Examination of a Bayesian Joint Modeling Approach for Handling Missing Moderators in Meta-Regression

Abstract

Meta-regression is used to understand the role of moderators in a meta-analytic model. However, during data extraction it is common for the data to not be clearly presented, incomplete, or missing. Consequently, missing study and participant characteristics arise, which can make it difficult to estimate meta-regression models. This dissertation examines a Bayesian conditional joint modeling (CJM) method for handling missing moderators in meta-regression using a series of conditional distributions. The use of CJM has been proposed in the meta-analysis literature to predict missing moderators (Hemming, Hutton, Maguire, & Marson, 2010). However, its performance has yet to be empirically studied. This dissertation investigated the CJM approach through a simulation study. Results suggest that the CJM approach performed similarly to listwise deletion when estimating the missing moderator but performed better when estimating the overall true effect-size.