A0989
Title: Hierarchical Bayesian framework for evidence synthesis across federated data networks
Authors: Fan Bu - University of Michigan (United States) [presenting]
Abstract: Evidence synthesis, or meta-analysis, is crucial for combining results from multi-site studies, especially when subject-level data cannot be shared. Traditional meta-analysis approaches relying on mixed effects models with normality assumptions are often inadequate when single sites have small sample sizes or rare outcome events. In addition, bias can arise due to systematic error when using observational data sources. To address both challenges, a Bayesian evidence synthesis framework is proposed that performs likelihood-based meta-analysis and meanwhile corrects for bias. A Bayesian hierarchical model is introduced that jointly analyzes profile likelihoods instead of point estimates, which well accommodates small sample sizes and rare outcomes. Bias correction is achieved by learning an empirical bias distribution from a large number of negative control outcomes that are not associated with exposures of interest. A hierarchical Dirichlet process prior is employed to flexibly characterize bias globally and locally while borrowing information across data sources. A Hamiltonian Monte Carlo algorithm is implemented to efficiently carry out posterior inference. Through a series of simulation studies and a case study comparing Type-2 diabetes treatments across a multi-national data network, it is demonstrated that the approach can provide more precise estimates and coherent results for improved interpretability.
