Dear JH,
if you go with a FWE-corrected voxel-level p for the F contrast, you should use the same FWE-corrected voxel-level p for post-hoc tests (meaning the same F threshold). You might want to go with that p devided by the number of conducted tests (might be .../6, if you test A > B, A > C, B > C and also into the other direction), but this is probably not essential (one does not necessarily correct for number of post-hoc tests in other contexts either).
In any case, these post-hoc tests should be limited to those voxels which showed up significantly in the F contrast. Otherwise the post-hoc tests are unrelated to the F contrast.
You should not start with a voxel-level FWE and then turn to any other uncorrected or cluster-corrected voxel-level threshold for post-hoc tests. This way you might well detect significant differences based on the initial .001, which are not significant on the voxel FWE .05 though. However, the voxels were identified based on that voxel FWE .05.
If you start with an initial, uncorrected voxel-level p combined with a cluster-corrected threshold, then post-hoc tests should focus on the cluster as a whole (something like extracting mean contrast estimates). Using a cluster correction means you should not interpret voxels anymore, but should interpret the combination of voxels = clusters. The threshold for post-hoc tests should reflect the initial voxel threshold though (or that p divided by the number of tests), as the voxels were initially identified as being significant based on that threshold (the cluster correction told you that the cluster was large enough not to interpret it as noise).
Note, this is my personal opinion. But if you think about the different thresholds, this seems the most valid approach to me. Unfortunately there seems to be no consensus / guidelines for a valid approach.
Best,
Helmut
|