3dDeconvolve (and all other AFNI programs) use an input mask file on a voxel-wise basis. Shifting the coordinates of the voxels in the mask makes no difference to the masking operation, since 3drefit doesn't change the actual voxel data in the mask.

If you actually want to shift the voxels, you have to change the dataset values in the .BRIK file. For example, shifting them down one voxel could be done with 3dcalc, or 3dZeropad, cleverly applied.

AFNI itself, on the other hand, displays dataset overlays based on their (x,y,z) coordinates, not on a voxel-wise basis. This is done so that you can overlay a low-res EPI-derived dataset on a high-res structural dataset. But you can't use an EPI-resolution dataset as a mask for a high-res dataset. Only the interactive AFNI program will allow the apparent merger of datasets of different resolutions, or whose voxels are not in a 1-to-1 relationship.

I hope this is at least as clear as mud.