 |
 |
"However, the uncorrected threshold prior to SVC may be (and usually is) different from that used in the whole-brain analyses. [...] Thus, peak-level statistics corrected for FWE over an a-priori ROI is relevant for inferences following SVC." I still cannot agree on that. The SVC in SPM is implemented to work on cluster statistics. 1) You can use different thresholds for different statistical comparisons / contrasts, as any threshold is arbitrary. But there's a lack of consistency then. The idea of a SVC is not to randomly choose some threshold or making something significant somehow, but to go with the thresholds you would have used for whole-brain and adjust these criterions to a small volume based on random-field theory (smoothness, resel counts). This implies that you must already have determined some thresholds based on RFT, otherwise the SVC is not justified. a) If you go with an uncorrected initial voxel threshold, then you do not rely upon RFT at all. So there's simply nothing to correct with a SVC. You should use a more liberal voxel threshold to further explore the a priori region. b) If you go with an uncorrected initial voxel threshold and correct for multiple comparisons based on the cluster-statistics, then SVC can be employed to adjust the cluster statistics. This approach is presented in Friston (1997, Human Brain Mapping), http://www.fil.ion.ucl.ac.uk/spm/doc/papers/karl_anat_specified.pdf . It does not make sense to adjust the voxel statistics because the uncorrected initial voxel threshold is unrelated to RFT. You would report T value of the peak, k and the SVC-derived pFWE on cluster level. If you wanted to additionally report the voxel p of the peak (which is typically not required as one reports the T or Z instead), then you would have to go with the uncorrected peak p. This p and the T are identical to the p and the T of the whole-brain statistics. 2) Now the FWE-correction on voxel level is also based on RFT of course. The idea for a SVC might be that the critical threshold for voxels should be smaller when decreasing the search volume. This corresponds to the approach described in Worsley et al. http://www.math.mcgill.ca/~keith/unified/unified.pdf (see the table on page 20 for different critical T values on voxel/peak level depending on the search volume/resels count). The results of the SVC should more or less correspond to a separate "whole-brain analysis" in which all those voxels outside of the search volume are exluded. However, the SVC is NOT implemented in SPM that way. If it were, then it wouldn't matter which voxel threshold to choose during the inital thresholding. But it DOES matter. The height threshold = critical T during SVC corresponds to the critical T determined based upon the "p value adjustment to control". So when going with FWE correction on voxel level and then turning to the SVC, the critical T is still the same although the search volume has been decreased, meaning SPM will not show any new voxels which might have been non-significant before. This also means SPM's SVC can't be used effectively that way (this is an implementation issue though, it would be possible, maybe there's even a hidden function). Another consequence is that it doesn't make much sense to threshold something with .001 uncorrected and then look at FWE-corrected peak statistics in the output, because they still rely upon the critical T derived from the .001. 3) For very small search volumes of a few voxels only the cluster statistics might not be very informative indeed. However, in that context is is questionable whether a SVC makes sense at all, as single voxels might easily reach significance within a very small volume. These findings tend to be interpreted as "we found something in basolateral complex of the amygdala". Now, just based on the MNI coordinates and some brain atlas this might be correct. But when considering the generally rather poor resolution of fMRI it is questionable whether interpretations like that make sense and whether these small clusters are relevant or whether you should additionally use some minimum size of k voxels to regard something as reportworthy. For small volumes I would rather argue to conduct a ROI analysis for the small structure as a whole and extract averaged contrast/beta estimates, which should increase SNR. This depends on the hypothesis though: Do you expect some effect within region x or do you expect the region to show the effect? 4) Concerning SVC, there is definitely a bias in literature I'd say. I regularly stumble across studies reporting significant SVC findings, but I don't recall any non-significant SVC findings. It seems people tend to drop the a priori region then or SVC is used a posteriori. Of course, this does not speak against the SVC method as such.