Estimating relative risks in multicenter studies with a small number of centers - which methods to use? A simulation...

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Multicenter study analyses require accounting for center clustering for valid inference.
• Adjusting for center effects in binary outcomes is challenging with small sample sizes or few events per center.
• Evaluating statistical models for center effect adjustment in challenging scenarios is crucial.

Purpose of the Study:

• To evaluate the performance of various statistical models for center effect adjustment in multicenter studies with binary outcomes.
• To compare generalized estimating equation (GEE) log-binomial and Poisson models, generalized linear mixed models (GLMMs), and a Bayesian binomial GLMM.
• To assess model performance under conditions of few centers, small sample sizes, or few events per center.

Main Methods:

• Conducted a simulation study with ≤30 centers and ≤50 subjects per center.
• Employed both randomized controlled trial and observational study designs to estimate relative risk.
• Compared GEE and GLMM models against a non-adjusted log-binomial model using bias, RMSE, and coverage metrics; used neutral priors for Bayesian GLMM.

Main Results:

• Frequentist methods showed little bias; binomial GLMM had poor convergence (27%-85%) but performed well otherwise.
• GEE models required small sample corrections for robust standard errors and proper 95% confidence interval coverage.
• Bayesian GLMM demonstrated the smallest RMSE and good coverage across scenarios, despite slightly more bias in smallest samples.

Conclusions:

• Recommend adjustment for center in multicenter studies with binary outcomes and few centers.
• Suggest using GEE log-binomial or Poisson models with small sample corrections.
• Alternatively, recommend a Bayesian binomial GLMM with informative priors for robust analysis.