[FieldTrip] Oscillatory power normalization

[Eelke Spaak]

Hi Matt,

When you are comparing power across conditions, it is not really
necessary to apply an explicit correction for the dominant 1/f
component of the raw spectrum. Since this 1/f component is present in
both conditions, when you subtract power in one condition from power
in another condition (or compute the ratio, or log-ratio, or relative
change, or whatever), the 1/f will cancel out and you will only be
left with whatever is due to your experimental manipulation. This is
true because the contrast is done per frequency. (Note that comparing
activity versus baseline is just a special case of looking at a
contrast between conditions, so the same argument holds there.)

The only time when an explicit correction for 1/f is useful, is when
you want to look at raw power. The most dominant oscillatory features
(visual alpha, visual contrast induced gamma...) will usually be
evident in raw spectra without such a correction, by the way.
Correcting for 1/f can be done in many ways, the most simple one is
simply taking the logarithm of power, something like:

freqCorrected = freqUncorrected;
freqCorrected.powspctrm = log10(freqCorrected.powspctrm);

Or you could take the first derivative in the time domain (equivalent
to multiplying the spectrum with f, search for post by Robert on this
on the FT list). Or you could take the log of both the frequency- and
power axes, then fit a line, and subtract it, then transform back
(10^corrected data).

But, the main point is: in the vast majority of typical cognitive
experiments, correcting for 1/f is not needed.

Best,
Eelke

On 23 April 2012 05:54, Matt Mollison <matt.mollison at gmail.com> wrote:
> Hi FieldTrippers,
>
> In almost all the papers I've read involving oscillatory power, some kind of
> transformation is done to the data due to the 1/f power spectrum effect
> (power decreases as frequency increases). I'm mostly looking at
> within-subjects experiments (every subject behaved in all conditions)
> comparing conditions across subjects, but it seems like normalizing the
> power spectrum should apply in any case (especially if any kind of
> parametric stats are done—right?).
>
> Anyway, it's not apparent to me how to use FT functions like ft_freqanalysis
> to make these transformations (e.g., log10 normalization, dB normalization
> [EEGLab does this], vector length normalization, etc.; the only thing I see
> is in ft_sourcedescriptives, but I'm not doing source analyses), and it
> confuses me why this is the case. I can't find much discussion regarding the
> 1/f issue on the FT wiki or the mailing list. This seems like an important
> step that is missing from any frequency analysis workflow. Am I missing
> something (meaning I just don't see the option), am I misunderstanding
> something (meaning I'm incorrect in this assumption), or is this an issue
> that needs to be fixed?
>
> Thanks,
> Matt
>
> --
> Univ. of Colorado at Boulder
> Dept. of Psychology and Neuroscience
> matthew.mollison at colorado.edu
> http://psych.colorado.edu/~mollison/
>
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip