White,S.,Hill,E.,Winston,J.,Frith,U.(2006).AnisletofsocialabilityinAsp - PDF document

White,S.,Hill,E.,Winston,J.,Frith,U.(2006).AnisletofsocialabilityinAspergerSyndrome:judgingsocialattributesfromfaces.Brainand(1),69Willis,J.,Todorov,A.(2006).Firstimpressions:makingupyourmindaftera100-msexposuretoaface.PsychologicalScience,592…Winston,J.,O’Doherty,J.,Dolan,R.J.(2003).Commonanddistinctneuralresponsesduringdirectandincidentalprocessingofmultiplefacial,84Winston,J.S.,O’Doherty,J.,Kilner,J.M.,Perrett,D.I.,Dolan,R.J.(2007).Brainsystemsforassessingfacialattractiveness.…Winston,J.S.,Strange,B.A.,O’Doherty,J.,Dolan,R.J.(2002).AutomaticandintentionalbrainresponsesduringevaluationoftrustworthinessofNatureNeuroscience(3),277…Yang,T.T.,Menon,V.,Eliez,S.,etal.(2002).Amygdalaractivationassociatedwithpositiveandnegativefacialexpressions.,1737EvaluatingfacetrustworthinessSCAN(2008)127 at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from judgmentsoftrustworthinesscorrelated0.75withjudgmentsofattractiveness,0.76withjudgmentsofaggressivenessand0.63withjudgmentsofintelligence.Thesehighcorrelationsmakeitdifficulttodisentanglethecontribu-tionsoffaceevaluationonspecificdimensionstoneuralresponses.Forexample,Winstonetal.(2007)recentlyfoundanon-linearamygdalaresponsetofacialattractiveness.However,giventhehighcorrelationbetweenfacetrust-worthinessandattractiveness,itispossiblethatthisresponsewasdrivenbythesharedvarianceofattractivenesswithtrustworthiness.Thestandardapproachistostatisticallycontrolforthesharedvarianceamongvariousjudgments,butthisapproachcanreducethestatisticalpowerofexperimentsand,inmanycases,itwouldbedifficulttodecideonanaprioribasiswhatjudgmentsshouldbecontrolled.Thealternativetothisapproachistoexperi-mentally,ratherthanstatistically,unconfoundcontribu-tionsofdifferentdimensionsoffaceevaluationtoneuralresponses.Thisalternativeapproachisfeasibleifthevaria-tionoffacesonthedimensionsofinterestcanbemodeled,asweshowedhere.Suchmodelscanproduceanunlimitednumberoffacesvaryingonspecificdimensionsandthefacescanbeorthogonalizedonthedimensionsofinterest.Adolphs,R.,Sears,L.,Piven,J.(2001).Abnormalprocessingofsocialinformationfromfacesinautism.JournalofCognitiveNeuroscience…Adolphs,R.,Tranel,D.,Damasio,A.R.(1998).Thehumanamygdalainsocialjudgment.,470Amodio,D.M.,Frith,C.D.(2006).Meetingofminds:themedialfrontalcortexandsocialcognition.NatureReviewNeuroscience(4),268…Ballew,C.C.,Todorov,A.(2007).Predictingpoliticalelectionsfromrapidandunreflectivefacejudgments.ProceedingsoftheNationalAcademyofSciencesoftheUSA(46),17948Bar,M.,Neta,M.,Linz,H.(2006).Veryfirstimpressions.…Blair,I.V.,Judd,C.M.,Chapleau,K.M.(2004).TheinfluenceofAfrocentricfacialfeaturesincriminalsentencing.PsychologicalScience,674…Blanz,V.,Vetter,T.(1999).Amorphablemodelforthesynthesisof3DProceedingsofthe26thannualconferenceonComputergraphicsandinteractivetechniques.LosAngeles:AddisonWesleyLongman,pp.…Breiter,H.C.,Etcoff,N.L.,Whalen,P.J.,etal.(1996).Responseandhabituationofthehumanamygdaladuringvisualprocessingoffacial(5),875…Buchel,C.,Holmes,A.P.,Rees,G.,Friston,K.J.(1998).Characterizingstimulus-responsefunctionsusingnonlinearregressorsinparametricfMRIexperiments.(2),140…Cox,R.(1996).AFNI:softwareforanalysisandvisualizationoffunctionalmagneticresonanceneuroimages.ComputersandBiomedicalResearch,…Dannlowski,U.,Ohrmann,P.,Bauer,J.,etal.(2007).Amygdalareactivitypredictsautomaticnegativeevaluationsforfacialemotions.Research:Neuroimaging,13Eberhardt,J.L.,Davies,P.G.,Purdie-Vaughns,V.J.,Johnson,S.L.(2006).Lookingdeathworthy:perceivedstereotypicalityofBlackdefendantspredictscapital-sentencingoutcomes.PsychologicalScience,,383Engell,A.D.,Haxby,J.V.,Todorov,A.(2007).Implicittrustworthinessdecisions:automaticcodingoffacepropertiesinhumanamygdala.JournalofCognitiveNeuroscience,1508Fridlund,A.J.(1994).HumanFacialExpression:AnEvolutionaryViewSanDiego,CA:AcademicPress.Gallagher,H.L.,Frith,C.D.(2003).Functionalimagingof‘theoryofmind.’TrendsinCognitiveSciences,77Gobbini,M.I.,Haxby,J.V.(2007).Neuralsystemsforrecognitionoffamiliarfaces.,32Gobbini,M.I.,Leibenluft,E.,Santiago,N.,Haxby,J.V.(2004).Socialandemotionalattachmentintheneuralrepresentationoffaces.,1628nekopp,J.(2006).Oncemore:isbeautyintheeyeofthebeholder?RelativecontributionsofprivateandsharedtastetojudgmentsoffacialJournalofExperimentalPsychology:HumanPerceptionand,199Kim,M.P.,Rosenberg,S.(1980).Comparisonoftwostructuralmodelsofimplicitpersonalitytheory.JournalofPersonalityandSocialPsychology,375…Knutson,B.(1996).Facialexpressionsofemotioninfluenceinterpersonaltraitinferences.JournalofNonverbalBehavior,165…Little,A.C.,Burriss,R.P.,Jones,B.C.,Roberts,S.C.(2007).Facialappearanceaffectsvotingdecisions.EvolutionandHumanBehavior,18Mitchell,J.P.,Cloutier,J.,Banaji,M.R.,Macrae,C.N.