Hello Cheng,

I have thought about this for a minute and this is what I came up with:

The ROIs in the AAL atlas are not functionally homogenous, i.e. they represent signals from different functional subunits of the brain.  So adding them together may result in the phenomena that you see.  This can be a justification for using the first eigenvariate from a SVD decomposition to summarize voxel time courses, but that will introduce problems of its own.  For example, if a ROI covers two functional subunits of the brain, then the eigenvariate will favor the one with the larger variance and exclude the other.  The result is that the ROI time course will only appear functionally connected to one of the subpopulations of the ROI.  On the other hand if you use the mean time course, as long as the time courses from the subpopulations do not add destructively, it will remain correlated (although likely a dimenished correlation) with both subpopulations.

I would bet that this phenomena, while it still exists, is less pronounced in the CC200 or CC400 time series which are more functionally homogenous, due to their smaller size and the fact that they were derived using functional information.

The justification for detrending data is to remove baseline scanner drift and lowpass filtering for removing some components of physiological noise (hopefully) as well as higher frequency scanner noise, that might induce false correlations in the data.

If you consider the new low frequency drifts to be an 'artefact' of the averaging process, then you might want to remove them.  I haven't seen this addressed in the literature, so I don't know what to refer you to in terms of a reference.

Personally, I wouldn't re-detrend or filter the data, and I would use the smaller and more homogenous ROIs.

I am very interested to see if anyone out there has an opinion on this point.

Cheers,
Cameron