Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
Time-Varying Causal Survival Learning
Authors: Xiang Meng, Iavor Bojinov,
Presenting Author: Xiang Meng*
The paper tackles a key challenge in causal inference: how to accurately estimate causal effects when the timing of treatment varies from patient to patient. Drawing on the well-known Stanford Heart Transplant study, we show how staggered adoption assumptions can be combined with survival analysis techniques to address this issue. In organ transplantation, for example, factors like donor availability and patient readiness often determine when a patient receives treatment, which can bias estimates if not handled correctly. By identifying the conditions that link staggered adoption designs to survival analysis, we demonstrate how existing survival methods can retain their causal interpretability under time-varying treatments. We further boost the precision and robustness of these estimates by incorporating double machine learning, which allows us to manage complex relationships between patient characteristics and survival outcomes. Through simulations and an analysis of heart transplant data, our approach outperforms traditional methods, reducing bias and offering theoretical guarantees for greater efficiency in survival analysis.