[FieldTrip] MVAR analysis on rank deficient EEG data (due to removal of ICA components)

Vink-5, J.J. J.J.Vink-5 at umcutrecht.nl
Wed Aug 15 09:24:28 CEST 2018

Dear Matti,

Thank you very much for the advice. It is much appreciated. 



Van: fieldtrip [fieldtrip-bounces at science.ru.nl] namens Matti Stenroos [matti.stenroos at aalto.fi]
Verzonden: dinsdag 14 augustus 2018 15:32
Aan: fieldtrip at science.ru.nl
Onderwerp: Re: [FieldTrip] MVAR analysis on rank deficient EEG data (due to removal of ICA components)

Dear Jord,

Indeed, if you drop components (= time-series & topography) away using
ICA, the rank of your data will go down. Matlab rank routine is,
however, not very good tool for assessing rank --- for example, even
mixing double and single precision floating point numbers in your
decomposition process may yield "full numerical rank", even though the
real rank is lower.

In my opinion, the easiest way to assess the rank is to use svd: just
take a measurement (or noise) covariance matrix C, do s = svd(C), and
plot s or log10(s). You should see a clear drop in the svd spectrum at
the point, when the "true rank" is reached. Then you can project the
rest out, and after that, also the rank function should show correct rank.


On 2018-08-10 11:14, Vink-5, J.J. wrote:
> Dear fieldtrip community,
> I'm currently working with 60 channel sensor space EEG data, which has
> been preprocessed using ICA, resulting in a rank deficient dataset.
> However, when I'm trying to evaluate the rank of my data using Matlab's
> rank function, it tells me that the data is full rank, which is somewhat
> confusing.
> The data clearly is not full rank, because when I want to perform an
> MVAR analysis it either tells me the data is singular or not positive
> definite. I read up a lot on rank deficiency and how this is introduced
> through ICA. My current understanding is that the reconstruction of the
> full 60 channel EEG data after removal of ICA components reduces the
> rank due to linear dependencies between the original channels and the
> reconstruction of the data without the removed ICA components.
> Consequently, people have suggested to remove channels in order to
> restore data rank. However, how do I know which channels contain these
> linear dependencies? Simply removing channels does not resolve this issue.
> Performing PCA prior to MVAR analysis could potentially resolve this.
> Does anybody know whether you can reconstruct sensor space directed
> connectivity from a PCA-based MVAR model?
> Best,
> Jord
> ------------------------------------------------------------------------
> /De informatie opgenomen in dit bericht kan vertrouwelijk zijn en is
> uitsluitend bestemd voor de geadresseerde. Indien u dit bericht
> onterecht ontvangt, wordt u verzocht de inhoud niet te gebruiken en de
> afzender direct te informeren door het bericht te retourneren. Het
> Universitair Medisch Centrum Utrecht is een publiekrechtelijke
> rechtspersoon in de zin van de W.H.W. (Wet Hoger Onderwijs en
> Wetenschappelijk Onderzoek) en staat geregistreerd bij de Kamer van
> Koophandel voor Midden-Nederland onder nr. 30244197. /
> /Denk s.v.p aan het milieu voor u deze e-mail afdrukt. /
> ------------------------------------------------------------------------
> /This message may contain confidential information and is intended
> exclusively for the addressee. If you receive this message
> unintentionally, please do not use the contents but notify the sender
> immediately by return e-mail. University Medical Center Utrecht is a
> legal person by public law and is registered at the Chamber of Commerce
> for Midden-Nederland under no. 30244197. /
> /Please consider the environment before printing this e-mail. /
> _______________________________________________
> fieldtrip mailing list
> https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> https://doi.org/10.1371/journal.pcbi.1002202
fieldtrip mailing list

More information about the fieldtrip mailing list