Estimating an adjusted risk difference in a cluster randomized trial with individual-level analyses

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Epidemiology

Background:

• Cluster randomized trials (CRTs) often report odds ratios for binary outcomes.
• The CONSORT statement recommends reporting both relative and absolute intervention effects.
• Estimating absolute intervention effects like risk difference (RD) in CRTs requires careful methodological consideration.

Purpose of the Study:

• To evaluate methods for estimating risk difference (RD) in cluster randomized trials (CRTs).
• To compare conditional (GLMM) and marginal (GEE) approaches using Gaussian, binomial, and Poisson distributions.
• To assess bias, standard error estimation, type I error, and coverage rates.

Main Methods:

• A simulation study was conducted within the framework of CRTs.
• Methods included generalized linear mixed models (GLMM) and generalized estimating equations (GEE).
• Distributions considered were Gaussian, binomial, and Poisson, with g-computation for binomial/Poisson RD estimation.

Main Results:

• All methods demonstrated no bias in risk difference estimation.
• GEE approach experienced convergence issues under specific conditions (low ICC, few clusters, small cluster size, many covariates, low prevalence).
• Gaussian distribution (both approaches) and GEE (binomial/Poisson) showed satisfactory standard error estimation; GEE outperformed GLMM in type I error and coverage.

Conclusions:

• The Gaussian distribution is recommended for its simplicity in estimating RD in CRTs.
• The GEE approach is generally preferred over GLMM due to better performance in type I error and coverage.
• GLMM may be used when GEE encounters convergence problems.