Testing different noise models output from fmriprep - unclear results with ICA-AROMA

Something seems strange here, adding in the motion parameters and their first derivatives into the aCompCor model made the correlation with FD stronger, which is strange since FD is just a linear combination of the motion parameters. Similarly, adding in the 6 motion parameters induces a higher correlation with FD than the base model.

I wonder if there’s any additional filtering steps and when they might be? I know that bandpassing just the BOLD data prior to regression actually induces artifacts and does a poor job accounting for the covariates.