Dear all,
I have a question about the function that estimates the General Linear
Model (spm_spm()), and more precisely about the following lines :
/%-Modify ResMS (a form of shrinkage) to avoid problems of very low variance
%--------------------------------------------------------------------------
try
if ~strcmpi(spm_get_defaults('modality'),'fmri')
ResMS = spm_read_vols(VResMS);
ResMS = ResMS + 1e-3 * max(ResMS(isfinite(ResMS)));
spm_write_vol(VResMS, ResMS);
clear ResMS
end
end
/
I understand the idea of avoiding problem of very low variance, but I
have some concerns with the use of the max, which makes the value of
ResMS sensistive to the presence of a single outlier voxel. In fact, I
experienced the problem by analyzing a population with two slightly
different preprocessing, and I obtained completely different results (in
one case, I got several detection while in the other case, I got not
detection). This was due to the fact that in the second case, there was
a few number of voxels with a high residuals value. Would it not be
preferable to consider the mean (or more preferably the median) instead
of the max in order to avoid such a problem ?
I also wonder why this correction is not necessary for fmri data.
Thanks,
Vincent
|