Hello AFNI community,
I have searched and found some content related to my question in prior discussion threads. However, none have completely addressed my question, which is related to the use of ROI masks and 3dClustSim. We are interested in several specific anatomically defined brain regions based on a priori hypotheses. The volumes of each of these regions vary widely (i.e., small amygdala mask, large dorsolateral PFC mask). In order test our hypotheses, we restricted our analysis based on the anatomically defined ROI masks, and completed a voxel-wise analysis within those ROIs. For multiple comparison correction, we could calculate (3dClustSim) a voxel extent threshold for a given p value using the volume of all anatomically defined masks combined. However, that approach would bias against significant findings in small regions (e.g. amygdala), while being biased toward ROIs that contain more voxels (e.g. PFC). It also does not account for regional differences in smoothness of the data, like those recently discussed in Cox et al., 2016. Specifically, that using the same clustersize threshold everywhere in brain data can result in higher false positive rates than expected in the smoother areas and lower false positive rates than expected in the less smooth areas. An alternative approach would be to calculate the smoothness (3dFWHMx) and the voxel extent threshold (3dClustSim) for each ROI separately. That is, we could calculate the smoothness and voxel extent threshold for the amygdala, PFC, and other ROIs separately. Although we think this second approach could be justified given our a priori hypotheses, others might argue this approach does not adequately address multiple comparisons concerns. We were hoping to get the community’s thoughts on how to balance concerns related to both type I and type II error, and see if there are other suggestions for addressing these issues with regard to correcting for multiple comparisons.
Thanks,
Adam