Estimating effective sample sizes using standard errors from appropriate analyses of cluster-randomized trials

BACKGROUND AND OBJECTIVES: Cluster-randomized trials (CRTs) typically provide less information than similarly sized trials with individual randomization due to correlation within clusters, ie, the effective sample size (ESS) is typically lower than the total number of participants. Ignoring clustering in analyses can lead to overly precise results (too small P values and too narrow confidence intervals) and bias. ESSs from CRTs can be valuable for interpretation or meta-analysis. ESSs can be estimated using intracluster correlation coefficients (ICCs), but these are often not known and must be assumed. METHODS: Two approaches for estimating ESSs from CRTs using standard errors from appropriate analyses taking clustering into account, but not requiring ICCs, are presented. The first approach scales event and participant counts by the ratio between the variances from appropriate analyses and simple analyses not taking clustering into account. The second approach uses an optimization procedure to modify event proportions or group means and counts. Code and explanations to implement both approaches are provided. RESULTS: Using data from three illustrative example CRTs, both approaches are compared to not taking clustering into account and results from appropriate analyses. Both approaches avoid inappropriately high precision when group-level summary data are analyzed. The second approach also corrects potential bias in estimates that can be introduced by ignoring clustering. CONCLUSION: These approaches may be used to aid interpretation and evidence synthesis of CRTs and can be used where results from appropriate analyses of CRTs are presented, even if ICCs are not reported.