Running 3dvolreg twice won't help.
You can subtract the reference volume from each volume in the time series and see how much left over signal you have before and after registration. However even though the difference volumes are related to the registration cost functional, they are not a quantitative estimate of the error. You can also look at the rmsold and rmsnew parameters that 3dvolreg writes out. But those parameters are not of much use for deciding whether an alignment is acceptable or not.
Did your subjects move suddenly by several mm or over a long duration? If the movement is over a long duration, then that is OK.
cheers.
ziad