Reasonable Doubt Uncertainty in Education Science and Law

18
 Rea sonable Do ubt: Uncertaint y in Educati on, Science and Law Page 1 of 17 PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) Copyright Br itish Academy, 201 4. Al l Righ ts Res erved. U nder the t erms of the licence agreement , an i ndi vid ual u s er m ay p ri nt out a PD F of a single chapter of a monograph in BASO for personal use (for details see: htt p://britishaca demy.un iversityp res s s cholarship.com/pa ge/privacy-p olicy/pri vacy -p olicy-and-l egal-not ice  ) . Subs cri be r: Po nticia Un iversidad Cato li ca del Pe ru (PUCP); date: 30 Apri l 2015 University Press Scholarship Online British Academy Scholarship Online Evidenc e, Inference and Enqui ry P hilip Dawid, William Twin ing, and Mim i Vas ilaki Pr i nt pu bli cati on d ate: 2011 Print ISBN-13: 9780197264843 Pu bli shed to Br i ti sh Academy Sch ola rs hi p Onli ne: January 2013 DOI : 10.5871/b aca d/97 80197264843. 001.00 01 Reasonable Doubt: Uncertainty in Education, Science and Law  TON Y GAR DN ER -MEDWI N DOI:10.5871/bacad/9780197264843.003.0017 Abstract and Keywords Thi s chap ter cha mpi ons the eli ci tati on and use of pro babil i stic measur es of u ncer tai nty. I t compares two different views (roughly corresponding, respectively, to the Bayesian and frequentist approaches to statistical inference) as to how probabilities should be used as evidence. It argues that, while the former may be a logical ideal, the latter may be more appropriate in the special circumstances of the courtroom. Keywords: unc er tainty, educa tion, probabilities, ev idence, statis tic al infe re nc e Abstract Th e u se of ev i de nce to res ol ve uncer ta i nti es is k ey to ma ny ende avour s, most conspicuous l y science and law. Despite thi s, the logi c of u ncer tai nty i s seldom taught expli ci tl y, and o ften se ems mi su nde rs tood . Traditi onal edu cati onal practi ce even fails to encourage students to identify uncertainty when they express

description

Razonamiento

Transcript of Reasonable Doubt Uncertainty in Education Science and Law

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 1 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    UniversityPressScholarshipOnlineBritishAcademyScholarshipOnline

    Evidence,InferenceandEnquiryPhilipDawid,WilliamTwining,andMimiVasilaki

    Printpublicationdate:2011PrintISBN-13:9780197264843PublishedtoBritishAcademyScholarshipOnline:January2013DOI:10.5871/bacad/9780197264843.001.0001

    ReasonableDoubt:UncertaintyinEducation,ScienceandLawTONYGARDNER-MEDWIN

    DOI:10.5871/bacad/9780197264843.003.0017

    AbstractandKeywords

    Thischapterchampionstheelicitationanduseofprobabilisticmeasuresofuncertainty.Itcomparestwodifferentviews(roughlycorresponding,respectively,totheBayesianandfrequentistapproachestostatisticalinference)astohowprobabilitiesshouldbeusedasevidence.Itarguesthat,whiletheformermaybealogicalideal,thelattermaybemoreappropriateinthespecialcircumstancesofthecourtroom.

    Keywords:uncertainty,education,probabilities,evidence,statisticalinference

    AbstractTheuseofevidencetoresolveuncertaintiesiskeytomanyendeavours,mostconspicuouslyscienceandlaw.Despitethis,thelogicofuncertaintyisseldomtaughtexplicitly,andoftenseemsmisunderstood.Traditionaleducationalpracticeevenfailstoencouragestudentstoidentifyuncertaintywhentheyexpress

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 2 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    knowledge,thoughmarkschemesthatrewardtheidentificationofreliableanduncertainresponseshavelongbeenshowntoencouragemoreinsightfulunderstanding.Inourinformation-richsocietytheabilitytoidentifyuncertaintyisoftenmoreimportantthanthepossessionofknowledgeitself.

    Inbothscienceandlawtherearefundamentallydifferentkindsofuncertaintyatissue.Thereisuncertaintywhetheraparticularhypothesisiscorrect,andthereisuncertaintyaboutobservabledatathatmaybegeneratedifaparticularhypothesisiscorrect.Bothareexpressedintermsofprobabilities.Eachhasitsowndomainofapplicationanditsownlogic,buttheinterrelationshipiscomplexandsometimesmisunderstood.Hypothesisprobabilitiesarealwaysopentoerrorthroughpossiblefailuretotakeaccountofrealisticalternatives,whiletheproperinferencesthatcanbedrawnfromdataprobabilities(ofteninthecontextofsignificancetesting)arequitelimitedandeasilyover-interpreted.

    Whenconsideringthesetwokindsofprobabilityinacourtoflawitispossibletointerpretthephrasereasonabledoubtindifferentways.Itcanbeseenasaddressingdatauncertainty:whethersuchincriminatingevidencemightwithreasonableprobabilityarisetoconfrontaninnocentperson.Or(themoreconventionalview)itcanbeseenassomesortofthresholdlevelontheprobabilitythatthedefendantisguilty(ahypothesisprobability).Eachtypicallyinvolveselementsofsubjectivejudgement,butfewerissuesanduncertaintiesarisewhenconsideringthedataprobabilityanditisarguedthatthisisoftenthemorecriticalandproperissueforajurytoaddress.ThishasparticularrepercussionsforcasesinvolvingidentificationofasuspectthroughtrawlofaDNAorotherdatabase.

    (p.466) UncertaintyandmisconceptionineducationalassessmentASAUNIVERSITYTEACHER(inmedicalscience)Ihavetriedtoraisestudentsawarenessofuncertaintiesintheirownknowledge.Inconventionaleducationalassessmentstudentsareoftenmotivatedtohideuncertainties.Thisisperverse,becauseitisobviousthatitisagoodthingtobeabletodistinguishbetweenreliableandunreliableaspectsofonesknowledge.Studentsshouldinfactberewardedforacknowledginguncertaintiesthattheyhave.Byrewardingthem,forexamplewithsimplebutcarefullydesignedmarkschemes,theycanbeencouragedtoreflectonthenatureofanydoubts,andontheevidencethattheymaybeabletobringtobeartoresolvethesedoubts.Insteadofdoingthis,conventionalmarkingusuallyencouragesstudentstoblufftheirwaythrough:ittreatsluckyguessesinthesamewayaswelljustifiedknowledge,andfirmmisconceptionsasnoworsethanacknowledgedignorance.Theresultisago-for-itcultureinwhichdecisionsmaybetakeninthelightofamarginalorsuperficialpreferenceforsomeoption,withlittlefurtherthought.

    Thedangersofsuchbehaviourareveryobviousinafieldsuchasmedicine,wherelivescanbelostthroughreluctancetoacknowledgeuncertainty.Eveninordinarydiscoursehowever,itisasignofweaknesstoshyfromproperawarenessandacknowledgementofonesuncertainties.Itisdisturbing,whentalkingtostudentsfamiliarwithessaywritingwhereyoumightthinkthediscussionofuncertaintieswouldbeparamountthatthey

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 3 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    oftensaytheywouldnotnormallymentiondoubtsaboutafactorargumenttheyaretemptedtoinclude.Anyonewhohasmarkedexamessaysknowsthatinpracticeaqualification,perhapsaquestionmarkinthemargin,willmakelittledifferencetoonesjudgementifthepointmadeiscorrect,whileitmakesabigdifferenceifuncertaintyisacknowledgedaboutsomethingincorrect.

