Comparing various Bayesian random-effects models for pooling randomized controlled trials with rare events.

Area of Science:

• Biostatistics
• Medical Research Methodology

Background:

• Meta-analysis of rare events poses significant methodological challenges due to low event counts.
• Bayesian methods are frequently employed for rare events data, offering advantages in incorporating prior information and handling zero events without continuity corrections.

Purpose of the Study:

• To compare the statistical performance of different Bayesian models for pooling rare events data.
• To evaluate the impact of weakly informative priors (WIP) versus non-informative priors (NIP) on Bayesian meta-analysis of rare events.

Main Methods:

• A simulation study compared four parameterizations of the binomial-normal hierarchical model, a beta-binomial model, and generalized linear mixed models.
• The simulation varied treatment effects, sample size ratios, and heterogeneity levels, using odds ratios for rare events.
• Model performance was assessed using bias, root mean square error, interval width, coverage, Type I error, and empirical power.

Main Results:

• Weakly informative priors (WIP) consistently outperformed non-informative priors (NIP) within identical model structures.
• A specific model incorporating the treatment effect parameter in the control arm's risk model demonstrated poor performance.
• Rare events meta-analysis is inherently underpowered, emphasizing the need to report statistical power in empirical studies.

Conclusions:

• The choice of prior distribution significantly impacts the reliability of Bayesian meta-analysis for rare events.
• Certain model structures are less suitable for rare events data, necessitating careful selection.
• Acknowledging and reporting the limited power of rare events meta-analyses is critical for accurate interpretation of results.