We relocate 297,400 events recorded by the Southern California Seismic Network (SCSN) between 1975 and 1998 using spatially varying station terms to improve relative location accuracy. Our method uses existing SCSN P and S picks, a smooth one-dimensional velocity model, and an iterative grid search approach based on the L1 norm. We apply empirical corrections for three-dimensional structure by computing station timing corrections that continuously vary as a function of source position. Station terms for each event are obtained by smoothing the residuals from nearby events using a natural neighbor (Delaunay) tessellation of the seismicity and then iterating until a stable set of locations and station terms is achieved. Our approach achieves relative location accuracy comparable locally to master event methods but can be applied uniformly over large regions. Median estimated standard errors for our final locations are 328 m in horizontal position and 741 m in depth. Our locations exhibit much less scatter, particularly in depth, than those of the SCSN catalog and a greater tendency to align into linear and planar features suggestive of fault structures. Our results appear comparable to, and in some cases better than, previous SCSN relocation studies using joint-hypocenter-velocity inversion techniques. Plots of daytime versus nighttime events permit discrimination between clusters of natural and artificial seismicity. We observe no simple relationship between the maximum depth of seismicity and surface geology. ¿ 2000 American Geophysical Union