A spike should really be an uncorrectable artifact.
Most commonly, a spike might be due to a fairly quick (even if
small) subject motion. Such a motion is usually different over
slices, as therefore would not be fixed by volume registration.
In such a case, you might want to minimize corruption of TRs
that are adjacent to the spike when time shifting.
That is the motivation for running 3dDespike first.
Look closely at your data and try to get a feel for the effects
of these steps. Then decide based on your own circumstances.
- rick