Fig. 6.

Global signal regression (GSR) shifts the distribution of resting state fMRI signal correlations across pupil-linked arousal states. (a and b) GSR introduces negative correlations in resting state functional connectivity. For each pair of nodes defined using the shen268 parcellation scheme, the Pearson’s correlation coefficient (R) was calculated using the average BOLD signals in each node from individual rest 1 (a) and rest 2 (b) data. Signal correlations from all pairs of nodes across 27 subjects are then collected to generate the distribution of R. (c–f) A shift of R distribution by GSR is observed in true pairs of fMRI-pupillometry data (c and d, n=27) and across randomized state datasets (e and f, n=702). In (c), the amount of shift induced by GSR was larger in datasets stratified as high arousal state than in datasets stratified as low arousal state. (g) Without using GSR, there are more positive signal correlations at high relative to low arousal state. (h) In two randomized state datasets, GSR introduces a common shift of the signal correlation distributions toward negative correlations. (i) When GSR is applied, the signal correlation distributions at the two arousal states overlap. (j) The same pattern is found in the randomized state datasets. ***: Uncorrected p=0, two-sample t-test for Fisher’s Z transformed R values.