The performance of odds ratio estimation under different scenarios in Bayesian meta-analysis: A simulation study

Area of Science:

• Biostatistics
• Statistical Modeling

Background:

• Frequentist meta-analysis has limitations in small-sample or rare-event scenarios.
• Bayesian methods offer a flexible framework for meta-analysis.

Purpose of the Study:

• To evaluate Bayesian meta-analysis methods for odds ratio (OR) estimation.
• To assess the impact of heterogeneity and prior distributions on OR estimation accuracy.
• To explore interactions between study characteristics and estimation performance using CHAID analysis.

Main Methods:

• Implemented a Bayesian framework with four heterogeneity priors: half-normal, exponential, half-Cauchy, and inverse-gamma.
• Conducted simulation studies across 1,152 scenarios varying number of studies, event rarity, randomization ratios, and baseline risks.
• Utilized Chi-square Automatic Interaction Detection (CHAID) analysis to identify key factors influencing model performance.

Main Results:

• Prior specification and number of studies in meta-analysis (NSMA) significantly impact estimation accuracy, particularly for rare events.
• CHAID analysis identified NSMA as the most crucial factor for estimation reliability.
• Event type and randomization ratio also showed notable influence under specific conditions.

Conclusions:

• Prior selection is critical in Bayesian meta-analysis for accurate odds ratio estimation.
• CHAID analysis is a valuable tool for understanding complex interactions and improving interpretability in meta-analysis.
• The findings emphasize careful consideration of priors and study characteristics for reliable Bayesian meta-analyses.