JournalofNeuroscienceMethodsxxx (2011) xxx–xxx
ContentslistsavailableatSciVerseScienceDirect
Journal
of
Neuroscience
Methods
j o ur n a l ho me p ag e :w w w . e l s e v i e r . c o m / l o c a t e / j n e u m e t h
A
novel
test
to
determine
the
significance
of
neural
selectivity
to
single
and
multiple
potentially
correlated
stimulus
features
Robin
A.A.
Ince
a,1,
Alberto
Mazzoni
b,1,
Andreas
Bartels
a,d,
Nikos
K.
Logothetis
a,e,
Stefano
Panzeri
b,c,f,∗aMaxPlanckInstituteforBiologicalCybernetics,Spemannstrasse38,72076Tübingen,Germany
bRobotics,BrainandCognitiveSciencesDepartment,IstitutoItalianodiTecnologia,ViaMorego30,16163Genova,Italy cBernsteinCenterforComputationalNeuroscience,Tübingen,Germany
dVisionandCognitionLab,CentreforIntegrativeNeuroscience,UniversityofTübingen,72076,Tübingen,Germany eDivisionofImagingScienceandBiomedicalEngineering,UniversityofManchester,ManchesterM139PT,UnitedKingdom fInstituteofNeuroscienceandPsychology,UniversityofGlasgow,Glasgow,UnitedKingdom
a
r
t
i
c
l
e
i
n
f
o
Articlehistory: Received30June2011
Receivedinrevisedform7November2011 Accepted8November2011 Keywords: Vision LFP Neuralcoding Naturalmovies Mutualinformation Oscillations Gammaband
a
b
s
t
r
a
c
t
Mutualinformationisaprinciplednon-linearmeasureofdependencebetweenstochasticvariables, whichiswidelyusedtostudytheselectivityofneuralresponsestoexternalstimuli.Herewedefine anddevelopasetofnovelstatisticalindependencetestsbasedonmutualinformation,which quan-tifythesignificanceofneuralselectivitytoeithersinglefeaturesortomultiple,potentiallycorrelated stimulusfeatureslikethoseoftenpresentinnaturalisticstimuli.Ifthevaluesofdifferentfeaturesare correlatedduringstimuluspresentation,itisdifficulttoestablishifonefeatureisgenuinelyencodedby theresponse,orifitinsteadappearstobeencodedonlyasasideeffectofitscorrelationwithanother genuinelyrepresentedfeature.Ourtestsprovideawaytodisambiguatebetweenthesetwopossibilities. Weuserealisticsimulationsofneuralresponsestunedtooneormorecorrelatedstimulusfeaturesto investigatehowlimitedsamplingbiascorrectionproceduresaffectthestatisticalpowerofsuch inde-pendencetests,andwecharacterizetheregimesinwhichthedistributionofinformationvaluesunder thenullhypothesiscanbeapproximatedbysimpledistributions(Chi-squareorGaussian).Finally,we applytheseteststoexperimentaldatatodeterminethesignificanceoftuningofthebandlimitedpower (BLP)ofthegamma[30–100Hz]frequencyrangeoftheprimaryvisualcorticallocalfieldpotentialto multiplecorrelatedfeaturesduringpresentationofnaturalisticmovies.WeshowthatgammaBLPcarries significant,genuineinformationaboutorientation,spacecontrastandtimecontrast,despitethestrong correlationsbetweenthesefeatures.
© 2011 Elsevier B.V. All rights reserved.
1. Introduction
Mutual information (Cover and Thomas, 1991) (hereafter referredtoasinformation)isameasureofdependencebetween stochasticvariables.Informationquantifiesthereductionin uncer-taintyaboutonevariable(forexample,asensorystimulus)after observationof another (forexample, a neural response). Infor-mation is a popularmetric toquantify, on a scale that allows meaningful comparisons acrossdifferent systems,how much a giventypeofaneuralresponseistunedtooneormorestimulus parameters(BorstandTheunissen,1999;Inceetal.,2010;Optican
∗ Correspondingauthorat:Robotics,BrainandCognitiveSciencesDepartment, IstitutoItalianodiTecnologia,ViaMorego30,16163Genova,Italy.
Tel.:+393316933490;fax:+390107170817. E-mailaddress:stefano.panzeri@iit.it(S.Panzeri). 1 Theseauthorscontributedequallytothiswork.
andRichmond,1987;Strongetal.,1998).Itsmainadvantageisthat
itisbasedonthefullprobabilitydistributionsoftheconsidered variablesandcanthereforequantifytheeffectofanynon-linearor arbitrarilycomplicateddependency.
Thecalculationofneuralinformationvaluesandtheevaluation of theirsignificance is madedifficultbythefact thatin neuro-physiologicalexperimentsthejointprobabilitiesofstimulusand neuralresponseshavetobeestimatedempiricallyfromalimited numberofdatasamples(“trials”).Limitedsamplinginducesboth asystematicerror(bias)anda varianceintheestimationofthe neuralinformation values. While thelimited samplingbiashas beenstudiedextensivelyintheneuralliterature(Paninski,2003;
Panzerietal.,2007;Victor,2006),theevaluationofthe
statisti-calsignificance of aninformation valuecomputedfromlimited experimentaldatahasbeenlessthoroughlyinvestigated.The sta-tisticalsignificance ofaninformationvaluetellsuswhetherwe can, at a particular confidence level or p-value, rejectthe null hypothesisthatthestimulusandresponseareindependent.The 0165-0270/$–seefrontmatter © 2011 Elsevier B.V. All rights reserved.
2 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx significance of information values is traditionally calculated by
employingaformofbootstraptest(Opticanetal.,1991).The stim-uliandresponseareshuffledacrossexperimentaltrialstoremove anydependence,andtheinformationvalueiscalculatedformany repetitionsofthisshufflingproceduretoobtainthedistributionof valuesunderthenullhypothesisofindependence.However, sev-eralproblemsrelated totheestimationofsignificance havenot yetbeenconsidered.The firstproblemis whetherit ispossible toemployparametricmodelsforthenullhypothesisdistribution. Thesecondproblemregardstheeffectofsamplingbias correc-tionsonthestatisticalpowerofsuchtests.Thethirdproblemisto determinewhichofthemanystimulusfeaturescharacterizing nat-uralisticstimuliactuallyinfluencetheconsideredneuralresponse. Thisisdifficultbecauseindynamicandcomplexstimuli,likethe onesfound in a natural environment and mimicked by movie stimuli,differentstimulusfeatures areoftenstronglycorrelated
(FelsenandDan,2005;Geisler,2008;Simoncelli,2003;Simoncelli
andOlshausen, 2001).Anabrupt movementofan observer,for
instance,whichleadstosuddenappearanceordisappearanceof objects,willnecessarilyproduceconcurrent,andhencecorrelated, changesinalargenumberoffeatures(Bartelsetal.,2008).Ifthe neuralresponsecarriessignificantinformationabouteachoftwo differentbut correlated features, it doesnot follow necessarily thattheneuralresponseisgenuinelyinfluencedbybothofthem. Theresponsecouldinsteadbeinfluencedbyonlyoneofthe fea-tures,withthedependenceontheotherfeatureappearingbecause ofthecorrelations betweenthem.Understandinghowto deter-minetrueneuronalselectivityinthesecasesisessentialtoallow meaningfulinvestigationoftheneuralrepresentation ofnatural stimuli.
Hereweinvestigatetheseissuesbyfirstcomparingbootstrap basedsignificancetestresultsforsinglefeatureinformationvalues includingdifferentbiascorrections,andpresentingcomparisons withananalyticexpressionforthenullhypothesisdistributionas wellasparametricfitsofsimpledistributionstosmallbootstrap samples.Thislaysthebasistounderstandhowbesttoperforma statisticaltestofindependencebasedonmutualinformation.We theninvestigatetheextenttowhichdependenciesbetween vari-ablescomplicatetheestimationofgenuinetuningtoeachfeature. Weconstructstatisticaltestsforinferringtrueselectivityinthe presenceofmultipleco-varyingfeatures.Wevalidatethese meth-odsbytestingthemonrealisticallysimulatedneuronalresponses tomultiplecorrelatedfeatures,andfinallyweillustratethemby investigatingtheextenttowhichthegammabandlimitedpower (BLP)oflocalfieldpotentials(LFPs)recordedfromtheV1areaof anesthetizedmacaquespresentedwithnaturalisticmoviesis gen-uinelytunedtooneormoreofseveralvisualfeaturescharacterizing themovies.We foundthat orientation,spacecontrastand time contrastareeach genuinelyencoded bythegammaBLP,inthe sensethattuningtoeach ofthesefeaturescannotbeaccounted forbytuningwithrespecttotheother,correlatedfeatures. More-over,wefoundthatcorrelationbetweenthesedifferentstimulus featuresacrossdifferentframesinthemovieincreasedtheoverall amountofinformationthatthegammaBLPcarriedaboutthevisual stimuli.
2. Testingthesignificanceofmutualinformationvalues 2.1. Mutualinformationandlimitedsamplingbias
Supposethatasensorystimulusspecifiedbymultiplefeatures ispresentedtoasubjectwhilerecordingneuralresponses.Oneway toquantifywhetheraneuralresponse,R,isselectivetoastimulus S,istomeasuretheinformationbetweenthesetofstimuliandthe setofassociatedneuralresponses(deRuytervanStevenincketal.,
1997;Fairhalletal.,2001;Panzerietal.,2003),whichisdefinedas
follows: I(R;S)=
s P(s) r P(r|s)log2P(r|s) P(r) (1)whereP(s)istheprobabilityofpresentationofstimuluss,P(r|s) istheprobabilityofobservingresponserwhenstimulussis pre-sented,andP(r)istheprobabilityofobservingaresponseracross allstimuluspresentations.Inthefollowing weassumethatthe values ofboth theneuralresponseR andthestimuli Sare dis-crete,oralternativelyarediscretizedintoafinitenumberofbins tofacilitatetheexperimentalsamplingofprobabilities. Informa-tionisnon-negativeandquantifiestheaveragereductionofthe uncertaintyaboutthestimulusthatcanbegainedfromobserving asingle-trialneuralresponse.Herewemeasureitinunitsofbits, onebitcorrespondingtoareductionofuncertaintybyafactorof two.Ifinformationislargerthanzero,theninprinciplethe stim-ulusfeatureandneuralresponsearenotindependent,indicating thattheresponsereflectstheparticularstimuluspresented.
