[FieldTrip] Cluster-based permutation tests for between-subject design

Eric Maris e.maris at psych.ru.nl
Tue Nov 4 12:22:12 CET 2014


Hi Fred,


What you ask can definitely be done in Fieldtrip. The analyses you need are 
very similar to those of Vladimir Litvak, for who I implemented the statfuns 
indepsamplesRegrT and depsamplesRegrT. To give you the recipe, I would have 
to go back to the code, because it is such a long time ago. Anyhow, you now 
know where to look, and I trust that you can get a long way yourself.

Best,

Eric Maris


  > -----Original Message-----
  > From: Frédéric Roux [mailto:f.roux at bcbl.eu]
  > Sent: dinsdag 4 november 2014 11:59
  > To: FieldTrip discussion list
  > Subject: Re: [FieldTrip] Cluster-based permutation tests for between-
  > subject design
  >
  > Dear all,
  >
  > I would like to follow up on this, as I am currently struggling
  > myself with the coding of the design matrix of my analysis.
  >
  > I have MEG data from n = 95 participants aged between 12 and 27 years
  > on a WM-task with 3 WM-load conditions.
  >
  > I would like to investigate the effects of age (continuous variable
  > ranging from 12 to 27) and WM-load (categorical variable ranging from
  > 1-3) on delay-activity in different frequency-bands.
  >
  > What would be the correct way to perform such an analysis in
  > fieldtrip?
  >
  > Would this be akin to a mixed-effects design?
  >
  > Best regards,
  >
  > Fred
  >
  > Frédéric Roux
  >
  > ----- Original Message -----
  > From: "Eric Maris" <e.maris at psych.ru.nl>
  > To: "FieldTrip discussion list" <fieldtrip at science.ru.nl>
  > Sent: Friday, October 31, 2014 4:56:16 PM
  > Subject: Re: [FieldTrip] Cluster-based permutation tests for
  > 	between-subject	design
  >
  >
  >
  >
  >
  > Dear Dylan,
  >
  >
  >
  > I replied to your question on the Fieldtrip faq page that you refer
  > to. Although the problem can be solved, it is not implemented yet,
  > because it requires the implementation of three new statfuns:
  > Hotelling’s T-square for dependent samples, Hotelling’s T-square for
  > independent samples, and Wilk’s lambda. A good statistics book (e.g.,
  > Johnson & Wichern) will show you how to do this yourself. Otherwise,
  > you have to be patient.
  >
  >
  >
  > Although your statistical problem can be solved, with neurobiological
  > data, it is almost never wise to statistically test interaction
  > effects in designs more complicated than the 2-by-2 factorial design.
  > In these more complicated designs, you always end up with F-tests,
  > which do not inform you about the pattern in the data that is
  > responsible for the interaction effect.
  >
  >
  >
  > Best,
  >
  >
  >
  > Eric
  >
  >
  >
  >
  >
  >
  >
  >
  > From: Dylan DeLosAngeles [mailto:dylan.delosangeles at gmail.com]
  > Sent: dinsdag 28 oktober 2014 13:22
  > To: FieldTrip discussion list
  > Subject: Re: [FieldTrip] Cluster-based permutation tests for between-
  > subject design
  >
  >
  >
  >
  > Thank you, Eelke.
  > I'm sorry I am not picking this up more quickly.
  >
  >
  >
  >
  >
  > Your examples of passing either multiple structs or cell arrays to
  > ft_freqstatistics deals with one group of subjects in multiple
  > conditions, or two groups in one condition, respectively. I have two
  > groups, both doing multiple (11) conditions and I would like to know
  > how to handle that in ft_freqstatistics. What I require seems to be
  > similar to what Eric Maris wrote about here:
  >
  >
  > http://fieldtrip.fcdonders.nl/faq/how_can_i_test_an_interaction_effec
  > t_using_cluster-based_permutation_tests
  >
  >
  >
  >
  >
  > but instead of a 2-by-2 factorial design, I am using a 2-by-11
  > factorial design. As such, I am not sure how this translates when
