Upload
vuque
View
215
Download
3
Embed Size (px)
Citation preview
1
Incentives,Innovation,andImitation:SocialLearninginaNetworkedGroup
ThomasN.Wisdom
SubmittedtothefacultyoftheUniversityGraduateSchool
inpartialfulfillmentoftherequirements
forthedegree
DoctorofPhilosophy
intheDepartmentsofPsychologyandCognitiveScience,
IndianaUniversity
August2010
2
AcceptedbytheGraduateFaculty,IndianaUniversity,inpartialfulfillmentofthe
requirementsforthedegreeofDoctorofPhilosophy.
DoctoralCommittee
_________________________________________________
RobertL.Goldstone,Ph.D.
_________________________________________________
ElinorOstrom,Ph.D.
_________________________________________________
KevinE.Collins,J.D.
_________________________________________________
JasonGold,Ph.D.
_________________________________________________
EliotR.Smith,Ph.D.
3
Acknowledgements
IwouldverymuchliketothankRobGoldstone,LinOstrom,KevinCollins,EliotSmith,and
JasonGold for invaluable advice and assistance in shaping this research; ToddGureckis,
Michael Roberts, Drew Hendrickson, and everyone in the Percepts & Concepts Lab for
feedbackandadvice;AllenLee,AndyJones,ZoranRilak,XianfengSong,andSarahGrantfor
helping to program and design the experiments; and Frances Kidwell, Itai Hasak, and
BennisPavisianforassistancewithrunningtheexperiments.Thisworkwasfundedinpart
byanNSFIGERTtraineeship.
4
ThomasN.Wisdom
Incentives,Innovation,andImitation:SocialLearninginaNetworkedGroup
Humans’extraordinarytalentsforlearningfromtheirenvironmentsandfromeachother
arethebasisofculturalandtechnologicaldevelopment,butfactorsaffectingtheuseof
theseskillssuchastime,informationdifferences,groupsize,andmaterialincentivesare
notyetcompletelyunderstood.Weusedaseriesoflaboratoryexperimentstoinvestigate
thecauses,consequences,anddynamicsofsociallearningstrategiesemployedbygroupsof
peopleincomplexsearchenvironments,andhowindividualimitationandinnovation
behaviorsaffectresultsatthegrouplevel.Intheseexperiments,participantsplayeda
simplecomputer‐basedpuzzlegamewithothers,inwhichguesseswerecomposedofsets
ofdiscreteunitsthathadbothlinearandinteractiveeffectsonscore,andeachplayercould
viewandimitateentireguessesorpartsofguessesfromothersinthegroup.Players
receivedround‐basedscorefeedbackaboutthequalityoftheirownguesses,andinsome
cases,others'guesses.Ourresultsshowedthatparticipantsusedseveralsociallearning
strategiespreviouslystudiedinotherspecies,aswellasstrategiesstudiedinthecontextof
innovationdiffusion,suchasimitationbiasestowardsolutionssimilartoone’sown,and
towardincreasinglypopularsolutions.Wefoundthattheriskofexploringinalargeand
complexproblemspacecausedparticipantstotakeaconservativeapproach,withsmall
amountsofinnovationandimitationusedtoacquiregoodsolutionsandmakeincremental
changesinthesearchforbetterones.Finally,wefoundthatimitation,ratherthanmerely
beingusedtocopyothersandavoidexploration,wasoftenusedbygroupmembersto
5
improveoneachothers’guesses.Contextualfactorsthatdisruptedordiscouraged
imitationgenerallyresultedinpooreroutcomesfortheentiregroup,becauseofareduced
capacityforparticipantstocreatesuchcumulativeimprovements.Theseresultsare
discussedinthecontextofknowledgeasacommons,withimplicationsforthepromotion
ofinnovationsandintellectualpropertypolicy.
6
TableofContents
1.(A)socialLearning...............................................................................................................................................7
2.Experiment1–PictureGame......................................................................................................................29
3.Experiment2–CreatureGameA...............................................................................................................53
4.ExplorationasaSocialDilemma................................................................................................................89
5.Experiment3–CreatureGameB(scorevisibility).........................................................................101
6.KnowledgeasProperty................................................................................................................................129
7.Experiment4–CreatureGameC(paymentandprotection)......................................................145
8.GeneralDiscussion.........................................................................................................................................188
References...............................................................................................................................................................195
7
1.(A)socialLearning
1.1.Introduction
Penguinsarethoughttobewarm‐bloodedanimalslikeotherbirds,butanoften‐told
talesuggestsotherwise.Asthestorygoes,agroupofpenguinswilloftenhesitatebefore
leavingtheicetoforageforkrill,becausealeopardsealmaybewaitinginthewaterto
forageforthem.Eachpenguinwouldprefertofindkrillassoonaspossibleandnevermeet
aleopardseal,buttheonlywaytobesurethewaterisfineisforatleastoneofthemto
jumpin.Thepenguins'simplesolutiontothisdilemmaistopushoneoftheirfellowsover
theedgeandobserveitforsignsofbeingdevoured(Dawson,1974).
Thankfully,humansocietieshavedevelopedinsuchawaythatgoingouttofind
fooddoesnottypicallyrequirenudgingourneighborsintothejawsofahungrypredator.
Butwedodependheavilyoninformationabouttheexperiencesofothersthatweexpectto
shareinsomeform.Infact,therearefewactivitiesthathumansparticipateinthatdonot
dependinsomewayonknowledgeobtainedfromothers.Thisisevidentuponcasual
reflectionabouthowpeoplegatherinformationandmakechoicesaboutthingslike
restaurantsormovies,candidatesforajoborpoliticaloffice,anewcitytoliveinoralarge
householdpurchase.Such"sociallearning"allowspeopletoobtaininformationabout
availableoptionswithoutundertakingthecostlyprocessofdirectlyevaluatingeachone,
thoughwithouttheincreasedaccuracythatsuchevaluationmightprovide.Thisprompts
thequestionofhowdecisionsaremadebetweentheoptionsoflearningaboutthe
environmentdirectlythroughexperience,orindirectlyviainformationprovidedbyothers.
Learningdirectlyfromtheenvironmentoftencarriesacostforthelearnerinterms
ofriskorresources,sowhenpossible,peoplewilloftenprefertoobtaininformation
8
indirectlythroughotherswithoutindividuallybearingthecostoflearningonone’sown.If
toomanypeopleavoidthecostsofdirectindividuallearningbutcontinuetomakeuseof
theinformationavailablethroughindirectsociallearning,anundesiredstagnationor
reducedadaptationtocomplexityandchangeintheenvironmentmayresult.
Ontheotherhand,whenlearnersareincompetitionforresources,eachlearner
mayhaveanincentivetopreventimitationbyothersoftheircostly,individually‐obtained
information.Adifferentproblemmayresultfromthisavoidanceofinformationsharing,
becauselearnersarepreventedfrombuildinguponandimprovingothers'solutions,which
canleadtorepeated"reinventionofthewheel,"anotherkindofstagnation.Ifanindividual
canpreventotherindividualsfromimitatingtheirsolutions,thenthecollectivebenefitsof
theirinnovationswillbeunderutilized.
Thesetradeoffsbetweenshort‐termself‐interestandgroupinterestcanbetreated
asanexampleofthewell‐studiedclassofphenomenaknownassocialdilemmas.Such
dilemmasapplytouseofnaturalresourcessuchasfisheries(Acheson,Wilson,&Steneck,
1998)aswellasartificialresourcessuchasirrigationsystems(Siy,1982).Alarge
literatureoftheory,fieldresearchandlaboratoryexperimentationhasbeenbuiltaround
thestudyofsocialdilemmas,andtheenvironmentalandinstitutionalfactorsthatcan
contributetotheirsolutions(Ostrom,1990).Recently,themethodsandframeworksused
toanalyzesocialdilemmasofphysicalresourceshavebeenproductivelybroughttobear
onthequestionofknowledgeasaresource(Hess&Ostrom,2007).
Thoughnotalwaysformallytreatedassuch,thesocialdilemmasofsociallearning
areaddressedinthepoliticaleconomyofknowledgeasatradablecommodity,intheform
ofintellectualproperty(IP).Intheory,IPlawandpolicyattempttobalancetheinterestsof
9
knowledgecreatorsandusersinordertofurthertheprogressofknowledge,giventheir
environmentsandincentives.However,thedistinctionbetweenknowledgecreatorsand
usersisnotalwaysclear,becauseknowledgecanbeusedasafoundationorscaffoldingfor
creatingfurtherknowledge.Sorestrictivepropertyrightsgrantedtoincentivizeknowledge
creatorscanactuallyhavetheeffectofimpedingtheprocessofcreation,reducingbenefits
forall.Thusadilemmaarisesintheconflictbetweentheprivateinterestsofindividuals
andthecollectiveinterestsoftheirgroupsintheuseofcurrentknowledgeandthecreation
offutureknowledge.
Thecomplexityofreal‐worldlearninganddecisionmakingmakesitdifficultto
teaseaparttheeffectsoffactorssuchasindividualdifferences,groupinfluence,and
environmentalcharacteristics,aswellaschangesinlearningandinteractionsovertime.
Theseriesofexperimentsreportedinthisdissertationwillattempttobridgetheindividual
andgroupperspectives,byexploringthedynamicinterrelationshipsofactions,incentives,
andinstitutionsinvolvedintheexchangeofideas,andtheirconsequencesforboth
individualsandgroups,throughaseriesofcontrolledlaboratoryexperimentsonhumans.
1.2.(A)sociallearning:Definitionsandbackground
BoydandRicherson(2005)defined"sociallearning"as"theacquisitionofbehavior
byobservationorteachingfromotherconspecifics."Sociallearninghasbeenstudied
extensivelyinhumans(e.g.Hurley&Chater,2005),aswellasnon‐humananimals,
includingforagingchoicesinstarlings(Templeton&Giraldeau,1996),foodpreferencesin
variousrodentspecies(Galef&Giraldeau,2001),andmatechoicesinblackgrouse
(Höglund,Alatalo,Gibson,&Lundberg,1995)."Asociallearning"canbedefinedconversely
10
astheacquisitionofbehaviorbyindividualinnovationorexplorationoftheenvironment,
withoutrecoursetoinformationfromothers.
Laterinthispaper,"asociallearning"and"innovation"willbeusedinterchangeably,
aswill"sociallearning"and"imitation."However,agreatdealofresearchhasgoneinto
attemptingtodefineanddisambiguatethevarietiesoflearningthatorganismsexhibit.
Beforewegoon,wewillsurveysomeofthisresearchandclarifywhichareaswewillbe
exploringindetail,andwhatwewillbeneglectinginordertoachievethisfocus.
1.2.1.Sociallearning/imitation
Galef(1988)reviewedmanystudiesofsociallearninginvariousspeciesinan
attempttoclarifythedefinitionsandexplanationsthathavebeenusedforthese
phenomena.Galefnotedinhisreviewthefutilityofusingtheword"imitation"inageneral
andunambiguousway,becauseoftheoverabundanceofusesanddefinitionsithas
acquiredovertheyears.Galefdiscussesthisabundancebeginninginthemodernerawith
Romanes(1884),Morgan(1900),andThorndike(1911),whosedifferenceswererootedin
differingviewsabouttheevolutionofhumans,andtheevidenceofprecursorsofhuman
abilitiesinotheranimals.
SimilarlytoBoydandRicherson(2005),Box(1984)suggested"sociallearning"asa
genericdescriptiveterm,todistinguishsocially‐influencedlearningfromlearningthatis
notinfluencedbyothers,thoughGalef(1988)notesthatthisdistinctionisnotnecessarily
clear,since"itisalwaystheindividualwholearns;"behaviorsmaybeinitiallyacquired
fromothersbutsubsequentlymaintainedbyindividuallearningoffavorableconsequences.
11
Wedonotintendtoattempttoresolvethisambiguity;itmustremaininthebackgroundto
helpexplainlearningbehaviorsinanyparticularcontextasamixtureofinfluences.
Itisimportanttodistinguishbetweentwomajordescriptivetermsforsocial
influence:(1)"socialenhancement,"inwhichothersinfluencetheperformanceof
behaviorsalreadyestablishedinanindividual'srepertoire(Galef,1988);and(2)"social
transmission,"inwhichsocialinteractionincreasesthelikelihoodthatanindividualwill
cometoindependentlyperformabehaviorinitiallyintherepertoireofanother,suchthat
thereisan"increasedhomogeneityofbehaviorofinteractantsthatextendsbeyondthe
periodoftheirinteraction"(Galef,1976).Forexample,inthemid‐20thcenturywhenitwas
commonformilktobedeliveredandleftoutsideofhomes,itwasnoticedthatincreasing
numbersofbirdswerelearningtoopenmilkbottlestodrinkthecreamfromthetop
(Fisher&Hinde,1949).Thiswasfoundtobeacaseofsocialenhancementratherthan
socialtransmission:laterexperimentsshowedthatforthepurposesoflearninghowto
opencontainers,allowingnaïveindividualstofeedfromcontainersopenedbyotherswas
justaseffectiveasallowingthemtoobserveaconspecificopeningacontainer.So
conspecifics'actvitiesfocusedothers'subsequentforagingonsimilarcontainers,whose
topswerevulnerabletoexistingfeedingbehaviorssuchaspecking(Sherry&Galef,1984,
1990).
Thephenomenathatthispaperwillexaminegenerallyfallunderthedescriptionof
"socialtransmission"ratherthan"socialenhancement;"weareinterestedinthespreadof
locallynovelinformationviapeerobservation,ratherthantheinfluenceofsocial
interactiononuseofinformationpossessedbyallgroupmembers.
12
Alargevarietyofexplanatoryterminologyhasdevelopedtorefertotheorized
behavioralmechanismsthatcouldunderlietheabovedescriptiveterms.Thiscollectionof
terminology,accordingtoGalef(1988),is"extensive,contradictory,andvague",and
thoughithasnotprovedadequateforexhaustivelyclassifyingoreffectivelyexplaining
sociallearningbehavior,ithasbeenhelpfulinsuggestingapproachesforexperimental
analysisofthemyriadwaysinwhichsocialinteractioninfluencesbehavior.Followingare
severalofthemorewidely‐usedtermsinthisarea.
"Stimulusenhancement"(Spence,1937)and"localenhancement"(Thorpe,1963)
refertosituationsinwhichobservationofademonstrator'sactions,orevidencethereofin
theenvironment,causetheobservertodirectagreaterproportionofitsbehaviortoward
thelocationorobjectofthedemonstrator’sactivity."Socialfacilitation"referstosituations
inwhichthepresenceofothers"energizesallresponsesmadesalientbythestimulus
situationconfrontingtheindividualatthemoment"(Zajonc,1965)."Contagiousbehavior"
isusedforsituationsinwhich"theperformanceofamoreorlessinstinctivepatternof
behaviorbyonewilltendtoactasareleaserforthesamebehaviorinothersandsoinitiate
thesamelineofactioninthewholegroup,"suchasyawninginhumans(Thorpe,1963).
"Copying"referstosituationsinwhichanobserverissensitivetothedegreeofsimilarity
betweenitsownbehavioranditsmodel's,anditsresponsesarereinforcedpositivelywith
greatersimilarityandnegativelywithdissimilarity(Miller&Dollard,1941;Thorndike,
1911).
"True"or"reflective"imitation,alsoknownas"observationallearning,"requires
thattheperformanceofanactissufficientlyinstigatedsimplybyobservingit,andinvolves
purposeful,goal‐directedcopying(Galef,1988).Bandura(1965)calledhumans'raretalent
13
forsuchimitation"no‐triallearning,"becauseitisevenfasterthantheone‐triallearning
observedinanimalswithastrongbuilt‐intendencytoformcertainassociations(e.g.
betweenthetasteofafoodandasubsequentstomachache).Evidencefor"true"imitation
inotheranimalshassofarbeenunconvincing(Davis,1973;Hall,1963)becauseofthe
difficultyofcontrollingforallofthealternativesociallearningprocesses(Galef,1988;
Heyes,1993).
Thisdissertationisconcernedmainlywithphenomenathatfallunderthe
descriptionof"socialtransmission,"sooftheaboveexplanatoryterms,wewillbelooking
mainlyto"true"imitation.Theexperimentsdescribedhereinwillbedesignedsuchthat
participantsaregivensufficientinformation,ability,andopportunitytopurposefully
observeandimitateothers.Theseaspectswillbeexplainedinmoredetailinlatersections.
1.2.2.Asociallearning/innovation
Asmentionedabove,ouruseoftheterm"asociallearning"isessentiallydefined
negativelyastheprocessofobtainingorgeneratinginformationfromsourcesotherthan
one'sfellowlearners.Theterm"innovation"isusedasashorthandoutoftheneedfora
conciseandfamiliardescriptorfortheimpliedindividualintroductionofinformationthat
islocallynovel(givenaverylargespaceofoptionstoexplore)amongagroupoflearners,
asopposedtothetransmissionofexistinginformationbetweenlearners.Using
"innovation"inthiswayinvolvessomeriskofconfusion,however,becausethistermisalso
usedheavilyinthebroaderstudyofcreativityanditsinstantiationinscientific,industrial,
andartisticpractice.(SeeSternberg(1999),KouniosandBeeman(2009),Paletzand
14
Schunn(2010),Dunbar(1999),andGarciaandCalantone(2002)forhintsastothebreadth
ofthisfield.)
Intheinterestofbrevity,thecomplexityandspecificityofanalysisinthisliterature
willbeomitted.Wewilltrytoretainsomeclaritybylimitingourdefinitionofinnovationto
theuseoflocallynovelsolutionelementsbyindividualsamplingfromtheenvironmentina
problem‐solvingcontext,asopposedtoobtaininglocallyknownsolutionelementsthrough
observationofothers.
1.3.Learningstrategiesandtradeoffs
GabrielTarde,oneoftheforefathersofsocialpsychology,consideredinnovation
andimitationtobe"thefundamentalsocialacts"(Tarde1903/1969).Innovation
(generatinglocallynovelapproachestosolvingproblems)producesadiversityof
solutions,thoughoftenatacostinresources(suchastimeorenergy)orrisk(ingivingup
theopportunitytoexploitmorereliablesolutionsinfavorofsomethingneworunknown).
Imitationallowsadecisionmakertoemploysolutionsdiscoveredandpassedonbyothers
withouthavingtodevelopthemindependentlyusingcostlytrial‐and‐errorlearning.When
adecisionmakeracquiresinformationfromothersaboutpossiblesolutionstoaproblem,
theresourcesnotexpendedininformationgatheringcanbeusedforotheraspectsof
problemsolving,thusimprovingperformanceoverall.
Ofcourse,dominationofacommunitybyeitherinnovationorimitationcanbe
problematic.Excessiveinnovationisunhelpfulbecausegoodideasarenotpropagatedand
extendedbyothers.Excessiveimitationismaladaptivebecausegoodbutsuboptimal
solutionsarespreadtotheexclusionofbetteralternativesthatareleftunexplored.A
15
commonassumptionisthatthecollectivecostsofexcessiveimitationwithinagroupare
particularlydetrimental–therearetypicallystrongincentivesforindividualstopursue
largelyimitativestrategiestoavoidrisk,whichmayleadtoadearthofgoodnewsolutions
thatwouldbenefiteveryoneinthegroup.Sothegroupmaybebetteroffifimitationisnot
universallychosen,buteachindividualisbetteroffbychoosingimitation.Thisideahas
strongparallelsinthelargebodyofresearchonsocialdilemmas,inwhichindividuals’self‐
interestsareinconflictwiththesharedinterestsofthemselvesandothersfacingthesame
problem.Researchonsocialdilemmashasconsistentlyshownthatwhensuchinteractions
arerepeatedovertime,andparticipantsareabletorecognizethedilemmatheysharewith
others,theyareoftenabletoadapttheirbehaviortoameliorateoravoidsocialdilemmas
(Ostrom,1990).
Laland(2004)reviewedstudiesofanimal(includinghuman)sociallearning
strategies,andtheirinteractionswithasociallearninginvariousfoodforagingbehaviors,
aswellasmatechoice,foodpreferences,andotherimportantdecisions.Ingeneral,itwas
foundthatthedefaultstrategy(thatis,thepreferredoptionwheneveravailable)isthe
simplestpossiblebehavior:exploitationofanexistingsolutionorknownresource
continuesuntilitisnolongerproductive,atwhichpointeitherinnovationorimitationcan
beusedtoacquireanewsolution.Mostofthestrategiesstudiedformakingsuchdecisions
canbedividedinto“when”and“who”strategies;thatis,heuristicsthatdictateunderwhat
circumstancessociallearningisfavoredoverasociallearning,and,whenitisfavored,those
thatdeterminefromwhichothersanindividualwilllearn.
16
1.3.1.Whentoimitate:Costlyinformation
Decidingwhentousesociallearningoftendependsonarelativeassessmentof
asociallearningasanalternative.Often,thereisgreaterrelianceonsociallearningwhen
acquiringinformationthroughasociallearningiscostly(Boyd&Richerson,1985).These
costscanbeintheformoftheresourcesrequiredtocarryoutasociallearning,theriskof
acquiringinformationthatwillnotbereliableinthefuture,andtheriskassociatedwith
learningabouthazardsdirectly(e.g.predators).Asmightbeexpected,sociallearningabout
predatorshasevolvedinawidevarietyofanimals(Mineka&Cook,1988).Returningtothe
AdeliepenguinsmentionedatthebeginningofChapter1,itisdifficulttoprovethatthey
intentionallypushoneanotherintothewatertotestforleopardseals(ratherthan
accidentallyjostlingoneanotherofftheedge),buttheydoavoidthewatertemporarily
whenoneoftheirfellowsiseaten(Dawson,1974).Ithasalsobeenshownthattheyare
sensitivetotheeffectsofpredatorsongroupsize:theyarelesslikelytoenterthewaterifa
groupreturningfromaforagingrunisunusuallysmall,indicatingthatithasbeenattacked
anddispersedbyaleopardseal(Müller‐Schwarze,1984).Perhapsmoresurprisingisthe
factthatpredatoravoidancelearninghasevolvedinsomecasesbetweendifferentspecies
vulnerabletothesamepredators(Mathis,Chivers,&Smith,1996).Thisstrategyis
beneficialbecauseitallowsminimalindividualcontactwithhazards,thoughifone
individuallearnsthe“wrong”hazards,imitationbyothersmaybeinefficientordangerous.
Arelatedstrategyistorelyonsociallearningwhenataskisdifficulttolearn
asocially.Suchastrategyhasbeenobservedintheprocessingoffoodswithphysicaland
chemicaldefensesbygorillas(Byrne&Russon,1998)andinanexperimentalvisual
identificationtaskinhumans(Baron,Vandello,&Brunsman,1996).Again,thisstrategycan
17
savevaluableriskandeffortforimitators,butiftheexamplestheylearnfromareflawed,
inefficientorharmfulbehaviorscanspreadinagroup.
1.3.2.Whentoimitate:Resolvinguncertainty
Another"when"strategyistocopyotherswhenoneisuncertainabouttherelative
valueofavailableoptions.Forinstance,BeckandGalef(1989)showedthatwhen
presentedwithtwonovelfoods,ratsshowedabiasfortheonethathadbeenconsumedby
conspecifics,asevidencedbytheodorontheirbreath;whenthefoodswerefamiliar,this
biaswasmuchweaker.Suchstrategiesareespeciallyusefulwhen,asabove,informationis
costlytogatherasocially–samplingfoodswithoutregardtoothers’decisionscanresultin
insufficientnutritionorevenpoisoning.
Anearlystudyofsocialinfluenceinhumansfoundthatinasimpleperceptualtask,
thelesscertainanindividualwasofhisownjudgment,themorelikelyhewastobe
susceptibletotheinfluenceofothers(Deutsch&Gerard,1955).Likewise,whenan
individualisuncertainabouttheappropriateresponseinaparticularsituation(becauseof
eitheralackofinformationorthefailureofpreviousresponses),itismorelikelythatthe
individualwillresorttoimitation(Thelen,Dollinger,&Kirkland,1979).Whilerandomly
tryingoutsocialbehaviorsisnotusuallypoisonous,socialmistakes(violationsofnorms)
canresultinostracism,whichhasbeenshowntocauseneurologicalresponsessimilarto
thoseofphysicalpain(Eisenberger&Lieberman,2004).Thisisthoughttobeduetothe
evolutionarynecessityofgroupinclusionforsurvival(MacDonald&Leary,2005).
However,ifnormsareharmful,theircostsmayoutweighthebenefitsofinclusioninthe
group.
18
1.3.3.Benefitsofimitationdependonselectivity
Itmaybethatonceadecisionismadetoimitateothersviaoneofthestrategies
above,individualsdonotmindwhotheycopy,aslongastheyavoidthecostsofasocial
learning.However,asimulationbyRogers(1988)ofagentsinanonstationary
environmentshowedthatifavoidanceoflearningcostsistheonlybenefitofimitation,then
theadditionofagentswhopersistentlyimitate(andchoosetargetsofimitationrandomly)
toapopulationofindividuallearnersdoesnotimprovetheaveragefitnessofthe
population.Thisisbecauseovertime,thecostsavoidedthroughimitationwillbebalanced
bycostsresultingfromtheuseofinformationthatisnolongeraccurate.Thusat
equilibrium,amixedpopulationofsocialandasociallearnershasthesameaveragefitness
asanasocial‐onlypopulation.(Asimilarconclusionwasreachedinatheoreticalanalysisof
foragingbehaviorbyGiraldeau,Valone,&Templeton(2002).)
BoydandRicherson(1995)andKamedaandNakanishi(2002)confirmedthese
results,butalsoextendedthemtoshowthatwhensociallearnerscanimitateselectively
(e.g.imitatingwhenindividualexplorationisrelativelyunreliableandthusmorecostly),
theoverallfitnessofthepopulationcanincrease,becausebothindividualandsocial
learningcanbecomemoreaccurate.
1.3.4.Whotoimitate:Frequency‐biasedstrategies
Frequency‐dependentimitationtakesintoaccounttherelativeprevalenceof
behaviorswhendecidingwhichtoimitate.Thisoftenentailsconformitytothebehavior
commontothegreatestnumberofobservedmodels.Suchbehaviorhasbeenobservedin
19
manyspecies,includingguppies(Lachlan,Crooks,&Laland1998),rats(Beck&Galef,
1989),andhumans(Hurley&Chater,2005).Granovetter's(1978)theoryofinnovation
diffusionisbasedona"threshold"proportionofgroupmembersnecessaryforan
individualtoadoptaninnovation.Suchconformitycanresultfromagoalofobtaining
informationabouttheutilityofasolutionfromitsapparentpopularity(informational
conformity),oravoidingtheappearanceofdeviancefromagroupanditsnorms
(normativeconformity).
1.3.5.Whotoimitate:Payoff‐biasedstrategies
Manyselectivestrategiesinvolveobservingthesuccessofmodelsinorderto
determinewhichtoimitate.Such"payoff‐biased"strategiescanuseanyofseveralcriteria,
dependingontheinformationaboutothers'successthatisavailable.Anindividualcan
simplyimitateany“successful”individual(whichrequiresonlysomeminimalthresholdor
criteriaforsuccess),ashasbeenshowninunsuccessfulbats'tendencytofollowpreviously
successfulbatstofeedingsites(Wilkinson,1992).Addingabitofcomplexity,onemight
comparethepayoffsofallindividualsandcopythemostsuccessful,asassumedinmany
theoreticalmodelsofhumandecisionmaking(Schlag,1998).Anindividualmayalso
imitateanydemonstratormoresuccessfulthanitself(whichrequiresevaluationof,and
comparisonbetween,theindividual'sownperformanceandthedemonstrator's).
Whilesuchstrategiesmayallowanimitatorgreatercertaintyinthepayoffofa
particularchoice,theymayalsorequiregreateffortininformation‐gatheringand
calculation,andtheserequirementsmaygrowexponentiallywiththesizeofasocialgroup
andtheirresponsivenesstoeachother.Thefeasibilityofsuchstrategiesisquestionedin
20
theoriesofboundedrationality,whichposit"fastandfrugal"heuristicsthatallowdecision
makerstoprosperwithlimitedinformationandcomputationalresources(Gigerenzer&
Goldstein,1996).Inpractice,suchlimitationsmaysubstantiallychangethecost‐benefit
calculationforcomplicatedstrategies,andleaddecisionmakerstofallbackonsimpler
strategiesorfacereducedperformanceinthefaceofexcessinformationprocessing
requirements.HeiseandMiller(1951)showedthatinadifficultcommunicationtask,the
removalofsomeparticipantcommunicationcapacityledtoincreasedperformance.The
broaderphenomenonofreducedperformanceinthepresenceoflargeamountsof
informationhasbecomeknownas"cognitiveoverload"(Sweller,1988).
Thoughmanymodelssuggestthatcopyingsuccessfulindividualscanbeaneffective
strategy,thebehaviorresponsibleforsuccessmaynotbeapparent,andthusunproductive
ormaladaptivebehaviorsmayaccompanysuccessfulones(Boyd&Richerson,1985).Even
ifaccurateinformationaboutagoodsolutionisavailable,itsuseorexploitationbytoo
manyusersmayreducethebenefitsavailabletoeach.Thussimplepayoff‐biasedstrategies
mayhavecostsintheformofcrowdingandoveruse,whicharenotremediableinthelong
runwithoutmorecomplexcontingentstrategiesthatrecognizetheeffectsthatotherusers
haveonthepayoffofsomesolutions.
1.4.Conformityandculture
Researchonhumansociallearningisfairlywidelydispersedacrossavarietyoffocal
behaviorsandmethodologies,whichisunderstandablegiventhemyriadpossiblefunctions
forsocially‐mediatedinformationinsuchasociallycomplexorganism.Belowisafocused
summaryofsomeofthisresearch.
21
1.4.1.Innovationdiffusion
Inhisstudiesofthediffusionofinnovationsacrosspopulations,Rogers(2003)
positsseveralfeaturesofaninnovationthatcanaffectitsadoption,amongthemthe
relativeadvantageobtainedbyreplacingaprevioussolutionwiththeadoptedinnovation,
andthecompatibilityoftheinnovationwithprevioussolutions.Theformerisaclear
analogtothesimplepayoff‐biasedimitationstrategiesmentionedabove,butthelatter
dealswithmorecomplexcostsofadoption.“Backwardcompatibility”withaprevious
solutionmaybedesiredtominimizethecostsofadaptationbytheadopter,suchas
learninghowtouseit,orreplacingitemsusedcomplementarilywithit.Compatibilitymay
alsobedesiredtomaintaincontinuedinteractionswithotherswhohavenotadoptedthe
innovationyet.Thisminimizessocialcosts,andmayallowthepreservationof"network
effects,"benefitsofaninnovationwhichscaleaccordingtothenumberofusers(Katz&
Shapiro,1985).Examplesofbackwardcompatibilityininnovationincludetheabilityto
playgamesfromanoldervideogamesystemwithanewerone,ortheabilityofusersof
newertechnologiessuchascellulartelephonesandvoice‐over‐IPtoconnecttousersof
land‐linetelephones.Whenthesecompatibilitiesarenotavailable,adoptionofnew
technologieswillpresumablyproceedmoreslowly.
1.4.2.Conformityasanadaptivebias
Imitationinhumansocialactivitiesbringstomind"conformity,"whichhasnegative
connotationsinbothcommonusageandinmuchofthesocialpsychologyliterature.Some
ofthesenegativeconnotationscanbetracedbacktoAsch's(1951)classicstudiesof
22
conformityinvolvedmeasuringtheeffectofagroupofunanimousconfederatesonthe
visualjudgmentsofnaïveindividualparticipants.However,theearlyemphasisonthe
prevalenceofundesirablebehaviorsintheresultsofexperimentssuchasAsch'shasbeen
moderatedinrecentre‐evaluationsofthedata(Hodges,2004),andcallsforamore
balancedexaminationofthecomplexityoftheenvironmentalmoderatorsandmotivesof
biaseslikeconformity(Krueger&Funder,2004).Forinstance,arecentmeta‐analysisof
125studiesusingtaskssimilartoAsch'sexperimentshowedthattheeffectofthesizeof
theconfederatemajoritygroupdependedonwhetherresponsesweregivenpubliclyand
face‐to‐faceorprivatelyandindirectly,whichisthoughttodeterminethetypeof
conformityprocess(normativeorinformational,respectively)thatdominatesthesituation
(Bond,2005).
Modelinghasshownthatthetendencytoimitateothersratherthaninnovate
(“conformitybias”)canbeadaptiveinuncertainenvironments(Boyd&Richerson,1985;
Kameda&Nakanishi,2002).Similarly,socially‐influenced"informationalherding"
behaviorhasbeenproposedasanexplanationforphenomenasuchasfadsandfinancial
panics(Bikhchandani,Hirshleifer,&Welch,1992;Eguíluz&Zimmerman,2000),butithas
beenarguedthatsuchbehavioristheresultofotherwiseadaptivelyrationalBayesian
reasoninginuncertainconditions(Anderson&Holt,1997;Banerjee,1992;Bikhchandani
etal.,1992).
KamedaandTindale(2006)arguethatconformitybiashasevolvedinhumansand
severalothertaxabecauseofitstendencytopromoteanetaveragebenefitforindividuals
ratherthanerror‐freeperformance.Culturalconventionsarethoughttobeaformoflarge‐
scaleconformitytobehaviorsthatevolvealongwiththeirassociatedpopulations,subject
23
toaccompanyingadaptivepressures(Boyd&Richerson,2005).Oftenthegroup‐levelor
population‐leveloutcomesassociatedwithdecisionsthatareindividuallyrationalarenot
immediatelyapparentorpredictable.Muchlikethesociallearningstrategiesstudiedby
Laland(2004),manycomplexculturalmechanismshaveevolvedtodealwiththisdilemma.
Whentheinteractionofagroupofindividualsinacertainenvironmentisfairlystableover
time,traditionscandevelopwhichdictateexpectationsfor"sensible"behaviorinthat
environment,anddiscourageinnovation.Traditionsmayberelatedtointeractionswiththe
environmentsuchashuntingorfarming,orinteractionswithotherindividualssuchas
marriagecustomsordominancehierarchies.Suchtraditionscanbereinforcedthrough
punishmentofthosewhoviolatethem,rewardsforthosewhofollowthem,orboth.
Traditionscanalsobestabilizedthroughsimpleconformity,thetendencytoimitate
behaviorsthatarecommoninagroup.Whenmembersofagroupimitate,adaptive
solutionstoproblemscanbeeffectivelypreservedwithinaculture.
1.5.Studyinginnovationandimitationinalaboratorysetting
1.5.1.Previouswork
Howcanweprojecttheinfinitespaceofhumaningenuityandsocialcomplexity
ontothecircumscribedstageofacontrolledexperiment?Asidefromflawsinthe
interpretationofAsch's(1951)studiesofconfederateinfluenceonthejudgmentsofnaive
individuals(Krueger&Funder,2004;Friend,Raferty,&Bramel,1990),alargerproblem
withthiskindofexperimentisinitstreatmentofconformityasastaticinfluencerather
thanpartofadynamic,interactiveprocess.Changesincultureandtechnologyinvolvelarge
numbersofagentsinteractingoverlongtimescales,andthegenerationandexchangeof
24
informationbetweenpeoplecannotrealisticallybetreatedasaseriesofisolatedevents.
Whileignoringsociallymediatedinformationinalaboratorysettingmayyieldaclearstudy
ofimportantcognitiveprocessesofindividualproblemsolvingandchoice,inmany
situationsitrepresentsanunrealisticconstraint,andmaylimittherelevanceor
applicabilityoftheconclusions.Asch,andthevastmajorityofthesubsequentresearch
inspiredbyhisstudiesofconformity,madethemethodologicalchoiceofcreatingjudgment
situationswithasinglehumansubjectsurroundedbyaccomplicesoftheexperimenter
whowerescriptedtogivespecificjudgments.Thismethodisjustifiedfromtheperspective
ofcreatingawell‐controlledexperimentalenvironmentforexploringfactorsaffecting
isolatedindividualchoicestoimitate.However,thecostsofthismethodareroughly
twofold.First,constrainingthejudgmentsofallbutoneparticipantinagroupmeansthat
mutualinfluencesininnovationandimitationcannotberevealed.Second,limitingthe
opportunitiesforresponsetoasingletrialforeachjudgmentmeansthatchangesinthese
mutualinfluencesaredifficulttostudy.Theimpactofindividualinnovationandimitation
choicesonthegroup’sperformancecanbestberevealedbyallowingallparticipantsina
decisionmakingtasktobenaturally,spontaneously,andrepeatedlyinfluencedbyone
another.Understandingthegroupdynamicsofimitationandinnovationisoneofthemain
goalsofourstudy,andsoweallowallgroupmembers’theopportunitytoinfluence,andbe
influencedby,eachotherovertime.
