You can run 3dDeconvolve with "-polort -1" to indicate there is no baseline model. In that case, the default tests "-fout" (etc.) are against 0. You can then construct the real tests you want using the "-glt" options. This would be more symettrical than treating one condition arbitrarily as the "baseline".
The issue arises, of course, because BOLD FMRI is capable only of measuring
differences. By custom, the difference is measured from "rest" (a subject of research in itself: [
jocn.mitpress.org]), but anything else is proper for this purpose. As a practical statistical issue, you will only get decent results if the "baseline" condition has a sufficient number of samples in the stimulus time series to be (at least) approximately balanced with the "active" conditions to which it is being compared.
bob cox