Including random centre effects in design, analysis and presentation of multi-centre trials

Area of Science:

• Clinical Trials Methodology
• Biostatistics
• Health Services Research

Background:

• Uncertainty exists regarding the necessity of adjusting for centre variation in large, diverse multicentre trials.
• A critical distinction lies between variations in outcomes (independent of treatment) and variations in treatment effects.
• This study re-analyzes the CRASH-2 trial to clarify the application of multi-level models for binary outcomes in multicentre settings.

Purpose of the Study:

• To clarify when and how to use multi-level models for multicentre studies with binary outcomes.
• To distinguish between variation in outcome and variation in treatment effect in large clinical trials.
• To re-analyze the CRASH-2 trial data to inform statistical design and analysis strategies.

Main Methods:

• Reanalysis of the CRASH-2 trial data, which randomized 20,127 trauma patients across 271 centres and 40 countries.
• Utilized logistic regression models with fixed effects for treatment, patient-level, and centre-level baseline covariates.
• Incorporated random effects to assess variation between countries and between centres within countries in both underlying risk of death and treatment effect.

Main Results:

• Significant variation was observed between countries and centres regarding the 4-week all-cause death outcome.
• No differences were found between countries or centres in the treatment effect of tranexamic acid.
• The average treatment effect remained unchanged after accounting for centre and country variation.

Conclusions:

• Distinguishing between underlying outcome variation and treatment effect variation is crucial; outcome variation is common, but treatment effect variation is not.
• Stratifying randomization by centre can resolve numerous statistical challenges in multicentre trials.
• Including random intercepts in the analysis may enhance statistical power and reduce bias in mean and standard error estimates.