Econometrica,Vol.69,No.6(November,2001),1597–1628 ANEVALUATIONOFECONOMETRICMODELSOF ADAPTIVELEARNING By Timothy C. Salmon 1 Thispaperevaluatestheeffectivenessoffoureconometricapproachesintendedtoiden- tifythelearningrulesbeingusedbysubjectsinexperimentswithnormalformgames.This is done by simulating experimental data and then estimating the econometric models on the simulated data to determine if they can correctly identify the rule that was used to generatethedata.Theresultsshowthatallofthemodelsexaminedpossessdifficultiesin accuratelydistinguishingbetweenthedatageneratingprocesses. Keywords: Learning,behavioralgametheory,reinforcementlearning,fictitiousplay. 1 introduction In recent game theoretic and experimental literature, learning theories have become increasingly popular. Theorists have been using them as a means of explaining how certain equilibria might be reached or why some are more plausible than others while experimentalists have been using them to describe how players’ choices evolve through the course of experiments. This popularity hasproducedalargenumberofdifferentlearningrules. There is, consequently, a growing literature concerned with trying to identify whichamongtheselearningrulesappearstobeusedbysubjectsinexperiments. A variety of experimental designs and games have been used but this paper will focus only on experiments involving normal form games. The fundamental question of interest here is whether any of the various econometric models are capable of accurately identifying the learning rule in use by a population or an individual.Itisimportanttonotethatthispaperdoes not intendtoaddressthe questiondirectlyofwhichlearningrule(s)bestdescribesreallearningbehavior, but only if aparticularruleisinuse,canthesemethodsaccuratelyandreliably identifyit. Toaddressthisquestion,themodelsfromsomeofthemoreprominentpapers in this area, Mookherjee and Sopher (1997), Mookherjee and Sopher (1994), CheungandFriedman(1997),andCamererandHo(1999),willbetestedusing simulatedexperimentaldata.Experimentaldatawillbesimulatedbyhavingcom- puter players choose according to several different specifications of learning 1 The author would like to thank Edi Karni, H. Peyton Young, various seminar attendees at The JohnsHopkinsUniversity,ColumbiaUniversity,TexasA&MUniversity,TheUniversityofCalifornia at Irvine, The University of Pittsburgh, and the anonymous referees for their helpful and insightful comments. The author would further like to thank Steve Holden for technical assistance with the estimation procedures. 1597