Pr > ChiSq
Change
COVERR
P2
COVERR
ParentalEncouragement
P1
56.19312
<.0001
-1.96473
A1
12.26622
0.0005
COVERR
0.46050
ParentalEncouragement
A2
12.08031
0.0005
-0.48351
COVERR
FamilySize
A2
11.22650
0.0008
-1.88205
COVERR
M2
A2
10.26408
0.0014
1.55895
COVERR
S2
S1
7.78117
0.0053
1.55314
COVERR
MentalAbility
A2
7.48800
0.0062
0.78161
COVERR
AchievementMotivation
A1
6.95709
0.0083
-0.52904
COVERR
P2
A3
6.54315
0.0105
0.76007
COVERR
A3
A2
6.21429
0.0127
-0.67173
140
ThistableshowsthecustomizedLMtestsoftheCORR_ERR set.Essentially,thistable
isthesameasoneofthestandardtablesproducedwiththeMODIFICATIONoption
becausebothtableshavethesameparameterregionCOVERR.
140
Parm
Type
Var1
Var2
LM Stat
Pr > ChiSq
Change
DV_DV
A2
DV_DV
A2
ParentalEncouragement
18.57947
<.0001
-2.31463
MentalAbility
17.20340
<.0001
DV_DV
0.95581
A1
ParentalEncouragement
15.86099
<.0001
1.88904
DV_DV
A1
MentalAbility
13.88705
0.0002
-0.75314
DV_DV
S2
MentalAbility
8.97262
0.0027
-0.38910
DV_DV
S3
AchievementMotivation
6.27192
0.0123
0.32436
DV_DV
S2
AchievementMotivation
6.19233
0.0128
-0.40928
DV_DV
ParentalEncouragement
MentalAbility
5.57439
0.0182
0.29376
DV_DV
ParentalEncouragement
AchievementMotivation
5.33788
0.0209
0.39218
DV_DV
FamilySize
AchievementMotivation
5.33730
0.0209
-1.20671
ThesecondcustomizedsetofLMtestssuggeststhataddingthedependentvariable(DV)
todependentvariable(DV)pathA2< ParentalEncouragementimprovesthemodelfit
themostamongstallpathsinthepath set.Thechisquareimprovementisabout19,
whichisstatisticallysignificant.
141
S2
S3
P1
1.
P2
P3
1.
M1
M2
M3
1.
Social Status
Parental
Encouragement
Family Size
Achievement
Motivation
Mental Ability
1.
A1
A2
A3
142
Addingthefirstpathsuggestedbythesecondsetofcustomizedsetisrepresentedby
thepathdiagramshownabove.ThepathinredshowsthatA2,whichisanindicatorof
AchievementMotivation,isnowalsoanindicatorofParentalEncouragement.Although
thefactorvariablefunctionalrelationshipispreservedinthissuggestedpathdiagram,
A2becomesfactoriallycomplex.ThealsoimpliesthatA2mightnothavebeenagood
(unique)measureofachievementmotivation.
142
---> AchievementMotivation,
ParentalEncouragement
AchievementMotivation
---> MentalAbility,
SocialStatus
---> S1 S2 S3
= 1. ,
ParentalEncouragement
---> P1 P2 P3 A2
= 1.,
AchievementMotivation
---> A1 A2 A3
= 1.,
MentalAbility
---> M1 M2 M3
= 1.;
run;
143
Nonetheless,youaddthisnewpathforA2,asshownintheabove PROCCALIScode.All
youneedtodoistoaddA2asoneoftheobservedindicatorsof the
ParentalEncouragementfactor.
143
After
Fit Summary
Chi-Square
Chi-Square DF
Pr > Chi-Square
Fit Summary
196.7455
59
Chi-Square
Chi-Square DF
168.2618
58
<.0001
Pr > Chi-Square
<.0001
0.0936
0.0896
0.7341
0.7675
RMSEA Estimate
0.1431
RMSEA Estimate
0.1291
0.8087
0.8468
144
Thesetwotablescomparethefitindicesbeforeandafteradding the
ParentalEncouragement>A2path.Themodelfitchisquareactuallydropsmorethan
19,whichwassuggestedbytheLMstatisticintheprecedingresults.Allotherfitindices
improvequiteabittoo.
Finally,acautionaboutallmodelmodification:youshouldvalidateyournewly
establishedmodelbynewdata.Thereasonisthatthemodelmodificationprocessis
subjecttothecapitalizationonchance.Usingaprincipledmodificationprocessbythe
customizedLMtestsmightnotavoidthechanceproblemcompletely.Confirmation
fromnewdataisalwaysrecommended.
144
145
Inthisworkshop,ImostlyusethePATHmodelinglanguagetofit SEM.Ialsobriefly
mentionedtheLISMODasaninterfacefortheLISRELmodel.There areactuallyquitea
fewmoremodelinglanguageinPROCCALIS:COSAN,FACTOR,LINEQS, MSTRUCT,and
RAM.Alloftheselanguagessupportmultiplegroupanalysisandmeanstructure
analysis.
