Residual sphere methods of imaging near-source structure are applied to earthquakes in the Mariana and other subduction zones of the Northwest Pacific. Between 17° and 20°N, the nearly vertical Mariana seismic zone provides a particularly good geometry for constraining the velocity structure and penetration depth of the subducted lithospheric slab. Path-corrected and smoothed residual spheres constructed from International Seismological Centre P wave travel times are dominated by negative residuals along the strike of the slab, even for the deepest focus events. Theoretical residual spheres are computed by tracing rays through three-dimensional thermal models of the slab for four representative earthquakes, whose epicenters are separated by less than 60 km and whose focal depths range from 149 to 624 km. The velocity structure derived from the thermal models is calibrated by inverting the observed travel times from two intermediate-focus earthquakes for the temperature coefficient of velocity. Ignoring the perturbations of phase transitions within the slab, we obtain ∂vp/∂T = −0.5 m/s/°K; including these perturbations, we obtain −0.4 m/s/°K. These values are consistent with laboratory measurements and observations of large-scale lower mantle structure, and they imply that ∂vp/∂T is a weak function of depth in the mantle. Thermal models which fit the observations are characterized by slab penetration into the lower mantle at least several hundred kilometers below the seismicity cutoff; decreasing the total penetration depth to less than about 1000 km significantly degrades the fit to the deep-focus events. The travel time data are insensitive to, but consistent with, penetration below 1000 km. Similar conclusions are obtained from the residual sphere analysis of deep-focus earthquakes in the Kuril-Kamchatka, Japan and Izu-Bonin seismic zones. We infer that subduction along the entire northwest margin of the Pacific plate involves the flow of upper mantle material into the lower mantle. These results have substantial implications for models of mantle convection and hypotheses regarding the nature of the 650-km discontinuity. In particular, our data are not consistent with stratified convection models which restrict the return flow associated with plate motions to be above 650 km nor do they favor the hypothesis that the seismic velocity jump at this level is caused by a chemical change.