Dear Horatio,
First of all, note that voxel-wise FDR correction is somewhat disputed, which is also the reason why it has been replaced by topological FDR on peak or cluster level in SPM (see http://dx.doi.org/10.1016/j.neuroimage.2008.05.021 and http://dx.doi.org/10.1016%2Fj.neuroimage.2009.10.090 ), although the "old" voxel FDR can still be enabled by modifiy spm_defaults. Voxel FDR is NOT topological FDR on peak level, which is often confused.
Leaving this aside, as you already try to account for multiple testing with the (corrected) voxel-wise FDR there's no need to go with another "statistical" threshold on cluster level. Your approach would be especially uncommon as you rely on uncorrected cluster p values, which are really hardly ever used. Thus there's no benefit, rather it makes it a little obscure as one would start to wonder what you were trying to do.
Just to make sure, if you go with a correction on voxel level then any of the remaining voxels is sig., be it 1 voxel or 100,000. In fact we would have to interpret every single voxel separately, it's just for simplicity that we turn to "set of interconnected voxels = clusters" afterwards, as it facilitates the discussion. And it's for the same reason that we often combine a corrected voxel threshold with an arbitrary extent threshold of k voxels, we just don't have to talk about single-voxel findings (although they might well reflect true effects). But in fact, if you really have concerns about single-voxel or very-small-cluster findings then you shouldn't go with a corrected voxel threshold but an uncorrected voxel threshold combined with a cluster correction (or turn to something else like threshold-free cluster enhancement).
Best
Helmut
|