Abstract
We study how to design experiments for the objective of choosing optimal polices. An experimenter wants to choose a policy to maximize welfare subject to budget or other policy constraints. The effects of counterfactual policies are described by a structural econometric model governed by an unknown parameter. The experimenter has access to some pilot data, and has the opportunity to collect additional data through an experiment. The joint experimental design and policy choice problem is a dynamic optimization problem with a very high-dimensional state space, since the chosen policy depends on the realized data. We propose a low-dimensional approximation to the solution and show it is asymptotically optimal under Bayes expected welfare. The method applies to policies allocating discrete as well as continuous treatments, such as cash transfers, prices, or tax credits, which may be targeted on the basis of covariates. We demonstrate the method using the conditional cash transfer program Progresa, showing how to design an experiment to help choose a policy aimed at increasing graduation rates and reducing gender disparities in education. Compared to the original Progresa experiment, the optimal experiment requires only one quarter as many observations to obtain equally effective policies.

View Samuel Higbee's Full Job Market Paper