Further details about recommended confounds for GLM


I am analyzing task fMRI data (2x2x2, multi-band,TR=800ms) and deciding which confounds to include in the GLM. After running across a post recommending, “6 motion parameters, FD, and aCompCor on run level and mean FD on group level (for both task and rest)”, I have a couple of follow-up questions. Confounds from fmriprep: which one would you use for GLM?

  1. Originally we had planned to use the 6 motion parameters + 6 motion derivatives while censoring any volumes flagged as a motion_outlier. Since the motion_outlier columns are determined by FD and DVARS values by default, is there a motivation for moving away from using DVARS?

  2. Is there any worry in using the same measure, FD, as a run level and group level confound?


I think the answers you get will vary across people, but here is my take.

If you are using task-based data, you may want to adopt one of the strategies described in Mascali et al., 2021, which came out since that thread was created. Similarly, see Parkes et al 2018 (and papers cited within) for resting state.

Not necessarily, and the two metrics are highly correlated so usually there are few, if any, volumes that DVARS would flag and FD would not (assuming equal levels of stringency of your FD and DVAR criteria). I do not think any reviewer would ding you for choosing FD alone vs FD+DVARS. However, keep in mind that the default FD and DVAR criteria for fmriprep may not be suitable for you data. For example, if you have few subjects, or young/clinical subjects, you may need to be more lenient.

I would say as long as you exclude the high motion subjects entirely, and that the distribution of remaining mean FDs across subjects are normal or not super skewed, it should be fine. Keep in mind that the first and second level models will use FD differently: the first level model will operate on a volume by volume basis and help your subject-wise statistical maps not be confounded by movement relative to that subject’s motion tendencies. The second level model will account for tendencies of motion between participants.

Consider two subjects with FDs ranging from 0-.2 and then .6-.8 (a very unruly subject! not very realistic, but just for demonstration). Their first level maps may be corrected in similar fashions since there is only a .2 window of FD to model, but of course we expect global biases in that high motion subject since noise is correlated with itself! So including mean FD as a second level covariate helps take care of that.

Hope this helps, and also happy to hear others’ thoughts on this too!