Modelling activity-diary data: complexity or parsimony

Abstract

The purpose of this thesis was to find an answer to two particular questions. At first whether simpler, and hence more parsimonious models would perform better, worse or approximately as well as complex models in the context of activity-diary data. And secondly: how well is the performance of nonlinear and semi-linear models, as compared to linear models at the selection of a transport mode? These semiand nonlinear models often lead to more parsimonious, but on the other hand also to more complex models (in terms of model definition, not in terms of the number of parameters). Another objective that fits within this second main topic of linear and nonlinear models, is concerned with testing parametric linear models on their goodness-of-fit. A test was developed to investigate lack-of-fit of a linear model based on a nonparametric classification tree. This test clearly shows the value of a nonlinear model, and how it can serve to improve a linear model. Of course, it is difficult to give clear recommendations on the choice of a particular model. Which model would be preferable? It raises many questions and there are several possible grounds for preferring one model above another. In transportation studies, predictive performance, interpretability, robustness and sensitiveness for policy measures are generally considered to be relevant criteria for model comparison. ...