Stef,
> Specifying an initial p-value of 0.001 cluster-extend thresholded (k=32) to achieve alpha = 0.05,
> I find effects of group and the covariate, but no interaction.
The conventional cluster-based adjustment for multiple testing can be excessively conservative ( [
afni.nimh.nih.gov] ). So, try setting voxel-wise p-value to 0.05 and see if you could identify some interesting results.
> The problem is also that some of the ROIs are quite small, for instance the midbrain ROI only contains 60 voxels.
If you have 6 or ROIs, it might be possible to adopt the Bayesian multilevel approach as described in the NeuroImage paper you mentioned earlier.
Gang