Leek - Adaptive Procedures in Psychophysical Research

download Leek - Adaptive Procedures in Psychophysical Research

of 14

Transcript of Leek - Adaptive Procedures in Psychophysical Research

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    1/14

    Perception&Psychophysics

    2001,63(8),1279-1292

    Thestudyofsensationandperceptioninhumansandotheranimalsisexpensiveandinvolvesanumberofpit-fallsanddifficulties.Fechner,in1860,firstrecognizedthatinnerconsciousnessmightbemeasurablebyoutwardbehavior.Itisthemeasurementofthatbehaviordirectly,interpretedasanindirectmeasureofperception,thatisthepurviewofpsychophysics.Overmorethanacentury,meth-odshavebeendevelopedandrefinedthatsupportthesys-tematicexplorationwithinsensorysystemsofthelimitsofdetectionanddiscriminationamongsimilarandconfus-ablephysicalstimuli.

    The measurement methodologies developed sinceFechnersrealizationhaveastheirprimarygoalthevalidreflectionofsensoryevents.Inordertohaveconfidenceinthevalidityandreliabilityofsuchmeasures,manysamplesofagivenbehaviortypicallymustbeobservedinastruc-turedandsystematicprocess,inresponsetocarefullycon-structedstimuli.Changesinstimulusstrengthorothercharacteristicsareassociatedwithchangesintheabilitytodetectordiscriminatesuchstimuli.Measuresofperfor-manceonpsychophysicaltasksasafunctionofstimulusstrengthorothercharacteristicsconstituteapsychometricfunction.Acompletecharacterizationofpsychometricperformanceasafunctionofchangesinstimulusstrengthmaybedevelopedbyusingthemethodofconstantstimuli,oneoftheclassicalpsychophysicalmethodologies.Withthisprocedure,asetofstimuliwithstrengthsspanningtherangeofsensationfromimperceptibletoconsistentlyper-ceptibleiscreated.Eachmemberofthestimulussetispre-

    sentedtoanobservermanytimes,atrandom,andanob-servationresponseisrequestedaftereachpresentation.Thepsychometricfunctionmaybesampledbyevaluatingthepercentageofpresentationsofeachmemberofthestimulussetthatisdetected.Thefunctionisassumedtobecontinuousalongthestimulusaxis,usuallywithmonoto-nicincreasesinperformancebeingassociatedwithin-creasingstimulusstrength.

    Figure 1showsanexampleofapsychometricfunction,measuringthenumberoftimesaparticularauditorystim-uluswasheard,dependingonwhatthestrengthofthestimuluswas.Inthisexample,amethodofdeterminingtheresponsetoagivenstimuluswasusedthatiscalledayesno method:Oneachstimuluspresentation,theob-servergivesoneofthosetworesponses,indicatingwhetherthestimuluswasperceivedornot.Othermethodshavebeendevelopedtomakethisdetermination,suchasthosethatinvolvetwoormoresequentialpresentationsthatdifferalongsomecharacteristicofinterest.Theobserverisaskedtoindicatewhichofthemultiplesequentialpresentationsoneachtrialwasatargetstimulus.Becausethetargetisal-wayspresentinoneandonlyoneofthepresentationsonatrialandtheobservermustselectoneofthepresentationsasaresponse,thesearecalledforced-choicemethods.Per-hapsthemostcommonoftheforced-choicemethodsisthetwo-alternativeforcedchoice(2AFC),althoughaswillbeshownlater,forced-choiceprocedureswiththreeorfouralternativesprovidemoresatisfactorymeasurementofpsychometricperformance.

    Althoughsuchafunctionisassumedtounderlietheperceptionofsensorystimuli,oftenonlyoneortwopara-metersofthefunctionwillsufficetosummarizepercep-tion.Themostcommonlydeterminedparameterisamea-sureoflocationofthefunctionalongthestimulusaxis,typicallyspecifiedasathresholdstimulusvalue.Thethresholdisdeterminedasalevelofdetection(ordis-

    crimination)performance,andfrequentlythecriterionperformanceforthresholdisselectedtobethemidpointof

    1279

    ThisworkwassupportedbyGrantDC00626fromtheNationalIn-stitutesofHealth. Theopinionsorassertionscontainedhereinarethe

    privateviewsoftheauthorandarenottobeconstruedasofficialorasreflectingtheviewsoftheDepartmentoftheArmyortheDepartment

    ofDefense. CorrespondenceconcerningthisarticleshouldbeaddressedtoM.R.Leek,ArmyAudiologyandSpeechCenter,WalterReedArmy

    MedicalCenter,6900GeorgiaAve.,NW,Washington,DC20307-5001(e-mail:[email protected]).

    Adaptiveproceduresinpsychophysicalresearch

    MARJORIER.LEEK

    WalterReedArmyMedicalCenter,Washington,D.C.

    Asresearchonsensationandperceptionhasgrownmoresophisticatedduringthelastcentury,newadaptivemethodologieshavebeendevelopedtoincreaseefficiencyandreliabilityofmeasurement.Anexperimentalprocedureissaidtobeadaptiveifthephysicalcharacteristicsofthestimulioneachtrialaredeterminedbythestimuliandresponsesthatoccurredintheprevioustrialorsequenceoftrials.In

    thispaper,thegeneraldevelopmentofadaptiveproceduresisdescribed,andthreecommonlyusedmethodsarereviewed.Typically,athresholdvalueismeasuredusingthesemethods,and,insome

    cases,othercharacteristicsofthepsychometricfunctionunderlyingperceptualperformance,suchasslope,maybedeveloped.Resultsofsimulationsandexperimentswithhumansubjectsarereviewedtoevaluatetheutilityoftheseadaptiveproceduresandthespecialcircumstancesunderwhichonemightbesuperiortoanother.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    2/14

    1280 LEEK

    thefunctionspanningtherangefromchanceperformancetoperfectperformance.Asecondsummaryparameterusedtodescribeperformanceistheslopeofthepsychometricfunction,whichisameasureofhowrapidlyperformancechangeswithagivenchangeinstimulusvalue.Often,sen-sorycapabilitiescanbeadequatelydescribedbyathresh-

    oldmeasurealonethatis,onesinglepointonthepsy-chometricfunction.However,inthemethodofconstantstimuli,thethresholdisextractedfromafullysampledfunction,makingthemeasurementofthissinglepointonthefunctionveryexpensiveintermsofexperimenttime.Ofnecessity,manytrialsareplacedatstimuluslevelsoftheunderlyingpsychometricfunctionthatarenotinfor-mativeaboutthreshold.Itisnecessarytohaveinforma-tionaboutperformanceattheseoff-thresholdlevelsinordertofullydevelopthepsychometricfunction,butinmanycasesthatadditionalinformationisneitherneces-sarytothegoalsoftheexperimentnorworththeextraex-

    perimentaltimeandeffort.Adaptivepsychometricproce-dureshavebeendevelopedtoaddressthismajorprobleminpsychophysicalmeasurementthatis,aninefficientplacementoftrialsalongthestimulusarrayinordertoex-tractarelevantmeasure.Anexperimentalprocedureisadaptiveiftheplacementofeachtrialalongthestimulusarrayisdeterminedbytheresultsofthetrialortrialsthathavegonebefore.Itisacharacteristicofalladaptivepro-ceduresthatknowledgeabouttheoutcome(e.g.,athresh-old)increasessystematicallyastheprocedureisinprogress.Thatis,theselectionofstimuliisdeterminedduringthecourseoftheexperiment,andstimulusplacementisdri-

    venbytheadaptivealgorithmtowardthedesiredmea-surementpoint.

    Adaptiveproceduresaredesignedtorapidlyextractrel-evantmeasurementsfromapsychometricfunctionthoughttounderlieperformanceonagivensensory/perceptual

    task.Generally,twotypesofmeasuresareofinterest:lo-cationalongastimulusaxis(e.g.,threshold)andtheslopeofthefunction(howrapidlyperformancechangeswithchangesinstimulusvalues).Often,onlyalocationmea-sureisrequired,buttherearesomeinvestigationswhosegoalisamorecompletedescriptionoftheunderlying

    function,requiringbothlocationandslopemeasures.Theproceduresthemselvesinvolvetwoseparateparts:place-mentoftrialsalongastimulusaxisandanalysisofthedataobtainedtoextractcharacteristicsoftheunderlyingpsy-chometricfunctions.

    Thechallengeofadaptivepsychophysicsistomakerel-evantobservationsonthepsychometricfunctionwithmaximumefficiencywithoutsacrificingaccuracy.Adap-tivemethodsofmeasurementhavebeendevelopedwiththegoalofpreservingaccuracyandreliability,whilemaxi-mizingefficiencyandminimizingsubjectandexperimentertime.Thisarticlewilltracethedevelopmentofmodern

    adaptiveprocedures,withspecialattentiontothestrengthsandweaknessesofthreeofthemostcommonlyusedmeth-ods.Inaddition,somespecialapplicationswillbediscussed,includingtheuseoftheseprocedurestomonitorchangesinperformancethatareduetolearningorattentionallapsesbysubjects,aswellassomecharacteristicsofadaptivemeth-odsthatbecomeimportantwithapplicationtomorecom-plexapplicationsorinstudyingmultidimensionalstimuli,suchasspeech.ThereaderisdirectedtoTreutwein(1995)foranexcellent,quantitativedescriptionofpsychometricfunctionsandadaptivetechniques.

