Selecting relevant moderators with Bayesian regularized meta-regression

Area of Science:

• Statistics
• Biostatistics
• Meta-analysis

Background:

• Meta-regression is crucial for analyzing heterogeneity in diverse literature.
• A common challenge is the high number of potential moderators relative to studies, risking overfitting and non-convergence.
• Existing methods struggle with selecting relevant moderators from large candidate pools.

Approach:

• Introduced Bayesian Regularized Meta-Analysis (BRMA) using regularizing priors (LASSO, horseshoe) to shrink small coefficients, effectively selecting moderators.
• Compared BRMA against restricted maximum likelihood (RMA) random effects meta-regression via simulation.
• Developed open-source software implementations in R (pema package) and JASP.

Key Points:

• BRMA demonstrated superior predictive performance and better rejection of irrelevant moderators compared to RMA.
• While BRMA coefficients were slightly biased towards zero, residual heterogeneity estimates were less biased than RMA.
• BRMA performed well even with small sample sizes (as few as 20 studies).

Conclusions:

• BRMA offers a robust solution for meta-regression with numerous candidate moderators, especially in small sample settings.
• The method enhances model generalizability and reduces the risk of spurious results in meta-analysis.
• BRMA provides a valuable tool for researchers dealing with complex heterogeneous literature.