Using information-theoretic approaches for model selection in meta-analysis.

Area of Science:

• Biostatistics
• Meta-analysis
• Statistical modeling

Background:

• Meta-regression analyzes associations between effect sizes and study characteristics (moderators).
• Traditional model selection in meta-regression relies on hypothesis testing.
• Identifying the best approximating model for the data-generating mechanism is crucial.

Purpose of the Study:

• To explore alternative model selection methods in meta-regression based on information criteria and multimodel inference.
• To compare the performance of these alternative methods against conventional approaches.
• To evaluate the utility of information-theoretic approaches with both maximum likelihood (ML) and restricted maximum likelihood (REML) estimation.

Main Methods:

• Described information-theoretic approaches: information criteria, multimodel inference, and relative variable importance.
• Applied these methods to an illustrative meta-regression example.
• Conducted a simulation study to compare model selection performance across various conditions.

Main Results:

• Information-theoretic approaches often outperformed conventional testing-based methods in identifying the true model.
• These alternative methods were frequently among the best performers across simulated conditions.
• Performance with REML estimation was comparable or superior to ML estimation, depending on REML likelihood computation.

Conclusions:

• Information-theoretic approaches offer a superior alternative for model selection in meta-regression.
• These methods demonstrate robustness and effectiveness with both ML and REML estimation.
• Wider adoption of these advanced model selection techniques in meta-regression is recommended.