COMPUTATIONAL ISSUES IN THE STATISTICAL DESIGN AND ANALYSIS OF EXPERIMENTAL GAMES

One goal of experimental economics is to provide data to identify mode ls that best describe the behavior of experimental subjects and, more generally, human economic behavior. We discuss here what we think are the three main steps required to make experimental investigations of e conomic games as statistically informative as possible: finding the so lution of the experimental game under the postulated equilibrium or ot her economic models, selecting from a potential class of experimental designs the optimal one for discriminating between those models, and c hoosing an optimal stopping rule that indicates when to stop sampling data and accept one model as the best explanation of the data. Each st ep can be computationally intensive. We offer an algorithmic presentat ion of the necessary computations in each of the three steps and illus trate these procedures by examples from our research on learning model s in experimental games with incomplete information. These three steps of experimental design and analysis are not limited to experimental g ames, but the computational burden of implementing these algorithms in other experimental environments-for example, market experiments-requi res further considerations with which we have not dealt.