Global signal regression: the beauty or the beast?


Global signal regression: the beauty or the beast?

Issue: global signal regression in R-fMRI preprocessing.
Discuss initiated date: August 8, 2013.
Want other nodes to cite when publishing: yes
No R-fMRI issues is more controversial than global signal regression (GSR), GSR almost polarized the R-fMRI field. Nowadays, if you submit a paper with GSR applied, you will have more than 80% chance to get the reviewer comment “what’s the situation without GSR”. The other case (ask you to do GSR) happens less, but researchers still wondering what the results are if GSR is performed.
GSR used to be an common practice in R-fMRI preprocessing, and Fox et al., (2005) reported the beautiful anti-correlated intrinsic networks (default mode network vs. task positive network) in the brain, so as Frasson (2005).  
GSR has been brought into the center of a storm in 2009, since Murphy et al. (2009) reported the tendency of GSR to zero-center the distribution of correlation values. At group-level, studies reported that GSR could alter inter-individual differences (Gotts et al., 2013; Saad et al., 2013; Saad et al., 2012). The case is even worse, as Scholvinck et al. (2010) reported the global signal is related to neural activity  in anesthetized monkeys, and the correlation between global signal and grey matter signal is very high (Fox et al., 2009), even reached 0.98 (Yan et al., 2013b). More recently, Wong et al. (2013) found that the amplitude of the global signal exhibited a significant negative correlation with EEG vigilance across subjects. Together of these findings, it’s not easy to justify why to remove such a signal.
On the other hand, GSR is believed to afford increased tissue sensitivity (Fox et al., 2009) and decreased dependencies on motion (Satterthwaite et al., 2013; Yan et al., 2013a). A recent neurophysiological investigation demonstrated that both positive and negative BOLD correlations have neurophysiological correlates reflected in fluctuations of spontaneous neuronal activity, and GSR enhances the neuronal-hemodynamic correspondence overall (Keller et al., 2013). 
Given the fact that GSR might be a beauty and a beast, alternative corrections are proposed.
1. PCA-based correction: regress out the first PC instead of the global signal (Carbonell et al., 2011). However, since the global signal is highly correlated with PC1, this is not a far away alternative from GSR in mathematical sense. 
2. Median angle correction: this is a way to correct the median angle (each voxel is a vector and has an angle from PC1, the median value of the angles is the median angle) to a certain value (He and Liu, 2012), i.e., each participant has fixed median angle after correction.
3. Group-level mean regression: recently studies reported that adding the mean intrinsic functional connectivity value (mean iFC (Yan et al., 2013b) or global correlation (GCOR, Saad et al., 2013)) as a covariate in group analysis, has some similar effects as performing global signal regression on empirical data.
4. CompCor: this method (Behzadi et al., 2007) is taking the principle components (usually 5 PCs) of WM/CSF regions as regressors in nuisance regression step. This method cannot acquire the anti-correlated networks as GSR.
5. ANATICOR: this method (Jo et al., 2010) is essentially regressing out regional WM signals – those white matter voxels close to the target GM voxel.
As there is no consensus on the GSR issue, here we would like raise a discussion on GSR related topics. We hope the related researchers can post their opinions, propose their alternative solutions, and explain the justifications. 