(2006).Medialprefrontaldissociationsduringprocessingoftraitdiagnosticandnondiagnosticpersoninformation.SocialCognitiveandAffective,49Mitchell,J.P.,Macrae,C.N.,Banaji,M.R.(2005).Formingimpressionsofpeopleversusinanimateobjects:social-cognitiveprocessinginthemedialprefrontalcortex.,251…Montepare,J.M.,Dobish,H.(2003).Thecontributionofemotionperceptionsandtheirovergeneralizationstotraitimpressions.ofNonverbalBehavior,237…Oosterhof,N.N.,Todorov,A.(underreview).ThefunctionalbasisoffacePessoa,L.,Japee,S.,Sturman,D.,Underleider,L.G.(2006).Targetvisibilityandvisualawarenessmodulateamygdalaresponsestofearfulfaces.CerebralCortex,366…Phillips,M.L.,Young,A.W.,Senior,C.,etal.(1997).Aspecificneuralsubstrateforperceivingfacialexpressionsofdisgust.,495…Rosenberg,S.,Nelson,C.,Vivekananthan,P.S.(1968).Amultidimensionalapproachtothestructureofpersonalityimpressions.JournalofPersonalityandSocialPsychology,283…Said,C.P.,Baron,S.,Todorov,A.(Inpress).Nonlinearamygdalaresponsetofacetrustworthiness:contributionsofhighandlowspatialfrequencyJournalofCognitiveNeuroscienceSambataro,F.,Dimalta,S.,DiGiorgio,A.D.,etal.(2006).Preferentialresponsesinamygdalaandinsuladuringpresentationoffacialcontemptanddisgust.EuropeanJournalofNeuroscience,2355SingularInversions(2006).FaceGen3.1FullSDKDocumentation.http://facegen.com(lastaccessed5June2007).Talairach,J.,Tournoux,P.(1988).Co-planarStereotaxicAtlasoftheHuman.NewYork:Thieme.Todorov,A.(Inpress).Evaluatingfacesontrustworthiness:anextensionofsystemsforrecognitionofemotionssignalingapproach/avoidancebehaviors.In:Miller,M.,Kingstone,M.A.,editors.TheYearinCognitiveNeuroscience,Vol.1.Todorov,A.,Duchaine,B.(Inpress).Readingtrustworthinessinfaceswithoutrecognizingfaces.CognitiveNeuropsychologyTodorov,A.,Mandisodza,A.N.,Goren,A.,Hall,C.C.(2005).Inferencesofcompetencefromfacespredictelectionoutcomes.,1623Todorov,A.,Pakrashi,M.,Loehr,V.R.,Oosterhof,N.(Underreview).Evaluatingfacesontrustworthiness:automaticassessmentoffacevalence.Uleman,J.S.,Blader,S.,Todorov,A.(2005).Implicitimpressions.In:Hassin,R.,Uleman,J.S.,Bargh,J.A.,editors.TheNewUnconsciousNewYork:OxfordUniversityPress,pp.362…126SCAN(2008)A.Todorovetal at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from Incontrasttotheobservedlinearresponseintherightamygdala,theresponseinanareaintheleftamygdalachangedasaquadraticfunctionoffacetrustworthiness.Thatis,theamygdalaresponsewasstrongesttofacesonbothextremesofthetrustworthinessdimension.ThisreplicatesthefindingsofSaidetal.(inpress).Totheextentthatjudgmentsoffacetrustworthinessreflectsimilarityoffacialfeaturestohappyandangryexpressions,theleftamygdala’ssensitivitytotheextremesofthedimensionisconsistentwithotherstudiesfindingastrongeramygdala’sresponsetoemotionallyexpressivefaces,independentofthevalenceoftheemotion,thantoemotionallyneutralfaces(Breiteretal.,1996;Yangetal.,2002;Winstonetal.,2003;Pessoaetal.,2006).Itshouldalsobenotedthatthepatternofresponseintheleftamygdalasuggeststhattheamygdala’sresponsewasmoresensitivetodifferencesatthenegativethanatthepositiveendofthetrustworthinessdimension,althoughthenegativelineartrenddidnotreachsignificance.Itisinterestingtonoteinthiscontextthattherela-tivelypoordiscriminationbetweentrustworthy-anduntrustworthy-lookingfacesofbilateralamygdaladamagepatientsisduetoabiastoperceiveuntrustworthyfacesastrustworthy(Adolphsetal.,1998).Thatis,althoughthesepatientsshowanoverallpositivitybiasinjudgingfaces,thisbiasisespeciallypronouncedforfacesatthenegativeendofthetrustworthinessdimension(seealsoTodorovandDuchaine,inpress).ThisalsoseemstobethecaseforpeoplewithAspergersyndrome(Adolphsetal.,2001;Whiteetal2006).Thefindingsfrompatientandfunctionalneuroima-gingstudiessuggestthattheamygdalaismoretunedtodetectingdifferencesinthenegativethaninthepositivevalenceoffaces.Inadditiontoleftamygdala,weobservedaquadraticresponseintheMPFCandprecuneus.However,theseregionsshowedastrongerresponsetofacesinthemiddleofthetrustworthinessdimensionthantountrustworthyandtrustworthyfaces.Theseregionsarepartofthenetworksupportingsocialcognitionprocesses(GallagherandFrith,2003;AmodioandFrith,2006;Mitchelletal.,2005;Mitchelletal.,2006)andareactivatedbythepresenceoffamiliarfaces(Gobbinietal.,2004;GobbiniandHaxby,2007).Oneinterpretationoftheirpatternofresponseisthatthefacesinthemiddlerangeoftrustworthinessarerelativelymorefamiliarthanthefacesontheextremesofthedimension.Anotherinterpretationisthatitismoredifficulttoinfertheintentionsofthesefacesthanfacesattheextremesofthedimensionand,asaresult,thesefacesengageregionssupportingtheoryofmindinferences.Theseinterpretationsremaintobetested.Weshowedthatitispossibletoconstructamodelforrepresentingfacesonaspecifictraitdimensionandtousecomputermodel-generated3Dfacestosearchfortheneuralsubstrateoffaceevaluation.Thisapproachhastwodistinctadvantages.First,incontrasttocorrelation-basedexplora-toryapproaches(Engelletal.,2007)inwhichfacesareratedonatraitdimensionandthentheneuralresponsesareregressedontheseratings,itisatheoryvalidationapproach.Second,itallowstheinvestigatortohaveprecisecontroloverthefacialstimuliandtogenerateanunlimitednumberoffacesthatvaryonaparticulardimensionofinterest.Asnotedintheintroduction,traitjudgmentsfromfacesarehighlycorrelatedwitheachother.Forexample,forthesetofstandardizedfacesusedbyEngelletal.