A multilevel model framework for meta-analysis of clinical trials with binary outcomes

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Meta-analysis is crucial for synthesizing evidence from multiple studies.
• Conventional fixed and random effects models have limitations in handling complex data structures.
• Binary outcomes are common in clinical trials, necessitating specialized analytical approaches.

Purpose of the Study:

• To explore the application of multilevel models in meta-analysis for binary outcomes.
• To compare multilevel models with conventional methods for both summary and individual patient data.
• To investigate methods for estimating treatment effects and between-trial variance.

Main Methods:

• Framing conventional fixed and random effects models within a multilevel modeling framework.
• Utilizing maximum likelihood (ML) or restricted maximum likelihood (REML) estimation.
• Employing parametric bootstrap and bias-corrected bootstrap methods for confidence intervals and variance estimation, especially for individual patient data.

Main Results:

• Multilevel models accommodate both summary data (e.g., log-odds ratios) and individual patient data.
• Bootstrap intervals are preferred for summary data meta-analysis due to relaxed assumptions.
• Bias-corrected bootstrap provides unbiased estimation and accurate confidence intervals for individual patient data, particularly for between-trial variance.

Conclusions:

• Multilevel modeling offers a flexible and powerful approach to meta-analysis of binary outcomes.
• The framework allows for modeling trial effects as fixed or random, with options for incorporating covariance.
• This approach facilitates extensions beyond standard meta-analysis techniques, enhancing evidence synthesis.