Hi Anders
I would set it based on the characteristics of your scanner. The idea of the HP filter is to remove effects of scanner drift (mainly do to heating of the gradients via the passive shims), and thereby make the noise more white. This threshold can be determined on a phantom with gel, but for most modern scanners the 128s cutoff works fine. One reason not to let the cutoff value follow th paradigm is in the example of a poor fMRI design with lots of long randomized conditions, in this case it could seem a good idea to change the setting to e.g. 300s, to avoid false negatives, but in doing so you will most likely get a lot of false positives because the i.i.d. assumption is no longer met.
Best
Torben
P.S. You may want to have a look at my previous postings on this topic and my 2006 NeuroImage paper, an references therein.
Torben Ellegaard Lund
Associate Professor, PhD
The Danish National Research Foundation's Center of Functionally Integrative Neuroscience (CFIN)
Aarhus University
Aarhus University Hospital
Building 10G, 5th floor, room 31
Noerrebrogade 44
8000 Aarhus C
Denmark
Phone: +4589494380
Fax: +4589494400
http://www.cfin.au.dk
[log in to unmask]
Den Uge:35 02/09/2011 kl. 14.04 skrev Anders Eklund:
> Thank you, are there any general pointers for how to set the cutoff?
>
> Should the cutoff frequency be different for a boxcar design with periods of 40 seconds compared to a boxcar design with periods of 60 seconds?
>
> /Anders
>
> S.F.W. Neggers skrev 2011-09-01 10:54:
>> Op 02-09-11 00:05, Anders Eklund schreef:
>>> Hello,
>>>
>>> what kind of detrending is performed by default for single subject fMRI analysis in SPM? I can't see any settings for detrending. Is the highpass filtering done instead of (linear, quadratic, cubic) detrending?
>> Yes. Discrete Cosine Transformation regressors are 'added' (not shown) to your design matrix with a wavelength up to the cutoff frequency you enter for high pass filtering. There is no reason (imho) to assume this works less well then the polynomial description you suggest. I say 'added' as in fact the entire GLM equation is filtered this way, but that is roughly equivalent.
>>
>> Have a look at the relevant chapters in Statistical Parametrical Mapping (the book). It's all there.
>>
>> Cheers,
>>
>> Bas
>>>
>>> /Anders
>>
>>
|