[FieldTrip] Oscillatory power normalization

[Matt Mollison]

My questions essentially boil down to: what do people do for power
normalization when assessing statistical differences?

It gets more detailed below regarding examining event-related power changes
relative to a baseline (within-subjects, comparing two conditions, stimulus
onset = 0 ms). I didn't find much discussion of this on the list or the
wiki. Any references for these issues would also be appreciated.

(1) Does power data need to be baseline normalized for statistical tests
comparing conditions? Normalization would put power on equal footing across
all subjects, conditions, sensors, times, frequencies, etc., but it will
surely affect power during the stimulus period in a particular way. If so,
do the two (or more) conditions need to use the same baseline condition, or
can each trial be normalized to its own pre-stim baseline period (a la
ft_freqbaseline)? For either, it seems like you'd always need
keeptrials='yes' in ft_freqanalysis. However, it does not seem to get
normalized in the cluster_permutation_freq tutorial (within-subjects)---am
I missing something?

If we should normalize:
(2) I've read a number of papers that Z-transform stimulus period power
relative to pre-stim activity (subtract mean, divide by std) before doing
statistics. I've also read a lot that don't mention baselines, or e.g. do a
decibel [dB] transform. ft_freqbaseline does not have a Z-transform option.
There is ft_preproc_standardize, but this seems to operate at a lower level
than is usually recommended. Z-transforming seems like a good option, but
how can I use it in the FT pipeline for within-subjects analyses
(especially with keeptrials='no')? Alternatively, when should one use
'absolute', 'relative', or 'relchange'?

Regarding choosing the baseline period:
(3) It seems that the baseline period needs to precede stimulus onset by a
sufficient amount of time so that the stimulus period doesn't bleed into
the baseline; this time would be specific to both the frequency and either
wavelet width or taper window length. For example, at 4 Hz with wavelet
width=6 or a taper with 6 cycles per time window (t_ftimwin) the
wavelet/window would be 1500 ms long, and the end of the baseline must
precede stimulus onset by at least half this to keep them separate. At
lower frequencies this could get quite unruly (e.g., 1 Hz would require
ending 3000 ms before stimulus). Is this correct? Maybe that's why it's
better to have a single separate baseline condition. Anyway, the
timefrequencyanalysis tutorial seems to disregard this separation of
baseline and stimulus activity (as have many papers I've read), so maybe
I'm wrong about this being necessary.

Thanks for your time,
Matt Mollison

--
Univ. of Colorado at Boulder
Dept. of Psychology and Neuroscience
matthew.mollison at colorado.edu
http://psych.colorado.edu/~mollison/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20120309/0cd15afd/attachment-0001.html>