This paper develops a dynamic model of neighborhood choice along with a computationally light multi-step estimator. The proposed empirical framework captures observed and unobserved preference heterogeneity across households and locations in a flexible way. We estimate the model using a newly assembled data set that matches demographic information from mortgage applications to the universe of housing transactions in the San Francisco Bay Area from 1994 to 2004. The results provide the first estimates of the marginal willingness to pay for several non-marketed amenities-neighborhood air pollution, violent crime, and racial composition-in a dynamic framework. Comparing these estimates with those from a static version of the model highlights several important biases that arise when dynamic considerations are ignored.