Heydari, Mohammad (2012) A Full Bayes Approach to Road Safety: Hierarchical Poisson Mixture Models, Variance Function Characterization, and Prior Specification. Masters thesis, Concordia University.
Heydari__MSc_S2012.pdf - Accepted Version
Road safety is a major concern of every department of transportation. Allocating resources to improve safety requires the identification of hazardous sites (hotspots) and the assessment of safety countermeasures. For such tasks, reliable safety performance functions are noteworthy to predict collisions, prioritize improvements, and capture countermeasure effectiveness for overall road network management.
In this thesis, a case study from New Brunswick was used. Bayesian statistics were mainly applied in analyses by introducing Poisson mixture models in a hierarchical fashion. Poisson and Poisson mixture models were compared. Different characterizations of variance functions were verified. In a novel approach, the inverse of variance in Poisson-Lognormal models was examined to vary across sites as a function of site characteristics. In addition, accidents were analyzed by severity. Hierarchical Poisson-Gamma models presented the best fit. Traffic flow was the most influential factor in variance functions. Models with random variance structure provided the best fit, followed by those varying as a function of site characteristics. The interaction between precipitation and density of horizontal curves was statistically significant only for injury-fatality accidents - these contributing factors weren’t significant when considered separately.
Additionally, the effect of prior specifications in hierarchical Poisson-Gamma models was examined adopting a case study and a data simulation framework. Results showed that informative priors, especially for the inverse dispersion parameter, improve the accuracy of parameter estimates. Data with low sample mean and small sample size were dramatically affected by prior specification. However, hotspot identification and goodness-of-fit were not very sensitive to prior choice.
|Divisions:||Concordia University > Faculty of Engineering and Computer Science > Building, Civil and Environmental Engineering|
|Item Type:||Thesis (Masters)|
|Degree Name:||M.A. Sc.|
|Date:||13 April 2012|
|Thesis Supervisor(s):||Amador, Luis|
|Deposited By:||MOHAMMAD HEYDARI|
|Deposited On:||18 Jun 2012 14:49|
|Last Modified:||05 Nov 2016 04:46|
Repository Staff Only: item control page