Modeling unobserved heterogeneity in motor vehicle crash injury severity data
The American Association of State Highway Transportation Officials (AASHTO) has established a goal to halve the national number of highway fatalities by 2027. In order to fulfill the states' portion of the goal, efforts are needed on building sophisticated crash injury data analysis methodologies for reliable safety hazards identification in development of state and local safety programs. On account of the considerable amount of unobserved and omitted on-the-spot information in crash datasets used by agencies, the issue of unobserved heterogeneity in crash data modeling has been identified and has attracted growing attention in recent years. Prior studies on relationships between highway safety elements and crash injury severity outcomes have suggested that effects of contributing factors in different situations may be non-homogenous. However, little is understood about the dynamics. This dissertation aims to contribute to the literature by (a) investigating how effects of hazardous factors vary across road segments and over time periods and (b) how they would interact with the effects of other factors on crash injury severity outcomes, with accommodation of unobserved heterogeneity in different levels and without prespecified assumptions on probability distributions. The analysis went beyond the heterogeneous effects formulation and included the model estimation details based on Bayesian inference. This dissertation consists two studies: (1) A particular case of cross-sectional unobserved heterogeneity modeling for a safety intervention program was studied by using Indiana adolescent crash data. A Markov Chain Monte Carlo (MCMC) algorithm was developed for estimation and a permutation sampler was extended for model identification. (2) A general case of time-varying unobserved heterogeneity modeling was carried out based on Indiana rural interstate crash data. Reparameterization and partially marginalized conditional samplers techniques were designed to reduce autocorrelation between consecutive draws and to improve the convergence efficiency of chains in estimation simulation. The implications for implementation of regulation enforcement and highway infrastructure upgrade and maintenance were discussed. The empirical results can provide substantial insights to government agencies that are concerned about strategic programming of safety countermeasures to leverage safety intervention resources. The methodologies set forth herein should be of interest to individuals who are developing analysis tools for crash cause diagnosis in state and local transportation safety programs, and have the potential for valuable new insights into a wide variety of questions in discrete data modeling.
Mannering, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our