I have a regression design in which the regressor heights differ according to the choices the participants make in each run. In the first level I made a contrast of this regressor (e.g., [1 0 0 0]), in the second level I combined each subject's runs, and in the third level I combined all of my subjects with a whole-brain analysis. Now I want to test if the average BOLD activity in an a priori standard-space region of interest is significantly related to my regressor.

I've seen several instructions/tutorials recommending running featquery at the second level, and I want to make sure I understand why. Is this because featquery will then transform the standard-space ROI into each subject's native space? Does doing so create a more precise or accurate ROI mask?

If I were to run featquery on my third-level, which output(s) could I test for significance? My pe and cope should give the same results because my contrast just selects the regressor, but am I correct in thinking that a one-sample t test of either one should answer my question of whether or not the average activity in the ROI is significantly linearly related to my regressor? And are the tstat or zstat outputs meaningful when I run featquery at the third level? I.e., do they give the average t/z statistic within my ROI from the third-level .gfeat/stats directory, or do they evaluate new statistics based on the ROI instead of the whole-brain analysis and its corrections?

Thank you!