Abstract:

The objective of this paper is both to examine the performance and to show properties of statistical techniques used to estimate learning models on experimental game data. We consider a game with unique mixed strategy equilibrium. We discuss identification of a general learning model and its special cases, reinforcement and belief learning, and propose a paramaterization of the model. We conduct Monte Carlo simulations to evaluate the finite sample performance of two kinds of estimators of a learning model's parameters. Maximum likelihood estimators of period to period transitions and mean squared deviation estimators of the entire path of play. In addition, we investigate the performance of a log score estimator of the entire path of play and a mean squared deviation estimator of period to period transitions. Finally, we evaluate a mean squared estimator of the entire path of play with observed actions averaged over blocks, instead of behavioral strategies. We propose to estimate the learning model by maximum likelihood estimation as this method performs well on the sample size used in practice if enough cross sectional variation is observed.

Website