Random-effects meta-analysis models for pooling rare events data: a comparison between frequentist and bayesian methods.

Background: Standard random-effects meta-analysis models for rare events exhibit significant limitations, particularly when synthesizing studies with double-zero events. While methodological advances in both frequentist and Bayesian frameworks now offer robust alternatives that bypass continuity corrections, the comparative performance of these approaches-especially between Bayesian and frequentist paradigms-remains understudied.

Methods: This study evaluates the performance of ten widely used meta-analysis models for binary outcomes, using the odds ratio as the effect measure. The evaluated models comprise seven frequentist and three Bayesian approaches. Simulations systematically varied key parameters, including control event rates, treatment effects, study numbers, and heterogeneity levels, to compare model performance across four metrics: percentage bias, 95% confidence/credible interval width, root mean square error, and coverage. The methods were further illustrated through applications to two published rare events meta-analyses.

Results: The results show that the beta-binomial model proposed by Kuss generally performed well, while the generalised estimating equations did not. In cases where heterogeneity is not large, all models tended to have a good performance except for the generalised estimating equations. When the heterogeneity is large, none of the compared models produced good performance. The Bayesian model incorporating the Beta-Hyperprior proposed by Hong et al. performed well, followed by the binomial-normal hierarchical model proposed by Bhaumik.

Conclusions: In summary, the beta-binomial model proposed by Kuss is recommended for rare events meta-analyses, and the Bayesian model is a promising method for pooling rare events data.