Coseismic subsidence along the Cascadia subduction zone causes abrupt relative sea‐level (RSL) rise that is recorded in coastal stratigraphy and foraminiferal assemblages. RSL reconstructions therefore provide insight into the magnitude, nature, and frequency of great earthquakes that can constrain deformation models and quantify the seismic risk faced by coastal populations. These reconstructions are commonly generated using transfer functions that are calibrated from counts of modern (surface) foraminifera and corresponding elevation measurements. We developed four transfer functions of increasing complexity to explore how and why the composition of the modern dataset and the choice of transfer‐function type affects subsidence reconstructions. Application of these four models to stratigraphic contacts (mud abruptly overlying peat or soil) representing the A.D. 1700 Cascadia earthquake and a field experiment that simulated subsidence show that a Bayesian transfer function (BTF) calibrated using a large modern dataset (19 sites from California to Vancouver Island) and incorporating prior information from stratigraphic context produces systematically larger subsidence estimates than a weighted‐averaging transfer function calibrated using a smaller modern dataset (8 sites in Oregon) that does not leverage stratigraphic context. This difference arises from (1) training set composition, (2) taxa–elevation relationships in the BTF that are not assumed to be unimodal, and (3) stratigraphic prior information that compensates for postdepositional, downward mixing of postearthquake foraminifera into pre‐earthquake sediment, which biases reconstructions at some sites toward smaller subsidence. Our reconstructions support a heterogeneous rupture model for the A.D. 1700 earthquake, but indicate that slip estimates in patches from Alsea Bay to Netarts Bay (Oregon) and from Netarts Bay to Vancouver Island should be increased.