Use of generalized linear mixed models for network meta-analysis

Area of Science:

• Biostatistics
• Statistical Modeling
• Evidence Synthesis

Background:

• Traditional pairwise meta-analysis has limitations in synthesizing evidence from multiple treatments.
• Network meta-analysis (NMA) has emerged as a powerful statistical method to address these limitations.
• Bayesian NMA, specifically the Lu and Ades model, offers a flexible framework for complex data structures.

Purpose of the Study:

• To demonstrate the implementation of the Lu and Ades Bayesian network meta-analysis model within a frequentist generalized linear mixed model (GLMM) framework.
• To illustrate correct estimation of random effects through covariate centering in GLMMs for NMA.
• To show the flexibility of GLMMs for incorporating covariates (moderators, confounders) for meta-regression in multi-treatment comparisons.

Main Methods:

• Implementation of the Lu and Ades (mixed treatments comparisons) model using frequentist generalized linear mixed models.
• Demonstration using two examples focusing on covariate centering for accurate random effects estimation.
• Exploration of dummy coding and contrast basic parameter coding schemes for consistent results.

Main Results:

• Centering covariates for random effects estimation within trials yields correct random effects estimation in the GLMM framework.
• Both dummy coding and contrast basic parameter coding schemes produce identical results when random effects are correctly specified.
• The GLMM approach facilitates straightforward incorporation of covariates for meta-regression in multi-treatment comparisons.

Conclusions:

• The frequentist generalized linear mixed model provides a viable and flexible approach for implementing Bayesian network meta-analysis models like Lu and Ades.
• Correct specification of random effects estimation, particularly through covariate centering, is crucial for accurate results.
• This GLMM-based NMA method is easily extendable to various outcome types (continuous, counts, multinomial) and facilitates meta-regression.