[FieldTrip] Granger causality - parametric and non-parametric calculation

[Alik Widge]

You may also find this informative:
http://www.sciencedirect.com/science/article/pii/S1053811915003316

Disclaimer: I have not personally tried this yet, although it's on our list
for a near future.

Alik Widge
alik.widge at gmail.com
(206) 866-5435


On Thu, Oct 29, 2015 at 2:49 AM, Schoffelen, J.M. (Jan Mathijs) <
jan.schoffelen at donders.ru.nl> wrote:

> Hi Daniel
>
> > I am currently computing the Granger causality between LFP signals from
> different brain regions. As it turns out, this is a more challenging
> endeavour than I had previously thought.
>
> Has anyone said to you it would not be challenging ;o)?
>
> > While testing parametric and non-parametric approaches to computing
> Granger causality, I came across some behaviour which seems odd to me. I
> hope someone can shed light on these issues:
> >
> > 1. Using the parametric approach (ft_mvaranalysis -> ft_freqanalysis ->
> ft_connectivityanalysis) I found that ft_freqanalysis can’t deal with MVAR
> data that contains a trial dimension. The first error is actually that it
> can’t find the field ‘label’, and this is true: if you run ft_mvaranalysis
> with cfg.keeptrials = ‘yes’, the resulting structure lacks the ‘label’
> field. However, if I manually copy the ‘label’ field from the original data
> to the MVAR data, ft_freqanalysis stops at some point where it uses an
> array that lacks one dimension.
>
> OK, this could (or could not) be a general issue with the code. The part
> of FieldTrip that deals with AR-modelling is quite old, and it could be
> that more recent changes in other parts of the code (e.g. concerning with
> data bookkeeping) destroyed some of the functionality. None of the
> developers are actively using AR-models in their daily research nowadays,
> so it could be that this code is a bit stale. This being said, in general I
> don’t think it is a good idea to fit an AR-model to a single trial worth of
> data. This would probably lead to not so meaningful estimates. It would be
> more meaningful to use a jackknife approach, where a variance across trials
> (or some measure that quantifies the extent to which a single trial differs
> from the rest) can be obtained with a leave-one-out approach.
>
> > 2. Using the non-parametric approach (ft_freqanalysis ->
> ft_connectivityanalysis) I stumbled across the problem that the
> implementation of Wilson’s algorithm that computes the factorisation of the
> spectral density matrix doesn’t allow non-integer frequencies nor non-equal
> distances between frequencies. Is this an inherent limitation of this
> algorithm?
>
> Yes.
>
> > 3. Again for the non-parametric approach: If I use trial-resolved FREQ
> data for ft_connectivityanalysis, the trial dimension is lost at line 389
> if I use data contain a ‘fourierspctrm’ field. If I use data containing a
> ‘powspctrm’ field, it takes literally hours at line 392 to ‘fix’ the CSD at
> line 585 in ft_checkdata (which uses ‘fixcsd’) for every trial. This is not
> resolved by using FREQ data which has both a ‘powspctrm’ as well as a
> ‘crsspctrm’ field.
>
> The input into ft_connectivityanalysis should be a frequency domain data
> structure containing either a ‘fourierspctrm’ (obtained with
> cfg.output=‘fourier’), or a crsspctrm/powspctrm (obtained with
> cfg.output=‘powandcsd’). Specifying cfg.output=‘pow’ does not make sense. I
> would expect it to crash, if it doesn’t we should consider making the error
> handling more strict. Also, for Wilson’s algorithm, the estimate of the
> cross-spectral density needs to be somewhat robust, and a single trial
> estimate in my opinion does not really make sense. The step at which the
> fourierspctrm representation is converted into the cross-spectrum ‘knows’
> this, and kicks out the trial dimension. The ‘fixcsd’ step in ft_checkdata
> can in deed be notoriously slow in some cases, which can be prevented to
> start from the right format of the data to begin with (as per
> ft_freqanalysis).
>
> > I’m not really sure whether I need trial-resolved Granger-causality, but
> in theory this should be possible. Of course, I could work around this
> issue by just using 1-trial FREQ data for ft_connectivityanalysis.
>
> Please do.
>
>
> Best,
> Jan-Mathijs
>
>
>
> >
> > Thanks in advance for your ideas!
> >
> > All the best,
> >
> > Daniel
> > --
> > Daniel Hähnke
> > PhD student
> >
> > Technische Universität München
> > Institute of Neuroscience
> > Translational NeuroCognition Laboratory
> > Biedersteiner Straße 29, Bau 601
> > 80802 Munich
> > Germany
> >
> > Email: daniel.haehnke at tum.de
> > Phone: +49 89 4140 3356
> >
> >
> > _______________________________________________
> > fieldtrip mailing list
> > fieldtrip at donders.ru.nl
> > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>
>
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20151029/c58687c1/attachment.html>