LEAPS-MPS: Empirical Bayes Evidence Synthesis Techniques for Hybridizing Observational and Experimental Data

When evaluating the efficacy of a health intervention, researchers may have access to two distinct types of data: observational data from sources such as electronic health records or insurance claims databases, and experimental data from randomized trials. Observational data are useful for their scope and representativeness, but the individuals who receive the intervention may fundamentally differ from those who do not, making causal estimation difficult. In contrast, experimental data use random treatment assignment to facilitate accurate causal effect estimation, but such experiments are often costly and time-consuming to conduct. Hence, regulatory agencies and statisticians have advocated for methodologies that integrate observational data with randomized trials. Despite increased research attention, there is little consensus on how to effectively combine these two data types for many common statistical procedures. This proposal aims to develop evidence-synthesis techniques to enhance causal effect estimation and inference and to design more efficient experiments. The primary toolkit for this research will be Empirical Bayes procedures, a flexible paradigm for weighting between competing estimators. The goal is to enable practitioners to estimate causal effects more efficiently in diverse settings. The principal investigator has identified three causal estimation tasks where current data-integration methods can be improved: 1) average treatment effect (ATE) estimation; 2) modeling of flexible functional estimands, such as dose-response curves and conditional average treatment effect (CATE) functions; and 3) the incorporation of multiple datasets of each type. Additionally, this proposal includes projects to narrow confidence intervals by incorporating observational data and to design more efficient and adaptive randomized trials that explicitly complement existing observational data. To demonstrate their efficacy and utility in practical settings, methods developed under this proposal will be deployed on two datasets: the Women’s Health Initiative, a study of the health effects of hormone therapy that includes observational and experimental components; and an air quality and Medicare insurance claims database maintained by the National Studies on Air Pollution and Health group. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.