The strong ground motion of the Loma Prieta Earthquake is synthesized using a known three-dimensional (3D) crustal model of the region, a rupture model determined under the assumption of laterally homogeneous structure, and Green's functions computed by superposition of Gaussian beams. Compared to results obtained assuming a laterally homogeneous crust, stations within 30 km of the hypocenter and 5 km of the fault trace are predicted to be strongly focused by factors up to 10 or more, while stations lying at the northeastern edge of the Santa Clara Valley are predicted to be weakly defocused by factors of 0.5. The focusing is caused by low velocities lying between the Zayante and San Andreas faults, and the defocusing is caused by a wedge of high-velocity material between the San Andreas and Sargent faults. The magnitude of amplitude variations resulting from known 3D structure is similar to that incorporated in site effects resulting from variations in surface impedance and shallow structure beneath receivers. The existence of these effects add to the uncertainty in estimates of future strong ground motion. The strong path dependence of the effects makes it possible to include them in estimates of future ground motions only by knowing both the 3D structure and the sites of localized high slip in future earthquakes. To match observed waveforms with the structure model of Eberhart-Phillips et al. (1990), the slip model of Beroza (1991) is modified by reducing the magnitude of the northwest maximum in slip by a factor of 0.7 and increasing the magnitude of the southeast maximum slip by a factor of 2.5. These modifications are in the range of variations in the estimated slip history published by investigators using different data sets and inversion procedures.