One-stage individual participant data meta-analysis models for continuous and binary outcomes: Comparison of...

Area of Science:

• Biostatistics
• Medical Informatics
• Clinical Trials

Background:

• Individual participant data (IPD) meta-analysis offers a robust method for synthesizing evidence from multiple studies.
• One-stage IPD meta-analysis models integrate data in a single step, accounting for within-study clustering and between-study heterogeneity.

Purpose of the Study:

• To evaluate the performance of restricted maximum likelihood (REML) and maximum likelihood (ML) estimation in one-stage IPD meta-analysis models.
• To assess these methods for synthesizing randomized trials with continuous or binary outcomes.

Main Methods:

• Simulation studies were conducted to assess ML and REML estimation performance in one-stage IPD meta-analysis.
• Models included stratified study intercepts or random study intercepts to handle clustering and random treatment effects for heterogeneity.
• Performance was evaluated for continuous and binary outcomes using different estimation and coding strategies.

Main Results:

• A t-distribution based approach generally improved confidence interval coverage for the summary treatment effect compared to a z-based approach for both ML and REML.
• For ML estimation with a stratified intercept, 'study-specific centering' of the treatment variable reduced bias in the between-study variance estimate.
• REML estimation demonstrated reduced downward bias in between-study variance estimates compared to ML, irrespective of treatment variable coding.

Conclusions:

• The choice of estimation method (REML vs. ML) and distributional approach (t-distribution vs. z-distribution) impacts the accuracy of summary treatment effects in one-stage IPD meta-analysis.
• Specific coding strategies for the treatment variable are important for ML estimation to minimize bias.
• REML offers advantages in reducing bias for variance estimation, though its application with binary outcomes requires careful consideration of pseudo-likelihood stability.