Dear Laura,
As you are able to run models with other con images from the same subjects it seems indeed to be the "error" related to voxels not surviving the effects of interest F test, meaning the model does not explain any variance at the default threshold .001 for any of the voxels. When turning to equal variance you should not detect any sig. voxels for a corresponding F test [1 0; 0 1] within that model, although this way, the model is estimated at least (as no voxels are preselected for sphericity correction), so you can explore it at more liberal thresholds and also set up a differential contrast.
You could still modify defaults.stats.fmri.ufp in spm_defaults, e.g. something like .05, which makes the initial F test more liberal, possibly resulting in voxels which can be used for non-sphericity estimation. You could report the chosen threshold then, e.g. Eklund et al. (2012, Neuroimage, 10.1016/j.neuroimage.2012.03.093 ) did so. I'm not sure whether anyone has tested how different initial thresholds affect non-sphericity correction though, and to keep things comparable it's probably better to go with the default threshold = pooling across voxels that show a sig. effect at this particular arbitrary threshold, and not e.g. pool across every voxel or across those surviving another arbitrary threshold. In principle you already have an answer anyway = no explained variance at .001.
Best
Helmut
|