Dear Stefan,
If high pass filters (HPFs) are used to remove very low
frequencies then the remaining serial correlation
can be removed by low-order AR models (see eg. [1]).
If you increase the cut-off period of the HPF then you
will change (I imagine, increase) the remaining
serial correlation.
SPM uses an AR coefficient of 0.2 (plus or minus a
small amount, determined via a Taylor series approximation).
So, if anything, the serial correlation might
be underestimated.
I would be surprised, however, if this had a large
effect on the final statistical inference.
But if you are very concerned about this issue
you could have a look empirically by using SPM5b
and Bayesian inference to estimate the optimal
AR model order (and optimal values of AR coefficients)
as a function of the HPF.
Best,
Will.
[1] W.D. Penny, S.J. Kiebel, and K.J. Friston. Variational Bayesian
Inference for fMRI time series. NeuroImage, 19(3):727-741, 2003.
Stefan Kaiser wrote:
> Dear colleagues,
>
> from the discussion on these pages I understood that the AR(1) algorithm is
> to be used in conjunction with a high-pass filter. I was wondering whether
> the specification of the high-pass filter cutoff period is of relevance for
> the correction of serial correlations.
> More precisely, our current paradigm requires a rather long cutoff period
> of 256s. Does this imply the risk of overestimating serial correlations
> (similar to not using a high-pass filter at all)?
> Help on this issue would be great.
>
> Stefan
>
>
>
> Stefan Kaiser, M.D.
> Department of Psychiatry
> Section Experimental Psychopathology
> University of Heidelberg
> Voss-Strasse 4
> 69115 Heidelberg
> Germany
>
>
>
--
William D. Penny
Wellcome Department of Imaging Neuroscience
University College London
12 Queen Square
London WC1N 3BG
Tel: 020 7833 7475
FAX: 020 7813 1420
Email: [log in to unmask]
URL: http://www.fil.ion.ucl.ac.uk/~wpenny/
|