(2007), 1012 0.150.10.0500.050.10.15210Face trustworthinessParameter estimatesA medial prefrontal cortexprecuneus Fig.3(A)Regionsinthemedialprefrontalcortexandprecuneusshowingsignificantquadraticeffectsasafunctionoffacetrustworthiness.Theresultsofaperformedonthecoefficientsofthequadratictrendregressorsontheindividualdata.(B)Parameterestimates(percentsignalchange)inthefunctionallydefinedMPFCasafunctionoffacetrustworthiness.(C)Parameterestimates(percentsignalchange)inthefunctionallydefinedprecuneusasafunctionoffacetrustworthiness.FortheplotsinpanelsBandC,thefaceswerebinnedintosixcategoriesaccordingtotheirtrustworthiness.Thelinesrepresentthebestfittingcurves. Table2BrainregionsrespondingsignificantlytofacetrustworthinessRegionsrespondinglinearlyVolume(mmxyztLeftputamen402161244.48Rightamygdala110261142.56RegionsrespondingquadraticallyPrecuneus47857394.34Medialprefrontalcortex458158194.30Leftamygdala271103.81-valueforthevoxelwithmaximumactivationintheclusterisreported.CoordinatesofthisvoxelarereportedinTalairachspace.EvaluatingfacetrustworthinessSCAN(2008)125 at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from theprecuneus(Table2andFigure3A).Bothoftheseregionsshowedanegativequadraticresponse.AsshowninFigure3BandC,theresponsewasstrongertofacesinthemiddlerangeofthetrustworthinessdimensionthantofacesattheextremesofthedimension.Evaluatingfacesontrustworthinessapproximatesthevalenceevaluationoffacesthatunderliesmultipletraitjudgments(OosterhofandTodorov,underreview).Inthisarticle,weusedamodel-basedapproachtotestfortheinvolvementoftheamygdalaintheimplicitevaluationoffacetrustworthiness.First,basedonbehavioraldata,webuiltaparsimoniousmodelforrepresentingfacetrustworthiness.Second,basedonthismodel,wegeneratednovelfaces.Third,weusedthesenovelfacesinanfMRIstudyandconfirmedtheactivationoftheamygdalaasafunctionofthetrustworthinessoffaces.Specifically,replicatingpreviousstudies(Winstonetal.,2002;Engelletal.,2007),astheuntrustworthinessoffacesincreasedsodidtheresponseinanareaintherightamygdala.Giventhatparticipantswereneverengagedinexplicitpersonevaluation,thisfindingprovidesfurthersupportforthenotionthatfacesarespon-taneouslyevaluatedontrustworthiness(Engelletal.,2007).Inadditiontorightamygdala,wealsoobservedalinearresponseforbilateralputamenandrightanteriorinsulaasafunctionoffacetrustworthiness.Astheuntrustworthinessoffacesincreased,sodidtheresponseintheseregions.etal.(2002)observedasimilarresponseintherightanteriorinsula.Theamygdala,putamenandanteriorinsulaareoftenactivatedintheprocessingoffacesexpressingnegativeemotions(Phillipsetal.,1997;Sambataroetal2006;Dannlowskietal.,2007).Thesefindingsprovideadditionalsupportforthehypothesisthatprocessingoffacetrustworthinessissubservedbythemechanismsunderlyingprocessingofemotionalexpressions(OosterhofandTodorov,underreview;Todorov,inpress). ARCB 1012 1012 Fig.2Amygdalaresponseasafunctionoffacetrustworthiness.(A)Amygdalaregionofastandardizedbrain.(B)Areaintherightamygdalashowingasignificantnegativelinearchange,thisareashowedthesamelinearresponseinEngelletal.(2007).(C)Areaintheleftamygdalashowingasignificantquadraticchange.Thestatisticalmapsshowtheresultsofa-testperformedonthecoefficientsofthelinearandquadratictrendregressorsontheindividualdata.(D)Parameterestimates(percentsignalchange)inthefunctionallydefinedrightamygdalaasafunctionoffacetrustworthiness.(E)Parameterestimates(percentsignalchange)inthefunctionallydefinedleftamygdalaasafunctionoffacetrustworthiness.FortheplotsinpanelsDandE,thefaceswerebinnedintosixcategoriesaccordingtotheirtrustworthiness.Thelinesrepresentthebestfittingcurves.124SCAN(2008)A.Todorovetal at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from andenteredintotheGeneralLinearModel(GLM).Themodelalsoincludedregressorsofnon-interest:timeseriesrepresentingsubjectheadmovement,time-dependentlinearandquadratictrendscausedbyscannerdriftandthepresentationofthe‘test’images.