Our technical team advises us that due to distortions and other image type incompatibilities, it is likely that the error in registering EPIs to anatomical T1s is larger than the subject movement between one scan and the next. This is probably an even worse problem with partial EPIs.
Our standard procedure is, after confirmation by visual inspection, to trust the registration of the EPI that is chronologically closest to the anatomical image. If there is visible misalignment, which almost never occurs, then other means of alignment can be used. But in general I have become convinced that always running registration algorithms across image modes as a standard part of preprocessing will add more error than it removes.