One more question: My data is 250 seconds long, and I am using the entire recording to calculate the NCM. Should I use only a portion?
It depends. How clean is the data. You absolutely should only use the cleanest piece of data you should check (visually) that it looks good the whole way with identical filter and projection settings as used when calculating. Any noise or artifacts should be avoided.
-Jeff
From: Jeff Eriksen <eriksenj at ohsu.edu<mailto:eriksenj at ohsu.edu>>
To: "mne_analysis at nmr.mgh.harvard.edu<mailto:mne_analysis at nmr.mgh.harvard.edu>" <mne_analysis at nmr.mgh.harvard.edu<mailto:mne_analysis at nmr.mgh.harvard.edu>>
Subject: [Mne_analysis] questions regarding noise-cov matrix (NVM)
I have been having some difficulties with the inverse results from my 256-channelEEG, some of which may stem from my noise-covariance matrix, and have come up with the following questions:
For MEG, the manual suggests recording some MEG without a subject in the loop ? room noise, basically. It also suggests for epilepsy work to use a segment of MEG (or EEG) that has no epileptiform activity. I am doing continuous EEG, with no discrete events to speak of, so I wonder if I should do a subject-less recording and use that as the basis of my noise-cov matrix. I would do that by applying the electrode Net to a styrofoam head covered with a saline soaked towel.
I don't think that would work (it may even create more problems). There are a couple of options you have
1 stick with a diagonal noise covariance
2 try using a different set of filtering on your EEG data to remove the brain activity but leave the other parts of noise e.g. 100Hz - 200Hz when calculating the noise covariance matrix
If you were to want to try this, you should definitely do some testing to see you are getting reasonable results (having some evoked data would help greatly for this testing). You should also definitely check the time frequency spectrum.
What is the use of the noise-cov supposed to accomplish in practical terms?
I'll leave this to experts better at explaining this than me. (someone please fill in!)
It has been suggested that I try the ?diagnoise switch if my inverse results are not matching my scalp topography (they are not, to my eye). I did that, and it seems to have helped a bit, but why? Why does the manual say that most users should never need to use this option?
Because in very general terms a full noise covariance matrix produces a much more elegant and better regularized inverse solution.
I would like to read the NCM matrix into Matlab to examine it, but I cannot understand the MNE Matlab routines (mne- and fif-) that allow me to do that ? too many undefined terms. Can someone give me an explicit set of instructions of how to look at the NCM in Matlab?
Well, could you explain what terms aren't defined? You only need one function to read the file the help output is below:
help mne_read_noise_cov
[cov] = mne_read_noise_cov(fname)
Reads a noise-covariance matrix from a fiff file
fname - The name of the file
The matlab structure is described in table 10.29 in the MNE manual.
Can I avoid the use of the NCM altogether, or set it to a diagonal identity matrix?
You can use the diagonal, though your solution won't be as well regularized as it could be.