METHODOLOGICAL APPROACHES TO INCORPORATE HETEROGENEITY IN TRAFFIC ACCIDENT SEVERITY
Open Access
- Author:
- Shyu, Ming-Bang
- Graduate Program:
- Civil Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 06, 2006
- Committee Members:
- Venkataraman Shankar, Committee Chair/Co-Chair
Martin T Pietrucha, Committee Member
Thorsten Wagener, Committee Member
Evelyn Ann Thomchick, Committee Member - Keywords:
- Covariance Heterogeneity Model
Nested Logit Model
Accident Severity Model
Heterogeneity
Heteroskedastic Extreme-Value Model - Abstract:
- Abstract Methodological Approaches to Incorporate Heterogeneity in Traffic Accident Severity Models Ming-Bang Shyu Chair of the Supervisory Committee: Associate Professor Venkataraman N. Shankar Scope of the Problem Fatal accidents exact a significant toll in terms of economic cost, in excess of 150 billion dollars yearly. Over 40,000 drivers, passengers, pedestrians and bicyclists are killed each year on United States highways. Traffic accidents cause what is termed in the medical literature as unintentional harm and injury. Only heart attacks, cancer, stroke and respiratory related illnesses cause more deaths in the United States than unintentional injuries. While traffic accident deaths may comprise less than 2 percent of all registered deaths annually, their impact on future income earners can be telling. As an example, over 40 percent of childhood deaths are due to unintentional injuries, with over 30 percent contributed by motor vehicle accidents. Among teenagers, the three leading causes of death are unintentional injuries, homicide and suicide. Nearly 52 percent of the 15-24 age group dies from unintentional injuries, with a significant number perishing in motor vehicle accidents. From an infrastructure standpoint, fatalities contribute to increase in lifecycle costs including transportation, social and emergency infrastructure. Given this backdrop, my goal in this dissertation is to parse out the contribution of infrastructure to motor vehicle related deaths. A 1993 study by the Carter Center estimated approximately 25,000 deaths annually to be behaviorally related. In a traffic accident context, common cited reasons that constitute behavior include speeding, driving under the influence of alcohol, driving without seat belt fastened, driving under fatigue, aggressive driving such as tailgating, and failure to yield. What are of interest in this dissertation are the impact of infrastructure in these deaths, as well as infrastructure impact in single and multi-vehicle collisions leading to death. For example, fixed object related collisions contribute to nearly 27 percent of all motor vehicle deaths, while multi-vehicle collisions contribute to almost 45 percent of motor vehicle deaths. I formulated the hypothesis that a variety of factors relating to human, roadway and vehicle effects are associated with motor vehicle accident injuries. I attempted to identify those that are strongly associated with injury severities. A focused study on single- and two-vehicle driver occupant only accidents using empirical data from the Washington State Patrol¡¦s accident database was conducted. I compiled over a 79-month period in Washington State from 1990 to 1996, detailed accident reports on over 127,000 cases. Objectives A multi-variate analytical framework that is robust and helps identify the marginal impact of important policy variables related to seat belt use, drunk driving enforcement and driving age related issues, while controlling for vehicle and roadway influences, was developed. It is also our objective to develop a framework with commonly available data without placing undue demands on data collection. Such a method will enhance the portability of our approach to be applicable to a variety of locales. Method Statistical methods relating to the analysis of ordinal and discrete outcomes were employed. The developed models also incorporated heterogeneity. Heterogeneity refers to effects that are not measured for various reasons. In our context, not measured implies not measurable, could be measured but was not measured for economic reasons, as well as unknown and hence not measurable. The impact of heterogeneity and correlation that exists in severity contexts is at the very least, loss in statistical efficiency of parameters in the model. As a result, strong associations can be imprecisely identified. Using a variety of techniques within this broad category of analysis, common denominator variables that were found to be strongly associated with driver only occupant severities were identified. These methods have been embraced by WSDOT as potential frameworks for implementing their safety project prioritization plan. Three model types known as extensions of the generalized extreme value model were examined. The multinomial logit is the simplest and most popular form. However, its structure impedes incorporation of heterogeneity. By definition, the multinomial logit assumes all outcomes are identically and independently influenced by random effects that are unobserved. As alternatives, in order to address the heterogeneity problem, the nested logit, the heteroskedastic logit and the covariance heterogeneity logit structures were examined. These structures are uniquely flexible in accommodating heterogeneity. The idea behind examining these structures is the need for robustly identifying a set of strong associations in terms of infrastructure variables. Results Factors relating to driver sobriety, seat belt use, human error in driving, vehicle type, type of collision and type of object struck appeared to strongly associate with injury. The findings reinforce in a single multi-variate framework insights from case-controlled studies on seat belt use and driver sobriety. Over 300,000 individual accidents were initially examined, and culled to include 127,000 accidents for final model development. Separate models of injury outcomes were developed for single-vehicle and two-vehicle accidents. Several hundred model specifications were tested prior to the finalization of model structures. Due to the variety of structures that are possible within the nested logit class of models, the modeling requirement extended to over a thousand specifications in order to identify the preferred structure. The nested logit analysis showed that after substantial testing heterogeneity and correlation effects are not clearly accommodated using a nested logit structure, thereby creating an argument for more sophisticated and flexible structures such the heteroskedastic and covariance heterogeneity models. Due to data constraints, multi-vehicle accidents involving three or more vehicles were not addressed. Furthermore, in two-vehicle accidents, vehicle mass difference effects are distinct, if they in fact exist as strong associations. Conclusion The benefits of this research are numerous ¡V it presents a multi-variate analytical framework that is robust by incorporating the heterogeneity issue in modeling and helps identify the marginal impact of important policy variables related to seat belt use, drunk driving enforcement and driving age related issues, while addressing critical infrastructure issues as well. For example, addressing the sensitivity of injury probabilities to the removal of fixed objects is a critical infrastructure issue. A decision maker can use the results of this model to estimate benefits in terms of societal cost reductions and compute the benefit cost of fixed object removals or collision protection. In addition, this research also highlights the importance of data types that need to be collected for robust policy development on traffic accident injury prevention. The nested logit model was suggested as the common denominator model to incorporate unobserved heterogeneity between PDO and PINJ. This important modeling capability has the potential to significantly enhance statewide consistency in infrastructure related decision making.