I have a couple questions about the use of AR(1) modelling in SPM2.
First, what publications describe the method as implemented in SPM2? I
found
"Classical and Bayesian Inference in Neuroimaging: Applications K. J.
Friston, D. E. Glaser, R. N. A. Henson, S. Kiebel, C. Phillips, and J.
Ashburner NeuroImage 16, 484--512 (2002)
but I wasn't sure that's the only or most direct reference.
Second, under what circumstances would using AR(1) drastically *decrease*
p-values? (Usually it seems it increases them.) Context is that I'm
exploring some data issues and I have zero conditions and two user-defined
regressors.
Third,
"The AR(1) characterisation is only appropriate when long-term
correlations due to drift have been modelled deterministically through
high-pass filtering or, equivalently by inclusion in the design matrix as
nuisance variables." (From the SPM2 release notes, e.g.
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind03&L=SPM&P=R102599&I=-3 )
I take it that means AR(1) should not be used without high-pass filtering?
TIA,
S
|