[FieldTrip] non-parametric methods for comparison of power spectra

[Gio Piantoni]

Hi Sangita,

I think that Fieldtrip can handle your problem just fine, if I
understand your point correctly.

If you run ft_freqanalysis with 'mtmfft', foilim [0 50] and
keeptrials='yes', you'll get a structure with dimension trials x
channel x frequency, where each element of the 3d matrix contains
power spectra. You can check the .dimord field, which will be
'rpt_chan_freq'. I don't understand your notation "channel x
frequency x power data". Do you mean a 'chan_freq' matrix with power
data in there? You'll need trials (or subjects) to calculate any
parametric or not statistics.

You can then use the output of ft_freqanalysis for your statistics. It
does not need any tweaking. The 1/f effect is not a problem because
you're testing a difference from zero, or a difference between two
conditions. So, while the mean over trials becomes smaller in absolute
values, the standard error becomes smaller as well in absolute terms,
resulting in roughly consistent t-values. In any case,
log-transformation might make your power data more normal, so that
will probably make your statistic more sound.

If this is not clear to you, please include part of your code and we
might help you further.

Hope this helps,

Gio

-- 
Giovanni Piantoni, Ph.D. student
Dept. Sleep & Cognition
Netherlands Institute for Neuroscience
Meibergdreef 47
1105 BA Amsterdam (NL)

+31 (0)20 5665492
g.piantoni at nin.knaw.nl
www.giovannipiantoni.com

On Tue, Jan 17, 2012 at 19:56, Sangita Dandekar
<sangita.dandekar at gmail.com> wrote:
> Dear Fieldtrippers,
>
> Forgive me if I have overlooked something in the FT documentation, but
> I was wondering if there are any non-parametric methods for comparison
> of power spectra in Fieldtrip?  When I say 'power spectra' I mean
> frequency vs power spectrum data without a time component (i.e. not
> the TFR data structures in Fieldtrip, but what one would get by
> applying an FFT to a time series and finding the magnitude of the
> power spectrum as a function of frequency).
>
> One (clumsy) way to do this would be to use the framework of FT's
> timelockstatistics function, but instead of entering channel x time x
> voltage data for each trial, use channel x  frequency x power data for
> each trial.
>
> I think the above should work, but am guessing it will probably be
> fine for detecting low frequency differences but not as sensitive for
> differences in the higher frequencies due to the 1/f power spectrum.
>
> Is there an alternative method that anyone can suggest? Does the
> suggested method above sound ok?  Maybe log transforming prior to
> application of the existing timelockstatistics framework?
>
> Thanks in advance for any help!
> Sangi
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip