1054-P: Reweighting All of Us Research Program Data to Correct for Sampling Bias—A Novel Machine Learning–Based Approach

Abstract

Introduction and Objective: The NIH’s All of Us (AOU) dataset provides a large, diverse sample but lacks representativeness due to volunteer- and site-based sampling. This study aimed to correct for selection bias in AOU through machine learning-based reweighting. Methods: AOU participants were matched to those from the National Health and Nutrition Examination Survey (NHANES, 2017-2020) using a graph neural network framework (GraphSAGE) combined with a clustering algorithm (KMeans). This method used 44 shared features, including sociodemographic and health variables, to cluster AOU and NHANES participants altogether into distinct groups, with each group containing a mix of participants from both datasets. NHANES complex survey design weights were then assigned to AOU participants within each group and rescaled. Results: The 8,921 NHANES and 291,631 AOU participants differed significantly across many characteristics. The average standardized mean difference between AOU and NHANES was substantially reduced from 0.123 (original AOU) to 0.077 (weighted AOU) (Table 1). The weighted estimates showed improved alignment with national estimates across most variables. Conclusion: Reweighting the AOU dataset enhanced its representativeness for key sociodemographic and health characteristics, broadening its utility for diverse epidemiological and health services research. Disclosure E. Staton: None. J. Lee: None. Z. Li: None. P. Li: None. C. Yang: None. J. Varghese: None. Y. Hong: Consultant; Weight Watchers International. I. Graetz: Research Support; Pfizer Inc, PRIME Education, LLC. H. Shao: None. Funding CDC (U18DP006711)

Ähnliche Arbeiten