    Itmayatfirstseemparadoxicaltorewardsomeoneforacknowledginguncertainty.PerhapsitrunscountertosocietysmodernguidelinesforHowtogetoninpolitics,business,andmaybesometimesevenscienceorlaw.Buttherewardsforcautioninthefaceofuncertaintyarefundamentaltobiologicalsurvival.Childrenandyounganimalslearntheselessonsoftenthroughgames,anditisnocoincidencethattherelevantmathematicscomesundertheheadingGameTheory(vonNeumannandMorgenstern1944).Of(p.467) coursecertaintyalsobringsitsownrewardsjustaslongasitiscertaintyaboutsomethingthatiscorrect.Topursueacricketinganalogy,thedecisiontotrytohitaballforsixisfine,solongasoneisprettysureonecansucceed:thepriceoffailuretendstobehigh.Ifoneisuncertainofsuccess,thenamoremodeststrokewillonaveragebebetterrewarded.Exactlythesameprincipleisemployedinthecertainty-basedmarkschemeweemployforself-testsatUCLandImperialCollegeinLondon(Gardner-Medwin1995,2006).Acknowledginguncertainty(67%estimatedprobabilityofbeingcorrect)gives1markforacorrectanswerand0ifincorrect.Claimingahighlevelofcertainty(80%)gives3marksor6,whileanintermediatelevelgives2marksor2.Thisisaproperormotivatingmarkscheme,inthesensethatastudentalwaysexpectstogainbygivinganaccurateindicationofhowreliablehe/shethinksthechosenansweris(Dawid1986).Withthisschemethereisnobenefittobegainedbythestudenttryingtoplaythesystem:pretendingtobeeithermoreorlessconfidentthanhe/shereallyis.Itisremarkablehowreadilystudentstaketothisscheme,appreciatingthegoodsenseandutilitarianvalueofsettlingformodestrewardsandpenaltieswhenunsure.

    Uncertaintyabouttheexpressionofknowledgeorunderstandingisanexampleofuncertaintyaboutideas.Muchofeducationaltestingrevolvesaroundideasthataregenerallyaccepted(atleastwithinanagreedframework)asdefinitelyeithertrueorfalse,thoughstudentswithpartialknowledgemayassignprobabilitieslessthan1andthosewithseriousmisconceptionsmayassignverylowprobabilitiestowhatiscorrect.Thereareofcoursemanyelementstoeducationalexperiencewherethereisnocertaintyandthetaskistogenerateordiscussideasthatareneitherknownnoragreedtobeobjectivelycorrectorincorrect.Herewecomeclosertotheissuesthatariseinscientificresearchandincourtsoflaw.

    Uncertaintyaboutideas:thedrivingforceinscienceScientistsworkwithtwofundamentallydifferentkindsofuncertainty,thoughtheydontalwaysdistinguishthemveryclearly.Thefirstisanintrinsicallysubjectiveprobability:Howcertainisitthataparticularideaorhypothesisiscorrect?Thisisobviouslyfundamentalwithinscience,buttheanswerstosuchquestionsarealwayssubjectiveinthesensethattheymaydifferevenbetweenwellinformedandintelligentindividuals.Bygainingadditionalevidenceonemayshiftsuchsubjectiveprobabilitiesinwaysthatall

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 4 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    agreearerationalandoccasionallyevenquantifiable.Indeed,asPopper(1959)pointedout,onecansometimesdemonstratethatanideaiscertainlywrong,thoughneverthatitis(p.468) certainlyrightasageneraltruth.Differentlevelsofscepticismanduncertaintyamongstindividuals,aboutwhatiscorrect,mayneverbeeliminated.Nevertheless,thereareofcoursegreatswathesofscientificconclusionsthatareregarded,bymoreorlessallwhohavestudiedthem,asessentiallycertain.Itisthepowerofsciencethattheaccumulationofevidence,givenenoughtimeandeffort,tendstobecomeoverwhelmingononesideortheotherofanyargument.However,evidencedoesnotalwaysdiminishuncertainty,andevenfirmlyembeddedideascanoccasionallybechallengedoroverturned,evenbysimplepiecesofevidence.Uncertaintyaboutideasisthemotivationforprogressinscience,despiteitsnebulousandlargelyunquantifiablenature.Itdrivesthetestingofpredictionsandthedevisingofalternativetheoriesandexperimentalchallenges.

    Darwinstheoryofnaturalselectionisaprimeexampleofanideathatwasconceivedwithahugeelementofuncertainty,notleastinDarwinsmind.Darwinsgreatsuccesswasthatheamassedavastamountofevidenceinitsfavourbeforegoingpublic:enoughevidencetodiminishhisownuncertaintyandtoshaketheconvictionofmanysceptics.Thereneverwas(andneverwillbe)atimewhenDarwinstheoryistotallycertain,beyondanymeaningfulthreshold.Ofcourse,nearlyeveryscientistbelieves,asIcertainlydo,inthistheory.Darwinsideaisasfirmlyembeddedasanyinscience,thankstotheprogressiveelucidationofbiologicalmechanismsandtheaccumulationofdatathatareall(sofarasIknow)consistentwiththecoreelementsandpredictionsofthistheory.Buttherecanbenocertaintythatnewideasandobservationscouldnotchangeit.AscientistsnaturalresponsetothequestionIsDarwinstheorybeyondreasonabledoubt?istosayWhatdoubt?Letsseeifwecandeviseatestforanyideasabouthowitmaybewrong.Scientistslovecoherentscepticism,justsolongastheideasaretestable.Thehistoryofscienceteachesthemtoregarddogmaticbeliefsaspointlessandarrogant.Thestrengthofthechieftenetsofscienceisnotthattheyarecertain,butthattooverturnthemwouldrequirenotonlyobservationsinconsistentwiththem,butalsoanewexplanationforalltheevidencethathashithertoappearedtosupportthem.Darwinachievedthisinrelationtothenotionofdivinecreation,asdidEinsteinwithclassicalmechanics.Scientistsseldomofcourseachievesuchdramaticrevolutionaryimpact;butevenwhenthepursuitofuncertaintydoesntoverturnorstrengthenexistingideas,itoftenthrowslightonfacetsthatwerepreviouslyunclear.

