A Bayesian sensitivity study of risk difference in the meta-analysis of binary outcomes from sparse data

Area of Science:

• Biostatistics
• Statistical Modeling
• Epidemiology

Background:

• Statistical meta-analysis commonly uses the normal random effect model for discrete data.
• Normal approximation may not accurately represent uncertainty in discrete data, particularly with rare events.
• Existing methods can yield poor results for sparse binary outcomes.

Purpose of the Study:

• To propose a Bayesian meta-analysis for binary outcomes with sparse data.
• To address limitations of normal approximation in meta-analysis.
• To introduce a method for assessing sensitivity to structural dependencies.

Main Methods:

• Development of a Bayesian meta-analysis procedure for sparse binary data.
• Application of suitable linking distributions for Bayesian robustness analysis.
• Sensitivity analysis of meta-analysis results concerning selected structure dependence.

Main Results:

• The proposed Bayesian procedure can be directly applied to sparse data without ad hoc corrections.
• Bayesian robustness studies are easily implemented by selecting appropriate linking distributions.
• The method was illustrated with a real-data example of binomial sparse data.

Conclusions:

• Bayesian meta-analysis offers a robust alternative for binary outcomes with sparse data.
• The approach improves upon normal approximation, especially in scenarios with rare events.
• The proposed method facilitates sensitivity analyses and direct application to real-world sparse data.