If you have some data to analyze with NO errors (in any data included in the analysis), you can do as James suggested to create a "timing" file with no times. Then you add the command line switch
-allzero_OK to indicate to 3dDeconvolve that a regression model that has no data is allowed -- normally, it will freak out (i.e., print an error message and quit) when that situation happens.
If only some runs are perfect and some runs DO have errors,
and you are analyzing the runs all together in one big analysis, then you don't need
-allzero_OK.