Dear SPMers,
I have been contemplating about the way SPM99 (didnt start SPM2b yet)
performs high pass filtering, and would like to have your opinion on this
matter.
When one wants to high-pass filter fMRI time series in SPM99 in order to get
rid of low freqeuncy scanner and neurophysiological noise, usually a set of
cosine functions with wavelengths chosen in such a way that n+0.5 (n=0:N, N
depending on cutoff frequency chosen for HP filtering) cosines fit the
number of scans (my own observations of the hidden regressors stored in
xX.K{1}.KH).
The latter approach seems to be widely agreed on by the SPM community (I
found hardly any comments in the archives on this one, maybe i used the
wrong search terms?). I have some doubts though that this approach really
filters out all the low frequency noise below the cutoff frequencies, since
phase isn't really modelled. To put it less mathematically, all hidden
regressors start at maximum amplitude (first scan), whereas the noise might
not. When for example performing a fourier transform (getting back to math
here) on a signal to view it in the frequency domain, one usually calculates
the sine and cosine terms (or complex natural exponents) for all frequencies
in the spectrum, thereby obtaining both amplitude and phase of the signal.
In that light, it might make sense to include the sine-waves as well as
hidden regressors to allow more phases of the same frequency to be regressed
out correctly.
Another concern I had was that including extra regressors, one loses degrees
of freedom (i.e. power) in the stats, especially when a hidden regressor
doesnt model much. This effect might be small of course for single subject
analyses with its high degrees of freedom. Nevertheless, as an alternative
approach one could simply perform direct filtering of time-series data for
each voxel before applying a regression analysis, instead of regressing them
out. One could for example take a polynomial filter (butterworth or so), or
calculate a fourier spectrum, set all ferquency amplitudes and phases below
the threshold to 0, and inverse fourier transform again (easy in matlab with
the fft function).
Thank you in advance for any comments!
Bas
--------------------------------------------
Dr. S.F.W. Neggers
dept. of Psychonomics,Helmholtz Institute
Utrecht University
Heidelberglaan 2
3584 CS, Utrecht, the Netherlands
Tel: (+31) 30 253 4582
Fax: (+31) 30 253 4511
E-mail: [log in to unmask]
Web: http://www.fss.uu.nl/psn/pionier
--------------------------------------------
|