I have a couple of questions regarding spatio-temporal cluster based permutation.
I am using a source space modeled on the fsaverage template, with 4098 sources/hemisphere (oct6).
I can successfully generate my adjacency matrix with adjacency = mne.spatial_src_adjacency(src) and run my analysis. I find some clusters, specifically one for each medial wall that do not reach significance. According to my hypotheses, it makes sense to find medial activation, but I do not have strong hypotheses with respect to laterality of such activation.
Therefore, I was wondering if there is a way to compute an adjacency matrix that allows clustering together the two hemispheres, so that I would get one (more significant) bilateral medial cluster instead of two separated clusters.
Related to this, I don’t fully understand the argument t_power of the function spatio_temporal_cluster_test. If I understand correctly, by default the clusters are computed uniquely based on spatial extent, whereas with t_power = 1, the clusters size is calculated based on spatial extent and significance value, is this correct?
I was considering the possibility of using the option spatial_exclude and focus my clustering on the anterior part of the brain, since I have strong hypotheses my activation should be somewhere there. How statistically “appropriate” is such approach? I agree that with a strong a priori hp it makes sense to restrict the clustering search space, but I think I have never seen it in papers (sorry if this is a stupid question, I am really not an expert!)
I was wondering if there is a way to compute an adjacency matrix that allows clustering together the two hemispheres, so that I would get one (more significant) bilateral medial cluster instead of two separated clusters.
The adjacency is a binary pairwise definition of which vertices are neighbors. If you look at the structure for a surface, it will be block diagonal, one block in the upper left for the left hemi and one in the lower right for the right hemi.
You could decide that vertices on the left hemi should be considered adjacent/neighbors with ones on the right hemi if they are within X millimeters of each other based on Euclidean distance. You could do this for example by passing the src[0]['rr'][src[0]['vertno']] and the src[1]['rr'][src[1]['vertno']] arrays to scipy.spatial.distance.pdist, thresholding it, and using it to replace the upper-right (and using its transpose to replace the lower-left) of the surface-based adjacency matrix that represents the cross-hemi parts, which are otherwise all zeros by definition.
We could make a function for this in MNE-Python if it seems useful to people.
t_power = 1, the clusters size is calculated based on spatial extent and significance value, is this correct?
t_power=1 (default) will use the sum of the t (or F or whatever your statistic produces) values of the points in the cluster. t_power=0 will essentially use the number of significant points, because the statistic will be taken to the zeroth power.
I agree that with a strong a priori hp it makes sense to restrict the clustering search space, but I think I have never seen it in papers
a priori assumptions are often used to restrict analyses and reduce the multiple comparisons problem, it seems reasonable to do it with clustering.