Dear Andy,
to start with, a comment on your first message "Because we can set p threshold after doing ROI but have to set it before SVC." This is actually incorrect. fMRI is very prone to messing around with the thresholds, because in most cases people probably have no idea how large their effects are going to be and whether the blobs are large enough to be thresholded at a conservative .05 FWE on voxel level or not. However, before running any sort of analysis you should have defined the thresholds of course.
Another comment on the thresholds for SVC. You should use thresholds which agree with your whole-brain threshold (see below). It is not invalid to look at whole-brain with some threshold and perform the SVC based on another, because any threshold is arbitrary, but it's certainly not compelling. If you decide to use different thresholds you definitely have to report them.
The idea of a SVC is to correct the statistics for a small volume, based on random field theory. Thus there are two options:
1) Apply some correction to FWE corrected cluster statistics (as presented in Friston, 1997, Human Brain Mapping)
2) Apply some correction to FWE corrected voxel statistics
For 1) you would typically look at whole-brain statistics with an initial uncorrected voxel threshold (like .001). To correct for multiple comparisons you would apply FWE correction on cluster level. You would then report only those clusters with an FWE corrected cluster p < .05. For your volume of interest, you would perform a SVC based on the same initial uncorrected voxel threshold. Thus, if there were no voxels below .001 before, the SVC won't show you anything either. However, if there were some voxels before, but the FWE corrected cluster was p > .05, the SVC might now result in an FWE corrected cluster p < .05 due to the lower search volume, which typically includes much less resels than the whole-brain volume.
2) is a little tricky, as it is not implemted in SPM in a straight-forward way, and I guess many papers don't perform this SVC correctly. If you look at your whole-brain results with an FWE corrected voxel threshold of .05, then you already correct for multiple comparisons on voxel level. Thus, you don't have to look at cluster statistics at all (typically one uses some extent threshold like k > 10 voxels though to avoid "single voxel" findings). Say you looked at your whole-brain statistics with .05 FWE, which equals to, say, T = 6.00 (height threshold below the glassbrain and FWEp at the bottom of the statistics output). If you click on "small volume" and define some mask or coordiante the glassbrain is still thresholded at T = 6.00 (corresponding to the number of resels in whole-brain volume). The FWEp and volume at the statistics output should have changed though, say FWEp is 4.00 now (actually this is the T value). This means that for your volume of interest, the FWE corrected p value of .05 corresponds to a height threshold T = 4.00. Load the contrast again, but instead of "FWE" choose "none" during "p value adjustment to control" and then enter 4.00. The whole-brain results are thresholded at this T now, but ignore these results. Instead, click on "small volume" and select the same mask or coordinate again. Look at the statistics output. If there are any voxels with an FWE corrected voxel p < .05 (aka peak-level), go ahead and report them (peak coordinates, T value, FWE corrected voxel p < .05, k - but ignore the cluster statistics). It is very important to go this way, otherwise the SVC "results" are based on the incorrect height threshold! It might take some time to get some nice figures, because you should mask those voxels outside of your volume of interest with T values ranging between 4.00 and 6.00 (in the current example).
Having said that, I think the SVC is somewhat problematic
1) in many cases people don't report enough information how the SVC was conducted (do they perform SVC on cluster level or voxel level)
2) if the search volume is small, then it contains only a very low number of resels, thus "correction" might be very close to "no correction" - this holds both for SVC on cluster level and on voxel level. Note that often, people don't report the k at all, so it remains unclear whether there was only a single sig. voxel.
3) Major limitation in my opinion: The SVC is not very informative. Assuming you're interested in this particular volume (otherwise there would be no need for a SVC), is it really sufficient to know there's something going on? I would rather prefer a proper ROI analysis in which you can also extract beta/contrast estimates for the different conditions.
Best,
Helmut
|