Power planning when cluster sizes are uncertain
A cluster-randomized trial will compare two implementation strategies using a binary patient outcome. The number of clusters is constrained, cluster sizes will vary, and neither the intracluster correlation nor the coefficient of variation in cluster size is known precisely. For sample-size planning, is it more defensible to use conservative fixed values for both quantities, a joint range of plausible scenarios, or a probability distribution that reports assurance rather than power at one assumed design effect? Which assumptions should determine the primary calculation?
