Reflection seismic data typically are undersampled. Missing near offsets can be interpolated in reflection seismic data with pseudoprimaries, generated by crosscorrelating multiples and primaries in incomplete recorded data. These pseudoprimary data can be generated at the missing near offsets but contain many artifacts, so it is undesirable simply to replace the missing data with the pseudoprimaries. A nonstationary prediction-error filter (PEF) can instead be estimated from the pseudoprimaries and used to interpolate missing data to produce an interpolated output that is superior to direct substitution of the pseudoprimaries into the missing offsets. This approach is applied successfully to 2D synthetic and field data. Limitations in conventional acquisition geometry limit this approach in 3D, which can be illustrated using a synthetic data set.