Stat 13, Intro. to Statistical Methods for the Life and ...frederic/13/F16/day07.pdf · Random sampling, random assignment, and blocking. 3. ... • Confounding variables are often

Stat 13, Intro. to Statistical Methods for the Life and Health Sciences.

1. Experiments and aspirin example. 2. Random sampling, random assignment, and blocking. 3. Blinding. 4. Portacaval shunt example. 5. Coverage, non-response bias, adherer bias, and clofibrate example. 6. More about confounding factors.

Finish reading chapter 4. http://www.stat.ucla.edu/~frederic/13/F16 .HW2 is due Oct 18 and is problems 2.3.15, 3.3.18, and 4.1.23.FreetutoringintheStatsClub,MondaysandTuesdaysfrom5:00-6:00pminFranz2258A

1

1.Experimentsandaspirinexample.

• Inanexperiment,theresearcherssettheleveloftheexplanatoryvariableforeachsubject.

• Theselevelsmaycorrespondtoatreatmentandcontrol.

• Welldesignedexperimentscancontrolforconfoundingvariablesbymakingthetreatmentandcontrolgroupsverysimilarexceptforwhattheexperimentermanipulates.

Aspirinexample.

Physicians’HealthStudyI (studyaspirin’saffectonreducingheartattacks.

• Startedin1982with22,071malephysicians.

• Thephysicianswererandomlyassignedintooneoftwogroups.

• Halftooka325mgaspirineveryotherdayandhalftookaplacebo.

Results• Intendedtogountil1995,theaspirinstudywas

stoppedin1988afterfindingsignificantresults.• 189(1.7%)heartattacksoccurredintheplacebo

groupand104(0.9%)intheaspiringroup.(45%reductioninheartattacksfortheaspiringroup.)

• Whataboutconfoundingvariables?Couldtheaspiringroupbedifferentthantheplacebogroupinsomeotherways?– Didtheyhaveabetterdiet?– Didtheyexercisemore?– Weretheygeneticallylesslikelytohaveheartattacks?– Weretheyyounger?

TheBigIdea• Confoundingvariablesareoftencircumventedin

experimentsduetotherandomassignmentofsubjectstotreatmentgroups.

• Randomlyassigningpeopletogroupstendstobalanceoutallothervariablesbetweenthegroups.

• Soconfoundingvariables,includingonestheresearchersdidn'tanticipate,shouldberoughlyequalizedbetweenthetwogroupsandthereforeshouldnotbeconfounding.

• Thus,causeandeffectconclusionsaresometimespossibleinexperimentsthroughrandomassignment.(Itmustbeawellrunexperiment.)

2.Randomsamplingandrandomassignment.

• Withobservationalstudies,randomsamplingisoftendone.Thispossiblyallowsustomakeinferencesfromthesampletothepopulationwherethesamplewasdrawn.

• Withexperiments,randomassignmentisdone.Thispossibleallowsustoconcludecausation.

• ThePhysician’sHealthStudy usedrandomassignment.Diditalsouserandomsampling?

• No,hardlyanyexperimentsuserandomsampling,butgettheirsubjectsinotherways.

• ThePhysician’sHealthStudysentoutinvitationlettersandquestionnairestoall261,248malephysiciansbetween40and84yearsofagewholivedintheUnitedStates.

• Ofthe59,285whowerewillingtoparticipateinthetrial,26,062weretoldtheycouldnotbecauseofsomemedicalconditionorcurrentmedicaltreatment.

• Sotowhatgroupcanwegeneralizetheresultsthattakingaspirincanreduceheartattacks?– Justphysiciansinthestudy?– Allmalephysiciansbetween40-84yearsold?– Allmalephysicians?– Allmalesbetween40-84yearsolds?– Allmales?– Everyonebetween40-84yearsold?– Everyone?

ArticleBaselineDemographicsAfterRandomAssignment

Parameter Placebo(n=129)

Uceris(n=128)

Meanage,years(range) 39.9(12–68) 37.6(13–66)Men 77(59.7) 70(54.7)Women 52(40.3) 58(45.3)Meandiseaseduration(yrs) 6.3 5.5Duration≤1 year,n(%) 23(17.8) 28(21.9)Duration>5 years,n(%) 51(39.5) 44(34.4)Proctosigmoiditis 64(49.6) 58(45.3)Left-sidedcolitis 44(34.1) 37(28.9)MeanbaselineUCDAIscore 6.2 6.5MeanbaselineEIscore 6.6 6.5Priormesalazine use 75(58.1) 66(51.6)Priorsulfasalazineuse 28(21.7) 33(25.8)

Sandborn WJ,TravisS,MoroL,JonesR,Gautille T,Bagin R,HuangM,YeungP,BallardED2nd Once-dailybudesonideMMX®extended-releasetabletsinduceremissioninpatientswithmildtomoderateulcerativecolitis:resultsfromtheCOREIstudy. Gastroenterology 2012Nov;143(5):1218-26

BlockingandRandomAssignment

• Thegoalinrandomassignmentistomakethetwogroupsassimilaraspossibleinallwaysotherthanthetreatment.

