关于部分中介与完全中介作用与课堂定义相同.docx
《关于部分中介与完全中介作用与课堂定义相同.docx》由会员分享,可在线阅读,更多相关《关于部分中介与完全中介作用与课堂定义相同.docx(8页珍藏版)》请在冰豆网上搜索。
关于部分中介与完全中介作用与课堂定义相同
http:
//203.208.35.101/search?
q=cache:
IERvvryqTfQJ:
来源:
google网页快照
DavidA.Kenny
January10,2008
Pleaseemailmesuggestionsandcorrections.
MEDIATION
WhatisMediation?
Baron&KennySteps
MeasuringMediationortheIndirectEffect
DesignIssues
SpecificationError
Extensions
LinkstoOtherSites
References
WhatIsMediation?
ConsideravariableXthatisassumedtoaffectanothervariableY. ThevariableXiscalledtheinitialvariableandthevariablethatitcausesorYiscalledtheoutcome. Indiagrammaticform,theunmediatedmodelis图片显示不出来
TheeffectofXonYmaybemediatedbyaprocessormediatingvariableM,andthevariableXmaystillaffectY. Pathciscalledthetotaleffect. Themediatedmodelis图片显示不出来
(Thesetwodiagramsareessentialtotheunderstandingofthispage. Pleasestudythemcarefully!
) Pathc'iscalledthedirecteffect. Themediatorhasbeencalledaninterveningorprocessvariable. CompletemediationisthecaseinwhichvariableXnolongeraffectsYafterMhasbeencontrolledandsopathc'iszero.PartialmediationisthecaseinwhichthepathfromXtoYisreducedinabsolutesizebutisstilldifferentfromzerowhenthemediatoriscontrolled.
Notethatamediationalmodelisacausalmodel. Forexample,themediatorispresumedtocausetheoutcomeandnotviceversa. Ifthepresumedmodelisnotcorrect,theresultsfromthemediationalanalysisareoflittlevalue. Mediationisnotdefinedstatistically;ratherstatisticscanbeusedtoevaluateapresumedmediationalmodel. Thereadershouldconsultthesectionbelowon SpecificationError.
Thereisalonghistoryinthestudyofmediation(Hyman,1955;MacCorquodale&Meehl,1948). Currentlymediationisaverypopulartopic. (Thispageaveragesover100differentvisitorsaday.) Thereareseveralreasonsfortheintenseinterestinthistopic. Onereasonfortestingmediationistryingtounderstandthemechanismthroughwhichtheinitialvariableaffectstheoutcome. Mediation(andmoderation)analysisareakeypartofwhathasbeencalledprocessanalysis. Moreoverwhenmostcausalorstructuralmodelsareexamined,themediationalpartofthemodelisthemostinteresting.
BaronandKennySteps
Ifthemediationalmodel(seeabove)iscorrectlyspecified,thepaths(c,a,b,andc')canbeestimatedbymultipleregression,sometimescallordinaryleastsquaresorOLS. Asdiscussedlater,othermethodsofestimation(e.g.,logisticregressionandstructuralequalmodeling)canbeused. Regardlessofwhichdataanalyticmethod(thegeneralassumptiononthispageisthatitismultipleregression)isused,thestepsnecessaryfortestingmediationarethesame. Thissectiondescribestheanalysesrequiredfortestingmediationalhypotheses[previouslypresentedbyBaronandKenny(1986)andJuddandKenny(1981)].
BaronandKenny(1986)andJuddandKenny(1981)havediscussedfourstepsinestablishingmediation:
Step1:
Showthattheinitialvariableiscorrelatedwiththeoutcome. UseYasthecriterionvariableinaregressionequationandXasapredictor(estimateandtestpathc).Thisstepestablishesthatthereisaneffectthatmaybemediated.
Step2:
Showthattheinitialvariableiscorrelatedwiththemediator. UseMasthecriterionvariableintheregressionequationandXasapredictor(estimateandtestpatha). Thisstepessentiallyinvolvestreatingthemediatorasifitwereanoutcomevariable.
Step3:
Showthatthemediatoraffectstheoutcomevariable. UseYasthecriterionvariableinaregressionequationandXandMaspredictors(estimateandtestpathb). Itisnotsufficientjusttocorrelatethemediatorwiththeoutcome;themediatorandtheoutcomemaybecorrelatedbecausetheyarebothcausedbytheinitialvariableX. Thus,theinitialvariablemustbecontrolledinestablishingtheeffectofthemediatorontheoutcome.
Step4:
ToestablishthatMcompletelymediatestheX-Yrelationship,theeffectofXonYcontrollingforM(pathc')shouldbezero(seediscussiononsignificancetesting). TheeffectsinbothSteps3and4areestimatedinthesameequation.
Ifallfourofthesestepsaremet,thenthedataareconsistentwiththehypothesisthatvariableMcompletelymediatestheX-Yrelationship,andifthefirstthreestepsaremetbuttheStep4isnot,thenpartialmediationisindicated. Meetingthesestepsdoesnot,however,conclusivelyestablishthatmediationhasoccurredbecausethereareother(perhapslessplausible)modelsthatareconsistentwiththedata. SomeofthesemodelsareconsideredlaterintheSpecificationErrorsection.
