Thanks Francoise.
We were basing our assumptions on this:
Paraphrasing:
The issue you are now facing relates to earlier warnings from @GilesColclough; the atlas has too many regions given the linear rank of the underlying data. The reason that the data has "low" rank in the first place (keeping in mind that 100+ dimensions is not actually that low!) is due to the ill-posedness of source-reconstruction in MEG. Basically: we are trying to estimate 3k+ signals (one for each voxel) from the correlated measurements of only ~250 sensors, and as the error message says, this gives us only 114 linearly independent signals to work with --- but the AAL parcellation assumes the existence of 115 independent regions.
There are essentially two ways forwards:
- If the AAL parcellation is not a must for your analysis, then consider using a parcellation with fewer regions (e.g. Desikan-Killiany);
- Otherwise, I would first remove all subcortical regions from the parcellation (the SNR in these regions is extremely low compared to cortical regions), and perhaps merge small regions together, or even with larger neighbours. For example, find parcels with very small label counts after the resampling, and either merge neighbours with small counts together, or else merge small regions with their smallest neighbour.
To clarify, we get 68 regional timecourses (Desikan atlas) and perform orthogonalizatoin (Colclough et al. 2015). And the algorithm fails with the message:
The ROI time-course matrix is not full rank.
This prevents you from using an all-to-all orthogonalisation method.
Your data have rank 60, and you are looking at 68 ROIs.
You could try reducing the number of ROIs, or using an alternative orthogonalisation method.
We search the 68 time series for those that are linearly independent.
We keep those and remove the rest.
We now have ~62 time series (depending on the subject).
We run the orthogonalization alogithm again.
This time it is successful.
So i'm not sure I understand.
Not all the timeseries seem to be contaminated with each other because we can remove them and have the algorithm work.
We therefore assume that reducing the number of regions of the atlas will ameliorate this issue.
Please help me understand why I am wrong.
PS. Maybe you mean we should use a different solution to minimum norm?