DirectevaluationoftheinformationdefinedinEq.(1)fromthe experimentalhistogramsofstimulus-responseoccurrences (usu-ally calledthe “plugin” estimate, see Nemenmanet al. (2004)) resultsinasystematicupwardbiasintheinformationestimate. Thisoccursbecauseoffinitesampling;thevariabilityofthe dis-tributionswillbesystematicallyunder-estimatedbecauseofthe limitedamountofdataavailable.It canbeshownthat in suffi-cientlywellsampledsituations(moreprecisely,inthesocalled asymptoticregimeinwhicheachresponseisobservedmanytimes duringtheexperiment)thebiasoftheplug-ininformation esti-matecanbeapproximatedbythefollowingexpression(Panzeri
andTreves,1996): BIAS[I(R;S)]=2N1ln2
s ( ˆRs−1)−( ˆR −1) (2)whereNisthetotalnumberoftrials, ˆRsisthenumberof possi-bleresponsesuponpresentationofstimulussand ˆR isthenumber ofpossibleresponsesacrossallstimuluspresentations.Thesum intheright-handsideisoverallstimulipresented.Eq.(2)is use-fultobuildanintuitionaboutthebiasproblem. Assumingthat ˆ
Rsisapproximatelyequalto ˆR,andconstantacrossstimuli,then thebias(Eq.(2))isproportionalto(ˆS −1)( ˆR−1)/N,where ˆSisthe numberofpossiblestimuli.IfwedefineNs=(N/ˆS)tobethe aver-agenumberoftrialsperstimulusthenwecanseethebiasofthe informationisproportionalto(( ˆR−1)/Ns)−(( ˆR −1)/N)andhence approximatelyproportionalto ˆR/Ns.WethereforeconsiderNs/ˆRas thekeyparameter,whichaffectsthemagnitudeofthebias;ifthis ratioislarge,thebiaswillbesmallandviceversa.
Thereareseveralmethodstoestimatethebias(Montanietal.,
2007;Nemenmanetal.,2004;Paninski,2003;Panzerietal.,2007).
Forexample,thebiascanbeestimatedfromEq.(2)(andthen sub-tractedout).Thisrequiresestimatingparameterssuchas ˆRs,the numberofpossibleresponsestostimuluss.Thiscanbeestimated eitherfromsimplycountingthenumber ofobservedresponses (irrespectiveoftheprobabilityofanyparticularresponse)(Miller, 1955).However,incasesoflimiteddatathissimplecountingis likelytounderestimatethenumberof possibleresponses,since possibleresponseswithlowprobabilitymaynotappear.Toaddress thisissueaBayesiancountingprocedurecanbeused(seePanzeri
andTreves(1996),referredtointhefollowingasthePTmethod).
Alternatively,thequadraticextrapolation(QE)procedure(Strong etal.,1998)calculatesinformationfromthefulldatasetaswell asfromsubsetsofthedataconsistingofN/2andN/4trials,and thenestimates thetrueasymptoticvalueof informationby fit-tingaquadraticcurvetothesedatapoints.Whilethesemethods
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 3 havebeenshowntobegreatlyeffectiveinreducingthesystematic
errorduetolimitedsamplingininformationestimation(in par-ticular,theyareveryeffectiveifNs/ˆRislargerthan1,seePanzeri
etal.(2007)forareview),theeffectofthesebiascorrectionson
estimatingthesignificanceofinformationvalueshasbeenlargely unexploredsofar.
2.2. Thestatisticalsignificanceofmutualinformation
Toconcludethataneuralresponseisselectiveforaspecific fea-tureitisnotsufficienttomeasurethattheinformationbetween thetwovariablesislargerthanzero.Toruleoutthepossibilitythat thisinformationvalueresultsonlyfromfluctuationsduetolimited samplingratherthanfromtruedependency,onemustalsoassess thestatisticalsignificanceofthismeasure.
Thestatisticalsignificanceofinformationvaluesistraditionally evaluatedthroughthemethodofsurrogatedata,whichisreferred tohereasthebootstrapmethod(OpticanandRichmond,1987). Theavailablestimuliandresponsedataareshuffledtoremoveany dependencebetweenthevariables. Thedistributionof informa-tionvaluescomputedacrossdifferentinstantiationsoftheshuffled datasetfollowsthe distributionof theinformation valueunder thenullhypothesisthatthevariablesarestatisticallyindependent, sinceanydependencehasbeenremovedbytheshuffling proce-dure.Afterselectingtherequiredsignificancelevel(p-value),the correspondingpercentileofthebootstrappednullhypothesis dis-tributioniscalculatedandthevalueobtainedfromtheun-shuffled dataiscomparedtothistodeterminestatisticalsignificance.
Whilethebootstrapprocedureisstraightforwardtoperform, itisproblematicwhensmallp-valuesarerequired,sinceaccurate characterizationoftheextremesofthenulldistributionrequires highnumbersofbootstraprepetitions.Besidesbeing computation-allydemanding,obtainingahighnumberofindependentshuffled combinationsofthesamedatasetrequiresaverylargeamountof data,whichcaninsomecasesbedifficulttocollect.Itistherefore interestingtoconsiderwhetherparametricmodelscanbefitto smallerbootstrapsamplepopulations,withoutcompromisingthe statisticalpowerofthetest.Infact,itisalongestablished,although apparentlylittleknownresultthatforlargenumbersofsamplesthe distributionofplugin(notbiascorrected)informationestimates underthenullhypothesisofindependence,whenmultipliedbya factorof2Nln2,followsachi-squaredistribution(Fanetal.,2000;
Wilks,1938):
2N ln2·I(R;S)∼2(( ˆR−1)(ˆS −1)) (3) Ifthestimulipresentedaredrawnfromasetcontaining ˆS dif-ferent stimuli, and theresponse is represented with ˆR discrete symbols,thenthedegreesoffreedomforthischi-squarenull distri-butionare( ˆR−1)(ˆS−1).Notetherelationshipwiththefirstorder biasestimateintheprevioussection.Themeanofthischi-square distribution,whenconverted backintotheinformationscaleby dividingby2Nln2,ispreciselythefirstorderestimateofthebias.
Thisnull hypothesis distribution is thesame as the asymp-toticnulldistributionforPearson’schi-squaretestofindependence
(Kullback, 1968). In fact, the scaled mutual information value
describedaboveisequaltotheteststatisticforalikelihoodratio testofindependence.Thechi-squareteststatisticcanbederivedas aquadraticapproximationtothislikelihoodratiotest,bytakinga Taylorseriesexpansion(Kullback,1968).However,withthe avail-abilityofmoderncomputingresourcescalculatinglogarithmsisno longerasubstantialchallengeandsothereislessofanadvantageto usingtheapproximatequadraticform.Anadditionalpointof inter-estisthatthemutualinformationstatisticisitselfthedominant terminafullBayesiantestwithuniformpriors(Wolf,1995).
However,notmuchisknownyetabouttheextenttowhichthe nullhypothesisdistributioncanbeapproximatedbyparametric
distributionsinconditionsofmoderatesampling,andaboutthe effectofbiascorrectionsupontheeffectivenessofthebootstrap test.Theseissuesareofhighpracticalimportance.Inthefollowing weinvestigatethembymeansofrealisticnumericalsimulationsof neuralresponses.
2.3. Investigationofperformanceofinformationsignificancetests usingsimulations
Toinvestigatethestatisticsofinformationvaluesandthe effi-ciencyoftestsofinformationsignificance,weconductedaseriesof numericalexperimentswithmodelsofsimulatedneuralresponses, which are bothrealistic and relevant totheneurophysiological analysisthat we report inSection 5. In brief,thesimulation is constructedtomatchthefirstandsecondorderstatisticsofthe gamma-rangeBLPofLFPsrecordedinprimaryvisualcortex(V1)in responsetovariationsofMichelsoncontrastduringthe presenta-tionofanaturalisticmovie.
Asymptotictheoremsofspectralestimationpredictthat spec-tralpowersareapproximatelychi-squaredistributed(Percivaland
Walden,1993;Reichetal.,2000).Hence,thecuberootoftheBLP
isapproximatelyGaussian(WilsonandHilferty,1931),aproperty thatwasconfirmedempirically(notshown)forthedatawe ana-lyzedinSection5(seeMagrietal.(2009)foranexplicittestof thisassumption).Therefore, ourmodelgeneratedLFPresponses byassumingthattheircubicrootfolloweda Gaussian distribu-tionwiththesamemean andvarianceas thecubic rootof BLP powerfromanexamplerecordingsite(seeSection5.1)inmacaque V1inresponseto8quantizedlevelsoftheMichelsoncontrastin thereceptivefieldoftheconsideredrecordingsiteduringthe pre-sentationofaHollywoodmovie.Forcomputinginformation,the simulatedBLPgammaresponseswerediscretizedintoR=4 equi-populatedbins(exactlyaswedoforrealdatainSection5).Werefer
totheAppendixAforfulldetailsaboutthesimulation.