IhavealsousedthedefaultML(maximumlikelihood)estimationmethodinthis
workshop,butPROCCALISsupportsmanyotherestimationmethodsaswell:GLS
(generalizeleastsquares),WLS(weightedleastsquares),ULS(unweightedleastsquares),
DWLS(diagonallyweightedleastsquares),andFIML(fullinformationmaximum
likelihood).
Ihaveonlyusedunstandardizedresultsinmostexamples,butPROCCALISalsoprovide
standardizedsolutionswithstandarderrorestimates.
Finally,analysisofmissingpatternswillbeavailablewiththe FIMLmethodinSAS/STAT
9.3.IhopetoaddmorefunctionalitiestoPROCCALISinthefuture.
145
Glossary
Manifest Observed variables (measured variables) in the data set.
Latent Unobserved variables.
Endogenous Dependent /mediating variables; at least one single-headed arrow points to it;
used as an outcome variable in an equation; can also be a predictor variable in other
equations.
Exogenous Independent variables; no single-headed arrows point to it; never used as an
outcome variable in the model; used only as a predictor in the model.
Factor A latent (unmeasured) variable that is treated as a hypothetical construct
(systematic source) in the model.
Error An exogenous term for uncertainty (unsystematic source) associated with an
endogenous manifest variable (or any endogenous variable, in a more general definition).
Disturbance An exogenous term for uncertainty (unsystematic source) associated with an
endogenous latent variable.
Path diagram representation
Rectangles: Observed / manifest variables.
Ovals / circles : Latent variables (factors, errors, and disturbances). Errors and
disturbances are not necessarily put into ovals/circles.
146
146
Glossary
Single-headed arrows: Directed paths, direct effects, path coefficients; specified in the
PATH statement.
Double-headed arrows that point to individual variables: Variance parameters of
exogenous variables or error variance parameters of endogenous variables; specified in
the PVAR statement.
Double-headed arrows that point to two distinct variables: Covariance parameters
between exogenous variables or error covariance parameters between endogenous
variables; specified in the PCOV statement.
Fit assessment
model fit chi-square statistic: Nonsignificance means that the theoretical model is
supported; not a very practical index because it almost always rejects all approximating
models that are practically useful.
AGFI (adjusted goodness-of-fit index) and Bentlers CFI (comparative fit index): Two
popular fit indices that indicate good model fit when their values are above 0.9.
SRMR (standardized root mean square residual) and RMSEA (root mean squared error
approximation): Two popular fit indices that indicate good model fit when their values
are below 0.05.
AIC, CAIC, and SBC: Information criteria for comparing competing models. The smaller
the better.
147
147
References
Arbuckle, J. (2008), AMOS 17.0 Users Guide, Crawfordville, FL: Amos Development
Corporation.
Bentler, P. M. (1985), Theory and Implementation of EQS: A Structural Equations
Program, Manual for Program Version 2.0, Los Angeles: BMDP Statistical Software.
Bentler, P. M. (1995), EQS, Structural Equations Program Manual, Program Version 5.0,
Encino, CA: Multivariate Software.
Bollen, K. A. (1989), Structural Equations with Latent Variables, New York: John Wiley
& Sons.
Fuller, W. A. (1987), Measurement Error Models, New York: John Wiley and Sons.
Holzinger, K. J. and Swineford, F. A. (1939), A Study in Factor Analysis: The Stability
of a Bi-Factor Solution, Supplementary Educational Monographs, No. 48. Chicago:
University of Chicago,Dept. of Education.
Jreskog, K. G. (1973), A General Method for Estimating a Linear Structural Equation
System, in A. S. Goldberger and O. D. Duncan, eds., Structural Equation Models in
the Social Sciences, New York: Academic Press.
148
148
References
Keesling, J. W. (1972), Maximum Likelihood Approaches to Causal Analysis, Ph.D. thesis,
University of Chicago.
Marjoribanks, K., ed. (1974), Environments for Learning, London: National Foundation
for Educational Research Publications.
McDonald, R. P. (1978), A Simple Comprehensive Model for the Analysis of Covariance
Structures, British Journal of Mathematical and Statistical Psychology, 31, 5972.
McDonald, R. P. (1980), A Simple Comprehensive Model for the Analysis of Covariance
Structures: Some Remarks on Applications, British Journal of Mathematical and
Statistical Psychology, 33, 161183.
Nowak, T. P., Hoffman, D. L., and Yung, Y. F. (2000), Measuring the Customer
Experience in Online Environments: A Structural Modeling Approach, Marketing
Science, 19(1), 2242.
Wiley, D. E. (1973), The Identification Problem for Structural Equation Models with
Unmeasured Variables, in A. S. Goldberger and O. D. Duncan, eds., Structural
Equation Models in the Social Sciences, New York: Academic Press.
149
149