[FieldTrip] Fwd: Questions about coherence statistics for group analysis related to J.Neurosci. 2011 and Science 2005 papers

[jan-mathijs schoffelen]

Hi Sophie,

I forward your message to the discussion list so that others may also participate in and benefit from the discussion

> I have been reading your J.Neurosci. 2011 and Science 2005 papers and I also read Maris' paper about "nonparametric statistical testing of coherence differences" and looked at Fieldtrip tutorial and browsed the fieldtrip mailing list archive about coherence statistics. However, i still have question, and hope you will have time to answer. 
> 
> Here a short description of our EEG-EMG experiment : we have 10 subjects, and we compute the corticomuscular coherence for each 
> EEG (64 channels) and EMG (4 channels, 2 muscles per hand) channels combination. Subjects are performing a visuotracking task and each trial contained 4 conditions (3sec long) . One trial is ~12 sec long ) . 
> 
> First, about your Science 2005 paper : in the materials and methods part, you z-transformed the coherence values and performed a non-parametric permutation test to test the significance of the difference between the prestimulus and the stimulus interval. Did you perform the test on the z-transformed coherence values (variable 'r' on your paper) ? 

Yes.

> Concerning the implementation in Fieldtrip while using the ft_freqstatistics function, is it the z-transformed coherence difference you compute in the statfun_xxx (and as output variable 's.stat' )? so as to perform the montecarlo permutation? We would like to compare each of the four conditions with the baseline period we defined in our experiment in the group subject. 

No. The computations implemented in FieldTrip's statfun_indepsamplesZcoh function are aimed at computing the Z-transformed coherence difference for a single subject, i.e. where the unit-of-observation is a trial. This is the method which is described in Eric Maris' J Neurosci Methods paper. In the other two papers you mention, we did statistical inference across subjects, i.e. each unit-of-observation was a subject. In order to achieve this, we computed for each subject the Z-transformed coherence by hand. These Z-coherence differences were tested against 0 using the montecarlo permutation approach. This is essentially the same as testing whether the distribution of coherence values across subjects is exchangeable across conditions. Practically this means that you need to do some work by hand:

For each subject you need to do something like this:

coh1 = ft_connectivityanalysis(cfg, freq1); % create coherence for subject X condition1
coh2 = ft_connectivityanalysis(cfg, freq2); %create coherence for subject X condition 2

coh1.cohspctrm = atanh(....) - atanh(...) etc -> compute the Z-transformed coherence difference between coh1.cohspctrm and coh2.cohspctrm;
coh2.cohspctrm(:) = 0;

cfg = [];
cfg.method = 'montecarlo';
cfg.statistic = 'diff';
cfg.parameter = 'cohspctrm';
...
stat=ft_freqstatistics(cfg, coh1, coh2, ..., ..., ..., )


> Second, about your J.Neurosci. 2011 ,as we would like to do comparison between conditions (comparable to what you did on your paper, i assume). In this case, as we want to know if the effect is relevant in our group of subjects, and from what  i understood from http://mailman.science.ru.nl/pipermail/fieldtrip/2009-May/002147.html one should not use the  "indepsamplesZcoh " statistic function when calling ft_freqstatistics . Is a function available soon in Fieldtrip release ? Or what modification should i do on indepsamplesZcoh function? (from what i understood, this function is also computing a z-transformation on the data) 

See above. i think in principle you could also take a 2-step approach. In other words, you could use the statfun_indepsamplesZcoh, to compute per subject and contrast the Z-transformed coherence (so you will have to use ft_freqstatistics with single trial data). Next, you can use the output to ft_freqstatistics for all subjects as an input to a second call to ft_freqstatistics to do the statistical inference across subjects. This still needs an extra set of input data-structures containing 0's as dummy condition. This is for the same reason as above.

> Another question about fieldtrip, it is preferable to input 2 grand average structure per condition as input in the ft_freqstatistics function? Or Should i input one structure per subject and per condition?

ft_freqstatistics accepts multiple input arguments so there is no reason to pass the data through ft_freqgrandaverage first.

> I have been very confused reading these papers and it is even harder to express my misunderstanding, so I hope my writing has been quite understandable.... 
> 
> Thank you in advance for any help with these issues.

best wishes,

Jan-Mathijs


> 
> 
> Sophie Chen - Ph.D Student in Neuroscience
> 
> Performance Motrice et Modélisation P2M
> Institut des Sciences du Mouvement http://www.ism.univmed.fr/
> UMR 6233 CNRS - Université Aix-Marseille
> Tel : +33 (0)4 91 17 22 06
> Office : D 2.32
> 

Jan-Mathijs Schoffelen, MD PhD 
Donders Institute for Brain, Cognition and Behaviour, 
Centre for Cognitive Neuroimaging,
Radboud University Nijmegen, The Netherlands
J.Schoffelen at donders.ru.nl
Telephone: +31-24-3614793

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20110922/2c16336f/attachment.html>