I am performing decoding analyses in source space data using LCMV beamformer and need insights. I have 2 questions:
I have T1 mri data for 11 subjects and performed the classic source reconstruction pipeline for them, and run cross_val_multiscore() for all of them. Then to perform statistical analysis, I use mne.stats.permutation_cluster_1samp_test(). My question is should I morph the source space of my subjects to the source space of fsaverage, to be more accurate in my statistical analysis?
I have 2 extra subjects for which i don’t have T1 data but that I want to use in my analyses. One option would be to warp them to fsaverage, but would that be a problem next if I plan to morph all source spaces to fsaverage (like in question 1)?
sorry to see that your post was without an answer for so long.
I am not sure I understand fully what you are trying to do, so please let me know if I am misinterpreting.
ad 1: My guess would be that this depends on whether you want to average across the brain output of the participants. Statistical accuracy should not increase with morphing (if anything, the opposite, as it might duplicate source points).
ad 2: Yes you can use a template here. Be aware though, that using a template together with a beamformer is not recommended for EEG data.
We have clustered labels into subregions (eg. visual, somatomotor, etc.) and perform decoding with cross_val_multiscore() wihtin these subregions. Then, we compute the statistical test using an array of accuracies with shape (n_subjects, n_time_points) for each of our subregions. I don’t want to necessarily increase my statistical accuracy but rather to understand whether this method is sufficient or if morphing is necessary, as some MNE tutorials suggest that morphing is required for group-level analyses.
Cool! I forgot to mention that I’m using MEG data– is it still not recommended?
morphing is done to be able to average across source points. These points end up at different locations in individual spaces, so you don’t know which points to average across subjects. Morphing takes care of this by morphing all points into one reference space. If you are using ROIs, this problem is not present if you end up with only one time series or value per ROI.
For MEG data it’s okay to do, however, be aware that the accuracy of your forward and hence also inverse solution will suffer by using templates. In other words: your outcome will be less reliable.
