meta送臨床試驗的分析32

Statistics542

IntroductiontoClinicalTrials

MetaAnalysis542-11-#1Meta-Analysis Alternatives? Occasionally Complementary? YesMeta-AnalysisCombinationofsimilarstudiesusingsimilarsubjectsandsimilartreatmentsandsimilares542-11-#2Figure2

OddsRatiosand95%ConfidenceLimits

forVariousStudiesandaPooledEstimate542-11-#3NewMethodofAnalyzingHealthDataStirsDebate

byLawrenceK.Altman542-11-#4Increasinguseofacontroversialstatisticalmethodtoevaluatemedicaltherapiesandsurgicalproceduresisbeginningtoaffectprofoundlythecareofpregnantwomenandpatientswithcancer,heartdiseaseandmanyothercommonconditions.Themethod,knownasmeta-analysispromisestoplananincreasinglyimportantroleindetermininghealthrisks,environmentalhazardsandnationalpolicyonpaymentformedicalcare.Backerssaytechniquecandrawbig,reliableconclusionsfromsmall,inconsistentfindings.

Meta-analysisisatermderivedfromtheGreekmeaningananalysisthatismorecomprehensive.Thelargernumbersobtainedbycombiningstudiesprovideagreaterstatisticalpowerthananyoftheindividualstudies.Researchersareoftenabletodrawmorereliableinferencesornewconclusionsfromthecombinedresultsthanfromthesmallerstudiesthatmaybeinconclusiveindividually.Inearlierapplicationsofmeta-analysis,researchersevaluatedintelligencequotients,governmentsocialwelfareprogramsandmanyothertopics.Meta-analysishascometomedicinelate,but“itisnowundergoingaboominpopularity,”saidDr.ThomasC.Chalmers,adistinguishedphysicianoftheDepartmentofVeteransAffairsinBostonandapioneerinmethodology.Themethodinvolvesananalysisofpreviousanalyses.Itcombinestheresultsofawiderangeofexistingsmallerstudiesandthenappliesoneofseveralstatisticaltechniquestodiscovermorepreciselywhatisknownfrompreviousresearch.Itmayalsoproduceaunifiedresultfromdiverse,apparentlycontradictorystudies.Thetechniquehasalreadyshednewlightontheeffectivenessofmedicaltherapies.Althoughithasnot,initself,revolutionizedanymedicaltreatmentithashelpedclearawaytheconfusioncausedbystudieswithscatteredandapparentlyconflictingfindingsandhasstrengthenandconfirmedfindingsfromtraditionalclinicaltrials.NYTimes8/21/90Reference:NIHProceedings542-11-#5MethodologicIssuesinOverviewsofRandomizedClinicalTrialsNIHConferenceMay1986StatisticsinMedicineVol6,No.3,1987WhatisthePurpose?a. Testingforatreatmenteffect(rejectingthenullhypothesis)b. Evaluatingasafetyissue(rareevents)c. Estimatingsizeoftreatmenteffectinsubgroupsd. Designofnewstudiese. Developpracticeguidelines542-11-#6IdealMetaAnalysisisRandomizedMulti-centerControlTrialSameprotocolSametreatmentSametypeofsubjectsSameemeasure542-11-#7IssuesinMetaAnalysisDifferencesAcrossStudiesin: a. Treatment b. ControlGroup/Population c. TimeSpan(Disease,BackgroundTherapy) d. eMeasuresPublicationBiasCompleteness/QualityofDataAccesstoData542-11-#8WhatStudiesShouldBeIncluded?AllexistingstudiesAllpublishedstudies"Non-flawed"trialsOtherselectioncriteria542-11-#9Meta-Analysis:When?(1)RetrospectiveAnalysesTestTreatmentEffectWhen:DefinitiveanswernotyetavailableNomorestudieslikelyNeedtosalvageavailableresultsDevelopPracticeGuidelinesDesignNewStudies542-11-#10Meta-Analysis:When?(2)ProspectiveAnalysesNotmendedBettertodesigninadvancepropermulti-centertrial(s)542-11-#11Meta-AnalysisMethodologyNotNewCombiningp-values,Fisher(1948)AnalysisofVariance,Fisher(1938)Combining2x2TablesMantel-Haenszel(1959)Cochran(1954)542-11-#12OddsRatio542-11-#13moreexplicitly

OR=ad/bcTCSaba+bFcdc+da+cb+dMethodsofMeta-Analysis542-11-#14Collapsingcanbemisleadingifthereisqualitativeinteraction.1.0CollapseDataRCT-1TCS155F8595OR=3.35RCT-2TCS515F9585OR=0.30CollapsedTC2020180180OR=1.02. GraphicalSeeFigure95%CIforeachstudy(ad/bc)exp{±1.96(1/a+1/b+1/c+1/d)}542-11-#15MethodsofMetaAnalysis542-11-#16ApparenteffectsoffibrinolytictreatmentonmoralityintherandomisedtrialsofIVtreatmentofacutemyocardialinfarction.StatinMed7:890:1988.542-11-#17Comparisonofmeta-analysisof12RCTsofi.v.mixeddrugs(double-blind)withi.v.metoprolol(double-blind)andi.v.atenlol(openstudy).StatinMed6(3):320,1987.542-11-#18Comparisonofmeta-analysisofmortalityin11RCTsandreinfarctionratesin10RCTsofi.v.streptokinasewithlargeco-operativestudy(GISSI).StatinMed6(3):320,1987.542-11-#19Comparisonofmeta-analysisof7smallRCTsofphenobarbitalinthetreatmentofneonatalintra-cranialhaemmorrhagewithonelargeco-operativestudy(3institutions).Endpointsaretotalinfantswithhaemmorrhageandtotalswithseverehaemorrhage(GradesIII-IV)only.StatinMed6(3):321,1987.OddsRatiosand95%ConfidenceLimits

forVariousStudiesandaPooledEstimate542-11-#20542-11-#213. Blocking(Peto-MH)OverallEstimate LetO=ai E=Ei Ei=(ai+ci)(ai+bi) ni V=Vi Vi=(ai+ci))(bi+di)(ci+di)(ai+bi ni2(ni-1) Z=O-E

C PooledOR OR=exp{(O-E)/V} 95%CI=exp{(O-E)/V±1.96/}MethodsofMetaAnalysis542-11-#224. AveragingP-values Fisher(1948) Pi=P-valueforithtrial Z=-2log(Pi)~2with2Ndf5. AveragingTestStatistics e.g.wi=niMethodsofMetaAnalysisMeta-AnalysisExamplesCardiologyPostMITreatments (e.g.,beta-blockers,aspirin)ThrombolyticTherapy (e.g.,streptokinase)Anticoagulants542-11-#23Registries/DatabasesByar(1980)BiometricsD'Ambrosia,Ellenberg(1980)BiometricsStarmeretal.(1980)BiometricsMantel(1983)StatisticsinMedicine542-11-#24Registries/DatabasesUseClinicalObservationalSeriesto:DescribeClinicalPracticeIdentifyRiskFactors"Evaluate"TreatmentHistoricalConcurrent542-11-#25DatabasesTreatmentEvaluationComparisonRequiresRiskFactorComparabilityMeasuredNotMeasuredorUnknownStatisticalModelsUsuallyNotAdequateAssociationvs.EstimationModelOnlyanApproximationSmallPortionofeExplained542-11-#26PotentialBiasesTimeTrends(DeclineinCHDDeath)AscertainmentChangesinDiagnosticCriteriaAvailabilityofTechnologySelectionBias542-11-#27Compliance“Adjustment