Generalized extreme value and mixed logit models: Empirical applications to vehicle accident severities.

摘要

This dissertation explores the usefulness of a flexible econometric method, namely the mixed logit model in its applicability to statewide vehicle collision severity modeling. The research motivation is formed out of the need to develop comprehensive framework for programming system-wide highway safety improvements. Prior evidence in published literature points to the suitability of nested logit structures for modeling accident severities, albeit in a conditional context. In this research, an alternate flexible approach for modeling severities is proposed. Unconditional severity models such as the proportion of severe accidents by frequency are considered in the investigation of the contemporaneous effects of roadway, environmental, and traffic factors on collision severities. Five distinct collision severities that can result from a collision are modeled (fatal injury, disabling injury, evident injury, possible injury, and property damage only). The advantage of the unconditional severity model approach is twofold. First, the models are directly implementable in a statewide safety programming framework that is built on frequency models. Second, the models can provide direct results on the costs and benefits of severe accidents. An empirical dataset pertaining to divided highways involving unbarriered roadway sections is used to demonstrate the suitability of the proposed modeling approach. The nested logit was estimated by full information maximum likelihood (FIML) techniques. A mixed logit was then estimated using the nested logit as a benchmark specification. The suitability of the mixed logit in terms of its ability to explain variabilities in roadway section characteristics and how they affect collision severity proportions is discussed in detail. Simulation based estimation methods unique to the mixed logit due to its non-closed form are also discussed. Specifically, the nature of draws such as random and Halton draws used for sampling differing parameter distributions were a central issue for examination. Findings from this dissertation will provide insight into the complex interactions and the effects of spatial, temporal, environmental, geometric and traffic flow factors affecting collision severity proportions. This, it is hoped, will not only extend the envelope of academic thought on severity modeling, but also provide much needed direction for decision makers in a statewide severity context.