Likelihood-based random-effects meta-analysis with few studies: empirical and simulation studies.

Area of Science:

• Biostatistics
• Medical Research Methodology

Background:

• Standard random-effects meta-analysis methods demonstrate poor performance with small sample sizes, a common issue in practice.
• The effectiveness of advanced modeling techniques in improving small-sample behavior remains unclear.

Purpose of the Study:

• To compare the performance of various meta-analytical methods in small-sample settings.
• To evaluate the impact of different modeling approaches and confidence interval adjustments on estimation accuracy and precision.

Main Methods:

• Evaluated likelihood-based methods, DerSimonian-Laird, Empirical Bayes, adjustment methods, and a fully Bayesian approach.
• Utilized normal approximation and Student-t distribution for confidence intervals.
• Incorporated linear mixed models and generalized linear mixed models (GLMMs) for binary meta-analyses.
• Empirically analyzed 40 meta-analyses from IQWiG reviews and conducted a simulation study.

Main Results:

• Frequentist methods with normal approximation showed inadequate coverage probabilities, particularly with unbalanced study sizes and heterogeneity.
• Bayesian methods generally provided better coverage than frequentist approaches, except in cases of extreme heterogeneity.
• Credible intervals from Bayesian methods were wider than unadjusted frequentist intervals but narrower than adjusted ones.
• GLMM-based confidence intervals were generally narrower than other frequentist methods; some methods were impractical due to numerical issues.

Conclusions:

• Caution is advised when applying meta-analytical methods to few studies, especially with heterogeneity and unbalanced sizes, due to compromised coverage or wide intervals.
• Bayesian estimation, using appropriate priors for heterogeneity, presents a viable compromise for improving meta-analysis results in small-sample scenarios.