Abstract

There is a dearth of flexible methods that leverage machine learning for estimating the causal effects of multiple time-varying treatments with censored survival outcomes using large-scale observational data. In this work, we develop a novel Bayesian causal inference framework with several key advancements. First, we introduce a continuous-time joint marginal structural model for censored survival outcomes based on restricted mean survival time (JMSM-RMST). Unlike commonly used proportional hazards structural models, JMSM-RMST does not rely on the proportional hazards assumption, providing unambiguous causal interpretations of the structural parameters. We then develop a novel Bayesian additive regression trees survival model (BMTree-Surv) that incorporates time-varying covariates. The BMTree-Surv model is used to accurately estimate time-varying weights for the JMSM-RMST model, effectively mitigating selection bias arising from time-varying confounding. Finally, we integrate these developments through a modularized Bayesian inference approach, enabling coherent inferences about the causal survival effects of time-varying treatments. We conduct extensive simulations to examine the practical operating characteristics of the proposed methods. Lastly, we apply our methods to a large-scale electronic health records dataset from the Yale New Haven Health System to estimate the causal survival effects of different antihypertensive treatment timing strategies in patients with severe hypertension.