Hi,
One possible solution, presuming you don't want to take Bob's approach of selecting a best fit from among competitor IRF's, is to use the approach outlined in educational materials afni07.pdf. This entails using waver and its -EXPR approach to generate three tent functions for a particular event. These tent functions are fit to your time series in overlapping combination using 3dDeconvolve.
After 3dDeconvolve you create synthetic IRF time courses from the fit coefficients and you can then compare these IRF's timepoint by timepoint using 3dRegAna or 3dANOVA.
The strength of this approach is that it will tell you where and WHEN activation differs among conditions in a particular region. For instance, if during condition 1 the amygdala has a fast/high magnitude response and under condition 2 the amygdala has a slow/low magnitude response, you might expect to see significantly greater activation for condition 1 early in the timecourse and for condition 2 late in the time course.
-jim