A general framework for the use of logistic regression models in meta-analysis

Area of Science:

• Biostatistics
• Medical Informatics
• Epidemiology

Background:

• Traditional meta-analysis methods often assume normal distribution of effect estimates, which may not always hold true.
• Existing methods for different meta-analytic scenarios (e.g., meta-regression, network meta-analysis, diagnostic test accuracy) can be complex and varied.
• Individual participant data (IPD) availability enhances meta-analysis but is not always feasible.

Purpose of the Study:

• To introduce and evaluate a one-stage random-effects logistic regression model for meta-analysis.
• To demonstrate the model's applicability across diverse meta-analytic settings, including those with summary data.
• To compare the proposed model with established meta-analysis techniques.

Main Methods:

• Utilized a one-stage random-effects logistic regression model.
• Applied the model to dichotomous event outcomes in meta-analysis, accommodating both individual participant data and summary contingency tables.
• Compared the model's performance against Bayesian network meta-analysis and bivariate/hierarchical summary ROC models for diagnostic test accuracy meta-analyses.

Main Results:

• The one-stage model maximizes the binomial likelihood, avoiding the assumption of normally distributed effect estimates.
• The model demonstrates versatility, applicable to meta-regression, network meta-analyses, and diagnostic test accuracy studies with minor modifications.
• Provides a potentially unifying framework for various meta-analytic approaches.

Conclusions:

• The one-stage random-effects logistic regression model offers a robust and flexible alternative for meta-analysis.
• This unified approach simplifies analysis across different meta-analytic scenarios, enhancing efficiency and potentially improving accuracy.
• The model's ability to handle various data structures and its avoidance of normality assumptions make it a valuable tool in evidence synthesis.