The earlyP wave coda (5–15 sec after the first arrival) of underground explosions at the Nevada Test Site is studied in the time domain using 2082 teleseismic short-period recordings, with the intent of identifying near-source contributions to the signals in the frequency range 0.2–2.0 Hz. Smaller magnitude events tend to have relatively high coda levels in the 0.4–0.8 Hz frequency band for both Yucca Flat and Pahute Mesa explosions. Coda complexity in this low-frequency passband is negatively correlated with burial depth for Pahute Mesa events but is only weakly correlated with depth for Yucca Flat events. Enhanced excitation of relatively long-period scattered waves for smaller, less deeply buried events is required to explain this behavior. Coda complexity in the 0.8–1.1 Hz band is positively correlated with magnitude and depth for Pahute Mesa events, but has no such dependence for Yucca Flat events. This may result from systematic variations between the spectra of direct signals and coda arrivals caused bypP interference for the largest events, all of which were detonated at Pahute Mesa. Another possible explanation is a frequency-dependent propagation effect on the direct signals of the larger events, most of which were located in the center of the mesa overlying strong lateral velocity gradients in the crust and upper mantle. Event average complexity varies spatially for both test sites, particularly in the 0.8–1.1 Hz band, providing evidence for frequency-dependent focussing or scattering by near-source structure. Both the direct arrivals and the early coda have strong azimuthal amplitude patterns that are produced by defocussing by mantle heterogeneity. The direct arrivals have stronger coherent azimuthal patterns than the early coda for Pahute Mesa events, indicating more pronounced deep crustal and shallow mantle defocussing for the direct signals. However, for Yucca Flat events the direct arrivals have less coherent azimuthal patterns than the coda, suggesting that a highly variable component of near-source scattering preferentially affecting the downgoing energy is superimposed on a pattern produced by mantle heterogeneity that affects the entire signal. This complicated behavior of the direct arrivals may be the result of triplications and caustics produced by the complex basement structure known to underlie the Yucca Flat test site. The presence of strong azimuthal patterns in the early coda indicates that source strength estimates based on early coda are subject to biases similar to those affecting estimates based on direct arrivals.