First, what do you want to test for with this model? It took me a bit of
time to find it again but you might find this comment from a decade ago
Otherwise, for this particular model, when specifying non-sphericity
options, you should set:
* subject factor: Independence: Yes, Variance: Equal
* condition factor: Independence: No, Variance, Unequal
which will correspond to three variance components.
On 04/06/2019 13:08, Janina Seubert wrote:
> Hi there,
> my question relates to a flexible factorial design that fails to estimate, and I can't figure out why.
> I have an fMRI study with 29 subjects and am trying to set up a flexible factorial design with unequal subject variance and equal variance between conditions, which does not run when I put in all subjects but will run fine when I put in certain combinations of subjects (eg. the first 20 is ok, but 1-10 and 21-29 is not. Also I can leave out subject 17 from the analysis of the first 20 and it will still run fine, but it crashes if I leave out another subject, for example)
> The paradigm consists of 2 regressors that are a linear and a quadratic modulator of the same variable (so definitely collinear, which I thought was ok, but maybe it is not?). If I run separate regressions on the linear and the quadratic regressor, that also works fine, but I would prefer to have them in one model if possible.
> I get the following error message, not sure if that's informative at all?
> Failed 'Model estimation'
> Error using spm_est_non_sphericity (line 208)
> Please check your data: There are no significant voxels.
> In file "/Users/janina.seubert/spm12/spm_est_non_sphericity.m" (v6913), function "spm_est_non_sphericity" at line 208.
> In file "/Users/janina.seubert/spm12/spm_spm.m" (v7120), function "spm_spm" at line 433.
> In file "/Users/janina.seubert/spm12/config/spm_run_fmri_est.m" (v7057), function "spm_run_fmri_est" at line 34.
> I am wondering if someone might have a suggestion on how I might go about to determine if there are any specific subjects that are causing a failure to converge, and if so, which? Or would it be wise to just abandon the idea to use this model altogether?
> I am attaching the design matrix for the estimated model with 20 subjects in case that's informative, as well as the design matrix with subject 17 removed (that's the one that is all black in the 20 subject model). That one looks as if it failed to consider unequal subject variance altogether, judging by the colors, although it ran without error. As a side note, I am also wondering why that is, if anyone has an idea?
> Sorry for the long question, any feedback on what I might do to troubleshoot this would be much appreciated!
Guillaume Flandin, PhD
Wellcome Centre for Human Neuroimaging
UCL Queen Square Institute of Neurology
London WC1N 3BG