    (p.469) UncertaintyaboutexpecteddataThesecondkindofuncertaintyisconceptuallymorestraightforwardandquantitative,thoughoftenunderpinnedbycomplexmathematics.Thisisuncertaintyaboutthedatathatwillariseinanexperiment,giventhattherelevantmechanismsandprinciplesareeitherfullyunderstoodorfullydefinedbyahypotheticalmodel.Statisticalanalysisofsuchuncertaintyispartofanyscientifictraining(andthebaneofmanystudents!).Itgivesaframeworkforevaluationofnewresults,sinceanovelclaimisoflittlevalueifthenoveltyorinterestoftheevidencecouldquitelikelyhavearisenwithconventionaloruninterestingassumptions.Thecommonestformatisthatofsignificancetesting,initiated

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 5 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    byFisher(1925).TheassumptionsonwhichasignificancetestisbasedaredescribedasanullhypothesisH0.Apointofpotentialinterestisencapsulatedinastatisticfromtheobservations,quantifyingadifferencefromthemeanormedianvalueexpectedonH0.Theoutcomeofthetestistheprobability(P-value)basedonH0that,withtheproceduresadoptedintheresearch,atleastaslargeavaluewouldbeobserved.

    SincetheuseandmisuseofsignificancetestsandP-valueshasahugeandcontroversialliterature(seeforexampleveryapproachablereviewsbyRoyall(1997)andSenn(2003)),Ishallpresentherejustapersonalperspectiveonsomeoftheseissues.Afundamentalpoint,clearIhopeinthedescriptionabove,isthataP-valueisaprobabilityofobservingdata,conditionalonaspecifichypothesis.Itdoesnotexpressuncertaintyabouthypotheses.P-valuesarehugelyvaluableinsciencebecausetheyquantifythedegreetowhichdataareconsistentwithsomeconventionalorpostulatedidea.PerhapsthemostusefulroletheyperformiswhentheP-valueislargeandtheresultisnot-significant:thismeansthatthepointofinterestinthedatawouldbequitelikelytoariseevenifH0istrue,sothereislittlereasontopayattentiontothisaspectofthedataasanykindofchallengetoH0.Onecanconcludethisquitestraightforwardly,withoutanyconsiderationofalternativehypothesesortherathernebulousbusiness,discussedinthelastsection,ofaddressingtheprobabilitiesthathypothesesmaybecorrect.

    AhazardintheinterpretationofP-valuesarisesfromterminologyoftenusedtoexpressconclusionsbasedonthem.LowP-valuesareoftensaidtojustifyrejectionofH0ataparticularsignificancelevel.Thisisonlytrueinaveryrestrictedformalsense,whenaP-valueisusedasadecisioncriterionbetweentwoactionsthatwouldbeappropriateifthereweresolidreasonstoacceptorrejectH0.ForexampleonemightdecidetoignorethedatabecauseitisconsistentwithH0,orstartresearchingalternativehypothesesifthedatawouldbesurprisingonthebasisofH0.Iftherearealternativehypotheses(p.470) alreadyformulated,withsomebasisforassigningthemrelativeprobabilities,thenaresultwithalowP-valuedoesntnecessarilyevenargueagainstH0,letalonejustifyrejectingit.AninterestingresultmaybesoextremethatitisveryunlikelyonH0,butitmaybeevenlesslikelyonwhateveralternativehypothesesareconsideredplausible.Asanexample,supposethattentossesofacoinyieldnineheadsandonetail.Doesthisresult,surprisinglyfarfrom50:50(P-value=2percent)argueforrejectionofahypothesis(H0)thatthisisafaircoin?Iftheonlyalternativesseemtobethatacoinmighthavetwoheadsortwotails,thenthedataclearlyshowsbothofthesefalse,supportingH0.However,ifoneenvisagesthepossibilitythatacoinmightbebiassedsoastobemorelikelytolandoneway,andconsidersthisplausible,thenthedatalendmoresupporttothisideathantoH0.WhetheroneendsupconsideringH0moreorlesslikelythanattheoutsetdependsontheinitialprobabilitiesoneassignstoanyofthese(andpossiblyother)ideasonespriorprobabilities.Inscience,especiallyatthefrontiers,thealternativestoawell-formulatedhypothesisareoftensimplyamatterofspeculation.InthatsituationallonecansensiblysayaboutdatawithalowP-valueisthattheparameterofinterestissurprisinglyfarfromexpectationonthebasisofH0(orsimilarhypotheses)andthatotherhypotheses,notnecessarilyveryplausibleoreventhoughtof,couldhaverendereditlesssurprising.

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 6 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    TheborderlinebetweenuncertaintyabouthypothesesanddataAusefulextensionofsignificancetestsisthecalculationofso-calledconfidencelimits:arangeofnullhypothesesthatwouldyieldP-valuesaboveaspecifiedlevelforaresultofinterestinthedatainotherwords,theywouldmaketheresultunsurprising.Thispracticehelpstoclarifyhowusefultheobservationsare:howwelltheyservetodiscriminatebetweendifferentvaluesofaparameterwithinahypotheticalmodel;butitstillsaysnothingabouttheprobabilitythatoneshouldassigntothecorrectnessofanyofthesemodels.Thecommonlyusedterminologyisagain,inmyview,misleadingbecausetheterm95percentconfidencelimitsforaparameterinamodelcaneasilybeunderstoodtomeanthatthereisa95percentprobabilitythatthetrueparameterlieswithintheselimits.Thisissimplynotso,asisevidentifyouthinkaboutanexperimentthattriestomeasureaneffectthatyouconsidertobealmostcertainlynon-existent(perhapsextrasensoryperceptionorhomeopathy).Whatevertheprotocolfortheexperiment,youshouldexpecton5percentofoccasionsthataresultwillbeobtainedforwhichanileffectisoutsidethecalculatedconfidencelimits.Ifyouhappentoexperienceoneof(p.471) theseresultsitwouldobviouslybeinappropriatetoadoptabeliefthatthereisa95percentprobabilitythatthepostulatedeffectisreal.Perhapsconsonancelimits(KempthorneandFolks1971)wouldbetterconveythetruemeaning,whichisarangeofmodelsforwhichthedatawouldberelativelycommonplacefallingwithinthe95percentmostlikelyoutcomes.

    Ingeneralitisimportanttomakeacategorydistinctionbetweentheprobabilitythatdatawillariseandtheprobabilitythatahypothesisaboutthegenerationofdataiscorrect.Sometimes,however,thereisasimplerelationshipbetweenthetwo.Forexample,anexperimentmayestimate(bysamplemeasurements)themeanvalueofaparameterthatisconsideredattheoutsetequallylikelytotakeanyofawiderangeofvalues:perhapssomecharacteristicofanewlyencounteredsubstance,objectorspecies.Then(withafew,oftenjustifiable,assumptions)1theposteriorprobabilitiesofdifferenthypothesesaboutthetruemeanarethesameasthedataprobabilitiesthatwouldbecalculatedforsamplemeans,onthehypothesisthatthetruemeaniswhatwasactuallymeasured.Returningtoterminology,thismeansthatwhennohypothesisinitiallyhaspreferredstatus,thelimitscautiouslydescribedaboveas95%consonancelimitsaretruly95percentconfidencelimitstheyidentifyarangethatonecanbefairlyconfidentincludesthetruemean(with95percentprobabilitybasedononlythesedata).Exampleslikethiswhereonecanconvertdirectlyfromadataprobabilitytoahypothesisprobabilityarenotrare,buttheyleadtoconfusioniftheirlimitationsarenotproperlytakenintoaccount.

