We would recommend using a conservative HPF - just pick a very large cutoff period (longer than any fluctuation you're interested in). This will then have little effect on slow fluctuations, but it will remove the very slow drifts and linear trends, and your correlations with these are quite high so you should definitely remove them.

As for motion, this is trickier. You still have pretty strong correlations there, so you need to do something. However, it is not uncommon that putting in motion parameters as confounds will remove all of the effects of interest too. The two remaining options are fsl_motion_outliers (which is generally best if the motions are large are sporadic) and melodic (where you need to carefully separate out the artefactual and non-artefactual components - either manually or with FIX or AROMA). You can also combine these (fsl_motion_outliers and melodic). However, if there is a strong linear drift in motion that is driving some of this then the HPF above should fix a lot of that. So it may be that using an appropriately conservative HPF will let you fix things.

From: FSL - FMRIB's Software Library <[log in to unmask]> on behalf of Marie-Eve Hoeppli <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Thursday, 28 April 2016 17:53
To: "[log in to unmask]" <[log in to unmask]>
Subject: [FSL] dealing with correlations between movement parameters and regressors of interest

Hi,

I am running some analyses on BOLD fMRI multiband sequences. We recorded continuous ratings during these sequences. I first downsampled these ratings at a rate of 250ms. I then demeaned them and used those ratings as regressors of interest in my GLM model.

I ran a two-step analysis:

1) feat with only MCFLIRT and registration turned on, followed by extraction of CSF-WM regressors using the feat outputs

2) feat with Fieldmap correction, spatial smoothing, no HPF and CSF-WM regressors as nuisance regressors

I have some interesting signal, but also a lot of noise, mostly posterior. The noise really looks like something due to movements. So I ran some correlations between the motion parameters I got out of step 1 described above and my regressors of interest (downsampled at the same rate as the motion parameters this time). I obtained some low to moderate correlations (average r =.4) . I also checked for correlations between the regressors of interest and a linear drift and I obtained moderate to high correlations (average r = .6).

I had previously tried to denoise the data using Melodic and HPF, but then there was close to no signal left.

I would be extremely grateful if someone could advise me on how to denoise my data without losing my signal.

Thank you.

Best,

Marie-Eve