Performance of various tests in analyzing sample data
Method | Estimated difference | 95% CI for difference | p value for test of no difference |
---|---|---|---|
1. Two-sample t test (individual observations) | 0.517 | (0.230, 0.805) | 0.0005 |
2. Wilcoxon (individual observations) | 0.486 | (0.200, 0.769) | 0.0013 |
3. Two-sample t test (means) | 0.476 | (−0.318, 1.269) | 0.2235 |
4. Wilcoxon (means) | 0.425 | (−0.387, 1.286) | 0.2475 |
5. LMM | 0.476 | (−0.322, 1.275) | 0.2260 |
6. GEE | 0.476 | (−0.228, 1.180) | 0.1850 |
7. Rank-sum test for clustered data (Datta and Satten) | 0.2015 |
Table 2 summarizes the results of applying the above methods to the dataset pictured in Fig. 1. We tested the null hypothesis that there was no significant difference between group 1 and group 2 observations. Methods 1 (t test) and 2 (Wilcoxon) were conducted on individual observations, thereby ignoring clustering. Methods 3 and 4 were applied after the data were reduced to the means of each cluster, thereby eliminating clustering. Unlike the t and Wilcoxon tests, LMM, GEE, and the rank-sum test of Datta and Satten (2005) are able to explicitly account for clustering. CI, Confidence interval.