B0953
Title: Heterogeneous causal effects in underrepresented populations: federated and transfer learning approaches
Authors: Larry Han - Harvard University (United States) [presenting]
Rui Duan - Harvard University (United States)
Abstract: In causal inference, it is challenging to estimate treatment effects for underrepresented populations accurately. Data sharing can improve the power to estimate treatment effects, but privacy concerns often constrain the sharing of patient-level data between sites. We propose a novel causal inference framework to leverage multiple sites and multiple subpopulations to make inferences on the conditional average treatment effect (CATE) for an underrepresented target population of interest. The method leverages transfer learning and federated learning to data-adaptively incorporate summary-level information from source populations and sites to learn about the target population CATE. In extensive simulation studies, we show that the proposed method substantially improves the estimation accuracy of the CATE for underrepresented target populations, lowering RMSE by up to $80\%$. We illustrate our method through a real-world study of COVID-19 vaccine efficacy on infection, hospitalization, and mortality in underrepresented populations using data from multiple VA study sites. Our method reduces the RMSE of the CATE for underrepresented populations by $35-65\%$.