Policy choice in experiments with unknown interference

This paper discusses experimental design to estimate welfare-maximizing policies. We consider a setting where units are organized into large, finitely many independent clusters and interact over unobserved dimensions within each cluster. The contribution of this paper is two-fold. First, we construct a test for whether a welfare-improving treatment configuration exists and hence worth learning by conducting a larger scale experiment. Second, we introduce an adaptive randomization procedure to estimate welfare-maximizing individual treatment allocation rules valid under unobserved interference. We derive asymptotic properties of the marginal effects estimators and finite-sample regret guarantees of the policy. Finally, we illustrate the method's advantage in simulations calibrated to an existing experiment on information diffusion.