Bayesian meta-analysis: The role of the between-sample heterogeneity

Area of Science:

• Biostatistics
• Medical Statistics
• Clinical Trials

Background:

• The random effects model in meta-analysis assumes complete heterogeneity of treatment effects across studies.
• This assumption may not hold, as studies might exhibit clustering or partial heterogeneity.
• Assessing homogeneity before pooling is often inconsistent.

Purpose of the Study:

• To evaluate the performance of the fully heterogeneous random effects model when true heterogeneity is absent.
• To demonstrate the benefits of identifying and incorporating study clustering into meta-analysis.
• To propose a Bayesian approach for model selection and averaging in meta-analysis.

Main Methods:

• Simulated binary experiments were used to compare meta-inference under different heterogeneity models.
• A Bayesian model selection procedure was developed to estimate the true cluster model.
• Bayesian model averaging was employed to integrate clustering estimation into the meta-analysis.

Main Results:

• The fully heterogeneous model performed poorly when heterogeneity was not complete.
• Knowledge of the true cluster model substantially improved meta-analytic inference.
• The proposed Bayesian methods were applied to a real-world meta-analysis of aspirin trials.

Conclusions:

• The standard random effects model may be inadequate when treatment effect heterogeneity is not fully present.
• Identifying and accounting for study clustering is crucial for accurate meta-analysis.
• Bayesian model selection and averaging offer a robust framework for meta-analysis with potential clustering.