Dear Ralf,

sorry for the delayed response. I think the exclusion of stimulus-correlated fluctuations is a conceptual question and there is no right or wrong. In short: If you do it, you risk false positives, if you don�t do it, you risk false negatives.

Basically, you will have to ask yourself how a voxel that is not in gray matter can correlate with your task:

Reason 1): There is some physiological change induced by the task, and, for example, the induced change in heart rate creates a pulsatile flow change in the CSF as well, from which you extracted the �nuisance regressors�. If this is the case, using such CSF voxels that show a correlation to generate �nuisance regressors� might regress out actual neuronally-induced signal from gray matter voxels, and you end up with false negatives, finding no activation. On the other hand, voxels with CSF/GM partial volume effects might be erroneously considered �active voxels�, if you don�t include such nuisance regressors, because the task activation looks just so similar to the physiological changes. So if you omit task-correlated voxels, you might end up with false positives. I rather err on the side of false negatives (because usually strong claims are only made about positive findings in most publications), and therefore opt for not excluding any voxels from a well-defined mask because of correlations. But the �well-defined� is the crucial thing. You have to be sure no gray matter voxels end up in your mask for the noise rois extraction, otherwise you might regress out neuronally-induced signal, and the argument above does not hold.

Reason 2): There is some random correlation in some voxels of the ROI (since we have a lot of them) with your task. In this case you would end up regressing out the correlated part, if it ends up in the extracted principal components. I feel that if the number of voxels is large and they are physiological noise-dominated (or better: not dominated by the random correlation or task-based fluctuation), this effect is small, as long as your time series is not too short (compared to the number of nuisance regressors). If that is the case, i.e., if the degrees of freedom in your residual data becomes very small (i.e., number of regressors is approaching number of points in the fMRI time series), funny things can happen, and I highly recommend these two papers by, among others, Molly Bright and Kevin Murphy (as first and last authors) for further reading:

Is fMRI �noise� really noise? Resting state nuisance regressors remove variance with network structure
http://www.sciencedirect.com/science/art...

Potential pitfalls when denoising resting state fMRI data using nuisance regression
http://www.sciencedirect.com/science/art...

But I also think their main conclusion supports the idea that if you have good reasons to include a moderate number of nuisance regressors into your model (compared to the length of your paradigm), you are safe despite the random regressor correlations, too.

All the best,
Lars