Univariate or Multivariate Analysis for Better Prediction Accuracy ? A Case Study of Heterogeneity in Vehicle Ownership

In this research, we contrast different modeling frameworks that offer alternative ways of capturing observed/unobserved heterogeneity. The model systems compared are: ordered logit, residential location cluster based ordered logit model (exogenous segmentation), mixed ordered logit, latent segmentation based ordered logit model, and a joint copula based self-selection clustering model. While the comparison across single dependent variable models is straight forward, the comparison with the copula based model requires post-processing to generate marginal distribution for the choice of interest. The comparison exercise is conducted in the vehicle ownership context using O-D survey data of Greater Montreal Area (GMA), Canada. The performance of the alternative frameworks is examined in the context of model estimation and validation (at the aggregate and disaggregate level) using a host of comparison metrics. In all cases, the superior performance of the ordered part of the joint copula based model indicates that employing information from an additional dependent variable (such as residential location choice in our case) allows us to better understand and predict the main dimension of interest (vehicle ownership).