Patrick Burns wrote:
>And as Tim rightly pointed out yesterday, you do want to think
>about whether or not the permutation has implications that may
>not be warranted. But again if I understand your problem, I
>think you should be okay.
My response went to Stahel and Burns, not to S-news.
Here is the relevant part of that response:
A permutation test here would require assumptions that you may not
wish to make. You could test
H0: two populations are the same (correlation, mean, var, ...)
vs
Ha: one has higher correlation
(where observations 1:n are from the first population, and (n+1):(2n)
from the second population), then you could store all observations
in a data frame, with one variable being a dummy indicating which class,
then permute the dummy variable.
If you don't want to assume that the two populations are the same
under H0, then you could use bootstrap tilting. Do bootstrap sampling
stratified on the two groups, with statistic being the difference in
correlations, e.g.
bootstrap2(data, cor, treatment = the dummy variable)
then use limits.tilt(). If the confidence interval excludes zero,
the hypothesis test is significant.
bootstrap2() and limits.tilt() are in the resample library, at
www.insightful.com/libraries/downloads.
Tim Hesterberg
========================================================
| Tim Hesterberg Research Scientist |
| timh@insightful.com Insightful Corp. |
| (206)802-2319 1700 Westlake Ave. N, Suite 500 |
| (206)802-2500 (fax) Seattle, WA 98109-3044, U.S.A. |
| www.insightful.com/Hesterberg |
========================================================
|