The standard practice in modelling route choice and mode choice is to use Random-Utility models (e.g. logit or probit) with utility functions that are linear in parameters and often linear functions of the explanatory variables. However, there are reasons to suspect that non-linearities exist in the valuation of attributes in route-choice behaviour. The literature contains a number of instances where non-linearities have been incorporated in discrete-choice models in a pragmatic way, and more recently gives a plausible theoretical underpinning for non-linearities. This raises the question of whether adding threshold effects and non-linear transformations to conventional logit models can be shown to improve the models, and which method seems best. To investigate this the authors estimate a number of logit models with non-linear utility functions on a route-choice dataset collected by earlier researchers in The Netherlands, and show a practical method to estimate a logit model with piecewise linear utility functions. The results tend to show that the piecewise linear formulation gives reasonable results and is easy to use, but that the use of the Box-Cox transform must be used with care, especially if it is applied to more than one variable simultaneously.
Abstract