significance of coherence differences

[Jan Hirschmann]

Dear Claudia,

I have nothing to add really to Eric's mail, just wanted to let you know that I think what you're doing is just fine. During frequency analysis fieldtrip averages spectra over trials automatically (though it is not a weighted average and I didn't quite understand why you want to weight them). The normalization, too, is part of the automatic calculation of coherence by connectivityanalysis. The normalization is actually part of the definition of coherence. And as Eric says, the montecarlo method is appropiate when doing many comparisons (many channels, many frequencies). Maybe you would like think about the test statistic. I have never used the fieldtrip statisics functions but it sounds like your statistic is a traditional t-value. There are other statistics which are commonly used in coherence analysis (z-transform, see e.g. the paper of Maris and Schoffelen on coh diff) which may do a better job in capturing the effect. Though in nonparametric testing the validity of your stats does not depend on the statistic, they are just not all equally effective. 

Best,

Jan

-----Original Message-----
From: FieldTrip discussion list on behalf of Eric Maris
Sent: Thu 9/9/2010 10:02 AM
To: FIELDTRIP at NIC.SURFNET.NL
Subject: Re: [FIELDTRIP] significance of coherence differences
 
Dear Claudia,

> I too am confused about coherence and its statistical analysis..I am
> trying to
> calculate the coherence between POz and 47 other channels in a
> multisubject
> study (N=19) with a within-subject design to test (1) whether coherence
> in
> condition 'Figure' is the same as in condition 'No figure', and (2)
> whether
> coherence differs between 3 conditions 'Hits', 'Misses' and 'Correct
> rejections'.
> 
> Either way I am not sure about a couple of things:
> 
> 1) This is maybe a silly question but in your reply to Jan's post you
> state: "...You apply this test statistic to the condition-specific
> coherences,
> obtained by summing and normalizing the trial(taper)-specific cross-
> spectra..." - You mean averaging the trial-specific cross-spectra (and
> trial-
> specific powerspectra) right? Moreover does this mean that
> cfg.keeptrials can
> be set to 'no' when calculating the powerspectra and cross-spectra with
> freqanalysis(_mtmconvol)? This would be great news for me, since the 4D
> freqdata is too large to save..

For statistical testing in a within-subjects design (units-of-observation
are subjects and not trials) you don't need the trials. So, you can safely
put cfg.keeptrials to 'no'.

By the way, try using  freqanalysis_mtmfft (frequency analysis without time
resolution) before freqanalysis_mtmconvol (with time resolutions). This
makes life a lot easier and conceptually you're doing the same thing.


> 
> 2) Another related question comes from the fact that the experiment had
> 3
> different masking durations, so to combine the powerspectra belonging
> to the
> same condition (i.e. the Figs) but different mask durations we
> calculated
> weighted averages based on the least amount of trials in each Mask
> duration
> group. My question now would be whether the normalization to get the
> coherence values should be done before or after this weighing (In other
> words,
> should I weigh the cross-spectra or the coherence values, or does this
> not
> matter?)

I don't see why you have to deal with the normalization yourself. You can
use ft_freqdescriptives or ft_connectivityanalysis to calculate the
coherence values taking the freqanalsis output as its input.

In any case, to obtain coherence values from the cross-spectra, one should
always normalize (i.e., divide) by the square-roots of the power-spectra
calculated on the same trials.

I may be wrong, but it seems that you are making life more difficult than it
should be.


> 
> 3) Now for the statistics. Again in your reply to Jan's post you say:
> "...For a
> single channel pair and a single frequency bin, the appropriate
> statistic is the
> dependent (paired) samples t-statistic or, in a nonparametric
> framework, the
> Wilcoxon signed rank sum test." Does this mean it is not valid to use
> freqstatistics with cfg.method = 'montecarlo' and cfg.statistic =
> 'depsamplesT'
> (and cfg.correctm = 'fdr' or 'cluster') if I want to test whether the
> coherence
> between POz (my ref.channel) and the 47 other channels is the same in
> different conditions, using all my frequency and time bins? Do I need
> to make a
> selection?

What you propose IS valid to control for multiple comparisons. However, in
my reply to Jan, I consider a single statistical test ("... For a single
channel pair and a single frequency bin, ..."); so, no multiple comparisons
problem here.


Good luck,

Eric




> 
> Thank you for your time.
> 
> Sincerely,
> 
> Claudia
> 
> ----------------------------------
> The aim of this list is to facilitate the discussion between users of
> the FieldTrip  toolbox, to share experiences and to discuss new ideas
> for MEG and EEG analysis. See also
> http://listserv.surfnet.nl/archives/fieldtrip.html and
> http://www.ru.nl/neuroimaging/fieldtrip.

----------------------------------
The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/neuroimaging/fieldtrip.


----------------------------------
The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/neuroimaging/fieldtrip.