    ORIGINSOFADAPTIVEPSYCHOPHYSICALPROCEDURES

    Althoughadaptiveprocedureshavebeeninuseinsomeformformanyyears(seeLevitt,1992,forabriefdiscus-

    Figure 1.Exampleofapsychometricfunctionshowingpercentageofcorrectde-tectionsasafunctionofstimuluslevel.Theindividualdatapointsarefittedwithalo-

    gisticpsychometricfunction.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    3/14

    ADAPTIVEPROCEDURES 1281

    sion),thesystematicapplicationofadaptivealgorithmstothemeasurementofsensoryfunctionmaybetracedtoclinicalroots.ThedevelopmentofthenewdisciplineofaudiologytotesthearingduringandafterWorldWar IIin-cludedamethoddevisedbyHughsonandWestlake(1944)forsearchingquicklyforanauditorythresholdby

    startingwithaninaudiblesoundlevelandincreasingtheleveluntilapositiveresponsewasexhibitedbythepatient.HughsonandWestlakeemphasizedtheimportanceofsoundlevelsbeginningbelowalistenersthreshold,thenincreasinguntilthethresholdofsoundwasreached.Theirprocedureunderwentminormodif icationsbyCarhartandJerger(1959),whoproposedthatanauditorytestshouldbeginatarelativelyhighlevelinordertodemonstratethesoundforthelistener.Thelevelwouldthendropinfairlylargestepsuntilthesoundbecameinaudible(signaledbyanegativeresponsefromthesubject),whenthelevelwouldincreaseinasearchforthreshold.Oncethelistener

    indicatedthatthesoundwasaudible(apositiveresponse),thelevelwoulddropagain,andanotherincreasingthresh-oldsearchwouldbeinitiated.Thefinalthresholdwastakentobetheaveragelevel,duringanincreasingseries,atwhichthelistenerindicatedthatthesoundwasheard.Althoughbearingasuperficialsimilaritytooneoftheclassicalpsychometricmethodologiescalledthe methodoflimits,thisclinicalmethoddiffersinthatthethresholdsearchalwaysoccursontheascendingtrialsandthede-scendinglevelsareusedonlyassubjectpreparation andtogetintopositionfortheascendingthresholdsearch.Intheclassicalmethodoflimits,bothascendinganddescending

    trialsareusedtoidentifywherearesponsechangesfromoneofaudibilitytoinaudibilityorviceversa,andthethresholdisestimatedfromthoseboundarylevels.

    Anotherearlyadaptiveprocedurethathasenjoyedwidespreadusebothclinicallyandinresearchapplica-tionsisBekesytracking,originallydevelopedbytheau-ditoryscientistwhoseworkincochlearmodelingwonhimtheNobelprizein1962.Inthisprocedure,amechanicalarmwithapencilattachedisdrivenbyapatientlisteningtotoneschangingcontinuouslyinintensityand,some-times,infrequency.Asthepatientindicatesthatthetoneisheard,thependrawsloweronanintensity-scaledgraph,

    andthelevelofthetonedecreases;ifthepatientindicatesthatthetoneisnotheard,tonelevelincreases,accompa-niedbymarkshigheronthegraph.Thethreshold,definedasthatintensityatwhichatoneisjustbarelyheard,istakentobethemidpointoftheup-and-downtracingsofthemechanicalpen.

    Themorecommonlyusedadaptiveproceduresinre-searchtodayemergedfromtheseclinicalbeginningsandfrommethodologicalresearchinthe1940stothe1960s.DixonandMood(1948)wereamongthefirsttosystem-aticallyinvestigatethecharacteristicsandstatisticalprop-ertiesofsimpleadaptiveproceduresthatsearchforthresh-

    oldthroughacombinationofincreasinganddecreasingstimulussteps,respondingtonegativeandpositivesubjectresponses.Thistypeofprocedurehascometobecalledastaircase andformsthebasisforagreatdealofpsycho-

    metrictestingusedtoday.Staircaseproceduresdifferfromearlierclinicaltechniquesinthattheycollectanumberofthresholdestimatesfrombothascendinganddescendingseriesoftrials.Theyrefinethemethodoflimitsbynotre-quiringresponsestoacompletesetoflevelsandbyre-spondingwithchangesindirectionofthestaircaseaftera

    changeinthesubjectsresponse.

    MODERNADAPTIVEMETHODS

    Thecommoncharacteristicsofcurrentlyusedadaptivemethodsarethecollectionofsubjectresponsestoeachtrial,withasystematicmanipulationofthestimuluslevelalongtheexperimentaldimensionofinterest.Eachmethodresultsinaseriesofstimuluslevelspresentedoverthecourseoftheexperiment,alongwiththeassociatedsub-jectresponses.Experimentalvariablesth atmayimpacttheresultsofthemethodologyincludetheamountofdif-

    ferencebetweenstimulusvaluespresented(thestepsize),theinitialstartingvalueofthestimulus,theprocessthatguidesthesequenceofpresentationlevelsoneachtrial(thetrackingalgorithm),andthedecisionforendingtheprocess(thestoppingrule).Thegeneralgoalofeachprocedureistomeasurecharacteristicsofthesubjectsperformanceovertheshortestamountoftime,withoutsacrificingac-curacy.Eachmethodmaybemostappropriateinagivenexperimentalsituation,andthereisasubstantialliteraturecomparingtheabilitiesofthesemethodstoprovidebias-freeresultswithhighreliability.

    Adaptivemethodologiesthatcurrentlyenjoywide-

    spreadusemaybeplacedintothreegeneralcategories,de-finedbytheirsystemsforplacingtrialsalongastimulusarrayandbythemannerinwhicheachestimatesafinaloutcome.Thefirstcategorytobedescribediscalledpa-rameterestimationbysequentialtesting(PEST),anditischaracterizedbyanalgorithmforthresholdsearchingthatchangesbothstepsizesanddirection(i.e.,increasinganddecreasinglevel)acrossasetoftrials.Asecondtypeofadap-tiveprocedureshasbeencalledmaximum-likelihoodpro-ceduresbuttheirmoregeneralcharacteristicisthatsetsofstimulusresponsetrialsarefitwithsometypeofogivalfunctionandsubsequenttrialplacementandthresholdes-

    timationistakenfromthosefittedfunctions.Finally,acommonformofadaptiveprocedures,knowngenericallyasstaircase procedures,willbedescribed.Foreachofthesethreecategoriesofprocedures,anexampleofatrackinghistorywillbeshown,inordertounderstandthedifferencesinthemannerinwhichtheadaptiverulesleadtoanestimateofthreshold.

    ParameterEstimationbySequentialTestingThePESTprocedure,firstdescribedin1967byTaylor

    andCreelman,useschangesinstepsizetofocustheadap-tivetrackevermorefinely,stoppingthetrackwhenthees-

    timatehasbeenadequatelydefined.Thefinalestimateissimplythefinalvaluedeterminedbythetrialplacementprocedure.ThePESTalgorithmisdesignedtoplacetrialsatthemostefficientlocationsalongthestimulusaxisin

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    4/14

    1282 LEEK

    ordertoincreasemeasurementprecisionwhileminimiz-ingthenumberoftrialsrequiredtoestimateathreshold.

    Figure 2showsatypicalPESTadaptivethresholdtrack,modifiedfromHall(1981),carriedoutaccordingtothesug-gestionsofTaylorandCreelman(1967).Notethatanini-tiallevelandastepsizeareselectedtobeginatrack.After

    eachpresentationatafixedlevel,astatisticaltestisap-pliedtoindicatewhetherperformanceatthatlevelisbet-terorpoorerthanthetargetedperformancelevel(e.g.,75%correctdetections).Oncethatdeterminationismade,thelevelmaychangebythecurrentstepsize,andaseriesofpresentationsoccursatthenewlevel,againtestingaftereachpresentationwhetherthelevelshouldbechanged.InFigure 2,aninitialstepsizeof8 dBisused;afterfourpre-sentations,thelevelchanges;ninepresentationsatthenewlevelareneededtodeterminethatthecurrentlevelistoolow,andthetrackmovesbackup,butthistimewithastepsizehalfaslarge.Thenextchangeindirection(occurring

    atTrial 21)producesanotherhalvingofstepsize.Furtherchangesinstepsize(alwaysaccordingtoPESTrules)occurthroughoutthetrack,whichterminateswhenacriterionstepsizeisreached.Thelevelspecifiedbythisfinalstepsizeistakenasthefinalthresholdvalue.Theimportantcharacteristicsofthistypeofthresholdtrackarethatthestepsizeschangeaccordingtorulethroughoutthetrack,sothatthetrackexcursionstendtobecomesmallerasathresholdvalueisapproached,andthatthefinalthresholdestimateistakentobethefinalvalueinthetrack,withoutspecificallyconsideringperformanceonprevioustrials.

    TherulesforimplementationofPESTwereoriginally

    outlinedbyTaylorandCreelman(1967),butmanysubse-quentauthorshaveproposedmodifications.Theoriginalrulesincludewhentochangelevels,aprocessfordecid-ingthenextlevelinvolvingastepsizechangingthrough-outthetrackaccordingtorule,astoppingrulebasedontheapproachofthedecreasingstepsizetoacriterionvalue,andaruleforestimatingthefinalthresholdmea-sure,typicallythelastlevelindicatedbythetrackingrules.

    TaylorandCreelmanalsodescribedametric,termedthesweatfactor, thatcouldbeusedtoevaluatetheefficiencyofagivenpsychometricprocedure.Likeningthesweatfactortoameasureoftheamountofworknecessarytoreachacertainlevelofprecisioninthemeasuringalgo-rithm,theydefinedthesweatfactorastheproductofthe

    numberof trials andthe variance of the measures.Throughsimulations,theydeterminedthevarianceofthePESTthresholdsandthemeannumberoftrialsnecessarytoachievethatlevelofvariability.Comparingresultsofthoserealisticsimulationswithanidealsweatfactor(gen-eratedfromasimulatedthresholddevicewithcompletestatisticalknowledgeoftheprobabilitiesassociatedwitheachstimuluslevel)producedameasureofefficiencyofthePESTprocedure,whichTaylorandCreelmancalcu-latedtobeabout40%50%.

    PESTwasdesignedtobeasefficientaspossibleintheplacementoftrialsalonganarrayofstimuluslevelsand

    toforceconvergenceofanadaptivetrackonagivenper-formancelevel(i.e.,threshold)asrapidlyasisconsistentwithaccuracyandreliabilityofmeasurement.Theorigi-nalPESTprocedurecalledfortrialplacementthroughoutthetrackbasedonastatisticaldeterminationofperformanceatthecurrentlevel,incomparisonwithexpectedperfor-manceatatargetedlevel,andathresholdestimatethatwassimplythefinalvalueofthetrack.AlthoughsomemodificationstotheoriginalPESTtrackingrulesweresuggestedbyFindlay(1978)inordertoencouragemorerapidconvergenceofthetrack,laterdevelopmentsemerg-ingfromtheuseofPESThavechangeditstwobasicchar-

    acteristicsoftrialplacementandthresholdestimation.Hall(1981)proposedahybridprocedurethatfollowed

    PESTrulesfortrialplacementalongthestimulusaxis,butinsteadoftakingthefinalvalueofthetrackasthreshold,attheendofthetrack,heusedperformanceonalltrialstoconstructapsychometricfunction,fromwhichathresholdvaluewasextracted.Thevalueofthishybridprocedurewasthatefficienttrialplacementcouldproceedinaprescribed

    Figure 2.AdaptivetrackfollowingthePESTprocedure.Thesedecibelvaluesarerelativetoanarbitrarythresholdof0 dB,shownwiththehorizon talline.Thesedataaremod ifiedfrom Figure 1

    inHall(1981).

