Evaluating Meta-Regression Models with Simulation Studies and Machine Learning Driven Imputation

Abstract

Despite the large depth of location and time dimensions in typical meta-regression datasets, the majority of the modeling only occurs at the study-level with little to no consideration for more precise location and/or time groupings. This dissertation addresses this gap by developing simulation frameworks to evaluate how alternative modeling strategies perform when meta-regression methodology specifications are applied to subject-level meta-regression data. This work advances the robustness of empirical research by clarifying how each method performs when confronted with complex heterogeneity. It also contributes to a growing trend in the medical and social sciences toward strengthening methodological reliability in evidence synthesis and provides a foundation for future extensions that integrate real-world data. By systematically evaluating existing methods, extending them with traditional parametric and machine learning driven imputation approaches, and clarifying the risks of misspecified models, this dissertation advances methodological transparency and reliability in evidence synthesis. The results provide applied researchers in economics, medicine, and the social sciences with practical tools for choosing and validating meta-regression models under realistic data conditions. This research also lays the groundwork for future projects that extend the simulation frameworks to incorporate model diagnostics for heterogeneity, temporal-only heterogeneity, and real-world empirical applications. In doing so, this dissertation positions meta-regression as a more trustworthy tool for synthesizing knowledge in fields where reliable evidence is essential for policy and practice.