Meta-CART: A tool to identify interactions between moderators in meta-analysis.

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Univariate moderator analysis is standard in meta-analysis, limiting the detection of complex interactions between study characteristics.
• Standard meta-regression struggles to identify interactions among multiple moderators influencing treatment effects.

Purpose of the Study:

• To propose and evaluate extensions to meta-CART for improved moderator analysis in meta-analysis.
• To address limitations of previous meta-CART by incorporating study sample sizes and avoiding effect size dichotomization.

Main Methods:

• Developed extended meta-CART incorporating study accuracy weights and regression trees to avoid dichotomization.
• Introduced new pruning rules for classification and regression trees (CART).
• Evaluated performance of meta-CART versions using Monte Carlo simulations.

Main Results:

• Meta-regression trees weighted by random-effects and employing a 0.5-standard-error pruning rule demonstrated superior performance.
• The required sample size for effective meta-CART depends on the number of study characteristics, interaction magnitude, and residual heterogeneity.

Conclusions:

• The enhanced meta-CART approach offers a more robust method for identifying interactions in moderator analysis.
• The findings provide guidance on the optimal implementation of meta-CART for reliable meta-analysis results.