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    5/14

    ADAPTIVEPROCEDURES 1283

    manneraccordingtothePESTrules,but,intheend,allthedatagatheredduringtheprocedurewereusedtoconstructthefinalpsychometricfunction.Throughsimulationsandexperimentaltrialswithhumansubjects,Hall(1981)demonstratedthatthehybridprocedurecouldovercomemanyoftheshortcomingsidentifiedinprevioususeof

    PEST:Theprocedurewasrelativelytolerantofsubjectlapses,notaffectedsignif icantlybyinappropriatestartinglevelsorstepsizes,andprovidedestimatesofbothathresholdvalueandaslopeofanassumedpsychometricfunction.

    Maximum-LikelihoodAdaptiveProceduresAlthoughHalls(1981)hybridprocedurechangedthe

    PESTmethodoffinalthresholdestimate,furthermodifi-cationsofPESTchangedtherulesforstimulusplacementaswell.Thesemodificationsmaybeclassifiedintoasec-ondcategoryofadaptiveprocedures,characterizedby

    stimulusplacementoneachtrial,drivenbyconsultingthecurrentbestestimateoftheentireunderlyingpsychomet-ricfunctionaftereverystimulusresponsetrial.Astheadaptivetrackgrowsinlength,theestimatedfunctionbe-comesbetterdefinedbythecollectionofdatapointsgen-eratedfromprevioustrials.Aftereachtrial,thesetofstim-uluslevelsandtheproportionofcorrectresponsesassociatedwitheachlevelarecombinedtoformapsychometricfunc-tion,asisshownschematicallyinFigure 1.Theindividualpointsarefittedwithanogivalfunctionofsomekind(Fig-ure 1showsalogisticfunction)andacurrentestimatedthresholdlevelisextracted.Anewpsychometricfunction

    isgeneratedaftereachtrialorsetoftrials,andsubsequenttrialsareplacedatatargetedperformancelevelonthemostup-to-datefunction.Amaximum-likelihoodfittingalgorithmistypicallyusedwiththistypeofprocedure.

    ThelinkbetweentheoriginalPESTandthemaximum-likelihoodadaptiveproceduresmaybeclearlyseeninpa-persbyPentland(1980)andWatsonandPelli(1983).PentlanddevelopedwhathecalledthebestPEST,totakeadvantageofthestrengthofthemaximum-likelihoodproceduresinthecontextofaPESTadaptivetrack.As-sumingalogisticpsychometricfunctionwithagivenslope,Pentlandsprocedureseekstomaximizetheinfor-

    mationprovidedbyeachtrialbyplacinglevelsatthemostcurrentestimateofthe50%pointontheassumedpsy-chometricfunction.InsimulationcomparisonswiththeoriginalPESTandtwootheradaptiveprocedures,Pent-landsmaximum-likelihoodprocedureprovedtobethemostefficient,requiringtheleastnumberoftrialstoreachagivenlevelofprecision.InPentlandsbestPEST,levelsaresetaccordingtoafittedfunctionaftereachtrial(usingallprevioustrials),andafixednumberoftrialsispre-sented.Thethresholdestimateissimplythelastvaluees-timatedasthe50%pointoftheultimatepsychometricfunction.WatsonandPelli,intheirQUESTprocedure,ad-

    vocatedtheuseofallinformationavailablefromprevioustrialsinthetrack,supplementedbypriorknowledge(fromtheliterature,previousexperiments,etc.)tosetthenexttestlevel.However,adistinctionwasmadebetweenthe

    useofpriorinformationtodrivethetrackandthefinales-timateofathreshold,whichusedonlythedatawithinthetrack.Foranadaptivetrackconsistingof128trials,Wat-sonandPellireportanefficiencyof84%fortheirQUESTprocedure,ascomparedwith40%50%efficiencyfortheoriginalPEST.

    Maximum-likelihoodadaptiveproceduresareattractivetoinvestigatorsbecausetheymakefulluseofalltrialsinanexperimentinordertodetermineathreshold,ratherthanestimatingthresholdonlyfromthelevelsvisitedattheendofanadaptivetrack,asintheoriginalPESTpro-cedure.Inmostapplicationsoftheseprocedures,bothafunctionshape(e.g.,alogisticorWeibullfunction)andaslopevaluemustbeassumed,sothatfromtrialtotrial,thefunctionmovesitslocationalongastimuluslevelarrayinordertofindthefunctionleadingtoathresholdestimate.Mostofthedevelopmentoftheseprocedureshasoccurredinthecontextofvisionorauditoryresearch,butLinschoten,

    Harvey,Eller,andJafek(2001),inthisissue,havedemon-stratedthevalueofmaximum-likelihoodadaptivemeth-odsinstudyingtasteandsmell.Theyassumedalogisticpsychometricfunctionandreportedthatthemethodsworkedwellinestimatingthresholdswithprecisionandspeed.Althoughforwell-studiedpsychophysicaltasks,informationconcerningthefunctionunderlyingperfor-mancemaybeknownfromtheliterature,additionalnon-adaptivemeasuresmightbenecessarytoestablishthefunc-tionbeforemaximum-likelihoodadaptiveproceduresmaysafelybeimplemented.ThiswastheapproachtakenbySaberiandGreen(1997)toevaluatetheuseofmaximum-

    likelihoodadaptiveproceduresinsomemeasuresofaudi-torydiscriminationoftimeandfrequency.

    AnillustrationofthistypeofadaptivemethodologymaybetakenfromGreen(1993),whodevelopedamaximum-likelihoodadaptiveprocedureinvolvingayesnopsycho-metrictaskthatpromiseshighlyefficienttrialplacementandthresholdestimation.GreensmethodissimilartotheQUESTprocedureofWatsonandPelli(1983),aswellastosomeotherimplementations,butGreensproceduredoesnotcarryasmanypriorassumptionsasdoesQUEST,andhasalesstheoreticallybasedschemefortheplace-mentoftrials.InGreens procedure,aparticularpsycho-

    metricfunctionisassumed(e.g.,alogisticfunction),andarangeofstimulusvaluesthoughttoincludethethresh-oldtobeestimatedisidentified,perhapsthroughpilotworkorconsultingtheliterature.Asetofcandidatepsy-chometricfunctionsiscomputedonthebasisoftheas-sumedformofthefunctionandthepossiblestimulusval-ues.Eachofthecandidatefunctionsisfittedtoallthedatacollectedtothatpointaftereachstimuluspresentation,andthelikelihoodassociatedwitheachfunctioniscom-puted.Themostlikelypsychometricfunctionisthenvis-itedatthetargetperformanceleveltodeterminethestim-ulusleveltobeusedonthenexttrial,followedbyanother

    updatingofthecandidatefunctionprobabilities.Thefinalestimateofthresholdisextractedfromthemostlikelypsy-chometricfunctionaftersomenumberoftrialsorinaccordwithsomestoppingrule.Figure 3showsatypicaladaptive

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    6/14

    1284 LEEK

    track,followingGreens(1993)maximum-likelihoodpro-cedure.Thetrackischaracterizedbyawideexcursionoverthefirstfewtrials,butarapidconvergencetoathresholdstimuluslevel.Greenassertedthatareliablethresholdes-timatecouldbegeneratedusingthismethodinasfewas12trials;Leek,Dubno,He,andAhlstrom(2000),describ-

    ingastoppingrulebasedonacriterionvariabilityintheadaptivetrack,reportedthat,typically,24trialswouldpro-ducehighlyreliablethresholdestimates.IntheexampledepictedinFigure 3,thethresholdlevelappearstostabi-lizeafterabout2025trials.Theprocedure,developedasayesnotask,hasbeenextendedtoaforced-choicepro-cedurebyDaiandGreen(1992)inastudyofauditoryin-tensitydiscrimination,andforfrequencyandintensitydiscriminationbyHe,Dubno,andMills(1998).Furtherassessmentoftheuseofthisprocedurefordifferentex-perimentaltasksandwithdifferenttypesofsubjectpopu-lationshasalsobeenreportedbyLeeket al.(2000).

    StaircaseProcedures

    BothPESTandthemaximum-likelihoodproceduresinvolvefairlycomplexstimulusplacementrulesand,insomecases,developmentofthresholdestimatesfromthetrackingdata.Themaximum-likelihoodproceduresalsorequiretheassumptionofaparticularformoftheunder-lyingpsychometricfunction,whichisnotwellestablishedforsomepsychometrictasks.Thesimplicityandflexibil-ityofadaptivestaircaseshaveledtotheiradoptionastheproceduresofchoiceinmanylaboratories.Thesemeth-odsgenerallyusethepreviousoneormoreresponseswithin

    anadaptivetracktoselectthenexttrialplacement,thenprovideathresholdestimateinavarietyofways,mostcommonlybyaveragingthelevelsatthedirectionrever-salsintheadaptivetrack(i.e.,theturnaroundpoints).Sim-pleupdownstaircasescallforareductioninstimuluslevelwhenthesubjectsresponseispositive(e.g.,Ihearatone)andanincreaseinstimuluslevelwhenthere-sponseisnegative.Figure 4Ashowsanexampleofasim-pleupdownadaptivetrack.Beginningatalevelabovethreshold,positiveresponsesleadtocontinueddecreasesinstimulusleveluntilanegativeresponseoccurs.This

    triggersareversalinthedirectionofthetrack,andlevelsonsubsequenttrialsincreaseuntilthenextchangeinre-sponse.Thesimpleupdownstaircaseproceduretargetsthe50%performancelevelonapsychometricfunctionthatextendsfrom0%correctperformanceatchanceto100%correctperformance.Inotherwords,thetracktargetsthe

    stimuluslevelforwhichtheprobabilityofacorrectre-sponseequalstheprobabilityofanincorrectresponseor,equivalently,thelevelatwhichthetrackwouldmoveupordownonthestimulusaxiswithequalprobability.Thevalueofthistypeofprocedureisintheveryfewassump-tionsnecessaryforitsimplementation.Incontrasttothemaximum-likelihoodmethods,noformofthepsychome-tricfunctionneedbeassumed,andthereisnoneedforcomplicatedcomputationandfittingproceduresbetweentrials.Furthermore,incontrasttoPEST,thetrialplace-ment,stepsize,andstoppingdecisionsareallrelativelysim-pleandstraightforward.Theonlynecessaryassumption

    foruseofthesemethodsisamonotonicrelationshipbe-tweenstimuluslevelsandperformancelevels.

