This project continues research begun under an earlier NSF grant on the dynamics and learning in multistage games. Since experiments, oligopoly experiments, bargaining experiments, committee experiments, and experiments in many other settings studied by social scientists are conducted by multi-stage games, a better understanding of these dynamics is critical to the kinds of inferences that can be made from the data. This project develops a unified theoretical and econometric approach with a structure that permits the rigorous analysis of diverse experimental data from multi-stage games. The theoretical development involves modification of standard game theoretic analysis by introducing random errors in actions (trembles) and beliefs (home-made priors and non-Bayesian updating). Different specifications of these random components, including assumptions about the degree to which subjects take errors into account in their own decision making, lead to several classes of models which can be rigorously evaluated and compared using data from simple experimental games. These alternative models can also be used to rigorously analyze - for the first time - data from earlier experiments that offer descriptive evidence about what appear to be systematic departures from fully rational behavior.
