Adaptive treatment assignment in experiments for policy choice

The goal of many experiments is to inform the choice between different policies. However, standard experimental designs are geared toward point estimation and hypothesis testing. We consider the problem of treatment assignment in an experiment with several non-overlapping waves, where the goal is to choose among a set of possible policies (treatments) for large-scale implementation. The optimal experimental design learns from earlier waves and assigns more experimental units to the better-performing treatments in later waves. We propose a computationally tractable approximation of the optimal design that we call "exploration sampling," where assignment probabilities are an increasing concave function of the posterior probabilities that each treatment is optimal. Theoretical results and calibrated simulations demonstrate improvements in welfare, relative to both non-adaptive designs as well as bandit algorithms. An application to selecting between different recruitment strategies for an agricultural extension service in Odisha, India demonstrates practical feasibility.