> So then I would be running a t-test on values following a t-distribution...
> which is about fine with large n because then the t-distribution is similar
> to a normal distribution, isn't it?
Yes, that should be fine in asymptotic sense.
> To be clear, each map is not scaled by *its* sos (or sqrt(sos)). Rather the sos
> is computed over the maps (i.e. over subjects), for each voxel separately.
Yes, that's what I understood.
> But if I understand you correctly that approach is fine, no?
Well, not so obvious to me. The voxel-wise scaling by sqrt(SOS) is definitely different between your 'within-subject' vs 'between-subject' maps, and that means the scaling would change the pairwise comparison as you've already seen. Since I have no grasp regarding how to interpret the scaled comparison, I will leave the interpretation issue to you.
Gang