Theextensiveliteratureon“groupdynamics”(e.g.Hare,1976;Forsyth,2006)
providesinsightintomanyinterestingprocessesofinteractioningroups,butmuchofitis
focusedoninterpersonalvariablesandspecificsettingssuchasclassroomdiscussion,jury
deliberation,orworkplacecollaboration.Whilethesearecertainlyfascinatingand
25
worthwhiletopics,theyoftenprioritizerealismandapplicationtothesesettingsatthe
expenseofcontrolandgeneralizability,andthusforourpurposestheyremainanindirect
influence.Amoredirectinspirationisderivedfromstudiesofthe“emergence”ofcollective
behaviorfromrelativelysimpleinteractionofindividuals(Holland,1999).Mason,Jones,
andGoldstone’s(2008)studyofgroupexplorationofaone‐dimensionalspaceservesasa
foundationforourstudyinamorecomplexdomain,andwewillwilluseitforcomparison
toourresultslater.
1.5.2.Experimentaldesignandgoals
Conceptually,wecandistinguishbetweentheoptions(behaviors,tools,foods,etc.)
availabletobeexplored,andthelearningstrategies(individualorsociallearningofvarious
types)usedtoexploreandexploitthoseoptions.Wecanalsothinkofchangesinsuch
strategiesovertimeasaformoflearningatahigherlevel.Inthecontextofanexperiment,
wewouldliketobeabletoclearlydesignandmanipulatetheoptionsandstrategies
availabletoparticipants,withoutexcessivelyconstrainingbehavior.
Manylearningtaskscanbeseenascombinatorialpuzzles‐‐choosingacombination
ofmoreorlesswell‐knownpartsorsubtaskstocreateamorecomplexbehavior.Such
combinatorialproblemsareanalogoustoexplorationofalargespaceofpossible
combinations.Ifparticipantshavetheabilitytosampleoptionsfromtheenvironment
throughasociallearningstrategies,oradoptthemfromothersthroughsociallearning
strategies,wecanobserveseveralimportantphenomena:varyingcombinationsof
elementsinparticipants'tentativesolutions,varyingmixturesoflearningstrategiesin
participants'choices,andvaryingproportionsofparticipantstrategiesingroups
26
attemptingtosolveproblemssimultaneously.Thesevariationscanbeobservedwithin
groupsovertime,orbetweengroups(e.g.thosewithdifferentnumbersofmembers,or
havingaccesstodifferentamountsorkindsofinformation).
Withthisinmind,weconstructedtwoexperimentalparadigmswithrecognizable
parallelsinreal‐lifeexploratoryandlearningactivities,inwhichparticipantscansample
andsharecombinationsofelementsfromalarge(butfinite,well‐defined,andstably
related)set,suchthatallchoicescanbeevaluatedaccordingtoanobjectivefunction,while
allowingforalargedegreeofflexibilityandvariationinchoicesandlearningstrategies
overtimeandacrossindividualsandgroups.
Incontrasttoseveralpreviousstudies(e.g.Rogers,1988;Boyd&Richerson,1995;
Kameda&Nakanishi,2002),weusedataskenvironmentthatdoesnotchangeovertime,
andenabledprovisionofregular,reliablescorefeedbackaboutparticipants'ownsolutions,
aswellasthoseoftheirpeers.Thesefeaturescouldbeexpectedtodiscourageexploration
oftheproblemspace,inthatthereisnopenaltyforperpetualexploitationofexisting
solutionsorrepeatedimitationofothers'solutions.Inaddition,theproblemspacewas
designedtobelargerandmorecomplexthaninthosemodels,inordertoprovidegreater
realismintheformofalargernumberofoptionsforexploration,aswellassubstantial
uncertaintyaboutoptimalsolutionsandstrategies.
However,byimplementingtheabovechangesandexplicitlyinstructingparticipants
tofocusonagoalofmaximizationofcumulativeperformance,wehopedtosimplifythe
taskwhilemimickingcharacteristicsofreal‐worldproblem‐solvingenvironments.Inour
task,solutioninformationcannotbecomeoutdatedinthesenseofinaccuracy,butits
performancecanbecomerelativelylesssatisfactorygivencontinuingopportunitiestofind
27
bettersolutions.Thisisanalogoustopracticaltechnologicalprogress,inthatolder
technologiesoftenstillfunctionadequatelyfortheirdesignedpurpose,butnewer
technologiescanprovideimprovementssuchasextrafeaturesormorereliable
performance.Ifthegoalistomaximizecumulativeperformance,someriskmaybe
rationallyjustifiedtoenablegreaterfuturerewards.
Thisapproachisofcoursenotideallysuitedforansweringallquestionsrelatingto
humansociallearning,butitallowsustoapproachavarietyofissueswhichmight
otherwisebetooabstract,varied,andsubtletoreasonablyapproachinacontrolled
laboratoryexperiment,suchastheroleofimitationinconformity,theeffectofindividual
differencesinrisktoleranceorincentiveresponsivenessinthediscoveryofinnovations,
andthecharacteristicsoflearningasacomplex,dynamic,socialendeavor.
1.5.3.Generalpredictions
Basedonthesociallearningstrategiesandtradeoffsdiscussedearlier,wecan
discusssomepreliminarypredictionsaboutparticipants'behaviorandperformance.First,
giventhateveryparticipantwassubjecttothesamechoicepossibilitiesandpayoff
structure,andthepayoffstructurewasfixedforagivengame,weexpectedthatimitation
wouldbemorecommonthaninnovation,becausetheformeroffersimmediaterisk‐free
returns,whilethelatteroffersmoreriskyandvariablereturns.Imitationwouldbebiased
primarilytowardhigher‐scoringsolutions(accordingtopayoff‐biasedstrategies),and
secondarilytowardrelativelysimilarsolutionstotheimitator'sown(duetocompatibility
strategies).Asforoutcomes,anintuitivepredictionbasedontheabovebehavioral
expectationsisthatthehighratesofimitationandsmallamountsofinnovationwouldlead
28
tostagnantgrouplearningprocesses,andalackofsubstantialimprovementsin
performanceovertime.Thispredictionisinlinewithanegativeviewofimitationasstatic,
idleinformationalconformity,butatoddswiththepotentialforgroupstosolvecollective
actionproblemsgivensufficientinformationandincentives(Ostrom,1990).These
predictionswillberevisitedandelaborateduponlaterinthisdissertation.
29
2.Experiment1–PictureGame
2.1.Experiment1overview
Inordertoexplorethedynamicsofinnovationandimitationingroups,andto
attempttoarbitratebetweenthecompetingpredictionsregardingindividualincentives
andgroupconsequences,wedesignedanexperimentinwhichnetworkedgroupsofpeople
exploredahighdimensionalproblemspaceoveraseriesoftime‐limitedrounds.
Participantsreceivedscorefeedbackabouttheirguessesaftereachround,andpassively
sharedinformationabouttheirguessesandscoreswithothers.Participantscouldobserve
andcopythemostrecentguessesoftheirneighborsastheyformedtheirownguesses.
Thedifficultyoftheproblemwasmanipulatedacrossconditionsbychangingthe
sizeoftheproblemspace.Arangeofgroupsizeswastestedtoinvestigatetheeffectsofthe
amountofsocialinformationonindividualperformanceandstrategies.Thein‐game
interactionbetweenplayerswaslimitedtothesimplepassiveexchangeofguess
informationonly,soastoallowexaminationofimitationandinnovationbehavior
unencumberedbymorecomplexsocialinteractions,suchasdirectcommunicationamong
players.
2.1.1.Predictions
Performancewaspredictedtoimproveoverroundswithineachgame,andacross
gameswithinanexperimentsession,duetoscorefeedbackandsimplelearningeffects.
Conversely,solutionturnover(theamountofchangeinasolutionbetweenrounds),and
thediversityofsolutionswithinagroup,wereexpectedtodecreaseovertimeas
30
participantsfoundmorepartsofthecorrectsolution,narrowedtheirsearchtosmaller
areasoftheboard,andrefinedtheirsearchstrategies.
Imitationwasexpectedtobemorecommonofpeerguesseswithhigherscores,
becauseutilitiesofguessesareexplicitandthusimitatorscanmaximizetheirexpected
utilitiesbychoosingthehighestscoringguesses.Imitationwasalsoexpectedtobefocused
onguessesthatweremoresimilartotheimitator'sownpreviousguess,asnotedabove.
Participantswhoimitatedmostoftenwereexpectedtoobtainbetterscoreperformance
(relativetothosewhoimitatedlessoften),becausetheywouldpresumablyimitategood
solutionswhiletakingfewerrisks.Theaveragescoresofparticipantsinlargegroups
(greaterthan4participants)wereexpectedtobegreaterthanthoseinsmallgroups,
becausethelargergroupswouldproduceagreaterdiversityofsolutions,andthusbeable
tosearchtheproblemspacemoreefficiently,contingentontheabilityofparticipantsto
makeuseofthisincreaseddiversitythroughselectiveimitation.
2.2.Experiment1Methods
145participantswererecruitedfromtheIndianaUniversityPsychologicaland
BrainScienceDepartmentundergraduatesubjectpool,andweregivencoursecreditfor
takingpartinthestudy.Participantspopulatedeachsessionbysigningupatwillfor
scheduledexperimentswithamaximumcapacityof9persons,andweredistributedacross
39sessionsasshowninTable2.1.
31
Table2.1:DistributionofParticipantsAcrossGroupSizes
Groupsize 1 2 3 4 5 6 7 8 9
#Sessions 8 6 9 3 3 5 1 2 2
#Participants 8 12 27 12 15 30 7 16 18
Weimplementedtheexperimentusingcustomsoftwareruninawebbrowser,and
eachparticipantusedamousetointeractwiththeexperimentalgame.Allparticipants’
computerscommunicatedwithagameserver,whichrecordeddataandupdatedscores
andteaminformationforparticipantsattheendofeachround.Thetaskwasaround‐
basedpicture‐matchingpuzzlegamewithscorefeedbackgivenaftereachround.Thegoal
picturethatparticipantsattemptedtomatchwasarandomlygeneratedsplinequantizedto
agridofsquarepixels.Thesquaresmakingupthesplinewerecoloredblack,andthe
remainingsquarescoloredwhite(seeFig.1forexamples).
Figure2.1:ExamplesofrandomlygeneratedgoalpicturesinExperiment1.
Theparticipant’sgameboardwasagridofthesamedimensionsasthegoalpicture,
witheachsquareinitiallycoloredwhite.Thecolorofeachsquareonthegameboardcould
betoggledbetweenblackandwhitebyclickingitwiththemouse.Eachparticipant’s
32
displayincludedtheirowngameboardandmostrecentscore(givenasthenumberof
squares,bothblackandwhite,markedcorrectlyoutofthetotalnumberofsquaresonthe
board),theirneighbors’gameboardsandscores,andindicationsofthecurrentroundin
thegameandtheamountoftimeremaininginthecurrentround(seeFig.2).Playerscould
copyaneighbor’smostrecentsolutiontotheirownatanytimeduringthegamebyclicking
thechosenneighbor’sboardwiththemouse.Eachgameconsistedof24roundsof10
secondseach1.Afterthelastroundineachgame,participantswereshowntheirguesses
andscoresforeachround,alongwiththegoalpicture,andabuttontoclickwhenthey
werereadytobeginthenextcondition.Whenallparticipantshadclickedthisbutton,the
nextconditionbegan.
Figure2.2:Exampleofaparticipant’sdisplay.
1 Inapilotstudy,we foundthatusingagamestructurewithhalf thisnumberofroundsandtwice the timeperround
yieldedfairlylowaverageandfinalperformancemeasures,andagreatdealofobservedparticipantidlenesstowardthe
endofeachround,whichpromptedthechangetoprovidingmorefrequentfeedbackinthecurrentgamestructure.
33
Participantswereinstructedtomaximizetheirscoresoverallroundsbymatching
thehiddengoalpictureascloselyaspossible.Theywerealsoinformedthatthepicture
theyweresupposedtomatchineachgamewasrandomlygeneratedandnot
representativeofanyparticularobject,shape,orsymbol,andwasgenerallynot
symmetrical;thattheblacksquareswereallconnectedtoeachothervertically,
horizontally,ordiagonally;andthatthenumberofblacksquareswassmallrelativetothe
sizeofthegrid.Followingtheinstructions,theparticipantsweregiventhefirstcondition,
afterwhichtheexperimenterconfirmedthateachoftheparticipantsunderstoodthe
mechanicsofthegame,andansweredanyquestionsthatarose.Theparticipantsthen
playedtheremainderoftheconditions,withanorderwhichwasrandomizedwithineach
session.Attheendofeachgame,participantswereshownthegoalpicture,alongwiththeir
guessesandscoresineachround.
Aparticipant’sscoreineachroundwasdefinedasacell‐by‐cellcomparison
(overlap)betweentheparticipant’sguessforthatroundandthehiddengoalpicture(i.e.
thenumberofcellswhichthetwopictureshadincommon),dividedbythetotalnumberof
squaresinthegoalpicture,togiveapercentagewhichcouldbecomparedbetween
conditionsofvaryinggridsize(seeFig.3).Animprovementwasdefinedasaninstanceofa
participantobtainingascorehigherthanallpriorscoresofallplayerswithinaparticular
condition.Eachparticipant’snormalizedimprovementsharewasdefinedastheir
individuallyachievedproportionofthetotalimprovementsachievedbyallparticipantsin
asession,multipliedbythenumberofparticipantsinthesession.Avalueof1indicateda
“fair”share,e.g.aparticipantachievedonethirdoftheimprovementsinathree‐person
34
session.Aparticipant’sscorerankinaparticularroundwasdefinedastherankoftheir
score(onebeingthebest)amongallscoresinthegroupinthatround;individualswiththe
samescorehadthesamerank.Turnoverforeachround(afterthefirst)wasameasureof
theamountofchangebetweenaparticipant’sguessesoversuccessiverounds.Itwas
definedconverselytosimilarity,exceptthatthetwopicturescomparedwerethe
participant’sguessesfromthecurrentroundandthepreviousround.
Figure2.3:Score/overlapcalculation:thefirsttwopictureshave20outof25squaresin
common(shownindarkgreyontheright),sotheyhaveanoverlapof80%.
Imitation(ameasureofwhetheraparticipantcopiedaneighbor’sguessina
particularround)wasdefinedasfollows:
€
Ipr =1:maxi(overlap(Gp,r,Gi,r -1)) > overlap(Gp,r,Gp,r -1)0 :maxi(overlap(Gp,r,Gi,r -1)) ≤ overlap(Gp,r,Gp,r -1)
;p ≠ i
WhereGp,ristheguessofParticipantpforRoundr,Gi,r1istheguessofNeighborifor
theRoundpriortoRoundr,andoverlapisthecomparisondescribedaboveforthescore
calculation.Inotherwords,aparticipanthasimitatedinaparticularround(Ipr=1)ifthe
35
participant’sguessisclosertothemostsimilarneighbor’spreviousguessthantothe
participant’sownpreviousguess.
Diversity(ameasureofthespreadofgroupmembers’guessesovertheproblem
spacewithinaparticularround)wasdefinedasfollows:
€
Dr =1−majority(Gspr)
p∑
s∑
StotPtot
WhereGspristhebinaryvalue(blackorwhite)ofsquaresintheguessofparticipant
pinroundr,Stotisthetotalnumberofsquaresinthegameboard,Ptotisthetotalnumberof
participantsinthegroup,andmajorityisabinaryfunctionthatconveyswhetherthevalue
ofGsprisinagreementwiththemajorityofparticipantsinthegroupforthatsquareinthat
round(0=notinmajority,1=inmajority).Diversityasdefinedaboveisconstrainedtobe
withinthe0to1range,andhighervaluesofdiversityindicatemoredeviationof
individuals’guessesfromthemajorityguesses.
Thenumberofsquaresinthegameboardwasmanipulatedacrosstwoconditions:
inthesmallboardsizecondition,thegameboardwas7squaresoneachsideforatotalof
49squares,andinthelargeboardsizecondition,thegameboardwas9squaresoneach
sideforatotalof81squares.Thelargerboardwashypothesizedtobemoredifficultto
fullysearch.Therewere4repetitionsofeachcondition,foratotalof8gamesineach
session.Theprobabilitydistributionofscoresamongallpossiblegameboardstatesineach
oftheboardsizeconditionsdescribedaboveisshowninFig.4.Thesizeofthegroup
participatingineachsessionwastreatedasacovariate;groupsizerangedbetween1and9.
36
Anotherfactorconsideredwasthe(randomized)positionofeachconditioninan
experimentsession;thiswascalledthegameorder.
(a)Boardsizeof49 (b)Boardsizeof81
Figure2.4:Distributionofscoresforallpossiblegameboardstatesineachboardsize
condition.Notethatduetothenumberofpossiblegameboardstatesineachcondition
(approximately5.6x1014and2.4x1024,respectively)themeanoverallandfinalscores
appeartofalloutsideofthedistributions,butinfactarejustveryrarescoresintheupper
tails.
2.3.Experiment1Results
Formostanalyses,dependentvariableswereaveragedacrossallparticipantswithin
agrouptogivemeasuresforthegroup’saggregateactivity.Inthismanner,the
fundamentallevelofanalysiswasthegroup,nottheindividual,anddependenciesbetween
individualswithinagroupdonotleadtoelevatedTypeIstatisticalerrors.
37
2.3.1.Boardsize/difficulty
Intheaggregate,participantsachievedfinalscoresof.893(i.e.89.3%oftheway
fromtheworstscoretothebestscore),andaveragescores(overallrounds)of.833%.
Meanfinalscoreswereslightlybutsignificantly(about2percentagepoints)lowerinthe
largerboardsizecondition(t(38)=‐2.88,p=.006;seeFig.2.4).Theaverageguessturnover
rateperroundwas7.3%ofthegameboard,andparticipantsengagedinimitationon
25.8%ofguesses.Therewerenosignificantdifferencesinturnoverorimitationrate
betweenthetwoboardsizeconditions.
2.3.2.Rounds
Thedatawereaveragedacrossallparticipantsandallconditionsineachgroupto
givedependentvariablemeasuresforeachgroupwithineachround.Linearmixed‐effects
modelswereusedtoexaminetrendsacrossroundsforeachdependentvariable,witha
randomeffectofgroupmembershipontheslopeoverrounds.Apreliminaryexamination
ofguesscontentconfirmedexperimenterobservationsthatparticipants’firstround
guessesoften(approximately18.5%ofthetime)consistedofallwhitesquares,becausethe
resultingscorewouldrevealhowmanysquareswerecorrectlymarkedaswhite,andthus
howmanyblacksquareswereinthesolution.Thiswasacleverandusefultacticfor
participants,buttendedtoskewtrendsacrossrounds.Forthisreason,thefirstroundwas
excludedfromanalysis.
Analysisofscoreversusroundshowedastronglysignificantpositivetrend
(F(1,857)=139.91,p<.0001,β=0.577;seeFig.2.5a).Similarly,astronglysignificant
negativetrendwasobservedforturnoverversusround(F(1,857)=169.06,p<.0001,β=‐
38
0.527;seeFig.2.5b).Asignificantnegativetrendwasalsofoundforimitationrateversus
round:participantstendedtoimitateeachotherlessoftenaseachgameprogressed
(F(1,713)=14.37,p=.0002,β=‐0.182;seeFig.2.5c).Guessdiversitywassubjectedtoa
similaranalysisafternormalizingitforparticipantgroupsize,whichwasaccomplishedby
dividingallvaluesbythemeandiversityvalueinthesecondroundfortheappropriate
groupsize,whichwasgenerallyatornearthemaximumduetothefirst‐roundblank‐board
phenomenonnotedabove.Theanalysisshowedthatthediversityofguessesinagroup
decreasedsignificantlyoverthecourseofagame(F(1,713)=33.38,p<.0001,β=‐0.415;see
Fig.2.5d).
(a) (b)
39
(c) (d)
Figure2.5:(a)Meanscoreincreased,while(b)turnover,(c)imitationrate,and(d)guess
diversitydecreasedasmoreroundswereplayedwithinagame.
2.3.3.Gameorder
Similarlinearmixed‐effectsmodelswereusedtoexaminetrendsfordependent
variablesacrossgameorderwithinsessions,averagedacrossallparticipantsandall
roundsineachgame.Onceagain,participantgroupwasusedasarandomeffectineach
model.Analysisofscoreversusgameordershowedasignificantpositivetrend
(F(1,272)=52.69,p<.0001,β=0.437;seeFig.2.6a),whileasimilaranalysisofturnover
showedasignificantnegativetrend(F(1,272)=23.08,p<.0001,β=‐0.305;seeFig.2.6b).
Imitationincreasedsignificantlyacrossgameorder(F(1,246)=11.86,p=.0007,β=0.214;see
Fig.2.6c),whileguessdiversitydecreasedsignificantlyacrossgameorder(F(1,111)=7.27,
p=.0081,β=‐0.282;seeFig.2.6d).
40
(a) (b)
(c) (d)
Figure2.6:(a)Meanscoreand(c)imitationrateincreased,while(b)turnoverand(d)
guessdiversitydecreasedasmoregameswereplayedwithinanexperimentalsession.
2.3.4.Groupsize
Amildupwardlineartrendwasobservedforscoreversusparticipantgroupsize
(F(2,36)=4.56,p=.0395),aswellasamarginalquadratictrendwhichpeakedatagroupsize
of4(F(2,36)=4.33,p=.0446;seeFig.2.7a).Similar,strongerupwardlinear(F(2,28)=24.07,
p<.0001)andquadratic(F(2,28)=16.94,p=.0003;seeFig.2.7c)trendswerealsofoundfor
41
imitationrateversusgroupsize.Nosignificanttrendswerefoundforturnoverorguess
diversityacrossgroupsize,althoughbothdisplayedsubstantialvarianceacrossgroup
sizes,andbothseemedtobegenerallyinverselyassociatedwithscore(seeFig.2.7b&
2.7d).
(a) (b)
(c) (d)
Figure2.7:(a)Meanscoreand(c)imitationrateshowedsignificantquadratictrendsacross
participantgroupsizes,while(b)turnover(d)guessdiversityshowednosignificant
trends.
42
2.3.5.Scoredifferenceandrankinimitation
Analysesofthetargetsofimitationshowedthatnearlyallinstancesofimitation
wereofthosewithhigherscoresthantheimitator’s(seeFig.2.8a),implyingthatimitation
behaviorwasgenerallypurposefulandnotrandom.However,therewasastrongbiasfor
imitatingthetop‐scoringsolutioninsmallergroupsthatweakenedsubstantially
(indicatingthatitwasmoredifficult)inlargergroups(seeFig.2.8b).
(a) (b)
Figure2.8:(a)Nearlyallimitationwasofguesseswithhigherscoresthantheimitator’s,
and(b)therewasastrongbiastowardimitatingtop‐scoringparticipants,whichweakened
inlargergroups.
2.3.6.Learningstrategy
Tofurtherinvestigatetherelationshipbetweenstrategyandperformance,we
performedregressionanalysesofscoreversusmeanratesofimitationandturnoverfor
43
individualsandgroups.Notingthepeakingtrendsforscoreandimitationacrossgroupsize
insection2.3.4,andthedifferenceinimitationtargetsacrossgroupsizeinsection2.3.5,we
splitgroupsizesapproximatelyinthemiddleoftherangecoveredintheexperiment,into
thosewith4orfewerparticipants,andthosewith5ormore.Alinearregressionofmean
individualscoreversusmeanindividualimitationrateshowedasignificantpositive
relationshipforthoseingroupsizesof4orless(F(1,49)=7.69,p=.008,β=0.368),butnone
ingroupsof5orlarger(seeFig.2.9a).Likewise,asignificantpositiverelationshipwas
foundbetweenmeangroupscoreandmeangroupimitationrateingroupsof4orsmaller
(F(1,16)=9.92,p=.006,β=0.619)butnoneingroupsof5orlarger(seeFig.2.9b).Acrossall
groupsizes,therewasasignificantpositiverelationshipbetweenanindividual’sscoreand
themeanimitationrateofallothergroupmembers,excludingtheindividual
(F(1,135)=11.68,p<.001,β=.282;seeFig.2.9c);thatis,regardlessofwhatanindividualdid,
she/hewaslikelytohaveahigherscoreiftheothersinher/hisgroupimitatedmoreoften.
Similaranalysesofscoreversusmeanturnovershowedstrongnegativerelationshipsfor
allgroupsizes,atalllevels:forindividualscoreversusindividualturnover
(F(1,143)=198.9,p<.0001,β=‐0.763;seeFig.2.10a),groupmeanscoreversusgroupmean
turnover(F(1,29)=34.59,p<.0001,β=‐0.738;seeFig.2.10b),andindividualscoreversus
groupothers’meanturnover(F(1,135)=40.0,p<.0001,β=‐0.478;seeFig.2.10c).In
addition,forimitativeguesses,weestimatedavaluefor“innovation”bycalculatingthe
proportionofanimitativeguesswhichwasdifferentfromboththeimitator’sprevious
guessandtheguessthatwasimitated.Thecorrelationofscoreswiththisvaluewerenearly
identicaltothosefoundforturnoverabove.
44
(a) (b) (c)
Figure2.9:Forsmallergroups(<5participants),higherscoreswereassociatedwithhigher
imitationratesforboth(a)individualsand(b)groups;however,theserelationshipsdidnot
holdforlargergroups.(c)Forallgroupsizes,regardlessofaparticularindividual’s
imitationrate,theindividual’sscoretendedtoincreaseastheimitationrateofothersinthe
groupincreased.
(a) (b) (c)
Figure2.10Higherscoreswereassociatedwithlowerturnoverratesfor(a)individualsand
(b)groups,aswellasfor(c)individualsrelativetotheturnoverratesofothersinthegroup
(regardlessoftheindividual’sturnoverrate).
2.3.7.Improvements
45
Anexaminationofparticipants’normalizedimprovementshareshoweda
distributionwithanunequalskew;approximately57.7%ofallparticipantsachievedless
thana“fair”improvementshareof1,whileasmallminorityachievedmuchhighershares
(seeFig.2.11).Inordertocomparethedistributionsofimprovementsumswithan
outcomegeneratedfromarandomprocess,weconstructedaPoissondistributionof
improvementsumsforeachparticipantgroupwithlambda(meanvalue)equaltothemean
improvementsumforthatgroup,andfoundtherangeofvalueswhichcontained50%of
thedensityinthisartificialPoissondistribution.Inover80%ofgroupswithmorethanone
participant,the50%densityrangefromtheassociatedPoissondistributioncontainedless
than50%ofthedensityoftheactualindividualimprovementsums,indicatingthatthey
hadagreaterskewthanwouldbeexpectedbychance.
Figure2.11:Histogramshowingtheunequaldistributionofimprovementsacrossthe
participantswithingroups.(Avalueof1indicatesanevenshare,e.g.anindividualachieved
one‐thirdofthetotalimprovementsinathree‐persongroup.)
46
Meanoverallscoreshowedastrongpositivecorrelationwithimprovementshare
(F(1,105)=46.89,p<.0001,B=0.350)Themeanturnoverrateforguessesthatresultedin
improvementswassignificantlysmallerthanthatofnon‐improvements(0.055for
improvementsvs.0.074fornon‐improvements;t(2040)=12.11,p<.0001).Nosignificant
differencewasfoundformeanimitationrate.
2.3.8.Guesssimilarity
Acomparisonbetweenthesimilarityofimitators’mostrecentguessestothose
whichtheyimitated,andtothosewhichtheydidnotimitate,revealedthattherewas
significantlygreatersimilaritytoimitatedguessesthantonon‐imitatedguesses(.777for
imitatedvs..723fornon‐imitated;t(4914)=‐18.23,p<.0001;seeFig.2.12a).Inotherwords,
imitationtendedtobebiasedtowardguessesthatweremoresimilartotheimitator’sown
priorguess.Thisdifferenceheldoverallroundswithinagame(seeFig.2.11b),even
thoughmeanguessdiversitydecreasedoverroundssuchthatsolutionsgenerally
converged(seeFig.2.5d).Nosignificanttrendswereobservedinlinearregressionsof
similarityversusimitatedscorerank,orthescoredifferencebetweenimitatorandimitated
participants.
47
(a) (b)
Figure2.12:Similaritybiasforimitation.(a)Imitators’previousguessesshowedgreater
similaritytotheguessestheyimitatedthantothosetheydidnotimitate.(b)Thebias
towardimitatingmoresimilarguesseswasconsistentacrossallroundsinagame.
2.4.Experiment1Discussion
2.4.1.Dynamicsandstrategies
Thelargerboardsizehadasignificantnegativeeffectonfinalscores,which
confirmeditsuseasaproxyforproblemdifficulty,butthischangeindifficultyhadno
significanteffectontheotherdependentvariables.However,weobservedrevealing
patternsinparticipants’behaviorthatgavesomecluesabouttheirstrategies.
Increasingmeanscoresacrossroundsandgameordershowedthatparticipantsin
groupslearnedthetaskandtheirdrawingsconvergeduponthecomputer’s“secret”picture
overroundsofonegameandoverthecourseoftheentiresession.Participants
accomplishedtheirimprovementsthroughtheuseoffairlyconservativestrategies,as
evidencedbythelowmeanturnoverrate.Furthermore,thedynamicsofthesestrategies
48
causedsolutionstobecomeincreasinglyentrenchedoverthecourseofthegame.This
happenedintwoways(whichmayhavebeenmutuallyreinforcing):participants’ratesof
imitationandgeneralturnoverdecreasedacrossrounds,andtheimitationthatdidoccur
wasbiasedtowardmoresimilarsolutions.Thisentrenchmentcarriedovertothegroup
levelaswell,shownbythedecreasinggroupsolutiondiversityacrossrounds.Ofcourse,
thisresultislikelypartiallyduetoparticipantsconvergingtowardthegoalpicture,butthe
averagefinalscoreofapproximately89%ofthemaximumsuggeststhatgroupmembers’
solutionsoftenconvergedbeforefindingtheoptimalsolution.
Theproblemspaceusedinthistaskisquitelarge(ontheorderof5x1014possible
solutionsforthesmallerboardsize),andanychangetoasolutionbymorethanonepixel
caneasilyresultinasituationwherescore‐decreasingchangescanceloutscore‐increasing
changes,whichmakesscorefeedbackdifficulttointerpret.Inaddition,thenatureofthe
problemispurelylinear–eachcorrectsquareaddsthesameamounttothetotalscore,and
therearenointeractionsthatmakethesearchproblemmorecomplex,orincentivized
largechangesoversmallones.Thusitmakessensethathighscoreswereconsistently
associatedwithlowturnover,andthatmeanturnoverwassignificantlylowerforguesses
thatresultedinimprovements.Ratherthanlarge,revolutionarychanges,participantsmade
small,incrementalimprovementsbychangingonlyafewcells,typicallyjustone.These
smallchangesallowedparticipantstomakeaccurateinferencesabouttheireffectson
score.
Theunequallyskeweddistributionofimprovementswithineachgroupshowedthat
notallparticipantswereskilledatfindinggoodnewsolutions,thoughimitationallowed
someparticipantstotakeadvantageofotherparticipants’innovationsandmaintainhigh
49
meanscores.Thefactthataverageturnoverwashigherfornon‐improvementguesses
shows,however,thatnon‐improverswerenotjustidlywaitingtoimitateothers’
improvements.
2.4.2.Benefitsofimitation
Imitationwasbiasedtowardhigher‐scoringandmore‐similarguesses,asexpected.
Thelatterallowedparticipantsawaytotakeadvantageofothers'goodsolutionswhile
maintaininglowturnoverandhighercontinuitywiththeirownpreviousguesses,
preservingthevalueoftheirexistingknowledgeoftheproblemspace.
Theassociationofhigherscoreswithgreaterimitationratesatboththeindividual
andgrouplevelsinsmallergroups(whowerebetterabletodistinguishthetop‐scoring
guesswhenimitating)showsthatimitationisnotnecessarilyharmfultoinnovationand
performanceimprovements.Inaddition,theassociationofhighindividualscoreswithhigh
imitationratesbyothersinthegroup(regardlessoftheindividual'sbehavior)indicatesa
systemicbenefitforimitationthatdoesnotaccordwithaviewofimitationasapurelyself‐
benefitingact.Itmaybethat,regardlessoftheintentionsofindividuals,imitationbenefits
thegroupbyactingasafilterforpropagatingandpreservingthebettersolutionsavailable
inagroupatagiventime,aswasfoundinarecentcompetitionofsociallearningstrategies
inasimulatedenvironment(Rendelletal.,2010).Thoughitwasreasonabletoexpect
improvementstobeassociatedwithalowerimitationrate(becausethosewhoonlyimitate
otherscannotdobetterthanthosetheyimitate),wefoundthattherateofimitationwas
aboutthesameamongimprovementsandnon‐improvements,whichmeansthatimitators
wereoftenbuildingontheguessestheyimitatedtocreateimprovements.Improvements
50
wereoftenachievedbyimitatingarelativelysuccessfulparticipant’ssolutionandthen
slightlytweakingthissolution.Oncetweaked,theimprovedsolutionwasthenavailableto
otherparticipants,includingtheindividualwhowasoriginallyimitated.
2.4.3.Imitationinformationoverload?
Therewasanunexpectedlylowerbenefitforimitationinlargergroups,asshownby
thelackofassociationbetweenimitationandscoreinlargergroups,whichwaspresentfor
smallergroups.Largergroupscanbethoughttoprovidemoreinformationaboutthe
distributionofsolutionstotheirmembers,becausetherearemoremodelsforeachgroup
membertoobserve.Thiscanalloweachofthemtomakemoreinformeddecisionsabout
whomtoimitateandwhatchangestomaketotheirguesses.However,itmayhavebeen
thatforlargergroupsizes,theamountofinformationprovidedwasmoredifficulttosearch
andcompare,whichledtomorerandomimitationdecisions(asindicatedbytheweaker
biastowardimitatingthetop‐scoringguess)andthuspoorerconvergenceongood
solutions.Itisunlikelythatthiswaspurelyastatisticalartifactofrandomchoiceamong
moreoptions,becausetherewasauniversaltendency(acrossallgroupsizes)toimitate
better‐scoringpeersthanoneself,arelativelyeasierthingtoaccomplishthanfindingthe
bestscore,butdecidedlynon‐random.So,thoughscoreinformationwasreadilyavailable,
itmayhavebeensubjecttocognitiveloadeffects(Sweller,1988).Largergroupswouldbe
expectedtoshowagreatervarianceinsolutionqualitybychance,butanincreasing
inabilitytoproperlydistinguishgoodsolutionswouldcanceloutthisbenefit.
51
2.5.Conclusions
Thoughourassumptionsaboutparticipants'imitationstrategies(favoringhigher‐
scoringandsimilarguesses)wereempiricallysupported,therelatedpredictionsofpoor
performancewerenotborneout.Therewasaconsistentbenefitforindividualstobein
high‐imitationgroupsregardlessoftheirownbehavior,andimitationwasalsoassociated
withbetterindividualandgroupperformancewhenitcouldbedoneselectivelyand
accurately.
Inshort,thetheoreticalimitation‐relatedsocialdilemmadidnotcauseatragic
outcome,whichisconsistentwiththebenefitfor"conformitybias"foundinprevious
modelsandexperimentsofsociallearning(Boyd&Richerson,1985;Kameda&Nakanishi,
2002),andwiththeliteratureonsocialdilemmaswithrepeatedinteractionsand
recognitionofacollectiveactionproblems(Ostrom,1990).Infact,thedegraded
performanceoflargergroupswaslikelyduetoanimpedimenttoimitation.
2.5.1.Modificationsandmotivations
Thefirst‐gamediscontinuityfortrendsinscore,turnover,andimitationindicated
thatthefirstgameineachsessionmighthavefunctionedasapracticegame,duringwhich
participantswerestilllearningthemechanicsoftheexperiment.Thissuggestedthat
explicitpracticeshouldbeincludedinfutureexperimentsbeforedatagatheringbegins,in
ordertoobtainmoreconsistentdata.Inadditiontothefirst‐roundblankboardstrategy
notedabove,participantswereoccasionallyobservedattemptingtocreateimagesonthe
gridcorrespondingtopatterns,symbols,andalphanumericcharacters,despiteinstructions
tothecontrary.Thisisanunderstandableresponsegivennaturalhumancreativity,the
52
presumablylowvisualorconceptualinterestinthegoalpictures,andpotentialsuspicion
aboutexperimenterdeception.Thesebehaviorsaddedsomenoisetothedata.Thelinear
natureoftheproblemandlackofinteractiveeffectsbetweensolutionelementsmayhave
resultedinalackofvariationfromtheslowandincrementalstrategiesweobservedinthis
experiment.
Theseissuessuggestseveralmodificationsforsubsequentexperiments:(a)amore
engagingtask,with(b)lesspossibilityforperformingthetaskinawaythatdepartsfrom
itsintendedpurpose,(c)aclearerandmoreintuitiveinterfaceandinstructions,(d)more
frequentfeedbacktoparticipantsabouttheirperformanceduringthegame(i.e.more
rounds),(e)moreexplicitrecordingofparticipantchoicesforimitation,innovation,etc.,
and(f)asmaller,morediscreteproblemspace(toimprovebothparticipant
comprehensionandtractabilityofanalyses)withinteractiveeffectsbetweensomesolution
elements(toobservepossiblevariationsinexplorationstrategiesfromthoseinthis
experiment).
Inthisstudywefoundthatthebehaviorofisolatedindividualsattemptingtosolvea
complexproblemdiffersmarkedlyfromthatofpeopleconnectedingroups,andthat
differencesinthesizeofagroupcanhavesignificanteffectsonbehavior.Overall,thereare
strongimplicationsinthesedatafortestingthepredictionsofpastworkintheareaof
groupproblem‐solvingbehavior,andpotentialapplicationstomanyreal‐worldproblems.
