Longitudinal microeconomic and machine learning models of skeletal activity generation

Abstract

This paper examines skeletal activity generation behaviour in the Greater Toronto and Hamilton Area (GTHA) with a longitudinal approach investigating the 20-year period between 1996 and 2016. The analyses include the estimation of cross-sectional and joint/pooled models of “Work”, “Work Business” and “School” activities using microeconomic theory and machine learning algorithms, where the population is partitioned into three mutually exclusive subsamples, namely “Worker”, “Student” and “Both”, to answer two questions in particular: (1) “Do parametric models of activity generation exhibit stability over time?”; (2) “How different/similar are microeconomic and machine learning models in terms of model specification?”. Strong stationarity is observed in microeconomic model parameters of skeletal activities where year-to-year variations in general remain statistically insignificant. This supports the use of pooled models of activity generation for forecasting applications in the region utilizing full information accrued over the years. Remarkable parallelism is detected between microeconomic and machine learning models. Through variable importance and partial dependence calculations, it is shown that the models built with the two approaches utilize the same explanatory variables to describe the decision-making process in numerous modelling exercises.