Mne_analysis Digest, Vol 113, Issue 24

✉️ Mailing List Archive (read-only)
1

We are just getting started with using this connectivity measure (wpli), so thanks for all of the help and information. I believe that a lot of the analysis code was from the MNE website, so it should be fairly familiar (I have the pseudo code and code snippet below). I have attached an image that shows what our lowest negative values (for ~30subjects) are with the wpli_debiased connectivity matrix and the percentage of connections that are negative . If the values should vary from -1 to 1, I would assume the percentage of negatives to be closer to 50% of the connections (since for every lead roi there would be a lag roi as well), but across subjects we are seeing about a 0-10% of the connections are negative. Attached are two example connectivity matrices (from Desikan-Killiany parcels) with the wpli_debiased measure as well as our code for the analysis. I will also upload the numpy matrices to the github site.

Psuedo Code:
Raw data is loaded from a resting state tSSS processed fif dataset
Make inverse operator
Make events of 1 second
Epoch data
Create dSPM (cov was created from emptyroom data - since no prestim period in resting data)
Load Labels >> Reorder labels to be L then R (vs. LRLRLR..)
Extract Labels
Create Connectivity using 3 measures (coh, pli, wpli)
Save out each conn measure as a numpy array.

2

We are just getting started with using this connectivity measure (wpli),
so thanks for all of the help and information. I believe that a lot of the
analysis code was from the MNE website, so it should be fairly familiar (I
have the pseudo code and code snippet below). I have attached an image
that shows what our lowest negative values (for ~30subjects) are with the
wpli_debiased connectivity matrix and the percentage of connections that
are negative .

Just to be clear, since it looks like these might be being used
interchangeably a bit in these discussions -- "WPLI" and "debiased WPLI2"
are expected to give different sets of values.

For WPLI (our implementation at least), the values should be strictly
between 0 and 1, with no negative values. WPLI is therefore biased
(non-zero mean for even pure noise data). Can you confirm this?

For debiased WPLI squared, the debiasing produces some small negative
values. I suspect in your original email, you might have been talking
about *debiased
wPLI* giving you small negative values, even though you mentioned using
*WPLI*...?

If the values should vary from -1 to 1, I would assume the percentage of

negatives to be closer to 50% of the connections (since for every lead roi
there would be a lag roi as well), but across subjects we are seeing about
a 0-10% of the connections are negative.

It looks like you're saying you're seeing generally small negative values
for debiased WPLI, which is to be expected. However, they should not be
uniformly spaced from -1 to 1. Rather, the mean (for pure noise) for this
measure should be near zero, with generally small values (like the negative
ones you report). Positive values thus mean "more likely to be real
connectivity". The distribution of values for data with true connectivity
should have a long right (positive) tail, which sounds like what you
describe. Can you confirm?

HTH,
Eric

P.S. Andrea, from what I see in the code, we already implement the absolute
value operation, unlike in the FieldTrip code.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170615/a648c1bc/attachment.html