    Whenalternativehypothesesarewelldefinedanddohavespecialstatusinanexperiment,forexamplerivalcausalexplanationsofthedata,thenupdatingofprobabilitiesforthealternativesrequiresBayesiananalysis(Lindley1972).Aswesawabove(inthecoinexample)attemptstomakesuchinferencesusingP-valuescanbequitemisleadingsinceanunlikelyresultonahypothesisH0mayactuallybeevidenceinfavourofH0.Bayesiananalysisinvolvesidentifyingpriorprobabilitiesforthevarioushypothesesandupdatingtheseinthelightofthedata,usinglikelihoodratios(therelativeprobabilitiesthatthedatawouldbeobservedonthedifferenthypotheses).Thetechnicalitiesneednot

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 7 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    concernushere,buttheprocedureisdemonstrablysuccessfulinsituationswherealternativehypotheseshaveclearpriorprobabilities.Evenwhenthisisnotthecase,andpriorprobabilitiesarelittlemorethanguesses,sufficientdatacansometimesleadtohighlyreliableinferences(seeforexample,(p.472) MacKay2003).Themainreservationisthatconclusionscanbewhollymisleadingifthecorrecthypothesishasbeenomittedfromconsiderationorassignedaninappropriateprobability.Ifthisisamajorrisk,asitcanofcoursebeatthefrontiersofresearch,thenthemorelimitedlogicbywhichsignificancetestschallengeindividualhypothesescanbeamorecomfortablebasisforscientificprogress.

    AnimportantapplicationofBayesiananalysisisinmedicaldiagnosis.Hereonestartswithfairlyobjectiveprobabilitiesthatdifferenthypothesesmightbetrue,forexamplethatapersonwithapersistentcoughmayhaveeachofanumberofdifferentdiseases.Thentheseprobabilitiesareupdatedonthebasisofdatafromthepatientshistoryandtestscarriedout,usinginformationabouthowfrequentlythedatawouldariseinthepopulationforeachoftheconditions.Thisprocessmaysimplyoccurinthedoctorsmind,leadingeventuallyeithertoconfidenceaboutspecificdiagnosesortoresidualuncertainty.Alternatively,aBayesiancomputeralgorithmmayprocessthedatatogenerateexplicitposteriorprobabilities.Whichisbetterisdebatable,giventhepossibleunreliabilityofthedoctorscompilationofinferencessetagainsthisorherabilitytotakeintoaccountsubtleaspectsofthedataignoredbyacomputeralgorithm.Butthereisnodisputingtheappropriatenessofthisstrategytoresolvemedicaluncertaintiesandtohelpmakedecisions.

    UncertaintyincriminaltrialsTheuncertaintiesinacriminaltrialhavesuperficialsimilaritiestotheproblemsofmedicaldiagnosis.Theendpointisadecisionthatturnsgradeduncertaintiesintospecificactions:whethertoconvictadefendant,orapplytreatmentforaspecificdiagnosis.Therelativeutilitiesofrightandwrongdecisionsmayinpartdeterminewhatlevelofdoubtisacceptableforafinaldecision.Inmedicinethisuncertaintyisnormallyexplicit,sincethepatientusuallymakesthefinaldecisionaboutwhethertoaccepttreatment,andmustthereforebeinformedofdoubtsandlikelyoutcomes.Inlaw,thecriterionforconviction(beyondreasonabledoubt)issubstantiallyopentoajurysinterpretation,withonlypartialknowledgeoftheutilitiesinvolvedsinceasentenceforconvictionmayonlybedeterminedatalaterstagebyajudge.Butthesearenotprofounddifferences.

    Legalcasesoftenhavemanylayersofcomplexity,forexampleconcerningfacts,motives,identification,intentionandwitnesscredibility.Finaldecisionsmayhingeonacomplexsynthesisofmanytypesofdoubt(Cohen1977;Andersonetal.2005),anditisfarfromclearthatmathematicalapproaches(p.473) arehelpfulincomingtoappropriateconclusions.Alternativeconceptionsofprobability(forexampleCohensinductiveprobabilitybasedonqualitativeeliminativereasoning)maysometimesbemorerational(oratleastmoremanageableforajury)whendecidingwhetherfactsarebeyondreasonabledoubtinatrial.Suchissuesaremuchdebatedinthelegalliterature(seeforexampleTillersandGottfried2006)withcogentargumentsfrommanyperspectives.

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 8 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    However,thedebatethatfollowsnaturallyfromthediscussionhereonthenatureofuncertaintiesinscienceisnotaboutwhetherprobabilitiesshouldbequantifiedbutwhetherweshouldbedealingwithuncertaintiesabouthypothesesoraboutdata.Ideally,ofcourse,onewouldlikecourtdecisionstobebasedoncertaintyratherthanuncertainty,butthisisrarelythecase.Iftheevidencedoessupportcertainty,thenconcludingthatdefendantDdefinitelycommittedcrimeC(anassertionaboutahypothesisprobability)isequivalenttoconcludingthattheevidencecouldnothavearisenifDwereinnocent(anassertionaboutadataprobability):eachconclusionfollowsfromtheother.Butifuncertaintiesarepresent,thesebecomedifferentquestions.

    Ihavearguedelsewhere(Gardner-Medwin2005)thatitcanberationalforajurytocometotheconclusionthatthehypothesisofguiltisveryprobablycorrect,whileacquittingonthegroundsthattheevidencecouldwithreasonablelikelihoodhavearisenforaninnocentperson.Itmayatfirstseemthattheseissuesareoppositesidesofthesamecoin,whichwouldmakemyassertioneitherparadoxicalormerelyamatterofsettingdifferentthresholdsforjudgingthetwoprobabilities.Iwillfirstsetoutacaricatureexample,basedoncasesthatfirstdrewmyinterest(asdiscussedbyDawid2002)toshowthatthisisnotthecase.

    Supposemultipleinfantsinafamilyhavediedincircumstancesconsistentwitheithersuddeninfantdeathsyndrome(SIDS:araremedicalconditionthatleavesnospecificsignspostmortem)orinfanticidebythemother.Thismayreasonablyleadtosuspicionofcrime,andindeedcaseshavecometocourtwithlittlemoreinthewayofpertinentevidence.Itisobviouslyrelevanttoaskquestionsanalogoustothoseunderlyingtestsofsignificance:Howlikelyisitthatsuchevidencewouldariseinaninnocentfamilywithcomparablegenetic,medicalandsocio-economicbackground,andhowoftenwouldsuchcasesbeexpectedtoariseinaninnocentpopulation?Suchquestionscannotbeansweredwithprecision,butcompetentexpertsshouldbeabletogivereasonablerangesfortheanswers.UnfortunatelythesupposedlyexperttestimonyinrecentUKcaseswasnotcompetent,raisingseriousconcernsabouthowexperttestimonyshouldbevalidated(RoyalStatisticalSociety2002);butthisdoesnotdetractfromtherelevanceofproperanswers.Theconclusionmightbethattheriskofthedeathsthathavebroughtthismother(p.474) tocourtarisinginherfamilywithoutcrimewasatleastaprobabilityP,andthatsimilarmultipledeathsinfamilieswithnogreaterriskfactorsformultipleSIDSmightbeexpectedtoarisesomewhereintheUKonceeveryXmonths.Ajurymightreasonablydecidetoacquitsimplyonthebasisofsuchtestimony,justifyingtheirdecisionbysayingthatifjuriesconvictonsuchevidencetheymightberesponsibleforconvictinginnocentmothersattherateofoneeveryXmonthsintheUK,whichtheydeemunacceptable.Theacceptabilitycriterionisofcourseasubjectivematter:howmanyfalseconvictionsofthissortmightbeacceptableinagivenpopulationperyear,decadeorcenturyintheinterestsofjustice.Butitisnomoresubjectivethanwhatconstitutesreasonabledoubtonanyotherbasis.

