Revisiting residential self-selection and travel behavior connection using a double machine learning

Residential self-selection (RSS) confounds the connection between the built environment and travel behavior. Existing studies have used endogenous switching regression models to quantify the proportions of the built environment itself and RSS in the observed behavioral difference between different environments. However, the models are sensitive to model specification and assume pre-defined (mostly linear) relationships among variables. This study applies a double machine learning approach to fill the gap. The empirical context is to jointly model residential choice of Bus Rapid Transit (BRT) neighborhoods and weekly driving distance of household owning cars in Jinan, China. The results showed that the RSS effect accounts for about 40% of the observed difference in driving distance between the households living inside and outside of BRT neighborhoods. This results also emphasizes the necessity of relaxing the linearity assumption in the research on the relationships among the built environment, RSS, and travel behavior.