[FieldTrip] Problems with permutation based cluster analysis (ft_freqstatistics)

Schoffelen, J.M. (Jan Mathijs) jan.schoffelen at donders.ru.nl
Thu Oct 19 08:19:35 CEST 2017


Hi Alea,

I think that it all looks OK. I guess that the P5-PO1-O1 cluster is spatially disjoint from the bulk of the suprathreshold data points. Don’t be mislead by the topographical image, where both the contour lines, and the 2D placement of the electrodes might reflect a picture that is different from what you would expect.

Note, that providing cirange as extra information is not the most informative. Better would have been to show posclusterslabelmat. This data field provides for each sample, whether and to which ‘cluster’ it belongs.

Best wishes,
Jan-Mathijs


J.M.Schoffelen, MD PhD
Senior Researcher, VIDI-fellow - PI, language in interaction
Telephone: +31-24-3614793
Physical location: room 00.028
Donders Centre for Cognitive Neuroimaging, Nijmegen, The Netherlands


On 18 Oct 2017, at 23:05, Alea Skwara <acskwara at ucdavis.edu<mailto:acskwara at ucdavis.edu>> wrote:

Hi All,

I have been conducting a permutation based cluster analysis using the method = 'montecarlo' correctm= 'cluster' options in ft_freqstatistics to examine longitudinal changes in spectral power. Everything runs just fine and it identifies clusters successfully.

However, when I examine the results of the analysis, some of the stats strike me as strange. Electrodes that have the highest F values (I am comparing across 3 timepoints) do not test as significant, while electrodes with lower F values do. At first I thought this was because I had a single "hot" electrode that was not part of a larger cluster, but looking more closely at the stats, it appears that this is not the case.

I have tried double-checking my neighbor definition method and electrode locations to assure that it is not an issue with neighbors being incorrectly registered. I have also tried adjusting the minimum cluster size using the minnbchan option to ensure it was not an issue of a cluster not reaching the minimum extent. Neither of these options seem to have any impact on the result.

At this point I'm at a loss. Is there something I'm misunderstanding about the algorithm?

Code below and plot below, channel-by-channel stats attached.

Thanks!
Alea

One funky thing to note about this code is that I am inputting average values into the analysis (averaged across time as participants have slightly differing lengths of data, and within bands as I am running this analysis on IAF-based bands, thus "7" is a dummy code for "mean beta band power", not 7 Hz)


%% THREE CONDITIONS
 %Set up the analysis cfg for 3 conditions (pre- to mid- to post-)
cfg = [];
cfg.channel          = {'all'};
cfg.latency          = 'all';
cfg.frequency        = [7 7]; %frequency band indicator
cfg.avgoverfreq      = 'no'; %working on bands not frequencies, so nothing to average
cfg.method           = 'montecarlo';
cfg.statistic        = 'depsamplesFmultivariate'; %three time points
cfg.correctm         = 'cluster';
cfg.clusteralpha     = 0.1;  %change to .1 because we're computing over averaged bands and time (losing power)
cfg.clusterstatistic = 'maxsum';
cfg.minnbchan        = 2;
cfg.tail             = 1;  %this has to be 1 for depsamplesF, 0 otherwise
cfg.clustertail      = 1;  %this has to be 1 for depsamplesF, 0 otherwise
cfg.alpha            = 0.05; %this has to be 0.05 for depsamplesF, 0.025 otherwise
cfg.numrandomization = 10000; %to get a min p-value of 0.0001 to account for 500 tests and correction by Bonferroni method.
%specify which electrodes can form clusters
cfg_neighb.method = 'distance';
cfg.neighbours = ft_prepare_neighbours(cfg_neighb, r1t1r_struct);
%cfg.neighbours = cfgManish.neighbours;
%cfg = getNeighbours(cfg); % using fixed neighbours
cfg.layout = layout';


% Set up the design matrixfor 3 conditions
subj = 27; %will differ based on condition

design = zeros(2,3*subj);
for isub = 1:1:subj
    design(1,isub) = isub;
    design(1,subj+isub) = isub;
    design(1,2*subj+isub) = isub;
end
design(2,1:subj) = 1;
design(2,subj+1:2*subj) = 2;
design(2,2*subj+1:3*subj) = 3;

cfg.design   = design;
cfg.uvar     = 1;
cfg.ivar     = 2;


% Run the analysis
[r1r_stat_beta] = ft_freqstatistics(cfg, r1t1r_struct, r1t2r_struct, r1t3r_struct);


%Plot the results
cfg1 = [];
cfg1.alpha  = 0.1;
cfg1.parameter = 'stat';
%cfg1.zlim   = [0 18.1];
cfg1.layout = layout';
cfg1.marker = 'o';
cfg1.subplotsize = [1 1];
cfg1.saveaspng = 'r1r_betacluster';

ft_clusterplot(cfg1, r1r_stat_beta);


***Note non-significant hot spot over left parietal***



<r1r_betacluster_fig1.png>


Alea C. Skwara
Graduate Student | Saron Lab
http://mindbrain.ucdavis.edu/labs/Saron

Center for Mind and Brain | University of California, Davis
267 Cousteau Place | Davis CA 95616
Cell: (828) 273-8595







<clusterstats_beta.xlsx>_______________________________________________
fieldtrip mailing list
fieldtrip at donders.ru.nl<mailto:fieldtrip at donders.ru.nl>
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20171019/54ba2185/attachment-0002.html>


More information about the fieldtrip mailing list