Behavior in a dynamic decision problem: An analysis of experimental evidence using a bayesian type classification algorithm

It has been long recognized that different people may use different strategies, or decision rules, when playing games or dealing with other complex decision problems. We provide a new Bayesian procedure for drawing inferences about the nature and number of decision rules that are present in a population of agents. We show that the algorithm performs well in both a Monte Carlo study and in an empirical application. We apply our procedure to analyze the actual behavior of subjects who are confronted with a difficult dynamic stochastic decision problem in a laboratory setting. The procedure does an excellent job of grouping the subjects into easily interpretable types. Given the difficultly of the decision problem, we were surprised to find that nearly a third of subjects were a “Near Rational” type that played a good approximation to the optimal decision rule. More than 40% of subjects followed a rule that we describe as “fatalistic,” since they play as if they don’t appreciate the extent to which payoffs are a controlled stochastic process. And about a quarter of the subjects are classified as “Confused,” since they play the game quite poorly. Interestingly, we find that those subjects who practiced most before playing the game for money were the most likely to play poorly. Thus, lack of effort does not seem to account for poor performance. It is our hope that, in future work, our type classification algorithm will facilitate the positive analysis of peoples’ behavior in many types of complex decision problems.