Data and Noise Covariance Matrix for LCVM Analyses

💬 Support & Discussions
1

Hi everyone,

I’m new to source reconstruction analyses, so I am happy for any advice on a potentially minor topic:

I have a question regarding the data and noise covariance matrix needed to perform LCVM analyses.

I want to downsample the data heavily before performing my final analyses on single trial source time courses. Thus, I am wondering if
a) it is recommended to compute the covariance matrices on the original data with a high sampling rate or
b) if I should compute the covariance on already downsampled data so that it matches the sampling rate of the data for the final analyses.

I already compared both covariance matrices: the pattern looks similar and the matrix values are highly correlated, but the absolute values are still a bit different.

Or does it not matter at all?

  • MNE version: 1.7
  • operating system: Ubuntu 22.04

Philipp

2

Ideally, both covariance matrices should be very similar, given that the heavily downsampled data still capture all the frequency content of the signal. In practice, this will not be the case, since also a lot less samples are used to estimate the covariance.

I would recommend using the covariance of the original data since that should be closer to the “true” covariance matrix. Most of the source reconstruction methods don’t rely at all on any temporal information, so the results of downsampling EEG → localization should be equal to localization → downsampling of source signals.

1 Like
3

Great, thank you for the quick reply! Then, I will go for the original data!