Theresultspresentseveralintriguingareasforfurtherstudy,whichwedecidedtoexplore
usingataskmodifiedaccordingtotheguidelinesabove.
53
3.Experiment2–CreatureGameA
3.1.Experiment2Overview
InordertoaddresssomeoftheissuesinExperiment1notedabove,wedesigneda
newtaskthatincorporatedthematicelementsofpopulargamessuchasvirtualpetsand
fantasysportsleagues(thoughtheparametersofthetaskwerelimitedtomakeitmuch
simplerthaneitherofthese).Theparadigminvolvesparticipantstryingtomaximizethe
scoreofachosensubsetofindividualunitsfromalargersetoveraseriesofrounds.
Participantscanseetheselectionsofothersineachroundandimitatetheminwholeorin
part,andtheunitsthatparticipantschoosefromareassignedindividualaswellaspairwise
interactionpointvalues.Thesizeandcomplexityoftheproblemspace(andthusthetask
difficulty,asinExperiment1)weremanipulatedviathesizesofthesetofselectableunits
andthesubsetthatcouldbechosenandevaluatedatonetime,aswellasthenumberof
pairwiseinteractions.Weincludedthoroughwrittenandoralinstructionstoparticipants,
aswellasopportunitiesforhands‐onpracticewiththetaskbeforedatacollectionbegan.
Wechangedtheinterfacetomakescorefeedbackclearer(providingemphasisonthesize
anddirectionofscorechangesaftereachround).Wealsounambiguouslydefinedand
determined(ratherthanhavingtoestimate)instancesofimitation,innovation,andso
forth,byrecordingthesourceofeachplayer’sselectionsintheinterface.
3.1.1.Predictions
Thenewtaskalsoallowedustogatheradditionalexplicitdataaboutthe
proportionsofthesourcesofparticipants’solutionelementchoices,includingInnovation
(theintroductionofsolutionelementswithoutpriorassociatedscorefeedback),and
54
Retrieval(thereinstatementofpreviouslyusedsolutionelementsafterchangingthem),
whichwerenotavailableasseparateclassificationsinExperiment1,aswellasImitation
andRetention(roughlytheinverseofturnoverfromExperiment1).
Imitationwasexpectedtobegreateringreaterdifficultyconditions(reflectingthe
greaterriskinalargerandmorecomplexproblemspace),andInnovationwasexpectedto
decrease.Ingeneral,InnovationwasexpectedtoberelativelyrarecomparedtoImitation,
duetoitsrelativelyhigherrisk.RetentionandRetrievalwereexpectedtoincreaseover
rounds(andthusInnovationandImitationwereexpectedtodecrease),asparticipants
narrowedtheirsearchtosmallerareasoftheproblemspace.Weexpecteddecreasesin
InnovationandRetrievalandanincreaseinImitationinlargerparticipantgroupsizes,
becauseparticipantswouldbeabletorelyonalargerpoolofgoodpeersolutionsto
imitate,whichwouldreducetheneedtoexploreorrelyontheirownprevioussolutions.
Thenewtaskwasgenerallyintendedtomaintainthesameoverallcharacterofthe
taskinExperiment1,whilecorrectingafewissues.Thereforeweexpectedthattheresults
wouldshowthesamegeneralpatternsobservedinExperiment1,withtheexceptionof(1)
smoothertrendsofdependentvariablesacrossgameorder(asopposedtothe
discontinuityinthefirstconditioninExperiment1)duetobetterinstructionandhands‐on
demonstrations;and(2)roughlymonotonictrendsacrossgroupsize(asopposedtothe
peakedquadratictrendsinExperiment1),duetointerfaceimprovementsallowing
participantstosearchandcomparepeersolutionandscoreinformationmoreefficiently.
55
3.2.Experiment2Methods
201participantswererecruitedfromtheIndianaUniversityPsychology
Departmentundergraduatesubjectpool,andweregivencoursecreditfortakingpartinthe
study.Participantspopulatedeachsessionbysigningupatwillforscheduledexperiments
withamaximumcapacityof9persons.49sessionswererun,and10sessions(containinga
totalof48participants)werediscardedduetonetworkorsoftwareproblems.The
remaining153participantsweredistributedacross39sessionsasshowninTable3.1.
Table3.1:DistributionofparticipantsacrossgroupsizesinExperiment2
Groupsize 1 2 3 4 5 6 7 8 9
#Sessions 8 6 5 5 5 2 4 3 1
#Participants 8 12 15 20 25 12 28 24 9
Weimplementedtheexperimentusingcustomsoftwareruninawebbrowser,and
eachparticipantusedamousetointeractwiththeexperimentalgame.Allparticipants’
computerscommunicatedwithagameserver,whichrecordeddataandupdatedscores
andteaminformationforparticipantsattheendofeachround.Inthegameitself,
participantsattemptedtomaximizethenumberofpointsearnedbytheirchosensubsets
(“teams”)fromaset(“league”)ofcreatureiconsover24rounds.Thedisplayincludedan
areafortheparticipant’sowncurrentteam,anotherareathatcouldbetoggledtoshowthe
participant’spreviousroundteamortheirbest‐scoringteamsofarinthegame(alongwith
theassociatedscore),aleagueareawhichshowedalloftheicons(potentialteam
members)thatwereavailableforselection,andindicationsofthecurrentroundinthe
gameandtheamountoftimeremaininginthecurrentround.
56
Ifasessionincludedmorethanoneparticipant,eachparticipant’sdisplayalso
showedtheteamandassociatedscoreforallotherparticipantsinthepreviousround.
Iconscouldbecopiedfromanypartofthedisplaytoaparticipant’scurrentteamby
dragginganddroppingthemwiththemouse,exceptforthosealreadyontheparticipant’s
currentteam,whichwerefadedinthedisplayandnon‐clickable.Thecurrentteamcouldbe
replacedentirelybyanotherteambyselectingthescoreboxabovethelatterasa“handle”
anddraggingittothecurrentteamarea.Theorderingofpeers’teamsineachparticipant’s
displaywasnotkeptconstantacrossconditions,toavoidimitationbasedonpastbehavior.
AscreenshotoftheparticipantinterfaceisshowninFig.3.1.
Figure3.1:ExampleofexperimentinterfaceinExperiment2.
57
Atthebeginningofeachsession,playersweregivenahands‐ondemoofthegame
(includingthevariouswaystomovecreaturestoone’scurrentteam),andfurtherinformed
aboutthemechanicsofthegameandwhattoexpectintheremainderoftheexperiment
session,includingthefollowinginformation.Eachgameconsistedof24rounds,andeach
roundwas10secondslong,asinExperiment1.Scorefeedbackwasgivenaftereachround:
iftheparticipant’sscorehadimprovedfromthepreviousround,thenumericalscore
displaycounteduptothenewscoreandturnedgreen,andifithadworsened,thedisplay
counteddowntothenewscoreandturnedred.Attheendofeachgame,thedisplay
showedtheplayer’sfinalscore,alongwithatableofthescoresofeachplayerineach
roundofthegame,whichwassortedbyaveragescore.Theplayer’sownscoreswere
highlightedtoshowtheirrelativeperformancewithoutplacingcompetitiveemphasisonit.
Playerswereinstructedtodotheirbesttomaximizetheirteams’scoresoverall24rounds.
Atthebeginningofeachgame,eachplayer’steamwasarandomselectionofcreatureicons
fromtheleague.
Eachgroupplayed8games,ofwhichhalfhadalargeleagueandteamsize(48and
6,respectively),andhalfweresmaller(24and5).Thesetwoparametersettingswere
intendedtovarythelevelofdifficultyofthegame,withtheformerbeingmoredifficult.
Thiswasbecausealthoughthescoredistributionandcombinatoricsmadehigher
numericalscorespossible,italsomadehigh‐scoringteamsrarerthaninthelattercase:for
thesmallerleaguesize,about4%ofallpossibleteamshadscoresgreaterthan70%ofthe
maximum,whileforthelargerleaguesizethisfigurewas0.4%(seeFig.3.3).Thatis,with
thelargerleagueandteamsize,itwashardertofindgoodsolutionsbecauseofthe
relativelylargenumberofpoorersolutions.
58
Ineachgame,eachiconwasassociatedwithacertainpositivenumberofpoints,and
severalspecialpairsoficonswereassociatedwithseparatescorebonusesorpenaltiesthat
capturedinteractionsbetweenicons.Thescoreforateamwascomputedbysummingthe
individualpointvaluesforeachicon,andthenaddingorsubtractingthevalueofany
specialpairspresent.Thepairsdidnotoverlap,andthedistributionwasdesignedtobe
challenging:pairswhichgavelargepositivebonusesweredistributedamongiconswith
smallindividualpointvalues,andpairswhichgavelargenegativepenaltiesweregenerally
foundamongiconswithlargeindividualpointvalues(seeFig.3.2).
(a)Leaguesizeof24
(b)Leaguesizeof48
Figure3.2:Pointdistributionforindividualicons(boxes)andinteractionbonusesand
penalties(ovals),forleaguesizesof(a)24and(b)48.
59
Individualpointvaluespericonrangedfrom1to8points,andpairinteraction
valuesrangedfrom‐20to20points,sothatthepossiblescorerangesforthelargeand
smallleagueandteamsizecombinationswere[‐6,60]and[‐6,51],respectively.Foreaseof
comparisonandanalysis,allscoreswerenormalizedtotherange[0,1]accordingtothe
rangeofscorespossiblewiththeassociatedleagueandteamsizecombination.The
combinationsofsuchindividualandpairvaluesresultedintheprobabilitydistributionof
scoresamongallpossibleteamsforeachleaguesizeshowninFig.3.3.
Participantswerenotgiveninformationaboutthemaximumscore,thescore
distribution,orthepositionoftheinteractionterms,thoughtheycouldpotentiallyhave
beendeducedduringplay.Theicons’displaypositionandassociationswiththepoint
distributionwereshuffledrandomlyforeachgame,sothattheirappearanceand
placementinthedisplaydidnotgivecluesastotheirpointvaluesduringthecourseofan
experimentsession.
(a)Leaguesizeof24 (b)Leaguesizeof48
60
Figure3.3:Distributionofscoresforallpossibleteamsineachdifficulty(leaguesize)
condition.
Ineachround,thefollowingdatawereautomaticallyrecordedforeachplayer:the
iconsonthecurrentteamattheendoftheround(orchoices),thesourceofeachicon,and
theresultingscore.Thesourceinformationindicatedwhethereachiconwasunchanged
fromthepreviousround(Retained),copiedfromtheplayer’sownpreviousroundteam
afterinitiallybeingremovedfromtheteam(Returned),copiedfromtheplayer’sownbest‐
scoringteamsofar(Retrieved),chosenfromtheleaguedisplay(Innovated),orcopiedfrom
anotherplayer’steam(Imitated).WhenImitationwaschosen,thepersistentidentifierof
thecopiedplayerwasrecordedtoallowfurtheranalysesofimitationdecisions.
InthecaseofaplayerreplacingtheentireteamwithImitatedicons,allchoiceswere
recordedasImitated,evenifoneormoreofthemwerealreadyontheplayer’sprevious
roundteam.(ThesamewastrueofreplacinganentireteamwithRetrievedicons,or
removinganiconandthenputtingitbackontheteamviaaLeaguechoice.)However,a
comparisonwiththeplayer’spreviousroundteamallowsthedeterminationofacorrected
choicesourceforeachicon,effectivelyrevealingapotentialincreaseintheRetainchoice
source,andapotentialdecreaseforallotherchoicesourcesexcepttheReturnchoice
source.Inessence,theoriginaluncorrectedchoicesourcesrevealtheliteralactionsof
players,whilethecorrectedchoicesourcesrevealtheireffects.Scoreimprovementsand
scoreranksweredefinedidenticallytothoseinExperiment1.
SimilartoExperiment1,guessdiversityforaparticularroundwasdefinedasthe
proportionoficonsintheleaguerepresentedononeormoreparticipants’teamsduringa
61
givenround.Thisvaluewasnormalizedbytheaverageexpectedvalueofthisproportion
foreachparticipantgroupsize,generatedbyaMonteCarlosimulationassuming
independentrandomteams.(Itshouldbenotedthatthiscalculationandtheassociated
solutionspacesinExperiments1and2areofdifferenttypesandsizesduetothenatureof
thetaskineach.)
3.3.Experiment2Results
Participantsachievedameanoverallscoreof.596(i.e.about60%ofthewayfrom
theworstscoretothebestscore),andameanfinalscoreof.690.Theaverageguesschoice
sourceproportions(originalandcorrected)areshowninTable3.2.TheverylowReturn
ratesuggestedthatanaverageturnoverratedefinedequivalentlytothatinExperiment1
couldbeapproximatedasoneminusthemeanRetentionproportion,or26.1%.(Allfurther
analysesofchoicesourcerefertothecorrectedversions,exceptwherenoted.)
Table3.2:AverageOriginalandCorrectedchoicesourceproportions
ChoiceSource Imitation Innovation Retention Retrieval Return
Original 20.6% 14.0% 59.5% 6.5% 0.5%
Corrected 9.8% 13.7% 73.9% 2.6% n/a
Difference ‐10.2% ‐0.3% +14.4% ‐3.9% n/a
Fortheinitialanalysis,dependentvariableswereaveragedacrossallparticipants
andallroundstogivemeasuresforthegroup’saggregateactivityineachgame.Thesedata
wereanalyzedusingarepeated‐measuresANOVA,treatingeachgroupasasinglesubject,
62
withdifficulty(leaguesize)asawithin‐groupsfactor,thegroupsizeusedasacovariate,
andtheparticipantgroup(sessionidentifier)includedasarandomeffectinthemodelfor
eachdependentvariable.Datafromthefirstroundofeachgamewasexcludedfrom
analysesofchoicesource,becauseseveralchoicesources(Imitated,Retrieved,Retained,
andReturned)areonlydefinedrelativetoapreviousround.Theaboveanalysesrevealed
significantmaineffectsofdifficultyandgroupsizeonscoreandseveralchoicesources,and
asignificantmaineffectofgroupsizeonguessdiversity,aswellasasignificantinteraction
effectbetweendifficultyandgroupsizeonguessdiversity.Theseresults,aswellas
dependentvariabletrendsoverrounds,gameorder,choicesourceuse,scoreperformance,
andguesssimilarityaredescribedindetailbelow.
3.3.1Leaguesize/difficulty
Participantsachievedmeanoverallscores(averagedacrossallrounds)andmean
finalscoresforeachLeagueSizeasshowninTable3.3andFig.3.3.Scoreswerenormalized
asdescribedabove,andpercentilesweredefinedasthepercentageofallpossibleteams
withlowerscoresthantheassociatedmeanscore.
63
Table3.3.Meanscore,guessdiversity,andchoicesourceproportionsineachcondition.
LeagueSize
MeanOverallScore
(Percentile)
MeanFinalScore
(Percentile)
MeanGuess
Diversity
MeanImitation
MeanInnovation
MeanRetention
MeanRetrieval
24 .606(90.1%) .693(95.7%) 67.6% 6.8% 16.8% 69.5% 6.3%
48 .539(93.7%) .634(98.5%) 63.4% 7.1% 14.9% 72.1% 5.3%
Diff. .067* .059* 4.2%* +0.3% 1.9%* +2.6%* ‐1.0%
*significantdifferences
Atwo‐samplet‐testwasusedtoexaminethemaineffectofdifficultyoneach
dependentvariable.Itwasfoundthatrelativetothelowerdifficultycondition,thehigher
difficultyconditionresultedinsignificantlylowerInnovation(t(306.8)=2.71,p=0.007),and
significantlygreaterRetention(t(307.5)=‐2.02,p=.0445),butnosignificantdifferences
werefoundforImitationorRetrievalbetweendifficultyconditions,andtrendsforallwere
unchangedforuncorrectedchoicesources.Thehigherdifficultyconditionresultedin
significantlylowerguessdiversity(t(233.8)=2.16,p=.0320),andsignificantlylowerscores
(t(305.8)=7.55,p<.0001;seeFig.3.3).Thescorepercentilesachievedareactuallyhigher
forthelargerLeagueSize,butthisismostlikelyduetothedifferenceintheshapesofthe
twoscoredistributionsused.
3.3.2.Rounds
Linearmixed‐effectsregressionmodelswereusedtoexaminetrendsacrossrounds
foreachdependentvariable,witharandomeffectofgroupontheslope.Analysisofscore
vs.roundshowedastrongpositivetrend(F(1,919)=897.05,p<.0001,β=0.717;seeFig.3.4).
64
Theaverageimprovementinscoreacrossroundswithinagamewas23.7%,andtrendsin
scoresoverroundswerepositiveforallgroupsizes.Guessdiversityshowedasignificant
decreaseacrossrounds(F(1,735)=188.62,p<.0001,β=‐0.404;seeFig.3.4),andsuchtrends
overroundsweremorestronglynegativeforincreasinggroupsize.
Asforchoicesources,Imitationshowedasignificantdecreaseoverrounds
(F(1,681)=126,p<.0001,β=‐0.453),asdidInnovation(F(1,857)=70.78,p<.0001,β=‐0.277).
Theoverallincidencerateofimitationdecreasedsimilarly.Retentionincreased
significantlyacrossrounds(F(1,857)=21.43,p<.0001,β=0.214),asdidRetrieval
(F(1,857)=9.67,p=.0019,β=0.128;seeFig.3.5).Usinguncorrectedchoicesourcesrevealed
nosignificantchangestothesetrends,exceptforslightdifferencesinslope.
Figure3.4:Meanscoreincreasedandmeanguessdiversitydecreasedasmoreroundswere
playedwithinagame;strongereffectswereobservedforlargerparticipantgroupsizes.
65
Figure3.5:MeanproportionsofRetentionandRetrievalincreasedandImitationand
Innovationdecreasedasmoreroundswereplayedwithinagame.
3.3.3.Gameorder
Similarlinearmixed‐effectsmodelswereusedtoexaminetrendsacrossgameorder
withinsessionsforeachdependentvariable.Scoredisplayedasignificantupwardtrend
acrossgameorder(F(1,272)=14.69,p=.0002,β=0.186;seeFig.3.6a).Theaverage
improvementinscoreacrossgameorderwithinasessionwas5.6%.Guessdiversity
displayedacorrespondingdownwardtrend(F(1,216)=20.02,p<.0001,β=‐0.180;seeFig.
3.6b).Nosignificanttrendswerefoundfororiginalorcorrectedchoicesourceproportions
overgameorder.
66
Figure3.6:Asmoregameswereplayedwithinanexperimentalsession,(a)Meanscore
increased,and(b)thediversityofguessesdecreased.
3.3.4.Groupsize
Trendsacrossparticipantgroupsizeforeachdependentvariablewereexamined
usinglinearmixed‐effectsmodels,withthegroup(sessionidentifier)usedasarandom
effectontheintercept.Scoreshowedasignificantupwardtrendacrossgroupsize,withan
averagescoredifferenceof11%betweenisolatedparticipantsandthoseinthelargest
groupsizeof9(F(1,37)=73.62,p<.0001,β=0.466;seeFig.3.7a).Guessdiversitydecreased
significantlyforlargergroups(F(1,29)=38.25,p<.0001,β=‐0.663;seeFig.3.7b).
Asforchoicesources,Imitationincreasedsignificantlyforlargergroups
(F(1,29)=22.35,p=.0001,β=0.565),andRetentionincreasedaswell(F(1,37)=12.09,
p=.0013,β=0.433),whileInnovationdecreased(F(1,37)=28.95,p<.0001,β=‐0.563),and
Retrievaldecreasedaswell(F(1,37)=12.46,p=.0011,β=‐0.464;seeFig.3.8).Original
(uncorrected)choicesourcesshowednotrendforRetentionacrossgroupsize,but
67
otherwisemaintainedthesamepatternofresultsnotedabove,withslightdifferencesin
slope.
Figure3.7:(a)Asparticipantgroupsizeincreased,meanscoresinagroupincreased,and
(b)thediversityofofferedsolutionsdecreased.
Figure3.8:Asparticipantgroupsizeincreased,meanproportionsofRetentionand
Imitationincreased,andInnovationandRetrievaldecreased.
68
GiventheresultsforImitationabove,itisimportanttodeterminewhetherthescore
advantageforlargergroupswassimplyanartifactofthegreaterchanceofobservinga
betterscorethanone’sowngiventhelargernumberofguessesbeingmade,leadingto
moreimitationandthushigherscores.Inordertodeterminewhetherthiswasthecase,we
calculatedthescoredifferencevariance(SDV):thevarianceofthedifferencesbetweenthe
top‐rankedparticipantandallotherparticipantswithineachround,averagedwithineach
game.Usingalinearmixed‐effectsmodelliketheothersusedforgroupsizeanalysesabove,
weconfirmedaslightbutsignificantupwardtrendofSDVacrossgroupsize
(F(1,29)=11.37,p=.0021,β=0.262).However,asimilaranalysisofImitationproportionvs.
SDVdidnotrevealanysignificanttrend,andcontrollingforSDVintheImitation
proportionvs.groupsizemodelabovedidnotalteritsignificantly.Inotherwords,the
greaterImitationinlargergroupsdidnotappeartobeduetoincreasedscorevariance.
3.3.5.Scoredifferenceandrankinimitation
Theanalysesinthissectionrefertooriginaluncorrectedchoicesources,because
theirintentistocaptureparticipants’awarenessofotherplayers’scores,notthecontentof
theirguesses.OfallguesseswithgreaterthanzeroImitationproportion,94.3%imitated
onlyoneotherparticipant,5.1%imitatedtwoparticipants,and0.6%imitatedmorethan
twoparticipants.Ofallinstancesofsingle‐participantimitation,82.4%involvedimitation
ofparticipantswhosescorerankwas1(thetopscoreinthegroup),10.7%whosescore
rankwas2,and7%whosescorerankwas3orbelow(seeFig.3.9a).Atthetimeofsuch
single‐sourceimitations,thescoreoftheimitatedparticipantwasgreaterthanthatofthe
69
imitatorin89.6%ofcases,equaltoitin2.6%ofcases,andlessthanthatoftheimitatorin
7.8%ofcases(seeFig.3.9b).Nosignificantdifferencesacrossgroupsize,round,etc.were
observedinthesevalues.
(a) (b)
Figure3.9:Therewerebiasestowardimitating(a)better‐scoringparticipantsthanoneself,
and(b)thebest‐scoringparticipant,andthesebiaseswereunaffectedbygroupsize.
3.3.6.Choicesourcestrategy
Thechoicesourcesofeachnon‐isolatedparticipantovertheentiresessionwere
analyzed,andeachparticipant’schoicesourcestrategywascategorizedaccordingtotheir
proportionofeachsource.Participantswhosechoicescontainedonesourceinanaverage
proportiongreaterthantheglobalaverageforthatsourceplusonestandarddeviation,
werelabeledwiththatstrategy.Forexample,aplayerwhoseguessesoverthecourseofa
sessionconsistedofagreaterproportionofImitatechoicesthantheaverageforallother
participantsintheexperiment,plusonestandarddeviation,werelabeledashavingan
70
overallstrategyof“Imitate.”Thosewhofittheabovecriteriaformorethanonechoice
source,ornone,werelabeledashavinga“Mixed”strategy.Thescoredistributionfor
playersineachstrategycategoryisshowninFig.3.10(a),withtheRetainstrategyscoring
thebest,followedbyMixed,Imitate,andRetrieve,withInnovatescoringtheworst.
Analysisoforiginaluncorrectedchoicesourcestrategiesshowedasimilarpattern,except
thattheImitateandRetainstrategiesswitchedplaces(seeFig.3.10(b).
(a) (b)
Figure3.10:Scorevs.(a)correctedchoicesourcestrategyand(b)uncorrectedchoice
sourcestrategy,showingthataconservativeandimitativestrategyresultedinthebest
performance.
Theabove‐mentionedfiguressummarizetheresultsofsimpleregressionanalyses
performedforscorevs.individualandgroupuseofeachchoicesource.Alinearregression
ofmeanindividualscorevs.meanindividualImitationguessproportionshoweda
significantpositivetrend–thegreateraparticipant’saverageproportionofImitation,the
bettertheparticipant’sscore(F(1,143)=8.64,p=.0038,β=0.239;seeFig.3.11a).Asimilar
71
positivetrendheldforRetention(F(1,143)=55.72,p<.0001,β=0.530;seeFig.3.12a).The
oppositewastrueforindividualscorevs.Innovation,whichdisplayedasignificant
negativetrend(F(1,143)=119.8,p<.0001,β=‐0.675;seeFig.3.13a),asdidRetrieval
(F(1,143)=10.93,p=.0012,β=‐0.267;seeFig.3.14a).Analysesofscorevs.original
uncorrectedchoicesourcesshowedsimilartrendsforImitationandInnovation,butnone
forRetentionorRetrieval.
Averysimilarpatternofresultswasshowninanalysesofmeangroupscorevs.
meangroupguessproportionforeachchoicesource,withupwardtrendsforImitation(i.e.
thehigheragroup’smeanImitation,thehigheritsmeanscore)andRetention,and
downwardtrendsforInnovationandRetrieval(allp<0.001;seeFigs.3.11b‐3.14b).The
onlydifferenceforanalysesofuncorrectedmeangroupchoicesourceswasthelackofa
trendforRetention.Likewise,averysimilarpatternofresultswasfoundforanalysesof
meanindividualscorevs.meangroup(excludingtheindividual)guessproportionforeach
choicesource.TherewasapositivetrendforImitation(i.e.themoreanindividual’sfellow
groupmembersimitated,thehighertheindividual’sscore),aswellasRetention,and
negativetrendsforInnovationandRetrieval(seeFigs.3.11c‐3.14c).Onceagain,theonly
differenceforanalysesofuncorrectedmeangroupchoicesourceswasthelackofatrend
forRetention.Alltrendsnotedaboveweregenerallymonotonic;thatis,therewereno
thresholdsorinflectionpointsbeyondwhichtherelationshipschanged.
72
(a) (b) (c)
Figure3.11:Higherscoreswereassociatedwithhigherimitationratesfor(a)individuals
and(b)groups,aswellasfor(c)individualsrelativetotheimitationratesofothersinthe
group(regardlessoftheindividual’simitationrate).
(a) (b) (c)
Figure3.12:Higherscoreswereassociatedwithlowerinnovationratesfor(a)individuals
and(b)groups,aswellasfor(c)individualsrelativetotheinnovationratesofothersinthe
group(regardlessoftheindividual’sinnovationrate).
73
(a) (b) (c)
Figure3.13:Higherscoreswereassociatedwithhigherretentionratesfor(a)individuals
and(b)groups,aswellasfor(c)individualsrelativetotheretentionratesofothersinthe
group(regardlessoftheindividual’sretentionrate).
(a) (b) (c)
Figure3.14:Higherscoreswereassociatedwithlowerretrievalratesfor(a)individuals
and(b)groups,aswellasfor(c)individualsrelativetotheretrievalratesofothersinthe
group(regardlessoftheindividual’sretrievalrate).
3.3.7.Improvements
AsinExperiment1,improvementsweretalliedforeachparticipantineachsession.
Ahistogramofnormalizedimprovementshareshowedthatjustoverhalf(54.5%)of
74
participantsachievedlessthanthe“fair”shareof1;however,thedistributionispeaked
stronglyaround1,andtherewerenoparticipantswhoobtainedzeroimprovements(see
Fig.3.9).JustasinExperiment1,inordertocomparethedistributionsofimprovement
sumswithanoutcomegeneratedfromarandomprocess,weconstructedaPoisson
distributionofimprovementsumsforeachparticipantgroupwithlambda(meanvalue)
equaltothemeanimprovementsumforthatgroup,andfoundtherangeofvalueswhich
contained50%ofthedensityinthisartificialPoissondistribution.Injustover60%of
groupswithmorethanoneparticipant,the50%densityrangefromtheassociatedPoisson
distributioncontainedlessthan50%ofthedensityoftheactualindividualimprovement
sums,indicatingthattheygenerallyhadonlyaslightlygreaterskewthanwouldbe
expectedbychance.
Figure3.15:Histogramshowingrelativelyequalachievementofimprovementswithin
groups.(Avalueof1indicatesanevenshare,e.g.anindividualachievedone‐thirdofthe
totalimprovementsinathree‐persongroup.)
75
The mean choice source proportions for guesses that resulted in score
improvementsandthosethatdidnotareshowninTable3.6.
Table3.6.Meanchoicesourceproportionsforimprovementandnon‐improvement
guesses.
ChoiceSource Imitation Innovation Retention RetrievalProportioninnon‐improvementguesses
9.9% 12.9% 74.9% 1.7%
Proportioninimprovementguesses
8.0% 20.2% 70.0% 1.4%
Difference 1.9%* +7.3%* 4.9%* ‐0.3%*significantdifferences
3.3.8.Guesssimilarity
Acomparisonbetweenthemeansimilarityofparticipants’mostrecentguessesto
thosewhomtheyimitated,andtothosewhomtheydidnotimitate,revealedaslightbut
significantdifference:.550forimitatedvs..503fornon‐imitated(t(5029)=‐7.10,p<.0001;
seeFig.3.20).Inotherwords,priortoimitation,theaverageimitators’guesswasmore
similartothatoftheimitatedparticipant(s)thantothoseofothers.Inaddition,the
differencebetweenmeansimilarityforimitatedandnon‐imitatedparticipantsremained
overrounds(seeFig.3.21).Nosignificanttrendswereobservedinlinearregressionsof
guesssimilarityvs.imitatedscorerank,orsimilarityvs.thescoredifferencebetween
imitatorandimitatedparticipants.
76
Figure3.16:Similaritybiasforimitation.(a)Imitators’previousguessesshowedgreater
similaritytotheguessestheyimitatedthantothosetheydidnotimitate.(b)Thebias
towardimitatingmoresimilarguesseswasconsistentacrossroundsinagame.
3.3.9Frequencyandmomentumbias
Inordertomeasurethebiasofparticipantstochooseaniconaccordingtoits
frequencyinneighbors’choices,wetalliedthenumberofplayersinthegroupwhoseteams
includedeachiconinthepreviousround(NR‐1),aswellasthenumberoftheremaining
playerswhoaddedittotheirteaminthecurrentroundviaImitationorInnovation.To
convertthesefigurestonormalizedfrequencies,thefirstnumberwasdividedbythe
participantgroupsize(N),andthesecondnumberwasdividedbythenumberof
participantswhodidnotpossesstheiconinthepreviousround(N–NR‐1).Inthiswaywe
wereabletomeasurethemeanprobabilityofImitationandInnovationforanyiconnot
alreadyincludedonaplayer’steam,basedonthefrequencyofitsappearanceon
neighbors’teamsintheplayer’sdisplay.
77
Thechanceprobabilityofimitation(resultingfromchoosinganiconatrandomfrom
amongallneighbors’teams)scaleswiththechoicefrequencyofaniconrelativetotheteam
size.Thechancelevelofinnovation(resultingfromchoosinganiconatrandomfromthe
leaguedisplay)isaconstantatoneovertheleaguesize.Sinceleagueandteamsize
conditionswerebalancedinallsessions,weusedtheaveragevalueofeachtocalculatethe
chancebaselines.Alinearmixed‐effectsanalysisofimitationprobabilityversuschoice
frequencyshowedapositivefrequency‐dependentImitationbiasthatwassignificantly
greaterthanchance(F(1,1128)=1648,p<.0001,B=.300;seeFig.3.17a),aswellasasimilar
butmuchsmallerfrequency‐dependentInnovationbias(F(1,1128)=268.7,p<.0001,
B=.062;seeFig.3.17b).Interestingly,theprobabilitiesofImitationandInnovationonly
roseabovechancewhenthemajorityofaparticipant’sneighborspossessedanicon(i.e.
whenChoiceFrequencywasgreaterthan0.5).
(a) (b)
78
Figure3.17:Therewerebiasestowardchoosingelementsthatweremorefrequently
representedonotherteamsin(a)Imitationand(b)Innovationdecisions,showingacopy
themajoritystrategy.
Inasimilaranalysisof“choicemomentum,”wetalliedthechangeinthenumberof
playerswhoseteamsincludedtheiconintheprevioustworounds(NR‐1‐NR‐2),aswellas
thenumberoftheremainingplayerswhoaddedittotheirteaminthecurrentroundvia
ImitationorInnovation.Toconvertthesefigurestonormalizedfrequencies,thefirst
numberwasdividedbytheparticipantgroupsize(N),andthesecondnumberwasdivided
bythenumberofparticipantswhodidnotpossesstheiconineitheroftheprevioustwo
rounds(N–max(NR‐1,NR‐2)).Inthiswaywewereabletomeasurethemeanprobabilityof
ImitationandInnovationforanyiconnotalreadyincludedonaplayer’steam,basedonthe
changeinfrequencyofitsappearanceonneighbors’teamsintheplayer’sdisplayoverthe
previoustworounds.
Thedistributionoffrequencychangesforalliconswasverynearlysymmetrical
aroundzero,suchthatanequivalentnumberofpositiveandnegativeproportionchanges
occurred,withsmallabsolutechangesmorecommonthanlargeones.Afterlog‐
transformingtheImitationprobabilitydatatoachieveanormaldistribution,at‐testof
Imitationprobabilityfornegativeandpositivechangesinchoicefrequencyshoweda
significantpositivemomentumbias(t(1236)=18,p<.0001;seeFig.3.18a),andasmallbut
non‐significantmomentumbiasforInnovation(seeFig.3.18b).Therewasalsoasignificant
diminishingofthepositivemomentum‐biasedImitationeffectacrossrounds,butno
changeacrossgameorderorparticipantgroupsize.
79
(a) (b)
Figure3.18:Therewerebiasestowardchoosingelementswhoserepresentationonother
teamswasincreasingin(a)Imitationand(b)Innovationdecisions.
3.4.Experiment2Discussion
3.4.1.Dynamicsandstrategies
AsinExperiment1,participants’increasingmeanscoresacrossroundsandgame
orderassuredusthattheylearnedthetaskandwerenotguessingrandomly.Theoverall
characteroftheirstrategiessharedtheconservativepatternofthoseinExperiment1,as
evidencedbythehighmeanproportionofRetention(whichincreasedacrossrounds);this
cautiousapproachwasaccentuatedinthehigherdifficultycondition.Likewise,the
correctedchoicesourceresultsshowedthatmanyguesselementsthatwereinitially
classifiedasImitationorRetrievalwereactuallycomposedoflargelyRetainedelements.
OurpredictionofahighermeanproportionofImitationrelativetoInnovationwas
borneoutinparticipants’intentions(asrecordedintheiruncorrectedchoicesources),but
contradictedintheireffects(correctedchoicesources)–theactualproportionofsolution
80
elementsdevotedtoInnovationwashigherthanImitation.Aspredicted,wedidfind
significantlylowerInnovationinthehigherdifficultycondition,butasinExperiment1,
therewasnosignificantdifferenceinImitationbetweenconditions.
Individualguessesbecameincreasinglyentrenchedovertime,asevidencedbythe
decreasingproportionsofinnovationandimitation,andincreasingproportionsof
retentionandretrieval,acrossrounds.Thisbehaviorisconsistentwiththecopywhen
uncertainstrategyinthatmoreimitationoccurredearlyonineachgamewhenparticipants
hadlessexperiencewiththecurrentproblem.Guessesbecameentrenchedatthegroup
levelacrossroundsaswell(asshownbydecreasinggroupsolutiondiversity)despite
decreasingamountsofimitation,becausetheremainingimitationwasincreasinglydriven
byconvergentbiasestowardgreaterguesssimilarity,higherchoicefrequency,andpositive
choicemomentum.Thesebiasesalsohelpexplainthedecreaseinguessdiversityinthe
greaterdifficultyconditionwithoutanaccompanyingincreaseintheincidenceofimitation.
WhereasBaron,Vandello,andBrunsman(1996)foundthatincreasingtaskdifficulty
increasedtheincidenceofimitation,inthisexperimentitappearstohaveinsteadchanged
thefocusoftheimitationthatoccurredtofavorincreasedgroupsolutionhomogeneity.
Thefactthataveragefinalscoresarelessthan70%ofthemaximumpossiblescore
impliesthat,especiallyinlargergroups,participantsaresettlingongoodbutsuboptimal
solutionsduetoinsufficientsearchofthemultimodal,“rugged”problemspace.Thisresult
agreeswiththefindingsofMason,Jones,andGoldstone(2008),thatfully‐connected
groups(liketheonesinthisexperiment)performedrelativelypoorlyonamultimodal
problemspace,whereasmoresparsely‐connectedgroups(latticesandsmall‐world
networks)foundoptimalsolutionsmorereliably,thoughmoreslowly.
81
Thelackofsignificantchangesinchoicesourceproportionsacrossgameorder
withinsessionsimpliesthatlearninginthecombinedcontextoftheproblemspaceandthe
groupoccurredanewforeachgame,withoutmajoradaptationsofthemembersofthe
grouptoeachotheroverthecourseofexperiment.Thismayhavebeenduetothelackof
communicationavailablefordiscussingorcoordinatingactionswithinthegroupduring
thesession(Ostrom,Gardner,&Walker,1994).However,guessdiversitydecreasedacross
gameorder,suggesting(aswiththedecreaseacrossdifficultyconditionsnotedabove)
mechanismsofconvergencethatmayhaveoperatedthroughtheabovebiasesinimitation.
