BasicsofStudyDesignJaniceWeinbergScDAssistantProfessorofBiostatisticsBostonUniversitySchoolofPublicHealthBasicsofStudyDesignJaniceW1BasicsofStudyDesignBiasandvariabilityRandomization:whyandhow?Blinding:whyandhow?GeneralstudydesignsBasicsofStudyDesignBiasand2BiasandVariabilityTheclinicaltrialisconsideredtobethe“goldstandard”inclinicalresearchClinicaltrialsprovidetheabilitytoreducebiasandvariabilitythatcanobscurethetrueeffectsoftreatmentBiasaffectsaccuracyVariabilityaffectsprecisionBiasandVariabilityTheclinic3Bias:anyinfluencewhichactstomaketheobservedresultsnon-representativeofthetrueeffectoftherapy

Examples:healthierpatientsgiventreatmentA,sickerpatientsgiventreatmentBtreatmentAis“newandexciting”soboththephysicianandthepatientexpectbetterresultsonAManypotentialsourcesofbiasBias:anyinfluencewhichacts4Variability:highvariabilitymakesitmoredifficulttodiscerntreatmentdifferencesSomesourcesofvariabilityMeasurementinstrumentobserverBiologicwithinindividualsbetweenindividualsCannotalwayscontrolforallsources(andmaynotwantto)Variability:highvariability5Fundamentalprinciple

incomparingtreatmentgroups:GroupsmustbealikeinallimportantaspectsandonlydifferinthetreatmenteachgroupreceivesInpracticalterms,“comparabletreatmentgroups”means“alikeontheaverage”Fundamentalprinciple

incomp6Whyisthisimportant?IfthereisagroupimbalanceforanimportantfactorthenanobservedtreatmentdifferencemaybeduetotheimbalanceratherthantheeffectoftreatmentExample:DrugXversusplaceboforosteoporosisAgeisariskfactorforosteoporosisOldersubjectsareenrolledinDrugXgroupTreatmentgroupcomparisonwillbebiasedduetoimbalanceonageWhyisthisimportant?Ifthere7Howcanweensurecomparabilityoftreatmentgroups?WecannotensurecomparabilitybutrandomizationhelpstobalanceallfactorsbetweentreatmentgroupsIfrandomization“works”thengroupswillbesimilarinallaspectsexceptforthetreatmentreceivedHowcanweensurecomparabilit8RandomizationAllocationoftreatmentstoparticipantsiscarriedoutusingachancemechanismsothatneitherthepatientnorthephysicianknowinadvancewhichtherapywillbeassignedSimplestCase:eachpatienthasthesamechanceofreceivinganyofthetreatmentsunderstudyRandomizationAllocationoftre9SimpleRandomizationThinkoftossingacoineachtimeasubjectiseligibletoberandomizedHEADS: TreatmentATAILS: TreatmentBApproximately?willbeassignedtotreatmentsAandBRandomizationusuallydoneusingarandomizationscheduleoracomputerizedrandomnumbergeneratorSimpleRandomizationThinkoft10ProblemwithSimpleRandomization:Mayresultinsubstantialimbalanceineitheranimportantbaselinefactorand/orthenumberofsubjectsassignedtoeachgroupSolution:Useblockingand/orstratifiedrandomizationProblemwithSimpleRandomizat11BlockingExample:Ifwehavetwotreatmentgroups(AandB)equalallocation,andablocksizeof4,randomassignmentswouldbechosenfromtheblocks1)AABB 4)BABA2)ABAB 5)BAAB3)ABBA 6)BABABlockingensuresbalanceafterevery4thassignmentBlockingExample:Ifwehavetw12StratificationExampleToensurebalanceonanimportantbaselinefactor,createstrataandsetupseparaterandomizationscheduleswithineachstratumExample:ifwewantpreventanimbalanceonageinanosteoporosisstudy,firstcreatethestrata“<75years”and“75years”thenrandomizewithineachstratumseparatelyBlockingshouldbealsobeusedwithineachstratumStratificationExampleToensur13AlternativestoRandomizationRandomizationisnotalwayspossibleduetoethicalorpracticalconsiderationsSomealternatives:HistoricalcontrolsNon-randomizedconcurrentcontrolsDifferenttreatmentperphysicianSystematicalternationoftreatmentsSourcesofbiasforthesealternativesneedtobeconsideredAlternativestoRandomizationR14BlindingMaskingtheidentityoftheassignedinterventionsMaingoal:avoidpotentialbiascausedbyconsciousorsubconsciousfactorsSingleblind: patientisblindedDouble

blind: patientandassessing investigatorareblindedTriple

blind: committeemonitoring responsevariables(e.g. statistician)isalsoblindedBlindingMaskingtheidentityo15HowtoBlindTo“blind”patients,canuseaplaceboExamplespillofsamesize,color,shapeastreatmentshamoperation(anesthesiaandincision)foranginareliefshamdevicesuchasshamacupuncture

