6/11/2015 StructuralEquationModelingStatwiki

StructuralEquationModeling
FromStatwiki

Structuralequationmodeling(SEM)growsoutofandservespurposessimilartomultipleregression,butinamore
powerfulwaywhichtakesintoaccountthemodelingofinteractions,nonlinearities,correlatedindependents,measurement
error,correlatederrorterms,multiplelatentindependentseachmeasuredbymultipleindicators,andoneormorelatent
dependentsalsoeachwithmultipleindicators.SEMmaybeusedasamorepowerfulalternativetomultipleregression,path
analysis,factoranalysis,timeseriesanalysis,andanalysisofcovariance.Thatis,theseproceduresmaybeseenasspecial
casesofSEM,or,toputitanotherway,SEMisanextensionofthegenerallinearmodel(GLM)ofwhichmultiple
regressionisapart.http://www.pire.org/

SEMisanumbrellaconceptforanalysessuchasmediationandmoderation.Thiswikipageprovidesgeneralinstructionand
guidanceregardinghowtowritehypothesesfordifferenttypesofSEMs,whattodowithcontrolvariables,mediation,
interaction,multigroupanalyses,andmodelfitforstructuralmodels.Videosandslidespresentationsareprovidedinthe
subsections.

Doyouknowofsomecitationsthatcouldbeusedtosupportthetopicsandproceduresdiscussedinthissection?
Pleaseemailthemtome(mailto:[email protected])withthenameofthesection,procedure,orsubsectionthat
theysupport.Thanks!

Contents
1Hypotheses
1.1Directeffects
1.2Mediatedeffects
1.3Interactioneffects
1.4Multigroupeffects
1.5MediatedModeration
1.6Handlingcontrols
1.7SupportingHypotheses
2Controls
3Mediation
3.1Concept
3.2Types
4Interaction
4.1Concept
4.2Types
5Modelfitagain
6Multigroup
7FromMeasurementModeltoStructuralModel
8CreatingCompositesfromLatentFactors

Hypotheses
Hypothesesareakeystonetocausaltheory.However,wordinghypothesesisclearlyastruggleformanyresearchers(just
selectatrandomanyarticlefromagoodacademicjournal,andcountthewordingissues!).InthissectionIofferexamples
ofhowyoumightworddifferenttypesofhypotheses.Theseexamplesarenotexhaustive,buttheyaresafe.

Directeffects

"Diethasapositiveeffectonweightloss"

"Anincreaseinhoursspentwatchingtelevisionwillnegativelyeffectweightloss"

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 1/7
6/11/2015 StructuralEquationModelingStatwiki

Mediatedeffects

Formediatedeffects,besuretoindicatethedirectionofthemediation(positiveornegative),thedegreeofthemediation
(partial,full,orsimplyindirect),andthedirectionofthemediatedrelationship(positiveornegative).

"Exercisepositivelyandpartiallymediatesthepositiverelationshipbetweendietandweightloss"

"Televisiontimepositivelyandfullymediatesthepositiverelationshipbetweendietandweightloss"

"Dietaffectsweightlosspositivelyandindirectlythroughexercise"

Interactioneffects

"Exercisepositivelymoderatesthepositiverelationshipbetweendietandweightloss"

"Exerciseamplifiesthepositiverelationshipbetweendietandweightloss"

"TVtimenegativelymoderates(dampens)thepositiverelationshipbetweendietandweightloss"

Multigroupeffects

"BodyMassIndex(BMI)moderatestherelationshipbetweenexerciseandweightloss,suchthatforthosewithalowBMI,
theeffectisnegative(i.e.,yougainweightmusclemass),andforthosewithahighBMI,theeffectispositive(i.e.,
exercisingleadstoweightloss)"

"Agemoderatestherelationshipbetweenexerciseandweightloss,suchthatforage<40,thepositiveeffectisstrongerthan
forage>40"

"Dietmoderatestherelationshipbetweenexerciseandweightloss,suchthatforwesterndietstheeffectispositiveand
weak,foreastern(asia)diets,theeffectispositiveandstrong"

MediatedModeration

Anexampleofamediatedmoderationhypothesiswouldbesomethinglike:

Ethicalconcernsstrengthenthenegativeindirecteffect(throughburnout)betweencustomerrejectionandjobsatisfaction.

