Gang, Thanks.
What Worsley does in fmristat (and I think other packages do this as well)
is setup the HRF so that its integral is always 1.0.
Then then convolve this with the design.
According to Keith, the beta you get here is the beta for a very long
block regardless of how long the actual block is. In other words,
if the integral of the HRF is 1, then if the block is long enough, when
convovled with this HRF the peak will stabilize at 1.0. Therefore,
100*beta/intercept (assuming raw data) will give you percent change
for a very long block. This allows meaningful contrasts between short
blocks and long blocks.
Not sure I understand the math here, but two people have independently
stated this to me so i think this is the way to go.
Be interested in AFNI's reaction to this.
Lee