HowtoBlindTo“blind”patient16WhyShouldPatientsbeBlinded?Patientswhoknowtheyarereceivinganeworexperimentalinterventionmayreportmore(orless)sideeffectsPatientsnotonneworexperimentaltreatmentmaybemore(orless)likelytodropoutofthestudyPatientmayhavepreconceivednotionsaboutthebenefitsoftherapyPatientstrytogetwell/pleasephysiciansWhyShouldPatientsbeBlinded17Placeboeffect–responsetomedicalinterventionwhichresultsfromtheinterventionitself,notfromthespecificmechanismofactionoftheinterventionExample:FisherR.W.JAMA1968;203:418-419

46patientswithchronicsevereitchingrandomlygivenoneoffourtreatmentsHighitchingscore=moreitchingTreatment ItchingScore cyproheptadineHCI 27.6 trimeprazinetartrate 34.6 placebo 30.4 nothing 49.6Placeboeffect–responsetom18WhyShouldInvestigatorsbeBlinded?TreatingphysiciansandoutcomeassessinginvestigatorsareoftenthesamepeoplePossibilityofunconsciousbiasinassessingoutcomeisdifficulttoruleoutDecisionsaboutconcomitant/compensatorytreatmentareoftenmadebysomeonewhoknowsthetreatmentassignment“Compensatory”treatmentmaybegivenmoreoftentopatientsontheprotocolarmperceivedtobelesseffectiveWhyShouldInvestigatorsbeBl19CanBlindingAlwaysbeDone?Insomestudiesitmaybeimpossible(orunethical)toblindatreatmentmayhavecharacteristicsideeffectsitmaybedifficulttoblindthephysicianinasurgeryordevicestudySourcesofbiasinanun-blindedstudymustbeconsideredCanBlindingAlwaysbeDone?In20GeneralStudyDesignsManyclinicaltrialstudydesignsfallintothecategoriesofparallelgroup,dose-ranging,cross-overandfactorialdesignsTherearemanyotherpossibledesignsandvariationsonthesedesignsWewillconsiderthegeneralcasesGeneralStudyDesignsManyclin21GeneralStudyDesignsParallelgroupdesignsGeneralStudyDesignsParallel22GeneralStudyDesignsDose-RangingStudiesGeneralStudyDesignsDose-Rang23GeneralStudyDesignsCross-OverDesignsGeneralStudyDesignsCross-Ove24GeneralStudyDesignsFactorialDesignsGeneralStudyDesignsFactorial25Cross-OverDesignsSubjectsarerandomizedtosequencesoftreatments(AthenBorBthenA)Usesthepatientashis/herowncontrolOftena“wash-out”period(timebetweentreatmentperiods)isusedtoavoida“carryover”effect(theeffectoftreatmentinthefirstperiodaffectingoutcomesinthesecondperiod)Canhaveacross-overdesignwithmorethan2periodsCross-OverDesignsSubjectsare26Cross-OverDesignsAdvantage:treatmentcomparisonisonlysubjecttowithin-subjectvariabilitynotbetween-subjectvariabilityreducedsamplesizesDisadvantages:strictassumptionaboutcarry-overeffectsinappropriateforcertainacutediseases(whereaconditionmaybecuredduringthefirstperiod)dropoutsbeforesecondperiodCross-OverDesignsAdvantage:t27Cross-OverDesignsAppropriateforconditionsthatareexpectedtoreturntobaselinelevelsatthebeginningofthesecondperiodExamples:TreatmentofchronicpainComparisonofhearingaidsforhearinglossMouthwashtreatmentforgingivitisCross-OverDesignsAppropriate28FactorialDesignsAttemptstoevaluatetwointerventionscomparedtoacontrolinasingleexperiment(simplestcase)Animportantconceptforthesedesignsisinteraction(sometimescalledeffectmodification)Interaction:TheeffectoftreatmentAdiffersdependinguponthepresenceorabsenceofinterventionBandvice-versa.FactorialDesignsAttemptstoe29FactorialDesignsAdvantages:Ifnointeraction,canperformtwoexperimentswithlesspatientsthanperformingtwoseparateexperimentsCanexamineinteractionsifthisisofinterestDisadvantages:Addedcomplexitypotentialforadverseeffectsdueto“poly-pharmacy”FactorialDesignsAdvantages:30FactorialDesignsExample:Physician’sHealthStudyPhysiciansrandomizedto:aspirin(topreventcardiovasculardisease)beta-carotene(topreventcancer)aspirinandbeta-caroteneneither(placebo)Stampfer,Buring,Willett,Rosner,EberleinandHennekens(1985)The2x2factorialdesign:it’sapplicationtoarandomizedtrialofaspirinandcaroteneinU.S.physicians.Stat.inMed.9:111-116.FactorialDesignsExample:Phy31BasicsofStudyDesignJaniceWeinbergScDAssistantProfessorofBiostatisticsBostonUniversitySchoolofPublicHealthBasicsofStudyDesignJaniceW32BasicsofStudyDesignBiasandvariabilityRandomization:whyandhow?Blinding:whyandhow?GeneralstudydesignsBasicsofStudyDesignBiasand33BiasandVariabilityTheclinicaltrialisconsideredtobethe“goldstandard”inclinicalresearchClinicaltrialsprovidetheabilitytoreducebiasandvariabilitythatcanobscurethetrueeffectsoftreatmentBiasaffectsaccuracyVariabilityaffectsprecisionBiasandVariabilityTheclinic34Bias:anyinfluencewhichactstomaketheobservedresultsnon-representativeofthetrueeffectoftherapy

