An alternate crash severity multicategory modeling approach with asymmetric property

Abstract

The logit model and its variations have been used extensively in the field of traffic safety in general, and crash severity analysis in particular. Attempts were made to overcome the logit's shortcomings and limitations by generalizing its binary form to a more relaxed and unconstrained setting. Such attempts include the addition of shape parameters in order to add more flexibility to the probability distribution, while maintaining the straightforwardness provided in the logit-type models, with the least computational effort. A well-known form that provides an extra parameter to the base logit is the scobit model. In this study, we explore several generalizations of the binary scobit model by applying the same conventional methods associated with the generalized logit forms, principally to cover the multinomial nature of crash severity outcomes. Those are the multinomial and the ordinal forms. Furtherly, we utilize mixed distributions to provide crash-specific random parameters with heterogeneity in means and variances. Crash severity dataset taken from Guangdong province, China, was used to compare the different forms. The multinomial scobit models provided better results in terms of sample and out-of-sample fit, with the cost of some complexity in the heterogeneous forms. Other forms did not show a substantial or consistent advantage over their logit counterparts. All models exhibit temporal instability when applied to multiple time periods.