My personal opinion is that this whole PPI business is not something delicate enough for strong inference anyway. As you've already noticed, the impulse response varies probably quite a lot across the brain and across conditions/tasks. Furthermore, deconvolution (e.g., what 3dTfitter does) can't guarantee you a robust and faithful recovery of the stimulus timing.
Having said that, you may consider experimenting different impulse response functions by fine-tuning the parameters of the gamma function in waver. Type "waver -help" to see those options.
Gang