If you have related comments, just simply reply this topic. As all the opinions will be summarized in a table (with direct links) later, it will be very helpful if you can use some templates (see an example at the end). All the other researchers can view your post and may comment with their responses. Hope we can reach consensus at some extent with this immediate, efficient and direct peer-viewed-system. 
[may serve as a template in commenting]
Opinion on GSR: 
Alternative method: 
Points want others to cite when publishing:
Behzadi, Y., Restom, K., Liau, J., Liu, T.T., 2007. A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage 37, 90-101.
Carbonell, F., Bellec, P., Shmuel, A., 2011. Global and system-specific resting-state fMRI fluctuations are uncorrelated: principal component analysis reveals anti-correlated networks. Brain Connect 1, 496-510.
Fox, M.D., Snyder, A.Z., Vincent, J.L., Corbetta, M., Van Essen, D.C., Raichle, M.E., 2005. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc Natl Acad Sci U S A 102, 9673-9678.
Fox, M.D., Zhang, D., Snyder, A.Z., Raichle, M.E., 2009. The global signal and observed anticorrelated resting state brain networks. J Neurophysiol 101, 3270-3283.
Fransson, P., 2005. Spontaneous low-frequency BOLD signal fluctuations: an fMRI investigation of the resting-state default mode of brain function hypothesis. Hum Brain Mapp 26, 15-29.
Gotts, S.J., Saad, Z.S., Jo, H.J., Wallace, G.L., Cox, R.W., Martin, A., 2013. The perils of global signal regression for group comparisons: a case study of Autism Spectrum Disorders. Front Hum Neurosci 7, 356.
He, H., Liu, T.T., 2012. A geometric view of global signal confounds in resting-state functional MRI. Neuroimage 59, 2339-2348.
Jo, H.J., Saad, Z.S., Simmons, W.K., Milbury, L.A., Cox, R.W., 2010. Mapping sources of correlation in resting state FMRI, with artifact detection and removal. Neuroimage 52, 571-582.
Keller, C.J., Bickel, S., Honey, C.J., Groppe, D.M., Entz, L., Craddock, R.C., Lado, F.A., Kelly, C., Milham, M., Mehta, A.D., 2013. Neurophysiological investigation of spontaneous correlated and anticorrelated fluctuations of the BOLD signal. J Neurosci 33, 6333-6342.
Murphy, K., Birn, R.M., Handwerker, D.A., Jones, T.B., Bandettini, P.A., 2009. The impact of global signal regression on resting state correlations: are anti-correlated networks introduced? Neuroimage 44, 893-905.
Saad, Z., Reynolds, R.C., Jo, H.J., Gotts, S.J., Chen, G., Martin, A., Cox, R., 2013. Correcting Brain-Wide Correlation Differences in Resting-State FMRI. Brain Connect.
Saad, Z.S., Gotts, S.J., Murphy, K., Chen, G., Jo, H.J., Martin, A., Cox, R.W., 2012. Trouble at rest: how correlation patterns and group differences become distorted after global signal regression. Brain Connect 2, 25-32.
Satterthwaite, T.D., Elliott, M.A., Gerraty, R.T., Ruparel, K., Loughead, J., Calkins, M.E., Eickhoff, S.B., Hakonarson, H., Gur, R.C., Gur, R.E., Wolf, D.H., 2013. An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data. Neuroimage 64, 240-256.
Scholvinck, M.L., Maier, A., Ye, F.Q., Duyn, J.H., Leopold, D.A., 2010. Neural basis of global resting-state fMRI activity. Proc Natl Acad Sci U S A 107, 10238-10243.
Wong, C.W., Olafsson, V., Tal, O., Liu, T.T., 2013. The amplitude of the resting-state fMRI global signal is related to EEG vigilance measures. Neuroimage.
Yan, C.G., Cheung, B., Kelly, C., Colcombe, S., Craddock, R.C., Di Martino, A., Li, Q., Zuo, X.N., Castellanos, F.X., Milham, M.P., 2013a. A comprehensive assessment of regional variation in the impact of head micromovements on functional connectomics. Neuroimage 76, 183-201.
Yan, C.G., Craddock, R.C., Zuo, X.N., Zang, Y.F., Milham, M.P., 2013b. Standardizing the intrinsic brain: towards robust measurement of inter-individual variation in 1000 functional connectomes. Neuroimage 80, 246-262.

