资源描述：
《graphical causal models》由会员上传分享，免费在线阅读，更多相关内容在工程资料-天天文库。
1、Chapter13GraphicalCausalModelsFelixElwertAbstractThischapterdiscussestheuseofdirectedacyclicgraphs(DAGs)forcausalinferenceintheobservationalsocialsciences.ItfocusesonDAGsÕmainuses,discussescentralprinciples,andgivesappliedexamples.DAGsarevisualrepresentationsofqualitat
2、ivecausalassumptions:TheyencoderesearchersÕbeliefsabouthowtheworldworks.Straightforwardrulesmapthesecausalassumptionsontotheassociationsandindependenciesinobservabledata.ThetwoprimaryusesofDAGsare(1)determiningtheidentiÞabilityofcausaleffectsfromobserveddataand(2)deriv
3、ingthetestableimplicationsofacausalmodel.ConceptscoveredinthischapterincludeidentiÞcation,d-separation,confounding,endogenousselection,andovercontrol.Illustrativeapplicationsthendemonstratethatconditioningonvariablesatanystageinacausalprocesscaninduceaswellasremovebias
4、,thatconfoundingisafundamentallycausalratherthananassociationalconcept,thatconventionalapproachestocausalmediationanalysisareoftenbiased,andthatcausalinferenceinsocialnetworksinherentlyfacesendogenousselectionbias.Thechapterdiscussesseveralgraphicalcriteriafortheidenti
5、Þcationofcausaleffectsofsingle,time-pointtreatments(includingthefamousbackdoorcriterion),aswellidentiÞcationcriteriaformultiple,time-varyingtreatments.IntroductionVisualrepresentationsofcausalmodelshavealonghistoryinthesocialsciences,Þrstgainingprominencewithpathdiagra
6、msforlinearstructuralequationmodelsinthe1960s(Blalock1964;Duncan1975).Sincethesebeginnings,methodologistsinvariousdisciplineshavemaderemarkableprogressindevelopingformaltheoriesforgraphicalcausalmodelsthatnotonlygeneralizethelinearpathdiagramsofyoreintoafullynonparamet
7、ricframeworkbutalsointegrategraphicalmodelswiththereigningpotentialoutcomesframeworkofcausalinference.Bestofall,methodologistshavedevelopedasystemthatisbothrigorousandeasytouse.Inrecentyears,graphicalcausalmodelshavebecomelargelysynonymouswithdirectedacyclicgraphs(DAGs
8、).Ontheirown,DAGsarejustmathematicalobjectsbuiltfromdotsandarrows.Withafewassumptions,however,DAGscanberigorouslyrela
