Investigation of Trends and Predictive Effectiveness of Crash Severity Models

This thesis describes analysis using ordinal logistic regression to uncover temporal patterns in the severity level (fatal, serious injury, minor injury, slight injury or no injury) for persons involved in highway crashes in Connecticut, focusing on the demographic split between senior travelers (65 years and over) and non-senior travelers. Existing state sources provide data describing the time and weather conditions for each crash and the vehicles and persons involved over the time period from 1995 to 2009 as well as the traffic volumes and the characteristics of the roads on which these crashes occurred. Findings indicate an overall increase in increased crash severity probability for seniors, as well as a distinct seasonal trend. Other time-dependant trends in the data were visible, but not significant. Additionally, this research investigates the use of partial proportional odds (PPO) as a statistical modeling technique to compromise between ordinal and multinomial modeling techniques and to avoid violating the key assumptions in the behavior of crash severity inherent in these two alternatives. The PPO technique is compared to ordinal and multinomial response models on the basis of model fit, covariate values and significance, and prediction accuracy. The results of the study show that the PPO model has adequate model fit and performs at least as well as either the multinomial or the ordinal model in terms of covariate significance and prediction accuracy, and thus is useful as method of predicting the severity of crashes