資源描述:
《partial identification and sensitivity analysis》由會員上傳分享,免費在線閱讀,更多相關(guān)內(nèi)容在工程資料-天天文庫。
1、Chapter18PartialIdenti?cationandSensitivityAnalysisMarkusGanglAbstractThischapterisconcernedwithmethodsofcausalinferenceinthepresenceofunobservedconfounders.Threeclassesofestimatorsarediscussed,namely,localidentiTcationusinginstrumentalvariables,sensitivityanalysis,andestimationofnonpa
2、rametricbounds.Ineachcase,theresponsetothecoreidentiTcationproblemistoretreatfromthestandardfocusonpointidentiTcationoftheaveragetreatmenteffect,yetthethreeapproachescharacteristicallydifferintermsofalternativequantitiesofinterestthatareconsideredempiricallyestimableundermorerestrictiv
3、ecircumstances.ThechapterdevelopsthebasicprinciplesunderlyingthethreeclassesofpartialidentiTcationestimatorsandillustratestheirempiricalapplicationwithananalysisofearningsreturnstoeducation.IntroductionDuetodifTcultiesofpracticalimplementationandethicalconcernsaboutthedesirabilityofsoc
4、ialexperimentationbutalsorootedingenuineepistemicinterestinpopulation-levelinference,observationaldesignsarethenorminmanyTeldsofempiricalsocialresearch.Whenaimingforcausalinference,socialscientistshencetypicallyrecordputativecausesandoutcomesofinterestthroughsurveys,observation,content
5、analysis,orotherdatacollectionmethodsbutdonotresorttoactivemanipulationofcausesinordertolearnaboutpotentialeffects.Asaconsequence,causalinferencerequiresresearcherstounderstandprocessesthatgovernthereal-worldincidenceofcausallyrelevanteventsandconditionsinordertosustainanyclaimthat,oth
6、erthingsequal,aparticularcauseofinterestistypicallyfollowedbysomeoutcome.Andsincethoserelevantconditionsoftreatmentassignmentneedtobeexplicitlymeasuredinobservationalstudies,causalinferenceinsocialresearchinevitablyinvolvesconsiderablesubject-matterinputaswellaspotentiallywidelydiffere
7、ntviewsonhowtoadequatelyidentifyaparticularcausaleffectofinterest.SocialscientistsareinfactacutelyawareofthedifTcultiesinvolvedininferringcausalrelationshipsfromobservationaldata.Researchersarebroughtuponwarningsthatcorrelationdoesnotequalcausation,andthatrelevantconfounderorsuppress