3.4.2.Groupsizeeffects(andlackthereof)
ThepredictedincreaseinImitationwithlargergroupsize(afteraccountingfor
artifactualscorevarianceexplanations),alongwithdecreasedInnovationandRetrieval
indicateabiastowardsociallearningthatscaleswiththenumberofmodelsolutionsto
compare,choosefrom,andintegrate,andtheaccompanyingincreaseinscoreindicatesthat
thiswasabeneficialstrategyforthistask.Conversely,thereductioninRetrievalwith
increasinggroupsizeindicatesagreaterdependencebyisolatedindividualsandthosein
smallergroupsonthebuilt‐in“memory”oftheBestScoreoptioninthegameasasourceof
reliablygoodsolutionsonwhichtobuild.Thecombinationoftheseresultsimpliesthatin
largergroups,thisfunctionofmemorymaybe“outsourced”tootherswhoimitateandthus
propagateandpreservegoodsolutionswithinthegroup.Thisphenomenonhasbeen
exploredpreviouslyasa“divisionofcognitivelabor”inthetheoryof“transactivememory”
ingroupcognition(Theiner,2009;Wegner,1986).Adifferentapproachtothecopywhen
uncertainstrategyisshownhere:Imitationisfavoredwhenthepayoffforinnovationis
82
relativelyuncertain,comparedtotheabundantinformationavailableaboutthecontent
andutilityofneighbors’guesses.
Weconfirmedtheintuitivepredictions(andthegeneralpatternfromExperiment1)
thatmostimitationeventswereofthetop‐rankedneighbor(confirmingtheuseofthecopy
successfulindividualsstrategy),andofneighborswithhigherscores(confirmingtheuseof
thecopyifbetterstrategy).Theactionsofimitatorswhochosenon‐top‐rankedoreven
lower‐scoringneighborstoimitatewerenotexplainedbysimilaritybetweentheirguesses,
andmayhavebeenduetorandomerrors.Thelackofinfluenceofparticipantgroupsizeon
thisresult(aswellastheroughlymonotonicincreasesinscoreandImitationacrossgroup
size,andthemoreequitabledistributionofimprovementshares)indicatesthatthetask
modificationsweimplementedinthisexperimentseemedtohavetheintendedeffectof
clarifyingthecomparisonofpeersolutionsandscores.
3.4.3.Choicestrategiesandcumulativeinnovation
Therelationshipevidentbetweenperformanceandchoicestrategy,inwhichabove‐
averageRetentionandImitationproducehigherscores,whileabove‐averageInnovation
andRetrievalproducelowerscores,reinforcestheevidencefromExperiment1thatthe
overallconservative(butnotregressive)approachnotedaboveisbeneficialforthistask.
However,acounterpointforthisseeminglysimpleresultisprovidedbythecomparisonof
choicesourceproportionsbetweensolutionswhichgeneratedimprovementsandthose
thatdidnot,whichshowedthatareplacementofsignificantamountsofImitationand
RetentionwithInnovationwasrequiredtocreatenewandimprovedsolutions.Thefact
thatsubstantialamountsofeachoftheabovethreechoicesourceswerepresentinsuch
83
improvedsolutionsshowsthatimprovementswerecumulative,relyingonindividuals’own
pastsolutionsaswellasborrowingfromothers.This,inturn,impliesthattheadaptive
valueofImitationinthiscontextisduetoitsfacilitationofselectivelearningandthe
generationofcumulativeimprovementswithlessriskyInnovation(Boyd&Richerson,
1995;Kameda&Nakanishi,2003).
3.4.4.Improvements,free‐riding,andanabsenceoftragedy
Theresultsregardingimprovementsandtheirdistributionwithingroupsindicate
thatparticipantssharedthetaskoffindingbettersolutionsmoreequitablythanifmost
participantswerepursuingapuresocialloafingstrategy.However,thedecreasein
InnovationandincreaseinRetentioninlargergroups,suggestadaptationsbygroup
memberstolimitriskyInnovationtowhatwasrequiredtoachieve“goodenough”results
giventheeffortsofothers.Infact,itmaybethatthelowerendofthedistributionof
Innovationthatactuallyoccurredwasnearlyoptimalfortheverythin‐taileddistributionof
scoresinthespaceofpossiblesolutions(seeFigure23).Only4.3%and1.6%ofpossible
solutionshavehigherscoresthantheparticipants’averagefinalscoreinthelowerand
higherdifficultyconditions,respectively.Thus,althoughtheresultsshowinginequalityin
individualimprovementsharesindicateasubstantialamountoffree‐ridinginthis
experiment,therewasnoassociated“tragedyofthecommons”forInnovations(Hardin,
1968;Ostrom,1990).ThisalsooffersaplausibleexplanationfortheincreasedInnovation
andcorrespondinglowerperformanceofparticipantsinsmallergroups‐‐havingfewer
fellowplayerstocopyfromalsoprovidesfewercluesastothedistributionofpossible
scores,whichpromptsfurtherriskyexplorationatahighercostinaverageperformance.
84
3.5.Experiments1and2–GeneralDiscussion
3.5.1.Factorsthatinfluenceimitation–whenandwhomtoimitate
ParticipantsinbothExperiments1and2displayedfairlyconservativestrategies,
typicallypreservingalargeproportionofeachguessfromroundtoround(throughlow
turnoverandhighretention,respectively).Theentrenchmentdynamicsseeninboth
experimentsshowedthischaracteristicintensifyingovertime,asincreasingproportionsof
guesseswerepreservedinlaterrounds.Asforthechangesthatweremade,Laland(2004)
describesseveralstrategiesthatareobservedacrossawiderangeofspecies.Thecopy
whenuncertainstrategyseemedtobethedefaultatthebeginningofbothexperiments,but
forthosewholearnedtherisksofinnovationfirsthand,thecopywhenasociallearningis
costlystrategywasaverylikelynextresort.Participantsalmostuniversallyemployedthe
copywhenbetterandcopythebeststrategiesaswell,andastrategyoffrequency‐
dependentimitationthatcloselyresembledcopythemajoritywasprevalentinExperiment
2.Participantsalsoshowedapositive“momentumbias”towardimitationofsolution
elementsthatwereincreasinginoverallfrequencyinthegroupratherthandecreasing.
Thisphenomenonhasalsobeenshowntooccurinanexaminationofbaby‐naming
decisionsbyparentsasrevealedby130yearsofsocialsecuritydata(Gureckis&Goldstone,
2009).
Characteristicsoftheproblemspaceandtheinformationenvironmentplayeda
substantialroleinthedynamicsofthesestrategiesandtheirconsequences.Theparticipant
groupsize‐dependentpatternsofimitationanditseffectsinExperiment1(andthelack
85
thereofinExperiment2)showedtheimportanceofclearandsharablesolution
information.
Itisusefultonotethatthetaskdesigninbothexperimentsallowedparticipantsto
pursuehybridstrategieswithinasingleround,inwhichtheyretainedsomepartsofa
solutionwhilechangingothersusingbothsocialandasociallearning.Oneparticularly
interestingwaythiswasaccomplishedwasthroughsimilarity‐biasedimitation.This
allowedtheimitatortomakeuseofsocialinformationwhilekeepingasolutionpartially
compatiblewithprevioussolutionsandexistingknowledgeoftheproblemspace,a
phenomenondiscussedatlengthinrelationtoinnovationpropagationbyRogers(2003).A
biastowardborrowingfromsimilarratherthandissimilarsolutionshasalsobeen
incorporatedintogeneralmachinelearningalgorithmsfeaturingmultipleagents
simultaneouslysearchingforsolutions(Goldberg,1989).Whenagentsborrowsolution
elementsfromotheragentspursuingsubstantiallydifferentsolutions,thereisastrongrisk
thattheresultingblendofsolutionswillbeasub‐optimalhybridnotwelladaptedtothe
nicheofeitheroftheoriginalsolutions.Byanalogy,twosolutionstopredationforasmall
mammalmightbeevolvelargeclawsforclimbtreeseffectivelyortodeveloplargewings
forflying.However,ahalf‐breedthatcombinesbothsolutionsmightwellenactneither
solutioneffectively.Likewise,giventhecomplexproblemsearchlandscapesusedinthe
experiments,participantsmayhavebeenbiasedtocopysolutionelementsfromsimilar
ratherthandissimilarsolutionstoensuregreatersolutioncompatibility.
86
3.5.2Group‐leveleffectsofimitationandinnovation
TheresultsofExperiment1suggestthatimitationcanbeindividually
counterproductivewhensocialinformationcannotbereadilycompared,integrated,and
adaptedtocreatenewsolutions.Thissituationmayariseforseveralreasons:(1)
evaluativeinformationiseitheractuallyunavailable,orpresentedinsuchawaythat
cognitiveloadeffectshinderitsuse;(2)theproblemspaceistoolargeorcomplexforparts
ofdifferentsolutionstobeeasilyanalyzedandcombined;or(3)theproblemspaceistoo
largeforsubstantialchangestobeeffectivelyevaluated.
TheresultsofExperiment2showthatwhentheabove‐mentionedproblemscanbe
avoidedorameliorated,imitationcanbeproductiveforindividualsaswellasgroups,
becauseitenablesthepreservationofgoodtentativesolutionsin“groupmemory”and
theirfurtherimprovementthroughcumulativeinnovation.Theseresultsalsoshowedthat
therisksofinnovationcanoutweighthebenefitsforbothindividualsandgroups,andthus
becomecounterproductivewhenusedtoomuch.Obviously,acompletelackofinnovation
willresultinalackofimprovements(becausestrategiesthatcombineimitatedelements
fromdifferentmodelswilloftenlosebeneficialinteractions),butthisexperimentsuggests
thatinalargeandcomplexproblemspace,atboththeindividualandgrouplevels,
innovationisbestusedsparingly,alongwiththeretentionofpreviousgoodsolutionsand
imitationofothers,toimproveoveralloutcomeswhilemaintaininghighaverage
performance.
Thesetworesultstakentogetherwiththereductionsindiversityovertimeimplya
viewthatisatoddswiththosepredictedfromasimpleTragedyoftheCommons(Hardin,
1968)orproducer‐scroungerdilemma(Kameda&Nakanishi,2002)interpretationof
87
sociallearning.Muchlike“conformist,”beinga“scrounger”oftencarriesanegative
connotationordenotation,suchas“socialloafing”(Latané,Williams,&Harkins,1979).
However,suchbehaviormaybeappropriatewhennotallgroupmembers’fulleffortsare
requiredtoproducesufficientbenefit.Inacomplexbutrelativelystableenvironment,the
bestoutcomeforthegroupmayresultfrommostgroupmembersconvergingona“good
enough”solutionquicklytoachievehighmeanperformance,andthenintroducing
productiveinnovationswhenpossible.Thus,insomecircumstancesatragic(oratleast
distinctlysuboptimal)outcomecanresultfromtoomuchinnovationandnotenough
imitation,ratherthantheotherwayaround,becauseinnovationisriskyandpossibly
redundant(andthuswastefulofresources),andimitationhelpstoconcentrateeffortsand
improvethethoroughnessofsearchintheproximityofknowngoodsolutions.Givensome
baselineinclinationtoaminimumamountofindividualexploration,thelimitingfactorin
improvingsearchperformancemaybetheamountofinformationsharingandcoordination
amongsearchers,whichallowthemtopoolboththebenefitsandtherisksofasocial
learning(Hess&Ostrom,2007).
Ofcourse,theresultsobtainedherearelikelytobehighlydependentonthe
problemspaceandtheinformationenvironmentinuse.Thoughthebenefitsofsocial
learninginatemporallystableenvironmentareoftenassumedtobeevident(e.g.Kameda
&Nakanishi,2002),thisstudyilluminatessomedetailsaboutthedynamicindividualand
group‐levelmechanismsbywhichthesebenefitscanaccrue(ornot,insome
circumstances).Givenimperfectindividualmemory,the“culturalknowledgepool”
(Kameda&Nakanishi,2003)requiresnotonlyprovisionofinformationbyasociallearning,
butalsoitsamplificationandpreservationthroughmechanismslikefrequency‐biased
88
adoption,or“conformisttransmission”(Boyd&Richerson,1985).Ourtask,inwhich
solutionshavemultiplecomponentswithepistaticrelationships,alsoallowedusto
examinehowsuchsolutionsarebuiltcumulativelyusingselectivelyvaryingproportionsof
differentinformationsources.Thisaddsrealisticcomplexitybeyondthatprovidedby
modelsandexperimentalsettingswithsimplerproblemstructuresorlessflexiblelearning
strategies.
89
4.ExplorationasaSocialDilemma
4.1Dilemmasofknowledge
Sofarwehavefoundevidenceforseveralimportantlearningstrategiesdiscussedin
previousmodelsandanimalresearch.Wehavealsoexploredtheideathatcollectivesearch
problemsliketheonesinourexperimentaltasksarepotentiallysusceptibletosocial
dilemmas(thoughwehavenotobservedevidenceforsuchadilemma).However,wehave
notsufficientlyelaborateduponthecharacteristicsofthe"resource"involved,howitis
createdandmaintained,orhowitmightbedegradedordestroyed.Atthispointitwillbe
usefultoreturntotheliteratureonsocialdilemmastopursuethisfurther.
4.1.1.Storiesandassumptions
Socialdilemmascanbesaidtooccur"wheneverindividualsininterdependent
situationsfacechoicesinwhichthemaximizationofshort‐termself‐interestyields
outcomesleavingallparticipantsworseoffthanfeasiblealternatives"(Ostrom,1998).
Contemporaryexamplesoflarge‐scalesocialdilemmasincludethefailuretolimitthe
productionofatmosphericpollutants,andtheoverharvestingofoceanfishstocks.Butsuch
dilemmascanariseinmanyeverydaysituations,fromtrafficjamsatrushhourtofibbingat
taxtime.Socialdilemmashaveoftenbeenframedintermsofthreemetaphoricalstories:
thePrisoner'sDilemma(Poundstone,1992),theTragedyoftheCommons(Hardin,1968),
andtheLogicofCollectiveAction(Olson,1965).
ThePrisoner'sDilemmaisbasedonahypotheticalsituationinwhichtwo
individualsaccusedofcommittingacrimetogetherarequestionedseparatelyandoffered
twochoices:confesstothecrimeorkeepsilent.Ifbothkeepsilent,theauthoritiescanonly
90
convictthemonminorchargesandthepenaltyforbothismild.Ifbothconfess,theyare
convictedforthecrimeandreceivesubstantialjailsentences.Ifonlyoneconfesses,hegoes
freewhiletheotherreceivesaharshsentence.Thissetofpossibilitiesissuchthat
whatevertheotherdoes,therational(ordominant)optionforeachindividualistoconfess.
However,ifbothindividualsfollowthisreasoning,thebestoptionforboth(keepingsilent)
isunavailable.(Itmaybethatwedonotwantcriminalstobeabletoescapesucha
dilemma,butthissituationcanbere‐framedinanynumberofotherways,e.g.anarmsrace
betweencountries,steroiduseamongcompetitiveatheletes,etc.)Theactionsofthetwo
agentsareinterdependentsuchthateithercanindependentlyincreasehisownpayoffat
theexpenseoftheother,butthebestoutcomecanonlybeobtainedifbothhavesomeway
totrusttheotherandjointlyincreasetheirpayoffs.
TheTragedyoftheCommons(Hardin,1968)isanotherhypotheticalsituation,in
whichapasture(orcommons)isavailableforusebymanyindividualstograzetheir
animals.Eachadditionalanimaladdedtothecommonsincreasesitsowner'spayoffinthe
shortterm,butreducestheamountoffodderavailableforalloftheotheranimals.Inthe
longterm,overgrazingcanleadtotheruinationofthepasture,sothatnofodderis
available.Themarginalbenefitofaddinganotheranimalisobtainedbyoneindividual,but
theshort‐termcostisspreadacrossallindividualsusingthecommons(anegative
externality).Thismeansthattheshort‐termbenefitsoutweighthecostsforeach
individual;ifallindividualsfollowthisreasoningsymmetrically,thelong‐termbenefits
(continuedgrazingcapacity)areunavailable.Again,thismodelisnotonlyimportantfor
cattlefarmersinagrarianvillages,asitcanbere‐framedtofitmanyothersituationsmore
relevanttomodernsociety(e.g.usinglawnsprinklerssuppliedbyacommonreservoir
91
duringadrought).Onceagain,theimportantpointisthattheactionsofmultipleagentsare
interdependentsuchthateachcanincreasehisorherownshort‐termpayoffatthe
expenseofothers,butthebestlong‐termoutcomecanonlybeobtainedifallfindsomeway
tocooperate.
TheLogicofCollectiveAction(Olson,1965)portraystheconflictinherentbetween
theself‐interestsofalargenumberoffirmssellingthesamegood.Themarketforthisgood
ispresumedtobeperfectlycompetitive,butina"disequilibrium"state,sothatprice
exceedsmarginalcostforallfirms.Eachwouldliketomaximizeprofitsbysellingasmuch
ofthegoodastheycanproduceatthehighestpricepossible.However,(undercertain
assumptionssuchashighbarrierstomarketentry,inelasticdemand,andsoforth),firms
thatincreaseproductionwillsellmorebutalsoincreasetheoverallsupply,whichwill
eventuallylowerprices.Anyfirmthatattemptstounilaterallyrestrictitsownoutputin
ordertolowersupplywillsimplyreduceitsrevenues,giventhatnofirmhasenough
marketpowerindividuallytoaffectprices.Allofthefirmshaveacommoninterestin
higherprices,buteachwouldrationallypreferthattheothersbearthecostoflowered
productioninordertogetit.(Onceagain,ifonehastroublefeelingempathyforthefirms
depictedhere,theframingcanbechangedtothatofworkerswhocannotallearna
sufficientwageifeachoffershisfullcapacitytoexploitativeemployers.)InOlson’s
formulation,theonlywaytokeeppriceshighisthroughsomeexternalintervention,such
asgovernmentpricesupports,tariffs,orcartelagreements,thecreationofwhichrequires
costlylobbyingororganization.Assumingthatthefirmssomehowforeseetheproblemand
bandtogethertocollectivelycreatetheprice‐supportingintervention(aslightdeparture
fromtheprevioustwomodels),aproblemremains.Becausethemarketforthegoodis
92
indivisible,eachfirmonceagainhastherationalincentivetoenjoythebenefitsofsuch
interventionwithoutcontributingtocreateormaintainit,ifsuchcontributions(lobbying
firmfeesorunionmembershipdues)arevoluntary.Thusthefirst‐orderdilemma
(agreeingtoorganizetocontrolprices)isatleasttemporarilysolved,butthesecond‐order
dilemma(maintainingthecostlyorganization)isvulnerabletoa“free‐riding”problem.
Eachoftheabovethreemodelsillustratesanimportantpointaboutsocial
dilemmas,andthenontrivialityoftheirsolutions.Thoughtheycaneachbeadaptedto
resemblefamiliarsituations,problemsarisewhenthesemodelsaretakentoberealistic
representationsofhumanbehavior.Withoutdenyingthatsocialdilemmascanhavetragic
outcomes,itmustbenotedthatthestoriesabove(andthetheoriestheyrepresent)involve
severaloverlysimplisticassumptionsaboutsocialdilemmas:(1)resourceusersarestrictly
selfishmaximizersofshort‐termgains,whowillnotcooperatetoovercomeasocial
dilemma,(2)itisarelativelysimpleanalyticaltasktochangetheincentivesofresource
usersbydesigningnewrules,and(3)centralizeddirectionandcoercionisrequiredto
successfullyovercomesocialdilemmas(Ostrom,1999).Ingeneral,thethreeassumptions
abovehavenotbeenborneoutbydatagatheredonhumanresponsestosocialdilemmas,
inthelaboratoryorinthefield.Peoplecaninfactrecognizethedilemmastheyface,and
changerulesandincentivestoavoidthem(Ostrom,1990).However,doingsocanbequite
acomplexanduncertainprocess,involvingmuchtrialanderror,andmaynotalwaysbe
successful(Sandberg,2001).Finally,centralizedsolutionshavegenerallynotbeenas
successfulasthosedesignedusingtheknowledgeandparticipationoflocalresourceusers
(Sneath,1998).
93
4.1.2.Typesofgoods
Socialdilemmasdealwithpotentialconflictsbetweenagentsoveravaluable
resourceorgood.Twopertinentdimensionsoverwhichgoodscanvaryaresubtractability
(whethertheuseofsomeportionofthegoodbyoneindividualprecludesitsuseby
others),andexcludability(theextenttowhichpotentialuserscanbepreventedfromusing
theresource).Subtractabilityisgenerallyabinarycharacteristic,andsubtractablegoods
arealsoknownasrivalrousgoods,becausepotentialusersarenecessarilyrivalswhenany
portionofagoodcanonlybeusedbyoneofthem,butisdesiredbyall.Whena
subtractablegoodisscarceandnomeasuresareinplacetoprioritizeitsuseamong
individuals(suchaspropertyrightsorotherrules),theexpectedresultisincreased
scarcityandconflict.However,thecharacteristicofsubtractabilitycanbeusedtomonitor
theuseofagood,whichcanhelpinencouragingorenforcingcooperationinadilemma.
Excludabilityisamorecontinuouscharacteristic,andneednotbeabsoluteinordertobe
effective.Physicalandtechnologicalmeansplayarole,butlegal,social,andculturalnorms
areimportantaswell;conceptsoffairness,justice,andtraditioncanallaffectthe
excludabilityofagood.
"Privategoods"arethosewhicharebothsubtractableandexcludable,suchasatool
oraloafofbread."Clubgoods"arethosewhichareexcludablebutnotsubtractable,suchas
accesstocopyrightedworks,orascenicviewfromfencedland.Privateandclubgoodsare
generallypresumednottoentailseriousdilemmas,becausetheirownersorproviderscan
excludepotentialusers,anduserscanavoidthosewhichareinefficientlyprovided.
Goodswhicharesubtractablebutforwhichexclusionisimperfectorcostlyare
knownas"commongoods"or"commonpoolresources"(CPRs).Instancesoftheseinclude
94
fisheries,forests,andthecanonicalexampleofpasturelands.Common‐poolresourcesare
typicallyself‐renewingatsomefiniterate,butadilemmamayobtainifindividual
incentivesdonotsustainappropriationoftheresourceandavoidoveruseordestruction.
TheyhavetypicallybeenthoughtofaccordingtotheTragedyoftheCommonsmodelas
describedabove.
Goodsthatareimperfectlyexcludableandnon‐subtractableareknownas"public
goods."Instancesoftheseincludegenerallyintangibleorindivisibleitemssuchas
televisionbroadcastsandthelightfromlighthouses,respectively.Theprimarypotential
dilemmaforpublicgoodsisthe"free‐riding"problem:howsuchgoodscanbesufficiently
providedandmaintained,sincetheycanbeusedbyindividualswhodon'tcontributeto
theirprovision(freeriders).Aconnectedproblemishowtoassurethosewhowishto
contribute(butdon'twanttowastetheircontribution)thatotherswilldosoaswell,
insteadoffree‐riding(theassuranceproblem).Thisdilemmaisoftenmodeledusingthe
LogicofCollectiveActionparadigmdescribedabove.
PublicGoodsandCommonPoolResourcesareconceptuallyandnaturallyrelated,in
thataCPRoftenrequiresaninitialorcontinuinginvestmentintheprovisionofinstitutions
andinfrastructurerelatedtoappropriation.Thisinvestmentallowscurrentandfuture
userstoparticipateintheappropriationoftheresource;thatis,theinputsthatcreatethe
benefitoftheinstitution'scontinuedexistencearerivalrouseveniftheresourceitselfis
not.Themoregeneralclassesofprovisionandappropriationproblemsaretypically
modeledandtreatedseparatelyinordertoelucidateusefulcharacteristicsinalleviating
eachtypeofproblem(Ostrom,Gardner,&Walker,1994).
95
4.1.3.Knowledgeasacommons
Thetaskofunderstandingknowledgeasagoodaccordingtotheabovetaxonomy
canbeabittricky,becauseknowledgecantakevariousformsthataresubjecttodifferent
usesandnorms.HessandOstrom's(2007)treatmentof"KnowledgeasaCommons"begins
bydefining"knowledge"as"alltypesofunderstandinggainedthroughexperienceor
study"(p.8).Inthisintangibleform,knowledgeappearstobeapurepublicgood.Firstly,it
isquitedifficulttoexcludepotentialusers:asdiscussedinpreviouschapters,humanshave
aprodigioustalentandinclinationforgainingknowledgefromeachother,andoncea
particularideahasbeenunderstood,itisdifficulttoun‐understand(or"derstand"(Currie,
2010))itbyanyreliablemeans,voluntaryorcoercive.Secondly,oneperson's
understandingofanideadoesnotprecludeorsubtractfromanyotherperson's;if
anything,thesharingofunderstandingcanreinforceandpreserveit.
Ofcourse,peoplehavefoundwaystoencodeandrecordthisintangible
understandinginvariousphysicalforms(e.g.paintedtombwalls,incisedclaytablets,
handwrittenpaperscrolls,printedbooks,pressedphonographrecords,celluloidfilm,and
documentsanddatabasesinanynumberofdigitalmedia)inordertoarchiveitandpassit
ontonewpotentialunderstanders.Thevaryingdegreestowhichthesetoolsfor
instantiationofknowledgearesubjecttophysicalandlegalconstraintiswhatcomplicates
thegovernanceofknowledgeasaresource.
Asphysicalobjects,printedmatterandaudiovisualrecordingscanbeprivately
ownedbyanindividualandmadeinaccessibletoothers,orheldbyapubliclibrary
collectionforfullandfreeusebyonepersonatatime,orkeptinamuseum'sdisplaycase
wheretheyareobservablebutnotfullyusable.Images,text,andsoundcanbeencoded
96
digitallyandcopiedmoreorlesseffortlessly(thoughtheymustalwaysbestoredinsome
physicalform).
Despitethenon‐rivalrousnatureofintangibleknowledge(andastransmissionand
storagebecomecheaper,perhapsanydigitizableknowledge)asapublicgood,some
resourcesrequiredtocreateit(money,tools,etc.)arecertainlyrivalrous.Totheextentthat
knowledgedependsonmorethanonepersonforitsprovision,theclassicalassurance
problemofcollectiveactionobtains.Totheextentthatthecreationofnewknowledge
dependsontheuseofpreviouslyexistingknowledgeand(1)thephysicalinfrastructureof
knowledgestorageandaccessrequiresresourcesformaintenance,or(2)knowledgecan
beappropriatedandaccessprevented,dilemmasofunderprovisionandoverappropriation
becomepossible.
Finally,ifa"commons"isdefinedsimplyasa"asharedresourcethatisvulnerable
tosocialdilemmas"(Hess&Ostrom,2007,p.13),itiseasytosee"knowledgeasa
commons."ForHessandOstrom(2007),analysisofanycommonsrequiresexaminations
ofequity("issuesofjustorequalappropriationfrom,andcontributionto,themaintenance
ofaresource"),efficiency("optimalproduction,management,anduseoftheresource"),
andsustainability("outcomesoverthelongterm").Wewillreturntothesethreecriteriato
judgeresultsofourexperimentsandthoseofotherstudies.
4.2.Influenceofinformationenvironmentonsociallearningstrategies
Sofarwehavenotfoundevidencefora"tragic"outcomeinunderprovisionof
individualexplorationinourparticipants'collectivesearchbehaviorovertime,extreme
inequityinprovision,orunsustainabilityinlonger‐termresults.Thoughparticipantswere
97
ratherconservativeintheirstrategies,innovationwasmorecommononaveragethan
imitation,implyingarelativelackofharmful"freeriding."Improvementsharewasfairly
equitablydistributedwithingroups,andfinalscoreswerequitehighintermsof
percentilesinthedistributionofpossibleoutcomes.However,participantsgenerallydid
notreachoptimalsolutions,sotherecouldbepossibilitiesforimprovingefficiencyand
absoluteperformancethroughchangesinparticipants'incentivesandinformation
environment.Theremayalsobeimportantfactorswhichcouldharmparticipants'abilityto
contributetotheproductionandsharingofinnovations.
4.2.1.Inefficiencyandinformation
Previousworkhasshownthatmakingsociallearningprocesseslessefficientcan
actuallyimprovethelong‐termperformanceofgroupsattemptingtosearchacomplex
problemspace,bothinlaboratoryexperiments(Mason,Jones,&Goldstone,2008)and
agent‐basedmodels(Lazer&Friedman,1997).Thereductioninefficiencywas
implementedinthesestudiesbychangingthedistributionoflinksinthesocialnetwork
connectingparticipants.Thishadtheeffectofslowingthespreadofinformationviasocial
learning,encouraginggreaterexploration,andmaintaininggreatersolutiondiversity
amonggroupmembers,thusavoidingprematuresettlingofthegroup'ssearchongoodbut
suboptimalregionsoftheproblemspace.Havingobservedsomegeneraldynamicsand
learningstrategiesofcollectivesearchintheexperimentsdescribedsofar,wewishedto
seewhetherthisseeminglyparadoxicaleffectwouldoccurwithconceptuallysimilar
changestoourexperimentalparadigm.
98
Ourpreviousexperimentshavealsoshownthatthesociallearningstrategiesthat
participantspursuedependontheinformationavailabletoeachparticipant,specificallyin
termsofthenumberofotherparticipantssharinginformationabouttheirtentative
solutions.Presumably,thisisbecauselargernumbersofpeerssharinginformation
increasetheoverallreliabilityoftheinformationavailabletoeach.Aspreviouslydiscussed,
thiscanbeinterpretedasaninstanceofthecopywhenuncertainsociallearningstrategy
(Laland,2004),becausegreatercertaintyaboutthevalidityofsocialinformationincreases
therelativeuncertaintyofasociallearning.Otherworkondiffusionofinnovationshas
shownthattheuseandinfluenceofsocialinformationdependontherelativeambiguityof
thebenefitofadoptingaparticularsolution(Granovetter,1978;DiMaggio&Powell,1983;
Abrahamson&Rosenkopf,1997;Gibbons,2004).Inamodelofsocialforagingbehaviorina
changingenvironment,KrebsandInman(1992)showedthatifthereisadelayinan
observer’srecognitionoftheforagingsuccessofademonstrator,thereisacorresponding
reductionintheinformation(andthusthebenefit)providedbythesocialinformation
providedbythedemonstrator;withalongenoughdelay,theobserverisbetteroffignoring
thedemonstratorandforagingasocially.Anotherresultfrommodelingisthatifpersonal
(asocial)andsocialinformationcannotbegatheredsimultaneously,theremaybeno
benefittousingsocialinformation,becausetheycannotbeeffectivelycombined;
furthermore,whenothers’actionsareavailabletobeobserved,butnottheresulting
success‐relatedcues,socialinformationshouldbeminimizedtoavoidinformation
cascades(Giraldeau,Valone,&Templeton,2002;Bikhchandani,Hirshleifer,&Welsh,
1992).
99
4.2.2.MotivationandconceptualdescriptionofchangesinExperiment3
Havingshowninpreviousexperimentsthatparticipantscanuseavarietyof
learningstrategiestosolveacollectivesearchproblem,wewantedtoseehowthese
strategieswouldshiftinresponsetochangesintheinformationenvironment.Ratherthan
changingthestructureofthesocialnetworkconnectingourparticipants(asinthework
citedjustabove),wedecidedtosimplyremoveaccesstoscoreinformationaboutothers'
solutions.Participantswouldstillbeabletoobserveandcopytheirpeers,butwouldhave
towaitfortheirownfeedbacktolearnthevalueofimitatedchoices;informationabout
others'solutionswasentirelyambiguous,whileinformationabouttheirownsolutionswas
entirelyunambiguous.Thisissimilartocontextsinwhichindividualanimalscanobserve
thebehaviorofothers,butnotthecuesoroutcomesthatmotivatethoseactions,or
situationsinwhichprivatefirmscanwithholdinformationaboutrevenues(i.e.thesuccess
oftheiractions)fromcompetitors.
Wewishedtousethisrelativelysmallmodificationtoaskquestionsabouthowthe
eliminationofinformationsupportingonekindofsociallearningwouldcauseparticipants
tochangetheirlearningstrategies,andhowsuchchangeswouldaffectperformance.Asthe
workaboveonimprovedcollectivesearchperformancethroughreducedcommunication
efficiencyindicates,impedingsociallearningmayincreaseasociallearningamongmost
participants.Suchanoutcomewouldpresumablyreducefree‐riding(improvingequityof
provision);butwouldthischangecauseparticipantstoactuallyfindbettersolutions,or
wouldthereducedefficiencyofaccesstopreviousknowledge(andtheinabilitytouseit
simultaneouslywithasocialinformation)reducetheefficiencyofknowledgeproduction?
Presumingthatsociallearningisnotdisplacedentirelybyasociallearning,howwouldits
100
prevalenceandusechange,andwhatinformationwouldbeusedtodirectremaining
imitationchoices?Finally,wouldtheseadaptationsinbehaviorimproveoverallindividual
andcollectiveoutcomesandimprovementsovertime(sustainability)?
Inordertoavoidceilingeffectsonperformanceandpresumablymakeiteasierto
distinguishtheperformanceofsuccessfulsocialandasociallearning,wechangedthe
problemspaceslightlyfromthepreviousexperimenttoshiftsomeofthemassofthescore
distributiontoalongerandfatteruppertail(increasingtheproportionofsolutionswith
higherscores).Thisallowedtheparalleldiscoveryofhigher‐scoringsolutionsamong
participantswithoutrequiringasmuchconvergenceofsolutioncontent;performance
couldbeincreasedwithoutnecessarilyconstrainingsolutiondiversity.
101
5.Experiment3–CreatureGameB(scorevisibility)
5.1.Experiment3Overview
Thetaskusedinthisexperimentwasthesameasthe“CreatureGame”of
Experiment2,withtwomajorchanges:(1)thescoresassociatedwithpeers’solutionswere
showninhalfthegamesineachexperimentsession,andhiddenintheotherhalf;(2)the
problemspacewaschangedbyaddingmorepositive‐scoringbonusinteractionsbetween
solutionelements,whichhadtheeffectofmakingtheuppertailofthescoredistribution
longerandfatter,sothattherewererelativelymoresolutionswithhighscores.
Modification(1)allowedfortheexaminationofdifferencesinstrategiesandperformance
associatedwithdifferencesintheavailablesocialinformation.Modification(2),thoughit
madedirectcomparisonswithExperiment2slightlymoretenuous,allowedparticipantsto
achievehighscoreswithoutnecessarilyconvergingintheirsolutions.
5.1.1.Predictions
Becauseofthesmallchangeswemadetothetask,weexpectedthatresultswould
bequitesimilartothoseofExperiment2whenpeers’scoreswereshown.Whenevaluative
informationaboutpeersolutionswasunavailable,participantswouldbeunabletobe
sufficientlyselectiveinimitation,andthusparticipantsemployinghighlyimitative
strategieswouldhaverelativelylowerscoresthanthosewithlessimitation‐heavy
strategies,participantswouldemploylessimitationandmoreinnovation,andsolution
diversitywouldincrease.Similarity‐biasedandfrequency‐biasedimitationstrategies
wouldbestrongerwhenpeerscoreswereinvisible,inordertocompensateforthelackof
directevaluativeinformation.Overall,ratherthanimprovingexplorationbehavior,the
102
impedanceofsociallearningbymakingpeerscoresinvisiblewouldresultinlowermean
scores(includingthoseofrelativelysuccessfulasociallearners)becausetheywouldbe
unabletoeasilytakeadvantageofgoodsolutionsfoundbyothersthroughselective
imitationandfurtherimproveuponthem.
5.2.Experiment3Methods
234participantswererecruitedfromtheIndianaUniversityPsychology
Departmentundergraduatesubjectpool,andweregivencoursecreditfortakingpartinthe
study.Participantspopulatedeachsessionbysigningupatwillforscheduledexperiments
withamaximumcapacityof9persons,andweredistributedacross65sessionsasshown
inTable5.1.
Table5.1:DistributionofparticipantsacrossgroupsizesinExperiment3
Groupsize 1 2 3 4 5 6 7 8 9
#Sessions 16 8 11 11 7 2 4 4 2
#Participants 16 16 33 44 35 12 28 32 18
ThetaskusedwasnearlyidenticaltothatofExperiment2,withthefollowing
changes.Tomoreeasilyfitthesessionintheone‐hourtimelimitrequiredforexperiments
usingoursubjectpool,thereweresixgamespersessioninsteadofeight.Inthreeofthese
games(theinvisiblescorescondition),thescoresofotherparticipantswerenotshown
alongwiththeirsolutionsfromthepreviousround;intheotherthreegames(thevisible
103
scorescondition),otherparticipants’scoreswereshown.Ofcourse,thisdistinctiononly
matteredinsessionsthatincludedmorethanoneparticipant.
