Dear SPM list,

I sent an email to the list a week ago regarding some problems I am having running a mixed-effects analysis in SPM2, but have had no reply. I send again in the hope that someone might have encountered a similar problem or have some useful advice. In summary, the trouble relates to the estimation of my mixed effects design, which may contain too many variance parameters in order to successfully proceed (see below for copied original email and more details).

I would appreciate any pointers in the right direction.

Many thanks

Belinda

-----Original Message-----
From: Belinda Liddell [mailto:[log in to unmask]]
Sent: Wednesday, July 27, 2005 4:22 PM
To: [log in to unmask]
Subject: [SPM] Mixed-effects analysis problems

Dear Karl and SPMer’s,

I’m attempting to run a mixed-effects analysis but have run into a problem that appears to be related to the estimation of the variance parameters.

The analysis consists of 10 subjects, where I have 3 conditions: the first 2 conditions are different for each subject (post hoc determination of volumes based on simultaneously recorded physiological data), while the 3^rd condition is the same for each subject.  I had thought that given this was an unbalanced design, that a mixed-effects analysis may be the most appropriate method.

I set up and estimated a fixed effects design to obtain my initial spm.m file. Given that each subject was different, I selected “no” to the question “Are these sessions replications”. Following this, I implemented the spm_mfx procedure, which was successful and resulted in obtaining a new spm.m file in a new directory. Following instructions, I tried to estimate this new spm.m file using the normal estimation process in SPM, but it would not work – and gave the following error message:

>>>>>>>>>>>>>>>>>>>>>>>>>>

SPM2: spm_spm (v2.66)                              09:29:33 - 21/07/2005

========================================================================

Initialising parameters                 :                   ...computingWarning: Divide by zero.

(Type "warning off MATLAB:divideByZero" to suppress this warning.)

> In C:\MATLAB6p5p1\toolbox\spm2\spm_spm.m at line 363

??? Index exceeds matrix dimensions.

Error in ==> C:\MATLAB6p5p1\toolbox\matlab\sparfun\spdiags.m

On line 102  ==>       a = [a; i i+d(k) B(i+(m>=n)*d(k),k)];

Error in ==> C:\MATLAB6p5p1\toolbox\spm2\spm_spm.m

On line 419  ==>                      s     = spdiags(1./sqrt(diag(s)),0,nScan,nScan);

??? Error while evaluating uicontrol Callback.

>>>>>>>>>>>>>>>>>>>>>>>>>>>

According to an email from Stefan Kiebal (18/2/04), an error like this indicates that there may be too many variance parameters, rendering estimation impossible. These may be increased in my design especially because I indicated that sessions were not replications. In his email, Stefan recommends controlling this by adjusting the selection of non-sphericity corrections. However, this is not possible in the context of a FFX analysis since there are no second level non-sphericity options.

I re-ran the FFX without correcting for serial correlations (AR1). I had thought this might reduce the variance parameters in the model but subsequently learned that spm_mfx has trouble running without pre-whitening the data.

I’m wondering if you might have any suggestions how I may be able to proceed with the mixed effects analysis under these circumstances. I also may have violated an assumption of the mfx process of which I am unaware. Any help that you may be able to give me would be greatly appreciated.

Thanking you in advance,

With best regards

Belinda Liddell