Advice on confounds in GLM


I know this is a problematic and thoroughly discussed question (I’ve spent a couple of days reading previous threads of people asking similar questions and some of the referred papers), and that there is no gold-standard for denoising. However, I’d really appreciate any comments on my current plan - this is a very confusing issue…

My data is task data (not resting-state). I have 12 fairly short runs per subject (~6 minutes per run), with a TR of 2 seconds, 2x2x2mm resolution.

The regressors I’m planning to add are the top 5 aCompCor regressors for each run, and all of the cosyneXX regressor in each run (there’s a different number of them in each run, between 3-5).

I’m using the non-aggressively ICA-AROMA denoised data (though I’m a bit worried about the quite aggressive smoothing of 6mm - see here). Hence, I’m not adding any motion parameters, motion outliers or Field Displacement regressors (I’m assuming AROMA will take care of that). I’m also a bit worried that it’s not smart to use ICA on data with such short runs. Anyone has experience with this? perhaps with short runs it’s better not to use ICA-AROMA?

I’ve also opted against using the global signal as this seems contentious.

I’d really appreciate any comments or suggestions.