Hi Marina
First off, thanks for the compliment of assuming that my little presentation
is useful. I'm glad it at least has sparked some questions if nothing more.
Second, although it is not explicitly clear in that presentation, the final
section (pages 42 onward) are basically conjecture on my part. I did make
this clear when I delivered that talk in person, but it didn't occur to me
to write that in the slides. I do not know what would be a statistically
sound method to perform statistics on the when and where questions of MNE
per subject or for groups. I considered the ideas presented from page 42
onwards as potential methods that need careful working out and future study.
Caveats aside, let me try to address your questions.
1. *"In the slides (page 46) it mentions that one first computes the MNE
signal and then computes a kind of dSPM. My limited understanding is that
the value "sum(s_i)" would come from the mne*.stc file. Since, I notice that
the MNE estimates are sums of squared dipole component strengths. Is that
so?"*
Yes. The s values in the equation from slide 46 are the dipole current
strengths from each direction (x, y and z) which should be available in the
stc file - but you'd need to check the MNE manual to be sure. If you only
have the sum of the squared dipole strengths, that's ok too if you want to
use the equation I wrote there, since it just saves you the step of doing
that summation yourself.
2. *"My next question is then how to calculate sum(var(s_i))? My
understanding is that I will need to calculate this by hand using the noise
covariance matrix and the inverse operator, unless these values are stored
somewhere."*
Yes, you will need to calculate it by hand, but how you calculate it is an
open question. If you believe that the noise covariance matrix method is the
way to go, then you can use that. There are some hidden assumptions in
there, but you have to commit to something, so perhaps that isn't a bad
option. I personally believe that the only noise covariance matrix that
makes any sense is the one based on the empty room measurement, but I will
caution that others smarter than me do not agree on this point. The noice
covariance and inverse operators are stored in MNE files. Check the manual
for how to grab them in matlab. Anyway, the important point is that you are
summing the variances across directions and across subjects. However you
achieve this is up to you.
3. *"A final question (and perhaps the most ignorant one), is why calculate
group dSPMs vs doing an ANOVA on the single-subject MNE estimates (or even
on contrasts of MNE estimates)?"*
This is interesting. The SNR that you achieve in your experiments will
depend on the details. If you have a low SNR, then group level statistics
may help to tease apart a subtle effect. That is one game that fMRI people
sometimes play, so why can't we? Another answer could be this: perhaps your
question isn't about individuals as much as about groups. If you are more
interested in individual responses then definitely there isn't much point in
doing a group dSPM. A third answer: perhaps you are happier using the
assumptions inherent in the group dSPM method versus the assumptions
inherent in the ANOVA method.
Hopefully I have conveyed a sense of confusion here. There are no clear cut
answers (that I have seen anyway) on how these computations should be done.
The suggestions I put in those slides are meant as guideposts for the
adventurous of heart to try out new directions. Keep in mind that if you do
use those ideas, you'll need to justify them, which is another project in
itself.
Good luck, and I encourage you to share with me and the group whatever you
settle on.
Daniel Goldenholz MD, PhD