I noticed a quirk and wasn’t sure what to think about it. Using the haxby data set to categorize faces vs houses with a svm. I tried using both the raw data and an ICA dimension reduction and it looks like the null shifts to the left when I look at the permutation test. I’d think the theoretical null should still be centered at .50 as long as there are 2 categories and equally sized groups. Any insight would be helpful.
My mistake this doesn’t seem to happen under the PCA transform, so maybe I’m not meeting one of the assumptions for ICA, but I’m still a little unsure how this would affect the permutation test.