The problem exist in the group analysis. I checked the iresp of several individual subjects and it is not just below average but also a negative-going response (i.e., betas going below baseline and then back up to baseline).
I used the general mean (across subjects and conditions). Would it make sense to use across-subjects means for each condition separately? My initial problem was large reaction time differences between conditions. It doesn't sound like it can change negative-going to positive-going HDRs, but from your answer it seems like it might help to decrease extreme cases. Does it make sense? Do you have any other idea for what am I going wrong?

Thanks a lot for all your help!
kallai