Inthiscase,theIViscustomerrejection,theDVisjobsatisfaction,burnoutisthemediator,andthemoderatorisethical
concerns.Themoderationisconductedthroughaninteraction.However,ifyouhaveacategoricalmoderator,itwouldbe
somethingmorelikethis(usinggenderasthemoderator):

Thenegativeindirecteffectbetweencustomerrejectionandjobsatisfaction(throughburnout)isstrongerformenthanfor
women.

Handlingcontrols

Whenincludingcontrolsinhypotheses(yes,youshouldincludethem),simplyaddattheendofanyhypothesis,"when
controllingfor...[listcontrolvariableshere]"Forexample:

"ExercisepositivelymoderatesthepositiverelationshipbetweendietandweightlosswhencontrollingforTVtimeanddiet"

"DiethasapositiveeffectonweightlosswhencontrollingforTVtimeanddiet"

Anotherapproachistostatesomewhereaboveyourhypotheses(whileyou'resettingupyourtheory)thatallyour
hypothesestakeintoaccounttheeffectsofthefollowingcontrols:A,B,andC.Andthenmakesuretoexplainwhy.

SupportingHypotheses

Gettingthewordingrightisonlypartofthebattle,andismostlyuselessifyoucannotsupportyourreasoningforWHYyou
thinktherelationshipsproposedinthehypothesesshouldexist.SimplysayingXhasapositiveeffectonYisnotsufficient
tomakeacausalstatement.Youmustthengoanexplainthevariousreasonsbehindyourhypothesizedrelationship.Take

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 2/7
6/11/2015 StructuralEquationModelingStatwiki

DietandWeightlossforexample.Thehypothesisis,"Diethasapositiveeffectonweightloss".Thesupportinglogicwould
thenbesomethinglike:

Weightisgainedasweconsumecalories.Dietreducesthenumberofcaloriesconsumed.Therefore,themorewediet,
themoreweightweshouldlose(orthelessweightweshouldgain).

Controls
LESSON:Controls(http://www.kolobkreations.com/Controls.pptx)

Controlsarepotentiallyconfoundingvariablesthatweneedtoaccountfor,butthatdontdriveourtheory.Forexample,in
DietzandGortmaker1985,theirtheorywasthatTVtimehadanegativeeffectonschoolperformance.Buttherearemany
thingsthatcouldeffectschoolperformance,possiblyevenmorethantheamountoftimespentinfrontoftheTV.So,in
ordertoaccountfortheseotherpotentiallyconfoundingvariables,theauthorscontrolforthem.Theyarebasicallysaying,
thatregardlessofIQ,timespentreadingforpleasure,hoursspentdoinghomework,ortheamountoftimeparentsspend
readingtotheirchild,anincreaseinTVtimestillsignificantlydecreasesschoolperformance.Theserelationshipsareshown
inthefigurebelow.

Asacautionarynote,youshouldnearlyalwaysincludesomecontrolshowever,thesecontrolvariablesstillcountagainst
yoursamplesizecalculations.So,themorecontrolsyouhave,thehigheryoursamplesizeneedstobe.Alsoyougeta
higherRsquarebutwithincreasinglysmallergainsforeachaddedcontrol.Sometimesyoumayevenfindthataddinga
controldrownsoutalltheeffectsoftheIVs,insuchacaseyoumayneedtorunyourtestswithoutthatcontrolvariable
(butthenyoucanonlysaythatyourIVs,thoughsignificant,onlyaccountforasmallamountofthevarianceintheDV).
Withthatinmind,youcantandshouldn'tcontrolforeverything,andasalways,yourdecisiontoincludeorexclude
controlsshouldbebasedontheory.

HandlingcontrolsinAMOSiseasy,butmessy(seethefigurebelow).Yousimplytreatthemliketheotherexogenous
variables(theonesthatdonthavearrowsgoingintothem),andhavethemregressonwhicheverendogenousvariablesthey
maylogicallyaffect.Inthiscase,IhavevalShort,apotentiallyconfoundingvariable,asacontrol,withregardstovalLong.
AndIhaveLoyRepeatasacontrolonLoyLong.IvealsocovariedtheControlswitheachotherandwiththeother
exogenousvariables.Whenusingcontrolsinamoderatedmediationanalysis,goaheadandputthecontrolsinatthevery
beginning.Covaryingcontrolvariableswiththeotherexogenousvariablescanbedonebasedontheory,ratherthanas
default.

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 3/7
6/11/2015 StructuralEquationModelingStatwiki

Whenreportingthemodel,youdoneedtoincludethecontrolsinallyourtestsandoutput,butyoushouldconsolidatethem
atthebottomwheretheycanbeoutoftheway.Also,justsoyoudontgetanycrazyideas,youwouldnottestforany
mediationbetweenacontrolandadependentvariable.However,youmayreporthowthecontroleffectsadependent
variabledifferentlybasedonamoderatingvariable.Forexample,valshortmayhaveastrongereffectonvalLongformales
thanforfemales.Thisissomethingthatshouldbereported,butnotnecessarilyfocusedon,asitisnotlikelyakeypartof
yourtheory.Lastly,evenifeffectsfromcontrolsarenotsignificant,youdonotneedtrimthemfromyourmodel(although,
therearealsootherschoolsofthoughtonthisissue).

Mediation
Lesson:TestingMediationusingBootstrapping
(http://www.kolobkreations.com/Mediation%20Step%20by%20Step%20with%20Bootstrapping.pptx)
VideoLecture:ASimpleGuidetoMediation(http://www.youtube.com/watch?v=9mf7nIAlH5c)
VIDEOTUTORIAL:MediationinAMOS(http://youtu.be/0artfnxyF_A)
Hairetal.:pp.751755

Concept

Mediationmodelsareusedtodescribechainsofcausation.Mediationisoftenusedtoprovideamoreaccurateexplanation
forthecausaleffecttheantecedenthasonthedependentvariable.Themediatorisusuallythatvariablethatisthemissing
linkinachainofcausation.Forexample,Intelligenceleadstoincreasedperformancebutnotinallcases,asnotall
intelligentpeoplearehighperformers.Thus,someothervariableisneededtoexplainthereasonfortheinconsistent
relationshipbetweenIVandDV.Thisothervariableiscalledamediator.Inthisexample,workeffectiveness,maybeagood
mediator.Wewouldsaythatworkeffectivenessfullyandpositivelymediatestherelationshipbetweenintelligenceand
performance.Thus,thedirectrelationshipbetweenintelligenceandperformanceisbetterexplainedthroughthemediatorof
workeffectiveness.Thelogicis,evenifyouareintelligent,ifyoudon'tworksmarter,thenyouwon'tperformwell.
However,intelligentpeopletendtoworksmarter(butnotalways).Thus,whenintelligenceleadstoworkingsmarter,then
weobservegreaterperformance.

Types

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 4/7
6/11/2015 StructuralEquationModelingStatwiki

Therearethreemaintypesofsimplemediation:1)partial,2)full,and3)indirect.Partialmediationmeansthatboththe
directandindirecteffectsfromtheIVtoDVaresignificant.Fullmeansthatthedirecteffectdropsoutofsignificancewhen
themediatorisadded,andthattheindirecteffectissignificant.Indirectmeansthatthedirecteffectneverwassignificant,
butthattheindirecteffectis.Thefigurebelowillustratesthesetypesofmediation.Pleaserefertothestepbystepguide
listedabovefordeterminingsignificanceofthemediation.