  > making difference data structures. I hope I am not missing something
  > obvious.
  >
  >
  >
  >
  >
  > Thanks again,
  >
  >
  > Dylan
  >
  >
  >
  >
  >
  > On Tuesday, 28 October 2014, Eelke Spaak < eelke.spaak at donders.ru.nl
  > > wrote:
  >
  > Dear Dylan,
  >
  > You don't want (or need) a single struct with a subj X chan X freq X
  > time powspctrm. Instead, it is often convenient to collect each
  > individual subject's struct in a *cell* array (rather than a struct
  > array). See, for example, here:
  > http://www.mathworks.nl/help/matlab/cell-arrays.html and here:
  > http://blogs.mathworks.com/loren/2006/06/21/cell-arrays-and-their-
  > contents/
  > .
  >
  > At the statistics step you should pass in multiple structs, each one
  > corresponding to one unit-of-observation, to ft_freqstatistics. This
  > can be done like so:
  >
  > stat = ft_freqstatistics(cfg, struct11, struct12,
  > struct13,...struct21, struct22, ...);
  >
  > or, using the cell arrays, like so:
  >
  > stat = ft_freqstatistics(cfg, groupA{:}, groupB{:});
  >
  > Make sure that each column in your design matrix describes one unit-
  > of-observation, in the order in which the structs are passed into
  > ft_freqstatistics.
  >
  > Best,
  > Eelke
  >
  > On 21 October 2014 03:41, Dylan DeLosAngeles <
  > dylan.delosangeles at gmail.com > wrote:
  > > Dear Eelke,
  > >
  > > Thank you for help regarding cluster-based permutation analysis of
  > two or more conditions.
  > >
  > > I am using time-frequency data (not time-lock structures). My first
  > problem seems to be getting my 12 subjects into the 4D powspectrum.
  > >
  > > My code below loads 12 subjects from the first group, but I end up
  > with a 1 x 12 struct where each struct's .powspctrm is 1 subject x 11
  > electrodes x 3 frequencies x 2049 time points, rather than one struct
  > with a 4D powspctrm with 12 subjects as rows x electrodes x freqs x
  > time points.
  > >
  > > for k = 1:Nmt, % states
  > > for i = 1%:Ng, % groups
  > > for j = 1:Ns, % subjects
  > >
  > > % load files
  > > p(j) = eeg3.eeg.load(fullfile(fpath,fname2));
  > >
  > > % convert to eeglab to get channel locations
  > > EEG(j) = eeg2eeglab( p(j));
  > > EEG(j) = pop_chanedit( EEG(j), 'lookup', chanlocfile);
  > >
  > > % preprocessing in fieldtrip
  > > d(j) = eeglab2fieldtrip( EEG(j), 'preprocessing');
  > >
  > > % specify length of time to use in config time =
  > > EEG(j).xmax-EEG(j).xmin;
  > >
  > > % setup configuration for freqanalysis cfg = []; % clear cfg
  > > cfg.output = 'pow'; cfg.channel = 'EEG'; cfg.method = 'mtmconvol';
  > > cfg.taper = 'hanning'; cfg.foi = 0.5:3; % delta cfg.toi =
  > 1:0.05:time;
  > > % length of each state cfg.t_ftimwin = 7./cfg.foi; % 7 cycles
  > > cfg.keeptrials = 'yes';
  > >
  > > % do freqanalysis
  > > freqdata(j) = ft_freqanalysis( cfg, d(j)); end end end
  > >
  > > My second problem is loading in the second group of 12 subjects and
  > what that will look like when I run ft_freqstatistics.
  > >
  > > Lastly, I just want to confirm what you said in your previous
  > email,
  > > that I should be using indepsamplesF for more than two conditions
  > (I
  > > have 11), and therefore my design should look like this;
  > > 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 (two groups)
  > > 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 10 10 11 11 (11 conditions)
  > >
  > > Any help would be appreciated.
  > >
  > > Kind regards,
  > > Dylan
  > >
  > >
  > >
  > >
  > > On Wed, Sep 24, 2014 at 3:29 PM, Eelke Spaak <
  > eelke.spaak at donders.ru.nl <mailto: eelke.spaak at donders.ru.nl >>
  > wrote:
  > > Hello Dylan,
  > >
  > > You can analyse a between-subjects design exactly as you would a
