Xiang Meng presents Time-Varying Causal Survival Learning

Abstract: Understanding the causal effects of time-varying treatments is critical in many domains, such as evaluating the impact of heart transplants on patient outcomes. Physicians face a challenge in determining whether transplants significantly improve patients' health and longevity, particularly when the timing of transplants varies across individuals. This timing variability, if not properly accounted for, can introduce bias and lead to misleading conclusions. Inspired by this problem, we propose a novel framework, Time-Varying Causal Survival Learning (TV-CSL), which extends the staggered adoption framework from econometrics to time-to-event variables. By leveraging techniques from survival analysis, we derive an efficient causal estimator that accommodates the dynamic nature of treatments. Additionally, we incorporate double machine learning methods to handle complex, non-linear relationships between covariates and outcomes, improving the estimator’s efficiency. Simulation results show that our estimator outperforms traditional methods by reducing bias and offering theoretical guarantees for improved efficiency.