quick reply - this doesn't seem to work so well when i include a large number (40+) of noise component time series recovered from melodic. (it regresses out a lot of useful stimulus and resting state activity).

however, if i cluster my noise component time series into ~5 clusters, and average within clusters to produce 5 separate noise regressors, it works quite well to regress out noise while retaining stimulus/resting state activity.

does anyone know why including so many regressors (40+) even though they are all noise (cardiac, motion, scanner drift, etc) removes much of my useful signal as well? it shouldn't be a dof issue, i have 800 time points per scan.