I am currently collecting both an sEEG dataset and an fMRI dataset to look compare correlation and anti-correlation in networks across modalities. There’s a lot of controversy over whether GSR should be used in fMRI when examining anti-correlations, as the extent to which this method may be inroducing them is questionable. I’m avoiding the use of GSR in my fMRI data for this reason, but I am less certain about how to proceed with the sEEG normalization method.
For my sEEG data, I am collecting from both cortical and subcortical areas which have highly different activation signatures. I tried using an average re-referencing scheme, but this seems inappropriate for my data given that issue. I then tried the “preferred” method, bi-polar re-referencing and noticed that there are A LOT of anti-correlationed regions and networks within my data. My concern is that the bi-polar re-referencing is introducing the analogous problem of GSR and fMRI into my sEEG data.
Any input on this problem would be greatly appreciated. Thanks again!
You are correct in that the average referencing scheme is analogous to GSR. Bipolar referencing is more analogous to spatial smoothing in MRI, where noise is tackled at a more local level. I suppose if you have very noisy data, such that the global level of noise is more prominent than trends in local fields, bipolar referencing would be similar to average referencing. Does that help at all?
I’m more concerned that bi-polar re-referencing is having an effect similar to GSR on my data in that it is inducing spurious anti-correlations. When subtracting the average of the two electrodes, it is shifting the weights such that one channel is highly anti-correlated with its neighbour. This appears to be a spurious anti-correlation relationship simply by virtue of subtracting the average of the two electrodes. Given that the phenomenon of anti-correlation is the object of this study, I’m wondering whether this method poses a fundamental problem. It may be the case, though, that the overwhelming majority of correlation between two neighbouring electrodes is noise, and as such, it is reasonable to see these anti-correlation relationships as genuine.
Yes, a problem is that there is not a ground truth. But, perhaps it may be worth running some kind of component decomposition on the signal, and see if components that are likely noise are correlated across channels.