Policy Learning with Confidence

Abstract: This paper proposes a framework for selecting policies that maximize expected benefit in the presence of estimation uncertainty, by controlling for estimation risk and incorporating risk aversion. The proposed method explicitly balances the size of the estimated benefit against the uncertainty inherent in its estimation, ensuring that chosen policies meet a reporting guarantee, namely that the actual benefit of the implemented policy is guaranteed not to fall below the reported estimate with a pre-specified confidence level. This approach applies to a variety of settings, including the selection of policy rules that allocate individuals to treatments based on observed characteristics, using both experimental and non-experimental data; and the allocation of limited budgets among competing social programs; as well as many others. Across these applications, the framework offers a principled and robust method for making data-driven policy choices under uncertainty. In broader terms, it focuses on policies that are on the efficient decision frontier, describing policies that offer maximum estimated benefit for a given acceptable level of estimation risk.