[may serve as a template in commenting controversial issues]

Opinion on GSR: neutral

Alternative method: vascular signal correction (VACOR)

Justification: the global correlation and the global signal may be driven by the vascular system, as all the brain regions share the same vascular system.

Points want others to cite when publishing: 1) vascular signal correction; 2) multiple-level global signal regression.


Global signal regression is controversial by receiving both mathematical and empirical criticisms.  To alleviate the mathematical criticisms, several mathematical alternative methods have been proposed. 

I personally would like to explore more mechanisms underlie such a procedure – why it gives the sensible results. My preliminary findings intrigued me that global signal is actually a vascular signal, as all the brain regions share the same vascular system. Thus I would like to propose an alternative method: vascular signal correction (VACOR) – take several principle components of the venous sinuses mask to regress out.

Given these findings, now I am wondering whether and at what extent GSR could remove vascular effects, although this method is not optimal.

1.     We looked at the correlation between the global signal and all the voxels’ time series (GS correlation).

Figure 1. Correlation with global signal.

From figure 1, the GS is mostly correlated with gray matter voxels, a same observation as (Fox et al., 2009). Interestingly, we observed CSF regions are at the opposite direction of the GS, while the nearby subcortical regions are at the same direction with the GS. The WM voxels don’t show significant correlation with the GS.

2.     Then we leveled-up the threshold for this correlation map, we observed a pattern pretty similar to the venous sinuses system (Figure 2).

Figure 2. GS correlation and the venous sinuses system.

This finding raises a possibility for the observation that all the GM voxels’ BOLD signals are highly correlated: driven by the vascular effect. [Here we would like to seek potential collaboration with who have both R-fMRI data and susceptibility weighed imaging (SWI) data / time of flight (TOF) data, please see this post for initiating collaborations].

3.     A potential issue could be that the global signal is driven by the voxels with high standard deviation / variance / amplitude, as a recent paper (Vigneau-Roy et al., 2013) reported that vascular density is significantly highly correlated with amplitude of low frequency fluctuations (ALFF) (Zang et al., 2007). To address that issue, we did multiple-level GSR and have multiple-level GS correlation maps to check.

4.     Multiple-level GSR. [Here is the second pointyou can skip it if already got tired, :)].

4.1.   Level-0 GS correlation. First, the time series of each voxel is z-standardized: subtract the temporal mean, and then divided by standard deviation. Then the global signal is averaged across the brain, which will no long be affected by the amplitude differences across voxels (unit length for each voxel). The GS correlation map was generated based on the z-standardized time series of all voxels and the GS, resulted in level-0 GS correlation map. The map is almost identical to Figure 1 and 2, i.e., the GS correlation is not purely contributed by the high variance voxels. Of note: the average value of this map, is actually the mean iFC or GCOR (Saad et al., 2013) between any pair of voxels.

4.2.   Multiple-level GS correlation: we do GSR based on the z-standardized 4D time series and the GS acquired by step 4. After GSR, we z-standardize all the voxels again to force the length of the each vector to 1. We calculate the level-1 global signal – which is a vector with the smallest angle with all other voxels after GSR. Figure 3 demonstrated the level-1 GS correlation. Interestingly, the subcortical system is at the opposite direction of level-1 GS, and the venous sinuses system is at the same direction with level-1 GS.

Figure 3. Level-1 GS correlation.

4.3.   While keep doing multiple level GS correction, seems the venous sinuses system is always at the direction of GS, while the motor network is at the opposite direction of the GS.

Figure 4. Level-2 GS correlation.

Figure 5. Level-3 GS correlation.

Figure 6. Level-4 GS correlation.

Figure 7. Level-5 GS correlation.

5.     Summary

5.1.   Vascular effects may generate the global signal effect! We may need to do vascular signal regression (based on SWI or TOF images) instead of GSR. The proposed alternative (vascular signal correction, VACOR) including: 1) map the vascular system based on SWI or TOF images, 2) Extract the PCs of the vascular system, 3) regressed out several (e.g., 5) PCs at the nuisance regression step.

5.2.    Motor related network (putamen, thalamus, motor cortex, SMA, insula) is the network most stable to multiple level GS regression. This network should be a cluster distant from other networks in high-dimensional space. We may figure out how the clusters of networks distributed in the high-dimensional space by clustering analysis.