Thereisonelesscommonformofmediationcalled"competitive"mediation.Inthiscase,thedirecteffectbetweenIVand
DVisneutralizedwhenthemediatorisabsent.Whenthemediatorisadded,thedirecteffectbecomessignificant(oftento
theresearcherssurprise)andisusuallyintheoppositedirectiontheorized,whiletheindirectpathisobservedtobe
significantandinthetheorizeddirection.Insuchcases,theIVhasdualeffectsontheDVwhichcanonlybeseparatedwhen
themediator,actingsomewhatlikeaprism,bifurcatesthecompeting(andneutralizing)effects.Zhaoetal.(2010
"reconsideringBaronandKenny")discussthis.

Interaction
VIDEOTUTORIAL:TestingInteractionEffects(http://www.youtube.com/watch?v=K34sF_AmWio)
LESSON:InteractionEffects(http://www.kolobkreations.com/Interaction.pptx)

Concept

Infactorialdesigns,interactioneffectsarethejointeffectsoftwopredictorvariablesinadditiontotheindividualmain
effects.Thisisanotherformofmoderation(alongwithmultigrouping)i.e.,theXtoYrelationshipchangesform(gets
stronger,weaker,changessigns)dependingonthevalueofanotherexplanatoryvariable(themoderator).So,forexample

youlose1poundofweightforeveryhouryouexercise
youlose1poundofweightforevery500caloriesyoucutbackfromyourregulardiet
butwhenyouexercisewhiledieting,theyoulose2poundsforevery500caloriesyoucutbackfromyourregulardiet,
inadditiontothe1poundyouloseforexercisingforonehourthusintotal,youlosethreepounds

So,themultiplicativeeffectofexercisingwhiledietingisgreaterthantheadditiveeffectsofdoingoneortheother.Hereis
anothersimpleexample:

Chocolateisyummy
Cheeseisyummy
butcombiningchocolateandcheeseisyucky!

Thefollowingfigureisanexampleofasimpleinteractionmodel.

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 5/7
6/11/2015 StructuralEquationModelingStatwiki

Types

InteractionsenablemorepreciseexplanationofcausaleffectsbyprovidingamethodforexplainingnotonlyhowXaffects
Y,butalsounderwhatcircumstancestheeffectofXchangesdependingonthemoderatingvariableofZ.Interpreting
interactionsissomewhattricky.Interactionsshouldbeplotted(asdemonstratedinthetutorialvideo).Onceplotted,the
interpretationcanbemadeusingthefollowingfourexamples(inthefiguresbelow)asaguide.MymostrecentStatsTools
Packageprovidestheseinterpretationsautomatically.

Modelfitagain
YoualreadydidmodelfitinyourCFA,butyouneedtodoitagaininyourstructuralmodelinordertodemonstrate
sufficientexplorationofalternativemodels.Themethodisthesame:lookatmodificationindices,residuals,andstandardfit
measureslikeCFI,RMSEAetc.Theonethingthatshouldbenotedhereinparticular,however,islogicthatshould
determinehowyouapplythemodificationindicestoerrorterms.

Ifthecorrelatedvariablesarenotlogicallycausallycorrelated,butmerelystatisticallycorrelated,thenyoumay
covarytheerrortermsinordertoaccountforthesystematicstatisticalcorrelationswithoutimplyingacausal
relationship.
e.g.,burnoutfromcustomersishighlycorrelatedwithburnoutfrommanagement
Weexpectthesetohavesimilarvalues(residuals)becausetheyarelogicallysimilarandhavesimilarwording
inoursurvey,buttheydonotnecessarilyhaveanycausalties.
Ifthecorrelatedvariablesarelogicallycausallycorrelated,thensimplyaddaregressionline.
http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 6/7
6/11/2015 StructuralEquationModelingStatwiki

e.g.,burnoutfromcustomersishighlycorrelatedwithsatisfactionwithcustomers
WeexpectburnCtopredictsatC,sonotaccountingforitisnegligent.

Lastly,remember,youdon'tneedtocreatetheBESTfit,justgoodfit.IfaBESTfitmodel(i.e.,oneinwhichall
modificationindicesareaddressed)isn'tlogical,ordoesnotfitwithyourtheory,youmayneedtosimplysettleforamodel
thathasworse(yetsufficient)fit,andthenexplainwhyyoudidnotchoosethebetterfittingmodel.Formoreinformationon
whenitisokaytocovaryerrorterms(becausethereareotherappropriatereasons),refertoDavidKenny'sthoughtsonthe
matter:David'swebsite(http://davidakenny.net/cm/respec.htm)

Multigroup
VIDEOTUTORIAL:TestingMultigroupModerationusingChisquaredifferencetest(prettycomplex...)
(http://www.youtube.com/watch?v=mirI5ETQRTA)
VIDEOTUTORIAL:TestingMultigroupModerationusingCriticalRatios(Theeasyway)
(http://www.youtube.com/watch?v=ZMYS90AU8bs)
LESSON:MediationversusModeration(http://www.kolobkreations.com/Mediation%20and%20Multi
group%20Moderation.pptx)

Multigroupmoderationisaspecialformofmoderationinwhichadatasetissplitalongvaluesofacategoricalvariable
(suchasgender),andthenagivenmodelistestedwitheachsetofdata.Usingthegenderexample,themodelistestedfor
malesandfemalesseparately.Theuseofmultigroupmoderationistodetermineifrelationshipshypothesizedinamodel
willdifferbasedonthevalueofthemoderator(e.g.,gender).Takethedietandweightlosshypothesisforexample.Amulti
groupmoderationmodelwouldanswerthequestion:doesdietingeffectweightlossdifferentlyformalesthanforfemales?
Inthevideosabove,youwilllearnhowtosetupamultigroupmodelinAMOS,andtestitusingchisquaredifferences,and
usingcriticalratios.Really,usingcriticalratiostakesaboutaoneminuteafterthemodelissetup,anditinvolvesnoroom
forhumanerror,whereasusingthechisquaremethodcantakeupwardsof30minutesanditinvolvesalotofroomfor
humanerror.So,Irecommendtheeasymethod!

FromMeasurementModeltoStructuralModel
VIDEOTUTORIAL:FromCFAtoSEMinAMOS(http://www.youtube.com/watch?v=nULF6BGVw0)

Manyoftheexamplesinthevideossofarhavetaughtconceptsusingasetofcompositevariables(insteadoflatentfactors
withobserveditems).ManywillwanttoutilizethefullpowerofSEMbybuildingtruestructuralmodels(withlatent
factors).Thisisnotadifficultthing.Simplyremovethecovariancearrowsfromyourmeasurementmodel(afterCFA),then
drawsingleheadedarrowsfromIVstoDVs.MakesureyouputerrortermsontheDVs,thenrunit.It'sthateasy.Referto
thevideoforademonstration.

CreatingCompositesfromLatentFactors
VIDEOTUTORIAL:ImputingCompositeVariablesinAMOS(http://www.youtube.com/watch?
v=dsOS9tQjxW8)

Ifyouwouldliketocreatecompositevariables(asusedinmanyofthevideos)fromlatentfactors,itisaneasythingtodo.
However,youmustremembertwoveryimportantcaveats:

Youarenotallowedtohaveanymissingvaluesinthedataused.ThesewillneedtobeimputedbeforehandinSPSS
orexcel(IhavetwotoolsforthisinmyStatsToolsPackageoneforimputing,andoneforsimplyremovingthe
entirerowthathasmissingdata).
Latentfactornamesmustnothaveanyspacesorhardreturnsinthem.Theymustbesinglecontinuousstrings
("FactorOne"or"Factor_One"insteadof"FactorOne").

Afterthosetwocaveatsareaddressed,thenyoucansimplygototheAnalyzemenu,andselectDataImputation.Select
RegressionImputation,andthenclickontheImputebutton.ThiswillcreateanewSPSSdatasetwiththesamenameasthe
currentdatasetexceptitwillbefollowedbyan"_C".Thiscanbefoundinthesamefolderasyourcurrentdataset.

Retrievedfrom"http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&oldid=5140"

Thispagewaslastmodifiedon22September2014,at13:48.

http://statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling&printable=yes 7/7