I would expect -2difMAD to be reasonable for a
single volume, even one computed as scaled
variance.

But ignoring that for a bit...

Both of those sets of numbers look a little off
to me, thought that depends on how many and what
types of interpolations were applied to get the
data registered and in standard space.

When extra interpolations are not applied (i.e.
all registration and the warp to standard space
are applied as a single transformation, to avoid
multiple interpolations), then the estimated
smoothness tends to be around the applied one,
plus 1 voxel. So if your original voxels were
3 mm^3, the estimated smoothness will often be
around 11 mm, give or take. If multiple
transformations and interpolations were applied
to the data, the 13-14 mm estimates might be
accurate.

But something is fishy, because using -2difMAD
has been reasonable in my limited experience,
and should certainly not return values below
your applied blur.


To be sure, did you apply each subject's brain
mask when estimating their smoothness?

- rick