Propensity Score Weighting with Generalized Boosted Models to Explore the Effects of the Built Environment and Residential Self-Selection on Travel Behavior

Abstract

Many studies have examined the association between the built environment, residential self-selection, and travel behavior. However, few studies have quantified the relative contribution of the built environment itself. Using the 2012 Nanjing Household Travel Survey data, this study applied hierarchical clustering and propensity score weighting to study the effects of the built environment and residential self-selection on travel behavior. First, residents’ household locations were classified into four built environment patterns using hierarchical clustering based on six built environment variables by loosely following the “5Ds” (i.e., density, diversity, design, destination accessibility, and distance to transit). Second, a powerful machine learning method, generalized boosted model (GBM), was employed to obtain propensity scores. Propensity score weighting, which is more effective for multiple treatments than matching or stratification, was used to control for residential self-selection. Lastly, the observed effect (OBE), the average treatment effect on the population (ATE), and the built environment proportion (BEP) were calculated for the walking trip frequency, bicycle trip frequency, public transit trip frequency, and vehicle kilometers traveled (VKT) of six pairs of built environment patterns. The results show that a high-density, mixed-use, walkable, and transit-accessible built environment is associated with more walking trips and lower VKT but has no impact on bicycle trips and has an inconsistent impact on public transit trips. The effects of some built environment variables on bicycle and public transit trips are tangled. The residential self-selection effect has the greatest impact on VKT (BEP: 48%–77%), followed by the walking trip frequency (BEP: 62%–74%) and the public transit frequency (BEP: 90%–107%).