Hi Gang,
That is funny that you refer me that paper because it is only after reading that paper that I chose to email you in the first place. That paper seems to deal only with the correct way of selecting a p value and describing the appropriate was to test for conjunctions. For the analyses in that paper they used Familywise correction to deal with false positives.
I use cluster-size correction via AlphaSim to deal with false positives. My question relates to when cluster correction should take place in a conjunction analysis.
The difficulty with your suggestion of carrying out only steps 1-3 (below), without any clustering as step 4, is that you may encounter a scenario in which there is 1 voxel of overlap for tests 1 and 2. Is this 1 voxel meaningful? Is it reportable?
1. Voxel-wise thresholding for statistical test #1, followed by cluster-wise thresholding
2. Voxel-wise thresholding for statistical test #2, followed by cluster-wise thresholding
3. Identifying spatial overlap among above-threshold voxels for tests 1 and 2.
That is why I suggested the alternative, which is to threshold individual voxels for tests 1 and 2 (without doing cluster correction), then look for overlap, and then finally ask where there are clusters of overlapping voxels that meet the minimum size requirements dictated by AlphaSim (see below).
1. Voxel-wise thresholding for statistical test #1
2. Voxel-wise thresholding for statistical test #2
3. Identifying spatial overlap among above-threshold voxels for tests 1 and 2.
4. Cluster-wise thresholding on the spatial overlap map
Given that I need to use a minimum cluster size at some point in the conjunction analysis to correct for false positives, does it make more sense to carry it out only for step 4 (as described immediately above)?
Christine