Since the spatial correlation of the events in a 3-D dataset are truly three dimensional, the best approach is to use three dimensional PEF's. Schwab 1998 notes that true 3-D PEF's more effectively whiten actual 3-D data than do ensembles of rotated 2-D filters. Events which appear incoherent in a 2-D slice may actually have hidden coherence in 3-D.
In 3-D, the way in which the data is sorted becomes an important issue. Among many different ways of sorting, two simple 3-D gathers for the cross-spread geometry (Figure 1) are common shot gathers (CSG) and common receiver-line gathers (CRLG). Define a CSG as the collection of records from all receiver lines in the survey for one shot, and similarly, a CRLG as the collection of records from one receiver line for all shots. Normally the shot spacing is far less than the receiver line spacing, so CRLG's are better sampled crossline than CSG's. This fact allows the use of fairly compact 3-D filters to predict the data in a CRLG, since the time delay between events is not large in the shot line direction. Still, the details involved in the choice of gather are strongly data-dependent, so these assertions are not exhaustive.
3-D prestack seismic data is five-dimensional, and redundant over one or more of these, depending on the survey design. A tantalizing possibility is to exploit this redundancy by reusing PEF's and predicted signal panels from nearby gathers as starting guesses.