Intuitionsuggeststhathowdifficultitistodeterminewhether aninformationvalueissignificantdependsboth onthenumber oftrialsavailable(themoredatathereare,theeasieritisto sam-pletheprobabilities)andalsoontheamountofinformationinthe neuralresponses(thelessinformationthereisintheresponses,the moretrialsmaybeneededtodeterminingsignificance).To investi-gatetheseeffects,weintroducedaparameter˛thatmodulatedthe informationinthesimulatedresponsesbytransformingthemean andstandarddeviationparametersasfollows:
Pmodel
s =Psdata+(1−˛)(Psdata−Psdata) (4)
where Pmodel
s is the stimulus conditional parameter (mean or standarddeviation)ofthemodelforstimuluss,Pdata
s isthe
cor-respondingparametermeasuredfromthedata,Pdata
s istheaverage oftheparameteracrossstimuliand˛controlsthemodulationof theinformationinthesimulatedresponses.If˛=1,thesimulated responseshavethesameproperties(andinformation)asthe mea-sureddata.Decreasing˛decreasestheinformationinthemodel responses,untilfor˛=0themodelresponsesarenotstimulus mod-ulatedanymore,andhencetheycarrynoinformation.Weshow resultsfor˛=1(fullinformation),˛=0.5and˛=0(noinformation). We first investigated the statistical power of the bootstrap testindetectingsignificantinformationincaseswherethemodel responseshadtrueinformationaboutthesimulatedstimuli.We used100randombootstrapstoestimatethenullhypothesis distri-bution,whichcorrespondstorelativelylargesamplingcompared towhatcanbedonefeasiblyinanalysesofrealdatasets,andthus representagoodbenchmark.Fig.1Ashowshowthepercentage of correct detectionsofsignificant information,across different realizationsof thesimulation, depends onthenumber oftrials perstimulus.For ˛=1 (informationin simulationmatching the
4 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx
Fig.1.Determiningthesignificanceofmutualinformation.ModelLFPgammaBLPresponsestoeightdifferentvaluesofMichelsoncontrastweresimulated(seetextfor details).Thenumberoftrialsgeneratedperstimuluswasvaried;foreachnumberoftrials100realizationsofthesimulationweregeneratedandthecalculationswere performedoneachrealization.Thex-axesshowtheNs/Rratio,thecriticalparameterforinformationestimates(seetext,hereR=4).Panel(A)showsthepercentageof realizationsthatwereclassedassignificantusingavarietyofmethods.Solidlinesshowthefullinformationmodel(˛=1);dottedlinesshowthe˛=0.5model;dashed linesshowthenoinformation(˛=0)model.FortheQE(redlines)andPT(greenlines;partiallyobscuredbyoverlappingbluecurve)biascorrectionmethodsthecorrected informationvaluewascomparedtothebootstrapdistributionofcorrectedvaluescalculatedfrom100shufflingrepetitions.Forthechi-squaremethod(blacklines)the analyticalnullhypothesisdistributionwasusedandnobootstrapshufflingwasperformed.Panel(B)showsthemean(errorbarsshowstandarddeviation)over100realizations ofthep-valuefromtheKolmogorov–Smirnovtestofthegoodnessoffitofthe100bootstrapsamplestotheanalyticchi-squaredistribution.Panel(C)showsthepercentage ofstimulusconditionalresponsebinswithlessthan5observations.
oneinrealresponses),thebootstrapmethodreliablydetected sig-nificancedownto32 trialsper stimulus(which correspondsto Ns/R=8).Thisshowsthatdetectingsignificancerequiresmoretrials thanwouldbenormallyrequiredforbiasfreeinformation estima-tion(biastechniquesareshowntobeeffectiveifNs/ˆRislargerthan
one,seePanzerietal.(2007)).Whentheinformationcontentwas
reduced(˛=0.5),moresampleswererequiredtoreliablydetect significance.Thismatchestheintuitionthatwithenoughtrialsan arbitrarilysmalldependence couldbereliablydetected,while a verystrongdependencecouldbedetectedwithfewsamples.
Theanalysisofthemodelwithnoinformation(˛=0,Fig.1A, dashedlines)showsthefalsepositiverateofthetests.Wefound thatforallmethodsthiswasstableandclosetothespecified sig-nificancelevel(p=0.05).
Acrossallinformationlevelssimulated,theuseofbias correc-tionswasneverhelpfulinthebootstrapcalculationofsignificance: itreduced(QE)ordidnotaffect(PT)thestatisticalpowerofthetest. Thereductioninstatisticalpowerislikelytooccurbecauseofthe increasedvarianceofthebiascorrectedestimates;mostmethods tradeoffreducedbiasintheestimateforanincreasedvarianceof theestimator.Inthefollowing,wewillthereforeconcentrateon usinguncorrectedpluginestimatorstodeterminesignificance.
We then examined the effectiveness of using the para-metric analytical chi-square distribution of values under the
nullhypothesis,rather thantheempiricalnon-parametric boot-strap distribution. Fig. 1B shows that the analytic chi-square approximation,withdegrees offreedomequal to( ˆR−1)(ˆS−1), described thebootstrap distributionof information values well (Kolmogorov–Smirnov(KS)test,p>0.3)providedthattherewere atleast32trialsperstimulus,i.e.Ns/R>8.Inthissamplingregime, since the chi-squaredistribution fits thebootstrap distribution well,thechi-squarecanbeusedtobothreducecomputationaltime andtocomputesignificanceforsmallp-valueswhichcouldbe dif-ficulttotestwiththebootstrapmethod.Weverifiedthat,inthis samplingregime,theanalyticchi-squaretestwasasgoodasthe bootstraptestinbothdeterminingtruepositivesandavoidingfalse negatives.
Withfewertrials(Ns/R<8)theanalyticalchi-squarenolonger modeledtheobtainedbootstrapdistributionsaccurately.Inthese cases,theresponseswereunder-sampled,witha largenumber ofbins havingoccupancyofless than5 (Fig.1C).These results areinagreementwithconventionalstatisticalpracticeregarding theapplicationofchi-squaretests(Cochran,1954;Larntz,1978), whichsuggestsusingthechi-squareapproximationonlywhenthe numberofexpectedobservationsunderthenullhypothesisof inde-pendenceforeachpossiblestimulusconditionalresponseisgreater than1,andisgreaterthan5foratleast80%ofstimulusconditional responses.Insituationswherethiswasnotthecase,thestatistical
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 5
Fig.2.Parametricmodelsofnullhypothesisbootstrapdistribution.ModelLFPgammaBLPresponsestoeightdifferentvaluesofMichelsoncontrastweresimulated(seetext fordetails).Thenumberoftrialsgeneratedperstimuluswasvaried;foreachnumberoftrials100realizationsofthesimulationweregeneratedandthecalculationswere performedoneachrealization.Thex-axesshowtheNs/Rratio,thecriticalparameterforinformationestimates(seetext,hereR=4).Resultsareshownforthefullinformation model(˛=1),the˛=0.5modelandthenoinformation(˛=0)model.Foreachrealizationachi-square(greenlines)orGaussian(redlines)distributionwasfittoeither5 (dottedlines),10(dashedlines)or20(solidlines)shuffledbootstrapvalues,andthisdistributionwasusedtoestimatethesignificanceofthemeasuredinformationvalue.
testbasedonthechi-squaredistributionperformedpoorly,witha markedincreasedfalsepositiverate.
Inmanysituationsitisnotpossibletocomputealargenumber ofbootstrapsandsoitisdifficulttocomputesignificanceforsmall p-values. Besides using the analytic chi-square approximation, anotherpossiblewaytocomputesmallp-valuesignificanceunder suchcircumstancesistofitparametricmodelstonullhypothesis distributionsbasedonsmallnumbersofbootstrapsamples.These empiricallydeterminedparametricmodelscan thenbeusedto evaluatethesignificanceoftheobservedinformationvalue.InFig.2
wetestedtheeffectivenessofsuchprocedures,byfittingeither chi-squareorGaussiandistributionstodistributionsof informa-tionvaluesobtainedfromsmallnumbers(5–20)ofindependent bootstrapsamples. Eveninunder-sampled situationswhere, as discussedabove,theanalyticchi-squareapproximationfailsfor Ns/R<23 (shown byincreased false positiverate inFig. 2C and smallKStestp-valuesinFig.1B),goodresultswerestillobtained withtheseparametricfits.Thechi-squarefitperformancewasvery closetothefullnon-parametricbootstrap(95thpercentileof100 bootstrap samples), even whenonly 5 bootstrap sampleswere used.Thisprovidesa largecomputationalsaving.The Gaussian modelalsoperformedwell,but20bootstraptrialswererequiredto maintainaclosefittothenon-parametricbootstrapinthe under-sampledregime,andcontrolofthefalsepositiveratethroughout thesamplingrangesconsidered.Althoughthedegreesoffreedom arehigh,suggestingthechi-squareisclosetoaGaussian,when fit-tingtheGaussianmodeltwoparametersmustbeestimatedfrom thedataratherthanjustoneforthechi-squaredistribution.This
couldexplaintheincreasedperformanceofthesimplermodelfor lownumbersofsamples.
Insummary,ouranalysisofthetestsofindependencegavea numberofpracticallyusefulresults:(i)whileaccurateestimatesof informationvaluesrequirebiascorrection,whentestingfor signif-icanceofinformation,bestresultsareobtainedwithuncorrected plug-inestimates.(ii)Theanalyticchi-squarenullhypothesis dis-tributionallowsaveryrapidandhighlyprecisetestevenwithsmall p-values providedsamplingissufficient(Ns/R>8inour simula-tion).Theregimeinwhichtheanalyticchi-squareisaccuratecan bedeterminedinrealdatabyusingtheKStestbetweenempirical bootstrapandanalyticchi-squaredistributionsforsub-samplesof differentsizes.(iii)Eveninsub-sampledsituationswherethe ana-lyticchi-squareapproximationisnotaccurateanditisnotpossible tocomputemanybootstrapsamples,itisstillpossibletoobtain goodresultsbyempiricallyfittingachi-squaredistributiontoa smallnumberofbootstrapsamples.
2.4. Theeffectofnumberofbinsusedinthediscretizationofa continuoussignalupontheevaluationofsignificanceof information
Animportantstepintheinformationanalysisofdiscretized ana-logsignals,suchastheLFPBLPresponse,isthebinningofdata intoanumberofcategoricalbins.Twoissuesregardingthe bin-ningprocedurerequirefurtherdiscussion.The firstissueis the binningalgorithm tobechosen.Here,we settheboundariesof thebinssothattheyareapproximatelyequallyoccupiedoverall
6 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx
A
B
Fig.3.Effectofnumberofbinsoninformationestimationandstatisticalsignificance.ModelLFPgammaBLPresponsestoeightdifferentvaluesofMichelsoncontrastwere simulated(seetextfordetails).Panel(A)showsaverageinformationestimatesover100realizations(errorbarsshowstandarddeviation)ofthesimulationwiththesame numberoftrialsastheexperimentaldata(3460trialsperstimulus)asafunctionofnumberofbinsusedtodiscretizetheresponses.Informationisestimatedwithout biascorrection(‘Plugin’,redline),withPanzeri–Trevesbiascorrection(‘PT’,blueline)andwithquadraticextrapolation(‘QE’,greenline).Panel(B)showsthepercentage ofrealizationsfoundtobestatisticallysignificant(100realizations,p=0.05,bootstrapmethod),asafunctionofthenumberoftrialsperstimulus,fordifferentnumbersof responsebins.Thecorrespondinginformationvaluesforeachcurveareindicatedbycorrespondingcoloredmarksinpanel(A).Theverticalreddottedlineindicatesthe numberoftrialsavailableintheexperimentaldatasetandusedforthesimulationsinpanel(A).
theavailabletrials(theunconditionalormarginaldistributionsare approximatelyuniform).Thisavoidssamplingproblemscausedby binshavingverylowprobabilityofoccupancy,forexampleif equal-widthbinsareusedtorepresentaGaussiansignal,thosesampling thetailsofthedistributionhaveaverylowprobabilityof occur-rence.Thisproceduremakestheinformationessentiallyareduced rankstatistic,sincetherelativeposition(rank)ofaparticulartrial determineswhichbin itisin.Anadvantageofthisprocedureis that,likeotherrankstatistics,itisnon-parametricandrobustto outliers.Thesecondissuetobeconsideredisthat,whateverthe binningmethodused,thenumberofbinsusedfordiscretization isanimportant,yetpartlyarbitrary,parameterthatmaygreatly affecttheinformationcalculation.
Thechoiceofthenumberofbinsisacompromisebetweenthe needtousea fineenoughdiscretization inordertocaptureall possibleinformativefeaturesinthedata,andthenneedtousea dis-cretizationcoarseenoughthattheprobabilityofeachdiscretized responsecanbesampledaccuratelywiththeavailabledata.Forany givendataset,selectingmorebinsmeansrestrictingtheaverage numberoftrialsperbin.Inthefollowingweusecomputer simu-lationtoillustratehowtoempiricallyevaluatewhatistheoptimal trade-offbetweenthesetwoconflictingconstraintsinorderto eval-uatethesignificanceofinformationvalues.
Fig.3Ashowshowestimatesofthemutualinformationdepend onthenumberofresponsebinschosenforthesimulateddataset generatedwiththemodeldescribedinSection2.3.InFig.3Athe simulateddatasethadexactlythesamenumber oftrialsasthe realLFPdataanalyzedinSection5(3460trialsineachofthe8 stimuluscategories).Westudiedthedependenceofinformation valuesuponthenumberofbinsusingbias-correctedinformation estimates,becausethedependenceofbias-uncorrectedestimates onthenumberofbinsismaskedbythelinearincreaseofthebias withthisparameter(seeEq.(2)andthebias-uncorrected informa-tioncurveinFig.3A).Forverylownumbersofbins,increasingthe numberofbinsincreasestheinformation,becausethemorefinely
discretizedresponses bettercapture theunderlyingcontinuous distributions.However,bias-correctedinformationvaluesplateau overalargerangeofbinnumbers,inwhichtheinformationvalues arestableandlargelyindependentofthebiascorrectionprocedure used.Thismeansthat,withintherangeofnumberofbinsatwhich informationplateausthesamplingissufficienttoallowanaccurate estimationoftheinformation,butincreasingthenumberofbins doesnothaveaneffect;therearealreadysufficientbinstocapture therelevantfeaturesofthedistribution.Foranaccurateestimate oftheinformationvalue,anyparameterwithintheplateauregion canbechosen.Whenestimatingthesignificanceofthe informa-tion, it is intuitive that using more bins reduces the available samplesperbin,andhencethestatisticalpowerofthetest,but there must be sufficient bins for information to be detectable.
Fig.3Bshowsthepercentageofrealizationsinwhichsignificance wasdetected(p=0.05,bootstrapmethod)forthesimulated sys-temdescribedabove, asa functionoftheamountof generated data.Inthis case,usingjust 2binsprovidesthegreatest statis-ticalpower,sincetheinformationwithtwobinsisalreadylarge enoughandisalargefractionoftheplateauvalueofinformation (Fig.3A).
Theseresultssuggestsomepracticalconsiderationsthatmay guidethechoiceof thenumber ofbinsfor significance testing. Wheneverpossible,bias-correctedinformationestimatesofhow informationscaleswiththenumberofbins,orempirical obser-vationabouttheshapeandwidthofthedistributions,shouldbe usedtoevaluatetherangeofnumberofbinsthatgivessufficiently highinformationvalues. Thesmallestnumberof binssufficient topreservealargefractionoftheinformationshouldbeusedfor theinformationsignificancetestinginordertomaximizethe sta-tistical powerofthesignificance tests.Whenever possible,it is alsoimportanttoverifythestabilityoftheresultsovera range ofbinnumbersbecauseinsomecasesusingtoofewbinsmaylead toverylargelossesofinformation withrespecttotheirplateau values.
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 7 3. Themultiplefeaturetuningprobleminthecontextof
informationtheory
Ifasetofstimuliaredefinedbymultiplefeatures,Fi,wecan considerthemutualinformationbetweentheneuralresponseand eachindividualfeature:
I(R;Fi)=
fi P(fi) r P(r|fi)log2P(r|fi ) P(r) (5)wherefirepresentsvaluesofthefeatureFiandtheprobabilities areasinEq.(1).However,ifthesefeaturesarenotindependently distributedacrossdifferentpresentationsofthestimuli,apositive valueofinformationI(R;Fi)aboutfeatureFidoesnotnecessarily implythattheneuralresponsespecificallyencodesthatfeature. TheneuralresponsemaynotbeselectivetoFi,butinsteadbe selec-tivetoanotherfeature(sayFj),whichiscorrelatedwithFi.When usingcorrelatedstimuli,animportantproblemishowto distin-guishthetrueselectivitytoafeaturefromthatacquiredonlyfrom itsrelationstootherfeatures.
3.1. Multiplefeaturetuningandinformationcomplementarity
Hereweaddresstheissueofdetermininggenuinefeature selec-tivityinthepresenceofmultiplecorrelatedfeaturesbyconsidering, forsimplicity,thecaseinwhichthestimulusisdefinedbytwo dif-ferentcorrelatedfeatures,F1 andF2.Theamountofdependency betweenthetwo featuresismeasured, ininformationtheoretic terms,bytheinformationI(F1;F2)betweenthem:
I(F1;F2)=
f1,f2 P(f1f2)log2 P(f1f2 ) P(f1)P(f2) (6)whereP(f1f2)isthejointprobabilityoffeaturevaluesf1andf2being presentedinthesamestimulus.Thetwofeaturesaredependentif I(F1;F2)>0.Iftheneuralresponseisselectivetobothfeatures(i.e. I(R;F1)>0andI(R;F2)>0,withI(R;F1)>I(R;F2),howcanwemake surethatthegivenneuralresponsetrulyencodesbothfeatures, anddoesnotencodeF2onlybecauseitcorrelateswiththetruly encodedfeatureF1?Onewaytoaddressthisproblemistolook attheinformationtheresponseconveysaboutF2 atfixedvalues ofF1.SincethevalueofF1 isfixed,anydependencebetweenthe responseandF2 representsgenuinecomplementaryinformation thatcannotbeexplainedbythedependencybetweenthefeatures. Thisquantityiscalledtheconditionalmutualinformation(CMI) andisdefinedas:
I(R;F2|F1)=
f1 P(f1) f,r2 P(rf2|f1)log2 P(rf 2|f1) P(r|f1)P(f2|f1) (7)whereP(rf2|f1)isthejointprobabilityofobservingresponserand featuref2atfixedf1.Inasituationwhere
I(R;F2|F1)=0 (8)
itcanbeproved(CoverandThomas,1991)thatthefollowing equa-tionholdsforallvaluesoff2,f1andr
P(f1,f2,r)=P(f2)P(f1|f2)P(r|f1) (9) ThisillustratesthefactthatthevariableF2 affectstheneural responseonlythroughitscorrelationwithF1.IfinsteadI(R;F2|F1)>0 thenF2stillhasaneffectonthevariableratfixedF1andsoP(f1,f2,r) cannotbeexpressedbytheproductintheright-handsideofEq.(9). Insummary,thefeatureF2hasagenuineeffectupontheneural response,beyondtheindirecteffectofcorrelationswithF1,ifand onlyifI(R;F2|F1)>0.
Itisimportanttonotethat,byusingthechainrule(Coverand
Thomas,1991),theCMIcanberewrittenas:
I(R;F2|F1)=I(R;F2F1)−I(R;F1) (10) whereI(R;F1F2)istheinformationthattheneuralresponsecarried aboutthejointpresenceofF1andF2,definedasfollows:
I(R;F1F2)=
f1f2 P(f1f2) r P(r|f1f2)log2P(r|f1f2 ) P(r) (11)ThereforetheconditionI(R;F2|F1)>0isequivalenttothe condi-tion
I(R;F2F1)−I(R;F1)>0 (12)
Therequirementthatthejointinformationbehigherthanthe informationcarriedaboutthevariableF1 impliesthattheneural responsecarries“true”informationaboutF2aboveandbeyondthe informationthatitcarriesaboutF1,orinotherwords,thatthe infor-mationthattheneuralresponsecarriesaboutF2iscomplementary tothatitcarriesaboutF1(Panzerietal.,2010).Alternatively,due tothesymmetryofinformation,theaboveconditionmeansa bet-terpredictionoftheresponsecouldbemadebasedonobservation ofbothfeaturesthancouldbemadebasedontheobservationof F1alone.ItisintuitivethatgenuinecodingofF2requiresthatthe responsecarryinformationaboutF2complementarytothatwhich itcarriesaboutF1.Theaboveequationsformalizethisintuition.
Thisframeworkcaninprinciplebeextendedtothecaseinwhich thestimulusisdefinedbyanarbitrarynumbernoffeatures(F1,..., Fn).ThestimulusfeatureFn isgenuinelyencoded bytheneural responseiftheinformationI(R;Fn|F1...Fn−1)betweenresponse RandfeatureFn,conditionaltoallothervariablesissignificantly positive:I(R;Fn|F1...Fn−1)>0.However, whiletheextension to multiplefeaturesisstraightforwardinprinciple,theformalismmay bedifficulttoapplytomanyfeaturesbecauseofthedifficultiesin evaluatingmultivariateresponseprobabilitiesfromfiniteamounts
ofdata(Panzerietal.,2007).
3.2. Effectofcorrelationoffeaturesonestablishinginformation complementarity
Wementionedabovethattheproblemofcorrelationsbetween stimulusfeaturescomplicatestheproblemofdeterminingwhich stimulusfeaturesaregenuinelyencodedbytheneuralresponse. In order todevelop a methodsuited totackle this problemon experimental data, we investigate theoretically how statistical dependenciesbetweenfeaturesinfluencetheamountofjoint infor-mationcarriedaboutthefeatures.Forsimplicitywewilllimitour discussiontoastimulusdefinedbyonlytwofeatures,F1andF2.
