Primary Submission Category: Policy Learning
Optimal Policy Learning for Recurrent Outcomes via Instrumented Difference-in-Differences: An Application to T2DM Treatment
Authors: Ritoban Kundu, Ashkan Ertefaie, Sean Hennessy, James Flory,
Presenting Author: Ritoban Kundu*
Learning reproducible, generalizable optimal treatment policies for chronic diseases requires large, representative populations observed over extended periods. While administrative health data offer an attractive foundation, their utility is often compromised by unmeasured confounding. We address this by proposing a novel framework based on Instrumented Difference-in-Differences (iDID) to estimate optimal policies for recurrent event outcomes subject to a terminating event. The iDID design is particularly advantageous as it relies on fewer and weaker assumptions than conventional instrumental variable or difference-in-differences methods. A key feature of our approach is that it explicitly addresses the fundamental challenge of avoiding policies that trivially reduce recurrent adverse events by increasing mortality, a common pitfall in policy learning for chronic disease settings. We develop a multiply robust estimator that remains consistent if any one of several subsets of nuisance models is correctly specified. Theoretical results establish the estimator’s consistency and large-sample behavior. Simulations demonstrate that our estimator outperforms existing approaches in finite samples. We apply this method to a national Medicare dataset to optimize first-line Type 2 Diabetes strategies for minimizing disease-related hospitalizations.