Revisiting the causal relationship between the built environment, automobile ownership, and mode choice using double machine learning

This study addresses endogenous challenges in modeling the built environment, automobile ownership, and mode choice. While conventional methods like Hackman 2-stage models and structural equation modeling attempt to address these confounds, their reliance on fixed-parameter assumptions constrains accuracy and fails to capture nonlinear relationships. Using a double machine learning model, by controlling for endogeneity introduced by car ownership, we analyze the potential causal and nonlinear associations between the built environment, car ownership, and mode choice in Wuhan, China. Results demonstrate that built environment attributes directly explain 15–20 % of transit mode variations and 40–50 % of active travel decisions after addressing endogeneity, with significant heterogeneity between car-owning and car-less households. The effects of the built environment are more modest compared to those derived from single-layer modeling. Built environment variables exhibits nonlinear effects on car ownership, transit and active travel choice, with effect magnitudes and patterns varying across car-owning and car-less groups. This study highlights the potential of using machine learning to accurately capture the complex causal relationship between the built environment and travel behavior in observational data.