Dear Israel,

I don't think that what you observe is due to your Matlab update.
Could you have been using two different versions of SPM5 (updates 1111
and 1782)?

This is probably due to an over complete specification of the covariance
components that may have been disclosed by a change to spm_reml, which
now returns the hyperparameter estimates after the update (as opposed to
before). ((c) Karl)
It means that you don't have enough degrees of freedom to estimate some
of the parameters in ReML and you should try to estimate your model
again after turning off the non-sphericity correction (set Variance to
Equal in the factors from your Flexible design).

The same conclusion could be drawn from this email:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0810&L=spm&P=28420

Best regards,
Guillaume.


Israel Christie wrote:
> Greetings SPMers,
> 
> I've been cursed with an inability to replicate a previously
> bulletproof PET analysis following a recent update to Matlab R2007a
> (Version 7.4.0.287).
> 
> The design incorporates two factors: Session (2 levels) and Task (5
> levels). The first 2 levels of task are performed once and the
> remaining 3 levels twice each (see the attached design matrix). The
> model is specified as a flexible factorial using the "Specify all"
> method.
> 
> The design is specified and estimated without error but the problem
> emerges when attempting to define any contrast involving the first two
> levels of the Task factor (see the attached contrast manager screen
> shot). In other words, you get invalid contrast errors using either
> the contrast manager or a batch file. These are contrasts that
> previously worked just fine (i.e., they were all estimable).
> 
> As best as I can tell, the problem lies in the fact the beta images
> for the 1st and 2nd levels of task for Session 1 are blank (i.e.,
> entirely blank or NaN). This is not the case for the beta images for
> the 6th and 7th scans, the corresponding and similarly non-repeated,
> tasks from the second session (see the attached beta image display
> screen shot).
> 
> The only potential solution I found in the archives was changing the
> BLAS, though two different BLAS (Atlas and Intel for my Pentium 4
> running Windows XP SP2) failed to correct the problem.
> 
> Any suggestions would be greatly appreciated.


-- 
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG