Do you mean that the trial only happened to be invalid sometimes, but valid for the rest of the experiment?
If that is the case, the '-censor' option is probably the valid way to go. It would not change anything about the design matrix. Instead it tells 3dDeonvolve that signal at those timepoints specified in the '-censor' option will be ignored for modelling. This will keep the time dependencies intact and at the same time the valid trials would still be modelled correctly.
Gang