Understanding household VMT generation: A comparative analysis with traditional statistical models and a machine-learning approach

Planners and policy makers have long been interested in predicting people’s travel behaviors, including the number of vehicle miles traveled (VMT) they generate. Reducing VMT has come to be seen as a key strategy for lowering greenhouse gas emissions and mitigating their associated health effects, as well as for increasing sustainability and equity within communities and on a global scale. Emerging machine learning methods such as boosted regression trees (BRT) allow for the identification of comparative influences of different factors as well as their nonlinear and threshold effects on travel outcomes, but studies comparing the results of these methods with traditional statistical regressions have been scarce. This study is the first to compare these methods’ indications of the impacts of land-use patterns on VMT generation using a large multiregional dataset. The results indicate that the two methods perform similarly in predicting whether households generate any VMT and in predicting the number of VMT generated by those households that generate any. The results of the BRT model validate the statistical models’ indications that built environmental variables contribute significantly to VMT production, while further allowing for the identification of nonlinear impacts. Key thresholds and nonlinear effects of land-use variables on household VMT generation are identified from the BRT model. Both models indicate that land-use patterns that are denser, more diverse, and have increased access to transit result in reductions of vehicular trips and overall VMT, while the BRT model provides effective thresholds for these variables useful for developing planning solutions.