Thanks @oesteban!
I don’t think I have an answer for you, but I can say the method of regressing WM/CSF post denoising is different from including the WM/CSF in the melodic mixing matrix. And I think the explanation you give is good, but I will re-phrase in my own words to see if that gets us anywhere.
regressing post denoising
In this scenario, the data will be made completely orthogonal to WM/CSF, That is, there will be no shared variance between the residual data and WM/CSF.
pro(?): the data will be completely independent of the WM/CSF signal.
con: you have to recalculate the WM/CSF measures (and you probably do not want to extract them from a smoothed image)
adding WM/CSF to Melodic mixing matrix
In this scenario, the data will somewhat independent of the WM/CSF, depending on how much variance was shared between the signal components and the WM/CSF. And then, depending on the specific data, a portion of the shared variance will be attributed to the signal components and a portion to WM/CSF. I don’t believe there will be a “winner-take-all” effect where either the signal components or the WM/CSF take all the shared variance, and leave nothing for the other. If the shared variance feels like a large concern, you can create a covariance matrix with the signal components and WM/CSF to see if there is decent variance shared (I don’t have a good idea on what constitutes a large amount of shared variance).
pros: you do not have to recalculate WM/CSF
cons: The data will not be independent with respect to WM/CSF
- this assumes non-aggressive denoising, if you do aggressive denoising, then the signal components will be made orthogonal to all other columns in the melodic mixing matrix.
EDIT: thought of a third option
make the WM/CSF orthogonal to the columns in the melodic mixing file
If you take the original WM/CSF signals, make them orthogonal to the noise columns within the melodic mixing file, and then regress the WM/CSF residuals from your data. This will ensure that the shared variance from the WM/CSF and the noise columns will not be introduced (I think), but any shared variance between WM/CSF and the signals will be given to the WM/CSF.
Alternatively, you could make the WM/CSF orthogonal to all the columns in the melodic mixing file, and then only unique WM/CSF variance will be regressed from the data.
So no need to recalculate the WM/CSF signal from the data.
@oesteban: I was working on a prerequisite to solve this problem in fmriprep to at least make the first scenario easier, I’ll ping the people there to see if I can continue working on it, or if it’s still too much of a moving target.
Best,
James