Hi all

I am working on a resting-state fMRI project using "Conn" for the analysis.  I would like to know if the "global signal"is included by default when running analyses in "Conn"? Is there a parameter where I can choose whether or not to include it?

In advance, thank you for your help.

Best Regards

Camilla Borgsted Larsen

Hi Camilla,

We were also wondering about this question - did you ever clarify?

My understanding is that since white matter is included as a covariate of no interest by default, some degree of global signal correction is occurring.

If anyone could clarify/expand on this it would be great.

Best wishes,

Dan Fitzgerald

Hi Dan,

I've looked into this when I started using conn and it does not use global signal regression. A paper emphasizing this is Chai, Xiaoqian J., et al., 2012, "Anticorrelations in resting state networks without global signal regression.".  Conn uses eroded white matter and csf masks and regresses out the principal component of the signal from those. This method in theory does not suffer from systematic introduction of negative correlation as pointed out by Murphy 2009 (which is the main concern with global signal regression) but it also retains some of the advantages of global signal regression by removing noise from white matter and csf. Maybe Alfonso can enlighten us more on this topic?

Kind regards,
Vincent.

Thanks Vincent, I could not have said it better. Global signal regression was first proposed as a method to counter the effects of motion-related and other physiological artifacts in fcMRI analyses. When unaccounted for, these effects tend to act as confounding effects typically  inflating/biasing the resulting connectivity measures. You can observe this effect if you look at the distribution of voxel-to-voxel connectivity values labeled as 'original' in the conn toolbox 'Preprocessing' tab, which typically appears heavily shifted towards the right (positive bias). So a first natural correction was to compute the global effects (from the average BOLD signal across the entire brain) and regress-out these effects at every voxel to somewhat 'center' this distribution. The problem with this approach, as pointed out by Murphy, was that it has the potential to introduce artifactual negative correlations between brain regions, mainly stemming from the fact that the average BOLD signal contains a mixture of movement/physiologial effects but also signals of neural origin. As a response to this critique, several labs started regressing-out instead average signals extracted only from CSF and white-matter areas in order to minimize neural sources from the resulting mixture, as it was shown that many motion-related as well as physiological effects tend to also be similarly present in these 'noise' areas (although in different proportions). The CompCor approach extends this basic idea by extracting multiple signals from each of these areas (instead of simply the average signal from each area), which serves as a richer representation of the range of subject-motion and physiological effects on the BOLD signal, offering better protection against the potential biases that these confounding effects could otherwise introduce, while still avoiding the inclusion of potential signals of neural origin in order not to artifactually introduce negative correlations into the resulting connectivity measures (Chai et al. goes into a lot of more detail and examples about these issues if you are interested). 

Hope this helps
Alfonso


Originally posted by Vincent Beliveau: