Dear Mike,
I have seen this before when the input images have dramatically different scale. If, say, some of the images are 1000-times more intense than the others, SPM's ReML will then scale down those images (depending on the design and how the badly scaled images are arranged relative to the design factors).
Try looking at the images all at once with the same color scale with CheckReg. Alternatively look at diag(SPM.xVi.V) (the SPM variable in SPM.mat of a troublesome analysis); these values should all be roughly the same or within an order of magnitude. If some are 100x or 1000x different from the others, this would explain the problem (and suggests maybe scalefactor artifacts would explain the differences).
(Are maybe some of the PET images in units of "ECAT" counts, and others in a fractional rate? ).
-Tom
On Mon, May 16, 2011 at 5:32 PM, Mike Sugarman
<[log in to unmask]> wrote:
Hello SPM experts,
I'm running a flexible factorial design with a library of PET scans. My design will set up properly, however when I attempt to estimate the model sometimes I get the error message
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate. RCOND = -1.000000e+00.
> In spm_inv at 22
In spm_reml at 119
In spm_spm at 860
In spm_getSPM at 235
In spm_results_ui at 263
And this message will repeat dozens of times. When I remove a few subjects, the problem typically resolves. Can anybody tell me what might be the issue with the scans that are causing this problem?
Thanks,
-Mike Sugarman
--
____________________________________________
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
