Methodological Approaches to Incorporate Heterogeneity in Traffic Accident Frequency Models
Open Access
- Author:
- Sittikariya, Sittipan
- Graduate Program:
- Civil Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 31, 2006
- Committee Members:
- Venkataraman Shankar, Committee Chair/Co-Chair
Martin T Pietrucha, Committee Member
Thorsten Wagener, Committee Member
Evelyn Ann Thomchick, Committee Member - Keywords:
- Median crossover accidents
Heterogeneity
Negative Binomial
Random effects Poisson
Bayesian modeling - Abstract:
- This dissertation addresses the impact of critical modeling issues on the statistical significance of specifications correlated with traffic accident frequencies. Also, this research intends to shed new light on the major factors contributing to the occurrence of median crossover accidents, which are particularly severe. Furthermore, the research aims to lay out methodologies that provide significant improvements in frequency predictions. The entire highway network of Washington State is used as the empirical setting. The database was developed over several years of observations and includes median crossover accident data, geometric and traffic volume information as well as weather data. In traffic accident databases that are comprised of cross-sectional panels, unobserved effects or heterogeneity, correlation due to shared unobserved effects through accident counts, and excess zero problems are common issues in the estimation of accident count models. I establish various model structures through classical frequentist and Bayesian approaches to explore these issues. Several hundred modeling specifications were assessed to determine the most meaningful and robust specifications. In the final analysis, the same set of specifications was used to compare the results across model structures. Preliminary results showed variabilities in statistical significance of key parameters; but the parameters themselves do not change significantly. Much of the variability in the standard errors of parameters results from error structure assumptions. For example, hierarchical structures, or multiplicative structures that are specified exogenously through Bayesian methods help improve predictions. The major contribution of this dissertation is the development of modeling taxonomies where assumptions under frequentist and Bayesian methods are incorporated through examination of model equivalencies. Such taxonomies are helpful in illustrating with greater clarity the relative usefulness of frequentist and Bayesian methods as they relate to all types of accident frequencies. In this sense, this dissertation offers a general framework. In addition, from a programming standpoint, the efficiency and consistency of highway safety prioritization can be broadly addressed through this framework.