    Levitt(1971)describedageneraltransformationpro-cedurefortargetingspecificlocationsonapsychometricfunction.Inthetransformedmethods,insteadofatracklevelchangeinresponsetoeverytrial,asmandatedforthesimpleupdownproceduretargetingthemidpointofthepsychometricfunction,sequencesofpositiveornegativeresponsesaredeterminedthatresultinanequalprobabil-ityofthetracksmovingineitherdirection.Forthesim-pleupdownprocedure,boththepositiveandthenega-tivesequencesconsistofonetrial,andthetracklevel

    movesaftereachresponse,targetingthe50%performancelevel.Totargetahigherperformancelevel,thesequenceforadownwardmovementmaybetwoormorepositiveresponses,andthesequenceforanupwardmovementmayremainatonenegativeresponse.Thisexampleistheex-tensivelyusedtwo-down,one-upprocedure,whichtargetsthe70.7%levelonthepsychometricfunction.Recallingthattheprobabilityofthedownsequencemustequaltheprobabilityofanupsequence,weseethatapositivere-sponsetotwoconsecutivetrialsmustoccurinordertomovethetrackdownward.Ifp istheprobabilityofapos-

    Figure 3.Adaptivetrackfollowingamaximum-likelihoodadaptiveprocedure,asdevel-opedbyGreen(1993 ).Decibelvaluesarerelativetothearbitrarythreshold of0 dB.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    7/14

    ADAPTIVEPROCEDURES 1285

    itiveresponseonagiventrial,thenp3p mustequal.50,andthereforethetargetprobabilityis.5 5 .707.Simi-larly,athree-down,one-uptransformationleadstoaper-formancetargetof.794(i.e.,p35 .50;the cuberootof.50is.794),asisshownintheexampletrackinFigure 4B.Asinthesimpleupdownstaircase,thethresholdsearchstartsabovethreshold,butinthiscase,adecreaseinstim-uluslevelrequiresthreesequentialpositiveresponses.A

    reversalinthetrackoccursafteronenegativeresponse,andagain,threepositiveresponsesarerequiredtobeginanotherdescendingrunoftrials.Inhis1971article,Levittoutlinedanumberofpossibletransformations,alongwiththeirtargetperformancelevels.Oneobviousimplicationis,ofcourse,thatthemorecomplicatedthesequencerule,themoretrialstypicallyrequiredinanadaptivetrackinordertoreachanestimateofthreshold.

    Althoughthetransformedupdownmethodsarewidelyused,onerestrictionthathasbeennotedisthatonlyasmallnumberoftargetlevelscanbeestimated.Kaernbach(1991)describedasimpleupdownprocedurethatcould

    beusedtoestimateperformanceatmanymoretargetlev-elsthanallowedbythetransformedmethods,byvaryingthestepsizesusedinthetwodifferentstaircasedirections.ThevalueofKaernbachsprocedure,ashedescribedit,

    wasinthesimplicityofthealgorithm,relativetothesometimesquiteelaboraterulesnecessaryforthetrans-formedprocedures,anditsabilitytotargetanyperfor-mancelevel,notjustthosethatcouldbeestimatedwithaspecif icsequenceofupanddowntrials.InKaernbachssimpleupdownweightingprocedure,aperformancelevelisanalyzedaccordingtothedesiredratioofuptodownsteps,andthestimuluslevelischangedafterevery

    trial.Kaernbachdescribedanexampleoftargeting75%correctperformancewitharatioofuptodownstepsizesof(1 p)/p or,inthiscase,.25/.75,or1/3.Inordertotar-getthatpointonthepsychometricfunction,thestimuluslevelshouldbechangedupwardafteranincorrectre-sponseanddownwardafteracorrectresponse,andthesizeoftheupwardstepshouldbethreetimesthesizeofthedownwardstep.UsingMonteCarlosimulations,Kaern-bach(1991)reportedamodestsavingsofabout10%ofexperimentertimewiththeweightedprocedureoveramoretraditionalstaircase,usingatwo-down,one-uprule(i.e.,targeting70.7%correct).Rammsayer(1992)

    evaluatedKaernbachs(1991)weightedmethod,usinghumansubjectsinanauditorytemporaldiscriminationtask.Hereportedthat,atthebeginningofadaptivetracks,theweightedupdownmethodwasmoreefficientthana

    Figure 4.Adaptivetracksfollowingastaircaseprocedure.(A)Simpleupdownstaircase;(B)transformedupdownstaircase,followingathree-down,one-upalgorithm.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    8/14

    1286 LEEK

    transformedfixed-step-sizeprocedurebutthat,fortrackslongerthanabout40trials,therewaslittledifferenceinef-ficiencybetweenthetwoprocedures.Becausetherewasaratherlargedifferenceinstepsizebetweentheupandthedownsteps,Rammsayernotedthathissubjectsre-portedanawarenessofthedirectionofthetrack,which

    mightbiastheoutcomeoftheprocedure.Rammsayersug-gestedthatthisproblemcouldbereducedbyinterleavingmorethanonethresholdtrack.However,itmightbepointedoutthatanysavingsinexperimentaltimewouldbelostifmorethanonetrackwerefoundtobenecessaryforrea-sonsotherthanefficiency.Kaernbachs(2001a)articleinthisissueusesthisweightingprocedureinanevaluationofthresholdtracking,usinganunforced-choice method.Hearguesthat,althoughthepsychometricpropertiesofincludingdontknowinthearrayofsubjectresponsesin-dicateonlyasmallimprovementovertheforced-choiceselections,subjectsaregenerallymorecomfortablenot

    havingtoindicateananswerwhentheyareveryunsure.

    ESTIMATESOFSLOPE

    FROMADAPTIVEPROCEDURES

    Adaptivemethodshavetendedtofocusonmeasuringonepointonthepsychometricfunctioninordertoesti-mateathresholdorlocationofthefunctionalongastim-ulusaxis.However,inmanycases,theslopeofthefunc-tionmaybeusefulinfullydef iningtheshapeofthefunctionor,fortheoreticalorclinicalreasons,inestablishingtherelationshipbetweenratesofchangeinperformancelevel

    andstimuluslevel.Inhisseminalpaper,Levitt(1971)dis-cussedtheoptimalplacementoflevelsalongthestimulusaxis.Ifameasureofthemeanofapsychometricfunctionisdesiredasanestimateofthreshold,itismostefficienttoplacetrialsasnearaspossibletothemidpointofthefunction.However,thestaircaseproceduresmaybeusedtoestimateaslopeofapsychometricfunctionbyplacingtrialsoneachsideofthemeanofthefunction.Therehavebeenanumberofinvestigationsofadaptiveproceduresmeanttoevaluatehowpreciselyandaccuratelytheslope,perhapsinadditiontoathreshold,canbedeterminedfromanadaptiveprocedure.

    Threeslightlydifferentprocedureshavebeenproposedrecentlythataredesignedtoefficientlyandpreciselyiden-tifytheslopeandthresholdofthepsychometricfunctionun-derlyingsubjectperformance.Theseproceduresbearsomesimilaritytotheearlierdescribedmaximum-likelihoodprocedures,inthat,startingwithasetofcandidatepsycho-metricfunctions(eitherexplicitlyorimplicitlydescribed),trialplacementoccursatthemostlikelythresholdvalue,aresponseiscollected,andthenthatresponseisincludedwithallotherresponsescollectedintheexperimentalruntogenerateanewsetofcandidatefunctions.Thenextpre-sentationlevelislocatedatthepredictedperformance

    levelmostlikely(atthatpointintheexperiment)toreflectthreshold(orsomeothertargetedparameter).Inthethreeproceduresdescribedbelow,meanttofocusonaslopees-timate,thecandidatefunctionsarenotexplicitlydefined,

    butratheratwo-dimensionalprobabilitydistributionisup-datedaftereachresponse,andthetwodimensionsrepre-sentthetwoparametersdefiningthepsychometricfunction.

    WattandAndrews(1981)describedaprocedurethatwasintendedtomaintaintheadvantagesofamethodofconstantstimuliindevelopingagoodestimationofthe

    underlyingpsychometricfunction,usingaprobitfittothedata,butemployinganadaptivechangeinstimuluspre-sentationlevelsinordertoincreaseefficiencyoverthemethodofconstantstimuli.Probit(Finney,1971)isamethodoffittingacumulativenormalfunctiontoasetofpsychometricdata,usingamaximum-likelihoodcriterion.Eachdatapointtobeincludedinthefitisweightedaccord-ingtoitsbinomialvariability,andthenumberoftrialsplacedatthatstimuluslevel.Thus,pointsonthepsycho-metricfunctionthatarecalculatedfrommanytrialsareas-sumedtobemorereliableand,therefore,aremoreimpor-tantinthefittingcalculations.InWattandAndrewss

    procedure,afewstimuluslevelsareselectedfortestingfromalargersetofpredeterminedvalues,anumberoftri-alsarepresentedusingonlythoselevels,andthenafunc-tionisfitusingtheprobitmethod.Anewsetofstimuluslevelsisselectedonthebasisoftheprobit-fittedpsycho-metricfunction,andafurtherblockoftrialsispresented.The thresholdandspread(slope)of thecumulative-normalpsychometricfunctionassumedbytheprobitfitconvergeonthevaluesthatbestreflectthesubjectsper-formanceonthesetsoftrials.WattandAndrewsadvo-catedtheuseofthisproceduretoimprovetheefficiencyofmeasuringacompletepsychometricfunction,suchas

    onecouldgeneratewiththemethodofconstantstimuli,withouttheneedtoassumeaslopevalue.

