Understanding household vehicle ownership in Singapore through a comparison of econometric and machine learning models

Abstract

Rising vehicle ownership trends have led to significant increases in negative externalities associated with transportation such as pollution and congestion. While empirical studies have typically used only econometric frameworks, we must ask the question: Can machine learning models outperform traditional econometric approaches? Using a socio-demographic dataset from Singapore, 22 feature vectors were constructed using appropriate transformations and imputing missing data to predict a 6-class categorical ordinal variable. In our comparison of six different supervised learning algorithms with the multinomial logit (MNL) model, we found that the neural network (NN) model was the most robust and performed the best while generalizing to the test dataset with a predictive accuracy score almost 10% better than the MNL. Consequently, we used an ordinal logit classification approach with neural network binary classifiers (OLC_NN) to address the imbalanced classification problem. This model was seen to perform the best in terms of all performance metrics, even in comparison with the ordinal logit (OL) model.

We also used the econometric models to obtain insights into the household vehicle ownership decision-making process. Singapore’s public transport system and strict regulatory practices influenced not owning any vehicle to be the most preferred alternative. A gender effect was also revealed, along with a strong indirect income effect through housing type and job sector. Additionally, the direct income effect was statistically significant and strongly positive in magnitude. An attitudinal aspect was noticed in households with young professionals, wherein they are strongly disinclined to own a car. Proximity to transit stations and taxi ownership were also found to be significant factors in influencing vehicle ownership negatively. This research paves the way for an integrated framework that incorporates both the econometric and supervised learning approaches to better predict the influence of disruptive changes.