Covariate adjustment in cluster randomised trials: a practical guide.

Abstract

Covariate adjustment can offer several potential benefits in the analysis of cluster randomised trials. These benefits include increasing statistical precision (ie, narrowing width of confidence intervals), as well as potentially reducing any bias arising from differential identification and recruitment across arms or missing outcome data. This article outlines a guideline for how to choose covariates to include in a prespecified adjustment plan for such trials. Recommendations include adjusting for covariates that have been included in any restricted randomisation; and adjusting for a prespecified set of covariates thought to be prognostic of the outcome, differential recruitment, or outcome missingness. When the prevalence of missing covariate or outcome data are non-negligible, a missing data technique such as multiple imputation (allowing for clustering), cluster mean imputation, or the missing indicator method, is recommended. In a case study, the proposed prespecified analysis plan includes adjustment for minimisation variables as well as four covariates thought to be prognostic of the outcome and potentially related to unblinded identification of participants after randomisation.