Credit: this is a project I am currently running, the results and conclusions here are preliminary. And if you developed a paper based on these findings, and published a paper quicker than me, I will be very grateful if you can give me some credit by citing this page.


Fox, M.D., Zhang, D., Snyder, A.Z., Raichle, M.E., 2009. The global signal and observed anticorrelated resting state brain networks. J Neurophysiol 101, 3270-3283.

Saad, Z., Reynolds, R.C., Jo, H.J., Gotts, S.J., Chen, G., Martin, A., Cox, R., 2013. Correcting Brain-Wide Correlation Differences in Resting-State FMRI. Brain Connect.

Vigneau-Roy, N., Bernier, M., Descoteaux, M., Whittingstall, K., 2013. Regional variations in vascular density correlate with resting-state and task-evoked blood oxygen level-dependent signal amplitude. Hum Brain Mapp.

Zang, Y.F., He, Y., Zhu, C.Z., Cao, Q.J., Sui, M.Q., Liang, M., Tian, L.X., Jiang, T.Z., Wang, Y.F., 2007. Altered baseline brain activity in children with ADHD revealed by resting-state functional MRI. Brain Dev 29, 83-91.


I found it because I was doing a rs-fMRI study without GSR. And I want to share something I found this morning.

In the new paper in CC. " Shu, H., Shi, Y., Chen, G., Wang, Z., Liu, D., Yue, C., ... & Zhang, Z. (2014). Opposite Neural Trajectories of Apolipoprotein E ϵ4 and ϵ2 Alleles with Aging Associated with Different Risks of Alzheimer's Disease. Cerebral Cortex, bhu237."

Researchers used a way called Global Negative Index[1] to determine the necessity of GSR.

In the method part of this article: "We calculated the Global Negative Index, the ratio of the number of voxels negatively correlated with the global signal to the total number of voxels, for each subject. All were greater than 0.03, suggesting that our data's global signal was irrelevant to nonneural noise and should not be regressed out"

I desire other's opinions concerning this way.

  2. [1]Chen G
  3. Chen G
  4. Xie C
  5. Ward BD
  6. Li W
  7. Antuono P
  8. Li SJ

 A method to determine the necessity for global signal regression in resting-state fMRI studiesMagn Reson Med 2012;68:1828-1835.


Hey all,

I have just published an article in Human Brain Mapping that discusses the potential of GSR to segregate face sensitive areas within the fusiform gyrus. Maybe some more evidence for the beauty of GSR?

The application of global signal regression (GSR) to resting-state functional magnetic resonance imaging data and its usefulness is a widely discussed topic. In this article, we report an observation of segregated distribution of amygdala resting-state functional connectivity (rs-FC) within the fusiform gyrus (FFG) as an effect of GSR in a multi-center-sample of 276 healthy subjects. Specifically, we observed that amygdala rs-FC was distributed within the FFG as distinct anterior versus posterior clusters delineated by positive versus negative rs-FC polarity when GSR was performed. To characterize this effect in more detail, post hoc analyses revealed the following: first, direct overlays of task-functional magnetic resonance imaging derived face sensitive areas and clusters of positive versus negative amygdala rs-FC showed that the positive amygdala rs-FC cluster corresponded best with the fusiform face area, whereas the occipital face area corresponded to the negative amygdala rs-FC cluster. Second, as expected from a hierarchical face perception model, these amygdala rs-FC defined clusters showed differential rs-FC with other regions of the visual stream. Third, dynamic connectivity analyses revealed that these amygdala rs-FC defined clusters also differed in their rs-FC variance across time to the amygdala. Furthermore, subsample analyses of three independent research sites confirmed reliability of the effect of GSR, as revealed by similar patterns of distinct amygdala rs-FC polarity within the FFG. In this article, we discuss the potential of GSR to segregate face sensitive areas within the FFG and furthermore discuss how our results may relate to the functional organization of the face-perception circuit.