Games and Economic Behavior 42 (2003) 116–136www.elsevier.com/locate/gebLearning by not doing: an experimental investigationof observational learningAntonio Merloa,cand Andrew Schotterb,∗aDepartment of Economics, University of Pennsylvania, 3718 Locust Walk, Philadelphia, PA 19104, USAbDepartment of Economics, New York University, 269 Mercer Street, New York, NY 10003, USAcCEPR, UKReceived 6 September 2000AbstractWe present experimental evidence suggesting that observational learning (i.e., learning-by-not-doing but by observing) may outperform learning-by-doing. 2002 Elsevier Science (USA). All rights reserved.JEL classification: C91; D83Keywords: Learning1. IntroductionCommon sense might suggest that if two identical rational economic agents who areattempting to maximize the same objective function have the same data or information setat their disposal, they will make the same choice. Perhaps for this reason it is hardly everasked whether the process through which this data was generated matters or is it sufficientto merely know the stock of information the agents have in their possession at the timethe choice is made. This question is clearly important for many economic situations. Forexample, think of an environment where a task is repeatedly performed and the expertise ofthe decision maker increases as a function of the cumulative stock of output produced as ina learning-by-doing environment. In such an environment, does the stock of observationsgenerated along the way summarize all the information needed to make efficient economic*Corresponding author.E-mail addresses: [email protected] (A. Merlo), [email protected] (A. Schotter).0899-8256/02/$ – see front matter  2002 Elsevier Science (USA). All rights reserved.doi:10.1016/S0899-8256(02)00537-7A. Merlo, A. Schotter / Games and Economic Behavior 42 (2003) 116–136 117decisions or is the actual experience that led to these observations important? In this paperwe present experimental evidence that addresses this question.One can envision learning in distinctly different ways. Observational learning, wouldhave a decision maker (the observer) simply looking over the shoulder of another decisionmaker (the doer) as he performs his learning-by-doing task (see (Jovanovic and Nyarko,1995) for an application of an apprentice-craftsman model similar, in spirit, to the onewe explore here). Here the observer simply keeps his eyes open but his hands in hispocket. Still after enough repetition both the observer and the doer would have the sameobservations and if they are rational and out to maximize the same objective function, wemight actually expect the doer to perform no worse than the observer since he or she hashad additional hands-on experience.A large number of animal studies support the hypothesisthat observational learning canbe efficient. For example, Terkel (1996) shows that young rats seem to learn how to skinpine cones quite easily by observing their mothers when they are young. John et al. (1969)demonstrates that cats can be trained to perform tasks by being placed in a cage divideddown the middle with a glass wall through which they can observe an animal alreadytrained in a particular task perform. After a certain amount of observation, the cat seemsperfectly able to perform the task despite the fact that it has never done it.In this paper we investigate the hypothesis that the process through which informationis gathered can affect the choices of experimental subjects drawn at random from identicalpopulations. What we find is that subjects can sometimes learn better by not doing but bywatching. A possible explanation is that when subjects perform an experiment or partici-pate in a market repeatedly and earn a small payoff each period, as is true for most markets,the feedback of the market tends to focus attention of the decision makers on the myopicstimulus-responseaspect of the problemand seemsto preventthem from learning the trade-offs existing in the experiment (see (Merlo and Schotter, 1999), henceforth M–S, for afuller exposition of this point). However, when such behavior is merely observed, attentionis again re-focused and performance is enhanced. To some extent, this explains the valueof coaches, theatrical directors, Sunday morning quarterbacks, and kibitzers in general.Another finding which we consider of interest is the fact that not only is observationpotentially performance improving, but how much better an observer does dependscrucially on who they observe. Put differently, those observers watching doers who didrelatively well made far better decisions than those who watched relatively poor doers.Good role models are crucial to good performance.In the remainder of this paper we proceed as follows: In Section 2 we discussthe decision task that our subjects performed and our baseline experimental design. InSections 3 and 4 we present the results of our baseline experiment. In Section 5 we reportthe results of two additional experiments run to answer potential criticisms of our baselineexperiment. Finally, in Section 6 we offer some conclusions.2. The decision task and the experimental designAs was true in M–S, all of the experiments performed to investigate our hypotheseswere of the tournament variety and similar to those of Bull et al. (1987) and Schotter118 A. Merlo, A. Schotter / Games and Economic Behavior 42 (2003) 116–136and Weigelt (1992). In those experiments, randomly paired subjects must, in each round,choose a number, e, between 0 and 100 called their decision number. After this numberis chosen, a random number is independently generated by each subject from a uniformdistribution over the closed interval [−a,+a]. These numbers, each player’s decisionnumber and random number, are then added together and a “total number” defined foreach subject. Payoffs are determined by comparing the total numbers of the subjects ineach pair and awardingthat subject with the largest total number a “big” payment of M andthat subject with the smallest total number a “small” payment m, M>m. The cost of thedecision number chosen, given by a convex function c(e) = e2/k, is then subtracted fromthese fixed payments to determine a subject’s final payoff. Hence, in these experimentsthere is a trade-off in the choice of decision numbers: higher numbers generate higherprobabilities of winning the big prize but also imply higher decision costs. Notice that inthe instructions, a