Adding run level regressors in FSL leads to error: "Warning: at least one EV is (close to) a linear combination of the others"


I’m conducting the second (run) level fMRI analysis using FSL. I am adding regressors at this stage to assess the impact of run-to-run variation in hallucination occurance and length (as reported by button press) in a sample of BPD patients. I have three regressors: 1) hallucination instance (amount of hallucinations in run) 2) mean hallucination length 3) total hallucination length.

However, some patients experienced no hallucination across any of the three runs. Therefore, the value of the regressor (pre-demeaning) for all runs is the same, 0. For these patients, FSL outputs the following error:
“Warning: at least one EV is (close to) a linear combination of the others… (Design matrix is rank deficient - ratio of min:max eigenvalues in SVD of matrix is [a very small number]. Contrasts involving these combinations will be set to 0”.
In response to this error, FSL skips these patients’ analyses.

I am looking for advice on how to avoid this error and ideally some background on why a rank deficient matrix is incompatible with this analysis.


TL;DR: FSL is upset when EVs do not change run to run.