Coherence!

[Robert Oostenveld]

dear Muthuraman

On 9 Jan 2007, at 11:08, Muthuraman Muthuraman wrote:
> %MTMFFT
> cfg.channel    = channelselection({'all' 'M1'}, data.label);
> cfg.channelcmb = channelcombination({'all' 'M1'}, data.label);

for DICS you will also need the cross-spectrum between all EEG
channels. Hence you should do cfg.channelcmb = 'all'. The code above
is fine if you want to look at coherence on channel level, which is
what you currently are looking at.

> cfg.foilim     = [2 30];
> cfg.taper      = 'hanning';
> cfg.pad        = 1;
> cfg.keeptrials = 'yes';
> cfg.variance   = 'yes';
> cfg.jackknife  = 'yes';
> freqmtmfft     = freqanalysis(cfg, data);
> fdmtmfft       = freqdescriptives([], freqmtmfft);

This uses 1 hanning taper, and you have one-second long data
segments. Hence the frequency resolution is 1/1s = 1Hz.

> %MTMWELCH
> cfg.method     = 'mtmwelch';
> cfg.output     = 'powandcsd';
> cfg.foi        = [2:2:30];
> cfg.t_ftimwin  = ones(1,length(cfg.foi)).*0.5;
> cfg.taper      = 'hanning';
> cfg.pad        = 1;

This uses a 500ms window per trial to do the estimation. That gives
you a 1/0.5 = 2Hz frequency resolution. The lower frequency
resolution implies that there is more frequency smoothing, so the
estimate will be different. If the true coherence peak is broad-band
(i.e. broader than 1 Hz), then the estimate will probably be better
(=higher), if not, then it will be worse (=lower).

> cfg.keeptrials = 'yes';
> cfg.variance   = 'yes';
> cfg.jackknife  = 'yes';
> freqmtmwelch   = freqanalysis(cfg, data);
> fdmtmwelch     = freqdescriptives([], freqmtmwelch);

You probably should do
   cfg = []
   cfg.variance   = 'yes';
   cfg.jackknife  = 'yes';
and then
   fdmtmwelch     = freqdescriptives(cfg, freqmtmwelch);
since the variance options do not apply to freqanalysis, but to
freqdescriptives instead.

> I have attached a figure with this mail[Coherence between C3 and
> M1] in which there is a difference in coherence values
> for the fieldtrip method and the welch periodogram method with and
> without laplacian transform
>
> Is there any input parameters i can change in the fieldtrip analysis
> so that i can have the similar values, or is there any major
> differences in the methods
> with fieldtrip and Welch periodogram method.

There is a difference in scalp-current-density (laplacian) data and
EEG potential data, so I also expect coherence to be different. It is
better to report channel level CSD/laplacian coherence than channel-
level ordinary potential difference, since SCD/laplacian is reference
free and potential is not. Note that you can also use FT to compute
coherence on SCD data, assuming that you can export the laplacian
data to a raw file. I don't know which software you used to compute
the SCD/laplacian and welsh coherence that you have in the figure,
therefore I also cannot comment on why the values are different. The
difference between the mtmfft and mtmwelsh can be explained (see
above) because the different frequency resolutions.

best regards,
Robert