-testwasperformedontheparameterestimatessuppliedbytheGLMforeachsubjecttotestforthesignif-icanceoflinearandquadraticestimatesacrossallsubjects.Wegeneratedgrouplevelstatisticalparametricmapsshowingvoxelsthatvariedlinearlywithfacetrustworthinessandvoxelsthatvariedquadraticallywithfacetrustworthi-ness.Themapswerethenthresholdedatanuncorrected-levelof0.001.Tofindouttheminimumclustersizeforcorrectedsignificanceof0.05,weconductedawholebrainMonteCarlosimulationofnull-hypothesisdata.Thesesimulationsdeterminedthattheminimumclustersizewas378mmBecausewemadeaprioripredictionsabouttheamygdala,wethresholdedthestatisticalmapsineachamygdalaatanuncorrectedvoxelwise-levelof0.05andthenconductedaMonteCarlosimulationineachamygdala.Thissimula-tionindicatedthataminimumclustersizeof135mmrequiredtoachievecorrectedsignificanceof0.05.BecausethisexperimentusedthesamedesignasthefMRIexperimentconductedbyEngelletal.(2007),wealsoconductedaconjunctionanalysisofthestatisticalmapsofthetwoexperimentsafterwesubmittedtheEngelletal.datatothesameGLManalysis.Fortheconjunctionanalysis,themapswerethresholdedatan-levelof0.05and,thus,theresultingconjointprobabilitywas0.0025.AMonteCarlosimulationineachamygdaladeterminedthattheminimumclustersizewas54mmforcorrectedsignificanceof0.05.Toindependentlyvalidatetheshapeoftheneuralresponseasafunctionoffacetrustworthiness,wedefinedregionsofinterests(ROI)andthenconductedanadditionalGLManalysistoextractthesignalchangeforeachface.Fortheamygdala,theROIweredefinedbytheintersectionofthestatisticalparametricmapswithananatomicalmaskoftheamygdala.Wealsocreatedfunctionalmasksforthoseregionsoutsidetheamygdalathatmetthecriterionforcorrectedstatisticalsignificance.Intheadditionalanalysis,foreverysubject,regressorsforeachfaceimagewereconvolvedwithanidealhemodynamicresponsefunctionandenteredintotheGLM.Themodelalsoincludedregressorsofnon-interest:timeseriesrepresentingsubjectheadmovement,time-dependentlinearandquad-ratictrendscausedbyscannerdriftandthepresentationofthe‘test’images.TheseGLMsprovidedtheparameteresti-matesforeachfacepresentedtoeverysubject.ItshouldbenotedthatthisanalysisdoesnotmakeanyassumptionsabouttheshapeoftheresponseasafunctionoffaceWithineachROI,wepulledthemeanpercentsignalchangeforeachface.Becauseeachfacewaspresentedonlyonce,webinnedthefacesintosixcategoriesof11faceseach,rangingfromthe11leasttrustworthyfacestothe11mosttrustworthyfaces.ThemeansignalchangeacrosssubjectswasplottedasafunctionofthesecategoriesintheROI(Figures2D,E,3BandC).Behavioraljudgments.Thetrustworthinessjudgmentsofthefacescollectedaftertheimagingsessionagreedwiththetrustworthinesspredictedbythemodel.Thecorrelationbetweenthelatterandthemeanbehavioraljudgmentswas0.001.Correlationanalysisatthelevelofindividualsubjectsshowedthatforallsubjectsbutone,thecorrelationbetweentheirjudgmentsandthemodeltrustworthinesswaspositive.Theaveragecorrelation0.35(s.e.0.06)wassignificantlyhigherthanzero,0.001.FMRIresults.ReplicatingEngelletal.(2007)findings,aclusterofvoxelsintherightamygdalashowedasignificantnegativelineartrendasafunctionoffacetrustworthiness(Figure2B).Theamygdalaresponsetofacesincreasedastheuntrustworthinessoffacesincreased(Figure2D).Aconjunc-tionanalysiswiththestatisticalmapforthelineartrendinetal.showedthatthisclusterwaslargelyoverlappingwiththeclustershowinganegativelinearresponsetothetrustworthinessoftherealfacesusedbyEngelletal.(103outof110mm).Therewasasmallclusterofvoxelsintheleftamygdala(31mm)showingthesamenegativelineartrend,butthisclusterdidnotpassthesignificancecriterionadjustedformultiplecomparisons.Theonlyotherregionbesidestheamygdalathatshowedasignificantlinearresponsetofacetrustworthinessandpassedthestatisticalthresholdcorrectedformultiplecomparisonswastheleftputamen(Table2).Similarly,totheresponseoftherightamygdala,theputamen’sresponseincreasedasthefaceuntrustworthinessincreased.Atareducedthresholdof0.01,alargecluster(1506mm)intherightputamenthatextendedintotherightanteriorinsulaalsoshowedanegativelinearresponse(0.05correctedformultiplecomparisons).Theanalysisofthequadratictrendshowedasignificantpositiveresponseinaclusterofvoxelsintheleftamygdala(Table2andFigure2C).AsshowninFigure2E,theamygdalaresponsewasstrongesttobothuntrustworthyandtrustworthyfaces,althoughtheresponsewasmoreelevatedforuntrustworthyfaces.However,thecoefficientforthenegativelineartrendwasnotsignificant.ThisfindingreplicatesthefindingsofSaidetal.(inpress)whofoundasimilarquadraticresponsefunctiontofacetrustworthinessintheamygdala.Inthepresentstudy,thequadraticresponsewasdetectableonlyintheleftamygdala.Theclusterintherightamygdalashowingaquadraticresponsewasverysmall(24mm)anddidnotpassthestatisticalthresholdcorrectedformultiplecomparisons.Theregionsotherthantheleftamygdalathatshowedaquadraticresponseandsurvivedthecorrectionformultiplecomparisonswerethemedialprefrontalcortex(MPFC)andEvaluatingfacetrustworthinessSCAN(2008)123 at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from Fourteen(sevenfemale)subjectsdifferentfromthesubjectsinthebehavioralstudyvolunteeredforthefMRIstudyandwerepaid$30fortheirparticipation.Theywerebetweentheagesof18and27(mean22.6).Allsubjectswereright-handed,hadnormalorcorrected-to-normalvisionandreportednohistoryofneurologicalillnessesorabnormalities.Weacquiredinformedconsentforparticipa-tionapprovedbytheInstitutionalReviewBoardforHumanSubjectsatPrincetonUniversity.Allsubjectswerefullydebriefedatthecompletionoftheexperiment.Facestimuli.Ninetynewfaceswerecreatedrandomlyusingthesameprocedureastheonedescribedinthe‘Method’sectionofstudy1.Thirty-threenon-ambiguousmalefaceswereselectedbasedonthesexjudgmentsofthreeindependentraters.Toincreasethevarianceoffacetrust-worthiness,weusedthese33facesasabasisof66newfaces:33trustworthyand33untrustworthyfaces(seeFigure1forexamples).Forthetrustworthyfaces,theshapecontrolswithpositivecoefficientweights(browridgeandcheekbones)wereincreasedwithabout2s.d.andthosewithnegativecoefficientweights(chinandnosesellion)weredecreasedwithabout2s.d.(becausetheshapecontrolsarecorrelated,wewereunabletomanipulatethemwithexactly2s.d.).Fortheuntrustworthyfaces,theshapecontrolsforeachfeaturewerechangedthesamedistancebutintheoppositedirec-tion.Toobtainacontinuousmeasureoftrustworthiness,wecomputedthepredictedtrustworthinessvalueforeachofthe66facesusingtheregressionmodelobtainedinstudy1(Table1).ThesevalueswerecenteredaroundzeroandusedtocreatetheregressorsforthefMRIanalysisasexplainedAsnotedintheintroduction,inresearchconductedafterstudy1,wehaveformallymodeledatrustworthinessdimensioninthe50-dimensionalspacedefinedbythe50symmetricshapecomponentsinFacegen(OosterhofandTodorov,underreview).ThepredictedvaluesfromtheregressionmodelforthefacesusedinthefMRIstudyandthepredictedvaluesfromthecomprehensivetrustworthinessmodelwerepracticallyindistinguishable.Thecorrelationwas0.99.Thus,theparsimoniousregressionmodelprovidedarobustrepresentationofthetrustworthinessoffaces.Thishighcorrelationisduetothefactthatweexpandedtherangeoftrustworthinessbymanipulatingthefacialfeaturesimportantforjudgmentsoftrustworthinessandthatthefeaturecontrolsarecorrelated.Thatis,changesinthefourfeaturesmanipulatedinthecurrentstudyarelinkedtochangesonanumberofotherfeatures.Inotherwords,changesinthefourfeaturesarelinkedtoall50underlyingprinciple(shape)components.Subjectswereinformedthattheywereparticipatinginastudyexaminingfacememory.Theyweretoldthattheywouldseesixblocksoffaceimages.Ablockconsistedof11faceimagespresentedinrandomorder.Theacquisitionrunbeganwitha12spresentationofafixationcross.Subsequently,eachfacestimuluswaspresentedfor1sinajitteredevent-relateddesign.TheISIwaschosenrandomlyfromanexponentialdistributionwithatargetmeanISIof3.5s.TheminimumISIwas1.5s.Subjectsweretoldto‘dotheirbest’torememberthefirst11faceimagesandthatthe12thimagewouldbea‘test’image.Theywereinstructedtoindicatewhethertheyrememberedthe‘test’imagefromthepreceding11faceimagesbypressingeithera‘yes’ora‘no’button.Eachblockwasseparatedbya12-srestperiodtoallowhemodynamicactivationtoreturntobaseline.Theorderofthefaceimageswasrandomizedforeachsubject.Stimuliwereprojectedontoascreenlocatedattherearoftheboreofthemagnet.SubjectswereabletoviewthesestimuliviaanangledmirrorattachedtotheRFcoilplacedabovetheireyes.Afterthescanningsession,subjectswereledtoacomputerandaskedtojudgethe66facesusedinthefMRIsessionontrustworthiness.Theorderofthefaceswasrandomizedforeachsubject.Eachfacewaspresentedatthecenterofthescreenuntilthesubjectresponded.Theresponsescalerangedfrom1(Veryuntrustworthy)to9(Verytrustworthy).Wewereunabletoobtainjudgmentsforonesubjectbecauseheneededtoleaveimmediatelyafterthescanningsession.Imageacquisition.Bloodoxygenationlevel-dependent(BOLD)signalwasusedasameasureofneuralactivation.Echoplanarimages(EPI)wereacquiredusingaSiemens3.0TeslaAllegrahead-dedicatedscanner(Siemens,Erlangen,Germany)withastandard‘bird-cage’headcoil2000ms,TE30ms,flipangle,matrix64).Byusing33interleaved3-mmaxialsliceswewereabletoachievenearwholebraincoverage.Priortotheprimarydataacquisitionscan,ahighresolutionanat-omicalimage(T1-MPRAGE,TR2500ms,TE4.3ms,flipangle,matrixsize256)wasacquiredforuseinregisteringfunctionalactivitytothesubject’sanatomyandforspatiallynormalizingdataacrosssubjects.Imageanalysis.AllfMRIdatawereanalyzedwithAnalysisofFunctionalNeuro-imagessoftware(AFNI;Cox,1996).Subjects’motionwascorrectedusingasix-parameter3Dmotion-correctionalgorithmfollowingslicescan-timecorrection.Datawerethenlow-passedfilteredwithafrequencycut-offof0.1Hzfollowingspatialsmoothingwitha6mmfullwidthathalfminimum(FWHM)Gaussiankernel.Thesignalwasthennormalizedtopercentsignalchangefromthemean.Totestforlinearandquadraticeffectsoffacetrustworthi-nessonneuralresponses,weusedapolynomialregressionetal.,1998).Wecreatedthreetimeseriesofinterest:azero-ordertimeseriesindicatingthepresenceofaface,afirst-ordertimeseriestestingforlineareffectsoftrust-worthinessandasecond-ordertimeseriestestingforquadraticeffectsoftrustworthiness.Boththefirst-orderandsecond-ordertimeserieswerecenteredaroundzeroandorthogonalizedtoeachother.Thethreetimeserieswerethenconvolvedwithanidealhemodynamicresponsefunction122SCAN(2008)A.Todorovetal at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from andfemalehumanfacesthatwerelaser-scannedin3D.UsingPrincipalComponentAnalysis(PCA),amodelwasconstructedsothateachfacecanberepresentedbyalimitednumberofindependentcomponents.Thecomponentsdonotcorrespondtospecificfacialattributesorfeatures.However,featurecontrols,whicharealineartransformationofthecomponents,resolvethisissue.Forexample,differentcontrolsallowforchangingthenose(e.g.flat/pointed)andtheeyebrows(e.g.down/upinnerbrowridge).Incontrasttotheprincipalcomponents,whichareuncorrelated,thefeaturesarecorrelatedand,thus,changingonecontrolvaluechangesothervalues.Forthisstudy,weworkedwiththe61symmetricshape(features)controlsofFacegenthattogethergivecompletecontrolovertheunderlyingFacestimuli.First,wegenerated96CaucasianfacesusingFacegen.Thefacesweregeneratedrandomlywiththefollowingconstraints.Facegen’sracecontrolsweresetsothatallfaceswereEuropean.Thiswasdonebecauseacompletelyrandomfacecanbeofanyrace(includingAfro-AmericanandAsian)andwewantedtoavoidjudgmentsaffectedbyethnicstereotypes.Additionally,facialattractivenesswasincreasedtomakethemmoresimilartothephoto-fittedrealfacesusedinEngelletal.(2007).Also,weintroducedabiastowardsmalefaces,becausemalefaceswithouthairlookmorenaturalthanfemalefaceswithouthair.Thisbiasresultedinmostlytypicalmalefaces,withsomefeminineandsomeextremelymasculinefaces.Bydefault,therandomlygeneratedfacesareemotionallyneutral.Facegenhasseparatecontrolsforaddingthebasicemotionalexpres-sions:anger,disgust,fear,sadness,happinessandsurprise.Foralloftherandomlygeneratedfaces,theseexpressionsweresettoneutral.Nevertheless,tofurtherensurethattheexpressionswereneutral,wealsosetthemouthshapecontrol,whichmovesthecornersofthemouthupanddown,toneutral.Second,wegeneratedanothersetof96facesderivedfromthefirstsetof96facesbymanipulatingtheeyebrows(loweringorraisingtheinnerbrowridge)andthemouth(thedistancebetweenthemouthandthenose)featuresofeachface.Thiswasdonebecauseapilotstudydeterminedthatthesetwofeaturesareimportantforjudgmentsoftrustworthiness.Thus,foreachofthefirstsetof96randomlycreatedfaces,anotherfacewascreatedbyadjustingeitherthebrowridgeinnerup/downcontrol(2s.d.),orthemouthup/downcontrol(2s.d.),orboth.Participantsweretoldthatwewereinter-estedinfirstimpressionsandthatthereisnorightorwronganswer.Eachofthe192faceswaspresentedonceandtheorderoffaceswasrandomizedforeachparticipant.Eachfacewaspresentedatthecenterofthescreenfor500msandwasprecededbya1000msfixationcross.Theinter-stimulusinterval(ISI)was1000ms.Theresponsescalerangedfrom1(Veryuntrustworthy)to8(Verytrustworthy).ThemeanjudgmentsaveragedacrossparticipantswereusedtofindthefacialfeaturesmostpredictivefortrustworthinessThetrustworthinessjudgmentsweresufficientlyreliable,0.80.Atthefirststageoftheanalysis,wecomputedthecorrelationsbetweenthemeantrustworthi-nessjudgmentsandeachofthe61featureshapecontrols.Weselectedthefourfacialfeaturesthatshowedthehighestcorrelationwithtrustworthinessjudgmentsindifferentfaceregions(Table1).Faceswithhighinnereyebrows,pro-nouncedcheekbones,widechinsandshallownosesellionlookedmoretrustworthythanfaceswithlowinnereye-brows,shallowcheekbones,thinchinsanddeepnosesellion.Wealsoselectedthesefeaturesbecausetheyshowedrelativelyweakcorrelationswitheachother,max(|Atthesecondstageoftheanalysis,weregressedthemeantrustworthinessjudgmentsonthefourfacialfeatures.Thisregressionanalysiswasbasedonthemeanjudgmentsoftheunambiguouslymalefaces,asjudgedbythreeindependentraters,becauseweusedonlymalefacesinthefMRIstudy.Thefourfacialfeaturesaccountedfor29.4%ofthevarianceoftrustworthinessjudgments.Thecoefficientsoftheregres-sionmodel(Table1)wereusedtopredictthetrustworthi-nessofanewsetoffacesusedinthefMRIstudy.Itshouldbenotedthatthesepredictedtrustworthinessvalueswererobustwithrespecttowhichfaceswereusedintheregressionanalysis.Aposthoccorrelationanalysisshowedthatthepredictedtrustworthinessvalueswereverysimilarifeitherallfaces(maleandfemale)wereusedtoestimatetheregressioncoefficients(0.99),orifonlymalefacesfromtheoriginalandunmanipulatedfacesetwereused(STUDY2:NEURALRESPONSESTOFACEInthisexperiment,weusedthesameimplicittaskasinetal.