When you write that you “added aCompCor50” to get your second plot, do you mean that there were two separate sets of confounds that you regressed out of the data (in two steps), or did you regress out all confounds in one single step? Given that you’re talking about the reintroduction of noise, I’m assuming you did the latter (one step). That latter step would be the correct choice (unless you separately orthogonalized the regressors).
I regressed out all confounds in one single step.
When using all aCompCor components that explain 50% of the variance, which mask were you using? (I’m not sure it matters, but if someone that knows more about AROMA chimes in the mask might be useful information)
In the figures I posted above (i.e, when I was adding 5 or 100 or 200 or all aCompCors), I was adding them in sequence (i.e., a_comp_cor_1, a_comp_cor_2, …). But it seems that if I include all aCompCors, it is using the aCompCors from the WM, CSF, and the combined mask?. Would this be redundant given that aCompCors from the combined mask contain information from both the WM and CSF masks? Perhaps this is (partially) the reason why I’m getting noise in my denoised data when using all aCompCors indiscriminately (i.e., from WM, CSF, and combined masks). However, I think another reason why I’m seeing noise is because aCompCors are derived from the unsmoothed data before it is smoothed and AROMA-corrected (as you point out in your post, and as many other posts have discussed).
The aCompCor components are calculated prior to denoising with unaggressive AROMA, so regressing out many aCompCor components from the
smoothAROMAnonaggrwould reintroduce some of the variability AROMA had taken out (Lindquist et al. 2019).
Yeah I’ve been reading the many threads on this potential issue. I’m aware that some tools exist to re-extract aCompCors from the unsmoothed AROMA-corrected data. But if I plan to use the smoothAROMAnonaggr output, I thnk I’d need to extract aCompCors from this file, not the unsmoothed file. Is this a correct interpretation?
To check this inference, you could make similar plots for outputs besides
smoothAROMAnonaggr.E.g., I’d expect less that regressing out aCompCor50 fromdesc-preproc_boldwouldn’t result in such terrible voxels, while regressing out the first 5 components might have slightly more of an effect. Though, again, using the broken stick threshold I’ve often found that no aCompCor components explain significant variance, which would mean a lack of much effect from regressing out just a few components is at least somewhat likely.
That’s a good idea, I’ll try that out.
The reason why this issue I’m having is particularly strange to me is because I have used fmriprep on previous datasets and did not have issues with using the outputted aCompCors as denoising regressors (using version 1.3.2). I also had never seen so many aCompCors outputted, which is why I was asking in my initial post if that was normal. Could you elaborate on why the number of aCompCors (from either WM, CSF, or combined masks) should equal the number of TRs? I’m not sure I understand why the number of principal components should equal the number of TRs. Why would a principal component be extracted for each TR?
Thanks so much for your help, I really appreciate it as someone that does not have much experience with this. Hopefully others find this thread enlightening as well.