Based on my reading of this paper, I have to
agree with Christine.
They seem to be defining the per-voxel p-value
of the conjunction to be the maximum p across
tests. So any clustering would come after that
per-voxel test, as opposed to what we do.
And it makes sense under the null-hypothesis
that at least one result does not hold. Part
of the null for a 5 test conjunction is that
4 of them could be true while the other is not.
So that 4 of them might look individually
significant does not gain one anything when
adding the 5th test to the conjunction.
It looks to me like that implies clustering on
the intersected thresholded datasets (any
clustering before then has no effect, clustering
the result would be the same as clustering the
clustered intersection).
- rick