In this paper, a general framework is proposed for the use of (quasi-) experimental data when choosing policies such as tax rates or the level of inputs in a production process. The data are used to update expectations about social welfare as a function of policy, and the policy is chosen to maximize expected social welfare. We characterize the properties of the implied decision function. For settings where experimentation is feasible, we characterize experimental designs maximizing expected social welfare, assuming policies are chosen based on the data. We discuss several economic settings which are covered by our framework, and apply our methods to data from the RAND health insurance experiment. In this application, we obtain much smaller estimates of the optimal copay (18% vs. 50%) than those obtained using a conventional sufficient-statistic approach. Our approach combines optimal policy theory, statistical decision theory, and nonparametric Bayesian methods. It explicitly takes into account economic theory, does not rely on restrictions of functional form or heterogeneity, and provides a transparent mapping from observations to policy choices. This mapping is optimal in finite samples.
