How reliable are treatment effects in clinical trials with dropout?

Dropout is widespread in clinical trials and real-world oncology studies, with up to half of patients leaving before the study ends due to side effects or other reasons. When such dropout is informative (i.e., dependent on survival time), it induces censoring bias that distorts causal survival analysis and leads to biased treatment effect estimates. This challenge is particularly acute when estimating conditional average treatment effects (CATEs), which are central to personalized medicine because they reveal which patients benefit most from treatment. In this paper, we propose an assumption-lean method to assess the robustness of CATE estimates in survival analysis when facing censoring bias. Specifically, we frame the underlying task through the lens of partial identification, which allows us to obtain informative bounds on the CATE under such conditions. Importantly, this approach helps identify patient subgroups where treatment is still effective despite potential censoring. We then show that our bounds converge to the true point estimates of the CATE when the censoring bias goes to zero. We further propose a novel model-agnostic meta-learner to estimate the bounds that can be used combined with arbitrary machine-learning models and that has favorable theoretical properties such as double-robustness and quasi-oracle efficiency. We finally demonstrate the effectiveness of our meta-learner across various experiments using both simulated and real-world data.