Choosing Monitoring Boundaries: Balancing Risks and Benefits Pamela Shaw [email protected] Department of Biostatistics, Epidemiology and Informatics University of Pennsylvania April 19, 2017
ChoosingMonitoringBoundaries:BalancingRisksandBenefits
DepartmentofBiostatistics,EpidemiologyandInformaticsUniversityofPennsylvania
April19,2017
Outline
• Whatdowemonitor?• Howdowemonitorit?• Achallengingexample• Somenewproposals• Discussion
2
Whatdowemonitor?• Clinicaltrialarchitectureistypicallydefinedbyaprimaryefficacyoutcome
• A fundamentalroleofaDSMBistoassessthebenefit/riskratio
• Priorstudiescanyieldalistofpotentialrisks/benefits– Maybesymptoms(nausea,pain,etc.)ormayberisksofsevereoutcomes(elevatedstroke,cancer,death)
– Mayalsohaveimportantsecondaryefficacyendpoints(e.g.fracturesinWHIHormoneReplacementTrial)
– Trialsstructurealsosetuptocaptureunanticipatedadverseevents
3
Someofthechallengestoassessingbenefit/risk
1. Multivariateoutcomesneedtobeconsidered– Theseoutcomesmaybeofvaryingseverity
2. Risksmaychangeovertime3. Risksmaybeinfrequent/rare4. Fornoveltherapies,risksmaybelargelyunknown5. Expecttheunexpected…1&2implythatinordertoevaluaterisk/benefitonehastoprioritizetheoutcomesandprioritizetheimportanceofearly/lateevents(explicitlyorimplicitly,formallyorinformally)
4
Twoapproachesformonitoringrisk/benefit
1. Multipleoutcomesassessedseparately– Primaryendpointmayhaveaformalmonitoringboundary– DSMBispresentedwithanalysesofseveralseparateendpoints:primary,key2nd-ry,importantsafetyoutcomes
– DSMBweighstotalityofevidence,asubjectivejudgmentismadeforoverallbalanceofrisk/benefit
2. Astatisticsummarizingrisk/benefitisassessed– Compositeendpointdeterminedpriortostartoftrial– Risk/benefitp-valuecalculated/comparedtoaboundary– Subjectivejudgmentstillneededtoweighttotalityofevidence
5
Issuesthatcomplicateevaluationofthebenefit/risktradeoff
• Severityofhealthoutcomesaffectedbythetreatmentmaybeverydifferent– Assessingoverallbenefitmeansgivingrelativeweightstothese
risks/benefits– Patients/cliniciansmayhavedifferingopinionsontheseweights
• Frequencyofhealthoutcomesaffectedbythetreatmentmaybeverydifferent– Whendoestheincreasedriskofarare,butserioussideeffect
offsetthebenefitofatreatment?• Toleranceofasideeffectdependsonwhetheritisina
healthypopulationorsickpopulation• Timingofendpointsmaydiffer:earlyharm,laterbenefitor
viceversa6
WHIExample• Women’sHealthInitiative(WHI)conductedtwohormonetherapy(HT)trials
• TrialswereuniqueintheamountofdatacollectedonHTpriortothetrialstart– Expecting40-50%decreaseinheartattacks– Observationalstudiesraisedconcernoverincreaseinbreastcancer
• AformalmonitoringplanwasputintoplaceforbothefficacyandharmforbothHTtrials– Considered8outcomesofroughlyequalimportance.Mostthoughttoberelatedtoefficacy
– Hadaglobalindexofbenefit/risk(Z=-1)(Wittes etal.2007;Freedmanetal.1996) 7
WHIHTmonitoringplan
• Primaryefficacyendpoint:Coronaryheartdisease(CHD)
• Primarysafetyendpoint:invasivebreastcancer• Formalanalysesusedweightedlog-rankstatistictofurtherdown-weightearlyevents– MotivatedbyexpectedearlyCHDbenefitandlateBCAharm.Also,drugneededtimetohaveaneffect
– Unweightedusefulincasetherewereearlyharms,don’twanttodownweightthem
8
WHITrial:Unexpectedoutcomes• Discrepancybetweenexpectedandobservedefficacyandsafetyendpoints– Earlyon,anincreasedriskofCHD/stroke/PEforactivearmemergedinbothtrials
