As results of first-level analysis of 4 out of 20 subjects, the calculated T values are all complex numbers. I tracked this issue back to standard errors being complex because of invalid covariance matrices.
Upon thorough exploration, I observed the following in an exemplar subject:
the problem disappears when I perform GLM without concatenating sessions. However, session concatenation is essential for my subsequent DCM analysis. And I don’t think there is problem with my session concatenation code because the same code worked well in another study.
the problem disappears when I perform GLM without segmentation and normalization, though there is concatenation step. However, segmentation and normalization are essential in my study.
Any insights you can provide would be greatly appreciated.