Confound regressors in GLM

I have run fmriprep and would like to run a GLM on the normalized functional data. However, I’m wondering if the confound regressors from fmriprep are appropriate since they have been computed in epi space.
Can I simply include fmriprep’s confound regressors in my GLM if the functional data have been normalized?

Yes. You’re looking to remove temporal signals that you have good reason to believe to be noise (of non-interest, anyway); this property does not go away under normalization. If normalization means a voxel that would have contained a signal of non-interest no longer correlates with that signal, the beta for that regressor will be smaller, but it will not damage the other uncorrelated signals in that voxel.

I don’t know if anybody’s done a study on whether ROI-based regressors change significantly post-normalization, but I would expect it to introduce a small amount of smoothing that could affect the results, but those results would remain highly correlated.

The concern generally is if your regressors of interest might correlate with a noise signal. It seems (citation needed) that calculating regressors as close to the original data as possible introduces less opportunity for task-related signals to infect ROIs that are expected to be pure noise, which is why fMRIPrep calculates all confounds in the original BOLD space (with STC + HMC + SDC).

And of course, some can only be calculated during preprocessing, like motion parameters and framewise displacement.

1 Like