You can perform SVC only after the initial voxel threshold (peak threshold, height threshold, "p value adjustment to control") has been selected. The SVC affects the cluster (extent) statistics only, not the voxel (height, peak) statistics. Therefore, it is necessary to report cluster statistics.
"Regarding reporting results, p < 0.05 FWE-corrected on the peak-level after SVC is usually used." I haven't checked that, but it might be the case. However, I don't see any circumstances for SVC in which an information about FWE-corrected voxel threshold makes sense, except if you want to cheat (or don't no about the method). For details see below.
1) If you decide to use a corrected voxel threshold of .05 FWE for whole-brain then any remaining voxel is significant. As you've already corrected for multiple comparisons on voxel level, there's no need to look at the cluster statistics at all. If you don't find any significant voxels on whole-brain for certain coordinates/regions and now run a SVC for these coordinates/region, you still won't find any voxels because the voxels still don't exceed the voxel threshold. So, SVC is useless in that context.
2) If you decide to use an uncorrected voxel threshold of say .001 for whole-brain you still have to correct for multiple comparisons, which is performed on cluster level then. In that case you might detect significant voxels (sig. in the voxel/height/peak statistics), with corresponding clusters failing to reach significance. Performing a SVC will reduce the search volume and the clusters might become significant. However, the SVC should NEVER be centered on a peak of the corresponding whole-brain statistics. If you have hypotheses about a certain region, then you have to use some anatomical mask or another contrast or some coordinates from other studies, if possible meta-analyses. You would then report something like "for whole-brain analysis an initial uncorrected voxel threshold of .001 was chosen. The cluster threshold was set to .05 (FWE-corrected). Due to a priori hypotheses a small volume correction was performed for region x, which was defined anatomically based on the corresponding label included in brain atlas y / which was defined as a sphere with a radius of z mm around the MNI coordinates a b c, as reported in study d". There's no real need to repeat the thresholds once more then for the SVC, because one would expect them to be identical to those of the whole-brain. For results section, something like "The SVC on region x resulted in a significant cluster with T = 4.56, k = 5, pFWE(cluster) = .032)". Unfortunately, many authors don't report k, and often it remains unclear what the p is refering to. I guess k is concealed frequently because it's likely just a few voxels, depending on the size of the volume defined. So it might well be significant, but on the same time it might be questionable whether it is relevant. I mean, nowadays you can find a study for almost any coordinate, so if you find some noise in your results just look for some paper to justify a SVC and turn the noise into something interesting...
3) Now, if you want to report whole-brain statistics with a corrected voxel threshold of .05 FWE and turn to a more liberal voxel threshold for SVC like .001 uncorrected, then you can try so of course. But this means you are changing the voxel threshold AND performing a SVC = reducing the search volume. It's not "just" a SVC. It is disputable why you have to modify both thresholds.
4) The other way round, starting with an uncorrected voxel threshold for whole-brain and then turning to a FWE-corrected voxel threshold (= more conservative = probably smaller clusters as voxels with smaller T values are lost) BUT conducting a SVC (= reducing the search volume = more liberal) at the same time is weird. If you want to perform a SVC, then stay with the uncorrected initial voxel threshold if you've already used it. Actually, some people might turn from uncorrected to FWE-corrected voxel threshold because one can "cheat" this way, assuming the reviewers have no idea about SVC. Imagine a very small cluster with a high T value at its peak. This cluster might still be too small to reach significance on cluster level when performing a SVC, but the peak might be signifcant on voxel level. So they maybe go with "SVC showed significant cluster with a p < 0.05 FWE-corrected on the peak-level". However, this "result" does not bear upon the SVC. The T value was already the same when looking at the whole-brain results.
Hope this helps and sorry for the long message,
Helmut
|