TherelationshipbetweentheCMI,I(R;F2|F1),andthecorrelation betweenvariablesI(F1;F2)canbedefinedasfollows(Adelmanetal.,
2003;Lüdtkeetal.,2008;Schneidmanetal.,2003):
I(R;F2|F1)=I(R;F2)+I(F1;F2|R)−I(F1;F2) (13) whereI(F1;F2)istheamountofcorrelationbetweenvariables with-outregard totheneuralresponse(seeEq.(6)),and I(F1;F2|R)is theamountof correlationbetweenthefeatures at fixedneural response,definedasfollows:
I(F1;F2|R)=
r P(r) f1,f2 P(f1f2|r)log2P(fP(f1f2|r) 1|r)P(f2|r) (14)Eq.(13)hasseveralimportantimplications.First,itshowsthatif thefeaturesareindependentlydistributed(i.e.I(F1;F2)=0)thenthe conditionI(R;F2)>0issufficienttoensurethatF2carries informa-tioncomplementarytothatofF1.Thisisbecausetheterm−I(F1;F2) istheonlyoneintheright-handsideofEq.(13),whichcanbe neg-ative.Moreover,andforthesamereason,ifthevalueofI(F1;F2)is
8 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx quantitativelymuchsmallerthanthatoftheinformationcarried
bytheneuralresponseabouteachoftheconsideredfeatures,itis safetoconcludethattheinformationabouteachfeatureisgenuine. Thisgivesausefulruleofthumbforinterpretationoftheresults. Second,Eq.(13)showsthatthepresenceofcorrelationbetween featuresdoesnotalwaystendtosuppressthecomplementarityof informationbetweenfeatures.Infactitcanbeshownthattheterm I(F1;F2|R)−I(F1;F2),sometimescalledtheinteractioninformation, canbeeitherpositiveornegativedependingonwhether correla-tionsbetweenfeaturesacrossallscenesareweaker orstronger thancorrelationsbetweenfeaturesatfixedneuralresponse(Pola
etal.,2003;Schneidmanetal.,2003).Thisinteractioninformation
isequaltothesynergy(Brenneretal.,2000b)usuallydefinedas Syn(F1,F2;R)=I(F1,F2;R)−I(F1;R)−I(F2;R) (15)
Together,usingEqs.(13)and(15)weseethat
I(R;F2|F1)=I(F2;R)+Syn(F1,F2;R) (16) Thisshows that theabsence of complementaryinformation (I(R;F2|F1)=0)impliesthepresenceofredundancy(negative syn-ergy) between F1 and F2. Similarly, if the features are coded synergistically,thecomplementaryconditionalmutual informa-tioncanbegreaterthantheunconditionalinformationconveyed aboutasinglefeature.Insuchcases,thecodingofgroupsoffeatures wouldbedifficulttofullycharacterizewithstandarduncorrelated stimulusdesigns.These theoreticalconsiderationsreinforcethe importanceofsuchmethodstoenablequantitativeinvestigation ofthecodingofcorrelatedfeaturesfoundinnaturalstimulation conditions.Indeed,ithasbeenproposedthatresponsestonatural stimuliareoptimizedtocopewithcorrelationsintheenvironment
(Barlow,1989;Danetal.,1996;Geisler,2008;OlshausenandField,
1996);themethodsdiscussedhereallowquantitative investiga-tionsoftheseideas.
4. Testingthesignificanceofcomplementarytuningto multiplefeaturesusingconditionalmutualinformation
Wenowconsidertheproblemofhowtoextendtheabovetests forsignificance ofmutual information toevaluatingthe signifi-canceofcomplementarityoftuningtomultiplefeaturesusingCMI. Thisproblemisdifficultbecausetheinclusionofmultiplefeatures exponentiallyincreasesthesizeofthestimulusspace,which corre-spondinglyreducesthenumberoftrialsavailableforeachstimulus combinationandcompoundsthesamplingdifficulties.The corre-lationsbetweenthefeatures canalsocausedifficultiesbecause theytendtoconcentratethejointdistributionsofthepresented combinationsof features, and the potentially uneven sampling introducedbycorrelationsamongfeaturesmustbecorrectlytaken intoaccountwhendesigningtheappropriatebootstrapshuffling andalgorithmsthatevaluatesignificanceofcomplementary infor-mationvalues.
4.1. Thestatisticalsignificanceofconditionalmutualinformation
Theconditionalmutualinformationcanbewrittenas I(R;F2|F1)=
f1
P(f1)I(R;F2|f1) (17)
whereI(R;F2|f1)istheinformationbetweenRandF2,conditioned onaspecificfixedvaluef1offeatureF1.Bysubstitutingintheabove equationP(f1)=Nf1/N,whereNf1 isthenumberoftrialsinwhich featureF1takesthevaluef1,oneobtains
I(R;F2|F1)= 1 2Nln(2)
f1 2Nf1ln(2)I(R;F2|f1) (18)AsdiscussedinSection2,whenusingpluginuncorrected infor-mationestimates,eachelementofthesumontheright-handside ofEq.(18)followsachi-squaredistributionwith( ˆR−1)( ˆF2−1) degreesoffreedom,underthenull hypothesisthatRandF2 are independentatfixedvalues of F1.Fromtheadditivity property ofthechi-squaredistributionwefind(assumingthenumberof observedresponses ˆRandfeaturevalues ˆF2arethesameateach fixedf1,seebelow)that2Nln(2)I(R;F2|F1)ischi-squaredistributed with ˆF1( ˆR −1)( ˆF2−1)degreesoffreedom.
Itisalsopossibletoevaluatethestatisticalsignificanceofthe conditionalmutualinformationusingthebootstrapapproachas describedinSection2.However,caremustbetakenwhen imple-menting theshuffling proceduretoensureit samplesfromthe requirednullhypothesis.Sincethenullhypothesisisthatfeature F2andresponseRareindependent,atfixedvaluesofF1,samples shouldbedrawnbyshufflingthecombinationofF2andRatfixed F1.ThesimplerprocedureofshufflingF2valueswithout consider-ingthecorrespondingF1valuewouldindeedbeincorrect(andwe verifiedthatitisalsolessaccurate,resultsnotshown)becauseit removesalltheinformationinF2,andnotonlythatadditionaltoF1. Whentheresponsefeaturesarehighlycorrelatedorwhenthe neuralresponseisselectiveonlytoaspecificcombinationof fea-tures,thenthenumberofrelevantresponsebinsforbothfeatures maybesmallerthantheproductofthenumberofrelevantbinsfor eachfeature.TherandompermutationofF2withoutconsidering thevalueofF1wouldthenmakethejointfeaturespaceatfixed responselargerthanintherealdata.Asaconsequencethenull hypothesisdistributionwillhavealargerbiasthanthetrueone
(Panzerietal.,2007),andalsoalargervariance.Thiswouldmake
theevaluationoftherequirednullhypothesisproblematic.Indeed, insuchcasestheaboveexpressionforthedegreesoffreedomof thechi-squareapproximationshouldbemodified.Insteadof tak-ing ˆF1( ˆR−1)( ˆF2−1),thevaluesof( ˆR−1)( ˆF2−1)observedforeach valuef1shouldbesummedtoobtainthetotaldegreesoffreedom. 4.2. Numericalinvestigationoftheperformanceofstatistical testsforinformationcomplementary
Toinvestigatethestatisticalpowerofthesetests,weperformed simulationswithasystemsimilartothatdescribedinSection2, butextendedtoincludetwocorrelatedfeatures.Again,themodel isbasedascloselyaspossibleonthestatisticsoftheexperimental dataconsideredinSection5,anditmatchesthefirstandsecond orderstatisticsofV1 gammaBLPLFPresponsestovariationsin spatialcontrastandtemporalcontrastduringpresentationof Hol-lywoodcolormovies.Thecorrelationsinthesimulatedfeatures alsomatchedthoseobservedinthemoviesusedintheexperiments. AsinSection2,thesimulatedgammaBLPresponsewasbinnedin R=4equi-populatedlevels,andeachsimulatedfeaturewasbinned into ˆFi=8equi-populatedclasses(seeAppendixAforfulldetails). Itisintuitivethatthesensitivityofanytestfor complementar-ityofinformationmustdependontheactualamountofadditional genuinecomplementary informationof thesecond feature(the largeritisit,theeasieritistodetect).Thusinoursimulationthe responsedependenceonF1(Michelsoncontrast)wasfixedinall simulations,butweconsideredtwodifferentcaseswithregardto thesecondcorrelatedfeatureF2 (temporalcontrast):inthefirst simulationitcarried anamountofadditionalinformationtoF1, andinasecondsimulationitdidnotcarryanyadditional informa-tion,althoughitwasinformativeperseduetoitscorrelationtoF1
(seeAppendixA).
We tested this bootstrap procedure in Fig. 4. We applied this procedure to uncorrected plugin estimates of information because,forthereasonsexplainedinSection2,theseuncorrected information estimates are better at determining significance.
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 9
Fig.4.Determiningthesignificanceofconditionalmutualinformation.ModelgammaresponsestoeightdifferentvaluesofMichelsoncontrastandeightdifferentvaluesof temporalcontrastweresimulated(seetextfordetails).Thenumberoftrialsgeneratedperstimuluswasvaried;foreachnumberoftrials500realizationsofthesimulation weregeneratedandthecalculationsperformedoneachrealization.Thex-axesshowtheNs/RratiowhereNsistheaveragenumberoftrialsforeachpossiblestimulusfeature pair(R=4).Panel(A)showsthepercentageofrealizationswheretheCMIbetweenresponseandtemporalcontrastgivenMichelsoncontrastwasfoundtobesignificantusing theuncorrectedbootstrap(redlines)andanalyticchi-square(blacklines)methods.Dottedlinesshowthemodelwithindependentrepresentationoftemporalcontrast; solidlinesshowthemodelwithnoindependentrepresentationoftemporalcontrast.Panel(B)showsthemean(errorbarsshowstandarddeviation)over100realizations ofthep-valuefromtheKolmogorov–Smirnovtestofthegoodnessoffitofthe100bootstrapsamplestotheanalyticchi-squaredistribution.
theCMIwascorrectlyfoundtobesignificant(i.e.percentageoftrue positives)dependedupontheNs/Rratio.NotethatinthiscaseNs istheaveragenumberoftrialsforeachpossiblestimulusfeature pair,andRisthenumberofresponses.Thebootstraptestwasable toreliablydetectthepresenceofsignificantcomplementary infor-mationforsystemswithNs/R>4(16trialsforeachpairofstimulus featurevalues,128trialsperF1stimulusfeaturevalue).
