[Eeglablist] paper published

Wed Dec 7 03:22:58 PST 2016

Dear Makoto,

We investigated this issue in our recent paper:

Van de Steen, F., Faes, L., Karahan, E., Songsiri, J., Valdes-Sosa, P. A., & Marinazzo, D. (2016). Critical comments on EEG sensor space dynamical connectivity analysis. Brain Topography, 1-12.

This work was conducted independently from and simultaneously as Brunner et al (2016). Both papers show how spurious connectivity on sensor space analysis can, in many cases, occur (not always though).

I'll try my very best to make it intuitively:

first of all, you should be aware that non negative (d)DTF, (r)PDC, time domain granger causality etc. all imply at least one non-negative off diagonal coefficient of the coefficient matrix (B) of the MVAR model.

To clarify: time series are modelled with a multivariate autoregressive model (MVAR) in which current values of time series are explained by a linear combination of past values of that time series itself and past values of other time series: X(t)= B1*X(t-1) + B2*X(t-2) + etc+ error where X(t) is a column vector contain the current values of all time series (X(t) = [X1(t) X2(t) ...]T). Diagonal elements of B relate past values (t-1, t-2 etc) of a time series to the current value (t) of the time series itself (e.g X1(t-1) to X1(t) )). Off diagonal elements relate past values of another time series to the current value (e.g X2(t-1) to X1(t)))

secondly, two independent time series ≠ two independent time series. what i mean is that there are different ways in which time series can be independent of one another. e.g two white noise processes vs. two independent autoregressive time series. In the former case the time series cannot be modelled with past values  of another time series nor can it be modelled with past values of that time series itself (all elements in B are zero). In the latter case, the situation is different. There you cannot model one time series with past values of the other but you can model it with past values of the time series itself (i.e. only non zero elements on diagonal of B). Now when you mix up (linear superposition) these time series (by volume conduction), the past contains information of both original time series.

since the mixing of time series is not identical (i.e. each EEG channel is a unique linear combination of sources), the inclusion of past values of another mixed time series can improve the prediction of the other mixed time series. This translates, in terms of the MVAR model, in a (some) non-negative off diagonal elements in B and thus also non negative DTF, PDC, etc. When mixing up white noise, no spurious connectivity will occur because the past wasn't informative in the first place.

You need to be aware that even though volume conduction implies instantaneously mixing of time series, the continuation of time does not unmix the past...

So basically these measures are not insensitive to volume conduction as claimed by some.

Hope it is a bit clear. Please do not hesitate to ask more questions

Kind regards,

Frederik

________________________________
Van: Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
Verzonden: woensdag 7 december 2016 03:46
Aan: Scott Makeig
CC: eeglablist at sccn.ucsd.edu
Onderwerp: Re: [Eeglablist] paper published

Dear Scott and list,

During the EEGLAB Workshop 2016, I asked this question to Tim Mullen during his lecture, without knowing this ongoing debate. Intuitively, it does not make sense to me why just linear mixing affects connectivity calculation, if dDTF or rPDC can suppress the spurious connections... Tim mentioned that there is good reason for this, but did not explain it during the lecture due to limited time. Can anyone give me an intuitive explanation why it is bad?

Makoto

On Wed, Nov 30, 2016 at 6:02 PM, Scott Makeig <smakeig at ucsd.edu<mailto:smakeig at ucsd.edu>> wrote:
Some of us noticed a paper published recently that claimed that effective connectivity measures between scalp EEG channels suffer no ill effects of volume conduction -- and immediately questioned its conclusions!  Clement Brunner of Graz mounted an effort to publish a rebuttal in the same journal, which has now appeared:

C. Brunner, M. Billinger, M. Seeber, T.R. Mullen, S Makeig, Volume conduction influences scalp-based connectivity estimates <https://sccn.ucsd.edu/~scott/pdf/brunner16.pdf> (a rebuttal). Frontiers in Computational Neuroscience,
doi:10.3389/fncom.2016.00121, 22 November 2016.

We have learned that another group is publishing a separate rebuttal ...

Scott Makeig

--
Scott Makeig, Research Scientist and Director, Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla CA 92093-0961, http://sccn.ucsd.edu/~scott

_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu<mailto:eeglablist-unsubscribe at sccn.ucsd.edu>
For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu<mailto:eeglablist-request at sccn.ucsd.edu>

--
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20161207/49771ee1/attachment.html>