Thedistributionofindividualpointvaluesfortheiconswasthesameasforthe
largerleaguesizeinExperiment2,butsevennewpositivebonusinteractionswereadded
betweenicons,andseveralexistinginteractionvalueswereshiftedtodifferentpairsof
icons,asshowninFig.5.1(comparetoFig.3.2b).Thesechangeshadtheeffectofincreasing
thecomplexityoftheproblemspace,aswellasincreasingthenumberofpossiblehigh‐
scoringteams.Asaresult,thepossiblescorerangechangedto[‐6,88],butallscoreswere
againnormalizedtotherange[0,1]foreaseofanalysis.(Notethatduetothisshift,
normalizedscorescannotbedirectlycomparedbetweenExperiments2and3.)The
combinationsoftheseindividualandpairvaluesresultedintheprobabilitydistributionof
scoresamongallpossibleteamsshowninFig.5.2(comparetoFig.3.3b).
Figure5.1:Pointdistributionforindividualicons(boxes)andinteractionbonusesand
penalties(ovals).
104
Figure5.2:Distributionofscoresforallpossibleteams.
5.3.Experiment3Results
5.3.1Overallmeans
MeandependentvariablesineachconditionareshowninTable5.2(seealsoFig.
5.2).Ofallgroupedparticipants,81.7%hadhighermeanscoresinthevisible‐scores
conditionthanintheinvisible‐scorescondition(seeFigure5.3).Isolatedparticipants
achievedmeanoverallandfinalscoresof.356and.395.
105
Table5.2.Meanscore,guessdiversity,andchoicesourceproportionsbycondition
Cond.OverallScore(Percentile)
FinalScore(Percentile)
GuessDiversity
Imitation Innovation Retention Retrieval
Visible .447*(92.6%) .523*(97.5%) 62.2%* 9.1% 13.4%* 75.0%* 2.4%*
Invis. .394*(89.1%) .475*(94.6%) 79.0%* 8.4% 15.0%* 71.0%* 4.2%*
IsolatedPartic. .356(81.4%) .395(89.1%) ‐‐ ‐‐ 29.2% 55.8% 13.7%
*significantdifferencesbetweenconditions
Figure5.3:Scattergramofindividuals’meanscoresineachcondition,labeledwiththeir
participantgroupsize.
106
5.3.2.Rounds
Linearmixed‐effectsmodelswereusedtoexaminetrendsacrossroundsforscore
andguessdiversity,witharandomeffectofparticipantgroup.Analysisofscoreversus
roundshowedastrongpositivetrendforgroupedparticipantsinthevisible‐scores
condition(F(1,1126)=521.82,p<.0001,B=.656,meantotalincrease=0.220),andaslightly
shallowerpositivetrendintheinvisible‐scorescondition(F(1,1126)=446.53,p<.0001,
B=.727,meanincrease=0.172;seeFig.5.4).Guessdiversityshowedasimilarlystrong
decreaseacrossroundsinthevisible‐scorescondition(F(1,1126)=304.78,p<.0001,B=‐
.443,meanchange=‐0.468),andaweakerdecreaseintheinvisible‐scorescondition
(F(1,1126)=97.31,p<.0001,B=‐0.453,meanchange=‐0.271;seeFig.5.4).Isolated
participants’scoresincreasedmuchless(thoughsignificantly)acrossrounds
(F(1,751)=30.26,p<.0001,B=.348,meanincrease=0.075;seeFig.5.4).
107
Figure5.4:Scoresincreasedandguessdiversitydecreasedmoreacrossroundsinthe
scores‐visibleconditionthaninthescores‐invisiblecondition.
Trendsforchoicesourcesacrossroundsshowedverysimilarpatternstothosein
Experiment2:ImitationandInnovationdecreasedsignificantly,andRetrievaland
Retentionincreasedsignificantly.Therewerenosubstantialdifferencesinslopesbetween
conditions,norsubstantialdifferenceswiththeslopesoverroundsfoundinExperiment2.
5.3.3.Gameorder
Similarlinearmixed‐effectsmodelswereusedtoexaminetrendsacrossgameorder
foreachdependentvariablewithinconditions.Forthisanalysis,thegameordervaluefor
eachgamewascorrectedtoitsorderwithinthecondition,i.e.Game1,2,or3ineach
condition.Scoredisplayedaslightbutsignificantincreaseacrossgameorderinthevisible‐
scorescondition(F(1,97)=16.57,p=.0001,B=0.264,meanchange=+0.048;seeFig.5.5),and
asmallbutnon‐significantincreaseintheinvisible‐scorescondition.Guessdiversity
displayedacorrespondingdecreaseacrossgameorderinthevisible‐scorescondition
(F(1,97)=62.09,p<.0001,B=‐0.263,meanchange=‐0.103),aswellasintheinvisible‐scores
condition(F(1,97)=70.79,p<.0001,B=‐0.363,meanchange=‐0.115;seeFig.5.5).
Asforchoicesources,changesovergameorderwithinconditionsweregenerally
slight.Imitationdecreasedslightlybutsignificantlyovergameorderinthevisible‐scores
condition(F(1,97)=6.24,p=.0141,B=‐0.121,meanchange=‐0.011),whileincreasing
significantlyintheinvisible‐scorescondition(F(1,97)=32.97,p<.0001,B=0.289,mean
change=+0.040;seeFig.5.6a).Innovationdecreasedsignificantlyacrossgameorderwithin
108
boththevisible‐scores(F(1,97)=26.82,p<.0001,B=‐0.173,meanchange=‐0.028)andthe
invisible‐scores(F(1,97)=41.04,p<.0001,B=‐0.226,meanchange=‐0.036;seeFig.5.6a)
conditions.Retentionincreasedsignificantlyonlyinthevisible‐scorescondition
(F(1,97)=29.97,p<.0001,B=0.252,meanchange=+0.041;seeFig.5.6b),andRetrieval
increasedsignificantlyonlyintheinvisible‐scorescondition(F(1,97)=14.42,p=.0004,
B=0.228,meanchange=+0.012;seeFig.5.6b).
Figure5.5:Scoreincreasedsignificantlyonlyinthevisible‐scorescondition,andguess
diversitydecreasedinbothconditions.
109
(a) (b)
Figure5.6:(a)ImitationandInnovationdecreasedsignificantlyinthescores‐visible
condition,whileImitationincreasedandInnovationdecreasedinthescores‐invisible
condition.(b)Retentionincreasedsignificantlyonlyinthevisible‐scorescondition,and
Retrievalincreasedsignificantlyonlyintheinvisible‐scorescondition
5.3.4.Groupsize
Trendsacrossparticipantgroupsizeforeachdependentvariablewithinconditions
wereexaminedusinglinearmixed‐effectsmodels,withtheparticipantgroupusedasa
randomeffectontheintercept.Scoreincreasedsignificantlywithgroupsizeinthevisible‐
scorescondition(F(1,63)=79.59,p<.0001,B=0.580),aswellasintheinvisible‐scores
condition(F(1,63)=15.45,p=.0002,B=0.309;seeFig.5.7),thoughthelattertrendwasnot
asstrong.Guessdiversityshowedacorrespondingdecreasewithincreasinggroupsizein
thevisible‐scorescondition(F(1,47)=68.83,p<.0001,B=‐0.699),aswellasaweakertrend
intheinvisible‐scorescondition(F(1,47)=17.28,p=.0001,B=‐0.430;seeFig.5.7).
110
Asforchoicesources,Imitationincreasedsignificantlyforlargergroupsinboththe
scores‐visible(F(1,47)=41.47,p<.0001,B=0.597)andscores‐invisible(F(1,47)=28.04,
p<.0001,B=0.500;seeFig.5.8a)conditions,andInnovationdecreasedsignificantlyfor
largergroupsinboththescores‐visible(F(1,47)=18.42,p=.0001,B=‐0.492)andscores‐
invisible(F(1,47)=14.04,p=.0005,B=‐0.436;seeFig.5.8a)conditions.Retentionincreased
forlargergroupsonlyinthescores‐visiblecondition(F(1,47)=5.91,p=.019,B=0.286;see
Fig.5.8b),whileRetrievalshowednosignificanttrendacrossgroupsize(seeFig.5.8b).
Figure5.7:Asparticipantgroupsizeincreased,meanscoresinagroupincreased,andthe
diversityofofferedsolutionsdecreased,withslightlyweakereffectsforbothinthe
invisible‐scorescondition.
111
(a) (b)
Figure5.8:Asparticipantgroupsizeincreased,(a)meanproportionsofImitationincreased
andInnovationdecreasedinbothconditions,and(b)Retentionincreasedonlyinthe
visible‐scorescondition,andRetrievalshowednosignificantchangeacrossgroupsize.
5.3.5.Differencesinimitation
Ofallinstancesofsingle‐participantimitation,thescoreoftheimitatedparticipant
wasgreaterthanthatoftheimitatorsignficantlymoreofteninthevisible‐scorescondition
thanintheinvisible‐scorescondition(t(74)=16.07,p<.0001;seeFig.5.9a);inthelatter
condition,theprobabilitywasabout54%,orapproximatelyatchance.Inaddition,there
wasasignificantlygreaterprobabilityofimitatingthetop‐scoringsolutioninthegroupin
thevisible‐scorescondition(t(80)=20.08,p<.0001;seeFig.5.9b).
Toexamineseparatelyhowoftenandhowmuchparticipantsimitatedoneanother,
wemeasuredthemeanproportionofguessesinwhichtherewasgreaterthanzero
Imitation(Imitationincidence),aswellasthemeanImitationproportioninsuchcases
(Imitationproportion).MeanImitationincidencewassignificantlyhigherinthevisible‐
112
scorescondition(F(1,229)=31.17,p<.0001),butthedistributionofmeanimitation
proportionswasweightedsignificantlymoreheavilytowardhighervaluesintheinvisible‐
scorescondition,asshownbyaKolmogorov‐Smirnofftestofequalityofdistributions
(D=0.1893,p<.0001;seeFig.5.10).Inotherwords,participantsinthescores‐invisible
conditioncopiedoneanotherlessfrequentlybutinlargeramountsatatime.
(a) (b)
Figure5.9:Inthevisible‐scoresconditiontherewerestrongbiasestowardimitating(a)
better‐scoringparticipantsthanoneself,and(b)thebest‐scoringparticipant,while
imitationbehaviorintheinvisible‐scoresconditionappearedessentiallyrandomwith
respecttoscoredifferenceandscorerank.
113
Figure5.10:ForguessesthatincludedatleastsomeImitation,participantsintheinvisible‐
scoresconditionhadhigherproportionsofImitationintheirguesses.
5.3.6.Choicesourcestrategy
AsinExperiment2,thechoicesourcesofeachnon‐isolatedparticipantoverthe
entiresessionwereanalyzed,andeachparticipant’schoicesourcestrategywascategorized
accordingtotheirproportionofeachsource.Participantswhosechoicescontainedone
sourceinanaverageproportiongreaterthantheglobalaverageforthatsourceplusone
standarddeviation,werelabeledwiththatstrategy.Forexample,aplayerwhoseguesses
overthecourseofaconditionconsistedofagreaterproportionofImitatechoicesthanthe
averageforallotherparticipantsinthatcondition,plusonestandarddeviation,were
labeledashavinganoverallstrategyof“Imitate.”Thosewhofittheabovecriteriaformore
thanonechoicesource,ornone,werelabeledashavinga“Mixed”strategy.Thescore
distributionforeachstrategycategoryineachconditionisshowninFig.5.11.
114
(a) (b)
Figure5.11:Scorevs.choicesourcestrategyin(a)visible‐scoresand(b)invisible‐scores
conditions,showingthataconservativehigh‐Retentionstrategyresultedinthebest
performance,thoughasimilarlyconservativehigh‐Retrievalstrategy(returningoftentoa
personalbest‐so‐far)showedgoodrelativeperformanceintheinvisible‐scorescondition.
Theabove‐mentionedfiguressummarizetheresultsofsimpleregressionanalyses
performedforscorevs.individualandgroupuseofeachchoicesource.Alinearregression
ofmeanindividualscorevs.meanindividualImitationguessproportionshoweda
significantpositiverelationship(thatis,thegreateraparticipant’saverageproportionof
Imitation,thebettertheparticipant’sscore),butonlyinthevisible‐scorescondition
(F(1,216)=8.12,p=.005,B=0.190;seeFig.5.12a);nosignificantrelationshipwasfoundin
theinvisible‐scorescondition.Theoppositewastrueforindividualscorevs.Innovation,
whichdisplayedasignificantnegativerelationshipinboththevisible‐scores
(F(1,216)=146.2,p<.0001,B=‐0.635;seeFig.5.12b)andinvisible‐scores(F(1,216)=67.88,
p<.0001,B=‐0.489)conditions.ApositiverelationshipheldforRetentioninboththe
115
visible‐scores(F(1,216)=64.6,p<.0001,B=0.480;seeFig.5.12c)andinvisible‐scores
(F(1,216)=13.81,p=.0003,B=0.245)conditions.Finally,apositiverelationshipwasfound
forRetrievalinonlytheinvisible‐scorescondition(SI:F(1,216)=12.73,p=.0005,B=0.236;
seeFig.5.12d).
AsinExperiment2,analysesofmeangroupscorevs.meangroupguessproportion
foreachchoicesourceshowedsimilarrelationshipsofthesamesignificanceanddirections
asthosenotedabove,aswellasanalysesofmeanindividualscorevs.meangroup
(excludingtheindividual)guessproportion,withtheexceptionoftheabsenceofa
relationshipwithRetrievalatbothlevels.Alltrendsnotedaboveweregenerally
monotonic;thatis,therewerenothresholdsorinflectionpointsbeyondwhichthe
relationshipschanged.
(a) (b)
116
(c) (d)
Figure5.12:Higherindividualscoreswereassociatedwith(a)higherindividualImitation
onlyinthevisible‐scorescondition,(b)lowerInnovationinbothconditions,(c)higher
Retentioninbothconditions,while(d)higherindividualRetrievalonlyintheinvisible‐
scorescondition.Bestfittinglinearregressionlinesareonlyshownwhenthelinear
relationwassignificant.
5.3.7.Improvements
AsinExperiment1,improvementsweretalliedforeachparticipantineachsession
andcondition.Histogramsofnormalizedimprovementshareshowedarelativelyequitable
distributionofimprovementswithingroupsinthevisible‐scorescondition,witha
distributionstronglypeakedneara“fairshare”of1(56%ofparticipantswerebetween0.4
and1.2),andonly6.4%ofparticipantshavingzeroimprovements;incontrast,therewasa
stronglyinequitably‐skeweddistributionintheinvisible‐scorescondition,withonly36.2%
ofparticipantshavingimprovementsharesbetween0.4and1.2,and21.1%withzero
improvements(seeFig.5.13).AKolmogorov‐Smirnofftestofequalityofdistributions
117
indicatedthatthesedistributionsweresignificantlydifferent(D=0.1789,p=0.002).Mean
overallscoreshowedastrongpositivecorrelationwithimprovementshareintheinvisible‐
scorescondition(F(1,168)=64.49,p<.0001,B=0.369),butthisrelationshipwasnotevident
inthevisible‐scorescondition.
Themeanchoicesourceproportionsforguessesthatresultedinscore
improvementsandthosethatdidnotareshowninTable5.3.Inbothconditions,the
proportionofInnovationchoiceswashigherforguessesthatyieldedimprovements
relativetonon‐improvements(invisible‐scores:t(733.20)=‐14.03,p<.0001;visible‐scores:
t(907.73)=‐17.14,p<.0001).Intheinvisible‐scorescondition,theproportionofImitation
choiceswassignificantlylowerforimprovementsthannon‐improvements
(t(916.77)=11.54,p<.0001),whileinthevisible‐scorescondition,theproportionof
Retentionchoiceswassignificantlylowerforimprovementsthannon‐improvements
(t(916.33)=9.34,p<.0001).Ofallimprovementsinthevisible‐scorescondition,24.2%
resultedfromguessesthatincludedImitation,versus12.2%intheinvisible‐scores
condition.In52.3%ofallimprovementsinthevisible‐scorescondition,thefocalplayer
imitatedatleastonepeerwhohadpreviouslyimitatedthefocalplayer,versus41.5%inthe
invisible‐scorescondition.Inotherwords,aplayerwhowasimitatedbyanotherplayer
oftenlaterimitatedthatsameplayerinthecourseofcreatinganimprovement,butthis
happenedsubstantiallymoreoftenwhenscoreswerevisible.
118
Figure5.13:Histogramsshowingrelativelyequitableachievementofimprovementswithin
groupsinthevisible‐scorescondition,andaninequitabledistributionintheinvisible‐
scorescondition.
Table5.3:Meanchoicesourceproportionsforimprovementandnon‐improvement
guessesineachcondition.
Condition Improvement %ofguesses Imitation Innovation Retention RetrievalNo 94.6% 9.1% 11.4%* 76.3%* 2.2%Visible
Scores Yes 5.4% 8.2% 19.4%* 69.5%* 2.1%No 95.6% 10%* 13.3%* 71.2% 4.4%Invisible
Scores Yes 4.4% 3.9%* 21.6%* 70.5% 3.5%*significantdifferenceswithincondition
5.3.8.Guesssimilarity
Acomparisonbetweenthemeansimilarityofparticipants’mostrecentguessesto
thosewhomtheyimitated,andtothosewhomtheydidnotimitate,revealedslightbut
significantdifferencesinbothconditions,butinoppositedirections.Inthescores‐visible
119
condition,thereweresimilarityvaluesof.563forimitatedvs..524fornon‐imitated
guesses(t(5084)=‐5.47,p<.0001;seeFig.5.14a).Inthescores‐invisiblecondition,there
weresimilarityvaluesof.316forimitatedvs..346fornon‐imitatedguesses(t(4267)=4.35,
p<.0001;seeFig.5.14b).Inotherwords,priortoimitation,theaverageimitators’guess
wasmoresimilartothatoftheimitatedparticipant(s)thantothoseofothersinthescores‐
visiblecondition,andlesssimilarinthescores‐invisiblecondition.
(a) (b)
Figure5.14:(a)Inthescores‐visiblecondition,imitators’previousguessesshowedgreater
similaritytotheguessestheyimitatedthantothosetheydidnotimitate,while(b)inthe
scores‐invisiblecondition,theoppositeeffectwasobserved.
5.3.9.Frequencyandmomentumbias
AsinExperiment1,wemeasuredthebiasofparticipantstochooseanicon
accordingtoitsfrequencyinpeers’choices.Toreiteratebriefly,wemeasuredthemean
probabilityofImitationandInnovationforanyiconnotalreadyincludedonaplayer’s
120
team,basedonthefrequencyofitsappearanceonpeers’teamsintheplayer’sdisplay,and
comparedthemtoexpectedchancebaselines.
Linearmixed‐effectsanalysisofimitationprobabilityversuschoicefrequency
showedapositivefrequency‐dependentImitationbiasthatwassignificantlygreaterthan
chanceinthevisible‐scorescondition(F(1,604)=943.25,p<.0001,B=.741),butsignificantly
lowerthanchanceintheinvisible‐scorescondition(F(1,604)=231.67,p<.0001,B=.470;see
Fig.5.15a).Therewasaslightpositivefrequency‐dependentInnovationbiasabovechance
inthevisible‐scorescondition(F(1,604)=181.20,p<.0001,B=.441),andbelowchancein
theinvisible‐scorescondition(F(1,604)=12.78,p=.0004,B=.131;seeFig.5.15b).
Wealsorepeatedtheanalysisof“choicemomentum,”bytallyingthechangeinthe
numberofplayerswhoseteamsincludedtheiconintheprevioustworounds,aswellas
thenumberoftheremainingplayerswhoaddedittotheirteaminthecurrentroundvia
ImitationorInnovation,andnormalizingforgroupsize.Afterlog‐transformingthe
Imitationprobabilitydatatoachieveanormaldistribution,at‐testofImitationprobability
fornegativeandpositivechangesinchoicefrequencyshowedasignificantpositive
momentumbiasinthevisible‐scorescondition(t(640)=‐14.192,p<.0001),andasmaller
positivebiasintheinvisible‐scorescondition(t(661)=‐9.98,p<.0001;seeFig.5.16a).A
slightpositivemomentumbiaswasfoundforInnovationinthevisible‐scorescondition,
butnocorrespondingsignificantbiaswasfoundintheinvisible‐scorescondition(seeFig.
5.16b).
121
(a) (b)
Figure5.15:Therewerebiasestowardchoosingelementsthatweremorefrequently
representedonotherteamsinthevisible‐scorescondition,andlessfrequentlyrepresented
onotherteamsintheinvisible‐scoresconditionfor(a)Imitationand(b)Innovation
decisions.
(a) (b)
Figure5.16:Therewerebiasestowardchoosingelementswhoserepresentationonother
teamswasincreasingin(a)bothconditionsforImitationand(b)onlythevisible‐scores
conditionforInnovationdecisions.
122
5.4.Experiment3Discussion
Aspredicted,resultsinthescores‐visibleconditionwerequitesimilartothosein
Experiment2,whiletheresultsinthescores‐invisibleconditiondifferedinsomewayswe
didnotpredict.
5.4.1.Differencesinperformance
HavingdemonstratedbenefitsforImitationinthepreviousexperiments,the
impedimenttosociallearningintroducedintheinvisible‐scoresconditionlowered
performanceaspredicted.Thus,thereductionintheefficiencyofsociallearning
implementedbyhidingpeers’scoresdidleadtoincreasedInnovationandsolution
diversity,butdidnotseemtoimprovecollectivesearchperformanceasinMason,Jones,
andGoldstone(2008)andLazer&Friedman(1997).Thisdifferencefromthestudiescited
abovewaslikelyduetothewaythatcommunicationefficiencywasreduced:whereasthey
decreasedtheconnectivityofthesocialnetworkthroughwhichinformationwas
exchanged,weleftthenetworkunchangedbuteliminatedanimportantpartofthe
informationthatparticipantsusedtoguideimitationdecisions.Ourresultsare
substantiallyinaccordancewiththefindingsofGiraldeau,Valone,andTempleton(2002),
whofoundthataninabilitytocombinetheuseofsocialandasociallearning
simultaneouslywouldresultinalackofbenefitforsociallearning;however,wedid
observesomebenefitforsociallearning,inthatparticipantsintheinvisible‐scores
conditionstillperformedbetterthanisolatedparticipants.
123
5.4.2.Differencesinstrategy
Overall,thechangeintheavailabilityofperformanceinformationseemedtoshift
participants’tacticsfrommakingsmallchangestotheirguesses,combiningasocialand
sociallearning(incrementaliststrategies)inthevisible‐scorescondition,tomakinglarger
jumpsaroundtheproblemspaceandoftenjumpingbacktopreviousknowngoodsolutions
(saltationiststrategies)intheinvisible‐scorescondition.Thisdifferencewasdiscerniblein
anumberofresults.First,whenscoreswereinvisible,proportionsofInnovationand
RetrievalwerehigherandRetentionwaslower,implyingthatonaverageparticipantswere
keepingsmallerportionsoftheirguessesfromroundtoroundandchangingoftenbetween
newandoldsolutionelements.ThoughtheoverallmeanproportionofImitationwasthe
sameinbothconditions,Imitationwasusedlessoftenandforalargerproportionofthe
guessintheinvisible‐scorescondition–participantsweresignificantlymorelikelytocopy
mostorallofapeer’ssolution.ThelowermeanRetentionintheinvisible‐scorescondition
impliesthatparticipantswereoftenjumpingtoapeer’sguessandeitherkeepingitifit
increasedtheirscore,orjumpingbacktotheirRetrievedpreviousbestguessifitdidnot.
Theabilitytomakeincrementalchanges,mixingelementsfromallsources,should
allowparticipantstoassesstheeffectofsmallerchangesandmakebetterjudgmentsabout
thequalityofindividualelementsandpairsfromotherparticipants;thustherewere
substantialincreasesinscoreacrossroundsandassociationsofRetentionandImitation
withhighscoresobservedinthevisible‐scorescondition.Arelianceonlargeriskyjumps
aroundtheproblemspacewouldlikelypayoffabouthalfthetimeforthemedian
participant,andthosewhojumpedbacktogoodprevioussolutionswouldloselessoverall
thanthosewhocontinuallyjumparound;thustherewereshallowerincreasesinscoreover
124
rounds,andanassociationofhigherscoreswithRetrievalbutnotImitationintheinvisible‐
scorescondition.Thissaltationiststrategyseemstohavebeenmoresuccessfulthannot
usingImitationatall,however,asshownbythesubstantiallylowerperformanceof
isolatedparticipants.
Theremovalofscoreinformationalsoaffectedtheuseofotherkindsofinformation
inparticipants’imitationstrategies;however,ratherthanstrengtheningsocially‐mediated
informationbiases(assuggestedbyAbrahamsonandRosenkopf(1997)andGibbons
(2004))suchasfrequencybiasorsimilaritybias,participantsactuallyshowedweakened
oroppositeinclinations.Theaboveinterpretationintermsofchangingthemagnitudeof
movementsintheproblemspaceholdshereaswell.Copyingasimilarsolutiontoyourown
isaninherentlyincrementaliststrategy,sothepresenceofthisbiasinthevisible‐scores
condition,andthepresenceofitsoppositeintheinvisible‐scoresconditioncanalsobe
explainedintermsofanoverallchangefromincrementalisttosaltationiststrategies.The
unpopularitybiasandreducedmomentumbiasweobservedmayhaveoccurredbecause
participantsknewthatimitationdecisionswereoftennotbasedonreliableperformance
information,andthusfrequency‐basedbiasesshouldbeavoidedtokeepfromjoining
informationherds(Banerjee,1992),assuggestedbyGiraldeau,Valone,andTempleton
(2002).
5.4.3.Learning?
Changesinstrategyacrossgameordercanbethoughtofaslearningoradaptation
tothetaskoversuccessivegames.Thechangesweobservedacrossgameorderimplythat
strategiesineachconditionwerestrengthenedoverthecourseofthesession.Inthe
125
visible‐scorescondition,thedecreasesinInnovationandImitationandincreasein
Retentionshowamoreincrementalistapproach,whileintheinvisible‐scorescondition,
theincreasesinImitationandRetrievalanddecreaseinInnovationdisplayedmoresolidly
saltationisttendencies.Itisfairlysimpletoseethatanincreasinglyincrementaliststrategy
thatincludedsomeImitationwouldshowanincreasingpayoffintermsofbeingableto
distinguishgoodindividualelementsandpairs,andtheimprovementinscoreacrossgame
orderinthevisible‐scoresconditionreflectsthis.Anincreasinglysaltationiststrategy,
however,isnotlikelytoshowbenefitsunlessthesolutionsofone’speersimprove,and
thuswesawnoimprovementinscoreacrossgameorderintheinvisible‐scorescondition.
Thereisnoobviousbetter‐performingstrategyinthiscontext,however,soparticipants
seeminglydoubleddownonthisone.
5.4.4.Convergent/locallyefficientsearch
Asseenintheincreasingscoreanddecreasingguessdiversitytrendsacrossrounds,
averageperformanceincreasedviatheconvergenceofgroupmembersonregionsofthe
problemspacethatcontainedhigh‐qualityteams.Thisconvergencecombinedwithasmall
amountofindividualexplorationcausedsuchregionstobeexploredmorethoroughlyand
stillbettersolutionstobefound.However,intheinvisible‐scorescondition,whenimitation
wasnotfocusedonasmallgroupofbetter‐performingneighbors(becauseperformance
informationwasnotavailable),similarguesses,orpopularsolutionelements,this
convergencehappenedmuchmoreslowly,searchwasmorediffuseandlessefficient,and
lowerperformanceresulted.Theweakertrendsofincreasingscoreanddecreasingguess
diversityacrossgroupsizeintheinvisible‐scoresconditionshowedthatthelackofscore
126
informationmadesearchersunabletoeffectivelytakeadvantageoftheincreasingnumbers
oftheirfellowsearchers.
5.4.5.Cumulativemutualimprovement
Thesignificantcorrelationofimprovementsharewithmeanscoresintheinvisible‐
scoresconditionsshowsthatindividualswhowererelativelymoreskillful(orlucky)were
rewardedwithproportionatelybetteroverallscorescomparedtoothers;thiswasbecause
theirfellowplayerscouldnoteasilycopytheirimprovementsandachievetheirscores,and
becauseknowingthatImitationwasunreliablemadesomeparticipantsmorelikelytoseek
outimprovementsontheirown.Inthevisible‐scoresconditionthiscorrelation
disappeared,butthemoreequitabledistributionofimprovementsshowedthatmore
participantswerecontributingtotheirdiscovery,andmeanscoresincreasedsignificantly
suchthatnearlyallparticipantsdidbetter.Inotherwords,whensociallearningwas
unimpededinthevisible‐scorescondition,highandlowindividualachievershad
approximatelythesamepayoffs,butabsolutepayoffswerehigherforallcomparedtothe
invisible‐scorescondition,inwhichhigh‐achieversgotabiggerpieceofasmallerpie.Thus
impedingsociallearningledtorelativelygreaterinequityandinefficiency,andpresumably
lowerlong‐termperformance(thoughwedidnottestthisexplicitlywithlongergames).
Thisadvantageformoreefficientsociallearningaccruedbecauseimitatorswerenot
merelyscroungers;thesubstantialproportionofImitationpresentinimprovementsshows
thatimitatedguesseswereoftenthebasisforfurthercumulativeinnovations.The
cumulativeinnovationhypothesisissupportedbythefactthatalargerproportionof
improvementsweretheresultofmutualImitationinthevisible‐scorescondition,inwhich
127
solutionelementswerepassedbetweenplayersviacopyingandbuiltintobettersolutions
intheprocess.Thisenabledamoreactivesharingofthe“labor”ofproducing
improvements,andincreasedperformancefromparticipantsoverall.Intheinvisible‐
scorescondition,thenecessityofadoptingothers’guessesinordertoobtaininformation
abouttheirperformanceallowedfeweropportunitiestoevaluatevariationsonthem;it
alsopreventedgroupmembersfromperformingthe“filtering”functionofcopyingand
consistentlyretainingonlysolutionelementsassociatedwithrelativelyhighscores,sothat
otherswouldhavelesschanceofobtaininglow‐scoringsolutionelementswhencopying.
5.5.Conclusions
InExperiment2,wenotedthelackofa“tragic”outcomeintheproductionanduse
ofhigh‐scoringsolutions,despitetheapparentincentiveforindividualstounder‐produce
innovationsandfree‐rideontheinnovationsofothersviaself‐interestedimitation.Inthis
experiment,weshowedthatitwaspossibletoinduceatragicoutcomebyreducingthe
capacityofindividualstomakeself‐interestedimitationdecisions,evenwhilethey
increasedtheirproductionofinnovation.
Itappearsthatlowerperformanceinthistaskwasnotduetoanunderprovisionof
individualinnovation,butalackofevaluativefilteringofsolutionelements,which
participantsinthevisible‐scoresconditiondidbychoosingtoimitateandretainsolution
elementsassociatedwithhighscores,andwhichinthescores‐invisibleconditionwasmade
muchmoredifficult.Theconsistentuseofthebetter‐performingsolutionsdoesnotrelyon
altruisticorpublicly‐mindedmotives,butsuchfilteringisimportantforsupportingothers'
successfulsociallearning,aswashighlightedinarecenttournamentofsimulatedsocial
128
learningstrategies(Rendellet.al2010).Insummary,wehaveshownthroughthis
experimentthatwhenknowledgeiscumulative,efficientandinformedappropriationisan
importantstepinfurtherprovisionofthepublicgoodofknowledge.
129
6.KnowledgeasProperty
6.1.Promotingprogress
Intwopartsofthesociallearningexperimentsdescribedthusfar(largergroup
sizesinExperiment1,andthescores‐invisibleconditioninExperiment3),wehaveseen
thatdespiteapparentlyadequateprovisionofinnovation,obstaclestoaccurate,strategic
imitationofthoseinnovationscanleadtounderperformanceinfurtherinnovation.Inthis
chapterwewillexploresomemeasurestakenintherealworldto(ostensibly)encourage
efficientinnovationand(eventual)imitation,andintroduceseveralimportantconceptsfor
understandingtheirimplicationsontherelatedpublicandcommongoodsofknowledge
discussedinChapter4.
6.1.1.TheProgressClause
Practicalresearchanddevelopmentarecostlyactivitiesintermsofmoney,effort,
time,risk,andopportunity.Theseactivitiesgenerallywillnotbeundertakenunlesstheir
resultshavesomeexpectedvaluetojustifytheircosts.The"ProgressClause"oftheU.S.
Constitutionempowersthelegislature"TopromotetheProgressofScienceandusefulArts,
bysecuringforlimitedTimestoAuthorsandInventorstheexclusiveRighttotheir
respectiveWritingsandDiscoveries."Manyothercountrieshaveenactedsimilarstatutes
andpoliciesprotecting"intellectualproperty"(IP),whichgenerallyallowtheoriginatorsof
certaininnovationstocontrolthedisseminationorproduction(andthusthebenefit)ofthe
productsoftheirinnovations.(Thisclausecoversthecreationofbothpatentsand
copyrights;inthischapterandthenext,wewillbefocusingonpatents,thoughmanyofthe
sameargumentscouldapplytocopyrights.)TheownerofapatentgrantedbytheU.S.
130
PatentandTrademarkOfficeisgiventherighttoexcludeothersfrom"making,using,
selling,offeringtosell,orimporting"thepatentedinvention(35U.S.C.§271),andthisright
canbereneweduntil20yearsafterthepatentapplicationwasfiled(35U.S.C.§154).In
exchange,apatentapplicantmustdescribetheinventioninenoughdetailsothataperson
whoisskilledinthefieldwouldbeabletomakeandusetheinvention(35U.S.C.§112).
Thedisclosurerequiredforapatentedinventioncanbeconsideredacontribution
toaknowledgecommons,inthattheinformationthatenablesoneto"makeanduse"the
inventioncanalsobeusedtoimproveit,ortocreateotherinventions.AsGhosh(2003)
notes,thecharacterizationofagoodasrivalrousdependsonthepropertyrightsthat
governit.TheexclusivecontroloveraninventionthatIPstatutesgivetocreatorscanmake
thatinventionusablebyonlyoneindividualorfirm,andsoartificiallytransformpractical
knowledgeintoasubtractableresource.ThustheexerciseofIPrightscanbeconsidered
appropriationofresourceunitsfromthecommons,whichcannotbedoneinan
unrestrainedwaywithoutthreateningtheviabilityoftheresource.
Thebalanceofrightsandresponsibilitiesintheseprovisionsimplicitly
acknowledgesthatindividualinnovationisencouragedforthepurposeoftheextending
thecollectivebenefitsofthegeneralprogressofknowledge.Thestandardassumption
abouttheintentofthe"ProgressClause"isthatitwasmotivatedbytheframers'viewthat
"encouragementofindividualeffortbypersonalgainisthebestwaytoadvancepublic
welfare"(MazervStein,1954).Thatis,publicbenefitistheprimaryintendedend,andthe
provisionofaprivateincentiveisthepreferredmeansforachievingit.Asdiscussed
previously,however,privatizationofaresourceisnotalwaysthebestwayofensuringits
equitable,efficient,andsustainableuse.
131
6.1.2.Measuringthepromotionofprogress
Theprogressofsciencedependsontheincentivesthatinventorshavetoinvestin
researchanddevelopment,aswellastodisseminatetheirinventionstothosewhocan
productivelyadopttheminsocietyatlarge.Thus,thesuccessofpatentstatutesinthe
promotionofscientificprogresscanbeusefullyevaluatedonthebasisoftheneteffectsof
theIPsystem,bothforthosewhodirectlyadvancetheprogressofscience,andforthe
generalpopulace.Thereareundoubtedlybenefitsthataccruetomanypatentholdersasa
resultoftheirexchangeofideasformoney,asthereareplainlybenefitstosocietyinthe
formofpatentedmoderntechnologies.ThequestioniswhethertherewardsoftheIP
systemoutweighitscostsforallinventors(boththosewhoareIPholdersandthosewhose
inventionsarenotprotectedbyIP),andthesocietyandeconomyinwhichtheywork.(The
lattercanbeconsideredanextensionoftheformer:aseconomiesandsocietiesgrow,new
technologicalneedsariseandthusnewopportunitiesforinnovation.)Herewewilltakea
shortdetourtoexaminepreviousattemptstoanswerthisquestionempirically,aswellas
thetheoriesunderlyingIPsystems;thesewillhelpconstrainandmotivateourfinal
experiment.
Though"rights"canseemratherabstract,proxieshavebeendevelopedforstudying
thequantitativeeffectsofvariousfactorsonoveralleconomicgrowthincross‐national
econometricstudies,includingprotectionsforpropertyrights.Generally,traditional
propertyrights(thosegoverninge.g.objectsorrealestate)haveastrongandunambiguous
relationshipwithgrowth(Svejnar,2002;Keefer&Knack,1995,1997).ParkandGinarte
(1997)performedaregressionusingindicesof"marketfreedom"(i.e.traditionalproperty
132
rights)aswellasIPrights,andconfirmedthattheformerhadastrongrelationshipwith
growth,butnotthelatter.Inafollow‐upstudy,GinarteandPark(1997)actuallyfound
evidenceforacorrelationbetween5‐yearlaggedR&DspendingandthestrengthofIP
rightsprotection,implyingthatinvestmentininnovationsledtothedevelopmentof
protectionsforthem("reversecausality").Incontrasttotraditionalpropertyrights,IP
rightsappeartohaveonlyanindirectorweakrelationshiptoeconomicgrowth;itmay
holdonlyforcertaingroupsofcountriesorcertainkindsofmeasures(Falvey,Foster,and
Greenaway,2006),andthecausaldirectionisunclear.
