Dear André,
I'm not sure what "For this analysis, 4 subjects produced the error message" means, did you obtain error messages on single-subject level or did you try to successively include/exclude these subjects from the second-level models?
In any case, as the error message states and as Volkmar has suggested, it is related to the non-sphericity estimation for that particular contrast, which is determined based on a selection of voxels that surpass an initial effects of interest F contrast. There are no sig. voxels, so there's nothing to pool across to determine the full covariance matrix. One workaround is to set variance to "equal" for the group factor, another alternative would be to modify defaults.stats.fmri.ufp in spm_defaults to lower the initial threshold.
Concerning your data, well, it seems there are no group differences for particular contrasts, at least with regard to the F contrast at the default .001 uncorrected (in principle you might detect lots of sig. differences with two one-sided T contrasts within the model at .001 uncorrected though, which would correspond to an F contrast at .002 uncorrected, or with lower initial thresholds).
You should not discard these subjects except there are very good reasons (artefacts), as this would be a serious confound = selecting data so that you detect group differences.
Best
Helmut
|