Fast and efficient robust Bayesian meta-analysis with spike-and-slab priors

Abstract

Meta-analyses are vital for synthesizing evidence across multiple studies, yet their validity is frequently undermined by publication bias. To address this issue, we developed robust Bayesian meta-analysis (RoBMA), which integrates multiple publication bias adjustment methods—such as selection models and PET-PEESE—using Bayesian model-averaging (BMA). Traditionally, RoBMA employed JAGS and bridge sampling to estimate individual models and combine them via BMA, providing a robust framework for evidence synthesis. In recent advancements, we expanded RoBMA to incorporate meta-regression and three-level models, enabling its application to a wide range of meta-analytic settings. However, these extensions substantially increase the model space, rendering traditional estimation methods computationally demanding. This talk reviews these methodological developments and introduces a significant enhancement: spike-and-slab model-averaging for both parameters and model components within RoBMA. Leveraging the spike-and-slab model-averaging markedly improves computational efficiency, facilitating applications to larger samples and more complex problems. The method retains RoBMA’s robustness while overcoming the computational bottlenecks of expanded model spaces, delivering fast and reliable estimates. We illustrate the practical benefits of this innovation through examples and discuss its implications for advancing meta-analytic research.