    Analternativeapproachistodecidesuchacasebasedontheprobabilitythatthehypothesisofguiltiscorrect,thatisthatthisdefendantisguiltywithinthemeaningofthe

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 9 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    law.Thisviewisoftenimplicitinthelegalliteratureandthemedia.ItwasclearlyexpressedintheletterfromtheRoyalStatisticalSociety(2002)totheLordChancellorintheUKconcerningthemultipleSIDScaseofRvs.SallyClark,andwasanalysedinrelationtothiscaseinmoredetailbyDawid(2002).Toquotetheletter:Twodeathsbymurdermaywellbeevenmoreunlikely[thantwodeathsbySIDS].Whatmattersistherelativelikelihoodofthedeathsundereachexplanation,notjusthowunlikelytheyareunderoneexplanation.Theimplication(madeexplicitbyDawid(2002)inSection2.3ofhispaper)isthatajuryshoulduseevidenceabouttheincidenceofthistypeofmurderinthepopulation,andconvictifthisissufficientlygreaterthanthatofSIDS,takingaccountofalltheknowncircumstances.Thisiscertainlyarationalwaytoinferprobabilityofguiltinlightoftheevidence,butithasuncomfortableconsequenceswhenusedasthebasisforconvictionoracquittal.Inparticular,itmeansthatadefendantwhowouldbeacquittedonthegroundsthattheevidenceisreasonablyconsistentwithinnocence(theargumentofthelastparagraph)mightfindherselfconvictedbecausesufficientlymanyotherpeoplehavecommittedthecrimeofwhichsheisaccused(Gardner-Medwin2005).Thisseemsethicallyimproperandprobablyunacceptableinlaw,becausethevoluntarycriminalactsandintentionsofpeopleinotherlegalcasescannotreasonablybeusedasanargumenttoestablishthataparticulardefendanthasbrokenthelaw.Suchevidencemightbeadmissibletodemonstrate,inunusualcases,thatpeopleareindeedsometimescapableofsurprisingbehaviour(forexample,amotherkillingherownchildren).Butaliberalsocietywouldnotbecontent,Ithink,withlegalpracticebasedonautilitarianprinciplethatifcrimeisrifethenthelawshouldconvictwithlowerstandardsofevidence,soastosuppresscrimeatthecostofimprisoningtheinnocent(Gardner-Medwin2005).Thiswouldbethe(p.475) actionofatotalitarianstate.Anenlightenedjurymustbepreparedtoacquitiftheevidencecouldplausiblyhavearisenwithoutguilt,howeverlikelyitmayseemthatthedefendantisguiltyonstatisticalgrounds.Ifcommoncrimesgounpunishedasaresult,thisneedstoberectifiedbyimprovingthequalityofevidenceratherthanbyloweringthethresholdforconviction.

    UncertaintyintrialswithevidenceofadefinitecrimeTrialsareofcourseusuallymorecomplexandlessamenabletoquantitativeanalysisthanthecaricaturediscussedabove.Commonlythereisclearevidenceofacrime,andamongstthemanyhypothesesthatmightaccountforthisevidencetherearetwoofprimaryconcern:thoseputforwardbytheprosecutionanddefence.Theprosecutioncaseentailsguiltonthepartofthedefendant,whilethatofthedefencemaynotidentifyatallwhoisguilty,merelyofferingalternativeexplanationforevidencebroughtagainstthedefendantandaddingfurtherevidencethatmaytendtoproveinnocence.Boththeprosecutionanddefencehypothesesaretypicallycompositeinthesensethatthereareunknownelementsforwhichtheremayatbestbeareasonablesetofassignableprobabilities(forexamplethatthevictimeitherwalkedorgotaliftfromAtoB).Thesituationissomewhatsimilartocomparisonoftwoscientifichypotheses,wheretherelativelikelihoodofthedataarisingthroughthealternativeexplanationscanprovidearationalbasisforincreasingbeliefinoneortheother.However,aswesawabove,therecanbefundamentalproblemsinthisprocess.

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 10 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    Thefirstproblemistheneedforpriorprobabilities,fromwhichtogenerateposteriorprobabilitiesinthelightoftheevidence.Withscientifictheorieswesawhowtheremaysimplybenowayofresolvingdifferencesofopinionaboutsuchpriors.Similarly,incourtthepriorsmaydependonsuchdubiouslysubjectivefactorsasajurorsexpectationonbeingsummonedtotakepart,suppositionsabouttherigouroftheprosecutionservices,orhowthedefendantisdressedincourt.Aspectsofthebackgroundhistoryofthedefendant,whichmayormaynotbeallowedtoemergeincourt,mayreasonablybeconsideredrelevanttopriorprobabilitiesandtherebytotheposteriorprobabilityofguilt,butconstraintsanddebateoftencentreroundthefairnessofadmittingsuchevidenceandthelegalbalancebetweenpotentialprejudicialandprobativevalue(seeforexampleRobertsandZuckerman2004:chaps.4,11).

    Asecondproblemabouttheuseofposteriorprobabilitiesarisesfromthecompositeandincompletenatureofthehypothesesunderconsideration.(p.476) Thedefencehypothesisseldomspecifieswhoactuallycommittedthecrime.Anyonefamiliarwithdetectivefiction(andprobably,thoughIdonothaveexperience,detectivefact)knowsthataprosecutioncasecanlookoverwhelminglystrongerthanadefencecaseuptothepointwhenanewideaoremergingfactsuddenlymakesplausibleahithertounconsideredsuspect,motiveoropportunity.Inanidealworldsuchadevelopmentwouldhavearisenandbeeninvestigatedbeforetrial,butmiscarriagesofjusticecanoccurifthisdoesnothappen.Thisisanalogoustothesimplecointossingexampleabove,wherethesameevidencewasseeneithertosupportornegatethehypothesisthatacoinisfair,dependingondetailsofthealternativehypothesesthatmayhavebeenthoughtof,andwhetherthesewereassignedsignificantpriorprobabilities.Miscarriagesofjusticecanveryeasilyoccurifthetruecriminalisabsentfromconsiderationorseemsabovesuspicion.

