Best approach for conducting source-level time-frequency analysis of ERPs

Greetings,

I'd like to perform TF analyses of ERPs in the source space, but I am not certain if these steps are the correct approach: compute the source for each trial, extract the time course for the ROI(s), apply the TF decomposition to the ROI time course, average the TF outputs for all trials within subject. Repeat for all subjects, then average TF results across conditions. Would you confirm if this is sensible?

Of course, extracting sources at the trial level also generates massive files, and I am using unconstrained sources since I only have a template brain... will I lose important information if I chose "Kernel only: shared" (within subject), or must I select "Kernel only: one per file"?

Would you have other suggestions for minimizing hardware and memory demands for this analysis?

Thank you!

Yes, this makes sense. It is all automated into one process call:
https://neuroimage.usc.edu/brainstorm/Tutorials/TimeFrequency#Scouts

will I lose important information if I chose "Kernel only: shared" (within subject), or must I select "Kernel only: one per file"?

This would produce exactly the same output, so go for the economic version.