Title: Augmenting Randomized Experiments Using Foundation Models

Abstract: Foundation models offer opportunities to improve the efficiency of randomized experiments, but replacing human participants with “digital twins” risks biased causal inference. We introduce hybrid augmented inverse probability weighting (H-AIPW), which adaptively combines predictions from multiple foundation models and external data with trial-based outcome models to estimate average treatment effects. Under appropriate sample splitting and regularity conditions, H-AIPW preserves valid inference despite inaccurate external predictions and achieves asymptotic variance no greater than that of any component estimator, including trial-only AIPW. Evaluations across eight social science experiments, simulations, and a clinical trial demonstrate precision gains, supporting foundation models as tools for augmenting randomized experiments.

Issa Dahabreh, MD ScD, is Associate Professor of Epidemiology and Biostatistics at the Harvard T.H. Chan School of Public Health, and Section Head for Epidemiology and Data Science at the Richard A. and Susan F. Smith Center for Outcomes Research. His research develops methods for causal inference, evidence synthesis, and the use of trial and external data to improve the design and analysis of studies informing clinical and public health decisions.