Hi Stefano,
At 02:58 PM 1/24/2005 -0500, Marenco, Stefano (NIH/NIMH) wrote:
>Hi Tom,
>I was a little surprised when I found the results in the attached .ps file.
>First I ran the comparison for negative effects with a cluster defining
>threshold of t=2.6 (the results are shown on page 4 of the ps file), then I
>remembered that I had used a smaller threshold earlier and I reran the
>results with a threshold of t=2.54 (page 7 of the ps file). Two
>non-overlapping areas of statistical significance emerged. It's not really
>clear to me how this can happen. If there are enough contiguous voxels to
>make a cluster emerge at the more stringent t-threshold, shouldn't these all
>be present in the more lax threshold test?
My guess is that the cluster you saw at t=2.6 and the other one one at
t=2.54 are two separate clusters. They both have near 0.05 p-values (0.0485
and 0.0385), so adjusting the cluster-defining threshold may have changed
the clusters' p-values slightly, and only one of the clusters happened to
produce p<0.05. To verify this, I suggest you change the FEW-corrected
p-value to p=0.1 instead of the default p=0.05, and check if you see both
clusters. I bet one of them has p<0.05 and the other has p>0.05.
Actually a recent posting by Tom talks about the uncertainty in p-values in
a permutation test. See
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0501&L=spm&O=D&F=&S=&P=25647
Hope this helps!
-Satoru
Satoru Hayasaka ==============================================
Post-Doctoral Fellow, MR Unit, UCSF / VA Medical Center
Email: shayasak_at_itsa_dot_ucsf_dot_edu Phone:(415) 221-4810 x4237
Homepage: http://www.sph.umich.edu/~hayasaka
==============================================================
|