megpalnar.m and statistical testing

Michael Wibral wibral at BIC.UNI-FRANKFURT.DE
Wed Sep 2 10:56:27 CEST 2009


Dear Fieldtrip developers & users,


I have some questions regarding planar gradiometers and averaging / statistical testing in fieldtrip and in general:

Computing the vector valued gradient of the field is a linear operation and - as such - fully exchangeable with any kind of other linear operation such as averaging over trials, subjects and computing differences between conditions.

This obviously does not hold for the pythagoras-like computation of the gradient vector's magnitude, because this is a non-linear operation.

As an example think of to sources in two subjects that are (roughly) in the same spot (sulcus) but, say, on opposing walls of this said sulcus. Therefore, the fields in these two subjects will have opposing signs. As a consequence their fileds will cancel in the average over subjects - as it should be, perhaps, because they are not the same source but represent individual variance. If I now take the planar gradient magnitude using megplanar and combineplanar in each subject these two sources will be in the same spot on the scalp and with the same sign (because magnitude is confined between 0 and +INF) and will not cancel. This is different from taking the planar gradient's magnitude after averaging - this latter planar gradient amplitude will be close to zero. So far this is not surprising as taking the magnitude is a nonlinear operation and not permuatble with linear operations.

My question is whether there any advice on where an when combineplanar should be used? Or would recommend it just for visualization purposes?

Thanks you very much for your help,
Michael





> -----Urspr√ľngliche Nachricht-----
> Von: "Eric Maris" <e.maris at DONDERS.RU.NL>
> Gesendet: 02.09.09 09:52:37
> An: FIELDTRIP at NIC.SURFNET.NL
> Betreff: Re: [FIELDTRIP] freqdescriptives/statistics (spatial clustering for planar gradiometers and magnetometers __ Neuromag)


