Implementing informative priors for heterogeneity in meta-analysis using meta-regression and pseudo data

Area of Science:

• Statistics
• Biostatistics
• Meta-analysis

Background:

• Standard meta-analysis methods struggle with imprecisely estimated between-study variance when few studies are included.
• Bayesian meta-analysis offers a way to incorporate external evidence on heterogeneity for more reliable effect size inference.

Purpose of the Study:

• To present a novel Bayesian meta-analysis method using data augmentation and meta-regression.
• To derive predictive inverse-gamma distributions for between-study variance to serve as priors.
• To compare the performance of this Bayesian method against frequentist and fully Bayesian approaches.

Main Methods:

• Data augmentation represents informative priors for between-study variance as pseudo data.
• Meta-regression is employed for the estimation of model parameters.
• Comparison with importance sampling, Markov chain Monte Carlo (MCMC), and DerSimonian and Laird procedures.

Main Results:

• Bayesian meta-analysis with data augmentation provides results comparable to MCMC and importance sampling.
• The proposed method, especially with restricted maximum likelihood estimation, yields results closer to MCMC than frequentist methods.
• The approach is implementable in standard statistical software, offering a less complex alternative.

Conclusions:

• The presented Bayesian meta-analysis method using data augmentation is accessible to applied researchers.
• This approach improves the robustness of meta-analysis, particularly when dealing with a small number of studies.
• The method is applicable to real datasets and can be extended to network meta-analysis.