Seismic hazard associated with Cascadia megathrust earthquakes is strongly dependent on the landward rupture extent and heterogeneous fault properties. We use 3-D numerical simulations and a seismic velocity model for Cascadia to estimate coseismic deformation due to ~M9 earthquake scenarios. Our earthquake source model is based on observations of the 2010 M8.8 Maule and 2011 M9.0 Tohoku earthquakes, which exhibited distinct strong-motion-generating subevents in the deeper portions of the fault. We compare our estimates for land-level change to paleoseismic estimates for coseismic coastal subsidence during the A.D. 1700 Cascadia earthquake. Results show that megathrust rupture extending to the 1 cm/yr locking contour provides a good match to geologic data, and along-strike variations in coastal subsidence can be produced by including strong-motion-generating subevents in the down-dip regions of the megathrust. This work demonstrates the potential to improve seismic hazard estimates for Cascadia earthquakes by comparing physics-based earthquake simulations with geologic observations.