I’m investigating a group difference, so my design matrix is [Group1 Group2 Gender Age LogICV]. All covariates are demeaned. The contrast matrix is thus [-1 1 0 0 0; 1 -1 0 0 0]. However, I got this warning running fixelcfestats:
fixelcfestats: [WARNING] Design matrix conditioning is poor (condition number: 110.248); model fitting may be highly influenced by noise
Of note, I have only 37 subjects (Group1=25, Group2=12). Running a correlation, I noticed a strong positive correlation (r = 0.8) between gender and ICV, which I assume is what we expect to happen.
Any comments?
Best,
Amir
UPDATE: There is no warning for FD since no logICV is included in the design matrix.
Yes, I know this collinearity is due to gender and ICV, but it is also the thing I expected to happen (males tend to have larger ICVs). My question is how to address this: leave the design matrix as it is, or remove one of the covariates (e.g., gender) and repeat the analysis? Statistically, I know that you can compare the fit of two linear models (in R for example) with an Anova, to see which model (with or without gender) fits better on your data, but have no clue whether it is feasible in this situation.
Okay, I understand. I am not aware of fit comparisons in MRtrix, sorry… Perhaps the lowest matrix condition number in fixelcfestats could be your (imperfect) criterion of selection. Hopefully MRTrix experts can advise you on that.
You can also try z-scoring your design matrix data before running fixelcfestats. Sometimes the poor matrix conditioning comes from data having vastly different scales.