Performance of location-scale models in meta-analysis: A simulation study.

Area of Science:

• Meta-analysis
• Statistical modeling
• Heterogeneity analysis

Background:

• Location-scale models in meta-analysis enable simultaneous examination of moderator effects on both the mean (location) and variance (scale) of true effect distributions.
• The complexity of these models presents challenges in fitting and lacks systematic examination of estimation and inference methods in meta-analysis.

Purpose of the Study:

• To compare different estimation methods, significance tests, and confidence interval construction methods for location-scale models in meta-analysis.
• To evaluate the statistical properties of these methods in the context of meta-analytic heterogeneity.

Main Methods:

• A Monte Carlo simulation study was conducted.
• Compared maximum likelihood and restricted maximum likelihood estimation.
• Evaluated Wald-type, permutation, and likelihood-ratio tests for significance.
• Assessed Wald-type and profile-likelihood confidence intervals for scale coefficients.

Main Results:

• Restricted maximum likelihood estimation yielded rejection rates closer to nominal levels and narrower confidence intervals.
• Permutation tests showed type I error rates closest to the nominal level; likelihood-ratio tests had the highest statistical power.
• Profile-likelihood intervals had lower coverage probabilities than Wald-type but were closer to the nominal 95% level.
• Dichotomous moderators resulted in slightly higher rejection rates and coverage probabilities than continuous moderators.

Conclusions:

• Location-scale models are valuable tools for modeling heterogeneity in meta-analysis.
• Despite potential challenges like parameter space constraints and non-convergence, these models offer a robust approach.
• Restricted maximum likelihood estimation and permutation tests show promising statistical properties for scale coefficient analysis.