– Lateron,divergenteffectsappearedforbreastcancer
• Debateensuedwhetherandhowthesafetyendpointbemodified(Wittes etal2007)
• Levelofsignificanceanddirectionofeffectvariedbasedonweightedvsunweightedlog-rankstatistics
9
WHIExample:Lessonslearned/affirmed
• Monitoringmultivariateoutcomesiscomplex• Reliablyassessingriskandharmsmeansknowingwhichendpointsarewhich
• Difficulttorelyonasinglep-valuewhenconsideringamultivariateoutcome
• Decision-makingisultimatelyasubjectiveactivity
10
WHIexamplehighlightsmonitoringduality• Pre-specifiedboundariesprotectagainstinflatingp-values
bydefiningriskcategoriesafteradifferenceisobserved• Formalboundarieshowevercan”lockthinking”andneed
tobeflexibleinthefaceofunexpectedrisks– Adesiretosticktopre-specifiedboundaries– Ironically,statisticianscanbequickertoditchtheboundariesthanclinicalcolleagues
• Desiretohaveaclear,data-drivenstatisticdothework,butinterpretationneedstobringinaglobalperspective– Datafromothertrials– Leaningsofothertrendsindata– Uncertaintyinassumptionsbehindmonitoringboundary
11
Needforbetterstatisticalapproachestoassessbenefit/risk?
Usualstatisticalapproacheshavesomelimitations:• Timetofirstignoressubsequentandpotentiallymoresevereoccurringendpoints
• HRcanoveremphasizeincreasesinsmallabsoluterisk
• HR–precisionlimitedbynumberofevents• Unoetal.2015discussadvantagesofriskdifference,percentiledifference,restrictedmeansurvivaldifferenceinnon-inferioritytrials
• Multiplicity
12
Manyrecentproposalsforassessingbenefit/risk*
–Win-ratio:Pocock etal.2012,FinkelsteinandSchoenfeld 1999;Bebu andLachin 2016,Oakes2016
– Severityranking:ShawandFay2016– TotalAssessment:Evansetal2015(DOOR),Berryetal.2013
– OutcomeWeighting:Bakal etal.2013– Proportionfavoringtreatment:Buyse 2010– Jointtest:FinkelsteinandSchoenfeld 2014
13
Approachestoassessingbenefit/risk
1. Createanaggregatescorefromaweightedsumofoutcomes– Interpretedasaglobalassessmentofpatientoutcome– Naturallyincorporatesmultipleevents
2. Orderoutcomesintermsofapreferredimportanceandrank/classifypatientsusingthehighestorderedoutcomepossible– Forcensoredeventtimesoftenmeansrankingpatientsoveracommonfollow-uptime
– Essentiallycreatesaweightedcombinationofscorestatistics,wheretheratesrelatetotheprobabilityoftheeventsofhigherorderbeingobserved
14
Differingopinionsonwhethertocreateseparatesafetyandefficacycomposites
• EvansandFollmann 2016advocateaunifiedcompositeofbenefitandriskasapragmaticendpointofeffectiveness
• Kipetal2008recommendagainstlumpingsafetyandefficacylimitsinterpretabilityinsettingofcardiovasculardisease• Frequentlydominatedbyasubclassofendpoints• Toosusceptibletoprovidingmisleadingevidence
• “Althoughnumerousapproachesandframeworkshavebeenproposedinrecentyears,thereisnosingleapproachorframeworkthatcanbeappliedandutilizedineverysetting.”(Ch 8,Jiang,He2016)
15
Winratio1
• Patientsintreatmentandcontrolgroupsareplacedintomatchedpairsaccordingtotheirriskprofiles
• Determineprioritizationofoutcomes– Example:twoendpoints:deathorMIHospitalization,consider
timetodeathfirstthentimetohospitalization• Withineachpair,atx subjectislabeledawinnerorloser
usingtheoutcomeofhighestprioritypossible– Comparetimetodeathifpossible;otherwisecomparetimeto
hospitalization;otherwisetied• Thewinratioistheratioofwins/lossfortreatmentarm
– P-valueandCIarereadilyobtainable
1.Pocock 201216
Usefulfeaturesofwinratio
