|
Activity Number:
|
469
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 9, 2006 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Biopharmaceutical Section
|
| Abstract - #307098 |
|
Title:
|
To Permute or Not Permute
|
|
Author(s):
|
Haiyan Xu*+ and Jason Hsu and Yifan Huang and Violeta Calian
|
|
Companies:
|
Johnson & Johnson Pharmaceutical R&D and The Ohio State University and H. Lee Moffitt Cancer Center & Research Institute and University of Iceland
|
|
Address:
|
2121 Town Court, N., Lawrenceville, 08648,
|
|
Keywords:
|
permutation test ; type I error rate ; multiple testing
|
|
Abstract:
|
Permutation test is a popular technique for testing a hypothesis of no effect, when the underlying distributions are unknown. To test for difference between two populations, a permutation test might be based on the difference of the sample means in the univariate case, and the maximum of such test statistics in the multivariate case. The null distribution is then estimated by permuting the observations between the two samples. We show that if the purpose is to test for equality of means, then a permutation test can have inflated Type I error rate unless the distributions are identical. We also show that if the purpose is to test for identical marginal distributions, then a permutation test can have inflated Type I error rate unless the joint distributions are identical. Implications of these results in multiple testing of microarray data will be discussed.
|
