Hi,
anything you do not model ends up in the error term of the GLM. Thus, if there is non-Gaussian noise in your data it will inflate your residuals. It seems to make sense that at rest you may have more RSNs kick in that during the task which will give you structured "noise" or RSN activations which you are seeing.
You may be able to extrect these components using melodic and i) include their time-course as a EV in your model or ii) filter them out. Both ways have their own advantages and disadvantages.
Cheers-
Andreas
________________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von Jeff Rudie [[log in to unmask]]
Gesendet: Donnerstag, 15. April 2010 18:55
An: [log in to unmask]
Betreff: [FSL] Regression of Task and working with Residuals
Hi FSL list,
I have been attempting to remove the effects of the task from an event related data set by working the residuals after applying my task model. To see if the task has been removed I have added the mean_func file to the residuals to make it no longer have nonzero values and then ran the same model again. What I have found is that at the group level, the main contrasts of rest>task and task>rest become flipped so that the residuals appear to have a pattern of task activations for the rest>task contrast. The max z-stat is reduced (from 7.9 to 6.5). I have tried using FLOBS (4 basis functions) instead of a double gamma HRF, and that seems to soak up more of the signal (from 7.9 to 4.6), but still I am wondering why I might be finding this inverse pattern in the residuals (GLM overfitting data) and if there might be any other way to better remove task signal the data.
Thank You,
Jeff Rudie
|