> Would it be appropriate to average each IRF time point's beta value
> (the data has previously been converted to % change from the voxel-wise
> mean rather than absolute intensity) across the subjects in a group to come
> up with a group average IRF? Would it be better to fit a curve to each
> subject's IRF estimate and then find average values for the curve fit parameters
> for the group? Is there a completely different method that is typically used?

HDR shape differences are sometimes ignored by relying on a summarized effect size such as the area under the curve (AUC) or the coefficient of the principal basis function (e.g., Gamma variate) with other coefficients (e.g., time and dispersion derivatives) being a prior assumed to contain little information. With the AUC approach the effects are summed to produce one value per condition, and then one adopts the same group analysis strategy as the HDR modeling with a fixed shape. However, the AUC approach can mask important information about HDR shape (e.g., HDR might have different signs or durations), contradicting the original goal of modeling the HDR shape.

Alternatively you can take the individual basis function coefficients of the reassembled HDR function to group analysis using 3dLME:

[afni.nimh.nih.gov]

And use the following in model.txt in your case with two or more groups:

Model:-1+Time:Group

Gang