Bayesian Hierarchical Models for the Combination of Data from Heterogeneous Sources

Abstract

Combining information from multiple data sources can yield more comprehensive andreliable results than relying on a single source. Comparative effectiveness research depends heavily on such integrated data, with the primary goal of synthesizing evidence to support researchers and inform decision-making aimed at improving population health. Although data integration is a promising approach for generating valuable evidence in biomedical research, there remains a lack of advanced methodological tools for accurate estimation in several settings commonly encountered in practice. In this dissertation, we develop methods to (1) adjust estimated treatment effects in aggregate-level data network meta-analysis to account for potential study-level bias, (2) estimate treatment effects in meta-analyses with longitudinal individual patient-level data in the presence of systematic outcome missingness, and (3) develop methods to address non-ignorable missingness in network meta-analyses of longitudinal individual patient–level data. These methods are developed under a Bayesian framework, which offers flexibility in modeling complex data structures and provides intuitive interpretation of uncertainty. Overall, this dissertation underscores the importance of leveraging all available evidence to obtain robust and clinically meaningful treatment effect estimates, thereby strengthening the reliability of comparative effectiveness research in real-world settings.