Among the SPMG3 basis functions, the 1st one models the basic shape of the HRF, the 2nd one (time derivative) presumably captures the differences in the latency of the peak response (where the peak occurs), and the 3rd one (dispersion derivative) catches the differences in the duration of the peak response (how wide the upstroke spans).
What you're going to do with the 3 basis functions depends on what you care about the HRF. If you're only interested in the basic shape, you can pretty much focus on the 1st basis function. On the other hand, if the subtle difference in terms of HRF shapes is of interest to you, you can use option -iresp in 3dDeconvolve to reassemble the whole HRF for each regressor, and take the reassembled HRFs for group analysis using 3dLME with the approach mentioned on the 2nd half of the following web page:
[
afni.nimh.nih.gov]
HTH,
Gang