Sensitivity in Sample Size Determination in Cluster Randomized Trials for Count Data

Music City Center 

For a balanced, cross-sectional parallel cluster randomized trial with count outcomes, we examined several methods to determine the number of clusters (N) necessary for a given power of the hypothesis test on the intervention effect. We applied the methods by either estimating parameter inputs using analytic derivations or leveraging empirical data. We compared methods across key parameters from a two-level Poisson generalized linear mixed model and developed a novel technique to evaluate the impact of parameter uncertainty. Using the analytic approach at 80% power, we conducted a simulation-based sensitivity analysis to estimate actual power. For the empirical approach assuming we had available control cluster data, we generated sampling distributions for N then conducted a sensitivity analysis. Power was most sensitive to the anticipated intervention effect. Except for a few cases, methods were equally sufficient. Given similar power between the two approaches, the empirical approach is sufficient, but the analytic approach is recommended as control cluster data are unneeded, sampling variability is not a concern, and implementation is simpler via straightforward formulae.

count outcome

cluster randomized trial

parameter uncertainty

sample size

sampling variability 

Biopharmaceutical Section