Neuromag stats

[Robert Oostenveld]

Hi Thomas

To start with a remark: I suggest that you switch to
timelockstatistics/freqstatistics/sourcestatistics, they are the new
interface to randomization testing with clustering. These new
functions have the same options (and more), but most importantly:
they are easier to understand (at least that is what we hope). From
now on I will refer to the new functions, but you can do most of the
stuff also with the old clusterrandanalysis function.

On 12 Sep 2006, at 1:01, Thomas Thesen wrote:
> I am trying to run clusterrandanalysis on single-subject data that
> have been
> collected on a Neuromag system. It seems that I am running into
> problems
> with the neighbouring channel geometry.

We do have (some) experience with that for frequency and time-
frequency (power) data, but then based on the virtual planar
gradients that we can compute for the CTF system (using MEGPLANAR).

I don't think that sofar anyone ever used it for ERP data. The
approach we use for power data is
PREPROCESSING -> MEGPLANAR -> FREQANALYSIS (with keeptrials) ->
COOMBINEPLANAR -> FREQSTATISTICS
i.e. the statistic is computed after summing the power in the two
planar channels at each sensor location. The power is added for each
trial.

For ERF data, COMBINEPLANAR is using Pythagoras rule (z^2=x^2+y^2) to
compute the combined amplitude of the two planar channels at each
location. In that computation, the sign of the field is lost. I.e.
you can combine planar ERFs, but you might get into problems with
different noise bias and with the interpretation of the combined data.

For the old 122 channel Neuroscan system, there are 122 channels at
66 locations. Since the channels are already planar MEGPLANAR is not
needed. After COMBINEPLANAR, you have a 66 channel data structure.

> The epochs are already in a format that went through
> timelockanalysis and
> gives the following structure:
>
> timelock_data =
>
>            avg: [102x102 double]
>            var: [102x102 double]
>        fsample: 250
>     numsamples: [297x1 double]
>           time: [1x102 double]
>         dofvec: [1x102 double]
>          label: {1x102 cell}
>          trial: [297x102x102 double]
>         dimord: 'rpt_chan_time'
>           grad: [1x1 struct]
>            cfg: [1x1 struct]

It seems that you have 102 channels in the data. You have a 306ch
Neuroscan system, with 102 magnetometers and 2x102 planar
gradiometers. Are you using the magnetometers (which should probably
work), the combined planar channels (102+102=102, should also work)
or what?

> When trying to run the data through clusterrandanalysis using:
> ...
> It gives the following error message:
>
> Selecting and formatting the data.
> selected 102 channels
> ...
> Calculating the neighbourhood structure of the channels.
> Obtaining the gradiometer configuration from a file.
> ??? Error using ==> exist
> The first input to exist is a string.
>
> Error in ==> read_fcdc_elec at 60

The read_fcdc_elec function should not get called there actually, so
something is wrong. Please try using the stand-alone (find attached,
I will ad it to the ftp release) NEIGHBOURSELECTION function and put
the result in cfg.neighbours. Furthermore, use TIMELOCKSTATISTICS.

> There seems to be some problem with the gradfile, but I am not sure
> what
> exactly.

Oh, now I see. You specified
   cfg.gradfile = data.grad.pnt.
You should do either
   cfg.grad = structure with gradiometer definition (*)
or
   cfg.gradfile = string, file containing the gradiometer definition
or neither of them, in which case data.grad will be used. In case
(*), the correct syntax would have been cfg.grad=data.grad, i.e. the
complete structure (including pnt/ori/tra/label).

I suggest that you try with NEIGHBOURSELECTION. It still might not
work for the neuromag data (we only used it on pseudo-planar CTF
sofar), but that can be extended. Once the initial neighbourhood
definition is set up, we can work out the other details about the
neuromag data (i.e. planar ERFs).

best regards,
Robert

-------------- next part --------------
A non-text attachment was scrubbed...
Name: neighbourselection.m
Type: application/octet-stream
Size: 4770 bytes
Desc: not available
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20060913/71f67a54/attachment-0002.obj>
-------------- next part --------------