  > > between-trials design (at least as far as the statistics step is
  > > concerned), in both cases the two conditions correspond to two
  > groups
  > > of observations, and not to the same group of observations measured
  > in
  > > two separate conditions (which would be a within-UO design). In
  > > FieldTrip, you would typically compute averages per subject, then
  > use
  > > an "indepsamplesT" (or indepsamplesF with >2 conditions) statistic
  > > (not depsamples). indepsamplesT only requires one row in the design
  > > matrix, indicating the condition.
  > >
  > > Note that if you have e.g. timelock structures in two (or more)
  > cell
  > > arrays, corresponding to the conditions, you can input them into
  > the
  > > statistics function as follows:
  > >
  > > stat = ft_timelockstatistics(cfg, tlCondA{:}, tlCondB{:});
  > >
  > > without having to call ft_timelockgrandaverage. In fact, the above
  > is
  > > the preferred way to do statistics now. (The same holds for
  > > ft_freqstatistics.)
  > >
  > > Hope that helps,
  > > Best,
  > > Eelke
  > >
  > > On 24 September 2014 02:32, Dylan DeLosAngeles <
  > > dylan.delosangeles at gmail.com <mailto: dylan.delosangeles at gmail.com
  > >> wrote:
  > >> Hello,
  > >>
  > >> So far, the tutorial on "Cluster-based permutation tests on
  > >> time-frequency data" has been very helpful.
  > >>
  > >> Out of the four combinations from the two UO-types (subjects and
  > >> trials) and the two experimental designs (between- and within-UO),
  > >> the tutorial covers statistics on data in two conditions in a
  > >> between-trials, in a within-trials and in a within-subjects
  > design.
  > >> However, I am wondering if there is any information about the
  > fourth type of experiment design: between-subjects.
  > >>
  > >> I have data for 2 groups with 12 subjects in each group. Both
  > groups
  > >> are measured during 11 conditions.
  > >> Can I approach this in a similar fashion to within-subjects design
  > >> (multiple subjects in multiple experimental conditions), such that
  > my
  > >> design is multiple groups in multiple experimental conditions. Is
  > it
  > >> a case of first averaging over all trials belonging to each of the
  > >> experimental conditions for each subject (as instructed in
  > tutorial),
  > >> and then averaging over all subjects in each group?
  > >>
  > >> Configuration code for setting up the design currently looks like
  > >> this; grp = 2; subj = 11; design = zeros(2, subj*grp);
  > >>
  > >> for i = 1:grp
  > >> design(1,i:2:end) = i;
  > >> end
  > >>
  > >> idx = 1;
  > >> for i = 1:subj
  > >> design(2,idx:idx+1) = i;
  > >> idx = idx+2;
  > >> end
  > >>
  > >> Is there anything else I need to take into consideration when
  > doing
  > >> these statistics?
  > >>
  > >> Thank you,
  > >> Dr Dylan DeLosAngeles
  > >> Research Fellow
  > >> Brain Signal Laboratory
  > >> Flinders University
  > >>
  > >> _______________________________________________
  > >> fieldtrip mailing list
  > >> fieldtrip at donders.ru.nl <mailto: fieldtrip at donders.ru.nl >
  > >> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
  > > _______________________________________________
  > > fieldtrip mailing list
  > > fieldtrip at donders.ru.nl <mailto: fieldtrip at donders.ru.nl >
  > > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
  > >
  >
  > _______________________________________________
  > fieldtrip mailing list
  > fieldtrip at donders.ru.nl
  > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
  > _______________________________________________
  > fieldtrip mailing list
  > fieldtrip at donders.ru.nl
  > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
  >




More information about the fieldtrip mailing list