InferenceonClinicalTrialswithOutcomesTruncatedbyDeathXiao-HuaAndrewZhou,Ph.DProfessorofDepartmentofBiostatisticsandAssociateDirectorofNationalAlzheimer’sCoordinatingCenter,UniversityofWashington,andChairProfessor,PekingUniversityApril24,20171/1Outline•Issueswithtruncationbydeathinclinicaltrials•Areal-worldexample•Definitionofcausalparametersofinterest•Modelidentifiability•Estimation•Applicationtoonereal-worldexamples2/1Truncationbydeath•Inmedicalstudies,researchersareofteninterestedinevaluatingriskfactorsforanon-mortalityoutcomesuchasmemorydecline.•However,thenon-mortalityoutcomemaybetruncatedbydeathifsomesubjectsdiebeforethefollow-upassessment,leavingtheirnon-mortalityoutcomestobeundefined.3/1Amotivatingexample•HealthandLifestyleSurvey(HALS),apopulation-basedprospectivecohortstudyconductedinEngland,ScotlandandWales.•ThereweretwovisitsintheHALS.ThebaselinesurveyHALS1wasconductedin1984-1985andinterviewed9003participantsontheirhealthstatuses,attitudestohealthandothermeasurementsrelatedtohealthandlifestyles.•Thefollow-upsurveyHALS2wasconductedin1991-1992,atwhichtime808ofthe7414originalrespondentstothebaselinesurveyhaddied.•ThesurvivingrespondentsofHALS1werere-surveyedtoexaminetheirchangesonmemoryscoresoverthe7-yearperiod.•WeareinterestedinestimatingtheeffectofsmokingonmemorydeclineusingobservationaldatafromtheHealthandLifestyleStudy(HALS).4/1ExistingMethods-naivemethod•Oneofthenaiveanalyticapproachesistoperformanalysisonlyonsubjectswhoarealiveoronallobserveddata.•Amajorlimitationofthisapproachisthatitfocusesonthehealthiestsubgroupofpopulation(thesurvivors)andtypicallyinducesbiasintheinferenceontheoriginalcohort.•Morefundamentally,directcomparisonsamongobservedsurvivorsarenotcausallyinterpretableastheycompareoutcomesfromdifferentsubpopulationsatbaseline.5/1ExistingMethods-missing-datamethods•Datatruncatedbydeatharesometimeshandledbyamethodformissingdata.•Revickietal.(2001)imputedthedatathatwouldhavebeenobservedifthesubjecthadnotdied.•However,there...