    Morerecently,King-SmithandRose(1997)reportedamethodspecificallydesignedtomeasuretheslopeofthepsychometricfunction.TheyagreewithLevitt(1971)thattheplacementoftrialsonthepsychometricfunctioncanbeselectedinordertomaximizeefficiencyofmeasure-menteitherofslopeorofthreshold,butnotofboth.Ifthresholdisthetarget,formaximumefficiency,trialsshouldbeplacednearthetargetperformancelevelonthepsychometricfunction,andtheclosertothecorrectper-formancelevel,themoreefficientwillbethemeasure-

    ment.However,tomaximizeefficiencyinmeasuringslope,pointsthatbetterdefinethespreadofthepsychometricfunctionareappropriate.Ifthefunctionitselfisasym-metricfunctionaboutitsmidpoint,twopointsequidistantfromthemidpointofthefunctionshouldbeselected.Anunbalancedfunctionwillrequireaslightlydifferentplace-mentoftrials.King-SmithandRosedevelopedanadap-tivemethodformaximumefficiencyinmeasuringeitherslopeorthreshold,makinguseofthebinomialvariabilityassociatedwitheachprobabilitylevelonanassumedpsy-chometricfunction.Stimuliforeachtrialareplacedwiththegoalofmaximizingefficiencybyminimizingthevari-

    abilityofestimatesaftereachsetoftrials.Themethodisadaptiveinthatastimuluslevelisdeterminedfromaprobabilitydensityfunctiongeneratedfromprevioustri-als.Thelevelmostlikelytocorrespondtothreshold(ora

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    9/14

    ADAPTIVEPROCEDURES 1287

    slopevalue)ispresented.Theresponseatthislevelisusedtoupdatetheprobabilityofagivenresponseasafunctionofthetruethreshold,expressedasalikelihoodfunction.TheinformationprovidedbythelikelihoodfunctioniscombinedwiththeinitialprobabilityfunctionbyBayesianmultiplication,resultinginanupdatedfunctiondescrib-

    ingtheprobabilitythateachintensityisthethresholdaftertheresponsetothatstimulus.Thisnewprobabilityfunctionisthenusedtobeginthenexttrialandtoestimateathresh-oldforoptimumplacementofthesubsequenttrial.Thisprocesscanbeappliedtosimultaneouslyconvergeonabestthresholdandbestslopebyusingtwo-dimensionalprobabilitydensityfunctionsandlikelihoodfunctionstogeneratethenextstimuluslevel.King-SmithandRosere-portedthatthesemethodsresultinrelativelyhigheffi-ciencyforthemeasurementofboththresholdandslope,whichcanbeimprovedtotheextentthatpriorknowledgeofeitheroftheseparametersmaybeincorporatedorthat

    assumptionsconcerningoneofthemmayallowexperi-mentalfocusontheother.

    King-SmithandRose(1997)notedthatitissomewhatmoredifficulttogetprecisionintheslopeparameterthaninathresholdparameter.KontsevichandTyler(1999)de-velopedanadaptiveprocesssimilartothatofKing-SmithandRose,reportingthattheirprocedurecouldproducereasonableprecisionintheestimateofthresholdinabout30trialsfora2AFCtaskbutthatabout300trialswerere-quiredinordertoestimateslopewithsimilarprecision.Aswiththeearlierstudy,thekeytoestimatingathresholdandslopeistouseeachtrialtoupdatetheposteriorprob-

    abilityofagiventwo-dimensionalprobabilitydistribu-tion.KontsevichandTyleridentifiedapotentialproblemwithselectingtheminimumvarianceofthisdistributioninthatthetwodimensionsofthresholdandslopearein-commensurateand,therefore,someweightingconventionmustbeimposed.Instead,theseauthorsproposedadif-ferentcostfactor,whichcouldbeminimizedinordertodeterminethenextstimuluslevelforpresentation.Thisminimizedfactorcorrespondstomaximuminformationgainaftereachtrial.Usingcomputersimulationsandpsy-chophysicalexperimentswithhumans,thisprocedurewasshowntoconvergetothresholdandslopevalueswithina

    relativelysmallnumberoftrials.Leek,Hanna,andMarshall(1992)investigatedtheutil-

    ityofusingperformanceonallthetrialsthatmakeupanadaptivethresholdtracktogenerateafullpsychometricfunctionfromwhichathresholdandslopevaluecouldbeextracted,reminiscentofHalls(1981)earliersuggestionofahybridPESTmaximum-likelihoodprocedure.Leeket al.(1992)usedcomputersimulationstodeterminetheprecisionofestimateofslopesthatcouldbeaccomplishedfromastaircaseproceduredesignedtotrackthresholdataparticularperformancelevel.Thesimulationsweremeanttodeterminewhetherbothslopeandthreshold

    couldbeobtainedbysimplyreconstructingapsychomet-ricfunctiononthebasisofthetrial-by-trialperformancewithinanupdowntransformedstaircasetrack.Aproce-

    durethatisoptimizedtoproducethresholdmeasurementsmightalsoprovideslopeinformation,withlittlelossinprecision.Experimentalrunsweregeneratedfollowingselectedstaircasealgorithmsbyconsultingaknownpsy-chometricfunctiononeachtrialtodeterminearesponse.Attheendofanadaptivetrack,thetrial-by-trialdatawere

    summarizedaccordingtoperformanceateachlevelvis-itedbythetrackandthenfittoapsychometricfunctionofthesameformastheoneconsultedinthesimulation.Thresholdsgeneratedbythetrackingalgorithmwerecomparedwiththresholdsextractedfromtheoriginalandreconstructedpsychometricfunctions.Slopeestimatesfromthefittedpsychometricfunctionswerecomparedwiththoseunderlyingperformanceontheadaptivetrials.Thefunctionsreconstructedfromthetrial-by-trialdatawereaccuratereflectionsoftheunderlyingfunctionsaslongasthetrackswereatleast200trialslong.Shortertracksresultedinestimatesofpsychometricfunction

    slopethatwerebiasedhigh(i.e.,slopestoosteep).Itwasnotedthatamaximum-likelihoodfitofthedatatothepsy-chometricfunctionprovidedgreaterstabilityofestimatesthandidanearliersetofanalysescarriedoutusingtheprobitfittingprocedure(Finney,1971).Theauthorscau-tionedagainstusingtheprobitprocedurewhenthepsy-chometricfunctionwastransformedsothattherangeofperformancewaslessthan0%100%.Thestatisticalpropertiesofsuchafit,describedfullybyMcKee,Klein,andTeller(1985),alterthevariabilityassociatedwitheachpointinthetransformedfunction,whichiscriticalasaweightingcomponentintheprobitprocedure.Thus,when

    thepsychometricfunctionistruncated,asinforced-choiceprocedures,McKeeet al.recommendedthattheprobitfitshouldbeavoided.TheLeeket al.(1992)re-constructionsofpsychometricfunctionsfromtrackingdataresultedinafindingregardingslopeestimatessimi-lartothatfrequentlyreported:Itispossibletoobtainac-curateandreliableslopeestimatesfromadaptiveproce-dures,butthecostismoretrialsandsubsequentlymoreexperimenttime.

    Slopesestimatedfromadaptivetrackingproceduresre-portedbyLeeket al.(1992)tendedtobetoohighunlessthetrackswerefairlylong.Thistendencyforanoveresti-

    mationofslopefromadaptivemethodshasbeenobservedbeforeandhasusuallybeenattributedtoanasymmetryinthedistributionoftrialplacementsalongtheunderlyingpsychometricfunction.Typically,moretrialsareplacedhigherthanthemidpointofthefunctionthanareplacedlower.However,Kaernbach(2001b),inthisissue,arguesthatthesourceoftheslopeoverestimationisnotthepoordistributionoftrials,but,instead,maybefoundintheadap-tivealgorithmsthemselves.Kaernbachshowsthat,fortri-alsplacedidenticallytotrialplacementwithinanadaptivetrackbutpresentedinrandomorderinsteadofaccordingtothesequenceofthealgorithm,slopesarenotoveresti-

    mated,lendingsupporttohisnotionthatthesequentialas-pectsofthetrialplacement,accordingtoadaptiverules,isthesourceoftheslopebias.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    10/14

    1288 LEEK

    Determinationofsloperequiressomeassumptionsabouttheformofthepsychometricfunction.Thereareanumberofogivalfunctionstochoosefrom,andcriteriaforanexperimenterschoicemaydependontheoreticalorcomputationalissues.Strasburger(2001a),inthisissue,pointsoutthatcomparisonsacrossstudiesandtasksare

    hamperedbythevarietyofpsychometricfunctionsthatareselectedandprovidesformulasforconvertingamonganumberofthemostcommonlyusedfunctions.Itissug-gestedinthatpaperthatforcomparisonofresultsacrossstudies,themaximumslopeofafunction,ortheslopeatthepointofinflectionoftheogivalfunction,wouldbeausefulmetric.Inasecondpaperinthisissue(Strasburger,2001b),characterrecognitionismeasuredusinga10-alternativeforced-choiceprocedurewithamaximum-likelihood/PESTprocedure,andthemaximum-slopemet-ricisusedtocompareresultsacrossstudies.

    Alsointhisissuemaybefoundthreearticlesaddress-

    ingthebestwaytosampleandfitthepsychometricfunc-tion.Althoughprobitfittingiscommonlyused,itisnotal-waysthemostappropriatechoice,becauseofchangesinbinomialvariabilitywithtruncationofthefunctionforforced-choiceproceduresandbecauseitassumesapar-ticularformofthefunctionthatis,thecumulativenor-mal.TwopapersbyWichmannandHill(2001a,2001b)takeuptheissuesofhowbesttosamplethefunctionandhowtodeterminethegoodnessoffitoftheassumedfunc-tion.MillerandUlrich(2001)describeamethodforfit-tingthepsychometricfunctionthatmakesnoassumptionabouttheunderlyingdistribution,asdoesthemorecom-

    monlyusedprobitanalysis.

    VIOLATIONOFASSUMPTIONS

    INADAPTIVEMETHODS

    Oneofthestrongestargumentsforusingadaptivepro-ceduresfortherapidandaccurateestimatesofcharacter-isticsofpsychometricperformanceisthatthereareveryfewrestrictionsthatmustbeaccommodated.Twosuchcommonlyacceptedrequirements,however,involvesta-bilityofthemeasurementovertimeandthemonotonicre-lationshipbetweenstimulusstrengthandperformance.