Again,thechi-squareapproximationwasaccurateprovidedthe samplingwassufficient.Forthesinglefeaturesystemshownin
Fig.1,thechi-squareapproximationstartedtofailatNs/R=23.For thisconditionalsystemNs/R=25isrequired.Thisisshownbythe increasedfalsepositive ratefortheanalyticchi-square approxi-mationwhennogenuinecomplementaryinformationispresent (Fig. 4A),and by thelow p-values for the KS-testbetweenthe chi-squareapproximationandthebootstrapdistributionforboth simulatedsystems(Fig.4B).Thefactthatmoretrialsarerequired herefortheanalyticchi-squaretoholdislikelyduetothestrong correlations betweenfeatures changing thestatistics of theF1 -conditionaldistributions;thef2valuesatfixedf1arenotnecessarily uniformlydistributedastheywereinthesinglefeaturemodel.We verifiedthisfactbydemonstratingthat insimulationsinwhich thecorrelationbetweenF1andF2wassettozero,thechi-square approximationwaseffectiveevenwithNs/R=23,asinthe individ-ualfeaturecase(resultsnotshown).
Insummary, ouranalysisoftheeffectivenessof testsof sig-nificanceofcomplementarity ofinformationrevealedthat:(i)a
bootstraptestbaseduponcomputinguncorrectedplugin informa-tionestimatesaftershufflingafeaturesconditionaltothevalueof other(s)isextremelyeffectiveatdetectingsignificancewithsmall amountsofbothfalsepositivesandnegatives(ii)Thebootstraptest canbereplacedbyananalyticalchi-squaretestprovidedsampling issufficient(Ns/R>32inoursimulation).Theregimeinwhichthe analyticchi-squareisaccuratecanbedeterminedinrealdataby usingtheKStestbetweenempiricalbootstrapandanalytic chi-squaredistributionsforsub-samplesofdifferentsizes.
5. Applicationoftheformalismtotheencodingofvisual featuresbygammaBLPinV1
Toillustratethefeasibilityoftheaforementionedformalismon neurophysiologicaldata,weusedittoanalyzeLFPsrecordedinV1 inresponsetorepeatedbinocularpresentationofacolormovie. Thisisausefulapplicationofourtestsforatleasttworeasons.
Thefirstreasonisthatthepreciseorigin,meaningandstimulus tuningoflocalfieldpotentialsis notyetfullyknownand isthe subjectofcontinuingstudies(Belitski etal.,2008;Berensetal.,
2008;Jiaetal.,2011;Katzneretal.,2009;KayserandKonig,2004;
Mazzonietal.,2008).Becauseofthis,thereisapressingneedfor
tools,whichareabletocorrectlyidentifyandquantifythetuning ofthesesignalstocomplexstimuli.
10 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx Thesecondreasonisthatthesedatacollectresponsesto
com-plexnaturalisticstimuli,whichcontainmanyvisualfeaturesthat co-varyovertime(Simoncelli,2003);forinstancemovingobjects mayenterorleavetheRForbeoccluded byotherfastmoving objects,orlocalilluminationmaychangedynamically,both caus-ingalargenumberoffeaturestochangeinacorrelatedfashion
(Bartelsetal.,2008).Evidentlysuchcorrelatedchangesmaypose
problemstostudiesexaminingtheselectivityofneuronal popula-tionstoindividualfeaturesofcomplexvisualimages.Duringthe presentationofsimplified,artificialvisualstimuli,theproblemcan bereadilysolvedbyvaryingdifferentvisualattributesinan uncor-relatedmanner(Brenneretal.,2000a;deRuytervanSteveninckand
Bialek,1988;Petersenetal.,2008;Rustetal.,2005;Touryanetal.,
2005;YamadaandLewis,1999).Insuchdesigns,demonstration
ofneuronalfeature-selectivitysimplyfollowstheselective stim-ulusparameterization.Yet,suchanuncorrelatedstimulusdesign cannotbeappliedwhenusingnaturalisticstimulisuchasmovies orsoundscapturingthecomplexityandcharacteristicsofa natu-ralenvironment.Analysingneuralresponsestonaturalisticstimuli isimportantforatleasttworeasons.First,naturalisticstimuliare likelytoengagecomplexpatternsofactivityincortical microcir-cuitsthatmaynotbeelicited inthepresenceofsimple stimuli optimizedforthestudyofbasicpropertiesofsingleneurons(Felsen
andDan,2005;Reinagel,2001).Second,naturalisticvisual
stim-ulihavebeenshown toevokemore reliableresponses(Hasson etal.,2009),andmorespecificinter-regionalcorrelations(Bartels
andZeki,2005),suggestingthatcodinginthevisualsystemmay
beoptimizedforprocessingnaturalisticstimuli(Danetal.,1996). Hereweprobetheabilityofourmethodstoquantifythestatistical significanceofinformationcarriedbyLFPsaboutagivenvisual fea-tureaboveandbeyondthatcarriedbyother,potentiallyco-varying visualfeatures.
5.1. Experimentalprocedures
Beforeproceedingwiththeanalysis,webrieflysummarizethe experimentalproceduresusedtorecordneuralresponsesto nat-uralisticcolormoviesinV1.Adetaileddescriptioncanbefound inpreviousstudies(Belitskietal.,2008,2010;Montemurroetal., 2008). All procedures were approved by the local authorities (Regierungspräsidium)andwereinfullcompliancewiththe guide-linesoftheEuropeanCommunity(EUVD86/609/EEC)forthecare anduseoflaboratoryanimals.Werecordedneuralactivitywith anarrayofextracellularelectrodesfromtheopercular(fovaland para-fovealfield-representations)portionofV1,of anesthetized macaquespresentedwithcommerciallyavailablecolormovieclips consistingof scenesfromTheLastSamurai(2003,WarnerBros. Pictures)andStarWars:EpisodeI–ThePhantomMenace(1999, Lucasfilm).Weanalyzedatotalof55recordingsitesfrom3different monkeys.Movieclipslasted150–210sandwerepresented binoc-ularlyataresolutionof640×480pixels(fieldofview:30◦×23◦, 24bittruecolor,60Hzrefresh)usingafiber-opticsystem(Avotec, SilentVision,Florida).Eachmoviewaspresented35–50timesper sessioninordertoadequatelysampletheprobabilitydistribution oftheneuralresponsestoeachpartofthemovie.
Theextracellularfield potentialwasfilteredtoextract Multi UnitActivity(MUA)andlocalfieldpotentialsusingstandardsignal conditioningtechniquesasdescribedindetailelsewhere(Belitski etal.,2008).TheMUAwasusedonlytodeterminethereceptive fields,asdescribedbelow.Thegammaband(30–100Hz)oftheLFPs wasextractedbymeansofzero-phaseshift, bidirectionalfilters (Kaiserfilterwithatransitionbandof1Hz,stop-bandattenuation of60dB,andpass-bandrippleof0.01dB).Theinstantaneouspower ofthisband-limitedsignal,shortenedinthefollowingasgamma bandlimitedpower,wascomputedasthesquaredmodulusofthe discrete-timeanalyticsignalobtainedviatheHilberttransform.
WefocushereonthegammaBLPasaneuralresponsesinceithas beenpreviouslyshowntobeinformativeaboutnaturalisticmovie stimulation(Belitskietal.,2008)andisknowntobemodulatedby visualfeaturessuchasspatialcontrast(HenrieandShapley,2005) andorientation(KayserandKonig,2004).
5.2. Extractionofvisualfeatures
Wefirstestimated theaggregatereceptive field(RF)of each recordingsitebyusingthereversecorrelationtechnique(Ringach
andShapley,2004),which measuresthesensitivity oftheMUA
responsetothetimecontrastofeachpixelofthescreen.When usingreversecorrelationwithspatiallycorrelatedstimuli(suchas thenaturalmoviesusedhere)theobtainedRFislikelybiasedby thestimuluscorrelationandisthereforelikelytobelargerthanthe
trueRF(Theunissenetal.,2001;Touryanetal.,2005).Forthis
rea-son,theRFchosenforthefeatureextractionanalysiswasfixedasa 1◦× 1◦regionaroundthepeakoftheneuralsensitivity,evenwhen theresultofreversecorrelationresultedinalargermap.Thissize ofRFwaschosenbecauseitistypicalofmultipleunitV1 recep-tivefields(Sceniaketal.,1999).AnexampleofsuchRFestimation isreportedinFig.5A.Reversecorrelationwasusedheretoobtain anestimateofthelocationoftheRFareaandnottoestimateits size.ToverifythattheprecisesizesetfortheRFwasnot criti-calfortheinformationresults,werepeatedtheanalysisselecting RFssizesof0.5◦×0.5◦and1.5◦×1.5◦.Wefoundthatthefeatures extractedfromRFsoftheselargerorsmallersizeswereessentially identicaltothoseobtainedwiththe1◦×1◦RF(correlationbetween thetimecoursesoffeaturesextractedwithdifferentRFsizeswas >0.95forallelectrodesandrecordingsessions,andinformation val-ueswerenotsignificantlydifferent,p>0.5).Asanadditionalcheck thatthelocationoftheRFwascorrectlydetermined,wecompared theRFpositionsobtainedbyreversecorrelationwiththose manu-allymappedbytheexperimenterwithsmallpolarstimuliduring theexperimentalsession,andwefoundexcellentagreement(not shown)forallsites.
We subsequently computed three different types of image features within each RF: spatial contrast (Michelson contrast, calculatedas thedifferencebetweenthemaximaland minimal luminance dividedby theaveragebetweenmaximal and mini-malluminance),orientation(computedasthedirectionofgradient ofcontrast(Kayseretal.,2003)),andameasureoftimecontrast, definedasthetotalamountofpixel-wiseluminancechangefrom frametoframe,thatis,theabsolutevalueoftheluminance differ-encebetweentwosuccessiveframes,calculatedforeverypixeland thenaveragedovertheRF(Bartelsetal.,2008).