Examples:healthierpatientsgiventreatmentA,sickerpatientsgiventreatmentBtreatmentAis“newandexciting”soboththephysicianandthepatientexpectbetterresultsonAManypotentialsourcesofbiasBias:anyinfluencewhichacts35Variability:highvariabilitymakesitmoredifficulttodiscerntreatmentdifferencesSomesourcesofvariabilityMeasurementinstrumentobserverBiologicwithinindividualsbetweenindividualsCannotalwayscontrolforallsources(andmaynotwantto)Variability:highvariability36Fundamentalprinciple

incomparingtreatmentgroups:GroupsmustbealikeinallimportantaspectsandonlydifferinthetreatmenteachgroupreceivesInpracticalterms,“comparabletreatmentgroups”means“alikeontheaverage”Fundamentalprinciple

incomp37Whyisthisimportant?IfthereisagroupimbalanceforanimportantfactorthenanobservedtreatmentdifferencemaybeduetotheimbalanceratherthantheeffectoftreatmentExample:DrugXversusplaceboforosteoporosisAgeisariskfactorforosteoporosisOldersubjectsareenrolledinDrugXgroupTreatmentgroupcomparisonwillbebiasedduetoimbalanceonageWhyisthisimportant?Ifthere38Howcanweensurecomparabilityoftreatmentgroups?WecannotensurecomparabilitybutrandomizationhelpstobalanceallfactorsbetweentreatmentgroupsIfrandomization“works”thengroupswillbesimilarinallaspectsexceptforthetreatmentreceivedHowcanweensurecomparabilit39RandomizationAllocationoftreatmentstoparticipantsiscarriedoutusingachancemechanismsothatneitherthepatientnorthephysicianknowinadvancewhichtherapywillbeassignedSimplestCase:eachpatienthasthesamechanceofreceivinganyofthetreatmentsunderstudyRandomizationAllocationoftre40SimpleRandomizationThinkoftossingacoineachtimeasubjectiseligibletoberandomizedHEADS: TreatmentATAILS: TreatmentBApproximately?willbeassignedtotreatmentsAandBRandomizationusuallydoneusingarandomizationscheduleoracomputerizedrandomnumbergeneratorSimpleRandomizationThinkoft41ProblemwithSimpleRandomization:Mayresultinsubstantialimbalanceineitheranimportantbaselinefactorand/orthenumberofsubjectsassignedtoeachgroupSolution:Useblockingand/orstratifiedrandomizationProblemwithSimpleRandomizat42BlockingExample:Ifwehavetwotreatmentgroups(AandB)equalallocation,andablocksizeof4,randomassignmentswouldbechosenfromtheblocks1)AABB 4)BABA2)ABAB 5)BAAB3)ABBA 6)BABABlockingensuresbalanceafterevery4thassignmentBlockingExample:Ifwehavetw43StratificationExampleToensurebalanceonanimportantbaselinefactor,createstrataandsetupseparaterandomizationscheduleswithineachstratumExample:ifwewantpreventanimbalanceonageinanosteoporosisstudy,firstcreatethestrata“<75years”and“75years”thenrandomizewithineachstratumseparatelyBlockingshouldbealsobeusedwithineachstratumStratificationExampleToensur44AlternativestoRandomizationRandomizationisnotalwayspossibleduetoethicalorpracticalconsiderationsSomealternatives:HistoricalcontrolsNon-randomizedconcurrentcontrolsDifferenttreatmentperphysicianSystematicalternationoftreatmentsSourcesofbiasforthesealternativesneedtobeconsideredAlternativestoRandomizationR45BlindingMaskingtheidentityoftheassignedinterventionsMaingoal:avoidpotentialbiascausedbyconsciousorsubconsciousfactorsSingleblind: patientisblindedDouble

blind: patientandassessing investigatorareblindedTriple

blind: committeemonitoring responsevariables(e.g. statistician)isalsoblindedBlindingMaskingtheidentityo46HowtoBlindTo“blind”patients,canuseaplaceboExamplespillofsamesize,color,shapeastreatmentshamoperation(anesthesiaandincision)foranginareliefshamdevicesuchasshamacupuncture

HowtoBlindTo“blind”patient47WhyShouldPatientsbeBlinded?Patientswhoknowtheyarereceivinganeworexperimentalinterventionmayreportmore(orless)sideeffectsPatientsnotonneworexperimentaltreatmentmaybemore(orless)likelytodropoutofthestudyPatientmayhavepreconceivednotionsaboutthebenefitsoftherapyPatientstrytogetwell/pleasephysiciansWhyShouldPatientsbeBlinded48Placeboeffect–responsetomedicalinterventionwhichresultsfromtheinterventionitself,notfromthespecificmechanismofactionoftheinterventionExample:FisherR.W.JAMA1968;203:418-419

46patientswithchronicsevereitchingrandomlygivenoneoffourtreatmentsHighitchingscore=moreitchingTreatment ItchingScore cyproheptadineHCI 27.6 trimeprazinetartrate 34.6 placebo 30.4 nothing 49.6Placeboeffect–responsetom49WhyShouldInvestigatorsbeBlinded?TreatingphysiciansandoutcomeassessinginvestigatorsareoftenthesamepeoplePossibilityofunconsciousbiasinassessingoutcomeisdifficulttoruleoutDecisionsaboutconcomitant/compensatorytreatmentareoftenmadebysomeonewhoknowsthetreatmentassignment“Compensatory”treatmentmaybegivenmoreoftentopatientsontheprotocolarmperceivedtobelesseffectiveWhyShouldInvestigatorsbeBl50CanBlindingAlwaysbeDone?Insomestudiesitmaybeimpossible(orunethical)toblindatreatmentmayhavecharacteristicsideeffectsitmaybedifficulttoblindthephysicianinasurgeryordevicestudySourcesofbiasinanun-blindedstudymustbeconsideredCanBlindingAlwaysbeDone?In51GeneralStudyDesignsManyclinicaltrialstudydesignsfallintothecategoriesofparallelgroup,dose-ranging,cross-overandfactorialdesignsTherearemanyotherpossibledesignsandvariationsonthesedesignsWewillconsiderthegeneralcasesGeneralStudyDesignsManyclin52GeneralStudyDesignsParallelgroupdesignsGeneralStudyDesignsParallel53GeneralStudyDesignsDose-RangingStudiesGeneralStudyDesignsDose-Rang54GeneralStudyDesignsCross-OverDesignsGeneralStudyDesignsCross-Ove55GeneralStudyDesignsFactorialDesignsGeneralStudyDesignsFactorial56Cross-OverDesignsSubjectsarerandomizedtosequencesoftreatments(AthenBorBthenA)Usesthepatientashis/herowncontrolOftena“wash-out”period(timebetweentreatmentperiods)isusedtoavoida“carryover”effect(theeffectoftreatmentinthefirstperiodaffectingoutcomesinthesecondperiod)Canhave