Using strong-motion data recorded in the Los Angeles region from the 1992 (M_w 7.3) Landers earthquake, we have tested the accuracy of existing three-dimensional (3D) velocity models on the simulation of long-period (≧2 sec) ground motions in the Los Angeles basin and surrounding San Fernando and San Gabriel Valleys. First, the overall pattern and degree of long-period excitation of the basins were identified in the observations. Within the Los Angeles basin, the recorded amplitudes are about three to four times larger than at sites outside the basins; amplitudes within the San Fernando and San Gabriel Valleys are nearly a factor of 3 greater than surrounding bedrock sites. Then, using a 3D finite-difference numerical modeling approach, we analyzed how variations in 3D earth structure affect simulated waveforms, amplitudes, and the fit to the observed patterns of amplification. Significant differences exist in the 3D velocity models of southern California that we tested (Magistrale et al., 1996; Graves, 1996a; Hauksson and Haase, 1997). Major differences in the models include the velocity of the assumed background models; the depth of the Los Angeles basin; and the depth, location, and geometry of smaller basins. The largest disparities in the response of the models are seen for the San Fernando Valley and the deepest portion of the Los Angeles basin. These arise in large part from variations in the structure of the basins, particularly the effective depth extent, which is mainly due to alternative assumptions about the nature of the basin sediment fill. The general ground-motion characteristics are matched by the 3D model simulations, validating the use of 3D modeling with geologically based velocity-structure models. However, significant shortcomings exist in the overall patterns of amplification and the duration of the long-period response. The successes and limitations of the models for reproducing the recorded ground motions as discussed provide the basis and direction for necessary improvements to earth structure models, whether geologically or tomographically derived. The differences in the response of the earth models tested also translate to variable success in the ability to successfully model the data and add uncertainty to estimates of the basin response given input “scenario” earthquake source models.