Defining Neighborhood Boundaries: Are Census Tracts Obsolete

Abstract

Residential neighborhoods are defined as convex geographical areas containing similar populations and roughly homogenous housing markets. Neighborhoods are relevant largely because confidentiality requires spatial aggregation of data collected at the household level. A hedonic model using individual sales transactions and their street addresses is combined with CART (Classification and Regression Trees) to define the optimal number of neighborhoods and to place neighborhood boundaries in one Connecticut town. There are about half the number of CART neighborhoods than there are census tracts. Moreover, the CART boundaries typically run behind the houses rather than down the middle of the street, and they reduce residual variation. The CART model is important to the submarkets literature, which aggregates neighborhoods into larger homogenous markets. Moreover, anisotropic spatial autocorrelation can be modeled with CART neighborhoods.