Hi all, here is a stats question. We are doing the following steps to perform and then threshold an MVPA:
-run classifier in subject space
-bring the accuracy maps for each subject into stereotactic space (using @auto_tlrc for anat, then align_epi_anat.py)
-performing a parametric group analysis (e.g., t-test against chance)
-thresholding the resulting significance map via FDR
HOWEVER, I am running into the following problem. When I warp the accuracy maps into tal, I end up with some "zeros" on the outside of the brain. This results in some very significant p-values, as it is thinking that the classifier was "0% accuracy" (i.e. significantly wrong). Thus, when I do my FDR, all of those significant p-values drive my threshold to be way more conservative than it ought to be.
What is the best way to fix this "problem"? Here are some of my ideas:
1. only include voxels inside the brain in the FDR, using a tlrc mask
2. only include voxels inside the brain in the FDR, but adjust the q to be .01 (or some other value?) so that it's more conservative
3. manually make the p-values outside the brain 1 (or maybe .5?) and include everything in the FDR
4. something else??
Thanks,
Nicole