(2007).Participantsostensiblyparticipatedinafacememorytask.Theywerepresentedwithblocksoffacesandaskedtoindicatewhetheratestfacewaspresentedintheblock.Thus,thetaskdidnotdemandexplicitpersonevaluation.Wetestedforbothlinearandquadraticeffectsasafunctionoffacetrustworthiness. Table1Zero-ordercorrelationsbetweenchangesinfacialfeaturesandjudgmentsoffacetrustworthiness,andregressionscoefficientsofchangesinfacialfeaturesaspredictorsoffacetrustworthinessFacialfeatureCorrelationRegressioncoefficientBrowridge(down/up)Cheekbones(shallow/pronounced)0.24Chin(wide/thin)Nosesellion(shallow/deep)0.05EvaluatingfacetrustworthinessSCAN(2008)121 at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from faceratherthanidiosyncraticperceptionsofthejudgenekopp,2006),Engelletal.arguedthattheamygdalaresponseisdrivenbystructuralpropertiesofthefacethatconveycuesforuntrustworthiness.etal.usedstatisticalprocedurestodisentanglethecontributionsofidiosyncraticperceptionsandconsensusjudgmentstotheamygdala’sresponsetofacetrustworthi-ness.Followingonthisexploratoryanalysis,thefirstobjec-tiveofthisarticlewastodevelopamodel-basedvalidationapproachfortestingtheroleoftheamygdalaintheevalua-tionoffacetrustworthiness.First,wedeterminedwhatfacialfeaturesareimportantforjudgmentsoftrustworthinessacrossparticipants.Second,webuilta3-dimensional(3D)computermodelforrepresentingfacetrustworthinessbasedonthesefeatures.Third,usingthismodel,wegeneratedtrustworthy-anduntrustworthy-lookingfaces(Figure1).Finally,usingfunctionalMagneticResonanceImaging(fMRI),wemeasuredhowneuralactivationchangesasafunctionofthetrustworthinessofthesemodel-generatedfaces.Thesecondobjectiveofthearticlewastotestnotonlyforlinearbutalsofornon-lineareffectsoffacetrustworthi-nessontheamygdalaresponse.Specifically,followingthecomputermodelingworkofOosterhofandTodorov(underreview)andthefindingsofSaidetal.(inpress),bothdescribedsubsequently,weexpectedthattheamygdalamightshowincreasedresponsetofacesonbothextremesofthetrustworthinessdimension.Inresearchconductedsubsequentlytostudy1,usingadata-drivenstatisticalmodelforfacerepresentation,OosterhofandTodorovbuiltamodelforrepresentingfacetrustworthiness.Theyarguedthatfaceevaluationofemotionallyneutralfacesisanovergeneralizationoffunctionallyadaptivesystemsfordetectionoftheemotionalstatesofothers(Knutson,1996;MontepareandDobish,2003).Specifically,judgmentsoftrustworthinessreflectdetectionofsubtlefacialfeaturesthatresembleemotionalexpressionssignalingapproach/avoidancebehavior(Todorov,inpress).Consistentwiththisargument,exaggeratingthefacialfeaturesinthenegativedirectionofthetrustworthinessdimensionproducedfacesexpressinganger,whereasexaggeratingthefacialfeaturesinthepositivedirectionofthedimensionproducedfacesexpressinghappi-ness.Theseexpressionssignaltotheperceiverwhethertheyshouldavoidorapproachthepersondisplayingtheemotion(cf.,Fridlund,1994).Giventhatseveralfunctionalneuroimagingstudieshavefoundincreasedamygdalaresponsetohappythantoneutralfaces(Breiteretal.,1996;Yangetal.,2002;Winstonetal2003;Pessoaetal.,2006),OosterhofandTodorov’sfindingssuggestthattrustworthyfacescanevokeastrongeramygdalaresponsethanfacesinthemiddleofthetrustworthinessdimension.Saidetal.(inpress)providedaconfirmationofthisprediction.Theymodeledbothlinearandquadraticcomponentsoftheamygdalaresponsetofacetrustworthi-nessandfoundthatthequadraticcomponentsprovidedabetterfitoftheamygdalaresponsethanthelinearcomponents.Theamygdalaresponsewasstrongertobothtrustworthyanduntrustworthyfacesthantofacesinthemiddleofthetrustworthinessdimension.However,con-sistentwiththepreviousfindingsoflinearamygdalaresponsetotrustworthiness(Winstonetal.,2002;Engelletal.,2007),theamygdalaresponsewasmoresensitivetodifferencesatthenegativethanatthepositiveendofthetrustworthinessdimension.Wesoughttoreplicatethisfindingwiththemodel-generatedfaces.STUDY1:CREATINGAMODELOFFACETheobjectivesofthefirststudywereto:(i)empiricallydeterminethefacialfeaturesimportantforjudgmentsoftrustworthinessand(ii)buildaparsimoniousmodelformanipulatingfacetrustworthinessbasedonthesefeatures.Usingadata-drivenstatisticalmodeloffacerepresentation(BlanzandVetter,1999;SingularInversions,2006),wegeneratedfaceswithneutralexpressionsandaskedpartici-pantstoratethesefacesontrustworthiness.Then,weregressedthemeantrustworthinessjudgmentsonthemodelvaluesofthefourfacialfeaturesshowingthehighestcorrelationwiththesejudgments.Consequently,theregres-sioncoefficientsestimatedfromthisanalysiswereusedtobuildaregressionmodelforpredictingthetrustworthinessofnovelfaces.WeusedthismodeltomanipulatethetrustworthinessofthefacesusedinthefMRIstudy.Twenty-oneundergraduatestudentspartici-patedinthebehavioralstudyforpartialcoursecredit.Statisticalmodeloffacerepresentation.WeusedtheFacegenModellerprogram(http://facegen.com)version3.1(SingularInversions,2006).Facegencreates3Dfaceswhoseshapeandtexturecanbeadjustedonmultipledimensions.ThefacemodelofFacegenisbasedonadatabaseofmale untrustworthy0 Fig.1ExamplesoffacesusedinthefMRIexperiment.Eachofthethreerowsshowstheuntrustworthy(ontheleft)andtrustworthy(ontheright)versionsofaface.Theirpositiononthetrustworthinessaxisindicatesthetrustworthinesspredictedbytheregressionmodel(seetextfordetails).120SCAN(2008)A.Todorovetal