[FieldTrip] depsamplesF

Fri Jan 28 10:43:42 CET 2011

Hi Tom,

To test main and interaction effects in your 2x2 within subjects design, you
have to perform 3 tests, each using the statfun desamplesT. Say you have the
output of ft_timelockanalysis for all four conditions: tlout_Ia, tlout_Ib,
tlout_IIa, tlout_IIb. Your then proceed as follows:

1.   Main effect of I versus II: calculate the mean of tlout_Ia.avg and
tlout_Ib.avg and put this is a a new struct variable tlout_I, which has the
same fields as tlout_Ia and tlout_Ib. Do the same with tlout_IIa.avg and
tlout_IIb.avg and make a new struct variable tlout_II. Then run
ft_timelockstatistics with input arguments tlout_I and tlout_II. With this
analysis you will test the main effect of I-versus-II.

2.   In the same way, you now test the main effect of a versus b. In your
calculations, the roles of (I,II) and (a,b) are now reversed.

3.   Interaction of I-vs-II and a-vs-b. Calculate the differences
(tlout_Ia.avg-tlout_Ib.avg) and (tlout_IIa.avg-tlout_IIb.avg), put them in
output structures and statistically compare them using
ft_timelockstatistics. With this analysis, you test the interaction of
I-vs-II and a-vs-b.

There is no need for Bonferroni correction or an adjustment of
cfg.clusteralpha (which does not affect the false alarm rate anyhow) and
cfg.alpha.

Best,

Eric Mari

Thank you very much for this helpful advice Olga. That would have also been
my impression until I found this discussion between Maya Zuckerman and Eric
Maris:

http://mailman.science.ru.nl/pipermail/fieldtrip/2010-December/003335.html
http://mailman.science.ru.nl/pipermail/fieldtrip/2010-December/003338.html

, which I may have misunderstood. I can see from the list that people are
interested in using depsampleF to investigate interactions between two
independent variables (e.g., background: I,II x stimulus: a,b) in
experiments with multiple participants. I can also see a several people have
had problems with this.

Given for datasets I_a, I_b, II_a, II_b I'd like to do something like:
cfg.design =????;
...
cfg.statistic ='depsamplesF';
...
 [stat] = ft_timelockstatistics(cfg, I_a, I_b, II_a, II_b);

to find the tests of of main effect clusters for each factor and interaction
clusters. How to specify up cfg.design could be one of the problems, as
could be what other cfg parameters need to be specified. I may be barking up
completely the wrong tree.

Assuming the interaction is significant, there are 4 differences that would
be of theoretical interest: I_a vs I_b, II_a vs II_b,I_a vs II_a,I_b vs
II_b). For instance, a "classical" interaction might exhibit the abolition
of significance of an effect ab seen at I at level II of that factor. 

So, if I understand the Zuckerman-Maris dialogue, if I used the permutation
test to test such I_a vd II_b differences as in the tutorial:

 cfg.method = 'montecarlo';
cfg.statistic = 'depsamplesT';
cfg.correctm = 'cluster';
cfg.clusteralpha = 0.05;
cfg.clusterstatistic = 'maxsum';
cfg.minnbchan = 2;
cfg.tail = 0;
cfg.clustertail = 0;
cfg.alpha = 0.025;
cfg.numrandomization = 500;
...
[I_avsb_stat] = ft_timelockstatistics(cfg, I_a, I_b);
[II_avsb_stat] = ft_timelockstatistics(cfg, II_a, II_b);

[IvsII_a_stat] = ft_timelockstatistics(cfg, I_a, II_a);
[IvsII_b_stat] = ft_timelockstatistics(cfg, I_a, II_b);

,I may have to either do some bonferonni correction:
-at the cluster level (cluster.alpha) 
-or at the level at which each t-test can be included in a cluster
(cfg.alpha). 
Please would you verify which, if at all?

If I am testing 4 such differences with 4 permutation tests then if the
bonferroni correction is to be at the cluster level, please should the
correction depend upon the number of clusters as well as the number of such
differences tested? 

Please, I look forward to any clarifications.

With best regards,

Tom Campbell.

  _____  

From: olga at graphicmind.info
To: tom_campbell75 at hotmail.com; fieldtrip at donders.ru.nl
Subject: Re: [FieldTrip] depsamplesF
Date: Thu, 27 Jan 2011 06:16:51 +0300

Hi, Tom, 

I guess if you do cluster analysis, which is based on permutation tests you
do not need any correction like.

Cluster-based statistics just deal with multiple comparison problem
differently (Monte-Carlo randomization, permutation tests and examine the
probability of your cluster among the random ones). Clusters may be formed
based on time, space/frequency adjacency.

Best Regards,

Olga.

On 27.01.2011, at 2:26, Tom Campbell wrote:

Dear Eric Maris, Robert Oostenveld and colleagues,

I write with some queries with reference to your previous correspondence on
the Fieldtrip listserv and would very much appreciate if you could please
answer them.

I am trying to use Fieldtrip to analyse timelocked ERP data from what is a
16(participant: [1:16]) X2(Background: congruent, incongruent)x2(Stimulus:
Animal, Vehicle) design. The code seems to runing well treating this as a
16(participant: [1:16]) X4(Visual stimulus: animal-congruent background,
animal-incongruent background, vehicle-congruent, vehicle-incongruent
)design with 4 conditions, though I haven't looked at the results of the
tests yet. If I then run cluster analyses of differences of theoretical
interest via depsamplest, please how would I bonferroni correct? I am
interested in what clusters exist in the background and stimulus main
effects and their background X stimulus interaction. Please is this possible
in fieldtrip to use depsamplesF to work with a Participant X "2-way" design?

Best regards,

Tom Campbell.

_______________________________________________
fieldtrip mailing list
fieldtrip at donders.ru.nl
http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20110128/85721485/attachment-0002.html>