    Mostexperimentersacknowledgethatabsolutestabilityoftheunderlyingpsychometricfunctiongenerallyisnotarealisticassumption,sincesubjectstypicallyexperiencesomeperceptuallearningduringthecourseofanexperi-ment,reducingthetruethreshold,oroccasionallyhavelapsesinattentionthatmayservetoincreasethemeasuredthreshold.Changesinthresholdacrossthemeasurementtrackmayalsoresultinshallowercalculatedslopesoftheunderlyingfunction.Violationsofasecondassumption,monotonicityofthepsychometricfunction,mayoccurifmembersofthestimulussetunderexaminationarenotho-mogenousalongagivenstimulusdimension.Stimuliwith

    greatercomplexityanddimensionalitysuchasspeech,forexamplearelikelynothomogenousand,therefore,posespecialproblemswhentestedadaptively.Therehavebeenattemptstomonitorviolationsoftheassumptionsof

    stabilityofthefunctionsandhomogeneityofthestimuliandtoassessthecostsofsuchviolations.

    TrackingThresholdChangesWithMultipleAdaptiveTracks

    Althoughapotentialproblemintheimplementationof

    adaptiveprocedures,itisalsooneofitsbenefits,asiden-tifiedbyLevitt(1971)andearlierpapers,thatadaptivetrackingproceduresmaybeusedtofollowchangesinthepsychometricfunctionoccurringduringthecourseofanexperiment.Forexample,asperceptuallearningoccursovermanytrials,thethresholdmaybeseentogradually(orsuddenly)improve,asreflectedintheshapesofthethresholdtracks.LeekandWatson(1984)usedthismethodtotracetheimprovementsindetectionofindividualtonesembeddedwithina10-tonepattern.Teninterleavedthresh-oldtracks,onetestingeachofthe10patterncomponents,wereexaminedtodeterminehowthetonefrequencyand

    temporalplacementwithinthepatternaffectedtheim-provementsindetectionthresholds.Incontrasttoimprove-mentsinthresholdovertime,adaptivetracksmayalsosig-nalsubjectfatigueordistractionoverthecourseofanexperiment.Hall(1983)suggestedamethodofidentify-ingashiftinasubjectsthreshold,bycomparingthesub-jectsresponseoneachtrialwiththeresponsepredictedfromtheestimatedpsychometricfunctionunderlyingper-formance.Totheextentthatthedifferencebetweenobtainedandpredictedperformanceisclosetozeroforeachtrialinthetrack,theestimatedpsychometricfunctionistakentobestablethroughoutthetrack.Leek,Hanna,andMarshall

    (1991)alsoproposedamethodfordeterminingwhetherthetruethresholdofasubjectwasshiftingduringthetimeofitsmeasurement,therebyproducinganinaccuratethresholdestimate.Theirapproachwastoinvestigatethepsychometricpropertiesofanunstableunderlyingpsy-chometricfunctionbysimulatingsystematicchangesinfunctionlocationandcomparingthevariabilitywithinandacrosstwointerleavedadaptivetracks.Thelogicwasthatifapsychometricfunctionwerechangingovertime,thevariabilityinlevelsvisitedwithinasingletrackwouldex-ceedthevariabilityobservedbetweentracksontrialsoc-curringclosetogetherintime.Inadditiontousingthese

    twosourcesofvariabilitytomonitorwhetherthefunctionisshiftingintime,Leeket al.(1991)showedhowtheacross-trackvariabilitymaybeusedtogenerateanestimateoftheslopeoftheunderlyingfunction.Incomputersimula-tionsandhumansubjectdata,boththeslopeestimatesandthestability-monitoringprocedurewereshowntoworkwellaslongastheshiftinthresholdsdidnotoccursorapidlythatthetrackscouldnotfollowthechanges.

    NonmonotonicPsychometricFunctions

    andHeterogenousStimuliEarlyinthecourseofdevelopmentofmodernadaptive

    methods,LevittandRabiner(1967)attemptedtoapplyanadaptiveproceduretothemeasurementoflevelsofspeechnecessaryforagivenlevelofperformance.Typically,speechtestingisaccomplishedbypresentingalistof

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    11/14

    ADAPTIVEPROCEDURES 1289

    speechstimuliatagivenlevelandaskinglistenerstore-peatthestimulitheyheard.Thespeechcommonlycon-sistsofstandardizedlistsofmonosyllables,consonantvowelnonsensesyllables,sentences,orrunningdiscourse.Often,itisdesiredtomeasureperformanceatanumberofdifferentpresentationintensitiesorindifferenttypesand

    levelsofcompetingnoise.Speechrecognitionasafunc-tionofincreasinglevelorincreasingsignal-to-noiseratioisanogivalpsychometricfunction.Speechtestingtypi-callyfocusesonsomeportionoftherisingpartofthefunction.LevittandRabinerappliedanupdownstaircaseandblocksofstimuliateachtestedleveltodetermine50%correctidentificationofthespeech.BodeandCarhart(1973)extendedthesefindingsbydevelopingwhattheycalledthedoubletprocedure,usingatransformedupdownadaptivemethod.Theyrantwosequentialtracks,targetingthe29.3%andthe70.7%correctidentificationsignal-to-noiselevel.Theaverageofthefinalthreshold

    levelsfromeachtrackconstitutedanestimateof50%cor-rectperformance.Steele,Binnie,andCooper(1978)usedthedoubletadaptiveproceduretostudytheimpactofvi-sualcues(lipreading)ontestsofspeechunderstanding,usingmonosyllabicwordsasstimuli.

    Althoughadaptiveprocedurescontinuetobewidelyusedinmeasuringspeechrecognition,thenatureofspeechstimulicreatesgreatervariabilityinmeasurementthanisobservedintestingmorehomogenoussetsofstim-uli,suchastonedetectionordiscrimination.Whenstim-uliarehomogenousinallcharacteristicsexcepttheoneundertest(e.g.,thelevelofapuretoneinafixednoise),

    theprocedurecanworkwell.Similarly,ifasetofspeechstimuliishomogenousinallfactorsaffectingintelligibil-ity,thelevelofthestimulieitherinquietorinnoisemaybeusedinanadaptiveprocedure.However,ifthestimuliarenothomogenous,thetrackingproceduremaybecom-promised.Forexample,aseriesofeasilyheardmonosyl-lablesmaydrivethelevelofthetracklow,butasubse-quentpresentationofaninherentlymoredifficultwordwilloccuratalevelthatistoolow,andthetrackmaynottrulyreflectoverallperformance.Heterogeneitywithinstimulussetsthusleadstoinappropriateplacementoftriallevels,greatervariabilityinthetrack,andpossibleconfu-

    siontothesubject.Dirks,Morgan,andDubno(1982)notedthisdifficultywhentestingidentificationofmono-syllablesandofspondeewords(i.e.,two-syllablewordswithequalstressoneachsyllable)inagroupofnormalhearingandagroupofhearing-impairedsubjects.Thespeechlevelwasheldconstantthroughoutanadaptivetrack,whereasthelevelofmultitalkerbabblewasvariedaccordingtoanadaptivealgorithmdesignedtotarget29.3%,50%,or70.7%correctwordidentification.Intheseprocedures,asimpleupdownprocedurewasfirstiniti-atedtofindthecorrectrangeofperformancelevelsforagivenindividual,afterwhicheitherthesimpleupdown

    algorithmwascontinued(50%correct)oroneofthetrans-formedupdownalgorithmswasimplementedtotargettheothertwoperformancelevels.Dirkset al.establishedthresholdsforeachtargetandstimulussetbutreported

    thatperformanceofthetrackswasindeedmorevariablewhenthestimuliweremonosyllables,ratherthanthemorehomogenoussetofspondeewords.Thisincreasedvari-abilitywasespeciallystrikingforthegroupofhearing-impairedlisteners,wholikelyexperiencedothersourcesofvariabilityassociatedwiththeirhearinglossaswell.It

    isimportant,therefore,inusingadaptivemethodswithsetsofspeechstimuli,tocontroltheheterogeneityofthestimulitotheextentpossibleandtobealerttoviolationofthemonotonicityassumptionacrosstrialsofanadaptivetrack.

    COMPARISONSOFADAPTIVEPROCEDURES

    Areanyoftheseadaptiveproceduresbetterthanothers,andhowdotheproceduresinteractwithotherpsychome-tricexperimentalchoices?Therehavebeenanumberof

    paperscomparingtheaccuracyandefficiencyofthepro-ceduresdescribedaboveinmeasuringthresholdsandslopesofpsychometricfunctions.Thesecomparisonshavebeenmadebyusingcomputersimulationsofexper-imentaltestsand,insomecases,evaluatingtheperfor-manceofhumanlisteners.

    Shelton,Picardi,andGreen(1982)evaluatedanadap-tivestaircase,amaximum-likelihoodprocedure,andPESTincollectingdatafromhumansubjects.Ineachcase,theychoseparametersfortheadaptiveproceduresthatwerecommonlyusedinpractice,measuringhumanperfor-manceonasimultaneous-andaforward-maskingaudi-

    torytask.Althoughtheproceduresdidnotproducelargedifferences,thereweresomecharacteristicsofeachthatmightsuggestonechoiceoranotherundercertaincircum-stances.Forshortadaptiveruns(i.e.,lessthan30trials),boththestaircaseandthemaximum-likelihoodproce-duresresultedinslightlybiasedthresholdestimates,al-thoughthebiascouldbemostlyovercomebyrandomiz-ingstartinglevelsforeachadaptiverun.Themaximum-likelihoodprocedure,however,wastheonlyoneofthethreemethodsthatcouldconvergewithinabout10trials,evenwithaslightlybiasedthreshold.Sheltonet al.sug-gestedthatthisproceduremightbemostusefulintesting

    infantsandanimals,wherethethresholdsmustbegath-eredrapidly.However,theypointedoutthatthemaximum-likelihoodmethodmaybeparticularlydifficultforinexpe-riencedlisteners,becausethereareveryfewsuprathresholdtrialsaffordedtothesubjects.

