Using aCompCor for nuissance regression

, ,
1

Hi,

I’ve sorted through discussions on here re: high numbers of aCompCor regressors output from fmriprep but still find myself a bit confused. My functional runs are 197 volumes (TR: 1.5s). I was originally planning to use the top 10 components from the combined mask and the cosine regressors in my GLMs, but then noticed how little variance is accounted for (~ 10%). If I were to use the 50% rule, I’d have upwards of 70 nuissance regressors which seems excessive.

I tried to use the Broken Stick logic for each participant and run, and again find that I’d need between ~30 and 70 nuissance regressors per run. I can see there’s not a great consensus on this, so I may just go with the top 5 CSF and top 5 WM …but wanted to check and see if anyone encountered the same issue.

Thanks!
Kylie

2

Hi Kylie,

I don’t have a solution, but I am running into the same question while trying to develop a post-fMRIPrep pipeline for our study data, so I thought I’d respond to this thread.

To provide one example, Esteban et al. (2020) seem to have used the following code, with dvars, framewise displacement, and the first 6 aCompCor components and the first 4 cosine components.


runinfo.inputs.regressors_names = ['dvars', 'framewise_displacement'] + \

['a_comp_cor_%02d' % i for i in range(6)] + ['cosine%02d' % i for i in range(4)]
3

Hi Aisa! Thanks a lot for your response. I ended up doing the first five components of the white matter and CSF and including both cosine regressors along with the 6 motion parameters. Just to let you know :slight_smile:

1 Like