• Canconsiderallobservedeventsonapatient– Allowsmoresevereeventstohavehigherpriority– Particularlyusefulincaseswherefirsteventisexpectedtobethelesssevereevent
• Potentiallyhigherpowerthananysingleendpoint– Particularlyiftreatmenteffectsimilaracrossendpoints
• Easytocalculateandmakeinference1,2,3
– UnpairedversionisavailableusingaU-statisticderivedfromallpossibletx-controlpair
171. Pocock 20122.Bebu andLachin,2016;3.Oakes2016
Winratioexample:TheSOLVDtrial(NEJM1991)
Background• SOLVDincludedaRCTofanoveltreatmentforpreventionof
mortality/hospitalizationinpatientswithcongestiveheartfailure(CHF)andweakleftventricleejectionfraction(EF)
• In1986-89,2569patientsrandomizedtoenalapril orplacebo
• Enalapril foundbeneficialformortality(p=0.0036)andtimetofirsthospitalization/death(p<0.0001)
Analysis• Consideredasubsetof662diabeticsubjects• Computeusualtime-to-first(TTF)endpoint• Computewinratioforcontrol-treatmentpatientspairs
formedusingabaselineCoxmodelriskscorefordeath18
SOLVDTrial:Time-to-firstanalysis
19
SOLVD:Winratio
• 343onPlaceboarm,319onactivearm– 24patientsgounusedinthepairedanalysis
• 145winsonactive;112winsonplacebo• WR=145/112=1.29(p=0.038)– 189rankedondeath:98winsforactive,91winsforplacebo
– 68rankedonhospitalization:47winsactive;21winsplacebo
20
Afewkeypointsaboutwinratio
• TheparametertheWRestimatesdependsonthecensoringdistributionsoftheendpoint– Importantconsiderationifearlyandlaterisks
• Trialsofdifferentlengthswillgenerallybeestimatingadifferenteffectestimate
• Whenpatientshavevaryingfollow-uplengthstheWRbecomesmoredifficulttointerpret– SOLVDfollow-up:1dayto4.6yearsinexample
• Ifdeathdeterminesseverity,thenisrankingbyotherlesssevereendpointsgaininginformationorameansofpotentialmisclassification?
21
Winratio:Gaininginformationfromhospitalizationormisclassifying?
22
HH
Censoringtime
Patient1diedat3years;Patient2censoredat2.5years;diedat4years
1
2
WinRatio:Gaininginformationfromhospitalizationormisclassifying?
23
HH
Censoringtime
Patient1diedat3years;Patient2censoredat2.5years;diedat4years
Thetruestateofinformationhereisthatthepatient1severityrelativetopatient2isintervalcensored.
1
2
ClinicalseverityrankingShawandFaySIM2016
• Rankindividualsaccordingtoclinicalseverity,usinginformationonboththesurrogateandtrueendpoint– Rankingfunctionofthetwoeventtimescanvarybysetting
• Settingofinterest:XDR-TB:sputumconversion/death– Rankpatientsbytimeofdeathifobserved
• Earlierisworse
– Ranktimetosputumconversionforthesurvivors• Earlierisbetter
– Conversiontimeirrelevantifpatientlaterdies• Performtwosampletestonaninterval-censoredclinicalseveritywhichincorporatesbivariatesurvivalinformation
24
ShawandFaySIM2016Ranking Values: Worst to Best
Time to Surrogate
Tim
e to
Dea
th
0 5 10 15 20 ∞∞
05
1015
20
12 23 3 34 4 4 45 5 5 5 56 6 6 6 6 67 7 7 7 7 7 78 8 8 8 8 8 8 89 9 9 9 9 9 9 9 9
10 10 10 10 10 10 10 10 10 1011 11 11 11 11 11 11 11 11 11 1112 12 12 12 12 12 12 12 12 12 12 1213 13 13 13 13 13 13 13 13 13 13 13 1314 14 14 14 14 14 14 14 14 14 14 14 14 1415 15 15 15 15 15 15 15 15 15 15 15 15 15 1516 16 16 16 16 16 16 16 16 16 16 16 16 16 16 1617 17 17 17 17 17 17 17 17 17 17 17 17 17 17 17 1718 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 1819 19 19 19 19 19 19 19 19 19 19 19 19 19 19 19 19 19 1920 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 20 2041 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21
1234567891011121314151617181920
25
