To investigate whether or not lithospheric slabs descending along subducting margins penetrate below the 670-km seismic discontinuity, we have analyzed 4040 P and PKIKP travel times from 14 intermediate- and deep-focus earthquakes in the northwestern Pacific and obtained estimates of the near-surface anomaly as a function of position on the lower focal hemisphere. Station and ellipticity corrections are applied, and any component of the anomaly explicable by hypocenter mislocation is removed by orthogonalization of the residual vector with respect to the location parameters; the resulting residual sphere is smoothed and interpolated by a stochastic filtering scheme to average out observational errors and the effects of small-scale heterogeneities far from the source. The smoothed residual spheres for six Sea of Okhotsk earthquakes deeper than 500 km are dominated by NE-SW trending troughs of negative anomalies having strikes and dips similar to the seismic zone at these depths; the troughs are bounded on the NW and SE by parallel ridges of positive anomalies, with peak-to-trough amplitudes averaging about 1.5 s. As the focal depth decreases, the pattern translates to the NW, so that for hypocenters near 200 km the negative trough has been replaced by a positive ridge. The axis of the anomaly pattern for the two Sea of Japan deep-focus earthquakes is rotated counterclockwise from Okhotsk, consistent with the nearly N-S strike of the Japan seismic zone. These correlations, and the fact that one nearby earthquake not within a subduction zone shows very little anomaly, suggest that the residual sphere anomalies are caused primarily by slab heterogeneity. Forward modeling experiments corroborate this conclusion. The thermal disturbance of the mantle is calculated for an assumed flow field by a finite difference algorithm, and from it a model of P velocity heterogeneity is constructed; theroetical travel time residuals are computed by tracing rays through this three-dimensional structure, and the event is relocated and the residual sphere smoothed using the same station set and by the same algorithm applied to the observations. To obtain a good fit to the deep-focus Okhotsk data requires penetration by the Kuril-Kamchatka slab to depths at least 900-1000 km. Penetratio to much greater depths is consistent with the data. This conclusion agrees with the S wave results of Jordan (1977); it implies the circulation of at least some upper mantle material into the deep mantle and argues against a rigorously stratified upper mantle/lower mantle convective system.