using cfg.correctm = 'cluster' and related memory issues.

[Eric Maris]

Dear Sameer,



> I am running into memory issues while trying to do non-parametric
> statistical testing for coherence differences for my Neuromag 306 data.
>
>   The statement causing overflow is line 277 in statistics_montecarlo:
> statrand = zeros(size(statobs,1), size(res,1));
> This would be a matrix of size 50631525 X 100 corresponding to some 40 GB!
> (Am I correct in assuming that the 100 is from cfg.numrandomization)?
>
> This statement is evoked if cfg.correctm = 'cluster';
> I could use cfg.correctm = max but I dont think this will do the right
> analysis.
>
> The one thing I can think of is changing cfg.channelcmb to try and reduce
> the number of combinations tried (but is not very preferable atleast right
> now). Is there anything else that comes to mind?
>
> The cfg I am using and other info is added below.


Using Fieldtrip for nonparametric statistical testing of the complete set of
between-channel coherences is one step too far. As a first step, one should
start comparing coherences with a single reference channel. I have described
this methodology in a recent paper in the Journal of Neuroscience Methods
(Maris et al., 2007).

Unfortunately, there is not a tutorial for this type of analysis, and the
help documentation does not bring you very far. This job is on my list, but
I have not yet found the time.

The bottom line is that testing coherence differences involving a reference
channel is certainly feasible, but the documentation is lacking,

Kind regards,

Eric Maris

dr. Eric Maris
NICI/Biological Psychology and
F.C. Donders Center for Cognitive NeuroImaging
University of Nijmegen
P.O. Box 9104
6500 HE Nijmegen
The Netherlands
T:+31 24 3612651 (NICI)
T:+31 24 3610754 (FCDC)
F:+31 24 3616066 (NICI)
E: maris at nici.ru.nl
MSc Cognitive Neuroscience :www.ru.nl/master/cns/

>
> Thank you for your time and help.
> sameer
>
> >>cfg = [];
> cfg.channelcmb = {'MEG' 'MEG'};
> cfg.method = 'montecarlo';
> cfg.statistic = 'indepsamplesZcoh';
> cfg.clusteralpha = 0.05;
> cfg.clusterstatistic = 'maxsum';
> cfg.clusterthreshold = 'nonparametric';
> cfg.minnbchan = 2;
> cfg.tail = 0;
> cfg.clustertail = 0;
> cfg.alpha = 0.05;
> cfg.numrandomization = 100; % make it larger for better testing
> cfg.precision          = 'single'; % experimenting to save memory
> cfg.label = CSD_LL_200_400.label ;
> cfg.channelcmb = CSD_LL_200_400.labelcmb;
>
>
> [statistics] = freqstatistics(cfg, CSD_LL_200_400 , CSD_RR_200_400);
>
> Note: I use Cross Spectral density instead of coherence as coherence does
> not keep trials.
>
> At the debug prompt after the error I have:
>
> K>> whos
>    Name                  Size                   Bytes  Class
> Attributes
>
>    Nrand                 1x1                        8  double
>    cfg                   1x1                 60159588  struct
>    dat              326655x127              331881480  single
> complex
>    design                1x127                   1016  double
>    res                 100x1x127               101600  double
>    statfull              1x1                405052324  struct
>    statfun               1x1                       16  function_handle
>    statobs        50631525x1                405052200  double
>    time_eval             1x1                        8  double
>    time_pre              1x1                        8  double
>
> ----------------------------------
> 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/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/fcdonders/fieldtrip.