Noise covariance matrix for several runs

Support & Discussions
1

Hi,

I have a general question about the ideal way to use noise covariance matrices when dealing with several runs per participant:

We are calculating them on the baseline periods of our epochs, to include subject related noise. So far so good…
But that means that the weighting of different channels can vary between runs, depending on the noise, correct?
So what’s the recommended way if you want to compare conditions between the runs:
Is it still valid to calculate the stcs (and eventually waveforms) the respective noise cov from each run?
Or should you use one noise covariance matrix (either from one run, or combine all runs to create one common matrix)?

Thanks in advance!

Laura

PS: I don’t think it’s relevant here, but I am using MNE 1.11 on Ubuntu 24

2

Hi Laura,

Using a covariance matrix estimate for each run separately is technically valid, however it can make comparisons between runs (or conditions split across runs if that your experimental design) less straightforward, because differences in the inverse solutions may partly reflect differences in the noise estimates rather than neural activity.

I suggest computing a single noise covariance matrix per subject, you could e.g. combine the baseline periods across all runs. This gives you more stable noise covariance estimates and ensures you are not introducing a bias in whitening, which is important if you plan to compare or average source estimates.

Two important questions here (that might justify per-run noise covariances):

  1. Do you plan to analyse the runs separately?

  2. Do you believe there is a systematic difference between runs regarding noise levels?

Hope this helps.

Cheers,

Carina

3

Hi Carina,

Thanks for your reply. That matches our thoughts on that topic as well.

In the experiments I had in mind, the analysis should happen separately (because the conditions/tasks vary), but it would also be interesting to compare some conditions across runs.

I would not expect systematic differences in noise level.

So it sounds like a common covariance matrix is the best option is this case?
But probably also if we had runs with the same conditions that we plan to combine for source analysis, correct?

Best,

Laura