> Hi Dahlia,
> 
> 
> My name is Eric Maris and, together with Robert Oostenveld, I implemented
> the permutation statistics in Fieldtrip. 
> 
> > 1. We're using Neuromag (Vectorview 306) data, and giving the appropriate
> > layout file (all channels). Does fieldtrip take into account which
> channels
> > are gradiometers and which magnetometers when it calculates adjacency for
> > clustering?
> > 
> > 2. How does fieldtrip calculate adjacency for channels? (time and
> frequency
> > are obvious...)
> > 
> > 3. How is the clustering performed?
> > 
> > We're getting a significant cluster but it doesn't look like a single blob
> > in the time-frequency domain (averaging over sensors) but like two very
> > distant blobs, which is very strange...
> 
> These are important points. As far as I know, you are the first to use the
> cluster-based permutation tests on Neuromag data. This means that you have
> to do some more work than the people who work with a system that only has
> axial gradiometers (like the CTF system). For axial gradiometers, Fieldtrip
> can calculate the adjacency in a straightforward way, using the positions of
> the coils in the helmet. Working with planar gradiometers and magnetometers,
> you have to define yourself the adjacency structure for the channels in your
> system (both gradiometers and magnetomers). Fieldtrip doesn't care about the
> way you define adjacency. You can give Fieldtrip any adjacency structure via
> the configuration field cfg.neighbours. (Type "help neighbourselection" to
> get info about how to define and adjacency structure.)
> 
> Defining adjacency on planar gradiometers requires some thinking. Things
> become easier if you (1) ignore the magnetometers, and (2) sum the power at
> the pairs of orthogonal planar gradiometers. This maps the data on a new set
> of channels for which adjacency is defined in a straightforward way. (Of
> course, you may also decide to retain the orientation information in your
> original planar gradiometers, but than you have to come up with a sensible
> definition of adjacency that takes this orientation information into
> account.)
> 
> Finally, a minor point, do not start with cluster-based permutation tests
> with both time and frequency resolution. If you are interested in
> oscillations, only keep the frequency resolution in you first analyses. You
> can later add the time axis if you want to know how the effect develops over
> time.
> 
> 
> Good luck,
> 
> Eric
> 
> 
> dr. Eric Maris
> Donders Institute for Brain, Cognition and Behavior
> Center for Cognition and F.C. Donders Center for Cognitive Neuroimaging
> Radboud University
> P.O. Box 9104
> 6500 HE Nijmegen
> The Netherlands
> T:+31 24 3612651
> Mobile: 06 39584581
> F:+31 24 3616066
> E: e.maris at donders.ru.nl
>  
> MSc Cognitive Neuroscience: www.ru.nl/master/cns/
> 
> 
> 
> 
> > 
> > On Tue, 1 Sep 2009 09:09:53 +0000, Jan-Mathijs Schoffelen
> > <j.schoffelen at PSY.GLA.AC.UK> wrote:
> > 
> > >Dear Dahlia,
> > >
> > >Hemant Bokil indeed uses the jackknife to obtain variance estimates of
> > >power (and coherence), see also for example his 2007 paper in
> > >J.Neurosci.Methods. More specifically, if your data is 'well-behaved',
> > >he has shown that by applying a specific correction to the power
> > >estimate, in combination of the jackknife, generates a test-statistic
> > >from a differential (i.e. contrast) power spectrum which has a
> > >standard normal distribution (i.e. the jackknife estimate of the
> > >variance is expected to be 1 and the estimate of the mean 0).
> > >Unfortunately, MEG data is hardly ever well-behaved, so we prefer to
> > >use non-parametric techniques to do statistical inference.
> > >Freqdescriptives in this respect still historically has the option of
> > >computing a jackknife estimate of the SEM of the powerspectrum/
> > >coherencespectrum, which can be used to compute a T-statistic across
> > >two conditions for example. However, I you would choose this path, you
> > >have to write some code which does this, because it is not in
> > >fieldtrip. The biascorrect option has been taken out altogether as far
> > >as I can see, (and had been designed only to correct the bias in the
> > >coherence spectra, and not in the power spectra if I remember
> > >correctly), and any reference in the documentation should be removed.
> > >Unfortunately, we did not yet have time to considerably clean up
> > >freqdescriptives, but this is quite high on the developer's to do list.
> > >The bottom line is: ignore biascorrect, and if you use the jackknife
> > >estimate of the SEM, you have to come up with some code of your own.
> > >Alternatively, you could look into freqstatistics and use
> > >cfg.statistic = 'indepsamplesT' / 'depsamplesT' if you want to do
> > >statistical inference.
> > >
> > >Best,
> > >
> > >Jan-Mathijs
> > >
> > >
> > >On 28 Aug 2009, at 06:34, Dahlia Sharon wrote:
> > >
> > >> Hi all,
> > >>
> > >> Is there a more detailed explanation of usage for the jackknife and
> > >> biascorrect options for freqdescriptives than the one in the
> > >> freqdescriptives reference
> > page(http://fieldtrip.fcdonders.nl/reference/freqdescriptives)?
> > >>
> > >> More specifically, are these options related to Bokil et al NeuroIm
> > >> 2007? How should they be employed to determine significance of
> > >> difference between conditions? (Is there somewhere a tutorial for
> > >> the use of these options analogous to the one about cluster-based
> > >> permutation testing?)
> > >>
> > >> Also, for the permutation analysis of TFRs
> > (http://fieldtrip.fcdonders.nl/tutorial/statistics?s
> > >> []=freqstatistics), if I don't want to employ the planar gradient
> > >> step (what exactly IS combineplanar? sorry I couldn't find it), can
> > >> I simply skip it and calculate the TFRs of the raw sensor data?
> > >>
> > >> Many thanks!
> > >> Dahlia.
> > >> ----------------------------------
> > >> The aim of this list is to facilitate the discussion between users
> > >> of the FieldTrip toolbox, to share experiences and to discuss new
> > >> ideas for MEG and EEG analysis.
> > >> http://listserv.surfnet.nl/archives/fieldtrip.html
> > >> http://www.ru.nl/fcdonders/fieldtrip/
> > >
> > >----------------------------------
> > >
> > >The aim of this list is to facilitate the discussion between users of
> > >the FieldTrip toolbox, to share experiences and to discuss new ideas
> > >for MEG and EEG analysis.
> > >
> > >http://listserv.surfnet.nl/archives/fieldtrip.html
> > >
> > >http://www.ru.nl/fcdonders/fieldtrip/
> > >
> > >
> > >
> > >
> > >
> > >----------------------------------
> > >The aim of this list is to facilitate the discussion between users of the
> > FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG
> > and EEG analysis. See also
> > http://listserv.surfnet.nl/archives/fieldtrip.html and
> > http://www.ru.nl/neuroimaging/fieldtrip.
> > >
> > 
> > ----------------------------------
> > The aim of this list is to facilitate the discussion between users of the
> FieldTrip
> > toolbox, to share experiences and to discuss new ideas for MEG and EEG
> analysis.
> > See also http://listserv.surfnet.nl/archives/fieldtrip.html and
> > http://www.ru.nl/neuroimaging/fieldtrip.
> 
> ----------------------------------
> The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/neuroimaging/fieldtrip.
> 



----------------------------------
The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/neuroimaging/fieldtrip.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Michael Wibral.vcf
Type: text/x-vcard
Size: 344 bytes
Desc: not available
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20090902/9972f10d/attachment-0001.vcf>


More information about the fieldtrip mailing list