• Sometimethereareknownconfoundersandyoucanblockon(controlfor)thesevariables.

• Forexample,ifoursubjectsconsistof60%femalesand40%males,wecanforceeachgrouptobe60%femaleand40%male,usingamatchedpairdesign.

• Blockingmakessensewhenthereareknownconfoundersyouwanttocontrolfor.Butrandomlyassigningsubjectstogroupsmakesthemassimilaraspossibleevenintermsofunknownconfounders.

3.Blinding.

Eveninexperiments,thetreatmentandcontrolgroupscanbedifferentinwaysotherthantheexplanatoryvariable.Thisisespeciallytruewhentheresponsevariableissomewhatsubjective.Painisanexample.Onestudyfoundthat1/4ofpatientssufferingfrompost-operativepain,whengivenaplacebo(justapillofsugarandwater)claimedtheyexperienced"significantpromptpainrelief".Thisshowsthatpeoplecannotjudgetheirownlevelsofpainverywell,andmaybeinfluencedbythebeliefthattheyhavetakenaneffectivetreatment.Thusinanexperimentwithsucharesponsevariable,researchersshouldensurethatwhoeverisrecordingtheresponsevariable(typicallythesubjecthim/herself)doesnotknowwhetherthesubjectreceivedthetreatmentorthecontrol.Thisiscalledblinding.

3.Blinding.Peoplemightnotbeabletojudgetheirownlevelsofpainverywell,andmaybeinfluencedbythebeliefthattheyhavetakenaneffectivetreatment.Thusinanexperimentwithsucharesponsevariable,researchersshouldensurethesubjectdoesnotknowwhetherheorshereceivedthetreatmentorthecontrol.Thisiscalledblinding.Inadouble-blind experiment,neitherthesubjectnortheresearcherrecordingtheresponsevariableknowstheleveloftheexplanatoryvariableforeachsubject(i.e.treatmentorcontrol).

4.Portacaval shuntexample.Thefollowingexampleshowstheimportanceofdoingarandomizedcontrolledexperiment.Theportacaval shuntisamedicalprocedureaimedatcurbingbleedingtodeathinpatientswithcirrhosisoftheliver.Thefollowingtablesummarizes51studiesontheportacaval shunt.Thepoorlydesignedstudieswereveryenthusiasticaboutthesurgery,whilethecarefullydesignedstudiesprovethatthesurgeryislargelyineffective.

DegreeofenthusiasmDesign HighModerateNoneNocontrols 24 7 1Controls,butnotrandomized 10 3 2Randomizedcontrolled 0 1 3

4.Portacaval shuntexample.Whydidthepoorlydesignedstudiescometothewrongconclusion?Alikelyexplanationisthatinthestudieswherepatientswerenotrandomlyassignedtothetreatmentorcontrolgroup,byandlargethehealthierpatientsweregiventhesurgery.Thisalonecouldexplainwhythetreatmentgroupoutlivedthecontrolgroupinthesestudies.

DegreeofenthusiasmDesign HighModerateNoneNocontrols 24 7 1Controls,butnotrandomized 10 3 2Randomizedcontrolled 0 1 3

5.Moreproblemswithstudies,andClofibrate example.Surveysareobservational.• Coverageisacommonissue.Coverageistheextenttowhich

thepeopleyousampledfromrepresenttheoverallpopulation.Asurveyatafancyresearchhospitalinawealthyneighborhoodmayyieldpatientswithhigherincomes,highereducation,etc.

• Non-responsebiasisanothercommonproblem.Poorcoveragemeansthepeoplegettingthesurveydonotrepresentthegeneralpopulation.Non-responsebiasmeansthatoutofthepeopleyougavethesurveyto,thepeopleactuallyfillingitoutandsubmittingitaredifferentfromthepeoplewhodidnot.

• Sameexactissuesinwebsurveys.

5.Moreproblemswithstudies,andClofibrate example.Non-responsebiasissimilartoadhererbias,inexperiments.Adrugcalledclofibrate wastestedon3,892middle-agedmenwithhearttrouble.Itwassupposedtopreventheartattacks.1,103assignedatrandomtotakeclofibrate,2,789toplacebo(lactose)group.Subjectswerefollowedfor5years.Isthisanexperimentoranobservationalstudy?

