Cluster‐Level Analyses to Estimate a Risk Difference in a Cluster Randomized Trial With Confounding Individual‐Level Covariates: A Simulation Study
Résumé
Cluster randomized trials (CRTs) may be analyzed using cluster‐level analyses. For binary outcomes, proportions are estimated for each cluster, and a risk difference can be estimated. The confidence interval is estimated using a Student distribution. However, in doing so, individual‐level characteristics are not adjusted for even though CRTs are known to be prone to recruitment/identification bias, possibly implying individual‐level confounders. With a simulation study, we compared cluster‐level analyses to estimate a risk difference for a two‐arm parallel CRT with individual‐level confounders and cluster‐level covariates. We considered the unadjusted (UN) method, two two‐stage procedure (TSP) methods considering a binomial or a Gaussian distribution, G‐computation (GC), and targeted maximum likelihood estimation (TMLE) methods. As expected, the UN method was biased. TSP methods were also biased for scenarios with a treatment effect when the number of clusters per arm was small. GC and TMLE methods were unbiased. For these latter methods, adjustment on only individual‐level covariates led to better performance measures (type I error rate, coverage rate and relative error of the standard error) than adjustment on both individual‐ and cluster‐level covariates. TSP, GC and TMLE had very similar results except in scenarios with a small number of clusters: Biased results for TSP methods and convergence problems for GC methods. In this case, TMLE should be preferred.