Sensitivity analysis for publication bias in meta-analysis of sparse data based on exact likelihood

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Synthesis

Background:

• Meta-analysis synthesizes multiple study findings, often using normal-normal random-effects models.
• Sparse data in meta-analysis (low event rates) challenges these models due to inaccurate normal approximations.
• Publication bias is a significant threat to meta-analysis validity.

Purpose of the Study:

• To develop a robust sensitivity analysis method for publication bias in meta-analyses with sparse data.
• To improve the accuracy and stability of inferences in such meta-analyses.
• To extend existing likelihood-based methods to generalized linear mixed-effects models.

Main Methods:

• Replaced the approximate normal within-study model with an exact generalized linear mixed model.
• Extended Copas's t-statistic selection function for likelihood-based sensitivity analysis.
• Applied the proposed method to real-world meta-analyses and simulation studies.

Main Results:

• The proposed method demonstrated superior performance compared to the standard normal-normal model's sensitivity analysis.
• The generalized linear mixed-effects model approach effectively reduced bias from data sparsity.
• The method proved accurate and stable in inference for sparse data meta-analyses.

Conclusions:

• The novel sensitivity analysis method effectively addresses publication bias in sparse data meta-analyses.
• Generalized linear mixed-effects models offer an improved approach over traditional methods for sparse data.
• This method provides valuable guidance for conducting reliable meta-analyses with limited event data.