hi Floris
One thing you could try is simply using 3dDeconvolve -nodata as you normally would but add in a censor file (via -censor) to censor out "volumes" where no data will be collected. Then, when you actually go to run the GLM on your data you can simply copy in volumes to the data you collected as placeholders when you run 3dDeconvolve so that the regressors generated by -stim_times GAM for instance line up with when the data were collected in time, but these copied volumes are censored out and do not contribute to the results.
I'm sure Rick will have some more ideas!
James