Print

Print
Dear Mike, 

Before estimation, the displayed design matrix is original matrix X (with the exception that DCT drift basis elements aren't shown).  After estimation, the 'filtered and whitened' design matrix (xX.xKXs.X = K*W*X) is shown.  If certain columns appear to be zeroed out by whitening, that is because the corresponding scans were found to have super-high variance.  (See my earlier post on errors with PET data).

I can't figure out why this is happening, though, as each session's data is scaled separately.  You can check out the scan-by-scan scale factors used to bring each session to a Grand Mean of 100 by looking at SPM.xGX.gSF; if there is crazy variation in this (e.g. super-low values for sessions 1 & 3) it would explain the effect.  Then the question is why are the computed globals so different over sessions... maybe crazy-large artifactual values in sessions 2, 4, 5,..,8?

Hope this helps with the detective work!

-Tom


On Wed, May 11, 2011 at 12:57 PM, Michael Thomas Smith <[log in to unmask]> wrote:
I'm interested in the correlation between a time series and the voxels in the brain. To find out more I made a GLM
(see http://www.sal.mvm.ed.ac.uk/images/problem1.png ).

Before estimation it looks fine (I think?).

During estimation however, 6 of the 8 columns are set to zero (see inset in bottom-left corner of the linked image). Just sessions 1 and 3 output beta images. There are no errors or warnings expressed during estimation, and the matlab output appears ok.

- I've confirmed the number of images in each session is equal to the number of values in the .txt regressor files.
- I've tried adding conditions to each session (just a single event near the start of each session) to see if it's the lack of conditions that's upsetting SPM. (this didn't make any difference).
- I've rebuilt the design matrix from scratch by hand, using the interface, but that didn't make any difference.

Looking forward to any suggestions!

Thanks,

Mike Smith

--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.



--
____________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick
Coventry  CV4 7AL
United Kingdom

Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax:  +44 24 7652 4532