at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from Evaluatingfacetrustworthiness:amodelbasedapproachAlexanderTodorov,SeanG.Baron,andNikolaasN.OosterhofDepartmentofPsychologyandCenterfortheStudyofBrain,MindandBehavior,PrincetonUniversity,Princeton,NJ08540,USAJudgmentsoftrustworthinessfromfacesdeterminebasicapproach/avoidanceresponsesandapproximatethevalenceevaluationoffacesthatrunsacrossmultiplepersonjudgments.Here,basedontrustworthinessjudgmentsandusingacomputermodelforfacerepresentation,webuiltamodelforrepresentingfacetrustworthiness(study1).Usingthismodel,wegeneratednovelfaceswithanincreasedrangeoftrustworthinessandusedthesefacesasstimuliinafunctionalMagneticResonanceImagingstudy(study2).Althoughparticipantsdidnotengageinexplicitevaluationofthefaces,theamygdalaresponsechangedasafunctionoffacetrustworthiness.Anareaintherightamygdalashowedanegativelinearresponse
astheuntrustworthinessoffacesincreasedsodidtheamygdalaresponse.Areasintheleftandrightputamen,thelatterareaextendedintotheanteriorinsula,showedasimilarnegativelinearresponse.Theresponseintheleftamygdalawasquadratic
strongestforfacesonbothextremesofthetrustworthinessdimension.Themedialprefrontalcortexandprecuneusalsoshowedaquadraticresponse,buttheirresponsewasstrongesttofacesinthemiddlerangeofthetrustworthinessdimension.faces;personperception;trustworthiness;amygdalaPeopleevaluatefacesonmultipletraitdimensions(Ulemanetal.,2005)andtheseevaluationspredictimportantsocialoutcomesrangingfromelectoralsuccess(Todorovetal2005;BallewandTodorov,2007;Littleetal.,2007)tosentencingdecisions(Blairetal.,2004;Eberhardtetal2006).Aslittleas100msexposuretoafaceissufficientforpeopletomakeavarietyofpersonjudgmentssuchastrustworthiness,competenceandaggressiveness(WillisandTodorov,2006).Infact,theminimaltimeexposureafterwhichpeoplestartdiscriminatingbetweendifferentcate-goriesoffacesmaybeaslittleas3338ms(Baretal.,2006;etal.,underreview).Althoughpeoplemakemultiplepersonjudgmentsfromfaces,thesejudgmentsarehighlycorrelatedwitheachother,reflectingthevalenceevaluationthatunderliespersonjudgments(Rosenbergetal.,1968;KimandRosenberg,1980).OosterhofandTodorov(underreview)showedthatjudgmentsoftrustworthinessapproximatethisvalenceevaluation.Inaseriesofstudies,theyelicitedspontaneouspersondescriptionsoffacesandthenidentifiedthemostfrequenttraitdimensionsusedtodescribefaces.Judgmentsonthesetraitsweresubmittedtoaprincipalcomponentsanalysis.Allpositivetraitjudgmentshadpositiveloadingsandallnegativetraitjudgmentshadnegativeloadingonthefirstprincipalcomponent,whichaccountedformorethan60%ofthevarianceofthesejudgments.Outof13differenttraitjudgments,judgmentsoftrustworthinessshowedthehighestcorrelationwiththiscomponent.Thiscorrelationwaspracticallyunchangedwhentheprincipalcomponentwasobtainedfromallothertraitjudgmentsexcepttrust-worthiness.Thecorrelationbetweentrustworthinessjudg-mentsandthiscomponentalinearcombinationof12othertraitjudgmentswas0.94,indicatingthatthesejudgmentsapproximatethevalenceevaluationunderlyingmultiplesocialjudgmentsfromfaces.Thisfindingisconsistentwithpriorfindingsabouttheinvolvementoftheamygdalaintheevaluationoffacesontrustworthiness.Adolphsetal.(1998)showedthatpatientswithbilateralamygdaladamagewereimpairedindiscrimi-natinguntrustworthy-fromtrustworthy-lookingfaces.Twosubsequentfunctionalneuroimagingstudiesconfirmedtheinvolvementoftheamygdalainfaceevaluationontrustworthiness.Winstonetal.(2002)showedthattheamygdalaresponsetofacesincreasedastheirperceivedincreased.Thiswasthecaseforbothexplicitandimplicit(agejudgments)evaluationoftrust-worthiness.Engelletal.(2007)usedanimplicittaskandreplicatedtheWinstonetal.findings.WhereastheperceivedtrustworthinessoffacesinWinstonetal.(2002)wasassessedbysubjectivejudgmentsoftrustworthinesscollectedafterthefMRIstudy,Engelletal(2007)usedconsensusjudgmentsoftrustworthiness(aver-agedacrossraters)obtainedbyanindependentsampleofparticipants.Theamygdalaresponsewasbetterpredictedbytheconsensusjudgmentsoftrustworthinessthanbytheparticipants’ownjudgmentsoftrustworthiness(collectedaftertheimagingexperimentasinWinstonetal.’sstudy).BecauseconsensusjudgmentsreflectpropertiesoftheReceived10October2007;Accepted28February2008AdvanceAccesspublication26March2008WethankValerieLoehrforherhelpwiththedatacollectionandAndyEngellandChrisSaidfortheircommentsonpreviousversionsofthispaper.ThisresearchwassupportedbyNationalScienceFoundationGrantBCS-0446846.CorrespondenceshouldbeaddressedtoAlexanderTodorov,DepartmentofPsychology,PrincetonUniversity,Princeton,NJ08540,USA.E-mail:atodorov@princeton.edu.doi:10.1093/scan/nsn009SCAN(2008)3,119127TheAuthor(2008).PublishedbyOxfordUniversityPress.ForPermissions,pleaseemail:journals.permissions@oxfordjournals.org at Dartmouth College Library on July 12, 2013http://scan.oxfordjournals.org/Downloaded from