Thanks, if I understand correctly, compute_covariance() takes the resting state (here my baseline) for computing the noise covariance. I followed the tutorial you linked above and the whitened data now looks good for Example 2.
I read that using baseline for the noise covariance suppresses activities resembling baseline in the source estimate. Does this indicate that some activities (oscillations?) in the baseline are causing the problem? Is it generally recommended to use resting state rather than empty room for noise covariance if I want to look at evoked data?
To complicate the question even more, the response I’m interested in is not really the onset but rather words within sentences (participants with problematic onset source reconstruction also have “bad” source reconstruction with these words). The baseline I’m using to calculate the evoked response to these words are response immediately before their onset, which are not technically resting states.
Example 1:
Example 2:
