Predictions of route flows in networks are sensitive to the route choice model applied as well as to the route choice sets on which the choice models operate. Studies have shown that there is a significant impact of size and composition of route sets on estimation performance of route choice models in terms of estimated parameter values and model quality. While these studies have looked at an aggregate level of modeling quality, a better insight into the causes of the impacts requires a much more detailed analysis of modeling quality at the level of choice prediction of individual routes. This paper addresses the issue of prediction quality of route choice in networks at a very detailed level in dependence of the size and composition of pre-defined route sets. For better control of experimental conditions, a small hypothetical network and generated trip data are used in combination with a systematic variation of size and composition of route sets such that all relevant properties of routes and route sets are known. A range of well-known route choice models is applied to study in a comparative way their sensitivity in predicting route choice probabilities as a function of the choice set properties. It appears that most route choice models do not appear to have sufficient robustness of choice prediction at the level of individual routes.
Abstract