A kilometer or a mile? Does buffer size matter when it comes to car ownership?

When modeling car ownership, the scale at which built environment (BE) and accessibility indicators are measured are often overlooked. Most existing research use fixed neighborhood (zonal) or sliding neighbor-hoods (buffers around households) of the same distance (e.g., 1-km or 1 mile) to estimate the impact of various neighborhood features. However, it is likely that 1) different features have different ranges of influence on behavior and 2) that the impact of some variables are highly scale sensitive. Our study builds a series of Gradient Boosting Machines decision trees and uses partial dependence plots to examine potential threshold effects of various built environment characteristics and accessibility to resources when measured at different scales. Household-specific walking-distance buffers ranging from 400 to 1600 m are tested. Results indicate that BE measures of population density, land-use diversity, and design have consistent impacts across the tested buffer-sizes and show mostly linear relationships with the probability of households being fully motorized below or above a threshold value. Accessibility to transit, bikesharing, carsharing, and local amenities, however, appear to be very sensitive to the size of the neighborhood considered. Their relative influence within the models and the shape of their partial dependence function can change drastically as buffer sizes vary. Furthermore, smaller buffers (<= 800 m) for accessibility measures exhibit clear threshold effects on car ownership, suggesting that shorter distances have more influence on car ownership. Further research is needed to validate the observed access distances and how many of each opportunity types are necessary for a significant impact on household car ownership decisions.

Automated summary

When modeling car ownership, the scale at which built environment (BE) and accessibility indicators are measured is often overlooked. This study uses Gradient Boosting Machines decision trees and partial dependence plots to examine the potential threshold effects of different BE characteristics and accessibility to resources at different scales. The results show that the impact of features and accessibility measures can vary depending on the size of the neighborhood.