SeverallinesofresearchhavebeenpursuedtoaddressthequestionofwhetherIP
rightsarebeneficialforinventors(andtheiremployersandinvestors).Studiesof
innovationpriortoandduringthe19thcenturyshowedthatmanyinventionsmadetheir
inventorsenormousprofitsandwerejudgedtobehighlyinnovativedespitebeing
unpatentedorhavingtheirpatentsinvalidated(Mokyr,1999;Moser,2002).Bessenand
Meurer(2008)andBoldrinandLevine(2008)reviewedabroadrangeof"natural
economicexperiments"ondiscretechangesinthepatentlawsofvariouscountriesinthe
20thcentury.Theyfoundthat,overall,thereisstrongevidencethatstrengtheningpatent
protectionleadstomorepatenting(asinthe"reversecausality"notedabove),butweakor
noevidencethatitincreasesinnovation.Infact,intherealmofsoftwarepatents,firmsthat
acquiredrelativelymorepatentssubsequentlytendedtoreducetheirR&Dspending
relativetosales(Bessen&Hunt,2007).BoldrinandLevine(2008)describeincreasesin
patentingactivityas"navigatingthepatentthickets,"inwhichfirmsthatfearinfringement
lawsuitsfromas‐yet‐unknownpatenteesacquirepatentsstrictlyforthepurposeoffiling
133
countersuitsandachievingsettlementswithouthavingtomodifytheirbusiness,thatis,to
avoidbeingforcedtoinnovatefurther.
BessenandMeurer(2008)examinedpatentrenewalbehaviorinordertoestimate
thevalueofpatentsapartfromtheirunderlyingtechnologies,andfoundthatnearly60%of
U.S.patentsfiledin1991werenotrenewedtofullterm,whichindicatesthatthemajority
ofpatentsfileddepreciatetoavaluelessthanthefewthousanddollarsoftheaverage
renewalfee.Asonemightexpect,thedistributionofpatentvaluesisskewed‐‐most
patentsareofrelativelylittlevalue,whileasmallnumberarequitevaluable.Avarietyof
measuresofpatentandinventionvalueadhereapproximatelytothe80‐20rule:80percent
ofthetotalvalueiscontainedin20percentoftheinventions(Harhoff,Scherer,&Vopel,
2003).BessenandMeurer(2005)foundthatgenerally,patentsgrantedtosmallinventors
(individuals,nonprofitsandcorporationswithlessthan500employees)aremuchless
valuablethanthosegrantedtolargerentities,about50%inthemean,andthedifference
betweenindividualsandorganizationsoverallisevengreater.Furthermore,theownership
ofthisvalueisfoundtoberatherlopsided:morethanhalfofthevalueofworldwide
patentsaccruestoasmallnumberoflargepharmaceuticalcompanies,andmorethantwo‐
thirdstofirmsinthechemicalandpharmaceuticalindustriesgenerally.
Overall,thereisstrongevidencethatpatentscandeliversignificantvalueto(some
of)theirowners,butthepatentsystemalsoimposescostsonpatentownersthrough
disputesovertherightstotheirinventions,aswellasonnonpatentinginventorsthrough
disputesrelatedtoinadvertentinfringementofothers'patents.Theprocessof"clearing"
rights(byeithermakingsurenotrespassoccursorobtainingalicense)isnotnecessarily
simpleformanypatents,becausetheboundariesofwhataninventiondoes,andbywhat
134
processormethoditisdone,canbearcaneandopentocostlylegalinterpretation(Moore,
2005).BessenandMeurer(2005),aftercontrollingforawidevarietyofvariables,found
thatincreasedspendingonR&Disactuallycorrelatedwithanincreasedriskofbeingsued
forinfringement,whichsuggeststhatinfringementisoftennotwillful,andoccursdueto
difficultiesindeterminingpatentboundaries.Acomparisonofaggregatelitigationlosses
incurredbyasampleofdefendantpublicfirmsduringtheyearsof1984‐1999withthe
aggregateincrementalprofitsderivedbypublicfirmsduringthesameperiodshowedthat
outsideofthechemicalandpharmaceuticalindustries,thecostsoflitigationclearly
exceededprofitsfrompatents(Bessen&Meurer,2008).Thusoverall,thepatentsystem
appearstoconstituteanetdisincentiveforinnovation.
Theresultsnotedabovesuggestthatinitspresentform,thepatentsystemisnot
promotingtheprogressoftechnologicaladvancement;thatis,theequity,efficiency,and
sustainabilityoftheknowledgecommonsdonotappeartobeimprovedbythepatent
system.Thismaysimplybeduetostructuralshortcomingsofthesystem,remediableby
adjustmentstopatentlengthorbreadth(Gilbert&Shapiro,1990).However,theremay
alsobemoresubtleissuesregardingthemotivationalprinciplesandinterdependenciesof
innovatorsthatarenotbeingadequatelyaddressedbyexistingpatenttheory,law,and
institutions.
6.2.Prospectingforprogress
Themonopolygrantedbyapatentisexpectedtocause"deadweightloss,"a
reductioninmutualvalueforproducersandconsumersthatoccurswhenconsumerswho
arewillingtopaymorethanthemarginalproductioncostofagood,butlessthanthe
135
patentee'smonopolyprice,arethuspreventedfrompurchasingthegood(Gilbert&
Shapiro,1990).Kitch(1977)introducedthe"prospecttheory"ofpatentstocommonlegal
usage,inwhichthecosttosocietyofdeadweightlossisexplicitlyjustifiedanddiscounted
byaperceivedneedtoincentivizenotjustcreationanddisclosureofinnovations,butalso
theirefficientmanagement.Kitchexpoundeduponthistheorywithananalogytothe
exploitationofmineralresourcesintheAmericanWest,buthewasrespondingtoBarzel's
(1968)callforasolutiontoacollectiveactionproblemamonginnovators.Barzelobserved
thatfirmswhowantedtousethe"freepublicgood"ofbasicscientificknowledgeto
produceinnovationsforcommercialuseeachhadanincentivetointroducesuch
innovationsassoonastheywereprofitable.Buthereasonedthatinnovators(andthus
societyatlarge)wouldcollectivelybebetteroffiffirmsdelayedtheintroductionof
innovationsuntilan"optimal"time,whenmarginalprofit(relativetoreliableoutside
investments)couldbemaximizedthroughthedevelopmentofsufficientproduction
capacityanddemand.InthesameyearasHardin's(1968)analysisoftheTragedyofthe
Commons,Barzel(1968)alsorecommendedenclosureoftheknowledgecommonsinthe
formofownershipofbasicknowledgeinordertopreventwhatheviewedasatragic
outcomeinitsuse.ButaswithHardin,theintellectualorigins,analogies,andassumptions
usedtounderstandthiscomplexsituationmustbecloselyexaminedforcluesastowhere
proposedsolutionsmaysucceedorfail.
Barzel'sanalysisofthis"optimal"usedependsonanumberofsimplifying
assumptions;forourpurposesthemostsalientarethat(a)innovationshaveaconstant
costandthuscanbeintroducedatanytime(ignoringthedynamicandcumulative
developmentofknowledge),(b)thereisonlyoneinnovatorassociatedwithagiven
136
innovation(ignoringcollaborativeinnovationprocesses),(c)therearenobenefitsofthe
innovationpassedontotheconsumer(ignoringtheexistenceofexternalitiesor
"spillovers"(Frischmann&Lemley,2007)uncapturablebytheproducer,suchas"network
effects"(Katz&Shapiro,1985)),and(d)theparametersofthesystemarewidelyand
accuratelyknown(ignoringthecomplexanduncertainnatureofinnovation).Kitch(1977)
argued,amongotherthings,thattheU.S.patentsystemwasstructurallypredisposedto
"prospect"uses,andthatgrantingbroadandlong‐termpatentrightsonthebasisofthe
"prospect"forefficientintroductionandmanagementofaninnovation(ratherthanonthe
basisofthepreviously‐accepted"rewardtheory"ofsimplyincentivizingproductionand
disclosureofinnovations)wouldresultinanetpositiveforsociety.
Implicitinthisapproachareotherassumptionsaboutinnovatorsandtheir
motivations.Primarily,theaboveanalysestreatinnovationasanactivitythatispursued
onlybyfirmswhoproducegoodsforthepurposeofsellingthemtoendusers.Themining
analogyusedbyKitch(1977)impliesthatmostpracticalinventionsrequireagreatdealof
investmentintheextractionandprocessingofthe"ore"ofbasicknowledge,andthusthat
firmswithexclusiveclaimswillbethebestinformedandequippedtocreateandmanage
innovations.However,thehistoryoftechnologyshouldteachusthatpeoplearegenerally
notskilledatanticipatingtheusesofnewinventionsortheneedsoffutureusers(Boyle,
2007);inventionssuchasmobilephones,DVDplayers,andfinancialdatabasesoftware
haveresultedinavarietyofusesandamagnitudeofeconomicandsocialbenefits
unimaginableanduncapturablebytheircreators(Baumol,2002).Thoughprospecttheory
positsadecentralized,entrepreneurialimplementationofinnovationatlarge,itcentralizes
controlofanyparticularinnovationinonefirm.Inpractice,holdersofpatentscandelayor
137
preventfutureinnovationsbyeitherrefusingtolicensetheirpatentstosubsequent
innovators,orchargingsufficientlyhighlicensingfeessothatsmallerinnovatorsare
preventedfromenteringthemarket(Merges&Nelson,1990).
Previousinventionscanaidinthecreationofnewinnovationseitheras“research
tools”thatenableafurtherinnovativestage,orasthebasisforanimprovedproduct(or
sequentialinnovation)onthe“qualityladder”(Hall,2007).Itisnotnecessarilythecase
thateithertheoriginalinnovatororonlythosewithdeep‐pocketedinvestorsarebest
suitedtoimproveuponoruseapreviousinnovationtocreateanother,butinsomecases
thesemaybetheonlyagentswhoareabletodoso.Additionally,innovatorsmaynothave
sufficientincentivetocreateorallowimprovementsontheiroriginalideas,iftheyfearthat
anewusewouldcompetewiththeoldone(Scotchmer,1991).Whenpreviousinnovations
necessaryforthecreationordevelopmentofanewinventionarepatentedbymultiple
parties,aphenomenoncalled“thetragedyoftheanticommons”canresult(Heller&
Eisenberg,1998).Inthissituation,thetragedyresultsnotfromalackofresourceoversight,
butfromanexcessofit;ownersoftechnologiesthatarecollectivelynecessaryforanew
innovationmayfailtoagreeonlicensingorrevenuesharingterms,andthusanyofthem
canvetofurtherprogress.
Tosummarize,wehaveseenthatpatentsystemsareintendedtoincentivizethe
creationanddisclosureofinnovations,aswellastheirefficientmanagement(undercertain
interpretations).Inpractice,however,patentscanalso(a)incentivizenon‐innovative
strategicorrent‐seekingbehavior,(b)reducetheincentivestocreateanddisclosethrough
theaccompanyingriskofcostlypatentlitigation,and(c)reducetheincentivetooptimally
managethroughpatentholders'abilitytodelayorprevent(butnotcaptureorprofitfrom)
138
manypositiveexternalitiesoftheirpatentedinnovations.Theabovefactorsalsoconstitute
disincentivestoimitateandbuildsequentialimprovementsfromtheinnovationsofothers.
Returningtotheconceptofaknowledgecommons,theseproblemsaresignsofan
institutionwhichisfailingtoensurecontributionandpreventoverappropriationfroma
resource(theoverallpoolofexistingknowledge)byitscommunityofusers.Inthelong
term,suchflawsthreatentheviabilityoftheresource.
6.3.Balancingmotivations
Anotherareaofresearchinwhichtheadequacyofexclusioninfosteringinnovation
becomessuspectisthecurioustendencyofpeopletotreatextrinsicincentives(suchas
monetarypayment)andintrinsicorsocialmotivations(suchaspersonalsatisfactionor
socialapprobation)differently:asnon‐additive,mutuallyexclusive,orcompletely
incommensurable.Forexample,givenanactivitythatproducesintrinsicrewardssuchas
volunteerserviceorblooddonation,theofferofmonetaryrewardsmayactually"crowd
out"themotivationtopursuetheactivityandreduceitsoveralllevel.Thus,undercertain
circumstancesthepricemechanismforraisingthesupplyofgoods,oneofthefundamental
pillarsofeconomics,maynothold.Thestudyofthisphenomenonhasproducedalarge
bodyofworkinbotheconomics(e.g.Titmuss,1970;Frey&Jegen,2001)andsocial
psychology(e.g.Deci,1971;Deci,Koestner&Ryan,1999).
HeymanandAriely(2004)positedthatsocialandmonetaryrewardsaremediated
byseparate"markets,”andthattheycannotbemixed.Theyperformedexperimentsin
whichtheymeasuredtheeffortputforthinresponsetovaryinglevelsandtypesofrewards
offeredtoparticipantsforperformanceofatask.Itwasfoundthateffortwasconsistently
139
highwhennorewardwasoffered,aswellaswhentokennon‐monetaryrewards(e.g.
candy)ofvaryingvaluewereoffered.Whencashwasofferedorthecashvalueofthenon‐
monetaryrewardswasmentioned,effortwasproportionaltothevalueofthereward
offered,andonlyapproachedthelevelreachedbynorewardsathighlevelsofmonetary
rewards.
FreyandJegen(2001)reviewedstudiesrelatedtothecrowding‐outeffectof
extrinsicrewardsonintrinsicmotivationsandfoundthatitwaspresentforavarietyof
norm‐relatedmotivationsandsettings:workeffortandreciprocityinsupplyinglabor,
altruisminserviceprovision,civicdutyandenvironmentalcareinmanagingacommon‐
poolresource,andtrustinlegalsystems.Deci,Koestner,andRyan(1999)performeda
meta‐analysisof128studiesonthiseffectandfoundthatawidevarietyofextrinsic
rewardsandrewardcriteriaunderminedintrinsicmotivationandinterestineffortful
tasks.
Thepresenceofsocialorintrinsicrewardsforsomecreativeandinnovativeactivity
isinasenseevidentbydefinition,fromthefactthatsuchbehaviorscontinueinamultitude
ofcontextswithoutanyexternalmonetaryreward,andeveninthepresenceofsubstantial
costs(e.g.varioushobbyistpursuits,amateurblogging,Wikipediacontributions,etc.)
Beyondthisevidence,theintrinsicmotivationsthatpeoplehaveforcreativeworkhave
alsobeenstudiedinsomedetailandavarietyofreasonsarefoundforparticipating:
curiosity,theenjoymentofchallengeornovelty,personalexpressionofvalues,buildingor
maintainingapositivereputationamongpeers,andsoforth(Lakhani&Wolf,2003).
Totheextentthatinnovatorsunderstandthedependenceoftheirworkonthatof
previousinnovators,andthatspilloverscanaccruetofellowinnovatorsandtosocietyat
140
largeasaresultoftheirefforts,themotivationsofreciprocityandpublicbenefitsmaybe
includedinthislistofintrinsicmotivations.Inthiscase,theopportunitytoexcludeothers
frombuildinguponone'sworkmaybeunattractive.Totheextentthatextrinsicrewards
canunderminethesemotivationsforcertainkindsofwork,someinnovativeactivitymay
simplynotoccur,oronlyatagreaterrelativecost.
6.4.Othermodelsofinnovation
Theseflawspromptustoconsiderothermodelsofinnovationanditsgovernance.
Theassumptionsofthe"reward"and"prospect"theoriesofinnovationbyproducers
excludeimportantclassesofinnovators:individualsorfirmswhocreateinnovationsfor
theirowninternaluse,andthosewhoproduceinnovationsinopencollaborationwith
othersinordertoshareinthecostsoftheircreationandthebenefitsoftheiruse(Baldwin
&vonHippel,2009;Strandburg,2009).
Innovationscreatedbyindividualsorfirmsfortheirowninternaluse(ratherthan
forsellingtousersataprofit)needonlytoprovideenoughbenefittocovertheeffortof
producingthem.Thismeansthatsmall,incremental,orspecializedinnovationsthatmight
notjustifyinvestmentorproductionbyanon‐userfirmcanbecreatedwithoutthe
provisionofexternalincentives.Studiesinawidevarietyofindustriessuchaschemical
products(Freeman,1968),scientificinstrumentsandsemiconductors(VonHippel,1976,
1977),andsportingequipment(Shah,2000)haveshownthatalargeproportionof
importantandnovelproductsandprocessesweredevelopedbyusers,andthatsubstantial
proportionsofusersengageindevelopingormodifyingproductstheyuse(Gault&von
Hippel,2009).Moreover,userinnovatorsoften"freelyreveal"(relinquishexclusiverights
141
andallowgeneralaccessto)theirinnovations(Harhoff,Henkel,&vonHippel,2003).They
dothisbothtoavoidthecostsanddifficultyofexcludingothersthroughIPrights(as
discussedabove)ortradesecrecy,andtotakeadvantageofprivatebenefitssuchasaccess
tothesequentialimprovementsofothers,reputationenhancement,orpositivenetwork
effectsresultingfromincreaseduseoftheirinnovation(Harhoff,Henkel,&vonHippel,
2003).However,userinnovatorscanalsomaintainsomeIPrightsandpursueotherlegal
measuresforthepurposeofavoidingexpropriationoftheirefforts(O'Mahony,2003).
Thisconceptcanbetakentoanotherlevelinopencollaborativeinnovation,in
whichagroupofusersactivelyrevealandcoordinatetheircreativecontributionstoeach
otherforintegrationinalarger‐scaleproject,effectivelypoolingtheirinvestmentswith
assurancesthatallwillshareintherewards.Suchprocessesareinusemostrecentlyand
visiblyinopen‐sourcesoftwaredevelopment(Benkler,2002),butthismodelof
developmenthasbeenusedinotherindustriesformanyyears(Allen,1983).Open‐source
contributorsdosobecausetheywanttoincluderatherthanexcludeothersfromtheuseof
theircontributions,oftenfornormativereasonsrelatedtothecultureofopen‐source
software,orforprivatebenefitssuchasthosementionedabove(Lerner,Tirole,&Pathak,
2006).Advancesincomputingpowerhaveenabledlargereductionsindesignand
developmentcostsforuserinnovation,andrelatedadvancesinnetworkedcommunication
andcollaborationsoftwarehaveloweredbarriersforthecoordinationofcontributionsin
collaborativeprojects(Benkler,2002).Thesedevelopments,alongwiththeincentivesthat
suchactivitiesprovideforbothinnovationandimitation(andrelatedsequential
innovationsandotherspillovers),makeitincreasinglyattractiveasamodeofscientificand
142
generalacademicpractice(Boyle,2007;Lougee,2007),aswellasmoresociallyoptimal
governanceofknowledge(Baldwin&vonHippel,2009).
6.5.Testingmodelsofincentivesandprogress
ThecomplexityofIPsystemsaswellastheirlargepotentialimpactoneconomic,
technological,andculturaloutcomesinsocietymakesthemfertilebreedinggroundsfor
theoriesaboutunderlyingissuessuchasincentives,conformity,creativity,theoptimal
balancingofindividualandcommoninterests,thenatureandstructureofknowledge,and
ahostofotherissues;thesetheoriesbynecessitycrossdisciplinaryboundariesoflaw,
politicalscience,economics,psychology,andphilosophy.Thepurposeofexperimentsin
understanding"intellectualproperty"andthepromotionofintellectualdiscovery,asinany
othersystem,istosimplifytheissuesathandinordertoshowmoreclearlyhowthey
operateandinteract,andinsodoing,topointoutwhatisnotwellunderstoodforfurther
study.
Thepreviousexperimentsinthisdissertationcanbeconceptuallycomparedtoopen
collaborativedevelopment,becauseeveryparticipant'scandidatesolutionsand(exceptin
onecondition)informationabouttheirperformanceareavailableforusebyothersinthe
group,withnocostorhindrance(thoughparticipantsdidnothaveachoiceinthematter).
However,therewardsforgoodandbadperformancearemateriallythesame‐‐course
creditwastheonlycompensationgiventoallparticipants.Wewishedtotesttheeffectofa
changetotheincentivesinthiscontextbypayingparticipantsasmallcashrewardbased
ontheirperformance,aswellasimplementingaverybasicandlimitedpatent‐likesystem.
Inthissystem,eachparticipantwouldbeallowed,atacertaincostandforalimitedtime,to
143
automaticallychargeotherswhocopytheirinnovations.Thepreviouslyexistingincentives
forinnovation(thepossibilityofreceivingthebenefitofanimprovedsolution)wouldbe
enhancedbythepaymentsthatsuccessfulinnovatorscouldgetfortheirsolutions;the
previouslyexistingincentivesforimitationwouldbereducedbythepaymenttothe
imitatedplayer.
Thischangewouldostensiblyincreasetheuseofinnovation,whichcouldresultin
thediscoveryofbettersolutions,whichcouldbecopiedbyotheruserswhilerewarding
originators.Butweknowfrompreviousexperimentsthatinnovationisrisky,andthegains
ofimprovementscanbecanceledoutbypoorer‐performingsolutionsdiscoveredinthe
process.Wealsoknowfromtheseexperimentsthatcumulativeimprovementsare
importanttothecollectiveperformanceofgroups‐‐ifthischangecausedlessefficient
imitationorlessuseofimitationoverall,thegainsachievedbysuccessfulinnovatorscould
failtopropagateandbearfruitinsequentialimprovements.Thereisalsotheriskthat
competitivemotivationsamongparticipantscouldleadtoexcessivepatentingto
intentionallykeepothersfromimitation,withoutregardtopotentialinnovatorgainsvia
transfersfromimitators.
Previouslaboratoryexperimentsbyothersregardingtheconsequencesofdifferent
systemsfordistributingandrewardingknowledgeproductiondonotgiveusreasonfor
confidenceintheperformanceofexclusionarypatentsystems.Onesuchstudycompared
"patent‐like”(winner‐take‐all,accordingtofirstdiscoveryofthebestsolutionelements)
and"market‐based"rewardsystems(inwhichparticipantscanbuyandsellsharesin
solutioncomponents,effectivelymakingbetsonthesuccessoftheirsolutions)forgroups
attemptingtosolvetheKnapsackProblem,andshowedsubstantiallyinferiorperformance
144
forthepatent‐likesystem(Meloso,Copic,&Bossaerts,2009).Anotherexperimentusinga
multi‐userinventiongametaskshowedthatboth“patent‐only”andmixed"patentand
commons"systemsofcumulativeinventionunderperformeda"commons‐only"systemon
severalmeasures,includinginnovation,productivity,andsocietalutility(Torrance&
Tomlinson,2009).
Wewilldiscusstheseexperimentsfurtherinthenextchapterincomparisonwith
ourownresultsinExperiment4.Inthisexperiment,wehavenotattemptedtoapproach
thelevelofversimilitudetorealpatentsystemsseenintheabovementionedstudies,
becauseweareinterestedintheeffectsofchangingincentivesongeneralizedsearchtasks,
andnotnecessarilyinallofthespecificdetailsofaparticularpotentialpatentsystem.We
expectthattheresultsofthisexperimentwillhavesomerelevancetopatentsystems,as
wellasothersituationsinwhichenvironmentalorinstitutionalfactorschangethe
incentivesofthosewhoexplore.ThegoalsforExperiment4aretocontinueourstudyof
thebasicmechanicsanddynamicsofsociallearningseeninthepreviousexperiments,by
examininghowtheyareinfluencedbyconcreteanddifferentialrewardsfordifferent
searchstrategies,andhowindividual‐levelprocessesaffectgroup‐levelresults.
145
7.Experiment4–CreatureGameC(paymentandprotection)
7.1.Experiment4Overview
ThetaskusedinthisexperimentwasthesameasinExperiment3,withfourmajor
changes:(1)inhalfofthegamesineachexperimentsession,weallowedparticipantsto
forfeitasmallnumberofpointsto“protect”theirguessessothatotherswhousedthesame
guess(oranyverysimilarguess)wereforcedtopayasmallnumberofpointstothe
protector;(2)welengthenedthetimeofeachroundslightly,sothatparticipantswould
havesufficienttimetotakenoteofwhethertheirguesswasprotectedbyanotherplayer
andchangeitiftheywished,andthenumberofgamesineachexperimentsessionwas
decreased,inordertofitthesessionintheallottedtime;(3)participantsreceivedasmall
cashpaymentaccordingtothenumberofpointstheyearnedinthegame;(4)participants
filledoutapost‐tasksurveyabouttheirperceptionsofthetask,andtheirownstrategies
andperformance.
Modification(1)allowedfortheexaminationofdifferencesinstrategiesand
performanceassociatedwithchangingtheincentivesforasocialandsociallearning.
Modification(2)wasintendedtoavoidnoiseinthedataassociatedwithunintendedor
hurriedchoices.Modification(3)allowedfortheexaminationofbehavioraldifferences
(frompreviousexperiments)associatedwithextrinsicmonetaryrewards.Modification(4)
offeredfurtherinsightintothemotivationsandjudgmentsofparticipants,whichcouldbe
correlatedwiththeirbehaviorinthetask.
7.1.1.Predictions
146
Inthegameswherethe“protection”featurewasunavailable,weexpectedlittle
changeinbehaviorandonlyasmallimprovementinperformance(comparedtothevisible‐
scoresconditionofExperiment3)duetothesmalladditionaltimeperroundandthe
expectationofcashrewards.Whentheprotectionfeaturewasavailable,weexpected
differencesinresultscomparedtowhenitwasnotavailable.Thepossibilityofreceiving
additionalpointsfromimitatorswaspredictedtocauseparticipantstoinnovatemorein
ordertofindrelativelyhighscoringsolutions,andthosewithhighscoringsolutionswould
tendtoprotectthemwhenpossible.Thoughthe“fee”forusingaprotectedsolutionwasset
quitelowrelativetotheaveragescoreearnedinasingleround,weexpectedthat
participantswouldavoidimitatingprotectedsolutions,thusreducingimitationbehavior
overall,anddirectingalargershareofimitationtounprotectedlower‐scoringsolutions.
SimilartoExperiment3,thiswasexpectedtoreduceorreversesociallearningbiasessuch
asfrequencyandsimilaritybias.Thecombinationofthesetwoinfluenceswouldslow
convergenceofsolutions,andkeepguessdiversityrelativelyhigh.Asforperformance,we
expectedthattheextraincentivetoexplorewouldresultinhighermaximumscores,but
thelargenumberoflowerscoresencounteredintheprocess,aswellasthedisincentiveto
builduponthe(protected)bettersolutionsfoundthroughthisincreaseinexploration
wouldlowersearchefficiency,thusmakingmeanperformancethesameorlowerthan
whenprotectionwasunavailable.
7.2.Experiment4Methods
159participantswererecruitedfromtheIndianaUniversityPsychology
Departmentundergraduatesubjectpool,andwereofferedcoursecreditfortakingpartin
147
thestudy,aswellasasmallcashpaymentaccordingtotheirperformance(meanpayment
wasapproximatelythreedollars).Participantspopulatedeachsessionbysigningupatwill
forscheduledexperimentswithamaximumcapacityof9persons,andweredistributed
across45sessionsasshowninTable7.1.Oneparticipant’sdata(fromoneofthe8‐person
sessions)wasexcludedduetoextremelyoutlyingperformance.
Table7.1:DistributionofparticipantsacrossgroupsizesinExperiment3
Groupsize 1 2 3 4 5 6 7 8 9
#Sessions 14 8 3 4 5 4 3 2 2
#Participants 14 16 9 16 25 24 21 16 18
Thetaskusedwasnearlyidenticaltothatofthevisible‐scoresconditionin
Experiment3,withthefollowingchanges.
Protectionavailability:inhalfofthegamesineachsession(theprotectionavailable
orPAcondition),participantswerepresentedwithachoiceforsevensecondsattheendof
eachround(justafterlearningthescoreassociatedwiththeirsolution)toforfeitfour
points(approximately.042intermsofnormalizedscore)to“protect”thesolution(ifthey
madenoselectionwhilethechoicewasavailable,thedefaultwasnottoprotect).Forthe
nextthreeroundsafterasolutionwasprotected,anyotherplayerwhoused(attheendof
theround)asolutionwithinoneelementdifferenceoftheprotectedsolutionwould
automaticallytransfertwopoints(approximately.021innormalizedscore)tothe
protector.Solutionswithinoneelementdifferenceofaprotectedsolutioncouldnotbe
protected.
148
Theprotectionfeeandusefee,respectively,wereintentionallysetatfairlylow
valuesinordertoensurethatthecostfortheseoptionswouldbevisiblebutwouldnot
undulydiscourageparticipantsfromusingthem,orexcessivelywarpincentivestoward
oneoptionortheother.Thewordingoftheoptionspresentedtoparticipantsattheendof
eachroundisshowninAppendix7.A.Weensuredthatparticipantsunderstoodthese
optionswithadditionalinstructionsandhands‐ondemonstrationsbeforedatagathering
began.Inordertoensurethatparticipantswereawareofthedecisionstheymade
concerningtheuseofprotectedteamsduringeachround,anypeers’teamsthatwere
currentlyprotectedwereshownwiththeirbackgroundhighlightedinbrightred.In
addition,whenaparticipantchangedtheirteamsuchthatitwasprotectedbyanother
player,itsbackgroundwasalsohighlightedinbrightred,andthemessage“Protectedby
anotherplayer”wasshownaboveit(thesenotificationsoccurredassoonasthechange
wasmade,notattheendoftheround).Ineachround,werecordedwhethereach
participantprotectedhisorherteamorusedanother’sprotectedteam,andhowmany
others(ifany)usedtheplayer’sprotectedteam.Intheotherhalfofthegamesineach
session(theprotectionunavailableorPUcondition),theprotectionoptionwasnot
availabletoparticipants,andgameswereplayedastheywereintheScoresVisible
conditionfromExperiment3.Gameswereplayedinarandomorderineachsession.
Longerroundsandfewergames:togiveparticipantssufficienttimetoevaluatethis
choiceandchangetheirguessesifdesiredintheprotectionavailablecondition,andtobe
abletoproperlycomparebehaviorbetweenconditions,welengthenedeachroundfrom10
to15secondsinbothconditions.Becauseofthelongerrounds(includingtheadditional
secondsgiventomaketheprotectionchoiceaftereachround),additionalpracticetimefor
149
thenewgamefeatures,andthepost‐tasksurvey(discussedbelow),wereducedthe
numberofgamesineachexperimentsessionfromsixtofourtofitthesessionintheone‐
hourtimelimitrequiredforexperimentsusingoursubjectpool.Ofcourse,thedistinction
betweengamesintheprotectionavailableandprotectionunavailableconditionsonly
matteredinsessionsthatincludedmorethanoneparticipant,butroundsandsessions
wereidenticalinlengthforallparticipantgroupsizes.
Payment:participantswereinformedbeforeeachsessionthattheywouldreceivea
smallcashpaymentof“afewdollars”whichwouldvaryaccordingtothetotalnumberof
pointstheyearnedindividuallyduringthesession.Wesetthepaymentperpoint
($0.00086)accordingtothedistributionofscoresfromExperiment3,suchthatthemean
paymentwouldbeapproximately$3.00.
Posttasksurvey:wedeviseda10‐itemquestionnaireforparticipantstocomplete
afterthefinalgameineachsessioninordertogatherself‐reportdataonparticipants’
attitudesandjudgmentsaboutthetaskandtheirstrategies,whichincluded8multiple‐
choicequestionswithresponseoptionsonaLikertscale,andtwofree‐responseitems.The
fulltextofthesurveyisshowninAppendix7.B.
Thedistributionofindividualpointvaluesandinteractionsfortheiconswasthe
sameasinExperiment3,andscoreswereagainnormalizedfrom[‐6,88]to[0,1]foreaseof
analysis.Notethatbecausethenormalizationdidnotincludetheincreasedrangepossible
fromthetransferofprotection‐relatedfees,itwaspossiblefornormalizedscorestogo
beyond[0,1],butthisdidnotoccur.Unadjustedscorereferstovalueswhichdonotinclude
thesubtractionofprotectionorusefeespaidortheadditionofusefeesreceived,
correspondingtotheplainvalueofthesolutionaccordingtoitselements;adjustedscore
150
referstothesamevalueswiththeabovefeesaddedorsubtracted(whenpresent).All
analysesofscorewillrefertounadjustedscoreexceptwherenoted;ingeneraltherewere
nodifferencesbetweenresultsforadjustedandunadjustedscores.
7.3.Experiment4Results
7.3.1Overallmeans
MeandependentvariablesineachconditionareshowninTable7.2(seealsoFig.
7.1).Forgroupedparticipants,meanoverallandfinalscoresineachconditionwerenearly
identicaltoeachotherandtothoseobservedinthevisible‐scoresconditioninExperiment
3.Onaverage,minimumscoreswerelowerandmaximumscoreswerehigherinthe
protectionavailablecondition,butneitherdifferencewassignificant.Isolatedparticipants
achievedslightlyhighermeanoverallandfinalscoresthanthoseinExperiment3.
Figure7.1:Distributionofscoresforallpossibleteams.
151
Table7.2:Meanscore,guessdiversity,andchoicesourceproportionsbycondition
(PU:ProtectionUnavailable,PA:ProtectionAvailable)
Cond.OverallScore(Percentile)
FinalScore(Percentile)
GuessDiversity
Imitation Innovation Retention Retrieval
PU .451(93.3%) .525(97.5%) 60.1%* 6.9%* 14.6%* 75.6%* 2.2%
PA .446(92.6%) .530(97.5%) 65.9%* 6.2%* 17.8%* 72.9%* 2.3%
IsolatedPartic. .376(86.1%) .466(94.0%) ‐‐ ‐‐ 26.1% 61.0% 12.2%
*significantdifferencesbetweenconditions
7.3.2.Rounds
Linearmixed‐effectsmodelswereusedtoexaminetrendsacrossroundsforscore
andguessdiversity,witharandomeffectofparticipantgroup.Analysisofscoreversus
roundshowedverysimilarpositivetrendsforgroupedparticipantsineachcondition(PU:
F(1,712)=208.87,p<.0001,B=.627;PA:F(1,712)=283.72,p<.0001,B=.670).Isolated
participants’scoresincreasedsignificantlyacrossroundsaswell,butsignificantlylessso
thanforgroupedparticipants(F(1,657)=95.52,p<.0001,B=.498,meanincrease=0.150;see
Fig.5.4).Guessdiversityshowedcorrespondingdecreasesacrossroundsinboth
conditions(PU:F(1,712)=169.33,p<.0001,B=‐0.458;PA:F(1,712)=165.94,p<.0001,B=‐
.485;seeFig.7.2).
152
Figure7.2:Scoresincreasedandguessdiversitydecreasedataboutthesameratewhether
protectionwasavailableorunavailable.
Trendsforchoicesourcesacrossroundsshowedverysimilarpatternstothosein
Experiments2and3,withapproximatelythesamemagnitudeandsignificance:Imitation
andInnovationdecreasedsignificantlyoverrounds,andRetrievalandRetentionincreased
significantly.Therewasnosignificantdifferenceinslopeforanychoicesourcebetween
conditions,norasubstantialdifferenceincomparisontotheslopeoverroundsforeach
choicesourcefoundinExperiment3.Protectiondecreasedacrossrounds(F(1,712)=17.54,
p<.0001,B=‐0.170),whileuseofprotectedteamsincreased(F(1,712)=49.26,p<.0001,
B=0.331;seeFig.7.3).
153
Figure7.3:Protectiondecreasedacrossrounds,whileuseofprotectedteamsincreased.
7.3.3.Gameorder
Similarlinearmixed‐effectsmodelswereusedtoexaminetrendsacrossgameorder
foreachdependentvariablewithinconditions.Forthisanalysis,thegameordervaluefor
eachgamewascorrectedtoitsorderwithinthecondition,i.e.Game1or2ineach
condition.Scoreshowednosignificantchangeacrossgameorderineithercondition,and
guessdiversitydisplayedadecreaseacrossgameorderinbothconditions(PU:
F(1,30)=5.94,p=.021;PA:F(1,30)=4.76,p=.037;seeFig.7.5).Therewerenosignificant
changesacrossgameorderforanychoicesourceineithercondition(seeFig.7.5).
Withintheprotectionavailablecondition,protectiondecreasedacrossgameorder
(F(1,26)=9.79,p=.0043,B=‐0.217,meanchg.=‐0.020),butuseofprotected teamsdidnot
changesignificantly.
154
Figure7.4:Scoredidnotchangesignificantlyineithercondition,andguessdiversity
decreasedinbothconditions.
(a) (b)
Figure7.5:(a)ImitationandInnovationdidnotshowsignificantchangesacrossgame
orderineithercondition,nordid(b)RetentionorRetrieval.
155
Figure7.6:Protectiondecreasedsignificantlyacrossgameorder,whileuseofprotected
teamsshowednosignificantchange.