5.3. Calculationofinformationconveyedaboutdifferentfeatures
Foreachrecordingsite,wecomputedtheinformationcarried bythegammaBLPaboutthecurrentvalueofthemoviefeatures inthecorrespondingRF.Forthiscalculation,thestimulussetwas createdasfollows(seealsoReichetal.(2000)).Wefirstdivided themoviepresentationintonon-overlappingadjacentwindows each10frameslong.Thiswindowlengthwaschosenbecauseit matchedthetimescaleofthefeatureautocorrelations.We aver-agedthevalueofeachfeatureoverthe10framesineachstimulus window.Unlessotherwisestated,wediscretizedtheresulting fea-turevalues acrossallwindowsinto8equallypopulatedclasses, andeachclasswasconsideredasadifferent“stimulus”valuefor theinformationcalculation.GammaBLPvalueswereaveragedover thesameintervalsandsubsequentlydiscretizedintoR=4equally populatedbins.We chosetheseparametersbecausetheyallow accuratesamplinggiventheamountofdataavailable.Wechose R=4basedontheconsiderationsinSection2.4,whichsuggest cau-tiouschoicesofthenumberofbins,especiallyconsideringthatthe
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 11
Fig.5.Exampleofmoviefeaturesandneuralactivity.(A)Foreachrecordingsite,thereceptivefieldwasdeterminedasthe1◦squareareainthescreenwherethecovariance betweenthemulti-unitactivityrecordedinthesiteandthetimecontrastwashighest.Thelocalfieldpotential(LFP)recordedinthesitewasthenstudiedasafunctionofthe visualfeaturesofthemoviecomputedwithintheassociatedreceptivefield.(B)Comparisonofthegammaactivityrecordedduringasingletrialinarepresentativerecording site(fromsessionD04nm1)withthemoviefeaturescomputedintheassociatedreceptivefield.Fromtoptobottom:LFPgammaband(30–100Hz),spatialcontrast,sineof theorientation,timecontrast.
samplingdemandsofconditionalmutualinformationaregreater thanforthesinglefeatureinformationconsideredthere.Wechose alargernumberofbinsforthestimulusfeatures,sinceouraimwas tocharacterizetheroleofcorrelationsbetweentheminevaluating tuningofneuronstomultiplefeatures.
Wecomputeddifferentsetsofinformationmeasurements.For eachinformationvalue,wefirstcomputedthesignificanceofthe uncorrectedplug-ininformationestimatewiththeanalytic chi-square approximationdescribed in Section 2.2for the absolute informationandinSection4.1fortheconditionalinformation.For thisanalysiswehad20,000–30,000trialsavailable(25,500±3100, mean±std).ThiscorrespondstoNs/R∼800fortheestimationof informationaboutanindividualfeature,andNs/R∼100forthe esti-mationofCMI.Thesenumbersarewellinsidetheregimeswhere thechi-squareapproximationsperformwell(Sections2and4). Wethencomputedanestimateoftheinformation,usingEq.(5)
fortheinformationbetweengammaBLPandfeatures,Eq.(6)for theinformationbetweenfeatures,andEq.(7)fortheconditional information,correctingforthebiaswiththePTcorrectionmethod
(PanzeriandTreves, 1996)describedby Eq.(2).Allinformation
measuresonexperimentaldatawerecomputedusingthe Infor-mationBreakdownToolbox2(Magrietal.,2009).
5.4. Results:complementaritycodingoforientation,spatial contrastandtimecontrastbytheLFPgammapower
Thetimecourseofthreeextractedmoviefeatures(spatial con-trast, time contrast and orientation)in the RF (Fig. 5A) on an examplerecordingsiteisshowninFig.5B,togetherwiththe asso-ciatedgammabandoftheLFPrecordedfromthatsite.Fromthis example,itisapparentthatthegammaBLPistuned,todifferent extents,tothesethreefeatures.Thisisconsistentwithprevious
2http://www.ibtb.org.
reportsaboutgamma-bandselectivityobtainedwithsimple
stim-uli(Frienetal.,2000;HenrieandShapley,2005).
Toquantifythesedependencies,wecomputedthemutual infor-mationbetweenthegammaBLPandthedifferentfeaturesandwe testedthesignificanceoftuningofeachofthe55LFPrecording sites.Tothisend,weusedtheanalyticalchi-squareinformationtest onthebiasuncorrectedpluginestimates(becausethisisthemost sensitivetestunderthesamplingconditionsanalyzed).Wefound thatallrecordingsitescarriedsignificantinformationatp<0.001 confidencelevelforallthreefeatures(Fig.6A).While responses weresignificantlymodulatedbyallstimulusfeaturesconsidered, we foundthat the populationaverage(computed withPTbias correction)oftheinformationaboutspatialortemporalcontrast wasmuchhigherthantheinformationaboutorientation(Fig.6B), showingthatorientationcausedweaker(althoughstillhighly sig-nificant)responsemodulationthantheotherfeatures.
Weevaluatedtherobustnessofthesignificancetestbyusing onlyasubsetofthemoviepresentations(andhenceforthofthe trials),similarlytowhatwasdoneinFig.1A.Wefound(Fig.6C) thatforthemostinformativefeatures(spatialandtemporal con-trast)significantresponsemodulationwasdetectedinalmostall recordingsites(>95%)evenwhenonly6moviepresentationswere considered(ofthe35–50available),andusingonlytwomovie pre-sentationswassufficienttodetectsignificantinformation about thesetwocontrastfeaturesinmorethan50%ofrecordingsites. Fortheorientationfeature,moredatawasrequiredtofind depen-denceinasizablefractionofrecordingsitesbecauseofthesmaller information.However,evenwithonlytwomoviepresentations, morethan30%ofrecordingsiteswerefoundtobesignificantly modulatedbyallthreefeatures.
AsdescribedinSection3,tounderstandwhetherthegamma BLPwastrulytunedtoallthreestimulusfeaturesandwasnot car-ryinginformationaboutsomeofthemonlybecauseofcorrelation toanothertrulyencodedfeature,wehadtocharacterizewhether ornotthedifferentfeaturesintheRFvaryindependentlyduring themovietimecourse.Wefoundthatthevisualfeaturesdidnot
12 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx
Fig.6.Informationcarriedbygammabandlimitedpower(BLP)aboutspatialcontrast,orientationandtimecontrast.(A)Informationaboutthevalueofthedifferentfeatures carriedbygammaBLPforeachrecordingsite,computedwith4classesforgammaBLP,8classesforfeaturesandnobiascorrection.Significancewascomputedwiththe chi-squareapproximation,p<0.001.SincethenumberofclassesforgammaBLPandfeaturesarefixed,theanalyticchi-squaresignificancethreshold(inbits)dependsonly onN,thenumberoftrialsavailable(Section2.2).Datapointsareorganizedintoverticallinessincedifferentnumbersoftrialswereproducedbydifferentexperimental protocols(dependingonthedurationandnumberofpresentationsofthemovie);experimentswiththesameNformeachverticalline.Thresholdp=0.001isindicated bydashedline.NumbersinlegendsindicatethefractionofrecordingsforwhichthegammaBLPconveyedsignificantinformation.(B)Informationaboutthevalueofthe differentfeaturescarriedbygammaBLP(mean±SEM,n=55recordingsites).Blacklineindicatesmeanvalueofsignificancethreshold(p=0.95ofaveragebootstrapover 400permutations).(C)FractionofrecordingsforwhichthegammaBLPconveyedsignificantinformation(chi-squareinformationtest;p<0.001)asafunctionofthenumber ofmoviepresentationsusedfortheanalysis.
varyindependentlyfromoneanotherinthemoviesectionused here:thepercentageofrecordingsiteRFswithsignificant infor-mationbetweenfeaturesatp<0.001confidencelevelwas71%,62% and73%forspatialcontrastwithorientation,temporalwithspatial
contrast and orientation with temporal contrast respectively (Fig.7A).Furthermore,thebiascorrectedpopulationaverage val-ues of information (Fig. 7B) were higher than those between features and neural response. This, as explained in previous
0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25 Information (bits) I(SC;O) I(TC;SC) I(O;TC) Threshold F1=SC F2=O F1=O F2=TC F1=TC F2=SC 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 I(F1; F2) (bits) p<0.001 71% 62% 73%
A
B
p=0.001 sig. thresholdFig.7. Mutualinformationbetweenvisualfeatures.(A)InformationbetweenfeaturescomputedforeachRFandvaluesaveragedover10framewindowsusing8feature classes.Significanceofinformationiscomputedwiththechi-squareapproximation,p<0.001.Thresholdp=0.001isindicatedbydashedline.Numbersinitalicsindicatethe fractionofrecordingsforwhichthemutualinformationbetweenthefeatureswassignificant.(B)Mutualinformationbetweenfeatures(mean±SEM,n=55RFs)computed withsameparametersof(A),andPTbiascorrection.Blacklineindicatesmeanofsignificancethreshold(p=0.95ofpopulationaverageofbootstrapover400permutations). Resultsshowthattheconsideredvisualfeaturesofthemoviesectionusedherearesignificantlycorrelated.
R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 13 4 6 8 10 12 14 x 10−3 0 0.05 0.1 0.15 0.2 0.25
Conditional Information (bits)
I(γ BLP; SC|O) I(γ BLP; SC|TC) I(γ BLP; O|SC) I(γ BLP; O|TC) I(γ BLP; TC|SC) I(γ BLP; TC|O) Threshold 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Bootstrap information over 400 permutations
Probability
8 feature classes
4 feature classes 12 feature classes
C
χ2 p<0.001 100% for allA
0 0.05 0.1 0 0.05 0.1 0 0.05 0.1B
I( γ BLP; SC|F1) I( γ; BLP; O|F1) I( γ; BLP; TC|F1) F1 SC O TC F1 SC O TC F1 SC O TC I(γ BLP; O|SC) I(γ BLP; O|TC) I(γ BLP; SC|TC) p=0.001 sig. thresholdFig.8.ConditionalinformationcarriedbygammaBLPaboutvisualfeatures.(A)InformationcarriedateachrecordingsitebygammaBLPaboutthevalueofeachfeature conditionedoneachoftheotherfeatures,computedwithnobiascorrection.Significance(p<0.001)isassessedwiththechi-squareapproximation,withthresholdvalue ofp=0.001indicatedbyblackdashedline.NumbersinlegendreportthepercentageofrecordingsforwhichthegammaBLPconveyedsignificantCMI.(B)Information carriedbygammaBLPaboutthevalueofeachfeatureconditionedoneachoftheotherfeatures(mean±SEM,n=55).InformationiscomputedwithPTbiascorrection.Black lineindicatesmeanofsignificancethreshold(p=0.95ofaveragebootstrap,400permutations).(C)DistributionofbootstrapvaluesofuncorrectedCMIfordifferentpairsof features(coloredlines,400randompermutations)comparedwiththecorrespondingchi-squaredistribution,fordifferentnumbersoffeatureclasses.Resultsareaveraged overallrecordingsfromsessionH06nm6.
sections,impliedthatthesingle-featureinformationanalysiswas notenoughtoestablishifallthreeconsideredfeaturesweretruly encodedbythegammaBLP.
