Title:Targeted Data Fusion for Causal Survival Analysis Under Distribution Shift

Abstract:Causal inference across multiple data sources has the potential to improve the generalizability, transportability, and replicability of scientific findings. However, data integration methods for time-to-event outcomes -- common in medical contexts such as clinical trials -- remain underdeveloped. Existing data fusion methods focus on binary or continuous outcomes, neglecting the distinct challenges of survival analysis, including right-censoring and the unification of discrete and continuous time frameworks. To address these gaps, we propose two novel approaches for multi-source causal survival analysis. First, considering a target site-specific causal effect, we introduce a semiparametric efficient estimator for scenarios where data-sharing is feasible. Second, we develop a federated learning framework tailored to privacy-constrained environments. This framework dynamically adjusts source site-specific contributions, downweighting biased sources and upweighting less biased ones relative to the target population. Both approaches incorporate nonparametric machine learning models to enhance robustness and efficiency, with theoretical guarantees applicable to both continuous and discrete time-to-event outcomes. We demonstrate the practical utility of our methods through extensive simulations and an application to two randomized trials of a monoclonal neutralizing antibody for HIV-1 prevention: HVTN 704/HPTN 085 (cisgender men and transgender persons in the Americas and Switzerland) and HVTN 703/HPTN 081 (women in sub-Saharan Africa). The results highlight the potential of our approaches to efficiently estimate causal effects while addressing heterogeneity across data sources and adhering to privacy and robustness constraints.