Proximal Learning for Individualized Treatment Regimes Under Unmeasured Confounding

Data-driven individualized decision making has recently received increasing research interest. However, most existing methods rely on the assumption of no unmeasured confounding, which cannot be ensured in practice especially observational studies. Motivated by proposed proximal causal inference, we develop several learning to estimate optimal treatment regimes (ITRs) presence confounding. Explicitly, terms two types proxy variables, are able establish identification results for different classes ITRs respectively, exhibiting tradeoff between risk untestable assumptions and potential improvement value function making. Based these results, propose classification-based approaches finding a variety restricted in-class their theoretical properties. The appealing numerical performance our is demonstrated via extensive simulation experiments real data application. Supplementary materials this article available online.