The varying perception problem: comparison of modeling approaches

Because choice probabilities of the well-known multinomial logit model only depend on attribute differences, the absolute values of the attributes do not have any effect on choice probabilities. Consequently, the model may have some counterintuitive results, particularly if choice sets differ substantially. Recently, researchers have proposed two alternative approaches to address this potential problem. One approach involves relaxing the assumption of independently and identically Gumbel distributed errors by assuming that a larger (dis)utility leads a larger error variance. The second approach substitutes the physical attributes by perceived values according to psycho-physical laws, assuming the perception of objective differences depends on absolute attribute intensity. This paper summarizes the methodologies, and explores the performance of a combination of these approaches using a numerical example and an empirical data set about revealed preferences regarding mode choice behaviour.