(Revised December 2, 2024)

Using data from 197 metro areas, we estimate the parameters of a dynamic, forward-looking neighborhood choice model where households have preferences over the racial composition of the neighborhood in which they live. Using multiple metro areas in the estimation sample enables us to develop a new, shift-share IV strategy to estimate the impact of the racial composition of neighborhoods on location choice that relies only on across-metro comparisons of similarly situated neighborhoods. Our neighborhood-level instrument is constructed by interacting national-average, across-neighborhood sorting patterns with respect to neighborhood-level topography with metro-level shares of households by demographic subgroup. For a given configuration of neighborhood-level topographic data, the instrument predicts variation in neighborhood-level racial shares that is attributable exclusively to variation in metro-level shares of demographic subgroups. We find that households in many different demographic subgroups have strong preferences to live in neighborhoods consisting mostly or entirely of households of the same race. These preferences are sufficiently strong that model simulations suggest that the current demographic composition of neighborhoods is not stable.