Hi,
I've been playing around with randomise on some (unrealistic)
simulated data, and have got voxel-wise (1-p)-values of more than 1,
which seems like it should be impossible...
I've put a very small (10x11x12x10) 4D NIfTI here:
http://www.cs.ucl.ac.uk/staff/gridgway/test.nii
which appears to give the following results:
$ randomise -i test -o test_randomise -1
$ avwstats test_randomise_vox_tstat1 -R -x
0.000000 1.068411 5 4 5
$ avwmeants -i test_randomise_tstat1 -c 5 4 5
3.851071 # fair enough
$ avwmeants -i test_randomise_max_tstat1 -c 5 4 5
0.418945 # fair enough
$ avwmeants -i test_randomise_vox_tstat1 -c 5 4 5
1.068411 # ???
$ avwmeants -i test -c 5 4 5
0.899183
2.682735
1.605658
1.076425
2.913476
-0.619041
0.911262
0.595414
1.170928
1.059577
As far as I understand the vox_tstat image, each voxel's p-value is
only dependent on the raw data for that voxel (as I have not selected
variance smoothing). If I put just this voxel's data into MATLAB and
resample (with replacement, sorry, 1024 or 10000 times) then I get a
p-value of about 0.003, which seems more reasonable than a negative one...
It's getting late though, so the chances of my having gone completely
mad are quite high...
Thanks,
Ged.
P.S. I'm using FSL 3.3.9 (randomise v0.11) on 32-bit RedHat Ent. Linux
Rel. 3.
|