We have recently run into a problem where a user constructed the individual subject beta maps with individual masks, then transformed to Tailairach/MNI space, then did the group analysis on this collection of datasets.

At the periphery of the brain, peculiar artifactual statistics showed up -- because such voxels were zero in many of the subjects but not all. In the limiting case, an edge voxel would be nonzero in only one subject. The statistics at such places are meaningless, of course. By themselves, these edge artifacts aren't a real problem, but when these statistics are converted to p-values and then are included in the computation of the FDR q-value, they bias the results.

We VERY STRONGLY recommend that you use a common mask at the group analysis level. It's simplest to do this using Rick R's afni_proc.py script at the individual subject level (with no mask at all), transform the beta results to Talairach/MNI, and only THEN use the same brain mask to zero out the unwanted voxels in all the subjects' datasets -- this mask being derived from the collection of individual subject masks (by intersection) or directly from the Talairach/MNI reference brain -- whatever floats your boat.