Hi,
> at present I am evaluating a validation study on our new superfast 3D
> parallel imaging fMRI acquisition. 1 scan lasts 500 ms, and hence I am a
> little worried that assumptions regrding serial correlations adopted in
> spm2's 'AR(1)+w' whitening method might not be entirely appropiate here.
> But when comparing it with good ol easy going 2D EPI with TR=2 sec, it
> is important to do adequate correction, otherwise thresholding would be
> far too lenient for the new scan sequence.
As you know the 'AR(1)+w' option in fact fits a AR(1) model with an
assumed rho coefficient of 0.2 by default - i.e. it assumes that the
AR(1) parameter is in fact somewhere near 0.2. There is also a
derivative term in the AR fitting which means that the estimation can
do a reasonable job for actual rho coefficients of something like 0.1
to 0.3 with the default of 0.2. Outside that range my impression was
that it starts to get pretty biased, at least from the simulations
that I was doing with JB a while back.
As you also know, SPM assumes that the rho coefficient is the same for
all 'activated' voxels in the brain.
I suppose you have a few options in your case. You could:
try and inspect the data and see what the actual rhos should be using
a more flexible model, and put that into SPM as the default instead of
0.2 (see below).
turn off the autocorrelation estimation (and therefore use OLS). At
least the parameter images will not be biased, and you can use these
at the second level.
try the fancy new Bayesian grand model in SPM5 which does have
flexible AR coefficient modelling which can also vary across the
brain.
FSL (ahem!)
If you want to get a feel for the actual AR coefficients, Marsbar has
Keith Worsley's full AR modelling as an option, so you could select an
ROI and get the AR coefficients that way. Something like:
D = mardo('SPM.mat');
R = maroi('my_roi.mat');
Y = get_marsy(R, D, 'mean');
D = autocorr(D, 'fmristat', 2); % AR model, order 2
E = estimate(D, Y);
The rho values are then in:
D.xVi.h
Best,
Matthew
|