Hi all,
Sebastian Hoefle <[log in to unmask]> wrote on Sat, 21 Jan 2012 14:23:01 -0200:
[…]
> I first run DCM on the whitened eigenvariates extracted from a spherical ROI (using spm_regions.m). For this run, I got positive values for the log-evidence (DCM.F). As I observed the artificial edge effects of the whitening in spm (see Ref1, Ref2, Ref3, Ref4 for more discussions in the SPM-Mailing list on this topic ), caused by the extraction of the ROI time series, I run the DCM model again, but now without whitening the data.
I recently ran into the same problem. Apart from the edge artefacts from the whitening in SPM see another difficulty for DCM: The preprocessing does not remove the mean of the voxel time courses, but the calculation of the first eigenvariate depends on it as it calculates a SVD on the covariance or kernel matrix.
So to do the VOI definitions for DCM I changed the following line in spm_regions.m in my personal SPM8 installation:
y = spm_filter(SPM.xX.K,SPM.xX.W*y);
into:
y = bsxfun(@minus, y, mean(y)); % remove mean
y = bsxfun(@rdivide, y, sqrt(var(y))); % whitening per voxel
y = spm_filter(SPM.xX.K,y); % temporal filtering
This whitening invalidates the variance estimation later on to some extend, but I still think that it is a good idea in case of motion artefacts in the data. The alternative would be to just do the mean remove and the filtering afterwards.
Ingo.
|