[FieldTrip] FW: 2x2 ANOVA at source level using cluster based permutation

[Maris, E.G.G. (Eric)]

Hi Philip,

Here is some advice:


1.      If you have a well-founded expectation for the direction of the effect (A>B, I>II, AI-BI>AII-BII) perform a one-tailed test. There is no sense in going back to a two-tailed test.

2.      Your clusteralpha=0.01 is more stringent than what most users would choose. I always go for 0.05 and never touch it (not playing around to find a value that gives me a “significant” result). (These high thresholds 0.01 and smaller come from Gaussian Random Field theory, and are not required for permutation inference.)

3.      Did you analyze your data at the single participant level (using trials as your replications/degrees-of-freedom)? To be hopeful for a significant result over participants, you need at least a few participants that show the effect at the single participant level and all over approximately the same area/frequency band/time interval. In addition, the other participants may not contradict this pattern by showing an opposite effect over the same area/frequency band/time interval.

Good luck,
Eric Maris



Van: philip Joadavi <p.joadavi at gmail.com>
Datum: dinsdag 28 maart 2023 om 14:43
Aan: "Maris, E.G.G. (Eric)" <e.maris at donders.ru.nl>, FieldTrip discussion list <fieldtrip at science.ru.nl>
Onderwerp: Fwd: 2x2 ANOVA at source level using cluster based permutation


Dear Dr.Maris/ Schoffelen /Oostenveld

I posted a question on the mailing list about 2x2 ANOVA at the source level (please see below my email).
It's been 2 weeks and I have not found if there is anything wrong with my code.  I would really really appreciate it if you could have a look at my code and let me know if I have done all the steps correctly.

Thanks a lot,
Philip



---------- Forwarded message ---------
From: philip Joadavi <p.joadavi at gmail.com<mailto:p.joadavi at gmail.com>>
Date: Wed, Mar 15, 2023 at 5:14 PM
Subject: 2x2 ANOVA at source level using cluster based permutation
To: FieldTrip discussion list <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>>

Dear all,

I would like to run ANOVA  to test the main and interaction effects on my beamforming results. (I used DISC beamformer)
I've already looked at the FAQ on the FieldTrip website and also the questions in the mailing list.

To check if I have done the steps correctly, I have also done the 2x2 ANOVA using SPM and I got significant results, but without FWE/FDR correction for multiple comparisons at the voxel level.

When I apply the steps described in the FieldTrip and also the mailing list (for cluster-based permutation) I could not get any significant clusters. This is surprising for me because I get always nicer results from permutation tests than I do not get from SPM.

I think I'm doing something wrong and I would appreciate any help.

So, as far as I understood, I can not use F statistics for 2x2 ANOVA, and therefore I used dependent sample T statistics for within-subject design as explained by Eric in the mailing list.

here is my code :
So, I have 2 conditions A and B, each with 2 levels (I & II).
For the averages, I used ft_ grandaverage for 2 conditions/levels, and I got a matrix of pow = [number of subjects x number of voxels] = [12x 411477]

Here is the code:

my data is arranged such that the columns are the conditions (A/B) and the rows are the levels (I/II).



% Main effect of levels (I/II)

I  = ft_sourcegrandaverage([],GA{1,1:2});

II = ft_sourcegrandaverage([],GA{2,1:2});



% Main effect of conditions (A/B)

A = ft_sourcegrandaverage([],GA{1:2,1});

B = ft_sourcegrandaverage([],GA{1:2,2});



% Interaction effect

cfg           = [];

cfg.operation = 'subtract';

cfg.parameter = 'pow';



INA = ft_math(cfg, GA{1,1}, GA{2,1}); % I - II (A)

INB = ft_math(cfg, GA{1,2}, GA{2,2}); % I - II (B)





%% prepare the hypothesis testing

cfg                  = [];

cfg.dim              = GA{1}.dim;

cfg.method           = 'montecarlo';

cfg.correctm         = 'cluster';

cfg.spmversion       = 'spm12';

cfg.parameter        = 'pow';

cfg.numrandomization = 'all';

cfg.correcttail      = 'alpha';



cfg.statistic        = 'depsamplesT';

cfg.alpha            = 0.05/2; % two tails

cfg.clusteralpha     = 0.01;

cfg.tail             = 0; % I also tried 1 for one tail

cfg.clustertail      = cfg.tail;

cfg.uvar             = 1; % subject

cfg.ivar             = 2; % conditions

cfg.design           = [repmat(1:numel(subjects),1,2); repelem([1 2],numel(subjects))];





stat{1}   = ft_sourcestatistics(cfg,I,II); % main effect (I/II)

stat{2}   = ft_sourcestatistics(cfg,A,B);  % main effect (A/B)

stat{3}   = ft_sourcestatistics(cfg,INA,INB); % Interaction



Side note: I also tried the one-tail option, but the results were not significant eighter.





I would appreciate any help!
Best,
Philip


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20230328/ca441b84/attachment.htm>