Esa,
> It wasn't possible in SnPM to get cluster level statistics using a
> corrected height threshold ("Corrected p-value for filtering") with p-
> values greater than 0.7.
I'm not sure I understand. Do you mean that none of your cluster
level p-values were smaller than 0.7?
If you see a dramatic difference in cluster size p-values between
SnPM and SPM, it's probably an indication of a problem; specifically,
the assumption of homogeneous smoothness probably is incorrect. If
the smoothness isn't constant across the brain, SPM p-values will be
liberal or conservative (depending on the local smoothness) and SnPM
p-values may be conservative. I am working on a solution to this
(non-constant smoothness) problem.
> We are wondering whether it is possible in SnPM to make a cluster level
> inference similar to SPM, e.g. using a height threshold that corresponds
> approximately to uncorrected p-value of 0.001? We are going to make
> statistical inference using corrected cluster level p-values at 0.05 level.
If you are *not* using variance smoothing, then yes, you can specify the
threshold with a p-value, and get results that will exactly correspond
to SPM (i.e. the cluster sizes will be identical, but the p-values
won't be).
If you are using variance smoothing, then this isn't so easy, as there
are no tabulated distributions for the pseudo-t statistic to convert
p-values to pseudo-t values. There's no perfect solution for now; start
with, say, 3 or 2.5.
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|