[FieldTrip] Need help with cluster-based permutation test with 3 groups of subjects
Stephen Politzer-Ahles
politzerahless at gmail.com
Fri Apr 14 03:27:24 CEST 2017
Hello Jose,
I haven't looked closely at your code, but it sounds like this is working
exactly as intended. A significant F-test doesn't mean that there must be
significant differences between all pairs of groups; hypothetically, an
F-test comparing 100 groups could be significant even if only 99 and 100
are significantly different from each other. This isn't something specific
to the cluster-based permutation test, this is just how F-tests work in
general.
As for getting different cluster extents with the 3-level F-test vs. the
pairwise t-test, this is also normal; you're comparing different things so
it's natural that the test statistics will come out slightly different.
Best,
Steve
---
Stephen Politzer-Ahles
The Hong Kong Polytechnic University
Department of Chinese and Bilingual Studies
http://www.mypolyuweb.hk/~sjpolit/
>
>
> Message: 1
> Date: Tue, 11 Apr 2017 12:08:44 +0200
> From: Jos? Antonio Uriguen Garaizabal <jose.uriguen at deusto.es>
> To: FieldTrip discussion list <fieldtrip at science.ru.nl>
> Subject: [FieldTrip] Need help with cluster-based permutation test
> with 3 groups of subjects
> Message-ID:
> <
CAMMf7X15byC6p3gqaie6wEuKekV2krZ0++keLLUPZ3pn2h-PAA at mail.gmail.com>
> Content-Type: text/plain; charset="utf-8"
>
> Dear all
>
>
> My name is Toni and I am working on EEG signal processing at University of
> Deusto, Bilbao, Spain. More specifically, right now trying to
> apply cluster-based permutation testing to determine whether there exist
> differences among 3 groups of subjects.
>
>
> So, for my testing, subject groups are E, EN and N, the 2 former being
> different types of patients and the latter being a control group. I am
> forming clusters based on proximity (in space) and each subject is
> characterized by a matrix of values that vary in 2D (per channel and
> another variable related but not equal to frequency).
>
> By means of a T-statistic (indepsamplesT) I can find differences in
between
> E and EN, but I find no clusters/differences in between E and N or EN and
> N, even though hypothetically (I think) they should exist. By means of an
> F-statistic (indepsamplesF) I can find differences among the 3 groups at
> the same time, then also between E and EN and no differences between E and
> N or EN and N...
>
> Am I missing something? I do not understand why the might exist a
> significant cluster when comparing all 3 groups that does not exist in the
> 1vs1 comparisons, even though the cluster is not the same I obtain when I
> compare E vs EN...
>
>
> For additional information, I attach how I run the test:
>
> cfg.method = 'distance';
> cfg.neighbourdist = 1;
> cfg.elec = ft_datatype_sens(subjEN{1}.elec);
> neighbours = ft_prepare_neighbours(cfg);
>
> cfg = [];
> cfg.channel = {'EEG'};
> cfg.latency = 'all';
> cfg.method = 'montecarlo';
> cfg.statistic = 'indepsamplesT';
> cfg.correctm = 'cluster';
> cfg.clusteralpha = 0.05;
> cfg.clusterstatistic = 'maxsum';
>
> cfg.minnbchan = 2;
>
> cfg.neighbours = neighbours; % same as defined for the
> between-trials experiment
> cfg.tail = 0;
> cfg.clustertail = 0;
> cfg.alpha = 0.025;
> cfg.numrandomization = 5000;
>
> nsubj = size(subjE,2)+size(subjEN,2);
> design = zeros(2,nsubj);
> design(1,:) = [1:nsubj/2 1:nsubj/2];
> design(2,1:size(subjE,2)) = 1;
> design(2,size(subjE,2)+1:nsubj) = 2;
>
> cfg.design = design;
> cfg.ivar = 2;
>
> [stat] = ft_timelockstatistics(cfg, subjEN{:}, subjE{:});
>
>
> When I compare the 3 groups
>
> cfg.method = 'distance';
> cfg.neighbourdist = 1;
> cfg.elec = ft_datatype_sens(subjEN{1}.elec);
> neighbours = ft_prepare_neighbours(cfg);
>
> cfg = [];
> cfg.channel = {'EEG'};
> cfg.latency = 'all';
> cfg.method = 'montecarlo';
> cfg.statistic = 'indepsamplesF'; %F-statistic
> cfg.correctm = 'cluster';
> cfg.clusteralpha = 0.05;
> cfg.clusterstatistic = 'maxsum';
>
> cfg.minnbchan = 2;
>
> cfg.neighbours = neighbours; % same as defined for the
> between-trials experiment
> cfg.tail = 1; %One sided
> cfg.clustertail = 1;
> cfg.alpha = 0.025;
> cfg.numrandomization = 5000;
>
> nsubj1 = size(subjE,2);
> nsubj2 = size(subjEN,2);
> nsubj3 = size(subjN,2);
> nsubj = nsubj1+nsubj2+nsubj3;
>
> design = zeros(2,nsubj);
> design(1,:) = [1:nsubj1 1:nsubj2 1:nsubj3];
> design(2,1:nsubj1) = 1;
> design(2,nsubj1+1:nsubj1+nsubj2) = 2;
> design(2,nsubj1+nsubj2+1:nsubj) = 3;
>
> cfg.design = design;
> cfg.ivar = 2; %2nd row is independent var
>
> [stat] = ft_timelockstatistics(cfg, subjE{:}, subjEN{:}, subjN{:});
>
>
> Then, subjE, subjN and subjEN are like this:
>
> Eall = subjE{:}
>
> Eall =
>
> struct with fields:
>
> label: {1?18 cell}
> fsample: 200
> elec: [1?1 struct]
> trial: {[18?115 double]}
> time: {[1?115 double]}
> cfg: [1?1 struct]
>
>
> Thanks in advance
>
> --
> *Jose Antonio Urig?en*
> PostDoc at Deustotech-LIFE (eVIDA Research Group)
> [image: Deusto] <http://www.deusto.es/>
> Universidad de Deusto / Deustuko Unibertsitatea / University of Deusto
> Facultad de Ingenier?a, 4? Planta
> Avda. Universidades 24. 48007 Bilbao
> Tel. +34 94 413 90 00 / Mov. +34 656 711 643
> Ext. 2980
> jose.uriguen at deusto.es <sanchez.alain at deusto.es>
> Web: evida.deusto.es
> *www.deusto.es <http://www.deusto.es/>*
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL: <
http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20170411/7a169976/attachment-0001.html
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20170414/eea51b3a/attachment-0002.html>
More information about the fieldtrip
mailing list