A0401
Title: Generalizing transfer learning: A flexible doubly robust estimation approach for missing data
Authors: Tianying Wang - Colorado State University (United States) [presenting]
Abstract: Modern large-scale datasets often face challenges with missing data, particularly binary covariates, which are prone to complete absence due to their reliance on specific questionnaires or detailed clinical evaluations. The absence of these covariates, often significant confounders, introduces bias in estimation and inference and reduces prediction accuracy. Traditional methods, reliant on mechanisms like missing at random, cannot address completely missing binary covariates. To overcome this, a doubly robust transfer learning framework is proposed that utilizes a novel sub-group shift assumption to effectively leverage data from source populations. The method, proven to be $n^{1/2}$-consistent and asymptotically normal, is validated through simulations and an application using UK Biobank data. A critical gap is filled in missing data methodologies and offers a practical solution for mitigating bias in modern studies.
