Accounting for centre-effects in multicentre trials with a binary outcome - when, why, and how?

Area of Science:

• Biostatistics
• Clinical Trials

Background:

• Accounting for centre-effects is crucial in multicentre randomised trials.
• Optimal analysis methods for binary outcomes in such trials remain unclear.

Purpose of the Study:

• To compare the performance of four analysis methods: fixed-effects, random-effects, generalised estimating equations (GEE), and Mantel-Haenszel.
• To identify the best analysis methods for multicentre trials with binary outcomes.

Main Methods:

• Re-analysis of the MIST2 randomised trial.
• A large-scale simulation study evaluating analysis method performance.

Main Results:

• Fixed-effects and Mantel-Haenszel methods dropped significant patient numbers due to over-stratification.
• Random-effects and GEE models included all patients, but GEE had convergence issues.
• Simulation showed random-effects and GEE (non-robust SEs) performed well across various centre numbers, offering nominal type I error rates and good power.
• Fixed-effects models yielded biased estimates with many centres, while Mantel-Haenszel lost power.

Conclusions:

• For trials with few centres, fixed-effects, random-effects, or GEE (non-robust SEs) are recommended.
• For moderate to large numbers of centres, random-effects or GEE (non-robust SEs) are the preferred methods.