Locally Asymptotically Minimax Statistical Treatment Rules Under Partial Identification

Policymakers often desire a statistical treatment rule (STR) that determines a treatment assignment rule deployed in a future population from available data. With the true knowledge of the data generating process, the average treatment effect (ATE) is the key quantity characterizing the optimal treatment rule. Unfortunately, the ATE is often not point identified but partially identified. Presuming the partial identification of the ATE, this study conducts a local asymptotic analysis and develops the locally asymptotically minimax (LAM) STR. The analysis does not assume the full differentiability but the directional differentiability of the boundary functions of the identification region of the ATE. Accordingly, the study shows that the LAM STR differs from the plug-in STR. A simulation study also demonstrates that the LAM STR outperforms the plug-in STR.