Notethatthestepsarestatedintermsofzeroandnonzerocoefficients,notintermsofstatisticalsignificance,astheywereinBaronandKenny(1986). Becausetriviallysmallcoefficientscanbestatisticallysignificantwithlargesamplesizesandverylargecoefficientscanbenonsignificantwithsmallsamplesizes,thestepsshouldnotbedefinedintermsofstatisticalsignificance. Statisticalsignificanceisinformative,butotherinformationshouldbepartofstatisticaldecisionmaking. Forinstance,considerthecaseinwhichaislarge,biszero,andsoc=c'. Itisverypossiblethatthestatisticaltestofc'isnotsignificant(duetothecollinearityofXandM)whereascissignificant.Itwouldthenappearthatthereiscompletemediationwheniffactthereisnomediationatall.
Following,Kenny,Kashy,andBolger(1998),onemightaskwhetherallofthestepshavetobemetfortheretobemediation. Certainly,Step4doesnothavetobemetunlesstheexpectationisforcompletemediation. Intheopinionofmostthoughnotallanalysts,Step1isnotrequired. However,notethatapathfromtheinitialvariabletotheoutcomeisimpliedifSteps2and3aremet. Ifc'wereoppositeinsigntoabsomethingthatMacKinnon,Fairchild,andFritz(2007)refertoas"inconsistentmediation,"thenitcouldbethecasethatStep1wouldnotbemet,butthereisstillmediation.Inthiscasethemediatoractslikeasuppressorvariable. MostanalystsbelievethattheessentialstepsinestablishingmediationareSteps2and3.
JamesandBrett(1984)havearguedthatStep3shouldbemodifiedbynotcontrollingfortheinitialvariable. Theirrationaleisthatifthereiscompletemediation,therewouldbenoneedtocontrolfortheinitialvariable. However,becausecompletemediationdoesnotalwaysoccur,itwouldseemsensibletocontrolforXinStep3.
MeasuringMediationortheIndirectEffect
Theamountofmediation,whichiscalledtheindirecteffect,isdefinedasthereductionoftheeffectoftheinitialvariableontheoutcomeorc-c'. ThisdifferenceincoefficientsistheoreticallyexactlythesameastheproductoftheeffectofXonMtimestheeffectofMonYorab;thusitholdsthatab≈c-c'. Thetwoareexactlyequalwhena)multipleregression(orstructuralequationmodelingwithoutlatentvariables)isused,b)therearenomissingdata,c)andthesamecovariatesareintheequation. However,thetwoareonlyapproximatelyequalformultilevelmodels,logisticanalysisandstructuralequationmodelwithlatentvariables. Forsuchmodels,itisprobablyinadvisabletocomputecfromStep1,butrathercshouldbeinferredtobec'+abandnotdirectlycomputed. NotethattheamountofreductionintheeffectofXonYisnotequivalenttoeitherthechangeinvarianceexplainedorthechangeinaninferentialstatisticsuchasForapvalue. ItispossiblefortheFfromtheinitialvariabletotheoutcometodecreasedramaticallyevenwhenthemediatorhasnoeffectontheoutcome!
Itisalsonotequivalenttoachangeinpartialcorrelations.
IfStep2(thetestofa)andStep3(thetestofb)aremet,itfollowsthattherenecessarilyisareductionintheeffectofXonY. Onewaytotestthenullhypothesisthatab=0istotestthatbothaandbarezero(Steps2and3). Ifsuchastrategywereusedandonewanteda.05probabilityofthecombinedtestthata=0andb=0,thenalphaforthetestsofaandbshouldloweredto.0253sothattheTypeIerrorprotectionrateiscorrect.
Muchmorecommonly,asingletestisusedandishighlyrecommended(MacKinnon,Lockwood,Hoffman,West,&Sheets,2002). ThetestwasfirstproposedbySobel(1982). Itrequiresthestandarderrorofaorsa(whichequalsa/tawheretaisthettestofcoefficienta)andthestandarderrorofborsb. TheSobeltestprovidesthestandarderrorofabcanbeshowntoequalapproximatelythesquarerootof
b2sa2+a2sb2
Otherstandarderrorshavebeenproposed,buttheSobeltestisbyfarthemostcommonlyreported. ThetestoftheindirecteffectisgivenbydividingabbythesquarerootoftheabovevarianceandtreatingtheratioasaZtest(i.e.,largerthan1.96inabsolutevalueissignificantatthe.05level). KristopherJ.PreacherandGeoffreyJ.Leonardellihaveanexcellentwebpagethatcanhelpyoucalculatethesetest(gototheSobeltest). Measuresandtestsofindirecteffectsarealsoavailablewithinmanystructuralequationmodelingprograms.TheseprogramsappeartousetheSobelformula.
ThederivationoftheSobelstandarderrorpresumesthataandbareindependent,somethingthatistruewhenthetestsarefrommultipleregressionbutnottruewhenothertestsareused(e.g.,logisticregression,structuralequationmodeling,andmultilevelmodeling). Insuchcases,theresearcherideallyprovidesevidenceforapproximatei