[FieldTrip] fieldtrip Digest, Vol 55, Issue 2

Nick Ketz nick.ketz at gmail.com
Tue Jun 2 17:53:42 CEST 2015

Mathis, you have hit on a much larger problem.  Tzvetan is right to 
suggest calculating each grid-point's connectivity with all other 
grid-points, which gives you a connectivity topography map for each 
grid-point.  The next logical step, which I've been struggling with, is 
to somehow cluster these topographies together, giving you clusters of 
neighboring grid-point topographies that are similar.  Is there any work 
being done to solve this problem?

I have been struggling with this problem for some time, that is: how to 
do cluster based permutation statistics in chan x chan bins of 
connectivity  measures.  Ideally I want to do this in a full chan x chan 
x freq x time cluster analysis.  I've searched on the list for related 
topics; posted, probably incoherently, my own issues; and generally 
scanned the list and the literature for any sort of relevant topics 
trying to broach this problem, but am still stuck.

The standard 3d chan x freq x time clustering uses a neighborhood 
structure to define what channels are considered neighbors, and uses 
numeric adjacency to cluster time and frequency. With chan x chan data 
however, there unfortunately isn't an intuitive way to think about 
neighborhood for the pairwise connectivity.  There are neighbors 
surrounding the reference grid-point, and there are neighbors in each 
grid-point's connectivity topography, i.e. neighboring grid-points that 
have similar/significant connectivity values.  How do you combine the 
two in a sensible way?  And even more challenging how do you do 
statistics on them?

I have been considering several approaches, including multi-dimensional 
scaling of these grid-point connectivity topographies, or just a 
correlation between topographies, and then using those values to select 
which reference grid-points to consider as similar enough to average 
together.  This then gives you a reference cluster of grid-points and 
allows you to do standard cluster permutation statistics on the average 
of the reference grid-points.  How exactly to determine which reference 
grid-points are 'similar enough' to be consider clusters is the tricky 
part, i.e. how can you statistically test if they should be considered a 
cluster or not?

I've been working in isolation on this problem for some time, and would 
love a discussion on the topic.  I'm hoping this message spurs other 
list lurkers to at least commiserate with me, but if possible to also 
post whatever approaches they have taken to solve the problem of 
clustering pairwise connectivity data.


On 6/2/15 4:00 AM, fieldtrip-request at science.ru.nl wrote:
> Hi Mathis, alternatively you could use ft_sourcestatistics. For this 
> you should reduce your 600 x 600 x freq matrix to freq of interest 
> first. Next, the resulting 600 x 600 matrix you?d reduce to 600 x 1 
> where each grid point has a mean connectivity value to all possible 
> grid points. >From then on you could use ft_sourcestatistics and 
> ft_sourceinterpolate to visualize. Good luck, Tzvetan
>> >Dear Fieldtrip Users,
>> >
>> >I am a first-year PhD student and have been lurking here for a long time (finding lots of useful answers), it's about time I ask my first question:
>> >
>> >I am trying to compare connectivity in source space (~25 subjects, between two conditions, ~600 virtual channels on a 1.5cm grid) using ft_freqstatistics with a cluster-based permutation test (see code below).
>> >
>> >The original input datasets have the dimensions: chan x chan x freq. I tried restructuring them to chancmb x freq (seehttp://mailman.science.ru.nl/pipermail/fieldtrip/2014-February/007620.html) before doing ft_freqstatistics, but the function then takes forever appending the (admittedly big) datasets.
>> >
>> >I then tried feeding them to ft_freqstatistics without restructuring, which didn't throw any errors. However, I noticed that the resulting clusters consist of adjacent cells in the chan x chan matrix, which doesn't make much sense since channels in adjacent cells are not necessarily spatially adjacent. I suspect that ft_freqstatistic assumes the 3-D input to contain a temporal dimension and therefore tries to build temporally adjacent clusters. Could any of the fieldtrip developers comment on this?
>> >Does anyone in the FT-community have any advice on how to statistically evaluate connectivity data using the permutation approach?
>> >
>> >Thanks for any input, best wishes,
>> >Mathis Kaiser

More information about the fieldtrip mailing list