A Bias-Corrected Bayesian Nonparametric Model for Combining Studies With Varying Quality in Meta-Analysis

Area of Science:

• Statistics
• Biostatistics
• Meta-analysis

Background:

• Bayesian nonparametric (BNP) methods enhance meta-analysis by relaxing distributional assumptions and managing random effects heterogeneity.
• Existing BNP models can account for clustering and multimodality but struggle with internal validity biases from varying study quality.
• Internal validity biases, including reporting bias and selection bias, can compromise the integrity of meta-analysis results.

Purpose of the Study:

• To introduce a novel bias-corrected Bayesian nonparametric (BC-BNP) meta-analysis model.
• To automatically correct for internal validity biases using only reported effect sizes and standard errors.
• To relax parametric assumptions on bias distributions and improve meta-analysis robustness.

Main Methods:

• Developed the BC-BNP model, a mixture of a parametric random effects distribution and a BNP model for bias.
• Evaluated the BC-BNP model using simulated datasets.
• Applied the BC-BNP model to two real-world case studies.

Main Results:

• The BC-BNP model effectively detects bias when present and aligns with standard models when bias is absent.
• Relaxing parametric assumptions for bias distribution yields consistent results with prior models (Verde et al.).
• BNP bias modeling can cluster studies with similar biases, offering deeper insights into heterogeneity.

Conclusions:

• The BC-BNP model offers a robust approach to meta-analysis by addressing internal validity biases.
• The model provides accurate results comparable to simpler models when bias is minimal.
• Implementation in the R package 'jarbes' facilitates practical application of the BC-BNP model.