7.3.4.Groupsize
Trendsacrossparticipantgroupsizeforeachdependentvariablewithinconditions
wereexaminedusinglinearmixed‐effectsmodels,withtheparticipantgroupusedasa
randomeffectontheintercept.Slopesacrossgroupsizedidnotshowdifferencesbetween
conditionsforanydependentvariable,exceptwherenoted.Scoreincreasedsignificantly
withgroupsizeinbothconditions(PU:F(1,43)=27.19,p<.0001,B=0.512;PA:
F(1,43)=36.69,p<.0001,B=0.568;seeFig.7.7),andguessdiversityshoweda
correspondingdecreaseinbothconditions(PU:F(1,29)=28.96,p<.0001,B=‐0.661;PA:
F(1,29)=34.48,p<.0001,B=‐0.672;seeFig.7.7).Asforchoicesources,Imitationincreased
significantlyforlargergroupsinbothconditions(PU:F(1,29)=37.72,p<.0001,B=0.680;PA:
F(1,29)=37.20,p<.0001,B=0.699;seeFig.7.8a),whileInnovationdecreasedsignificantly
156
onlyintheprotectionavailablecondition(F(1,29)=8.94,p=.006,B=‐0.437;seeFig.7.8a).
NeitherRetentionnorRetrievalshowedsignificantchangesacrossgroupsize(seeFig.
7.8b).
Protectiondidnotshowasignificanttrendacrossgroupsize,whileuseofprotected
teamsincreasedinlargergroups(F(1,29)=37.45,p<.0001,B=0.751;seeFig.7.9).
Figure7.7:Asparticipantgroupsizeincreased,meanscoresinagroupincreased,andthe
diversityofofferedsolutionsdecreased.
157
(a) (b)
Figure7.8:Asparticipantgroupsizeincreased,(a)meanproportionsofImitationincreased
inbothconditionsandInnovationdecreasedonlyintheprotectionavailablecondition,and
(b)RetentionandRetrievalshowednosignificantchange.
Fig.7.9:Asparticipantgroupsizeincreased,protectiondidnotchangesignificantly,butuse
ofprotectedteamsincreased.
158
7.3.5.Differencesinimitation
Acrossbothconditions,approximately94.7%ofallguessesincludingImitation
wereofasingleparticipant.Atthetimeofsuchsingle‐sourceimitations,thescoreofthe
imitatedparticipantwasgreaterthanthatoftheimitatorinapproximately90%ofcasesin
bothconditions(seeFig.7.10a).Ofallinstancesofsingle‐participantimitation,the
probabilityofimitatingthetop‐scoringsolutioninthegroupwassignificantlyhigherinthe
protectionunavailableconditionthantheprotectionavailablecondition(t(34)=5.27,
p<.0001;seeFig.7.10b).Ananalysisoftheprobabilityofprotectionacrossscorerank
showedthatthebesttwosolutionspresentinanygivenroundwerenearlyalways
protected(seeFig.7.10c).
Toexamineseparatelyhowoftenandhowmuchparticipantsimitatedoneanother,
wemeasuredthemeanincidenceofguessesinwhichtherewasgreaterthanzeroImitation
(Imitationincidence),aswellasthemeanuseofImitationinthosecaseswhereitwas
greaterthanzero(Imitationproportion).MeanImitationincidencewassignificantlyhigher
intheprotectionunavailablecondition(0.205vs.0.162;F(1,92)=19.37,p<.0001),but
meanImitationproportionwaslower(0.341vs.0.390;F(1,89)=7.28,p=.008),andthe
distributionofmeanimitationproportionswasweightedslightlybutsignificantlymore
heavilytowardhighervaluesintheprotectionavailablecondition,asshownbya
Kolmogorov‐Smirnofftestofequalityofdistributions(D=0.1116,p<.0001;seeFig.7.11).
Inotherwords,participantsintheprotectionavailableconditioncopiedoneanotherless
frequentlybutinlargeramountsatatime.
159
(a) (b)
(c)
Figure7.10:(a)Therewerestrongbiasestowardimitatingbetter‐scoringparticipantsthan
oneselfinbothconditions,and(b)asignificantlystrongerbiastowardimitatingthebest‐
scoringparticipantintheprotectionunavailableconditionthanintheprotectionavailable
condition;and(c)arelatedanalysisshowedthatthebesttwosolutionsinanygivenround
werehighlylikelytobeprotected.
160
Figure7.11:ForguessesthatincludedatleastsomeImitation,participantsinthe
protectionavailableconditionhadhigherproportionsofImitationintheirguesses.
5.3.6.Choicesourcestrategy
AsinExperiment2,thechoicesourcesofeachnon‐isolatedparticipantoverthe
entiresessionwereanalyzed,andeachparticipant’schoicesourcestrategywascategorized
accordingtotheirproportionofeachsource.Participantswhosechoicescontainedone
sourceinanaverageproportiongreaterthantheglobalaverageforthatsourceplusone
standarddeviation,werelabeledwiththatstrategy.Forexample,aplayerwhoseguesses
overthecourseofasessionconsistedofagreaterproportionofImitatechoicesthanthe
averageforallotherparticipantsintheexperiment,plusonestandarddeviation,were
labeledashavinganoverallstrategyof“Imitate.”Thosewhofittheabovecriteriaformore
thanonechoicesource,ornone,werelabeledashavinga“Mixed”strategy.Thescore
distributionforeachstrategycategoryineachconditionisshowninFig.7.12.
161
(a) (b)
Figure7.12:Scorevs.choicesourcestrategyin(a)protectionunavailableand(b)
protectionavailableconditions,showingthataconservativehigh‐Retentionstrategy
resultedinthebestperformanceinbothconditions,thoughhighlyImitativestrategiesalso
performedwell.
Alinearregressionofmeanindividualscorevs.meanindividualImitationguess
proportionshowedasignificantpositiverelationship(thatis,thegreateraparticipant’s
averageproportionofImitation,thebettertheparticipant’sscore)inbothconditions(PU:
F(1,142)=7.91,p=.006,B=0.230;PA:F(1,142)=3.96,p=.048,B=0.165;seeFig.7.13a).The
oppositewastrueforindividualscorevs.Innovation,whichdisplayedasignificant
negativerelationshipinbothconditions(PU:F(1,142)=120.30,p<.0001,B=‐0.677;PA:
F(1,142)=98.01,p<.0001,B=‐0.639;seeFig.7.13b).Apositiverelationshipheldfor
Retentioninbothconditions(PU:F(1,142)=56.24,p<.0001,B=0.533;PA:F(1,142)=53.38,
p<.0001,B=0.523;seeFig.7.13c).Finally,anegativerelationshipwasfoundforRetrievalin
162
bothconditions(PU:F(1,142)=4.88,p=.029,B=‐0.182;PA:F(1,142)=7.15,p=.008,B=‐
0.219;seeFig.7.13d).
AsinExperiment2,analysesofmeangroupscorevs.meangroupguessproportion
foreachchoicesourceshowedsimilarrelationshipsofthesamesignificanceanddirections
asthosenotedabove.Analysesofmeanindividualscorevs.meangroup(excludingthe
individual)guessproportionshowedsimilarrelationshipsaswell,withtheexceptionofthe
absenceofsignificantrelationshipsforRetentionorRetrieval.Plotsoftheseareomittedfor
clarity.Alltrendsnotedaboveweregenerallymonotonic;thatis,therewerenothresholds
orinflectionpointsbeyondwhichtherelationshipschanged.
Similaranalysesofprotectionshowednosignificantrelationshipwithscorefor
individuals(seeFig.7.14a),orattheothertwolevelsdiscussedabove(groupand
individualvs.groupothers),whilescorevs.useofprotectedteamsshowedasignificant
positiverelationshipforindividuals(F(1,142)=67.42,p<.0001,B=0.567;seeFig.7.14b),as
wellasforgroups(F(1,60)=72.99,p<.0001,B=0.740),andforindividualscoreversus
protectionbyothersinthegroup,excludingtheindividual(F(1,142)=40.86,p<.0001,
B=0.471).Allofthesamerelationshipswerefoundforbothadjustedandunadjustedscore.
Plotsoftheseareomittedforclarity.
163
(a) (b)
(c) (d)
Figure7.13:Higherindividualscoreswereassociatedwith(a)higherindividualImitation,
(b)lowerInnovation,(c)higherRetention,and(d)lowerRetrieval.
164
(a) (b)
Figure7.14:(a)Protectionshowednosignificantrelationshipwithscore,while(b)higher
useofprotectedteamswasassociatedwithhigherscores.
7.3.7.Improvements
Asinpreviousexperiments,improvementsweretalliedforeachparticipantineach
sessionandcondition.Histogramsofnormalizedimprovementshareinbothconditions
showedalessequitabledistributionofimprovementswithingroups(moreparticipants
withzeroimprovements,andfewerwithsharesnear1)relativetothescores‐visible
conditioninExperiment3(compareFigs.7.15and5.13).AKolmogorov‐Smirnofftestof
equalityofdistributionsindicatedthatthedistributionsineachconditionwerenot
significantlydifferent.However,theuppertailofthedistributionwasessentiallyshifted
upwards(i.e.thehighestimprovementsharesinthePUconditionwerereplacedbyeven
highersharesinthePAcondition),thoughthehighestachieversineachconditionwere
entirelydifferentsetsofindividuals.Meanoverallscorealsoshowedastrongpositive
165
correlationwithimprovementshareinthePAcondition(F(1,112)=49.98,p<.0001,
B=0.348),butthisrelationshipwasnotevidentinthePUcondition.
Themeanchoicesourceproportionsforguessesthatresultedinscore
improvementsandthosethatdidnotareshowninTable7.3.Intheprotectionunavailable
condition,guessesthatyieldedimprovementshadhigherInnovation(t(417)=‐9.78,
p<.0001)andlowerRetention(t(395)=6.01,p<.0001)relativetonon‐improvements.Inthe
protectionavailablecondition,therewashigherImitation(t(549)=‐2.78,p=.006),lower
Innovation(t(574)=2.86,p=.004),andlowerRetrieval(t(682)=3.37,p=.0008)for
improvementsthannon‐improvements.Ofallimprovements,aboutthesameproportionin
eachconditionresultedfromguessesthatincludedImitation(PU:.252vs.PA:.224).
In51.2%ofimprovementsintheprotectionunavailableconditionand43.9%inthe
protectionavailablecondition,thefocalplayerimitatedatleastonepeerwhohad
previouslyimitatedthefocalplayer.Inotherwords,aplayerwhowasimitatedbyanother
playeroftenlaterimitatedthatsameplayerinthecourseofcreatinganimprovement,and
thishappenedmoreoftenwhenprotectionwasunavailable.
166
Figure7.15:Histogramsshowingdistributionsofimprovementswithingroupsineach
condition.
Table7.3:Meanchoicesourceproportionsforimprovementandnon‐improvement
guessesineachcondition.
Condition Improvement %ofguesses Imitation Innovation Retention RetrievalNo 94.7% 7.9% 13.6%* 75.7%* 2.0%PU Yes 5.3% 9.1% 20.1%* 69.0%* 1.7%No 92.5% 7.0%* 16.8%* 73.0% 2.3%PA Yes 7.5% 10.0%* 14.4%* 73.9% 1.3%
*significantdifferencesbetweenimprovementsandnonimprovementswithin
condition
7.3.8.Guesssimilarity
Acomparisonbetweenthemeansimilarityofparticipants’mostrecentguessesto
thosewhomtheyimitated,andtothosewhomtheydidnotimitate,revealedslightbut
significantdifferencesinbothconditions.Intheprotectionunavailablecondition,there
167
weresimilarityvaluesof.576forimitatedvs..545fornon‐imitatedguesses(t(2511)=‐2.90,
p=.0037;seeFig.7.16a).Inthescores‐invisiblecondition,thereweresimilarityvaluesof
.491forimitatedvs..457fornon‐imitatedguesses(t(2051)=‐3.18,p=.0014;seeFig.7.16b).
Inotherwords,priortoimitation,theaverageimitators’guesswasmoresimilartothatof
theimitatedparticipant(s)thantothoseofothersintheprotectionunavailablecondition,
andthisrelationwasfoundbutwasslightlylessprominentintheprotectionavailable
condition.
(a) (b)
Figure7.16:(a)Intheprotectionunavailablecondition,imitators’previousguesses
showedgreatersimilaritytotheguessestheyimitatedthantothosetheydidnotimitate,
while(b)inthescores‐invisiblecondition,asimilarbutweakereffectwasobserved.
7.3.9Frequencyandmomentumbias
AsinExperiment1,wemeasuredthebiasofparticipantstochooseanicon
accordingtoitsfrequencyinpeers’choices.Toreiteratebriefly,wemeasuredthemean
168
probabilityofImitationandInnovationforanyiconnotalreadyincludedonaplayer’s
team,basedonthefrequencyofitsappearanceonpeers’teamsintheplayer’sdisplay,and
comparedthemtoexpectedchancebaselines.
Linearmixed‐effectsanalysisofimitationprobabilityversuschoicefrequency
showedapositivefrequency‐dependentImitationbiasthatwassignificantlygreaterthan
chanceintheprotectionunavailablecondition(F(1,258)=355.39,p<.0001,B=.722),but
significantlylowerthanchanceintheprotectionavailablecondition(F(1,258)=387.81,
p<.0001,B=.714;seeFig.7.17a);bothbecameapparentathighervaluesofchoice
frequencythaninExperiment3.Thereweresimilarpositivefrequency‐dependent
Innovationbiasesabovechanceinbothconditions(PU:F(1,258)=133.15,p<.0001,B=.517;
PA:F(1,258)=192.79,p<.0001,B=.601;seeFig.7.17b).
Wealsorepeatedtheanalysisof“choicemomentum,”bytallyingthechangeinthe
numberofplayerswhoseteamsincludedtheiconintheprevioustworounds,aswellas
thenumberoftheremainingplayerswhoaddedittotheirteaminthecurrentroundvia
ImitationorInnovation,andnormalizingforgroupsize.Afterlog‐transformingthe
Imitationprobabilitydatatoachieveanormaldistribution,t‐testsofImitationprobability
fornegativeandpositivechangesinchoicefrequencyshowedsimilarsignificantpositive
momentumbiasesinbothconditions(PU:t(304)=‐8.66,p<.0001;PA:t(299)=7.76,
p<.0001;seeFig.7.18a).SlightpositivemomentumbiaseswerealsofoundforInnovation
inbothconditions(PU:t(385)=‐3.40,p=.0007;PA:t(346)=‐5.54,p<.0001;seeFig.7.18b).
169
(a) (b)
Figure7.17:(a)Therewerebiasestowardchoosingelementsthatweremorefrequently
representedonotherteamsintheprotectionunavailablecondition,andlessfrequently
representedonotherteamsintheprotectionavailableconditionforImitationand(b)
biasestowardmorefrequentelementsforInnovationdecisionsinbothconditions.
(a) (b)
Figure7.18:Therewerebiasestowardchoosingelementswhoserepresentationonother
teamswasincreasinginbothconditionsfor(a)Imitationand(b)Innovationdecisions.
170
7.3.10.Surveyresponses
Thesurveydistributedtoparticipantsaftertheendofthelastgameineachsession
isincludedinappendix7.A.Thefirsteightquestionsonthesurveyweremultiple‐choice,
withpossibleresponsesona7‐choiceLikertscale,coded1‐7.Questions2,3,and4were
relatedtoprotectionandhadanadditionaloptionfor“doesnotapply,”whichwascodedas
0.Thelasttwoquestionswerefree‐response;oneaskedparticipantstodescribetheir
strategies,andmostresponsesweresomevariationona“copy‐if‐betterandthenmake
smallchanges”strategy;theotheraskedforanyfurthercomments,andmostresponses
wereeitherblank,orshort(generallypositive)commentslike“fun”orinteresting.”For
quickreference,theeightmultiple‐choicequestionsareshowninTable7.4.Stripchartsof
responseswithkerneldensityestimatesareshowninFig.7.19.
Table7.4:Multiple‐choicesurveyquestions,withresponsesassociatedwiththeextreme
lowandhighendsofthescale.Thosesignificantlyassociatedwithhigherindividualscores
areshowninbold.
increaseyourownscore(withoutregardtothescoresotherplayersmighthave)Q1 Duringthegame,didyoumostlytryto…?
scorehigherthanotherplayers
keepothersfromusingyoursolutionsQ2 If/whenyouprotectedyoursolutions,wasitmostlyto…?
getpaymentforothers'useofyoursolutions
copythewholeteamandpaythefeeQ3 Whenyouwantedtouseanotherplayer’sprotectedteam,didyougenerallypreferto…?
copyitandthenchangeittoavoidpaying
NeverQ4
Whenyoufoundateamwitharelativelyhighscore,didyoueverintentionallyleaveitunprotectedsothatotherscouldcopyitforfree? Everytime
171
VeryeasyQ5 Howdidyoufeelaboutthedifficultyofthetask?Veryhard
otherplayers’teamsQ6 Didyouchoosecreaturesmoreoftenfrom…?theleague
otherplayers’teamsQ7 Didyoufindthatyourscoreswerebetterwhenyouchosecreaturesfrom…?
theleague
LuckQ8 Doyoufeelthatsuccessinthistaskwasbasedmoreon…?
Skill
(a)
172
(b)
(c)
Figure7.19:Responsestothemultiple‐choicesurveyquestionsinTable7.4,plottedas
stripchartswithkerneldensityestimatesandmeanlines,separatedasfollowsforclarity:
responsesfromgroupedparticipantsto(a)non‐protection‐relatedquestionsand(b)
protection‐relatedquestions,and(c)responsesfromisolatedparticipantstoQ5andQ8.
173
Relationshipsbetweenplayerscoresandresponsestoeachquestionwereanalyzed
usingamixed‐effectslinearmodelwithparticipantgroupasarandomeffect.Higherscores
wereassociatedwithresponsesattheupperendofQ1(scorebetterthanothers)
(F(1,256)=22.68,p<.0001,B=0.236),theupperendofQ2(getpaymentfromothers)
(F(1,252)=6.39,p=.0122,B=0.140),thelowerendofQ3(copythewholeteamandpay)
(F(1,243)=14.91,p=.0001,B=‐0.210),thelowerendofQ4(neverleaveunprotected)
(F(1,242)=6.73,p=.0100,B=‐0.142),thelowerendofQ6(choosefromothers’teams)
(F(1,254)=19.88,p<.0001,B=‐0.225),andtheupperendofQ8(performancebasedonskill)
(F(1,254)=8.15,p=.0047,B=0.148).Inexaminationsofresponsetrendsacrossparticipant
groupsize,wefoundjustonesignificanttrend:responsestoQ2weresignificantlyhigher
(getpaymentfromothers)inlargergroups(F(1,29)=8.88,p=.0058,B=0.243).
174
7.4.Experiment4Discussion
7.4.1.SimilaritieswithExperiment3
Aspredicted,resultsintheprotectionunavailableconditionofExperiment4(E4PU)
weregenerallysimilartothoseinthevisible‐scoresconditionofExperiment3(E3SV).
PerformancewasslightlybetterinE4PUthanE3SVaspredicted,whichmayhavebeendue
totheincentiveofcashpayment,ortheextratimetomakedecisionsabouteachround's
candidatesolution,thoughunfortunatelywecannotdifferentiatebetweenthesetwo
factorsascauses.IsolatedparticipantsinE4alsodidbetterthanthoseinE3,which
reinforcesthisexplanation.ImitationwasmarkedlylowerandInnovationhigherinE4PU
thaninE3SV,whichmayhavebeenacarryovereffectfromthePAcondition,inwhichthese
differenceswereevenmoreextreme.TrendsofRetentionandRetrievalovergroupsize
thatwerepresentinE3SVwerenotpresentinE4PU,andtherewasanegativecorrelation
ofscorewithRetrievalinE4PUthatwasnotpresentinE3SV,butineachcasethemissing
relationshipwasnominallypresentintheotherexperiment,butnotsignificant.Theless
equitabledistributionofimprovementshareinE4PUrelativetoE3SVcouldbeattributed
toadifferentialcrowding‐out(FreyandJegen,2001)ofintrinsicmotivationtoachieve
improvementsforsomeparticipantsandnotothersbythecashpaymentofferedinE4.
7.4.2.Performance
PredictionsconcerningthePAconditionweregenerallyborneoutaswell.
Maximumscoreswereslightlyhigher,butoverallscoreswereactuallyslightlylowerinthe
PAcondition,thoughneitherdifferencewassignificant.This,alongwiththeincreased
Innovation,decreasedImitation,andincreasedguessdiversityrelativetoPU,accordswith
175
thepredictionsmadeforthiscondition:thegreaterincentivesforexplorationledto
increasesinInnovationandlessconvergence,andsomegoodreturnsintheformofslightly
moreimprovements.However,thegreatergeneralriskofexploration,aswellasless
efficientpropagationofbetterperformingsolutionsthroughImitation(becausethebest
solutionsweregenerallyprotected)ledtoawashingoutofthebenefitsofimprovements
bylowerscoresencounteredalongtheway,ratherthanfurtherimprovementofthebetter
solutionsthatwerefound.
7.4.3.Strategy
Aspredicted,theavailabilityofteamprotectionalsoaffectedtheuseofotherkinds
ofinformationinimitationstrategies:PAparticipantsshowedweakersimilaritybias,anda
reversaloffrequencybias,atleastforImitation.Theformercouldhavebeenduetoan
inclinationtowardsaltationiststrategiesasseeninE3SI,whilethelattercouldhavebeen
duetohigher‐frequencyiconstypicallybeingonprotectedteams.Thedecreasein
ImitationincidenceandincreaseinImitationproportionwhenprotectionwasavailable
indicatesthatthefeeforimitationwasadisincentiveforImitation,butwhenImitationwas
pursued,participantseitherkepttheimitatedguessunchanged(totakeadvantageofthe
presumedimprovementinscoreand“gettheirmoney’sworth”fortheusefee),orchanged
relativelyless(inorderto“inventaround”theprotectedsolutionwithoutasmuchriskof
loweringthescore).
Theassociationofhigherscoreswithhigheruseofprotectedteamswaslikelydue
tothefactthattheusefeewasquitesmallrelativetothepotentialbenefitinmostcases.
Thelackofarelationshipbetweenprotectionandscorecouldhavebeenduetogeneral
176
overuseofprotection;thetypicalresponsetoQ4onthesurveyindicatesthatformost
participantsthedefaultactionupondiscoveringahigh‐scoringsolutionwastoprotectit,
andthisisborneoutinFigure7.10c.InthePAcondition,theincreaseduseofImitationin
improvementsrelativetonon‐improvementsshowsthatforthosewhorecognizedthe
valueofusingprotectedteams,Imitationwasquiteaproductivestrategy.
Thedecreaseinprotectionandincreaseinuseofprotectedteamsoverroundscan
bothbeexplainedbythefactthatimprovementsavailablefordiscoverybecame
increasinglyrareasparticipantsdiscoveredbettersolutionsandmovedhigherinthescore
distribution.Protectiondecreasedbecausefewerandfewerguessesyieldedprotection‐
worthy,relativelyhigherscores,andtheuseofprotectedteamsincreasedbecausethe
dwindlingsupplyofscore‐improvingsolutionsstokedthedemandforusingthoseofothers
andmadeparticipantsmorewillingtopaytherelatedfee.Theincreaseinuseofprotected
teamsacrossparticipantgroupsizecanbeconsideredanextensionofthegenerallygreater
Imitationinlargergroups.
Thoughtherewasnosignificantchangeinscoreorchoicesourcesacrossgame
orderwithinconditions,thedecreaseinthenumberofgamesperconditionrelativeto
previousexperimentsmadeforlessdatainwhichtodetecttrendsofthissort.The
decreaseinprotectionovergameorderwaslikelyanadaptationtothefailureofthe
always‐protectstrategy(displayedinQ4responsesandthenear‐universalprotectionof
thebest‐scoringsolutions)toturnaconsistentprofit(asshownintheflatscoreversus
protectionresults).
7.4.4Self‐reportedmotivationsandmethods
177
Despitereceivinginstructionsthatthegoalofthetaskwasmerelytomaximizeone’s
owntotalscoreoverthecourseofeachgame,aswellasbeinginformedthatpaymentwas
contingentonlyonindividualperformance,responsestoQ1revealedthatmany
participantswerejustasmotivated(ormore)toincreasetheirperformancerelativeto
others.Thosewhoreportedsuchrelativeperformancemotivationalsotendedtoscore
higherthanothers,asdidthosewhosawperformanceinthegameasamatterofskill
ratherthanluckinQ8(thoughmostparticipantsbelieveditwasthelatter).Q2showedthat
participantsweresomewhatevenlydividedbetweenprotectingtokeepothersfromusing
theirsolutionsandgettingpaymentforthem,andresponsestoQ3showedthat
participantswerealsofairlyevenlydividedbetween“licensing”and“inventingaround”
others’protectedteams.
7.5.Conclusions
7.5.1.Similaritiesanddifferenceswithpreviousresults
Inthisexperiment,wefoundpatternsofresultsverysimilartothoseinprevious
experiments,buttherewerealsosomecuriousnewcombinationsoftheseeffects.
Specifically,weobservedinbothconditionstheby‐nowfamiliarresultthataconservative
approachtothistaskpaidoff:relativelyhighusageofImitationandRetentionwere
associatedwithhighscores,andInnovationandRetrievalwithlowscores;trends
displayingtheserelationshipsdifferedverylittleacrossconditions.
Besidestheresemblancebetweentheprotectionunavailableconditionof
Experiment4(E4PU)andExperiment3(E3SV)notedearlier,theprotectionavailable
condition(E4PA)showedsomeresultsthatresembledthoseoftheinvisible‐scores
178
conditioninExperiment3(E3SI)aswell.Forinstance,therewasareducedtendencyto
imitatetop‐scoringsolutionsinE3SIwasduetotheimpossibilityofdistinguishingthe
qualityofsolutions,andinE4PAduetothedisincentiveofpayingtheusefeefortop‐
scoringsolutions(whichweregenerallyprotected).Therewasalsoincreaseddiversityand
innovationrelativetothealternateconditionforboth,andadecreaseintheuseof
similarity‐andfrequency‐biasedimitation.However,thesechangesledtothediscoveryof
fewerimprovementsinE3SIrelativetoE3SV,whilethereweremorefoundinE4PA
relativetoE4PU,thoughthisdidnotleadtogreateroverallorfinalscores.Overall,it
appearsthattheextraincentiveforInnovationinE4PAledtosomeimprovementswhose
benefitswerecanceledoutbytheworsesolutionsencounteredintheexplorationprocess,
andtherewasdisincentiveforimitationwhichresultedinfewersequentialimprovements,
whichmayhaveimprovedaverageperformancewithlessrisk.
7.5.2.Similaritiesanddifferencesbetweenexperimentalandreal‐worldpatentsystems
Thisexperimentfeaturedapatent‐likesystemforgoverningtheuseofinnovations,
butwasinnowaymeanttoaccuratelyrepresentthecomplexitiesofarealpatentsystem.
Majordifferencesbetweenthistaskandsuchrealsystemsinclude:ourautomatic
“licensing”ofpatentedinnovations,withnoabilityforpatenteestorefusetogranta
license,negotiatethelicensingfee,orinfringeothers’patentswithoutlicensing;
unambiguouspatentscopeandnoabilitytovaryordisputethebreadthofpatented
features;nodistinctionbetweendiscoveryandproductionofinnovations;noabilityto
“opensource”innovationssuchthattheycouldbeusedfreelybyothersbutnotpatented;
andnoabilitytosellorotherwisetransferpatentrightsbetweenparticipants.Wefeltthat
179
testingtheeffectofachangeinincentivestoinnovateinamoregeneralsearchdomain,and
establishingcontinuitywithourpreviousexperimentsinthisdomainweremoreimportant
thanattemptingtoreplicatemorefeaturesofpatentsystems.However,anexperimentthat
didincludeallofthefeaturesnotedaboveactuallyfoundthatacompletelycommons‐based
(patent‐free)systemoutperformedbothapatent‐onlysystemandamixedpatent‐
commonssystem(Torrance&Tomlinson,2009).
Besidesthedistinctionbetweendiscoveryandproductionofinnovationsinthereal
world,thereisalsothefactthatphysicallyproducingmanyinnovativeproductsdepends
onrivalrousrawmaterialresources.Again,therewasnoprovisionforthisfactinour
experiment,butitwasincorporatedinanotherstudytocreateamarket‐basedmechanism
forpromotinginnovation(Meloso,Copic,&Bossaerts,2009).Similartoourexperiment,
participantsinthisstudyexploredvaryingcombinationsofoptionstosolveaproblem;the
primarydifferencewasthatinthe“patent”condition,amonetaryprizewasgiventothe
firsttodiscovertheoptimalsolution,andinthe“market”condition,participantsbought
andsoldsharesineachoption,andsharesofoptionscontainedintheoptimalsolutionpaid
amonetarydividendattheendofthegame.Melosoetal.(2009)foundnobenefitforthe
patent‐basedsystemoverthemarket‐basedsystem:theoptimalsolutionwasfoundinan
equalnumberofrunsofeachcondition.However,theoptimalsolutionwasfoundbya
greaterproportionofparticipantsinthe“market”condition.Theincentiveofprofiting
fromthesalesofsharesintheoptimalsolutionmotivatedmoreparticipantstofindthe
optimalsolutionthanaprizethatonlyoneofthemcouldwin.Thesharepriceofeach
optioninMelosoetal.’sstudyalsosignaledparticipantsaboutthegroup’scollectivebelief
initsvalue,whichsupportedalearningstrategysimilartothefrequencybiasseeninour
180
results,andthe“filtering”ofoptionsusedinthesociallearningstrategiestournamentof
Rendelletal.(2010).
7.5.3.Polycentricinnovation
Thereremainlargedifferencesbetweentheaboveexperimentaltreatmentsand
manyreal‐worldintellectualdiscoverycontexts,mostobviouslyinthescaleofdiscoveries,
andtheinvestmentsoftimeandotherresourcesrequired.Butthefactthatmany
predictionsfromthetraditional“reward”and“prospect”theoriesofpatentshavebeen
contradictedorremainunconfirmedintheaboveexperimentalresults,aswellascross‐
nationaleconometricstudies(Park&Ginarte,1997),large‐scalereviewsofdatarelatedto
patentingbehaviorandchangesinpatentlaw(Bessen&Meurer,2008;Boldrin&Levine,
2008),andavarietyofothermeasures(seeChapter6)shouldgiveuspauseincontinuing
toendorsesuchtheories.Theincreasinglywell‐studiedperformanceofalternatemodelsof
innovationbyindividualusersandgroupsengagedinopencollaboration(Baldwin&von
Hippel,2009;Strandburg,2009;Benkler,2006)showusthattherearemanypossibilities
formoreequitable,efficient,andsustainablewaysofcreatingbeneficialinnovationsand
supportingaknowledgecommons.
MazzoleniandNelson(1998)notethatthetheoriesusedtojustifyIP(thoseof
incentivesforcreation,disclosure,commercialization,andefficientexplorationof
innovations)arenotuniversallyapplicableandinsomecircumstancesconflictwitheach
other.Theyalsonotethatthesetheoriesmakeassumptions(implicitlyorexplicitly)about
severalimportantfeaturesofcontextsinwhichinnovationoccurs:(1)“Thenatureand
effectivenessofmeansotherthanpatentstoinduceinventionandrelatedactivities”(a
181
generaleffectivenessquestion)(2)“Whetherthegroupofpotentialinventorsislikelyto
workondiverseandnon‐competingideas,orwhetherthegroupislikelytobefocusedona
singlealternativeorasetofcloselyconnectedones.”(i.e.thestructureofthesearchspace)
(3)“Thedeterrenteffectofthepresenceofpatentsonunauthorizeduseofatechnology
andonthetransactioncostsinvolvedinlicensinganinvention.”(issuesrelatedto
competingfollow‐oninnovationsandanticommonsconcerns)(4)“Whetherthemultiple
stepsintheinvention,development,andcommercializationofanewtechnologytendto
proceedefficientlywithinasingleorganization,orwhetherefficiencyisenhancedif
differentorganizationsareinvolvedatdifferentstagesoftheprocess.”(issuesrelatedto
competitionandcentralization)(5)The“topographyoftechnologicaladvance,”i.e.“the
mannerinwhichinventionsarelinkedtoeachothertemporally,andassystemsinuse”
(alsodealingwithsequentialinnovationsandthestructureoftheproblemspace)
(Mazzoleni&Nelson,1998,p.1033).
Variationsintheseconditionsacrossdifferentinnovationcontextspromptachange
fromasking“Whichtheoryiscorrect?”to“Wheredoeseachtheoryapply?”Itiseasytosee
wherealternatemodessuchasuserinnovationoropencollaborationcouldgainpurchase
underthisframework,bymakingchangestotheassumptionsembodiedintraditionalIP
theories.Ratherthancentralizinginnovationdecisionsinpatent‐holdingindividualsor
firms,thesealternatemodesincreasetheflexibilityofthegovernanceofprovisionand
appropriationinaknowledgecommons,andsubjectthemtonormsdeterminedbythe
relevantcommunitiesofusersandinnovators.Asseenintheanalysesofpatent
performanceandvalueacrossindustries(Bessen&Meurer,2008),conditionsaround
inventionprocessesindifferentfields(e.g.pharmaceuticalsversussoftware)couldsupport
182
avarietyofdifferentpracticesandorganizationalstructuresaroundthediscoveryanduse
ofinnovations,withdifferentrulesgoverningactionssuchascoordinationanddisclosure
ofcontributionstothecommons.Someofthesepracticesarealreadyinuseinsituations
suchasanonprofitpatentpoolforAIDStreatmentsindevelopingcountries(UNITAID,
2010),androyalty‐freelicenserequirementsforworldwidewebtechnologystandards
(WorldWideWebConsortium,2004).Atthesametime,thereisaplaceforhigherlevelsof
organizationandgovernance.Asmentionedearlier,IPrightscanbeusedtoresolveorpre‐
emptlegaldisputesbetweenknowledgeusersorusercommunities(O’Mahony,2003).It
hasalsorecentlybeenarguedthattheWorldIntellectualPropertyOrganization
(historicallyacoalitionforthepromotionofIPrightsandenforcement)shouldbere‐tooled
asamoregeneralinternationaladministrativeforumforconsideringandimplementing
policiesrelatedtoevolvingmodesofinnovation(Strandburg,2009).Thisrelianceonmix
ofvariousformsandscalesofauthoritywithoverlappingjurisdictions(asopposedto
completelycentralizedordecentralizedauthority)hasbeenfoundtobeanimportant
factorinthegovernanceofpublicgoodsandcommon‐poolresources,inwhichithas
becomeknownas“polycentric”governance(Ostrom,Tiebout,&Warren,1961;Andersson
&Ostrom,2008;Ostrom,2008).
Thereareundoubtedlysomeinnovationsthatmayhavebeendelayedornotcreated
atallapartfromsomeformofpatentsystem,buttherearealsoinnovationsthatwere
mademostlywithoutit(suchasmanyopen‐sourcesoftwareprojectsandmuchacademic
research)whichwouldprobablynothavebeencreatedifeachcomponentwassubjectto
patentprotection.Ratherthanarguingfortheabolishmentofthepatentsystem,these
innovationsdemonstratethatapolycentricsystemofknowledgecommonsgovernanceis
183
possible,desirable,andinfactisalreadyuponus.Studiessuchasthosedescribedinthis
dissertationcanaidinunderstandingthecauses,consequences,anddynamicsofdiscovery
undervariousformsofgovernance,andthereforeinimprovingthepracticalprocessesof
discoveryforthebenefitofall.
184
7.A.ProtectionchoicewordinginExperiment4
Protectionchoicedialog:attheendofeachroundexceptthelast,participantsinthe
protectionavailableconditionwhoseteamwasnotalreadyprotectedbyanother
participantwereshownadialogboxwiththefollowingtext(inquotes)andbuttons(in
brackets):
“Doyouwanttoprotectyourteam?Thiswillcostyou4points.”
[ProtectMyTeam] [Don’tProtectMyTeam]
Protectionconfirmation:Iftheparticipantclickedthe[ProtectMyTeam]button,a
dialogboxwasdisplayedwiththefollowingtextuntilthenextroundbegan:
“Yourteamwilbeprotectedfor3rounds.Youhavebeencharged4points.”
Protectiondenied:Iftheparticipantclickedthe[ProtectMyTeam]buttonand
anotherparticipanthadprotectedthesameteamfirstduringthesamepost‐roundchoice
period,adialogboxwasdisplayedwiththefollowingtextuntilthenextroundbegan:
“Sorry,someoneelseprotectedthisteambeforeyou.”
185
Protectiondeclineconfirmation:Iftheparticipantclickedthe[Don’tProtectMy
Team]button,adialogboxwasdisplayedwiththefollowingtextuntilthenextround
began:
“Youchosenottoprotectyourteam.”
Previousprotection:Ifadifferentplayerhadprotectedthesameteamorateam
withinoneicondifferenceinapreviousround,thefollowingmessagewasdisplayed
insteadoftheProtectionchoicedialog:
“Yourteamthisroundcannotbeprotected.Waitingforotherstoprotecttheir
teams…”
186
7.B.Post‐tasksurveyinExperiment4
Thefollowingsurveywasdistributedtoparticipantsfollowingtheendofthelast
gameineachsession.
