Seismic compression is defined as the accrual of contractive volumetric strain in unsaturated soil during strong shaking by earthquakes. We document and analyze two case histories (denoted school site and site A) of ground deformation from seismic compression in canyon fills strongly shaken by the Northridge earthquake. Site A had ground settlements up to about 18 cm, which damaged a structure, while the school site had settlements up to about 6 cm. For each site, we perform decoupled analyses of shear and volumetric strain. Shear strain is calculated using one-dimensional and two-dimensional ground response analyses, while volumetric strain is evaluated from shear strain using material-specific models derived from simple shear laboratory testing that incorporates important effects of fines content and as-compacted density and saturation. Analyses are repeated using a logic tree approach in which weights are assigned to multiple possible realizations of uncertain model parameters. At the school site, predicted settlements appear to be unbiased. At site A, the analyses successfully predict the shape of the settlement profile along a section, but the weighted average predictions are biased slightly too low. We speculate that the apparent site A bias can be explained by limited resolution of the site stratigraphy, bias in laboratory-derived volumetric strain models, and/or uncertainty in the estimated earthquake-induced settlements.