In my opinion you should use the same initial voxel threshold for small-volume correction that you used for whole-brain analysis. As the name implies, the idea of SVC is to decrease the search volume. This makes sense only if you used an uncorrected initial voxel threshold like .001 in whole-brain, see below:
If you go with an uncorrected initial voxel threshold of .001 in whole-brain, then also use it for SVC. The voxel statistics of SVC will be identical to those of the whole-brain, but the cluster-ps are adjusted according to the new (smaller) search volume. Cluster ps/qs should be smaller than .05 FWE to be considered as significant.
If you go with a FWE-corrected voxel threshold of .05 FWE (in which a cluster-correction is not necessary, as you already correct for multiple comparisons on voxel-level) for whole-brain, then any of the surviving voxels can be considered as significant, even if it's a cluster of k = 1 voxels (although one typically uses a threshold of k = 10 or 20 voxels to avoid these small clusters). A SVC with a FWE-corrected voxel threshold of .05 FWE won't show you anything new, as the voxel statistics remain the same.
However, if you start with a FWE-corrected voxel threshold of .05 FWE for whole-brain and then turn to SVC with a more liberal (uncorrected) voxel threshold of say .001, then you don't only reduce the search volume but also lower the voxel threshold at the same time. So it's not looking at a smaller volume, but looking at a smaller volume with a more liberal threshold.
|