Hi all,
I am running resting-state fMRI analyses (group comparison GLM,
functional connectivity, ReHo, fALFF) using the Schaefer 400
17-network atlas on a clinical dataset of ~300 subjects.
After applying a brain mask to ROI extraction, several
ROIs (temporal pole and OFC) have low coverage in some
subjects due to well-known susceptibility-induced signal dropout
in these regions. I have already removed
ROIs where >10 subjects have NaN values. However, I still have
a handful of ROIs where 1–5 subjects have NaN values - not because
these subjects have bad data , but because their brain mask
does not cover <50% of these specific atlas parcels.
My questions:
-
Is imputation the solution for this issue? (i have 18 individuals with Nan)
-
if yes, for analyses with correlation matrices and ReHo, what imputation
strategy is most appropriate for this sparse missingness
(1–5 subjects per ROI out of 275)? Mean imputation per ROI, KNN imputation, Multiple imputation ? -
Are there papers specifically addressing sparse missing
ROI values in parcellated resting-state data and comparing
imputation strategies? I haven’t found in the literature how to address this issue
Thank you very much in advance!
Best,
Stepan