    Kollmeier,Gilkey,andSieben(1988)usedamathe-maticalmodel,aswellashumandata,tocomparetwostaircaseproceduresandthePESTprocedure,withbotha2AFCandathree-alternativeforced-choice(3AFC)exper-imentaltask.Theyalsoevaluatedbothmodelandhumanlistenersonasetoffixedlevel(nonadaptive)procedures.Theirhumanlistenerswereallexperiencedinthetask,

    whichwasdetectionofasignalembeddedinnoisethatis,simultaneousmasking.Themodelpredictedsimilarthresholdsfromadaptiveandnonadaptiveprocedures,butinpractice,humanlistenersactuallyproducedslightlybet-

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    12/14

    1290 LEEK

    terthresholdsinadaptiveprocedures.Themodelpre-dictedthata2AFCprocedureusedinastaircasetargeting71%correctwouldbetheleastefficientprocedure,whereasthe79%3AFCstaircasewouldbethemostefficient.Thehumandataalsosupportedthelatterasthemostefficientprocedure,butresultsweresomewhatvariable.Oneofthe

    problemsidentifiedwithhumandata,incontrasttothemodeleddata,wasthatthemodelunderestimatedthevari-abilityproducedbyhumanlisteners.Theauthorssug-gestedthatthereasonforthismightbeanunderlyingpsy-chometricfunctionthatisslowlyvaryingandsuggestedthatvariabilityinthetracksmaybepartitionedintoarapidtrial-to-trialvariability,combinedwithvariabilitycon-tributedfromaslowlyvaryingfunction.Hall(1983)hadearliermadesuchasuggestion,anditlaterwastakenupinsimulationsofunstablepsychometricfunctionsbyLeeket al.(1991).Kollmeieret al.suggestedthatthevariabil-itymightbecontrolledbycombiningthresholdsfromshort

    tracks,ratherthanusinglong,perhapsvarying,trackstoestimatethresholds.HicksandBuus(2000)agreedwiththisnotion,findingmoreconsistentthresholdsfromin-terleavingseveralshorttracksthanfromfollowingonelongtrackinvolvingthesametotalnumberoftrials.Insummary,Kollmeieret al.reportedthat,ifnootherex-perimentalconsiderationsdictateotherwise,themostef-ficientcombinationofmethodsisthe79%3AFC.Thisconclusionhasbeenreachedbyanumberofotherauthors.

    Kollmeieret al.(1988)alsofoundthatthresholdsfromadaptivetracksinhumanperformancetendedtobebiasedlow(i.e.,betterthresholds)thanwouldbeexpectedby

    fixedtrials.Eventhoughsomesimulationstudiessuggestthatthresholdsshouldbesimilarfromfixed-levelandadaptiveprocedures,thereareconsistentreportsofthresh-oldsofhumansubjectsbeingbetterinadaptiveproceduresmeanttotargetthesamelevelasfixed-levelmethods.Tay-lor,Forbes,andCreelman(1983)reportedthisindescrib-ingtheircomparisonsofPESTprocedureswithfixed-levelprocedures.Stillman(1989)comparedthresholdsmeasuredwiththeadaptiveproceduresandthresholdsfromfixed-leveltests,findingthattheadaptivemethodsalwaysproducedthresholdsthatwerelower,justaswasreportedbyKollmeieret al.(1988)andSheltonet al.(1982).

    Inotherwords,theadaptiveproceduresalwaysoveresti-matedsubjectsperformance,producingbetter(lower)thresholdsthanwereevidencedinanonadaptivetask.

    InStillmans(1989)study,bothinexperiencedandex-periencedsubjectswereused,andcomparisonsweremadeforresultsfromadaptiveandnonadaptiveproceduresandfortwostaircaseproceduresandaPESTadaptiveproce-dure.Thetaskwasa2AFCdetectionofa1-kHztoneinabandpassnoisecenteredat1kHz.Resultsindicatedsim-ilarthresholdsforthe79%staircaseandthePESTproce-dures(targeting80%correct)andsimilarvariabilitywithinthetwostaircaseprocedures.SheltonandScarrow(1984)

    alsomeasuredperformanceofhumanlistenerstodeter-minewhethersomeexperimentalchoiceswerebetterthanothers.Alltheirlistenerswereinexperienced,andtheyusedseparategroupsof10listenerseachforeachoffour

    conditions,staircaseandmaximumlikelihood,usingbotha2AFCanda3AFCprocedureineach.Thetaskwasde-tectionofatoneinnoise.TogetherwithSheltonet al.(1982),theseauthorsreportedthatthresholdswereessen-tiallyequivalentforallthe2AFCprocedurestested(stair-case,PEST,andmaximum-likelihood)andforthe3AFC

    staircaseandmaximum-likelihoodprocedures.Theydidobservesomedifferencesinvariabilityandefficiencyacrosstheprocedures,notingthatthe3AFCstaircasepro-videdthemoststablethresholdsacrossadaptiverunsbutthatthemaximum-likelihoodmethodproducedstabilityearlyinarun.Therefore,accordingtotheseauthors,ifprac-ticetrialsarenotpossibleorthenumberoftrialsislimited,themaximum-likelihoodprocedureshouldbefavored.

    SchlauchandRose(1990)primarilyusedsimulationswithasmallsetofhumandatatoinvestigatetheuseofstaircaseprocedureswith2-,3-,and4AFCtasks.Theymea-suredbothefficiency(variability)andthresholdbiasasa

    functionofnumberofintervals,stepsize,andthetargetoftheadaptivetrack(equivalently,thedecisionruleforchang-ingstimuluslevels).Theyidentifiedlessvariabilityinthresh-oldmeasurementsasthenumberofintervalsincreasedespecially,from2to3intervals,lesssobetween3and4intervalsandforthehigherperformancetarget(79%vs.71%).Eventakingintoaccountthegreaterexperimentaltimenecessarytopresentthelargernumberofintervals,the3AFCand4AFCprocedureswerestillmoreefficientthanthe2AFC.Theyalsoreportedgreatervariabilityinthresholdestimatesforlargerstepsizes.The2AFC71%targetwasmorebiased(i.e.,identifyingbetterperfor-

    mance)thanthe4AFC79%,andtherewasmorebiasforlargerstepsizes,especiallyforthe71%2AFCprocedure.SchlauchandRosesuggestedthatthisbiaswasaresultofbehaviornearchanceperformanceandtheeffectsofguessing.Theyalsofoundnoimprovementsinperformanceofthemethodsforadaptivetrackslongerthan100trials.Byfittingthetrial-by-trialdata,usingaprobitmethod(Finney,1971),thethresholdsrecoveredsomeofthebiasthatwasassociatedwithalltheadaptiveprocedures.Inordertoimproveefficiencyandreducebias,theseauthorsrecommendedfitt ingthetrial-by-trialdataintheadaptivetracktoestimateathresholdandtheuseofsmallstepsizes

    inthetrackingprocedure.Althoughthe4AFCproceduregavethebestefficiencyandtheleastbias,thetimetakentopresentfourintervalsoneachtrialmaystrainthemem-oryofsubjectsandmay,intheend,increaseexperimentaltimeeventhoughfewertrialsmightbenecessary.

    Insummary,thereislittletorecommendanyofthethreereviewedpsychometricproceduresfromthestand-pointoftheperformanceofthemethodsthemselves.Itseemsclearthatsomeexperimenterselectionsofvariousimplementationsofthemethodsmayincreaseordecreasethebiasandreliabilityoftheprocedures.Inparticular,the2AFCtaskisgenerallyapoorchoice,particularlywhen

    pairedwithastaircasetargetof71%.McKeeet al.(1985)providedacleardescriptionoftheimpactoftruncatingthe0%100%psychometricfunctionwhenusingaforced-choiceproceduresuchasthe2AFC.Insteadofthefunc-

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    13/14

    ADAPTIVEPROCEDURES 1291

    tionsspanningalargerangefromchanceperformanceat0%correcttoperfectperformanceat100%,thesetrun-catedfunctionsresultfromincreasedchancelevels(e.g.,50%for2AFC,33.3%for3AFC),andtherefore,therangeofthepsychometricfunctionisdecreased.Thevariabilityassociatedwitheachpointofthefunction,however,con-

    tributestothebiasandvariabilityofmeasurementac-cordingtothebinomialdistribution.Therefore,ingeneral,pointsfallinglowerthanthemidpointofthetruncatedfunctionsgenerallyhavegreatervariability.McKeeet al.suggestedthatmeasurementsarelikelytobemorereli-ableiftheyareontheuppersideofthemidpointofthefunction.Thebinomialvariabilityofthetruncatedpsy-chometricfunctionsmayaccountfortherelativelypoorerpsychometricperformanceofthe2AFC71%combination(targetlowerthanthe75%midpointofthe2AFCfunc-tion),withbetterperformancewhenthecombinationofforced-choicetaskandtargetperformancelevelleadtotri-

    alsplacedhigheronthefunction.Green(1990)addressedasimilarpoint,arguingthatthebestplacementofstimulustrialswasnearthetopendofthepsychometricfunction.

    SUMMARYANDCONCLUSIONS

    Threecategoriesofadaptiveprocedureswerereviewed.PESTproceduresdonotrequireassumptionsabouttheshapeoftheunderlyingpsychometricfunctionandpro-videarapidandsystematicconvergenceonathreshold.Maximum-likelihoodproceduresforplacingtrialsatop-timalstimuluslevelsandforprovidingthresholdandslope

    estimatesarecomputationallyintensiveandrequireas-sumptionsregardingtheshapeoftheunderlyingfunction.However,theyconvergeontargetedvaluesveryquicklyandmakegooduseofallthedatacollectedinatrack.Staircasemethodsrequireveryfewassumptionsandhavefairlysimplealgorithmsforplacementofstimuliandestimationofthresholdvalues.Theymaysupportanestimateofslope,solongassufficienttrialsarepresented.

    Someofthesemethodshaveslightadvantagesoveroth-ers,givenparticularexperimentalcircumstances.Forex-ample,whentestingmustbeaccomplishedveryquickly,asintestinganimalsorinfants,thefasterconverging

    maximum-likelihoodproceduresmightoffersomebene-fitoverlongerstaircaseprocedures.Thereisstrongcon-sensus,however,thatthepopular2AFCproceduresdonothavedesirablestatisticalproperties,particularlywhenpairedwithadaptiveproceduresthattargetrelativelylowperformancelevels(i.e.,belowthemidpointofthepsy-chometricfunction)andshouldbeavoided.Finally,stim-ulitestedinadaptiveproceduresshouldhavethecharac-teristicofhomogeneityandamonotonicrelationshipbetweenstimuluslevelandperformancelevel.Thishasbeenshowntobeproblematic(althoughnotfatal)intest-ingsomekindsofspeechrecognitionadaptively.

    Adaptivemethodsofferhighprecisionandreliabilityinpsychometrictesting,atasignificantsavingsintimeovernonadaptivetesting.Overthelast50years,refinementsandevaluationsoftheseprocedureshaveshowntheway

    toaselectionofexperimentalvariablesandparametersthatresultinlittlecostforthesavingsintime.Althoughthereareinherentbiasesinsomeofthemethods,thesecanbemostlycompensatedbyathoughtfulconsiderationofexperimentaltechniquesandparameters.

