Thanks for the quick reply, Eugene.
To use a concrete example, imagine I have an functional data set with 1) some volumes I want to treat as confounds because there was abrupt motion, and 2) other volumes I want to treat as confounds because the participant drifted beyond a desired threshold (possibly for several contiguous volumes). Let's assume that in case 2 there is nothing to distinguish these drift volumes from those that are acceptable (i.e., no difference in perceptual experience or task performance on the part of the participant).
If I understand correctly, for issue 1, it would be preferable (would you say necessary?) to use the single-volume-per-column method, because the nature of the artifact might vary widely between volumes. The idea here is that the best fit for a given volume will be a predictor limited to just that volume.
For issue 2, would it be acceptable (preferable?) to use the many-volumes-in-one-column method? If not, what is a reasonable way to disregard these volumes? Would the single-volume-per-column method be appropriate? Or is the use of a confound predictor / matrix not the right way to handle volumes that are undesirable for reasons that are not due to "true" artifact? I can imagine other situations where volumes might need to be disregarded for reasons other than motion, but which might not distinguish the signal in the undesirable volumes from volumes that are acceptable (e.g., temporary equipment failure that the participant is unaware of, such as a non-feedback-linked buttonbox failure).
Best regards,
Ruskin
|