Estimating the variance for heterogeneity in arm-based network meta-analysis.

Area of Science:

• Biostatistics
• Medical Informatics
• Clinical Epidemiology

Background:

• Network meta-analysis (NMA) is crucial for synthesizing evidence from multiple studies.
• Arm-based models are a common approach in NMA, often fitted using generalized linear mixed models.
• Full maximum likelihood (ML) estimation in these models can produce biased variance estimates for trial-by-treatment interactions, impacting heterogeneity assessment.

Purpose of the Study:

• To investigate alternative variance estimation methods for arm-based network meta-analysis models.
• To reduce bias in heterogeneity variance estimates compared to full maximum likelihood estimation.
• To evaluate the performance of penalized quasi-likelihood/pseudo-likelihood and hierarchical likelihood approaches.

Main Methods:

• Fitting arm-based network meta-analysis models using generalized linear mixed model procedures.
• Implementing penalized quasi-likelihood/pseudo-likelihood and hierarchical (h) likelihood for variance estimation.
• Comparing proposed methods with full ML via simulation studies and real data analysis.
• Exploring model modifications including sum-to-zero restrictions and baseline contrasts.

Main Results:

• Penalized quasi-likelihood/pseudo-likelihood and h-likelihood methods demonstrated reduced bias in variance estimation.
• These methods achieved satisfactory coverage rates in simulation studies.
• Alternative ML-based adjustments (sum-to-zero, baseline contrasts, residual ML-like) also reduced bias but had less optimal coverage.

Conclusions:

• Penalized quasi-likelihood/pseudo-likelihood and h-likelihood are recommended for variance estimation in arm-based network meta-analysis.
• These methods offer improved accuracy and reliability for assessing heterogeneity.
• The findings contribute to more robust statistical inference in network meta-analysis.