Series Editor BRIAN H. ROSS Beckman Institute and Department of Psychology University of Illinois, Urbana, Illinois AcademicPressisanimprintofElsevier 225WymanStreet,Waltham,MA02451,USA 525BStreet,Suite1800,SanDiego,CA92101-4495,USA Radarweg29,POBox211,1000AEAmsterdam,TheNetherlands TheBoulevard,LangfordLane,Kidlington,Oxford,OX51GB,UK 32 JamestownRoad,London,NW17BY,UK Copyright©2014,ElsevierInc.Allrightsreserved. Nopartofthispublicationmaybereproduced,storedinaretrievalsystemor transmittedinanyformorbyanymeanselectronic,mechanical,photocopying, recordingorotherwisewithoutthepriorwrittenpermissionofthepublisher PermissionsmaybesoughtdirectlyfromElsevier’sScience&TechnologyRights DepartmentinOxford,UK:phone(+44)(0)1865843830;fax(+44)(0)1865 853333;email:[email protected] requestonlinebyvisitingtheElsevierwebsiteathttp://elsevier.com/locate/ permissions,andselectingObtainingpermissiontouseElseviermaterial Notice Noresponsibilityisassumedbythepublisherforanyinjuryand/ordamageto personsorpropertyasamatterofproductsliability,negligenceorotherwise,or fromanyuseoroperationofanymethods,products,instructionsorideascontained inthematerialherein.Becauseofrapidadvancesinthemedicalsciences,in particular,independentverificationofdiagnosesanddrugdosagesshouldbemade ISBN:978-0-12-800283-4 ISSN:0079-7421 ForinformationonallAcademicPresspublications visitourwebsiteatstore.elsevier.com PrintedandboundinUSA 14 15 16 10 9 8 7 6 5 4 3 2 1 CONTRIBUTORS JeffreyAnnis DepartmentofPsychology,UniversityofSouthFlorida,Tampa,Florida,USA PaulAtchley DepartmentofPsychology,UniversityofKansas,Lawrence,Kansas,USA WilliamR.Aue DepartmentofPsychologicalSciences,PurdueUniversity,WestLafayette,Indiana, andDepartmentofPsychology,SyracuseUniversity,Syracuse,NewYork,USA ColinBla¨ttler ResearchCenteroftheFrenchAirForce(CReA),Salon-de-Provence,France GlenE.Bodner DepartmentofPsychology,UniversityofCalgary,Calgary,Alberta,Canada AmyH.Criss DepartmentofPsychology,SyracuseUniversity,Syracuse,NewYork,USA Andre´ Didierjean UniversityofFranche-Comte´ &InstitutUniversitairedeFrance,Besanc¸on,France VincentFerrari ResearchCenteroftheFrenchAirForce(CReA),Salon-de-Provence,France UlrikeHahn DepartmentofPsychologicalSciences,Birkbeck,UniversityofLondon,London, UnitedKingdom AdamJ.L.Harris DepartmentofCognitive,Perceptual&BrainSciences,UniversityCollegeLondon, London,UnitedKingdom GretaJames DepartmentofPsychology,UniversityofWaterloo,Waterloo,Ontario,Canada CharlesW.Kalish DepartmentofEducationalPsychology,UniversityofWisconsin-Madison,Madison, Wisconsin,USA JeffreyD.Karpicke DepartmentofPsychologicalSciences,PurdueUniversity,WestLafayette,Indiana,USA DerekJ.Koehler DepartmentofPsychology,UniversityofWaterloo,Waterloo,Ontario,Canada SeanLane DepartmentofPsychology,LouisianaStateUniversity,BatonRouge,Louisiana,USA ix x Contributors MelissaLehman DepartmentofPsychologicalSciences,PurdueUniversity,WestLafayette,Indiana,USA KennethJ.Malmberg DepartmentofPsychology,UniversityofSouthFlorida,Tampa,Florida,USA MichaelE.J.Masson DepartmentofPsychology,UniversityofVictoria,Victoria,BritishColumbia,Canada RichardM.Shiffrin DepartmentofBrainandPsychologicalSciences,IndianaUniversity,Bloomington,Indiana, USA JordanT.Thevenow-Harrison DepartmentofEducationalPsychology,UniversityofWisconsin-Madison,Madison, Wisconsin,USA CHAPTER ONE Descriptive and Inferential Problems of Induction: Toward a Common Framework Charles W. Kalish1, Jordan T. Thevenow-Harrison DepartmentofEducationalPsychology,UniversityofWisconsin-Madison,Madison,Wisconsin,USA 1Correspondingauthor:e-mailaddress:[email protected] Contents 1. Introduction 2 2. Theory-BasedandSimilarity-BasedInductiveInference 3 3. InductionasStatisticalInference:DescriptiveandInferentialProblems 6 4. InductiveandTransductiveInference:SampleandPopulationStatistics 9 5. UsingTransductiveInference 12 6. Summary:TransductiveandEvidentialTheoriesofInference 16 7. DistinguishingTransductiveandEvidentialInferences 17 7.1 PeopleandStatistics 20 8. DevelopingSolutionstoDescriptiveProblems 21 8.1 CorrelationsandAssociations 22 8.2 ComponentialAnalysis 22 8.3 TransitionProbabilities 23 8.4 AbsolutetoRelationalStatistics 24 8.5 GlobaltoSpecificRelations 25 8.6 SimpletoComplex 26 8.7 SummaryofSolutionstoDescriptiveProblems 27 9. SolutionstoInferentialProblems 27 9.1 TransductiveInference 28 9.2 BayesianInference 28 9.3 BetweenTransductiveandEvidentialInference 29 9.4 CommunicativeBias 30 9.5 IntentionalVersusIncidentalLearning 31 9.6 SummaryofSolutionstoInferentialProblems 31 10. SummaryandConclusions 32 References 34 Abstract Therearemanyaccountsofhowhumansmakeinductiveinferences.Twobroadclasses ofaccountsarecharacterizedas“theorybased”or“similaritybased.”Thisdistinctionhas PsychologyofLearningandMotivation,Volume61 #2014ElsevierInc. 1 ISSN0079-7421 Allrightsreserved. http://dx.doi.org/10.1016/B978-0-12-800283-4.00001-0 2 CharlesW.KalishandJordanT.Thevenow-Harrison organizedasubstantialamountofempiricalworkinthefield,buttheexactdimensions ofcontrastbetweentheaccountsarenotalwaysclear.Recently,bothaccountshave usedconceptsfromformalstatisticsandtheoriesofstatisticallearningtocharacterize humaninductiveinference.Weextendtheselinkstoprovideaunifiedperspectiveon inductionbasedontherelationbetweendescriptiveandinferentialstatistics.Mostwork inPsychologyhasfocusedondescriptiveproblems:Whichpatternsdopeoplenoticeor representinexperience?Wesuggestthatitissolutionstotheinferentialproblemof generalizingorapplyingthosepatternsthatrevealsthemorefundamentaldistinction between accounts of human induction. Specifically, similarity-based accounts imply thatpeoplemaketransductiveinferences,whiletheory-basedaccountsimplythatpeo- plemakeevidentialinferences.Incharacterizingclaimsaboutdescriptiveandinferential components of induction, we highlight points of agreement and disagreement betweenalternativeaccounts.Adoptingthecommonframeworkofstatisticalinference alsomotivatesasetofempiricalhypothesesaboutinductiveinferenceanditsdevelop- ment across age and experience. The common perspective of statistical inference reframes debates between theory-based and similarity-based accounts: These are notconflictingtheoreticalperspectives,butratherdifferentpredictionsaboutempirical results. 1. INTRODUCTION Inductionisafundamentalcognitiveprocess.Broadlyconstrued,any predictionorexpectationaboutempiricalphenomenarepresentsaninduc- tive inference. Within Psychology, learning, categorization, probability judgments,anddecision-makingareallcentralformsofinductiveinference. Other psychological processes may be treated as involving induction (e.g., perception,languagecomprehension).Therearelikelymanydifferentpsy- chologicalmechanismsinvolvedinmakinginductiveinferences,manyways people make predictions and form expectations. This chapter focuses on a paradigmcase:Learningfromexamples.Basedonexperiencewithalimited setofexamples,peoplegeneralizetonewexamples.Notallinductiveinfer- encesneedtakethisform(thoughbybeinggenerousaboutwhatcountsas an “example” and an “expectation” almost any induction may). However, learningfromexamplescapturesanimportantsetofphenomena,andcovers a broad enough range that characterizations may apply to other forms of inductive inference. This chapter further focuses on developmental questions. How do infants and young children learn from examples, and what changes across the lifespan? The development of inductive inference is a particularly important question because induction is both (potentially) a subject of DescriptiveandInferentialProblems 3 developmentandamechanismorsourceofdevelopmentalchange.Manyof the changes that occur over the lifespan may reflect learning from experi- ence: Children learn more about their world and culture and so become more adult-like in their inferences (e.g., Carey, 1985). Infantsclearlylearnfromexperience(e.g.,Rovee-Collier&Barr,2001). Atthesametime,therearemanydevelopmentalprocessesthatlikelyaffect the nature of such learning. As children acquire language, develop abstract representations, and are exposed to formal instruction, what and how they learn from examples changes. Whether there is continuity in processes of inductive inference, or whether development involves the acquisition of newformsofinferenceisamajorsourceofdebate.Debatesaboutthenature ofinductiveinferencehavealonghistoryincognitivedevelopment.Alter- native positions have been clearly articulated and defended with empirical results. One of the primary goals of this chapter is to provide a unified account of these alternatives. 2. THEORY-BASED AND SIMILARITY-BASED INDUCTIVE INFERENCE There are two primary approaches to inductive inference, similarity based and theory based. This basic dichotomy appears in many forms, with alternatives characterized in slightly different ways (e.g., “emergent” vs. “structured probability,” Griffiths, Chater, Kemp, Perfors, & Tenenbaum, 2010;McClellandetal.,2010).Insimilaritytheories,learningfromexamples involves forming associations or other representations of patterns of co-occurrence (e.g., Hampton, 2006, see papers in Hahn & Ramscar, 2001). Such accounts typically posit continuity in inductive inference, both phylogenetically and ontogenetically. They tend to invoke domain-general mechanisms and emphasize internalizing structure from environment. Changes in inductive inference are a result of changing experience: As the childformsdifferentassociations,comestorepresentmoreormorecomplex patterns in experience, their thinking changes. Alternative, theory-based approaches treat learning from examples as a form of hypothesis testing (Chater, 2010; Gelman & Koenig, 2003; Gopnik et al., 2004; Murphy & Medin, 1985). Such accounts often emphasize domain-specificity (in the hypotheses available) and are congenial to nativists (e.g., innate sources of hypotheses). Theory-based views involve some developmental discontinu- ities,atleastphylogenetically(itisunlikelythatsimpleorganismstesthypoth- eses). As hypothesis-testing seems to be a more complex cognitive process 4 CharlesW.KalishandJordanT.Thevenow-Harrison thanassociationformation,anaturaldevelopmentalhypothesisisthatinfants maystartmakingsimilarity-basedinductionsbutacquiretheory-basedinduc- tion at some point. Asthedescriptionsofferedaboveillustrate,similarity-basedandtheory- based views differ on a number of dimensions. While distinctions between the two approaches have organized much of contemporary research (see Feeney & Wilburn, 2008; Gelman & Medin, 1993; Pothos, 2005; Sloutsky&Fisher,2008;Smith,Jones,&Landau,1996),itisnotalwaysclear just where the critical differences lie. For example, similarity-based approaches tend to emphasize domain generality and continuity across development,butneednotdoso.Inmotivatingourproposalforaunifying framework,wefirstconsidersomealternativewaysofcharacterizingthetwo approaches to inductive inference. Similarity-basedtheoriesareoftencharacterizedby“bottom-up”build- ing of associations from basic, perceptual, experience (Smith et al., 1996). Theory-based accounts emphasize “top-down” application of conceptual structures or constraints to organize experience (Wellman & Gelman, 1992). In the developmental literature, similarity-based theories are often associatedwiththeviewthatyoungchildren’sinductiveinferencesarebased on apparent, perceptual features (see Keil, Smith, Simons, & Levin, 1998; Springer, 2001). Children learn from examples by forming associations between perceptual features. Theory-based views hold that even young children organize experience using abstract, theoretical, concepts, such as “cause” or “belief” (Carey, 1995; Wellman & Gelman, 1992). Children canlearnnotjustperceptualassociationsinexperience,butrelationsinvolv- ing nonperceptual properties as well (Mandler, 2004). This framing of the alternatives has led to substantial research about children’s representations of nonperceptual information (e.g., Gelman & Markman, 1986; Kalish, 1996;Wellman&Estes,1986;butseeSloutsky&Fisher,2008).However, we suggest that the perceptual versus abstract features distinction is largely orthogonaltowhetherinductionisbestcharacterizedassimilarityortheory based. For example, it is quite possible to learn similarity relations among abstract features. Aseconddimensionofdistinctionisrulesversusgradedrepresentations. Theory-based inferences are characterized as all-or-none judgments based on rules or criterial features (Sloutsky, Lo, & Fisher, 2001). For example, indeterminingthecategorymembership(andthusthebasisforfuturepre- dictions) of an animal, its parentage is particularly informative and other information (e.g., location) is largely irrelevant. The critical features may DescriptiveandInferentialProblems 5 be unknown: An underlying “essence” determines category membership and forms the basis for inductive inferences (Gelman, 2003). The point is that adistinctionis madebetweenthosefeatures that truly determinecate- gorymembership,orcauseobjectstohavethepropertiestheydo,andthose featuresthataremerelyassociatedwithotherfeatures.Theory-basedinduc- tiveinferencedependsonidentifyingthecritical(causal,essential)features. In contrast, similarity-based theories emphasize patterns of associations acrossanumberoffeatures.Anypatternofassociationcanbeusefulforpre- diction and inference: There is no distinction between “really” and “merely”associated.Featuresareusefulforpredictionbecauseoftheirinfor- mationalvalue:Doesobservingonefeatureaffecttheprobabilityofobserv- ing another? This perspective tends to emphasize graded or probabilistic judgments (Yurovsky, Fricker, Yu, & Smith, 2013). Multiple features or patternsofassociationcanbepresentatanyonetime(e.g.,ananimallooks likeadogbuthadbearparents).Inferenceinvolvescombiningthesefeatures (e.g.,weighting bypastdiagnosticity;see Younger,2003). Researchmoti- vatedbythiscontrastaddressesselectivityininductivejudgments(Kloos& Sloutsky, 2008; Sloutsky et al., 2001). Do children privilege some features over others when making inductive inferences? Can such preferences be tracedbacktopatternsofassociationordotheyinvolvebeliefsaboutcauses and essences (Gelman & Wellman, 1991; Kalish & Gelman, 1992)? For example, when a child judges that an animal that looks like a dog but has bear parents will have internal organs of a bear rather than a dog, are they usingaruleorprinciplethat“parentsmatter”oraretheybasingtheirjudg- ment on the past reliability of parentage over appearance? The question of thegradedversuscriterialbasisofchildren’sinferenceshasmotivatedsignif- icant research but is also largely orthogonal to the distinction we wish to draw. There are a number of other ways of distinguishing between theory- based and similarity-based inductive inference. For example, theories may involve conscious deliberate judgment, while similarity is unconscious andautomatic(seeSmith&Grossman,2008).Wesuggestthatallthesedis- tinctionsaresymptomsorconsequencesof amorefundamentaldifference. Theory-based accounts treat examples as evidential; similarity-based accounts treat examples as constitutive. In theory-based inference, the examples a person has encountered provide evidence for a relation (Gelman,2003;Gopnik&Wellman,1994;Murphy&Medin,1985).That allthedogsonehasseensofarhavebarkedprovidesevidencethatthenext dog observed will also bark. In contrast, for similarity-based views, the 6 CharlesW.KalishandJordanT.Thevenow-Harrison predictionaboutthenextdogisakindofreportofthatpastexperience.The characterizations of theory-based inference discussed above are a conse- quence of attempts to explicate evidential inferences in terms of scientific theories (see Gopnik & Wellman, 1992, 1994). Scientists use theories to interpret evidence, and evidence is used to develop and refine theories. Toassertthatyoungchildrentreatexamplesasevidenceistoassertthatthey dowhat scientists do.There is also atradition of formal approaches to evi- denceevaluationinthestatisticalandphilosophicalliterature.Aspsycholo- gists have adopted these formal approaches, a new characterization of theory-based inference has been developed (Gopnik et al., 2004; see Oaksford & Chater, 2007; Tenenbaum & Griffiths, 2001; Xu & Tenenbaum, 2007a). Theory-based inference is a type of statistical infer- ence. Similarity-based inference is also a type of statistical inference. This common grounding in statistical theory, induction as statistical inference, provides a unified perspective on theory-based and similarity-based accounts. We develop this unified perspective below and use it to identify justwhatisatissueinthedebatebetweentheory-basedandsimilarity-based views. This perspective leads directly to empirical tests of the two views. 3. INDUCTION AS STATISTICAL INFERENCE: DESCRIPTIVE AND INFERENTIAL PROBLEMS Makingastatisticalinferenceinvolvestwosteps:describingtheavail- able data and then generalizing. For example, after conducting an experi- ment, a researcher needs to describe her results. She may compute the mean and standard deviations of observations in the various conditions. Thosedescriptivestatisticsconveyinformationaboutpatternsinthesample, in the observed data. The researcher’s next step is to make some general claims basedon those descriptivestatistics. Shewants to estimatea popula- tion parameter or identify the generative process that produced the obser- vations.Thisstepinvolvescomputinginferentialstatistics(e.g.,at-test).Ina nutshell,similarity-basedapproachestoinductiveinferencefocusonthefirst step:Thedescriptiveproblemofcharacterizingpatternsinthedata.Theory- basedapproachesfocusonthesecondstep:Theinferentialproblemofesti- matingagenerativeprocess.Infleshingoutthischaracterizationofinductive inference,weintroduceanumbertermsanddistinctions,manyofwhichare illustrated in Fig. 1.1. The descriptive problem in inductive inference is noticing patterns in experience. Some patterns may be obvious, some less so. Children may