For a few of my subjects, their fmriprep (v23.1.2) reports for the heatmap of correlations among nuisance regressors in RESTING STATE (but NOT task runs) have “crossed out” correlations, whereas the majority of my subject reports do not (they are just blank). Seems to also always be between non_steady_state_outlier00 and either dvars, fd, rmsd or a combination of these last three.
Why is this happening? Is this a non-issue?
TIA.
Hi @ariveradom,
Short answer, a non issue.
This is likely because your first outlier might be happening on the first volume. In the three metrics with the cross, the first value is typically n/a, since metrics like FD are defined by volume-to-volume changes, so there is no initial value. I am guessing that since the only non-zero value happens on a N/A value of the other columns, it is leading to issues in the correlation function. Nothing to worry about.
Best,
Steven
Hi Steven, thanks for your explanation! I’ve encountered this issue as well. Do you think trimming the first few volumes might help, or are there other metrics we should consider to address potential confounds? Thanks again!
Hi @Wenwu_Sun and welcome to neurostars!
I am not quite sure I understand your concern. What is being plotted above isn’t really an issue, just that a correlation is being attempted with a non-value leading to empty cells in the correlation matrix. I am not sure I see why it should impact your nuisance regression choices.
More broadly there’s not a gold standard set of confounds. And it also can vary based on task vs rest, so I don’t think I can provide you an explicit answer.
Best,
Steven
Thank you, Steven, for your clear and prompt response! I have a follow-up question: Could selecting confounds with high correlations lead to overfitting? Also, would trimming the initial fMRI volumes (or the CompCor time series) help reveal the true correlations and guide us in choosing the appropriate regressors? Thanks a lot for your insights!
Hi @Wenwu_Sun
Overfitting is more a concern for choosing a higher number of confounds, sacrificing temporal degrees of freedom. GLMs are pretty well equipped to handle multicollinear regressors, though it is still better to remove redundant ones for better temporal degrees of freedom.
I am sorry I am not sure how to answer the second part because I don’t really know what you mean by true correlations.
Best,
Steven
MANAGED BY INCF
