We use a large instrumental dataset from the 2019 Ridgecrest earthquake sequence (Rekoske et al., 2019, 2020) to examine repeatable source‐, path‐, and site‐specific ground motions. A mixed‐effects analysis is used to partition total residuals relative to the Boore et al. (2014; hereafter, BSSA14) ground‐motion model. We calculate the Arias intensity stress drop for the earthquakes and find strong correlation with our event terms, indicating that they are consistent with source processes. We look for physically meaningful trends in the partitioned residuals and test the ability of BSSA14 to capture the behavior we observe in the data.

We find that BSSA14 is a good match to the median observations for M>4M>4⁠. However, we find bias for individual events, especially those with small magnitude and hypocentral depth≥7  kmdepth≥7  km⁠, for which peak ground acceleration is underpredicted by a factor of 2.5. Although the site amplification term captures the median site response when all sites are considered together, it does not capture variations at individual stations across a range of site conditions. We find strong basin amplification in the Los Angeles, Ventura, and San Gabriel basins. We find weak amplification in the San Bernardino basin, which is contrary to simulation‐based findings showing a channeling effect from an event with a north–south azimuth. This and an additional set of ground motions from earthquakes southwest of Los Angeles suggest that there is an azimuth‐dependent southern California basin response related to the orientation of regional structures when ground motion from waves traveling south–north are compared with those in the east–west direction. These findings exhibit the power of large, spatially dense ground‐motion datasets and make clear that nonergodic models are a way to reduce bias and uncertainty in ground‐motion estimation for applications like the U.S. Geological Survey National Seismic Hazard Model and the ShakeAlert earthquake early warning System.