We are preprocessing data using other methods, but then put the timeseries into CONN so we can do compcor on them and subsequent analysis.
I'm attaching a screenshot from the QA of histograms before and after denoising.
Also attaching a screenshot showing denoising parameters included...
in particular, WM and CSF selected for compcor.
filter set to include 0-100 (essentially to turn it off).
linear detrending also selected.

As you can see, there are around 60 correlation distributions here. Prior to denoising, the means are on the order of 0.1. After denoising they are ALL going to zero.
Would expect this from GSR, but that is exactly what we are trying to avoid and therefore using compcor.
Please advise... this is unexpected.
Thank you.

here is the screenshot with denoising parameters...

That is curious. It is fine if those distributions are all approximately centered after aCompCor, but their means/centers should not be exactly zero (as your are right that is the hallmark of GSR). Could you please:

1) display the distribution of QC_GCOR values across subjects (in QA plots, 'distribution of subject-level QC measures') to better look at the range of center values for those distributions

and 2) please display the QM and CSF masks for one or a few subjects just to check that these are not incorrectly extending beyond their corresponding tissue classes

Also, just checking, but it is often a good idea to also include motion and other (e.g. scrubbing) parameters as part of the denoising procedure instead of just the WhiteMatter and CSF components. Any particular reason why you would prefer not to do that in your dataset? From your figures there appear to be still some considerable level of variability in the shape of the FC distributions across subjects&sessions so I was just wondering...

Best
Alfonso
Originally posted by joshm:

thank you... will look at that.
we are not doing any other denoising since all prior denoising is being done with ME-ICA, which we are relying on to handle all other regressors.  We add in compcor per the 2018 Power paper (PNAS) that suggests ME-ICA plus additional post-ME-ICA global signal correction.
Regarding (2), that was a concern of mine as well - that perhaps grey matter is being included (and therefore giving a GSR-like denoising).  Will thresholding of CSF or WM change that?  What would be the next step if the masks are extending into grey matter?  I saw a previous discussion on the forum about thresholds, but I think the conclusion was it didn't matter too much.

Originally posted by Alfonso Nieto-Castanon:

Alfonso,
Thank you for your reply.
The WM and CSF masks look ok to us.  In any case, we are importing data to CONN that are in MNI space, and the WM/CSF masks are those in CONN (also MNI space).
Regarding the QA-GCOR plots, what exactly are we looking at?  Is each point on the plot the center value (mean) of the correlation histogram for a particular subject?
Thank you.

Dear functional conn comunity, 

I have been wondering the same thing. Denoising seems to shift the distributions towards having mean 0 (although in most cases there is a longer tail on the positive side).  This will induce "spurious" negative correlations as well as somewhat change the output of weighted graph theory measures i imagine.
I am not sure what part of the denosing may produce this (e.g., simple movement within the scanner may produce higher correlations). I might investigate myself if I find time; but in the mean time any comment or discussion is welcome. 

Happy connectivity
Peter