New methods in crash severity analysis



Journal Title

Journal ISSN

Volume Title



This dissertation aims to conduct a multidimensional analysis of crash severity. The new methods are introduced to cope with critical issues that may arise from a lack of important crash information. The first issue concerns unobserved heterogeneity that is present in occupant injury analysis. The second issue is related to selectivity bias. Selectivity bias can occur when the estimation of the crash injury rate ignores information from property damage only crashes. The final issue is related to omitted variable bias that occurs in crash frequency by severity analysis as a result of human factors being omitted in the estimation process. I conducted three studies to investigate all three issues. I conducted three studies to investigate all three issues. This dissertation is comprised of three major parts as follows:

In the first part, I investigate factors influencing occupant injury severities by using a mixed logit model with heterogeneity in means and variances. The study is focused on the most severe outcome for an occupant in hybrid vehicle involved crashes that occurred on the Washington State road system during year 2006-2010. The model estimation results show that a wide range of variables influence the most injured occupant. The spectrum of severities includes no injury, possible injury, evident injury, incapacitating injury and fatal injury. The number-of-occupants parameter and the intersection-location indicator parameter are found to be random with significant heterogeneity in both means and variances. Sources of heterogeneity include the ratio of hybrid to non-hybrid vehicle counts in the crash, vehicle weight to horsepower ratio range (maximum difference in ratio) for the crash, number of adult occupants aged 41–64 years, functional class, and vehicle type interactions. The results further show the potential of models that address unobserved heterogeneity to unravel important relationships in the analysis of highway injury severities.

The second part of this dissertation involves crash severity analysis in terms of number of occupants. The goal of this study was to explore the potential for selectivity bias in estimating injury rates if property damage only (PDO) crashes were ignored. Injury analysis in the traffic safety field has traditionally been based on the analysis of injury crashes alone. This study questions the validity of that assumption by analyzing crash-specific injury rate measure as the estimation outcome while incorporating the effect of shared unobservable effects that might induce selection bias. Crash injury rate, which is determined as the ratio of the number of injured occupants to total occupants involved in the crash is modeled as a Tobit function. To account for the effect of unobserved effects that influence the likelihood of property damage only crashes as well as injury crashes, the Heckman selection model using a copula-based approach is applied to analyze the likelihood of injury crashes and the crash injury rate in two-vehicle crashes on Washington State roads during the period 2014- 2015. The estimation results show a significant correlation between the likelihood of injury crashes and the crash injury rate, which confirms the importance of including PDO crash information in the estimation. Moreover, the result suggests that using a copula-based approach to relax normality assumptions provides for better prediction.

The final part presents a methodological bridge approach to integrate human factors from the crash-specific severity outcome level or “conditional level” into the analysis at the frequency by severity level or “unconditional level”. This approach has never been done based on a survey of the published safety literature, but is crucial for the development of holistic insights on the marginal effects of human factors, traffic flow roadway geometrics on crash frequency. Traditional traffic safety analysis is divided into two distinct and unconnected areas – the conditional severity analysis in which factors specific to crashes are analyzed with respect to the probability of injury. In contrast, at the unconditional analysis, geometric, traffic flow and environmental factors are analyzed with respect to the number of crashes occurring in a time period. As a result of the disconnect, the marginal effect of roadway geometrics and traffic flow can be biased and inaccurate, and in many cases, overestimated. This in turn affects safety policy in terms of safety interventions, because a geometric safety intervention may look too promising when in fact its effect may be diminished when one accounts for human factors.

Integrating the conditional and unconditional dimensions will allow for incorporation of human factors such as sobriety of driver, safety belt use, driver fault, occupant age and gender information, and vehicle type information. In the absence of these types of data, unconditional frequency analysis suffers from unobserved effects that can bias our inferences on the true effect of geometric attributes of a highway segment or an intersection location. Human factors in this study included impaired drivers, distracted drivers, and no-belted drivers by proportion. These factors were extracted from crashes and integrated long with geometric and traffic flow factors into the analysis of total crashes which happened on the Washington State road system during the period 2014-2015. The analysis was conducted at the one-mile segment level using a random parameter framework. The findings reveal that the human factors improved the model predictions significantly, especially for property damage injury, and minor injury. Among the human factors variables studies, the impaired drivers by proportion parameters were found to be random with respect to the likelihood of crash occurrences depending on whether the proportion is more than 8%. Furthermore, the paper also investigates factors influencing the random parameters’ means and variances. The area effects such as crashes that happened in a specific county are found to influence the random parameters means, while roadway types are found to influence the variance of the random parameters. The study also suggests that ignoring human factors could lead to omitted variable and specification bias.

Through the intensive investigation, this dissertation answers several unanswered dimensions of crash severity analysis. The proposed approaches provide significant improvement in model estimations by accounting for important issues such as unobserved heterogeneity, selection bias, and omitted human factor effects.



Unobserved heterogeneity, Random parameters, Copula, Selection bias