    Theseproblemsconcerningprobabilityofguiltareanalogoustotheproblemsthatmakescientistsshyfromdebateabouttheprobabilitythataparticularhypothesisistrue.Theprobabilityofguiltcannothoweverbeavoidedaltogether,becauseitwouldobviouslybeunreasonabletoconvictsomeonewithoutsomehowconcludingthatthereisahighprobabilityofguilt.Butwehaveseenbothethicalconcernsandlogicalproblemsabouttreatingthisasawell-foundedsufficientcriterionforconviction.Hence,asinscience,itisalsoconstructivetofocusonthemorelimitedbutinsomerespectsmoredirectinferencesthatmaybedrawnfromdataprobabilities.Incourt,arelevantprobabilityiswhetheraninnocentpersoncouldhavebeenbroughttocourttofaceatleasttheweightofincriminatingevidencethathasbeenseenforthedefendant.Thiscanbeasubstantialconcernwherethereisahighprofilecrime,intensepoliceeffortandtheopportunitytotrawlawidepopulation:itbecomesquitepossiblethataninnocentsuspectmaybefoundagainstwhomaconvincingcasecanbemade.ThemurderinLondonofJillDandoandthesubsequentconvictionandacquittalofBarryGeorgecometomind.Ajurymustbepreparedtorecognisethataseeminglystrongcasecanoftenbemadeagainstapersoninnocentofthecrimeatissue.

    Incourtthereisnotsuchacleardistinctiontobemadebetweenhypothesisanddata

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 11 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    probabilitiesasthereisinscience.Thisisbecausehypothesesandexplanationsofdataincourtarenotthewelldefinedstochasticmodelsthataretheusualbasisforcalculationofdataprobabilitiesinscience.Therearesubjectiveelementstotheprobabilitythatevidencemightarisewithoutguilt,forexamplehowlikelyitisthataparticularscenariowoulddeveloporthatwitnessesmaylie.Uncertaintyaboutwhetherevidencemightariseforaninnocentdefendantisthereforeinsomerespectsjustassubjectiveandarguableasisuncertaintyaboutthehypothesisofguilt.However,itinvolvesfewer(p.477) priorsandcanbejudgedwithinamorelimitedframework.Thisframeworkfocuseslargelyonthethingsthatcouldhappentoinnocentpeople,ratherthanonthebehaviourofcriminals.Anincidentaladvantage(thoughhardlyareasoninitselfforpreferringthisapproach)isthatsuchscenariosmaybeeasierforajurytoenvisage.

    Selectionofsuspects:theDNAdatabasecontroversyAhazardfamiliartostatisticiansisthepotentialuseofdatatwiceover,oncetoselectahypothesisofinterest(inthiscase,topickasuspectfromapopulation)andthenagainasevidencethatthehypothesis(guiltofthissuspect)istrue.Therecanbeanelementofthisinpoliceprotocols(asintheSIDS/murdercases,probablytheBarryGeorgecaseandmanycasesinvolvingDNAmatching).Theprocedurestoavoidlogicalerrorsduetodoubleuseofdataarenotalwaysappreciatedbyscientists,soitisunsurprisingthatjuriesmayhavedifficultyhandlingtheissuecorrectly.Theymayevennotbeawarethatasuspectcametoattentionthroughatrawlofpolicerecordsratherthanthroughaconnectionwiththecase,sincethefactthatasuspecthadapolicerecordmaybeconsideredprejudicialincourt(seee.g.Kaye2009forexamplesandcourtrulingsinrelationtoDNAtesting).Therearediverseviewsandheateddebateinlegalandstatisticalcirclesabouthowtohandleevidencethatservesthesedualroles.Forexample,theAmericanNationalResearchCouncilCommitteeonDNAForensicScience(1996),Stockmarr(1999)andDevlin(2000)advocatedowngradingtheweightofevidencewhenthesuspectisidentifiedbyasinglematchinadatabasetrawl,whileBalding(2002,2005),Dawid(2002),Kaye(2009)andothersclaimthattheuseofadatabasedoesnotdiminishtheevidenceindeedstrengthensit,albeitusuallyonlyslightly,byrulingoutthoseinthedatabasewhotestnegative.Thedifferencescancorrespondtolargefactorsintheestimatedprobabilityofafalseconviction:factorscomparable(asshownbelow)withthenumberofpotentialsuspects.AsDawid(2002)pointsout,differencesbetweenstatisticalapproachestouncertaintymustrarelyhavesuchseriouspotentialconsequencesforthoseaffected.

    Theissuescanbesomewhatclarifiedbyconsideringthesafetyofconvictionshowlikelytheproceduresaretoincriminateinnocentpersons.SupposethatthepersonwholeftDNAatacrimesceneisconsideredcertaintobethetrueperpetratorofacrime(TP),andthatthetechnicalprobabilityofthesamplematchingarandompersonotherthantheTPisaverysmallnumber(p),withnopossibilitythatamatchwouldfailtobeevidentiftheTPistested.Supposethereisaprimesuspectwhoisestimatedonthebasisofnongenetic(p.478) evidencetohavea50percentprobabilityofbeingguilty(s=50%:strongsuspicion,butcertainlynotenoughonitsowntoconvict).IftheDNAprofileofthissuspectistestedandmatchesthecrimesample,thentheprobabilitythatthismatchis

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 12 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    falseisp/(1+p),oralmostexactlyp.Convictiononthisbasishasasmallestimatedprobabilitypofbeingamiscarriageofjustice.Theriskisgreateriftheevidenceagainsttheprimesuspectisweaker(s0.5),thoughwithonlyasmallnumber(M)ofpossiblesuspectsthiswouldmakelittledifferenceforatestthatconfirmsaprimesuspect.2

    CasesinvolvingtrawlofaDNAdatabasearemorecomplex.Supposethereareinitiallynoprimesuspectsbasedonnon-geneticevidence,butalargenumberNofpotentialsuspects(perhapsallthemalesofaplausibleagewithinacityonagivenday).SupposeaDNAdatabasecontainsprofilesforarandomsubsetincludingDofthesepossiblesuspects,andthatasearchofthedatabasehasrevealedexactlyoneoftheseasmatchingtheDNAatthecrimescene.ThereisattheoutsetaprobabilityD/Nthatthedatabaseincludesthetrueperpetrator(TP),andthereforeaprobability(D/N)(1p)D1thathewouldgiverisetoauniquematch,takingaccountoftheprobability(1p)D1thatnoadditionalmatcharisesbychanceforanyofthe(D1)innocentpersonsinthedatabase.Butthereisalsoaprobability(1D/N)pD(1p)D1thattheTPisnotinthedatabaseandthatexactlyoneoftheinnocentpersonsinthedatabasedoesmatch.Giventhefactthatauniquematchhasbeenfound,theprobabilitythatthisistheTPis1/(1+p(ND)).Thereisacomplementaryprobability(approximatelyp(ND)ifthisis1)thatthisisafalsematch,typicallymuchgreaterthantheprobabilityofafalsematchwhenasingletestiscarriedoutonaprimesuspect.3Thefactorbywhichthisriskisincreasedcomparedwiththeillustrativeexampleinthelastparagraphis(ND),thenumberofpotentialsuspectsoutsidethedatabasepotentially100,000ormoreinsomecases(Kaye2009).4OfcourseamatchfoundinaDNAtrawl(p.479) leadstoinvestigationofthesuspectandfurthernon-geneticevidence.Insomecasesthismayleadtoconfidentconviction(e.g.ifthesuspectturnsouttomatchaCCTVimagefromthecrimescene)oreliminationoracquittalifthematchisclearlyfalse(e.g.ifthereisadefinitealibi).ButDNAevidenceoftenappearstothejurytobethestrongestevidenceinacaseandconvictionfollowingaDNAtrawlmayevenresultwhenallotherevidenceseemstoweighinthedefendantsfavour(Donnelly2005;Kaye2009).ItiscrucialinsuchcasesthatajuryshouldrecognisethattheDNAdatacomesfromatrawlandcarriesamuchgreaterriskoffalseincriminationthanifitwereconfirmingsuspicionofaprimesuspect.Withoutthis,suchtrialsmustbeconsideredunsafe.