To shed light on this issue, we computed, for each pair of features, the conditional information I(BLP;F2|F1) carried by thegammaBLP aboutfeatureF2 given thevalue of featureF1. Weestimatedthesignificanceofthisvalueusingthechi-square approximation for the CMI described in Section 4. We found that for every pair of features F1 and F2 the CMI was signifi-cant at the p<0.001level for all recording sites (Fig. 8A). We thencomputedpopulationaveragesofPTbiascorrectedCMI val-uesforeachpairoffeatures(Fig.8B).Again,orientation(Fig.8B, middlerow)had weakerconditionalmodulationthantheother features,althoughitremainedstatisticallysignificantwhen con-ditioned on either remaining feature. Interestingly, the values of conditional information were higher in all cases than the correspondingunconditioned single-featureinformation values, indicatingthatcorrelationsbetweenfeaturesinthenaturalmovie actuallyincreased(ratherthandecreased)thejointinformation thatthegammaBLPcarriedaboutthesemultiplefeatures.
Theoriginofthiseffectwasfurtherinvestigatedbycomputing I(F1;F2|BLP=1,2,3,4),i.e.byanalysinghowthemutual informa-tionbetweenthefeaturepairswasrelatedtothedifferentlevelsof LFPgammabandpower.Fig.9showsthatthemutualinformation
betweenfeaturepairsatfixedresponsewassignificantforall lev-elsofgammaBLP(chi-squareapproximationtestonuncorrected information;p<0.001),andwasalsosignificantlyvaryingamong differentlevelsofgammaBLP(ANOVAtest;p<0.001).Thefactthat thedegreeofdependencebetweenfeaturesdiffersacrosssections ofstimulithatevokedifferentresponsesexplainswhycorrelations betweenfeaturesincreaseinformation.Inthiscase,thedegreeof dependencebetweenfeaturescanbeusedtopredicttheneural responseandsoitincreasestheoverallinformationbetween fea-turesandresponses(seePanzerietal.(1999)orPolaetal.(2003)
foraproof).Moreover,theaverageoverallfourlevelsofgamma BLPofI(F1;F2|BLP=1,2,3,4)washigheroverallthanI(F1;F2)for allfeaturepairs,theaveragegainbeing+26%,+26%and+22%for I(SC;O),I(O;TC)and I(SC;TC)respectively.Inthiscase,according toEq.(13)statisticaldependenciesamongfeaturesmustincrease information.
Wethenevaluatedifagivenfeaturecarriedinformationthat wascomplementarytothatcarriedbybothotherfeatures,by cal-culatingI(BLP;F2|F1,F3)using4classesforthefeatures.Wefound thatthiswasalsosignificant(chi-squareapproximation,p<0.001) forallpermutationsoffeatures.Theresultsoftheabove compu-tationsshowthatthegammaBLPconveyedgenuineinformation abouteachsinglefeature,eventhoughtheywerestrongly corre-lated.
14 R.A.A.Inceetal./JournalofNeuroscienceMethodsxxx (2011) xxx–xxx 1 2 3 4 0 0.05 0.1 0.15 0.2 0.25 0.3 I(SC;O|γ BLP) bits
γ BLP level γ BLP level γ BLP level
0 0.05 0.1 0.15 0.2 0.25 0.3 bits 0 0.05 0.1 0.15 0.2 0.25 0.3 bits 1 2 3 4 1 2 3 4 I(O;TC|γBLP)
C
I(SC;TC| γ BLP)B
A
Fig.9.Informationbetweenvisualfeaturesatfixedresponse.MutualinformationbetweenfeaturesconditionedtofourdifferentlevelsofgammaBLP(mean±SEM,n=55). (A)Spatialcontrastandorientation,(B)orientationandtimecontrast,(C)spatialcontrastandtimecontrast.Informationiscomputedforvaluesaveragedover10frame windowswithPTbiascorrectionand8featureclasses.Asterisksindicatethatthemutualinformationissignificantforeachlevelofresponse(chi-squareapproximationtest onuncorrectedinformation,p<0.001).ANOVAtest(p<0.001)showsthattheinformationbetweenfeaturepairschangessignificantlyfordifferentlevelsoftheresponse.
We performed all the significance tests using the analytic chi-squareapproximationontheuncorrectedplugininformation valuesbecausetheyallowfastcomputationofhighlysignificant p-values,andbecausetheconsiderationsofSection3predictedthis testtobehighlyeffectiveunderthesamplingconditionsdescribed here.Toillustratetheaccuracyofthechi-squareapproximation,we comparedthedistributionofthebootstrapvalues(over400 per-mutations)oftheuncorrectedplugininformationestimateswith theanalyticalchi-squaredistributionwiththedegreesoffreedom predictedbytheequationsinSection4.1.AsshowninFig.8Cfor arepresentativesession,withthisdatasettherewasanexcellent agreementbetweenthebootstrapandthechi-squaredistribution. Inparticular,whenchangingthenumberofbinsintowhichthe featureswerediscretized,thebootstrapdistributionfollowedthe changeinthenumberofdegreesoffreedomofthechi-square dis-tributionpredictedinSection3:whenthefeatureswerebinned with4,8and12classestheaveragep-valuesoftheKStestwere 0.49±0.025,0.36±0.025,0.34±0.015respectively(mean±SEM overallrecordingsites).Theseresultsbothhighlightthevalueof theanalyticalapproximationsreportedinourstudyand demon-stratethatthechi-squarecouldbecorrectlyusedinthisdatasetto accuratelycomputesignificancevaluesdowntosmallp-values. 6. Discussion
Wehaveconsideredtheapplicationofmutualinformationas astatisticaltest toestablishdependencybetweenstimulus fea-turesand neuronalresponse variables.While accuratebiasfree informationvaluesarecrucialtoenablemeaningfulquantitative comparisons,inmanyapplicationsitissufficienttoestablishthe presenceofastatisticallysignificantdependence.Wehaveshown that for determining significance with the bootstrap method, uncorrectedplugininformationestimatesshouldbeused,andthat ifsufficient dataare available, the bootstrap procedurecan be bypassedcompletelyandsignificancedeterminedthroughan ana-lyticalformforthedistributionofvaluesunderthenullhypothesis. Thisallowsaccuratecalculationofp-valuesandsignificancewith greatlyreducedcomputationalrequirements.
Thisfinding,particularlytheaccuracyoftheanalyticchi-square approximation,hasimplicationsforanalysisofotherneurological signals,forexamplefunctionalmagneticresonanceimaging(fMRI) data.Anapproachoftentakenwithsuchdataisamass-univariate analysistodeterminedependenciesbetweenvariousstimulation conditionsandthetimecourseofthebloodoxygenlevel depen-dent(BOLD)signalateachvoxel.Theabilitytodirectlycalculate asignificancep-valueforanydependencecheaplyusingthe chi-squareapproximationsuggeststhatthemethodsdiscussedhere
couldbeeasilyappliedtosuchdata,evenwhentheyconsistoftens orevenhundredsofthousandsofvoxels.Theadvantagesofsuch anapproachisthattheyarecompletelynon-parametricanddonot requireanyconvolutionofthestimulustimecoursewitha hemo-dynamicresponsefunction(HRF).Thismeanstheycanbeappliedto situationswheretheHRFisnotwellcharacterizedorisnot guar-anteedtobeuniversalacrossconditionsandareas.Thismayfor instancebethecasewheninvestigatingtherelationshipbetween thefMRIBOLDresponseandthepowerofEEGorLFPresponses indifferentfrequencybands,orwhencollectingfMRIresponsesto complicateddynamictimevaryingstimuliwithmultiplecorrelated featuressuchasthemoviestimuliconsideredhere.Additionally, thegeneralandnon-linearnatureofmutualinformationallowsit tocaptureanytypeofdependency,incontrasttoconventional lin-earanalysis.Similarconsiderationscanbeappliedtotheproblemof estimatingsignificanceofactivationinothertypesofneuroimaging signals,suchasthesourcesreconstructedfromEEGorMEGdata
(Grossetal.,2001).
Whilewehaveshownthattheuseofbiascorrectionshasno beneficialeffectonestablishingthepresenceofastatistically sig-nificantstimulusresponsemodulation,oncethatsignificancehas beendetermined,bias correctedvalues are of coursecrucialto obtainanaccuratequantificationofthestrengthofthemodulation. Thebiascorrectedvaluesallowdirectquantitativecomparisons oftheamountsoftheinformationcarriedbydifferentsystemsor experimentswithdifferentnumbersoftrials,andoftheamounts ofinformationcarriedbyneuralcodescharacterizedbya differ-entresponsedimensionality(suchasforexample,spiketimesand spikecounts).Ourfindingssuggestthatitmaybeworth imple-mentingahybridapproach,whereuncorrectedinformationvalues areused(accordingtotheproceduresdescribedhere)todetermine thestatisticalsignificanceofanydependenceeffect,whilebias cor-rectedvalues arecomputedtoproduceanaccurateestimateof theeffectsize.Inpractice,sincemostbiascorrectiontechniques involvecalculationof theuncorrectedvalues thisapproach car-rieslittlecomputationaloverhead.However,caremustbetakento avoidstatisticalerrorsdueto‘non-independent’analysis(Vuletal., 2009),whicharisewhenreportingmeanvalueswherethemeanis calculatedonlyfromsamplesexceedingsomethreshold.Forthis reason,in theanalysisofrealdatapresented herewereported biascorrectedmeaninformationvaluesoverthewholepopulation ratherthantheinformationvaluesaveragedoverthe subpopula-tionwithsignificantinformation. Theaccuratedeterminationof thestatisticalsignificancecanofcoursebeusedasaselectiontool forothertypesofindependentanalysis,forexamplecalculationof receptivefieldsofsignificantlytunedneurons.