Front
188
8.GeneralDiscussion
8.1.Conclusionsandcontinuations
Whenoneobservesothersforcluesaboutbeneficialbehaviors,itisimportanttobe
attentivetothemanypossibledifferencesbetweenthecontextstheyencounterandthe
relatedsituationsinwhichoneexpectstofindoneself.Thesecanbeduetochangesinan
environmentovertime,ordifferenceswithinandbetweenenvironmentsduetothingslike
climate,culture,orthecapabilitiesbroughttothepresentfrompreviousexperience.
Similarly,theexperimentalresultswehavesharedinthisdissertationaresubjectto
caveatsaboutthecontextinwhichweobtainedthem,butwebelievetheyoffersome
insightintocommonconflictsandmeritsubstantialfurtherstudy.
Thegoaloftheseexperimentswastoexaminehowgroupsofpeoplefindsolutions
toproblems,byadaptingtheirbehaviortotheirenvironmentsandtoeachotherovertime,
undervaryingcombinationsofinformationandincentives.Situationsthatfitthis
descriptionareubiquitiousinthelivesofhumansandothersocialanimals,butnotyetwell
understood.Theyarefullofcomplicatingandconfoundingfactorsthatmakeitdifficultto
findtheimportantfactorsthatcausetheoutcomeswesee.Intheseexperiments,wepared
awaymanyofthecomplexitiesofreal‐worldsociallearningtoenablerelativelysimple
interactionsinacontrolledproblemenvironment,sothatwecouldobserveparticipants'
behaviorunencumberedbyatleastsomeofthesecomplicatingfactors.Forinstance,we
didnotallowparticipantstospendpointsonimaginarylawyerswithwhichtosueeach
otherforvirtualpatentinfringement(Torrance&Tomlinson,2009),nordidwebuildan
artificialforagingapparatusforthemtopeckat.Wehopedtofindcommonalitiesbetween
189
aspectsofdifferentkindsofexploration,suchasbirdssearchingforfoodandpeople
searchingforgoodideas,byabstractingfrommanyoftheirspecifics.
Thisquestforanidealabstractioncancreatetwokindsofflaws:first,itcanleave
outimportantqualitiesoftheeventsweareattemptingtomodel;second,thespecific
detailsthatarechosentoimplementthegeneralqualitieswewishtostudywillby
necessityintroducedistortionsoftheirown.Toconcludethisdissertation,wewilldiscuss
someoftherelevantaspectsofreal‐worldsociallearningandexplorationthatourtasks
didnottakeintoaccount,aswellassomeoftheartificialcharacteristicsthatmayhave
distortedparticipants'behavior,beforesummarizingthefindingsthatwebelievehave
validitybeyondourspecificparadigmsandshouldbeexaminedfurther.
8.2.Omissionsandcommissions
Weattemptedtoprovideappealing,understandable,andfairlyneutralproblemsfor
participantstosolve,whichdidnotundulyfavorcertaincognitiveassets(e.g.verbalor
mathematicalknowledge)thatcouldhaveintroduceddifferencesinbehaviorunrelatedto
ourdesiredphenomena.Inrealsearchsituations,ofcourse,historyandpreparationmatter
‐‐peoplebringallsortsofpreviousknowledgeandexperiencetobearonnovelsituations,
andtheseaffectoutcomesenormously.Thereisnosuchthingasa"pure"searchtask,
unmooredfromanyspecificskills.Traitsthatarepresumablyimportanttoperformancein
ourtaskssuchaslargeworkingmemoryandstrategicthinkingarebynomeansevenly
distributedamongthepopulation,norisenthusiasmforpuzzles,virtualpets,orfantasy
sportsleagues,elementsofwhichwetriedtoincorporateinordertomakethetask
interesting.
190
Inaddition,muchoftheknowledgeandexperiencethatpeoplebringtonewtasks
concernstheirfellowproblemsolvers,particularlytrustandotherbeliefsaboutothers'
capabilitiesandintentions.Peoplearepoliticalanimalsaswellassocialanimals,andthese
beliefs,andhowtheyarecreated,maintained,andmanipulated,areabsolutelycentralto
sustainedinteractionsingroups.Usingparticipantswho(presumably)knewlittleabout
eachotherbeforehandreducedthepossibilityoftheseimportantphenomenahaving
effectsonthecoordinationofbehavior.Thedevelopmentanduseofsuchinterpersonal
knowledgewasalsopreventedbylimitingparticipants'interactionsstrictlytothepassive
sharingofinformationabouttheirsolutions;theywerenotpermittedothermethodsof
communicationwitheachother(suchasachatinterface),ormanipulationofthe
environment(suchastaggingoficons).Wealsoomittedexplicitidentifyinginformation
aboutpeers,andpreventedimplicitidentificationbyavoidingconsistentpositioningof
eachparticipant'schoicesinthetaskdisplays.Capabilitiessuchasstableevaluationof
groupmembers'behavior,discussionandcollectivedeterminationofnormsare
particularlyknowntoaffectoutcomesinpublicgoodandcommon‐poolresourcedilemmas
(Ostrom,Gardner,&Walker,1994).Relatedly,eachgameinourtasklastedonlyafew
minutes,andthewholesessionwaslessthananhour;inreality,thelong‐term
sustainabilityofanysystem,letalonesomethingascomplexasaknowledgecommons,
cannotbeestimatedinsuchashorttime.
BeyondthespecificdetailsofpatentsystemsthatwementionedomittinginChapter
7,theeconomicvalidityofourtaskislimitedbythefactthattherewasnocostfor
explorationbeyondtheopportunitycostofforegoingotheroptions.Manyrealexploration
tasks(e.g.pharmaceuticaltrials,mineralexploration,orsoftwaredebugging)involve
191
substantialinvestmentsoftimeandotherresourcesbothupfrontandonacontinuing
basis.Therelativelyshortlengthandnarrowbreadthofprotection,amountsforprotection
andusefees,andtheoveralllevelofmonetarypaymentcouldraisesimilarobjections.
Becauseoftheabovedifferences,wemustbecautiousaboutgeneralizingour
resultstoobroadly,andourconclusionsmustbetestedinothersociallearningand
explorationparadigmstoobservetheeffectsoftheseandothercontextualfactors.
8.3.Summaryoffindings
Despitetheseconcerns,webelievethat(a)ourresultsmakeasubstantial
contributiontoknowledgeaboutthecauses,consequences,anddynamicsofsociallearning
andsearch;(b)thattheyhavesubstantialvaliditybeyondthetaskandparticipantsthat
generatedthem;and(c)thattheycanleadtofurtherfertileinvestigationinthelaboratory
andthefield.
Wefoundthatourhumanparticipantsusedseveralsociallearningstrategies
previouslystudiedinotherspecies(Laland,2004),suchascopyingbetter‐performing
peers,copyingsolutionelementsthataremorefrequentamongpeers,andcopyingwhen
uncertainaboutthereturnstoasociallearning.Wealsofoundevidenceforcopying
solutionssimilartoone'sown,whichhasbeenexaminedinstudiesofinnovationdiffusion
asbackwardscompatibility(Rogers,2003),aswellascopyingsolutionelementsthatare
increasingratherthandecreasinginfrequencyamongpeers,asobservedinarecentstudy
ofbaby‐namingdata(Gureckis&Goldstone,2009).Wealsofoundthat,ratherthansimply
fallingbackonalternatesociallearningstrategies(suchasfrequencybias)when
192
performance‐basedimitationwasunavailable,theuseofeachsourceofinformationwas
usedindifferentwaysaccordingtothespecificincentivesandrisksinvolved.
Participantsusedfairlyconservative,incrementaliststrategiestoexplorecomplex
problemspaces,sothatinitialsatisfactorysolutionscouldbeusedasabasisforfurther
developmentviasmallamountsofbothimitationandinnovation.Thesetacticsallowed
participantstoprogressivelyandcollectivelynarrowtheirsearchtowardbetterregionsof
theproblemspacewithouttakingexcessiverisksthatcouldhurtaggregateperformance.
Somesimplemodels(Rogers,1988;Giraldeau,Valantone,&Templeton,2002)
predictthatimitationwillnotimprovetheoverallperformanceofagroup,becauseagents
willsimplyuseittotakeadvantageoftheinformationprovidedbyothers,andavoidthe
costsofexploration.Incontrast,wefoundthatparticipantsoftenusedimitation(in
combinationwithinnovationandretentionofprevioussolutionelements)tocumulatively
improveononeanother'ssolutionsandenhancebothindividualandoverallgroup
performance.
Whenweintroducedcontextualfactorsthatcausedimitationtobeeitherdisrupted
(hidingperformanceinformationaboutpeers'solutions)ordiscouraged(chargingasmall
feeforusingasolutionthatapeerhadpreviouslydiscoveredand"protected"),theuseof
innovationincreased,butimprovedperformancedidnotresult.Intheformercase,
participantscouldnotuseimitationtopropagategoodsolutions,andtheincreased
explorationwasessentiallywastedbecausethesearchremainedunfocusedandinefficient.
Inthelattercase,theadditionalexplorationactuallyresultedinanincreaseddiscoveryof
improvedsolutions,butalsothediscoveryofworsesolutionsintheprocess,which
193
canceledoutthegains;thecostforimitationmeantthatparticipantsimitatedlessoften
andwerethuslesslikelytobuildononeanother'ssolutions.
8.4.Applicationsandfuturedirections
Inthecontextofscientificandculturalprogress,thereareparallelstothemodelsof
Rogers(1988)andGiraldeau,Valone,andTempleton(2002)mentionedabove,which
predictthatinnovationwillbeinshortsupplyandinefficientlyusedunlessthosewho
innovatecanexcludeothersfromimitatingthemwithoutpermission(e.g.Kitch,1977).
Ourworkshowsthattheabovemodelsareperhapsoverlysimplisticwhenitcomes
tohumansociallearningandexploration.Individualshaveacapacitytouseothers'
innovationsnotjusttoobtainastaticbenefitingoodperformance,buttocreateadynamic
benefitformanybyproducingsequentialimprovements;soadisincentivetoimitate(in
theformofexclusionrightsforinnovators)mayresultinlessproductiveinnovation,not
more.Theworkofothersinthisarea(e.g.Torrance&Tomlinson,2009;Bessen&Meurer,
2008;seeChapters6&7)confirmsthesefindingsandshowsthatthereareother
theoreticalandpracticaldisadvantagestograntingexclusiverightstoknowledge.
Thereisacountervailingmovementtotreatknowledgeasaresourcethatbenefits
fromcommunitymanagementatmultiplescales,ratherthanstrictprivateownershipor
centralizedstatecontrol(Ostrom&Hess,2007).Therearenaturallydebatesbetween
proponentsofthesevariousmethodsofmanagement,butitisclearthatthefutureof
knowledgeisinextricablyboundupinlearninghowtoequitably,efficiently,and
sustainablygoverntheproductsofitspastandpresent.
194
Tomasello(1994)contendsthatthecapabilitiesofhumansforselectiveand
cumulativesociallearningconstitutea“ratcheteffect”thatallowsculturetodevelopstably
acrossgenerations,andthatthiseffectmaybeuniquetohumans.Thoughtheexperiments
inthisdissertationpresentagreatlysimplifiedenvironmentforsuchsociallearning,they
confirmedandextendedseveralprevioustheoreticalandempiricalresultsinthefield.
Theconcernsaboutourexperimentsnotedabove(aswellasrelatedworkdiscussed
inpreviouschapters)suggestpotentiallyusefulvariationsofourcoretasks.Forinstance:
useofvariouspracticalknowledgedomainsforsearchandvariouslevelsofexpertise
amongparticipants;useofnoisy(ratherthansimplypresentorabsent)scorefeedback;
variationsinthestructureofthesocialnetworkthatconnectsparticipants;explicitcosts
forinnovationandmoresubstantialincentivesforperformance;andcommunication
amongparticipants,andthereforeopportunitiesforcoordinationoftheirsearchbehavior
andmorestakeholder‐drivengovernanceoftheuseoftheresultingimprovements.
Theresoundingrefrainofthelargebodyofresearchoncommonsdilemmasand
othercollectiveactionproblemsisthatmanycanbesolvedthroughcarefulinvestigation
andmodificationoftherelatedenvironmental,cultural,andinstitutionalvariables,but
thereis"nopanacea,"nouniversalsolutiononwhichwecanrely(Ostrom,2010).
Continuedstudyisnecessary,ofthestructureanddynamicsofhumancollaborative
exploration,thenatureofincentivesandinnovativeeffort,andtheirinteractionas
manifestedinthedynamicsofcreativityinoureconomyandculture.Wepresentourwork
inthehopethatotherswillalsofinditusefulinilluminatingoursharedproblems.
195
References
Abrahamson,E.,&Rosenkopf,L.(1997).Socialnetworkeffectsontheextentofinnovation
diffusion:Acomputersimulation.OrganizationScience,8(3),289‐309.
Acheson,J.M.,Wilson,J.A.,&Steneck,R.S.(1998).Managingchaoticfisheries.InF.Berkes
&C.Folke(Eds.),Linkingsocialandecologicalsystems:Managementpracticesand
socialmechanismsforbuildingresilience(pp.390–413).Cambridge,MA:Cambridge
UniversityPress.
Allen,R.C.(1983).Collectiveinvention.JournalofEconomicBehavior&Organization,4(1),
1‐24.
Anderson,L.,&Holt,C.(1997).Informationcascadesinthelaboratory.AmericanEconomic
Review,87(5),847‐862.
Andersson,K.P.,&Ostrom,E.(2008).Analyzingdecentralizedresourceregimesfroma
polycentricperspective.PolicySciences,41(1),71‐93.
Asch,S.E.(1951).Effectsofgrouppressureuponthemodificationanddistortionof
judgments.InH.Guetzkow(Ed.)Groups,Leadership,andMen.Pittsburgh:Carnegie
Press.
Baldwin,C.Y.,&VonHippel,E.A.(2009).Modelingaparadigmshift:Fromproducer
innovationtouserandopencollaborativeinnovation.HarvardBusinessSchool
FinanceWorkingPaperNo.10‐038.AvailableatSSRN:
http://ssrn.com/abstract=1502864
Bandura,A.(1965).Vicariousprocesses:Acaseofno‐triallearning.InL.Berkowitz(Ed.),
AdvancesinExperimentalSocialPsychology,Vol.II.NewYork:AcademicPress.
196
Banerjee,A.(1992).Asimplemodelofherdbehavior.QuarterlyJournalofEconomics,
107(3),797‐817.
Baron,R.S.,Vandello,J.A.,&Brunsman,B.(1996).Theforgottenvariableinconformity
research:Theimpactoftaskperformanceonsocialinfluence.JournalofPersonality
&SocialPsychology,71,915‐927.
Barzel,Y.(1968).Optimaltimingofinnovations.TheReviewofEconomicsandStatistics,
50(3),348‐355.
Baumol,W.J.(2002).Thefreemarketinnovationmachine:Analyzingthegrowthmiracleof
capitalism.Princeton,NJ:PrincetonUniversityPress.
Beck,M.,&Galef,B.G.,Jr.(1989).Socialinfluencesontheselectionofaprotein‐sufficient
dietbyNorwayrats(Rattusnorvegicus).JournalofComparativePsychology,103(2),
132‐139.
Benkler,Y.(2002).Coase'spenguin,or,Linuxandthenatureofthefirm.YaleLawJournal,
112,369‐446.
Bessen,J.,&Hunt,R.M.(2007).Anempiricallookatsoftwarepatents.JournalofEconomics
&ManagementStrategy,16(1),157–189.
Bessen,J.,&Meurer,M.J.(2005).Thepatentlitigationexplosion.BostonUniversitySchool
ofLawWorkingPaperNo.05‐18.Retrievedfrom
http://law.bepress.com/cgi/viewcontent.cgi?article=1532
Bessen,J.,&Meurer,M.J.(2008).Patentfailure:Howjudges,bureaucrats,andlawyersput
innovatorsatrisk.Princeton,NJ:PrincetonUniversityPress.
Boldrin,M.,&Levine,D.K.(2008).Againstintellectualmonopoly.Cambridge:Cambridge
UniversityPress.
197
Bikhchandani,S.,Hirshleifer,D.,&Welch,I.(1992).Atheoryoffads,fashion,custom,and
culturalchangeasinformationalcascades.JournalofPoliticalEconomy,100,992‐
1026.
Bond,R.(2005).Groupsizeandconformity.GroupProcesses&IntergroupRelations,8(4),
331–354.
Box,H.O.(1984).Primatebehaviorandsocialecology.London:Chapman&Hall.
Boyd,R.,&Richerson,P.J.(1985).Cultureandtheevolutionaryprocess.Chicago:University
ofChicagoPress.
Boyd,R.,&Richerson,P.J.(1995).Whydoescultureincreasehumanadaptability?Ethology
andSociobiology,16,125‐143.
Boyd,R.,&Richerson,P.J.(2005).Theoriginandevolutionofcultures.NewYork:Oxford
UniversityPress.
Boyle,J.(2007)Mertonianismunbound?Imaginingfree,decentralizedaccesstomost
culturalandscientificmaterial.InC.Hess&E.Ostrom(Eds.).Understanding
knowledgeasacommons(pp.123‐144).Cambridge,MA:MITPress.
Byrne,R.W.,&Russon,A.E.(1998).Learningbyimitation:Ahierarchicalapproach.
Behavioral&BrainSciences,21,667‐684.
Currie,N.(2010,March11).Derstand,understand,un‐understand.Retrievedfrom
http://imomus.com/index032010b.html
Davis,J.M.(1973).Imitation:Areviewandcritique.InP.P.G.Bateson&P.H.Klopfer
(Eds.),Perspectivesinethology.NewYork:Plenum.
Dawson,E.W.(1974).Adeliepenguinsandleopardseals:Illustrationsofpredation–
history,legendandfact.Notornis,21(1),36‐69.
198
Deci,E.L.(1971).Effectsofexternallymediatedrewardsonintrinsicmotivation.Journalof
PersonalityandSocialPsychology,18(1),105‐115.
Deci,E.L.,Koestner,R.&Ryan,R.M.(1999).Ameta‐analyticreviewofexperiments
examiningtheeffectsofextrinsicrewardsonintrinsicmotivation.Psychological
Bulletin,125(3),627‐668.
Deutsch,M.,&Gerard,H.B.(1955).Astudyofnormativeandinformationalsocial
influencesuponindividualjudgment.JournalofAbnormalandSocialPsychology,51,
629‐636.
DiMaggio,P.J.,&Powell,W.W.(1983).Theironcagerevisited:Institutionalisomorphism
andcollectiverationalityinorganizationalfields.AmericanSociologicalReview,48,
147–160.
Dunbar,K.(1999).Scientificcreativity:Itsrecognitionanddevelopment.InThe
encyclopediaofcreativity(Vol.1,pp.1379‐1384).NewYork:AcademicPress.
Eguíluz,V.M.,&Zimmerman,M.G.(2000).Transmissionofinformationandherdbehavior:
Anapplicationtofinancialmarkets.PhysicalReviewLetters,85,5659–5662.
Eisenberger,N.I.,&Lieberman,M.D.(2004).Whyrejectionhurts:Acommonneuralalarm
systemforphysicalandsocialpain.TrendsinCognitiveSciences,8(7),294‐300.
Falvey,R.,Foster,N.,&Greenaway,D.(2006).Intellectualpropertyrightsandeconomic
growth.ReviewofDevelopmentEconomics,10(4),700–719.
Forsyth,D.R.(2009).Groupdynamics(5thed.).Belmont,CA:Wadsworth.
Freeman,C.(1968).Chemicalprocessplant:Innovationandtheworldmarket.National
InstituteEconomicReview45(August):29‐57.
199
Frey,B.S.,&Jegen,R.(2001).Motivationcrowdingtheory.JournalofEconomicSurveys,
15(5),589‐611.
Friend,R.,Rafferty,Y.,&Bramel,D.(1990).ApuzzlingmisinterpretationoftheAsch
'conformity'study.EuropeanJournalofSocialPsychology,20,29‐44.
Frischmann,B.M.,&Lemley,M.A.(2007).Spillovers.ColumbiaLawReview,107,257‐301.
Galef,B.G.,Jr.(1976).Socialtransmissionofacquiredbehavior:Adiscussionoftradition
andsociallearninginvertebrates.InJ.S.Rosenblatt,R.A.Hinde,E.ShawandC.Beer
(Eds.),AdvancesintheStudyofBehavior(Vol.6,pp.77‐100).NewYork:Academic
Press.
Galef,B.G.,Jr.(1988).Imitationinanimals:History,definition,andinterpretationofdata
fromthepsychologicallaboratory.InT.Zentall&B.G.Galef,Jr.(Eds.),Social
learning:Psychologicalandbiologicalperspectives(3‐28).Hillsdale,NJ:Lawrence
Erlbaum.
Galef,B.G.Jr.,&Giraldeau,L.A.(2001).Socialinfluencesonforaginginvertebrates:Causal
mechanismsandadaptivefunctions.AnimalBehaviour,61(1),3‐15.
Garcia,R.,&Calantone,R.(2002).Acriticallookattechnologicalinnovationtypologyand
innovativenessterminology:Aliteraturereview.JournalofProductInnovation
Management,19,110‐132.
Gault,F.,&vonHippel,E.(2009).Theprevalenceofuserinnovationandfreeinnovation
transfers.MITSloanResearchPaperNo.4722‐09.AvailableatSSRN:
http://ssrn.com/abstract=1337232
Ghosh,S.(2003).Deprivatizingcopyright.CaseWesternReserveLawReview,54,387‐501.
200
Gibbons,D.E.(2004).Networkstructureandinnovationambiguityeffectsondiffusionin
dynamicorganizationalfields.AcademyofManagementJournal,47,938–951.
Gigerenzer,G.,&Goldstein,D.G.,(1996).Reasoningthefastandfrugalway:Modelsof
boundedrationality.PsychologicalReview,103(4),650‐669.
Gilbert,R.,&Shapiro,C.(1990).Optimalpatentlengthandbreadth.TheRandJournalof
Economics,21(1),106‐112.
Ginarte,J.C.,&Park,W.G.(1997).Determinantsofpatentrights:Across‐nationalstudy.
ResearchPolicy26,283‐301.
Giraldeau,L.,Valone,T.J.,&Templeton,J.J.(2002).Potentialdisadvantagesofusing
sociallyacquiredinformation.PhilosophicalTransactionsoftheRoyalSocietyof
LondonB,357(1427),1559‐66.
Goldberg,D.E.(1989).Geneticalgorithmsinsearch,optimization,andmachinelearning.
Reading,MA:Addison‐Wesley.
Granovetter,M.(1978).Thresholdmodelsofcollectivebehavior.TheAmericanJournalof
Sociology,83(6),1420‐1443.
Gureckis,T.M.,&Goldstone,R.L.(2009).Howyounamedyourchild:Understandingthe
relationshipbetweenindividualdecisionmakingandcollectiveoutcomes.Topicsin
CognitiveScience,1(4),651‐674.
Hall,B.H.(2007).Patentsandpatentpolicy.OxfordReviewofEconomicPolicy,23(4),568‐
587.
Hall,K.R.L.(1963).Observationallearninginmonkeysandapes.BritishJournalof
Psychology,54,201‐226.
Hardin,G.(1968).Thetragedyofthecommons.Science,162,1243‐1248.
201
Hare,A.P.(1976).Handbookofsmallgroupresearch(2nded.).NewYork:FreePress.
Harhoff,D.,Henkel,J.,&vonHippel,E.(2003).Profitingfromvoluntaryinformation
spillovers:Howusersbenefitbyfreelyrevealingtheirinnovations.ResearchPolicy,
32(10),1753‐1769.
Harhoff,D.,Scherer,F.M.,&Vopel,K.(2003).Exploringthetailofpatentedinventionvalue
distributions.InO.Grandstrand(Ed.),Economics,Law,andIntellectualProperty(pp.
279‐309).TheHague:Kluwer.
Heise,G.A.,&Miller,G.A.(1951).Problemsolvingbysmallgroupsusingvarious
communicationnets.JournalofAbnormalPsychology,46(3),327‐335.
Heller,M.A.,&Eisenberg,R.S.(1998).Canpatentsdeterinnovation?Theanticommonsin
biomedicalresearch.Science,280(5364),698‐701.
Hess,C.&Ostrom,E.(Eds.).(2007).Understandingknowledgeasacommons.Cambridge,
MA:MITPress.
Heyes,C.(1993).Imitation,culture,andcognition.AnimalBehaviour,46,999‐1010.
Heyman,J.,&Ariely,D.(2004).Effortforpayment.Ataleoftwomarkets.Psychological
Science,15(11),787‐93.
Hodges,B.H.(2004).Aschandthebalanceofvalues.BehavioralandBrainSciences,27,
343‐344.
Höglund,J.,Alatalo,R.V.,Gibson,R.M.,&Lundberg,A.(1995).Mate‐choicecopyinginblack
grouse.AnimalBehaviour,49(6),1627‐1633.
Holland,J.H.(1999).Emergence:Fromchaostoorder.NewYork:BasicBooks.
Hurley,S.&Chater,N.(Eds.)(2005).Perspectivesonimitation:Fromneurosciencetosocial
science.Cambridge,MA:MITPress.
202
Kameda,T.,&Nakanishi,D.(2002).Cost–benefitanalysisofsocial/culturallearningina
nonstationaryuncertainenvironment:Anevolutionarysimulationandan
experimentwithhumansubjects.EvolutionandHumanBehavior,23(5),373‐393.
Kameda,T.,&Nakanishi,D.(2003).Doessocial/culturallearningincreasehuman
adaptability?Rogers'questionrevisited.EvolutionandHumanBehavior,24(4),242‐
260.
Kameda,T.,&Tindale,R.(2006).Groupsasadaptivedevices:Humandocilityandgroup
aggregationmechanismsinevolutionarycontext.InM.Schaller,J.Simpson,&D.T.
Kenrick(Eds.),Evolutionandsocialpsychology(pp.317‐342).NewYork:Psychology
Press.
Katz,M.L.,&Shapiro,C.(1985),Networkexternalities,competitionandcompatibility.
AmericanEconomicReview,75,424‐440.
Keefer,P.,&Knack,S.(1995).Institutionsandeconomicperformance:Cross‐countrytests
usingalternativeinstitutionalmeasures.EconomicsandPolitics,7(3),207‐227.
Keefer,P.,&Knack,S.(1997).Whydon’tpoorcountriescatchup?Across‐nationaltestof
aninstitutionalexplanation.EconomicInquiry,35,590‐602.
Kitch,E.W.(1977).Thenatureandfunctionofthepatentsystem.JournalofLawand
Economics,20,265‐290.
Kounios,J.,&Beeman,M.(2009).TheAha!moment:Thecognitiveneuroscienceofinsight.
CurrentDirectionsinPsychologicalScience,18(4),210‐216.
Krebs,J.R.,&Inman,A.J.(1992).Learningandforaging:Individuals,groups,and
populations.TheAmericanNaturalist,140(Supplement:Behavioralmechanismsin
evolutionaryecology),S63‐S84.
203
Krueger,J.I.,&Funder,D.C.(2004).Towardsabalancedsocialpsychology:Causes,
consequences,andcuresfortheproblem‐seekingapproachtosocialbehaviorand
cognition.BehavioralandBrainSciences,27,313‐376.
Lachlan,R.F.,Crooks,L.,Laland,K.N.(1998).Whofollowswhom?Shoalingpreferences
andsociallearningofforaginginformationinguppies.AnimalBehaviour,56(1),181‐
190.
Lakhani,K.R.,&Wolf,R.G.(2003).Whyhackersdowhattheydo:Understanding
motivationeffortinfree/opensourcesoftwareprojects.MITSloanWorkingPaper
No.4425‐03.doi:10.2139/ssrn.443040
Laland,K.N.(2004).Sociallearningstrategies.Learning&Behavior,32(1),4‐14.
Latane,B.,Williams,K.,&Harkins,S.(1979).Manyhandsmakelightthework:Thecauses
andconsequencesofsocialloafing.JournalofPersonalityandSocialPsychology,
37(6),822‐832.
Lazer,D.&Friedman,A.(2007).Thenetworkstructureofexplorationandexploitation.
AdministrativeScienceQuarterly,52,667‐694.
Lerner,J.,Tirole,J.,&Pathak,P.(2006).Thedynamicsofopensourcecontributors.
AmericanEconomicsReviewPapersandProceedings,96,114‐118.
Lougee,W.P.(2007)Scholarlycommunicationandlibrariesunbound:Theopportunityof
thecommons.InC.Hess&E.Ostrom(Eds.).Understandingknowledgeasacommons
(pp.311‐332).Cambridge,MA:MITPress.
MacDonald,G.,&Leary,M.R.(2005).Whydoessocialexclusionhurt?Therelationship
betweensocialandphysicalpain.PsychologicalBulletin,131(2),202–223.
204
Mason,W.A.,Jones,A.,&Goldstone,R.L.(2008).Propagationofinnovationsinnetworked
groups.JournalofExperimentalPsychology:General,137(3),422‐33.
MazervStein,347US201(1954).
Mazzoleni,R.,&Nelson,R.(1998).Economictheoriesaboutthebenefitsandcostsof
patents.JournalofEconomicIssues,32(4),1031‐1052.
Meloso,D.,Copic,J.,&Bossaerts,P.(2009).Promotingintellectualdiscovery:Patents
versusmarkets.Science,323(5919),1335‐1339.
Merges,R.P.,&Nelson,R.R.(1990).Onthecomplexeconomicsofpatentscope.Columbia
LawReview,90,839‐916.
Miller,N.E.,&Dollard,J.(1941).Sociallearningandimitation.NewYork:McGraw‐Hill.
Mokyr,J.(1999).TheBritishindustrialrevolution:Aneconomicperspective.Boulder:
Westview.
Moore,K.A.(2005).Markmaneightyearslater:Isclaimconstructionmorepredictable?
Lewis&ClarkLawReview,9,231‐247.
Morgan,C.L.(1900).Animalbehavior.London:Arnold.
Moser,P.(2002).Howdopatentlawsinfluenceinnovation?Evidencefrom19th‐century
worldfairs.NationalBureauofEconomicResearch,WorkingPaperNo.9909.
O'Mahony,S.(2003).Guardingthecommons:Howcommunitymanagedsoftwareprojects
protecttheirwork.ResearchPolicy,32(7),1179‐1198.
Olson,M.(1965).Thelogicofcollectiveaction:Publicgoodsandthetheoryofgroups.
Cambridge,MA:HarvardUniversityPress.
Ostrom,E.(1990).Governingthecommons:Theevolutionofinstitutionsforcollectiveaction.
NewYork:CambridgeUniversityPress.
205
Ostrom,E.(1998).Abehavioralapproachtotherationalchoicetheoryofcollectiveaction.
AmericanPoliticalScienceReview,92(1),1‐22.
Ostrom,E.(1999).Copingwithtragediesofthecommons.AnnualReviewofPolitical
Science,2,493‐535.
Ostrom,E.(2008).Polycentricsystemsasoneapproachforsolvingcollective‐action
problems.WorkingpaperW08‐6.AvailableatSSRN:
http://ssrn.com/abstract=1304697
Ostrom,E.(2010).Beyondmarketsandstates:Polycentricgovernanceofcomplex
economicsystems.AmericanEconomicReview,100(3),641‐672.
Ostrom,E.,Gardner,R.,&Walker,J.(1994).Rules,games,&commonpoolresources.Ann
Arbor:UniversityofMichiganPress.
Ostrom,V.,Tiebout,C.M.,&Warren,R.(1961).Theorganizationofgovernmentin
metropolitanareas:Atheoreticalinquiry.AmericanPoliticalScienceReview,55(4),
831‐842.
Paletz,S.B.,&Schunn,C.D.(2010).Asocial‐cognitiveframeworkofmultidisciplinaryteam
innovation.TopicsinCognitiveScience,2(1),73‐95.
Park,W.G.,&Ginarte,J.C.(1997).Intellectualpropertyrightsandeconomicgrowth.
ContemporaryEconomicPolicy,15,51‐61.
Poundstone,W.(1992).Prisoner'sdilemma.NewYork:Doubleday.
Rendell,L.,Boyd,R.,Cownden,D.,Enquist,M.,Eriksson,K.,Feldman,M.W.,Fogarty,L.,
Ghirlanda,S.,Lillicrap,T.,&Laland,K.N.(2010).Whycopyothers?Insightsfromthe
sociallearningstrategiestournament.Science,328(5975),208‐13.
Rogers,A.R.(1988).Doesbiologyconstrainculture?AmericanAnthropologist,90,819‐831.
206
Rogers,E.M.(2003).DiffusionofInnovations(5thed.).NewYork:FreePress.
Romanes,G.J.(1884).Mentalevolutioninanimals.London:KeganPaul.
Sandberg,A.(2001).InstitutionalchallengesforcommonpropertyresourcesintheNordic
countries.Stockholm:Nordregio.
Schlag,K.H.(1998).Whyimitate,andifso,how?Aboundedlyrationalapproachtomulti‐
armedbandits.JournalofEconomicTheory,78(1),130‐156.
Scotchmer,S.(1991).Standingontheshouldersofgiants.JournalofEconomicPerspectives,
5(1),29‐41.
Shah,S.(2000).Sourcesandpatternsofinnovationinaconsumerproductsfield:
Innovationsinsportingequipment.MITSloanWorkingPaperNo.4105.Retrieved
fromhttp://opensource.mit.edu/papers/shahsportspaper.pdf
Sherry,D.F.,&Galef,B.G.,Jr.(1984).Culturaltransmissionwithoutimitation:Milkbottle
openingbybirds.AnimalBehaviour,32(3),937‐938.
Sherry,D.F.,&Galef,B.G.,Jr.(1990).Sociallearningwithoutimitation:Moreaboutmilk
bottleopeningbybirds.AnimalBehaviour,40(5),987‐989.
Siy,R.Y.(1982).Communityresourcemanagement:LessonsfromtheZanjera.QuezonCity,
Philippines:UniversityofthePhilippinesPress.
Sneath,D.(1998).StatepolicyandpasturedegradationinInnerAsia.Science,281,1147‐
1148.
Spence,K.W.(1937).Experimentalstudiesoflearningandhighermentalprocessesin
infra‐humanprimates.AnimalBehavior,34,806‐850.
Sternberg,R.J.(Ed.).(1999).HandbookofCreativity.Cambridge:CambridgeUniversity
Press.
207
Strandburg,K.J.(2009).Evolvinginnovationparadigmsandtheglobalintellectual
propertyregime.ConnecticutLawReview4,861‐920.
Svejnar,J.(2002).Transitioneconomies:Performanceandchallenges.TheJournalof
EconomicPerspectives,16(1),3‐28.
Sweller,J.(1988)Cognitiveloadduringproblemsolving.CognitiveScience,12,257–285.
Tarde,G.(1969).TheLawsofImitation.(ElsieClewsParsons,Trans.).Chicago:Universityof
ChicagoPress.(Originalworkpublished1903.)
Templeton,J.J.,&Giraldeau,L.‐A.(1996).Vicarioussampling:Theuseofpersonaland
publicinformationbystarlingsforaginginasimplepatchyenvironment.Behavioral
EcologyandSociobiology,38,105‐113.
Theiner,G.(2009).Makingsenseofgroupcognition:Thecuriouscaseoftransactive
memorysystems.InW.Christensen,E.Schier,andJ.Sutton(Eds.),ASCS09:
Proceedingsofthe9thConferenceoftheAustralasianSocietyforCognitiveScience
(pp.334‐342).Sydney:MacquarieCentreforCognitiveScience.
Thelen,M.H.,Dollinger,S.J.,&Kirkland,K.D.(1979).Imitationandresponseuncertainty.
JournalofGeneticPsychology,135,139‐152.
Thorndike,E.L.(1911).Animalintelligence.NewYork:Macmillan.
Thorpe,W.H.(1963).Learningandinstinctinanimals(2nded.).London:Methuen.
Titmuss,R.M.(1970).Thegiftrelationship.London:AllenandUnwin.
Torrance,A.,&Tomlinson,B.(2009).Patentsandtheregressofusefularts.Columbia
ScienceandTechnologyLawReview,10,131‐168.
208
UNITAID(2010).TheMedicinesPatentPoolinitiative.Retrievedfrom
http://www.unitaid.eu/images/projects/PatentPool/pp_factsheet_april2010_web.p
df
VonHippel,E.(1976).Thedominantroleofusersinthescientificinstrumentinnovation
process.ResearchPolicy5(3),212‐39.
VonHippel,E.(1977).Thedominantroleoftheuserinsemiconductorandelectronic
subassemblyprocessinnovation.IEEETransactionsonEngineeringManagement
EM24(2),60‐71.
Wegner,D.M.(1986).Transactivememory:Acontemporaryanalysisofthegroupmind.In
B.Mullen,G.R.Goethals(Eds.),Theoriesofgroupbehavior.NewYork:Springer.
Wilkinson,G.(1992).Informationtransferateveningbatcolonies.AnimalBehaviour,44,
501‐518.
WorldWideWebConsortium(2004).W3Cpatentpolicy.Retrievedfrom
http://www.w3.org/Consortium/Patent‐Policy‐20040205/
Zajonc,R.B.(1965).Socialfacilitation.Science,149,269‐274.