Greybox:Severityscoreforpatientwhoconvertedinweeks6-8,inclusive,butdroppedoutafterweek16.Intervalcensoringindisjointintervals
Furthermusingsontestsofseverityusingjointsurvivaldistribution
• Takeadvantageofalltheinformationregardingthesurvivaltime(notlimitedtocommonfollow-uptimesforpairs)
• Includingthetrueuncertaintyabouttheseverityofapatient
• Teststatisticforcompositestillhastheproblemthatthatparameterestimateddependsonlengthoftrial– FayandShawshowedthattheresultingteststatisticisaweightedsumofateststatisticondeathandonthesurrogate
26
DOORRankingEvansetal.2015
• Collectionofpossibilitiesofclinicaloutcomesofapatientsarerankedaccordingtopreferredtoleastpreferredoutcome– Rateallpossibleclinicalpathsonanordinalscale– Rating/rankingcanbedonebyexpertclinicalpanel,potentially
alsoincludingpatients– ThenaU-typestatisticcouldbeusedtoexamineiftheoutcome
ontx betterthanthatforapatient• Proposedablindedadjudicatedcommitteecouldevaluate
clinicalseveritybasedonpatientchart– Notpracticalforlargertrialsorveryreproducible
• Similarideasdiscussedbyanumberofauthors,includingChuang-Steinetal.1991
27
DOORhypotheticalexample(Evansetal.2015)
1. ClinicalbenefitwithoutAE2. ClinicalbenefitwithAE3. Survivalw/oclinicalbenefitorAE4. Survivalw/oclinicalbenefit+AE5. Death• Insettingofanti-infective,breaktiesusinglengthofantibioticregimen(DOOR/RADAR)
28
DOOR:AdvantagesandlimitationsAdvantages– Simpleandintuitivemeasure– Rankingcognitivelyeasierthanweighting– Canincorporatedifferentrankingsystems
Limitations– Varyinglengthoffollow-upcanbeachallenge– Lossofinformationthroughtiescanbeaproblemforordinal
– Willbedifficulttoadapttounexpectedbenefitsorrisks.Wouldneedtoreconveneoutcomerankingpanel
• Perhapsbestusedalong-sideIndividualcomponentsforinterpretation
29
Somepragmaticconsiderations• Forcompositeendpoints:creategroup(s)basedonsimilar
severity– Somesettingsmaywanttopoolsafetyandriskfornetclinical
benefit– Addedinterpretabilityifindividualsoutcomesoccurwithsimilar
frequency• Sensitivityanalysistoseeimpactofvaluesystems
– Ifusingoutcomeweighting,canbeusedidentifythe“valuebreakingpoint”
• PracticeRundecisionscenarios:Valuableexercisetohonetheneededvaluejudgements(somecanbepre-specified)andstatisticaldecisionboundaries
• Clearpresentationandvisualizationofdata(estimates)forDSMBreportwillaidinassessmentoftotalityofevidence
30
Conclusions• Nooneapproachwillworkforeverysetting• Goodtorememberallapproachesinvolvesubjectivity• Specificendpoint+compositesthatsummarizeeffectonmultipleendpointsseemslikeaflexibleandpowerfulcombination
• Statisticalpropertiesofcomposites needrigorousexaminationandthoroughnumericalinvestigationbeforestartoftrialforexpectedscenarios
• Practicerundecisionscenarios:Valuableexercisetotohonetheneededvaluejudgementsandstatisticaldecisionboundaries
• Apriordevelopmentofrisk-benefitstatisticandboundaryisausefuldecisiontoolbutcannotbeprescriptive
31
32
Thankyou!
References1. Bakal JA,Westerhout CM,CantorWJ,Fernández-Avilés F,WelshRC,FitchettD,GoodmanSG,
ArmstrongPW.Evaluationofearlypercutaneouscoronaryinterventionvs.standardtherapyafterfibrinolysisforST-segmentelevationmyocardialinfarction:contributionofweightingthecompositeendpoint.EuropeanHeartJournal.2013Mar21;34(12):903-8.Bebu
2. Bebu I,Lachin JM.Largesampleinferenceforawinratioanalysisofacompositeoutcomebasedonprioritizedcomponents.Biostatistics.2016Jan1;17(1):178-87.
