hi Rick,

pardon me for coming back on the same issue, but I am having second thoughts here..

1) Shouldn't the null hypothesis dataset indeed depend on the statistical model employed? We label 'noise' everything that our model can't explain, and in this perspective the smoothness of the residuals *should* vary with the statistical model employed (in fact, in afni_proc.py the smoothness is indeed estimated on the basis of the chosen statistical model, so change the model and the estimated smoothness will change). Now, why would this line of reasoning apply to individual analyses but not to group analyses?

2) Having a very good model does not necessarily imply that the residual image will be very smooth, since the model fit is done independently for each voxel while the smoothness is estimated spatially. Even if the model fits the data very well, it won't reasonably do so across the whole brain and therefore there will be patches of non-activated areas that can be quite spatially uncorrelated.

am i still not getting it??

thanks again for your patience

g.