Disentangling Effect Size Heterogeneity in Meta-Analysis: A Latent Mixture Approach

An important task of meta-analysis is to observe, quantify, and explain the heterogeneity across the reported effect sizes of primary studies. A primary issue that challenges this task is the myriad of subtle factors that could have contributed to the observed heterogeneity. We leveraged the recent advances in theoretical machine learning to develop a novel latent mixture-based method for disentangling effect-size heterogeneity in meta-analysis. Mathematical analysis and simulation studies were carried out to demonstrate that, when the observed heterogeneity stems from more than 1 factor, our method can attain a substantially higher statistical power than the traditional methods for moderator analysis without requiring researchers to make judgment calls on which factors to consider or correct for in analyzing the observed heterogeneity. We also conducted a case study with real-world data to show how our method may be used to address long-standing inconsistencies in the literature. Translational Abstract An important task of meta-analysis is to explain the heterogeneity among primary studies. However, it is often a challenge for researchers to delineate the myriad of subtle factors that could have contributed to the observed heterogeneity. We leveraged the recent advances in theoretical machine learning, specifically the efficient decomposition of Gaussian mixture distributions, to develop a novel latent mixture-based method for disentangling heterogeneity in meta-analysis. As demonstrated by mathematical analysis and simulation studies for moderator estimation, our method can attain substantially higher statistical power than the traditional methods without requiring researchers to make judgment calls on which factors to consider or correct for in analyzing the observed heterogeneity.