    REFERENCES

    Bod e,D.L.,&Ca r h a r t ,R. (1973).Measurementofarticulationfunc-tionsusingadaptivetestprocedures.IEEETransactionsonAudioand

    Electroacoustics,AU-21,196-201.Car h a r t ,R.,&Jer ger ,J.F. (1959).Preferredmethodforclinicalde-

    terminationofpure-tonethresholds.JournalofSpeech&HearingDisorders,24,330-345.

    Dai,H.,&Gr een ,D.M. (1992).Auditoryintensityperception:Suc-cessiveversussimultaneousacross-channeldiscriminations.Journal

    oftheAcousticalSocietyofAmerica,91,2845-2854.Dir ks,D.D.,Mor ga n,D.E.,&Dubn o,J.R.(1982).Aprocedurefor

    quantifyingtheeffectsofnoiseonspeechrecognition. JournalofSpeech&HearingDisorders,47,114-123.

    Dixo n,W.J.,&Moo d,A.M. (1948).Amethodforobtainingandana-

    lyzingsensitivitydata.JournaloftheAmericanStatisticsAssociation,43,109-126.

    Fech n er ,G.T.(1860).ElementederPsychophysik(Vol. 1).Leipzig:

    Breitkopf& Hrt el.[Alsoavailableas ElementsofPsychophysics.NewYork: Holt,Reinhart& Winston ,1966.]

    Findl ay,J.M.(1978).Estimatesonprobabilityfunctions:Amorevir-ulentPEST.Perception&Psychophysics,23,181-185.

    Finn ey,D.J.(1971).Probitanalysis(3rded.).Cambridge:CambridgeUniversityPress.

    Gr ee n ,D.M.(1990 ).Stimulusselectioninadaptivepsychophysicalprocedures.JournaloftheAcousticalSocietyofAmerica,87,2662-

    2674.Gr ee n, D.M.(1993).Amaximum-likelihoodmethodforestimating

    thresholdsinayesnotask.JournaloftheAcousticalSocietyofAmer-

    ica,93,2096-2105.Hal l ,J.L. (1981).Hybridadaptiveprocedureforestimationofpsy-chometricfunctions.JournaloftheAcousticalSocietyofAmerica,69,

    1763-1769.Ha l l ,J.L. (1983).Aprocedurefordetectingvariabilityofpsy-

    chophysicalthresholds.JournaloftheAcousticalSocietyofAmerica,73,663-667.

    He,N.-J.,Dubn o,J.R.,&Mil l s,J.H. (199 8).Frequencyandintensitydiscriminationmeasuredinamaximum-likelihoodprocedurefrom

    youngandagednormal-hearingsubjects.JournaloftheAcousticalSocietyofAmerica,103,553-565.

    Hicks,M.L.,&Buus,S.(2000).Efficientacross-frequencyintegra-tion:Evidencefrompsychometricfunctions.JournaloftheAcousti-

    calSocietyofAmerica,107,3333-3342.Hu gh son ,W.,&West l a ke ,H. (1944).Manualforprogramoutlinefor

    rehabilitationofauralcasualtiesbothmilitaryandcivilian.Transac-tionsoftheAmericanAcademyofOphthalmology&Otolaryngology,

    48 (Suppl.),1-15.Ka er n ba ch ,C. (1991).Simpleadaptivetestingwiththeweightedup

    downmethod. Perception&Psychophysics,49,227-229.Ka er n ba ch ,C.(2001a).Adaptivethresholdestimationwithunforced-

    choicetasks.Perception&Psychophysics,63,1377-1388.Ka er n ba ch ,C.(2001b).Slopebiasofpsychometricfunctionsderived

    fromadaptivedata.Perception&Psychophysics,63,1389-1398.Kin g-Smit h ,P.E.,&Rose,D.(1997).Principlesofanadaptivemethod

    formeasuringtheslopeofthepsychometricfunction.VisionRe-search,37,1595-1604.

    Kol l meie r ,B.,Gil k ey,R.H.,&Sieben ,U.K.(1988).Adaptivestair-casetechniquesinpsychoacoustics:Acomparisonofhumandataand

    amathematicalmodel. JournaloftheAcousticalSocietyofAmerica,83,1852-1861.

    Kon t sev ich ,L.L.,&Tyl er ,C.W.(1999).Bayesianadaptiveestima-tionofpsychometricslopeandthreshold.VisionResearch ,39,2729-

    2737.

  • 7/28/2019 Leek - Adaptive Procedures in Psychophysical Research

    14/14

    1292 LEEK

    Leek ,M.R.,Dubn o,J.R.,He,N.-J.,&Ahl st r om,J.B.(2000).Ex-periencewithayesnosingle-intervalmaximum-likelihoodproce-

    dure.JournaloftheAcousticalSocietyofAmerica,107,2674-2684.Leek ,M.R.,Han n a,T.E.,&Mar sha l l ,L.(1991).Aninterleaved

    trackingprocedur etomonitorunstablepsychometricfunctions.Jour-naloftheAcousticalSocietyofAmerica,90,1385-1397.

    Lee k,M.R.,Han n a ,T.E.,&Mar sh a l l ,L.(1992).Estimationofpsy-chometricfunctionsfromadaptivetrackingprocedures.Perception&

    Psychophysics,51,247-256.Leek,M.R.,&Wat son,C.S.(1984).Learningtodetectauditorypat-

    terncomponents.JournaloftheAcousticalSocietyofAmerica,76,1037-1044.

    Lev it t ,H. (1971).Transformedupdownmethodsinpsychoacoustics.JournaloftheAcousticalSocietyofAmerica,49,467-477.

    Lev it t ,H. (1992).Adaptiveproceduresforhearingaidprescription andotheraudiologicapplications.JournaloftheAmericanAcademyof

    Audiology,3,119-131.Lev it t ,H.,&Ra bin er ,R.L.(1967).Useofasequentialstrategyinintel-

    ligibilitytesting.JournaloftheAcousticalSocietyofAmerica,42,609-612.

    Linsc h ot en ,M.R.,Ha r ve y,L.O.,Jr .,El l er ,P.M.,&Ja f ek ,B.W.(2001).Fastandaccuratemeasurementoftasteandsmellthresholds

    usingamaximum-likeliho odadaptivestaircaseprocedure.Perception&Psychophysics,63,1330-1347.

    McKe e,S.P.,Kl ein ,S.A.,&Tel l er ,D.Y.(1985).Statisticalproper-tiesofforced-choicepsychometricfunctions:Implicationsofprobit

    analysis.Perception&Psychophysics,37,286-298.Mil l er ,J.,&Ul r ich ,R.(2001).Ontheanalysisofpsychometricfunc-

    tions:TheSpearmanKrbermethod.Perception&Psychophysics,63,1399-1420.

    Pen t l a n d,A. (1980).Maximum-likelihoo destimation:ThebestPEST.Perception&Psychophysics,28,377-379.

    Ra mmsay er ,T.H.(199 2).Anexperi mentalcomparisonoftheweightedupdownmethodandthetransformedupdownmethod.Bulletinof

    thePsychonomicSociety,30,425-427.Sa ber i,K.,&Gr ee n ,D.M.(1997).Evaluationofmaximum-likelihood

    estimatorsinnonintensiveauditor ypsychophysics.Perception&Psy-

    chophysics,59,867-876.Sch l a u ch ,R.S.,&Rose ,R.M.(1990).Two-,three-,andfour-interval

    forced-choicestaircaseprocedures:Estimatorbiasandefficiency.

    JournaloftheAcousticalSocietyofAmerica,88,732-740.

    Sh el t on ,B.R.,Pica r di,M.C.,&Gr een ,D.M.(1982).Comparisonofthreeadaptivepsychophysicalprocedures.JournaloftheAcousti-

    calSocietyofAmerica,71,1527-1533.Sh el t on ,B.R.,&Sca r r ow,I.(1984).Two-alternativeversusthree-

    alternativeproceduresforthresholdestimation.Perception&Psy-chophysics,35,385-392.

    St ee l e,J.A.,Binn ie,C.A.,&Cooper ,W.A.(1978).Combiningau-ditory andvisualstimuliintheadaptivetestingofspeechdiscrimina-

    tion.JournalofSpeech&HearingDisorders,43,115-122.St il l ma n,J.A.(1989).Acomparisonofthreeadaptivepsychophysical

    proceduresusinginexperiencedlisteners.Perception&Psychophysics,46,345-350.

    St r a sbu r ge r ,H.(2001a).Convertingbetweenmeasuresofslopeofthepsychometricfunction. Perception&Psychophysics,63,1348-1355.

    St r a sbur ge r ,H.(2001b).Invarianceofthepsychometricfunctionforcharacterrecognitionacrossthevisualfield. Perception&Psy-

    chophysics,63,1356-1376.Tayl or ,M.M.,&Cr eel ma n,C.D.(1967).PEST:Efficientestimates

    onprobabilityfunctions.JournaloftheAcousticalSocietyofAmer-ica,41,782-787.

    Tay l or ,M.M.,For bes,S.M.,&Cr eel man ,C.D.(1983).PESTre-ducesbiasinforcedchoicepsychophysics.JournaloftheAcoustical

    SocietyofAmerica,74,1367-1374.Tr eu t wein ,B.(1995).Adaptivepsychophysicalprocedures.VisionRe-

    search,35,2503-2522.Wat son,A.B.,&Pel l i,D.G.(1983).QUEST:ABayesianadaptive

    psychometricmethod.Perception&Psychophysics,33,113-120.Wat t ,R.J.,&An dr ew s,D.P.(1981).APE:Adaptiveprobitestimation

    ofpsychometricfunctions.CurrentPsychologicalReviews ,1,205-214.

    Wich man n,F.A.,&Hil l ,N.J.(2001a).Thepsychometricfunction:I.Fitting,sampling,andgoodnessoffit. Perception&Psychophysics,

    63,1293-1313.Wichman n,F.A.,&Hil l ,N.J.(2001b).Thepsychometricfunction:

    II.Bootstrap-basedconfidenceintervalsandsampling.Perception&Psychophysics,63,1314-1329.

    (ManuscriptreceivedJune22,2001;

    revisionacceptedforpublicationSeptember25,2001.)