CO1.2 - Estimation of a risk difference in a cluster randomized trial

Cluster randomized trials (cRT) are randomized trials in which randomization units are clusters of individuals, rather than individuals themselves. They are usually analysed using a conditional approach (with generalized linear mixed model - GLMM) or a marginal one (using generalized estimating equation - GEE). When the outcome is binary a logit link function is classically used. This leads to express the results as a relative effect, with an odds ratio. Relative effects aren't clearly understandable and lead to an over-optimistic appraisal of the results. The CONsolited Standards of Reporting Trials (CONSORT) statement recommends reporting both relative and absolute effect, which means, for binary data, to report a risk difference (RD). Presently, we lack guidelines regarding the best way to estimate a risk difference in a cluster randomized trial. The objective is to assess the statistical properties of different methods used to estimate an adjusted risk difference from clustered data, considering a two-parallel group cRT. We conducted a simulation study. We generated binary outcome and considered a two-parallel group cRT with multiple covariates at both individual and cluster levels. Individual level covariates were generated as confounding factors. We used GEE to estimate the intervention effect. We considered a Gaussian (G) distribution with an identity link function, thus estimating directly a RD. We also considered Binomial (B) and Poisson (P) distributions with associated logit and log link functions to estimate relative intervention effects, and then used the g-computation method to estimate a RD. We considered an exchangeable correlation matrix. Results show that whatever the method used, we did not observe any bias. Coverage rates are lower than 95% when a small number of clusters are considered. They increase when the number of cluster increases but remains lower than 95%. The empirical standard error (EmpSE) is minimized by the method using the Gaussian distribution. The three methods have negligible bias and similar coverage rates. All of them seem to be appropriate to estimate a RD. Method using the identity link with a Gaussian distribution seems to be preferable because it is easier to implement and it minimizes the EmpSE. Risk difference , Cluster randomized trial , Binary outcome , Generalized estimating equations Les auteurs n'ont pas précisé leurs éventuels liens d'intérêts.