Date of Award

2014

Publication Type

Master Thesis

Degree Name

M.A.Sc.

Department

Civil and Environmental Engineering

Keywords

Boosted Regression Tree, Heteroscedastic Ordered Logit, Injury Severity, Non-parametric Model, Parametric Model, Single- and Two- vehicle Crash

Supervisor

Lee, Chris

Rights

info:eu-repo/semantics/openAccess

Abstract

This study applies both parametric model (Heteroscedastic Ordered Logit (HOL)) and non-parametric models (Random Forest, Classification and Regression Tree (CART), and Boosted Regression Tree (BRT)) to analysis of driver's injury severity in single-vehicle and two-vehicle crashes on highways. The HOL model not only estimates quantitative effects of significant explanatory variables, but also captures heteroscedasticity (i.e. variation in the unobserved effects among observations) of the variables such as head-on collision, abnormal conditions and female drivers. On the other hand, the BRT model effectively captures nonlinear effects of continuous variables including truck percentage, AADT, driver's age and vehicle age on severe injury. It was found that the BRT model predicted driver's injury severity more accurately than the HOL and CART models for both single-vehicle and two-vehicle crashes. Based on the model results, some remedial treatments are discussed to reduce driver's injury severity in crashes on highways. It is recommended that both HOL and BRT models are used for more accurate prediction of crash injury severity.

Share

COinS