365 – Topics in Complex Survey Data Analysis
Comparison of Permutation Tests and GEE Methods for Group-Randomized Trials with Count Data
Ping Xu
University of South Florida
Brian Leroux
University of Washington
Group-randomized trials (GRT) often have a small number of clusters, creating challenges for identifying a valid analysis method with sufficient power. Previous studies found problems with two most popular methods- GEE and permutation tests. GEE suffers from anti-conservative tests even with small-sample corrections, while permutations suffer from conservativeness due to discreteness of test distribution. We compared the performance of small-sample adjusted GEE methods with three weighted permutation procedures for GRT with count data. TypeI error rates and power were estimated by simulation in 81 scenarios with number of clusters from 10 to 40, ranges of values for overall mean count, covariate effect, overdispersion parameter, and intracluster correlation coefficient. Permutation tests had valid typeI errors in all scenarios and had power at least as high as other tests. Average power to detect moderate treatment effects was 0.794 for weighted permutation test versus 0.787 for the best-performing GEE method. For large treatment effects, power was 0.905 versus 0.901. In conclusion, permutation tests showed a wider range of validity with no loss of power, compared to GEE methods.