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

5.Moreproblemswithstudies,andClofibrate example.Non-responsebiasissimilartoadhererbias,inexperiments.Adrugcalledclofibrate wastestedon3,892middle-agedmenwithhearttrouble.Itwassupposedtopreventheartattacks.1,103assignedatrandomtotakeclofibrate,2,789toplacebo(lactose)group.Subjectswerefollowedfor5years.Isthisanexperimentoranobservationalstudy?

Itisanexperiment.DoesClofibrate work?Clofibrate patientswhodiedduringfollowup

adherers 15%non-adherers 25%total 20%

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

--------------------------------------------------------------------------Placebo

adherers 15%nonadherers 28%total 21%

Thosewhotookclofibrate didmuchbetterthanthosewhodidn'tkeeptakingclofibrate.Doesthismeanclofibrate works?

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

--------------------------------------------------------------------------Placebo

adherers 15%nonadherers 28%total 21%

Thosewhoadheredtoplaceboalsodidmuchbetterthanthosewhostoppedadhering.

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

--------------------------------------------------------------------------Placebo

adherers 15%nonadherers 28%total 21%

Allinalltherewaslittledifferencebetweenthetwogroups.

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

--------------------------------------------------------------------------Placebo

adherers 15%nonadherers 28%total 21%

Adherersdidbetterthannon-adherers,notbecauseofclofibrate,butbecausetheywerehealthieringeneral.Why?

Clofibrate patientswhodiedduringfollowupadherers 15%non-adherers 25%total 20%

--------------------------------------------------------------------------Placebo

adherers 15%nonadherers 28%total 21%

Adherersdidbetterthannon-adherers,notbecauseofclofibrate,butbecausetheywerehealthieringeneral.Why?• adherersarethetypetoengageinhealthierbehavior.• sickpatientsarelesslikelytoadhere.

6.Moreaboutconfoundingfactors.• Byaconfoundingfactor,wemeananalternativeexplanation

thatcouldexplaintheapparentrelationshipbetweenthetwovariables,eveniftheyarenotcausallyrelated.Typicallythisisdonebyfindinganotherdifferencebetweenthetreatmentandcontrolgroup.Forinstance,differentstudieshaveexaminedsmokersandnon-smokersandhavefoundthatsmokershavehigherratesoflivercancer.Oneexplanationwouldbethatsmokingcauseslivercancer.Butisthereanyother,alternativeexplanation?

• Onealternativewouldbethatthesmokerstendtodrinkmorealcohol,anditisthealcohol,notthesmoking,thatcauseslivercancer.

6.Moreaboutconfoundingfactors.• Anotherplausibleexplanationisthatthesmokersareprobably

olderonaveragethanthenon-smokers,andolderpeoplearemoreatriskforallsortsofcancerthanyoungerpeople.

• Anothermightbethatsmokersengageinotherunhealthyactivitiesmorethannon-smokers.

• Notethatifonesaidthat“smokingmakesyouwanttodrinkalcoholwhichcauseslivercancer,”thatwouldnotbeavalidconfoundingfactor,sinceinthatexplanation,smokingeffectiveiscausallyrelatedtolivercancerrisk.

6.Moreaboutconfoundingfactors.• Aconfoundingfactormustbeplausiblylinkedtoboththe

explanatoryandresponsevariables.Soforinstancesaying“perhapsahigherproportionofthesmokersaremen”wouldnotbeaveryconvincingconfoundingfactor,unlessyouhavesomereasontothinkgenderisstronglylinkedtolivercancer.

• Anotherexample:left-handednessandageatdeath.PsychologistsDianeHalpernandStanleyCoren lookedat1,000deathrecordsofthosewhodiedinSouthernCaliforniainthelate1980s andearly1990sandcontactedrelativestoseeifthedeceasedwererighthanded orlefthanded.Theyfoundthattheaverageagesatdeathofthelefthanded was66,andfortherighthanded itwas75.Theirresultswerepublishedinprestigiousscientificjournals,NatureandtheNewEnglandJournalofMedicine.

6.Moreaboutconfoundingfactors.Allsortsofcausalconclusionsweremadeabouthowthisshowsthatthestressofbeinglefthanded inourrighthanded worldleadstoprematuredeath.

6.Moreaboutconfoundingfactors.• Isthisanobservationalstudyoranexperiment?

6.Moreaboutconfoundingfactors.• Isthisanobservationalstudyoranexperiment?Itisanobservationalstudy.• Arethereplausibleconfoundingfactorsyoucanthinkof?

6.Moreaboutconfoundingfactors.• Aconfoundingfactoristheageofthetwopopulationsin

general.Leftiesinthe1980swereonaverageyoungerthanrighties.Manyoldleftieswereconvertedtorightiesatinfancy,intheearly20thcentury,butthispracticehassubsided.Thusinthe1980sand1990s,therewererelativelyfewoldleftiesbutmanyyoungleftiesintheoverallpopulation.Thisaloneexplainsthediscrepancy.