This is an archived version of the Handbook. For the current version, please go to training.cochrane.org/handbook/current or search for this chapter here.

One way to avoid unit-of-analysis errors in cluster-randomized trials is to conduct the analysis at the same level as the allocation, using a summary measurement from each cluster. Then the sample size is the number of clusters and analysis proceeds as if the trial was individually randomized (though the clusters become the individuals). However, this might considerably, and unnecessarily, reduce the power of the study, depending on the number and size of the clusters.

Alternatively, statistical methods now exist that allow analysis at the level of the individual while accounting for the clustering in the data. The ideal information to extract from a cluster-randomized trial is a direct estimate of the required effect measure (for example, an odds ratio with its confidence interval) from an analysis that properly accounts for the cluster design. Such an analysis might be based on a ‘multilevel model’, a ‘variance components analysis’ or may use ‘generalized estimating equations (GEEs)’, among other techniques. Statistical advice is recommended to determine whether the method used is appropriate. Effect estimates and their standard errors from correct analyses of cluster-randomized trials may be meta-analysed using the generic inverse-variance method in RevMan.