If I compute the noise covariance on cleaned data (with bad electrodes removed), can I still estimate the sources (Minimum Norm Imaging (dSPM, unconstrained) ) using all electrodes in the source estimation step? If I mark an electrode as bad in even just one trial, Brainstorm excludes it from source estimation. As a result, I sometimes end up using only a small subset of electrodes (e.g., 20 out of 64), even though many of those electrodes were good in other trials.
Let me offer some thoughts and suggestions (with some questions remaining), though my experience is with MEG and not EEG.
Your questions applies separately to the noise covariance computation, and then for source estimation.
If you want to have all channels in your source estimation, for sure you'll need them all in the noise covariance. Is Brainstorm smart enough to use only the good trials on a per channel basis when computing this covariance? I don't recall but I think it might. It would make things easier and it would be valid to apply a full covariance to fewer channels, if Brainstorm supports it, selecting the correct rows and columns.
Otherwise, and for the source estimation step, I don't think it would be possible in Brainstorm to use a kernel with all channels to apply to a trial with some marked bad. In that case, you have the option of computing the kernel separately for each group of trials with different sets of good channels. This should be fine with minimum-norm-type kernels, where the data covariance is not used (unlike beamforming for example).
Another option would be to use interpolation to fill in the data of bad channels in those few trials. Does Brainstorm offers interpolation of bad channels?
Cheers