3. BerryJD,MillerR,MooreDH,Cudkowicz ME,VanDenBergLH,KerrDA,DongY,IngersollEW,ArchibaldD.TheCombinedAssessmentofFunctionandSurvival(CAFS):anewendpointforALSclinicaltrials.Amyotrophiclateralsclerosisandfrontotemporaldegeneration.2013Apr1;14(3):162-8.BuyseM.Generalizedpairwisecomparisonsofprioritizedoutcomesinthetwo-sampleproblem.StatisticsinMedicine.2010Dec30;29(30):3245-57.
4. Chuang-SteinC,Mohberg NR,SinkulaMS.Threemeasuresforsimultaneouslyevaluatingbenefitsandrisksusingcategoricaldatafromclinicaltrials.StatisticsinMedicine.1991Sep1;10(9):1349-59.
5. EvansSR,Follmann D.UsingOutcomestoAnalyzePatientsRatherthanPatientstoAnalyzeOutcomes:AStepTowardPragmatisminBenefit:RiskEvaluation.StatisticsinBiopharmaceuticalResearch.2016Oct1;8(4):386-93.
6. EvansSR,RubinD,Follmann D,Pennello G,HuskinsWC,PowersJH,Schoenfeld D,Chuang-SteinC,CosgroveSE,FowlerVG,Lautenbach E.Desirabilityofoutcomeranking(DOOR)andresponseadjustedfordurationofantibioticrisk(RADAR).ClinicalInfectiousDiseases.2015Jun25:civ495.
7. FinkelsteinDM,Schoenfeld DA.Combiningmortalityandlongitudinalmeasuresinclinicaltrials.StatisticsinMedicine.1999Jun15;18(11):1341-54.
33
References(2)8. FinkelsteinDM,Schoenfeld DA.Ajointtestforprogressionandsurvivalwithinterval-censoreddata
fromacancerclinicaltrial.StatisticsinMedicine.2014May30;33(12):1981-9.
9. FreedmanL,AndersonG,Kipnis V,PrenticeR,WangCY,Rossouw J,Wittes J,DeMets D.Approachestomonitoringtheresultsoflong-termdiseasepreventiontrials:examplesfromtheWomen'sHealthInitiative.ControlledClinicalTrials.1996Dec1;17(6):509-25.
10. JiangQ,HeW,editors.Benefit-RiskAssessmentMethodsinMedicalProductDevelopment:BridgingQualitativeandQuantitativeAssessments.ChapmanandHall/CRC;2016Jul7.
11. KipKE,Hollabaugh K,MarroquinOC,WilliamsDO.Theproblemwithcompositeendpointsincardiovascularstudies:thestoryofmajoradversecardiaceventsandpercutaneouscoronaryintervention.JournaloftheAmericanCollegeofCardiology.2008Feb19;51(7):701-7.
12. OakesD.Onthewin-ratiostatisticinclinicaltrialswithmultipletypesofevent.Biometrika.2016Jul25:asw026.
13. Pocock SJ,Ariti CA,CollierTJ,WangD.Thewinratio:anewapproachtotheanalysisofcompositeendpointsinclinicaltrialsbasedonclinicalpriorities.EuropeanHeartJournal.2012Jan1;33(2):176-82.
14. ShawPA,FayMP.Aranktestforbivariatetime-to-eventoutcomeswhenoneeventisasurrogate.Statisticsinmedicine.2016Aug30;35(19):3413-23.
15. UnoH,Wittes J,FuH,SolomonSD,Claggett B,TianL,Cai T,Pfeffer MA,EvansSR,WeiLJ.Alternativestohazardratiosforcomparingtheefficacyorsafetyoftherapiesinnoninferiority studies.AnnalsofInternalMedicine.2015Jul21;163(2):127-34.
16. Wittes J,Barrett-ConnorE,Braunwald E,ChesneyM,CohenHJ,DeMets D,DunnL,DwyerJ,HeaneyRP,VogelV,WaltersL.MonitoringtherandomizedtrialsoftheWomen'sHealthInitiative:theexperienceoftheDataandSafetyMonitoringBoard.ClinicalTrials.2007Jun1;4(3):218-34.
34
35