Prediction-Powered Data Fusion for Treatment Effect Estimation
The authors propose a framework to combine data from a small randomized controlled trial (RCT) and a large observational study (OBS) to estimate treatment effects without making strong assumptions about the OBS confounding. They developed AIPW-Fusion for the Average Treatment Effect (ATE) and two learners, DR-Fusion and R-Fusion, for Conditional Average Treatment Effects (CATE). This method preserves unbiasedness from the RCT while using the large OBS data to significantly boost estimation precision.