Hi Gaurav,

One would certainly want to cluster at the group level.

High significance of individual subject results is actually
irrelevant (usually) in the group analysis. All it means is
that a given subject beta that goes into the group analysis
is reliable.

What approach 1 would do is exclude subject voxels from the
group test that are not sufficiently reliable, limiting the
number of samples at each voxel, possibly to a small number
if the individual results do not overlap well. That would
lower the degrees of freedom, which would vary across space.

And note that this would not affect the significance at the
group level at all. It would still be necessary to cluster
or do other corrections for multiple comparisons.

On the flip side, Gang's 3dMEMA approach does take those
individual subject statistics into consideration in a related
way. The more reliable betas might have a stronger influence
on the group result.

But the basic key is that you still have to cluster/correct
at the group level anyway.


Regarding blur estimates, expect to get estimates per subject
and then average those together. You do not want the errts
time series averaged down to zero before running 3dFWHMx, as
that might play havoc with the estimates (the data might look
VERY blurry, I don't know). But I would not expect averaging
the errts datasets to increase the reliability of the blur
estimation. Averaging over the errts and then over subjects
is already (hopefully) generating a very robust estimate.

Note that for a blur estimate, less noise is not necessarily
better. The goal of 3dFWHMx is to get an accurate estimate
of the blur, it really has nothing to do with random clusters
at that point.

- rick