The development of an empirical random utility travel demand model, like the development of most other statistically based models, typically includes testing and comparing several different functional specifications of the model to determine which specification best explains the available data. This paper is concerned with comparisons based on prediction success tables and indices. It is shown by example that prediction success tables and indices can lead to selection of the incorrect model when a correctly specified model is compared with an incorrectly specified one. This can happen even with data sets sufficiently large to make the effects of random sampling errors negligibly small. Accordingly, it is concluded that prediction success tables and indices should not be used for model selection. Alternative selection procedures that are both reliable and easy to use are described.
Abstract