dSPM, normalization, and workflow

Hi all,

I'm looking at within-subject changes in EEG source activity, and would like to use dSPM (constrained sources) for source estimation.

However, when reading Tutorial 27: Workflows, section 'Constrained cortical sources', I see that it is recommended that we 'Estimate sources for each average (constrained, no normalization)' when computing subject averages. From my understanding, dSPM involves normalisation (with respect to noise covariance), and I'm unsure if this means that we shouldn't be using dSPM for subject averages (and subsequently, group analysis)? I've searched the discussion forum for possible answers, and it seems that others have been performing dSPM when computing subject averages -- I'm having difficulty reconciling that with what I'm understanding from the Workflows tutorial. I'm wondering if I'm missing something here, and would really appreciate if someone can point me in the right direction..

Thank you!

Tutorial 27 suggests to apply a normalization (Z-score with respect to baseline) only once, after averaging all the acquisition runs in one subject. Averaging normalized source maps within a subject is not recommended, it leads to different statistics.

If you apply a Z-score normalization at some point in your analysis, it does not matter whether you select dSPM or non-normalized current density maps in the options. Try applying a Z-score on both, you obtain exactly the same results. So in the context of a group analysis where you normalize you data wrt a baseline, the initial choice of using dSPM or not doesn't matter in terms of the type of normalization you obtain. But selecting dSPM can lead to wrong results if you average run-level averages.

If you have only one acquisition run per subject (or if all the runs can have exactly the same conditions of acquisitions and pre-processing steps, and can share their channel file), another approach is to rely only on the dSPM normalization across subjects (no Z-score normalization before group averages). I can't tell you whether this is a good idea or not, this is not what has been selected as the recommended procedure by our source modeling and statistics experts.

For displaying run-level or subject-level averages in source space, you can still compute dSPM maps, they will be much easier to explore and interpret. But prefer not applying this normalization for the final group statistics.

@Sylvain @John_Mosher @pantazis: any additional comment?

Hi Francois,

Thank you very much for your reply. It has provide some much-needed clarification for me.