Item – Theses Canada

​
OCLC number
429724842
Link(s) to full text
LAC copy
LAC copy
Author
Miranda-Moreno, Luis F.,1974-
Title
Statistical models and methods for identifying hazardous locations for safety improvements.
Degree
Ph. D. -- University of Waterloo, 2006
Publisher
Ottawa : Library and Archives Canada = Bibliothèque et Archives Canada, [2007]
Description
3 microfiches
Notes
Includes bibliographical references.
Abstract
Traffic safety in transportation networks is one of the main priorities for many government agencies, private organizations and the society as a whole. This is due mainly to the huge monetary and non-monetary costs caused by traffic accidents. Recognizing the need to reduce the costs of traffic accidents, many safety improvement programs sponsored and headed by national or provincial transportation agencies are continuously carried on in Canada and other countries. These programs are often targeted to hazardous locations or hotspots where the risk of accidents is unacceptably high and safety countermeasures are most warranted. Hotspots are usually identified through a screening process that first ranks the locations of interest based on their safety status and then selects a subgroup of locations for subsequent detailed engineering inspections. Although there is an extensive literature on the topic of hotspot identification, several important issues still remain. For instance, a number of statistical models and methods have been proposed for safety analysis in general and hotspot identification in particular; however, few studies have been conducted to systematically assess the accuracy and implications of these alternative models and methods. Moreover, past studies have mainly focused on identifying the appropriate ranking criteria with few focusing on the issue of how to establish selection rules for hotspot identification. The goal of this thesis research was to address some of these issues through a series of new methodologies and analyses. In this thesis, we have first investigated the implications of applying alternative Poisson random effect models and ranking criteria using the Empirical Bayes (EB) approach for hotspot identification. In particular, we have introduced the use of various models as an extension to the most commonly used Negative Binomial model for safety analysis. We have also introduced the use of two hierarchical Poisson/Lognormal models, as an alternative to the hierarchical Poisson/Gamma model. In order to provide a rigorous assessment of the performance of these alternative models and methods, we have developed a simulation-based methodology that represents a novel approach for comparing alternative accident risk models and safety analysis methods. The overall results of our investigation showed that, when working with datasets with a low mean accident frequency and small sample sizes, the impact of the hyper-prior assumption became evident. It was also found that under these critical conditions, the hierarchical Poisson/Lognormal models performed better than the Poisson/Gamma model when diffuse priors were assumed on the dispersion parameter. Based on a simulation study, we have also investigated the performances of the Empirical Bayes versus full Bayes methods. Based on this, we found that the full Bayes estimators performed better than the EB estimators when working with datasets with small number of sites and characterized by an overall low mean accident frequency. However, when the dataset is sufficiently large, these two approaches yielded practically the same results. We have also proposed a systematic methodology for identifying hotspots based on a threshold-based strategy which may be employed under various safety measures. The proposed methodology is based on the concept of multiple hypotheses testing, enabling one to control appropriate global error rates in optimizing the thresholds. We have also suggested two new safety indicators that integrate the total accident costs, and the "anticipated" cost-benefit ratio. We introduced a screening method based on the posterior distribution of ranks. Based on our analyses, we observed that this method performed very well over a broad range of models and multiple testing methods. Finally, the performance of alternative posterior approaches and error minimization objectives was also investigated. We found that the posterior distribution of ranks performed slightly better than the absolute posterior distribution of the safety measure of interest.
ISBN
9780494236789
0494236787