|
Abstract:
|
Cluster-randomized trials are often cost-efficient or may be required design for some study questions. However, data analysis and study planning for these studies is complicated by the dependence between observations within each cluster. When the number of observations varies among clusters, Eldridge et al. provide an approximate power calculation. We assessed this approximation via simulations using sample sizes distributed approximately normal, uniform, and bimodal degenerate with equal means and standard deviations. We found that as the kurtosis decreases, the approximation becomes increasingly anticonservative, though it is biased for surprisingly small coefficients of variation. Each row in the table is based on 10,000 simulations and uses 25 clusters in each arm with a delta of 0.208, so that the simulated power would be .8 if the estimate were unbiased.
Mu CV Kurt. Sim. Power (95% CI) Dist
50 .2 3.0 .78, .79 N
50 .2 1.8 .78, .79 U
50 .2 1.0 .77, .78 BMD
50 .5 1.8 .75, .76 U
50 .5 1.0 .75, .76 BMD
The coefficient of variation approximation should be used only with caution.
|
