Substantial risks associated with few clusters in cluster randomized and stepped wedge designs

Given the growing attention to quality improvement, comparative effectiveness research, and pragmatic trials embedded within learning health systems, the use of the cluster randomization design is bound to increase. The number of clusters available for randomization is often limited in such trials. Designs that incorporate pre-intervention measurements (e.g. cluster cross-over, repeated parallel arm, and stepped wedge designs) can substantially reduce the required numbers of clusters by decreasing between-cluster sources of variation. However, there are substantial risks associated with few clusters, including increased probability of chance imbalances and type I and type II error, limited perceived or actual generalizability, and fewer options for statistical analysis. Furthermore, current sample size methods for the stepped wedge design make a strong underlying assumption with respect to the correlation structure-in particular, that the intracluster and inter-period correlations are equal. This is in contrast with methods for the cluster cross-over design that explicitly allow for a smaller inter-period correlation. Failing to similarly allow for the inter-period correlation in the design of a stepped wedge trial may yield perilously low sample sizes. Further methodological and empirical work is required to inform sample size methods and guidance for the stepped wedge trial and to provide minimum thresholds for this design.