Ged,
> I was wondering how non-parametric methods (e.g. SNPM) relate to
> this issue -- would the same problem or any similar problems be
> expected?
The short answer is 'yes', similar problems would be expected.
Generically speaking, non-parametric permutation methods are nothing
more than a way for getting more trustworthy P-values. Problems of
roving peaks in your t images won't be fixed if you just change your
P-value threshold.
More subtly, though, SnPM allows you do use variance smoothing, which,
under low DF (say, < 20) will regularize your variance and make your
t-image more like your difference (contrast) image. This might make
for more satisfying results (it certainly seems to improve power).
My own $0.02 on your initial problem, though, is a question: Why were
you looking outside the brain in the first place? Every additional
voxel you examine increases the severity of the multiple testing
problem. By searching outside the brain you are allowing both
nonsensical results *and* are reducing the sensitivity of your
analysis. Hence, I'm in favor of using fairly tight explicit masks
which remove any non-brain voxels from the analysis in the first
place.
Hope this helps.
-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
|