Hi Ed,
> I must admit that I have very little a priori justification for using a
> particular convolution shape for specifying my statistical model. I was
> wondering if anyone could give me some help on weighing the relative merits
> of using a gamma versus a double-gamma HRF or perhaps even one of the other
> convolution shapes. I have used AFNI before so I am familiar with the use of
> a FIR convolution as well, though given the additional work needed to recover
> a single percent signal change statistic from a FIR convolution, I think I'd
> prefer to stick with a one of the others for now.
Most studies just use a "canonical" HRF shape -- that is a curve that
peaks around 6s and has a slight undershoot at the end. This is
implemented in FEAT as a double gamma function and problably should be
used unless you have a reason to change it. If, for instance, you expect
shape differents across brain region or if you already know that a
particular region in your study doesn't conform to the canonical shape
very well.
The original work on the shape of the HRF was done in visual
cortex I think but many studies have confirmed the same pattern despite
some differences in time-to-peak, the presence or absence of an initial
undershoot, and overall time of the HRF. THe one I remember was by
Aguirre et al. 1998 (NeuroImage, 8, 360-9). Their solution was to measure
a bunch of HRFs empirically, reconstruct the shapes, and then use PCA to
calculate a reasonable set of basis functions to deal with the
variability.
Good luck,
Joe
|