    HowarewetoreconcilethisconclusionwiththeseeminglycontraryBayesianargumentpresentedbyDawid(2002)andKaye(2009)thattheorderinwhichDNAandnon-geneticdataareobtainedisimmaterial.Theyarguethatsinceatrawleliminatessuspectsaswellasidentifyingonewhomatches,whencombinedwithotherevidenceitwillrationallyleadtogreaterprobabilityofguiltthanwouldbeinferredifthesamepersonhadbeentestedastheprimesuspect.ThisargumentisillustratedinFigure17.1forthesimplecaseconsideredabove,wherethereisstrongnon-geneticevidence(E)againstadefendant,eitheremergingattheoutsettojustifyconfirmatoryDNAtestingofthedefendantasaprimesuspectorelseemergingafterthesuspecthasbeenidentifiedinatrawl.Thisshowshowtheoddsonthesuspectsguiltareaccumulatedtosimilaroutcomes,simplyinadifferentorder.Thisnear(p.480) equivalenceofoutcomeshasbeenusedtochallengethenotionthatjuriesneedtobeinformedofthespecialfactsand

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 13 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    statisticalconsiderationswhensuspectidentificationoccursthroughtrawlsormultipletesting(Balding2002,2005;Kaye2009).

    Figure17.1 .ComparisonofBayesiantreatmentsofprobabilityofguilt.Arrowsshowsuccessivechangesinprobabilityofguiltandassociatedodds,forasuspectidentified(1)bynon-geneticevidence(E)followedbyaconfirmatoryDNAtest,or(2)throughtrawlinaDNAdatabasefollowedbyinvestigationanddiscoveryofE.Thenon-geneticevidenceonitsownleadstoaprobabilityofguilt(s)thatistakenheretobe50percent.N,Dandpareexplainedinthetext.Dashedarrowsshowthe(approximate)probabilityofaDNAmatchbeingfalsewitheachprocedure.Ifsuchamatchwereknowntobefalsethiswouldlead,givenassumptionsinthetext,tocertainacquittal.

    Iwouldwhollyacceptthisargumentifthecollectionandevaluationofnon-geneticevidencewererigorous,quantitative,freeofselectionbiasanduncertainties,andderivedfrominvestigationofallpossibleavenues.Butthisisunrealistic,sometimesevenforevidencebackedbyscience.Therealityisthatmuchevidencedependsonhunchesandqualitativeargument,andthoroughinvestigationhasoftenbeenrestrictedtojustoneorafewofthepossiblesuspects.Thedegreeofsuspicionappropriateforaprimesuspectisoftenhighlyuncertain,andthediscoveryofapositiveDNAmatchprovidesstrongvindicationofthepotentiallyuncertainargumentsthatledtothestatusasprimesuspect.WhenasuspectisinsteadidentifiedbyDNAtrawl,theargumentsconstructedaroundnon-geneticevidenceobtainedafteridentificationwilllackthattestofauthenticityandmustretainalltheircaveatsanduncertainties.Contrastforexample,twosituations.5In(A),SbecomesasuspectbecauseanunreliableinformantassertsSdidit,SisDNAtestedandprovestomatchthecrimescene.In(B),SissuspectedbecausetrawlofadatabaserevealsthatheisaDNAmatchandthentheinformantaddshisassertionSdidit.Inneithercasewilltheinformantsblandassertioncarrysignificantweight,butitisessentialthatajuryshouldknowwhichscenarioappliestothecase,becausetheprobabilitythattheprocedurewouldproduceaDNAmatchincriminatinganinnocentpersonismuchgreaterin(B)than(A),byafactorequaltothenumberofplausiblesuspectswithinthedatabase.Aworthwhileapproachisforajurytoconsiderintrawlcaseswhether,ifsubsequentnon-geneticevidencehadbeenestablishedattheoutset,itwouldhavejustifiedthedefendantbeingtestedastheprimesuspect.OnlythenistheprobabilityoffalseincriminationbyDNAaslowastherandommatchprobabilityp.Incase(B)abovethisconsiderationwouldhighlightthecrucialimportanceofknowingwhethertheinformantspost-trawlassertionwasmadewithorwithoutknowledgethatSwasbythenasuspect.

    Mythesishereisthatcourtsneedtoaddresstheprobabilitythatpoliceprocedures

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 14 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    (includingtrawlsforsuspects)couldhaveproducedevidenceagainstaninnocentpersonthatisatleastasincriminatingasthatpresented.Thisisaformofdataprobabilityandishowthecourtshouldultimatelyjudgethesafetyofaconviction.Itisajudgmentthattakesaccountofmorethanjustfactslinkingthedefendanttothecase,butsoitshould.Thedefendant(p.481) isincourtduetooperationoftheseprocedures,whichmustthereforebeunderscrutinyjustasmuchasthedefendanthimself.Ajudgmentthatfocusesonlyontheprobabilitythatthisdefendantisguiltywillnecessarilybedefi-cient.Itwouldessentiallyapportionprobabilitybetweenallpossiblesuspects,onlyoneofwhomisfullyinvestigatedandpresented,whilethetrueculpritandnatureofthecrimemaynotevenhaveenteredconsideration.Thereisastronganalogyheretothedifficultyofjudginghowlikelyitisthataparticularscientifichypothesisiscorrect:thisisalwaysuncertainbecausewesimplyneverknowifwehavehadtheluck,insightorimaginationtoconsiderthecorrectexplanationofthedataweobserve.

    ConclusionAcknowledgementandcharacterisationofuncertaintiesarekeyelementsintheapplicationofknowledgeandevidence.Alackofawarenessofthenatureorextentofuncertaintycanleadtoconfusionandinappropriatedecisionsinareasofeducation,scienceandlaw.Myfocusherehasbeenonadistinctionthatcutsacrossthemajordebatesthatoftenemergeinstatistics(aboutthedefinitionofprobabilityinfrequentistorBayesianterms)andinlaw(aboutthemeritsofquantitativeorqualitativeapproachestodecisionmakingincourt).ThedistinctionIamconcernedwithisbetweenprobabilitiesassignedtohypothesesandprobabilitiesfortheobservationofdata,conditionalonaspecifichypothesis.Thereisnotachoicebetweenaddressingjustoneortheother:eachhasitsownlogicanditsdomainsofrigorousapplicationinbothscienceandlaw.Botharecentraltotheprocessofestablishingknowledge,whichitselfconsistsofgradationsofbeliefinhypothesesjustified,atleastpartially,byevidencefromdata.Thelinkbetweenhypothesisanddataprobabilitiesisseldomrigorousineitherscienceorlaw.Therearegreyareaswhereconclusionsmaybemattersofopinion.Scienceingeneralgetsroundthisproblembyrelyingontheindefiniteaccumulationofevidencethatwillultimatelyswampallbutthemostextremedifferencesofopinion.Criminallawreliesmoreonthemoderatingeffectofajurydecisiontogenerateanoutcomethatisgenerallyacceptedasreasonable.Inthelegalcontexthowever,Iarguethattheusualfocusonprobabilitythatthedefendantisguilty(ahypothesisprobability)isinadequate.Clearbenefitsariseifaverdictisconsideredtobeultimatelyconstrainedbyadataprobability:howlikelyisitthatsuchincriminatingevidencecouldhavearisenforaninnocentperson?

    (p.482) Note.IamgratefultoD.KayeandD.Baldingandtotheeditorsofthisvolumeforconstructivecomments.

    References

    Bibliographyreferences:

    Anderson,T.,Schum,D.andTwining,W.(2005),AnalysisofEvidence,2ndedn.(Cambridge,CambridgeUniversityPress).

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 15 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    Balding,D.J.(2002),TheDNAdatabasesearchcontroversy,Biometrics,58:2414.

    Balding,D.J.(2005),Weight-of-EvidenceforForensicDNAProfiles(Chichester,JohnWiley).

    Cohen,L.J.(1977),TheProbableandtheProvable(Oxford,ClarendonPress).

    Dawid,A.P.(1986),ProbabilityForecasting,inEncyclopediaofStatisticalSciences,vol.7,ed.S.Kotz,N.L.JohnsonandC.B.Read(NewYork,Wiley-Interscience),pp.21018.

    Dawid,A.P.(2002,)BayessTheoremandweighingevidencebyjuries,inBayessTheorem,ed.R.Swinburne,Proc.Brit.Acad.113:7190

    Devlin,B.(2000),TheevidentiaryvalueofaDNAdatabasesearch,Biometrics,56:1276

    Donnelly,P.(2005),Appealingstatistics,Significance,2:468.

    Fisher,R.A.(1925),StatisticalMethodsforResearchWorkers(Edinburgh,OliverandBoyd).

    Gardner-Medwin,A.R.(1995),Confidenceassessmentintheteachingofbasicscience,AssociationforLearningTechnologyJournal,3:805.

    Gardner-Medwin,A.R.(2005),Whatprobabilityshouldajuryaddress?,Significance,2:912.Availableonlineat.

    Gardner-Medwin,A.R.(2006),Confidence-basedmarkingtowardsdeeperlearningandbetterexams,inInnovativeAssessmentinHigherEducation,ed.C.BryanandK.Clegg(London,Routledge,TaylorandFrancisGroup).

    Kaye,D.(2009),Roundinguptheusualsuspects:alegalandlogicalanalysisofDNAdatabasetrawls,NorthCarolinaLawReview,87:425503.

    Kempthorne,O.andFolks,L.(1971),Probability,StatisticsandDataAnalysis(Ames,IA,IowaStateUniversityPress).

    Lindley,D.V.(1972),BayesianStatistics:areview(Philadelphia,PA,SocietyforIndustrialandAppliedMathematics).

    MacKay,D.J.C.(2003),InformationTheory,InferenceandLearningAlgorithms(Cambridge,CambridgeUniversityPress).

    NationalResearchCouncilCommitteeonDNAForensicScience(1996),AnUpdate:theevaluationofDNAforensicDNAevidence(Washington,DC,NationalAcademyPress).

    Popper,K.(1959),TheLogicofScientificDiscovery(London,Hutchinson).

    Roberts,P.andZuckerman,A.(2004),CriminalEvidence(Oxford,OxfordUniversity

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 16 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    Press).

    Royall,R.M.(1997),StatisticalEvidence:alikelihoodparadigm(London,ChapmanandHall).

    RoyalStatisticalSociety(2002),LetterfromthePresidenttotheLordChancellorregardingtheuseofstatisticalevidenceincourtcases..

    Senn,S.(2003),DicingwithDeath:chance,riskandhealth(Cambridge,CambridgeUniversityPress).

    Song,Y.S.,Patil,A.,Murphy,E.E.andMontgomery,S.(2009),AverageprobabilitythataColdHitinaDNAdatabasesearchresultsinanerroneousattribution,J.Forensic.Sci.,54:227.

    Stockmarr,A.(1999),LikelihoodratiosforevaluatingDNAvidencewhenthesuspectisfoundthroughadatabasesearch,Biometrics,55:6717.

    Tillers,P.andGottfried,J.(2006),Acollateralattackonthelegalmaximthatproofbeyondareasonabledoubtisunquantifiable,Law,ProbabilityandRisk,5:13557.

    vonNeumann,J.andMorgenstern,O.(1944),TheoryofGamesandEconomicBehavior(Princeton,NJ,PrincetonUniversityPress).

    Notes:ProceedingsoftheBritishAcademy,171,465483.TheBritishAcademy2011.

    (1)Forexample,measurementsthatarenormallydistributedrandomvariableswithmean,wherethepriorprobabilitydistributionforisessentiallyuniformoverarangemuchlargerthanthemeasurementstandarddeviation.

    (2)Ifthepriorprobabilityofguiltofthematchingsuspectiss,basedonnon-geneticevidenceandeliminationofanypreviouslytestedsuspects,thentheprobabilitythatthismatchisfalseisapproximatelyp(1-s)/s(assumingthisitselfissmall).Evenifsfortheprimesuspectisconsideredunquantifiable,butthereareatmostMsuspectswhoareatallplausible,smustrationallybetakenasatleast1/Monthebasisthathisprobabilityofguiltmustbeconsideredatleastashighaseachoftheothers.Thissetsanupperboundontheprobabilityofafalsematchasp(M1).

    (3)AdifferentderivationofthesameconclusionisgivenbySongetal.(2009).

    (4)Notethatalargerandmorerelevantdatabase(largerD)reducestheriskthatauniquematchwillbeafalsematch.ThiscanbeseenasanargumentinfavourofmaintaininglargeDNAdatabases.Largedatabaseswouldalsoincreasethefrequencywithwhichdatabaseusewouldgeneratesuspectsforinvestigation.However,theremightbelegitimateconcernsthatinappropriatehandlingincourtofthestatisticalandtechnicalissuessurroundingdatabaseusemightleadtomorefalseconvictions.

  • Reasonable Doubt: Uncertainty in Education, Science and Law

    Page 17 of 17

    PRINTED FROM BRITISH ACADEMY SCHOLARSHIP ONLINE (britishacademy.universitypressscholarship.com). (c) CopyrightBritish Academy, 2014. All Rights Reserved. Under the terms of the l icence agreement, an individual user may print out a PDF of asingle chapter of a monograph in BASO for personal use (for detailssee: http://britishacademy.universitypressscholarship.com/page/privacy-policy/privacy-policy-and-legal-notice ). Subscriber: PontificiaUniversidad Catolica del Peru (PUCP); date: 30 April 2015

    (5)IamgratefultoD.Kayeforsuggestingconsiderationofthisscenario.

    Accessbroughttoyouby: PontificiaUniversidadCatolicadelPeru(PUCP)