Dear experts,
The typical voxelwise univariate analysis requires registering individual participant's data to a common space. The assumption of this procedure is that voxels containing true signals are distributed in the same way across participants. In cases where we cannot assume this to be true, ROI analyses in participant native space can be used, but this involves averaging across voxels, which is problematic when the region contain voxels showing opposite direction of signal change. To get the best of both procedures, here is what I came up with, which is modified from the original max-statistic procedure (described in Dr.Mumford's video: https://www.youtube.com/watch?v=zg0tmnGGwBI&ab_channel=mumfordbrainstats).
1. Use independent data to define an ROI (e.g. anatomically or localizer) in each participant's native space
2. Generate 1st-level voxelwise contrast images of the effect of interest for each participant (i.e. effect contrasts.)
3. Generate 1st-level voxelwise contrast images with permutation of condition labels for each participant (i.e. permutation contrasts)
4. On the 2nd-level, randomly sample a permutation contrast from each participant.
5. For each permutation contrast sampled, extract the voxel within the ROI that had the highest contrast value. Naturally this value will likely appear at different locations across participants. Using these values to generate a test statistic (e.g. one-tailed t). This is the step that differed from the original max-statistic procedure.
6. Repeat 4 & 5 to generate a 2nd-level null distribution and get a statistical significance threshold.
7. Rank order effect contrast values within the ROIs across participants
8. Mark voxels that exceeded the threshold determined in 6. These voxels are deemed significant regardless where they are in the ROI and from which participant they come from.
This procedure would allow us to make regional claim without averaging across voxels, which may wash out the signal if a region contains voxels showing opposite effects. But since I haven't done any simulation or derivation of this, it would be very much appreciated if I can get your opinion on this. Does it make sense? Is it obviously wrong in someway?
Thanks,
HY
|