Abstract
Individualized treatment rule (ITR) is a stepping stone to precision medicine. To ensure validity, ITRs are ideally derived from randomized trial data, but the use cases of ITRs extend beyond these trial populations. Transferring knowledge from experimental data to real-world data is of interest, while experimental data with selective inclusion criteria reflect a population distribution that may differ from the real-world target. In well-designed experiments, granular information crucial to decision making can be thoroughly collected. However, part of this may not be accessible in real-world scenarios. We propose a learning scheme for ITR that simultaneously addresses the issues of covariate shift and missing covariates with a quantile-based optimal treatment objective. Specifically, we compare the outcome uncertainty across treatment arms that is due to missing covariates and use it to guide treatment selection to reduce the likelihood of worse outcomes. The performance of this method is evaluated in simulations and a sepsis data application.