Hi Alfonso,

First, I noticed that CONN 15.f now provide an option to set threshold for compcor in "batch.Setup.cwthreshold". Thanks for rapid revision.

I agree with you on the issue of threshold selection: partial volume effect from GM or too restricted noise variance.

The main concern described in the original paper of CompCor Behzadi et al. (2007) NI was the partial volume effects. Thus in the paper, they used very strict threshold (p>0.99 and 2 voxels erosion for the anatomical CompCor (aCompCor).

The partial volume effect was also my concern, so I tried a couple of options for the thresholding:

A. threshold=0.50 with 1 voxel erosion
B. threshold=0.99 with 0 voxel erosion

with other parameters to be constant as:
- BPF=[0.01,inf] (to compare the denoised timeseries with the unprocessed timeseries)
- # of PCs=5 (WM), 5 (CSF)
- covariates of motion parameters and the 'effect of resting'

I created some figures that show: (1) Movement estimated from ART, (2) color-coded z-score of timeseries from GM, WM, CSF (with P>0.99) voxels from one file (3) and that of another file to compare. X-axis marks TR, which is 1.4 seconds. This visualization is quite similar to that in Power et al., 2015, NI.

First, the comparison between timeseries before denoising (wua+"rest.nii") vs. denoised result with thrs=0.5
https://dl.dropboxusercontent.com/u/24268284/compcor_2001/2001_timecourse_wuarest.nii_vs_cw0.50-1_bpf0.01-Inf_gs0.nii.png

And wuarest.nii vs. denoised result with thrs=0.99
https://dl.dropboxusercontent.com/u/24268284/compcor_2001/2001_timecourse_wuarest.nii_vs_cw0.99-0_bpf0.01-Inf_gs0.nii.png

finally comparison of two denoised results
https://dl.dropboxusercontent.com/u/24268284/compcor_2001/2001_timecourse_cw0.50-1_bpf0.01-Inf_gs0.nii_vs_cw0.99-0_bpf0.01-Inf_gs0.nii.png

I think that huge, i.e. 0.5mm ;), movement at around 290 and 370~400 TRs created big impact on the data (thus abrupt changes in all voxels).
And it looks that the effect of movement was not well corrected the denoising with THR=0.5.
It's not entirely removed, but it looks better with THR=0.99.

But more importantly, it seems like denoising THR=.5 introduced some changes where it was fine in the timeseries before denoising. If you see the comparison between wuarest.nii vs. cw0.50-1, magenta peaks in GM at around 130, 200, 240 TRs were even pronounced after denoising. Although I cannot say that the results with THR=0.99 is perfect, but it looks still better, at least to me.

Small perturbation can be removed after filtering in the following analysis step, but I just want to make a suggestion that more conservative threshold may be safer. This is to compare the results with THR=0.5 vs. THR=0.9, both with 1 voxel erosion:
https://dl.dropboxusercontent.com/u/24268284/compcor_2001/2001_timecourse_cw0.50-1_bpf0.01-Inf_gs0_vs_cw0.90-1_bpf0.01-Inf_gs0.png

This shows that THR=0.9 is still not very different from THR=0.5, unlike THR=0.99.

Finally about a bit irrelevant issue, but I want to say that none of the combinations was good enough to remove the effect of movement around 290 TRs, which can be a reason to consider global signal, or scrubbing. But it is also questionable if it's necessary to remove risking the possibility to introduce additional artifacts.

I also want to compare the covariance distributions for the different denoising parameters. Where can I find the distributions (something like the previews that I can see on the Denoising tab on the GUI) after running standard pipeline?

Best,
-SG

+PS: despiking after regression seems to work quite nicely in this case, at least in removing the impact of the movement around 290 TRs. But I think I still need to inspect the effect of the processing step of question further more. Power's suggestion of inspecting of removals distance-dependent correlation (which could be spurious but only introduced by motion), the distribution of the correlation coefficients (just like the preview already in the current version, which is my favorite functionality of this awesome software), and further inspections could give better ideas to users on which combination of denoising parameters would work on their own data.