[FieldTrip] LCMV giving conflicting results vs SAM and DICS

[Charidimos Tzagarakis]

Hi Eelke,
Thank you for looking into this!
The piece of code I commented out is indeed [790:836] (changing line 784
accordingly or alternatively adding a copy of 785-789 after 790).
I did try running ft_timelockanalysis with keeptrials='no' for all 3 calls
and that resulted in the same behaviour as before for LCMV (ie not the
"correct" one). I also tried it   setting keeptrials='no' in
ft_sourceanalysis at the same time (ie keeptrials='no' for both
ft_timelockanalysis and ft_sourceanalysis) which had the same outcome (I
actually had run these tests before posting but wanted to 2ble check).
I may however be confused here because there are some other possible
combinations with keeptrials since for each subject there is a total of 6
calls to functions that accept it as an input and the filter is estimated
with only 2 of them.
Do let me know if there is another combination I should run and I will do
that.
Best,
Haris

Charidimos [Haris] Tzagarakis MD, PhD, MRCPsych
University of Minnesota Dept of Neuroscience and Brain Sciences Center


On 24 October 2013 03:08, Eelke Spaak <eelke.spaak at donders.ru.nl> wrote:

> Hi Charidimos,
>
> Thanks for your elaborate e-mail. We discussed this issue in
> yesterday's FieldTrip meeting. It is very reasonable that you expect
> the three methods to produce the same results, and the code/interface
> indeed suggests that this would be the case. In other words, we need
> to clean up this part of the code to produce consistent results;
> however, this might take a while.
>
> We believe the discrepancy is caused by the fact that you compute the
> covariance while specifying cfg.keeptrials = 'yes', which is
> unnecessary while computing the filter. This results in a covariance
> matrix with dimensions trial X channel X channel, which then is
> subsequently averaged over trials during ft_sourceanalysis. Likely the
> filter is still in fact computed on the averaged covariance, as it
> should be. However, the subsequent projection of data through the
> filter will not first average the covariance, and will in essence only
> project the first trial through the filter. (This is because the dip{}
> fields will be struct arrays of dimension nTrial X 1, and in an
> assignment a(1).b = 2; a(2).b = 3; x = a; then x will be 2, so
> everything but the first element of a struct array is ignored in an
> assignment statement.)
>
> Could you try to compute the covariance with cfg.keeptrials = 'no' and
> then run LCMV again? This will still compute the covariance on the raw
> traces, not on the average. If our hunch is right, you should get the
> correct results then.
>
> As said, this is indeed messy and unclear, our apologies for that.
>
> Finally, could you tell us what part of the code you commented out to
> make the results consistent? We suspect it will be between lines 790
> and 835, but it would be helpful to know for sure.
>
> Best,
> Eelke
>
>
> On 23 October 2013 21:43, Charidimos Tzagarakis <haristz at gmail.com> wrote:
> > Hi Don,
> > Thanks for your reply. I do agree that some of these beamformers
> > are probably better suited than others to study beta desynchronisation,
> or
> > generally questions in the frequency domain. I in fact started from DICS.
> > The reason I looked at LCMV and SAM is to be sure that I get consistent
> > results (and also because my experiment can also address some additional
> > time domain questions, so I wanted to see what happens when my data goes
> > from one to the other). I therefore tried to set them up so I that the
> > results from all 3 are similar. So you are correct in pointing out that
> my
> > question is why SAM and LCMV produce different results. Based on their
> > description within the fieldtrip website and mailing list (if I have
> > interpreted these correctly) they should be treating the covariance
> matrix
> > in the same way (to get a true evoked response covariance matrix you
> would
> > normally need the extra call to ft_timelockanalysis that I show at the
> last
> > piece of code in my post). Also , if the covariance matrix is indeed the
> > issue, it seems that the change in how it is treated in the
> > ft_sourceanalysis code (based on the "hack" I described) is unique to
> LCMV
> > (none of the other beamformer options share that piece of code) and is
> > active only when LCMV is given a precomputed filter which is (I think)
> > unusual.
> > Hence my question!
> > Best,
> > Haris
> >
> > Charidimos [Haris] Tzagarakis MD, PhD, MRCPsych
> > University of Minnesota Dept of Neuroscience and Brain Sciences Center
> >
> >
> >
> > On 23 October 2013 13:31, Rojas, Don <Don.Rojas at ucdenver.edu> wrote:
> >>
> >> Haris,
> >>
> >> Sorry - In my last post, I mistakenly put DICS and the Fieldtrip
> >> implementation of SAM into the same frequency domain category. In
> Fieldtrip,
> >> SAM is a time-domain technique and is not the same as the
> implementation of
> >> SAM that has been used in the published literature for beta ERD. So,
> are you
> >> then wondering why the two time-domain approaches produce differing
> results?
> >> That probably does depend on how the covariance matrix is calculated.
> >> Although I still think it is a bad idea to use a time-domain beamformer
> on
> >> motor beta ERD/ERS.
> >>
> >> Best,
> >>
> >> Don
> >>
> >>
> >>
> >> _______________________________________________
> >> 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
> _______________________________________________
> 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/20131024/31bec2fd/attachment.html>