A temporal investigation of crash severity factors in worker-involved work zone crashes: Random parameters and machine learning approaches. (June 2021)

• Record Type: Journal Article
• Authors: Mokhtarimousavi, Seyedmirsajad; Anderson, Jason C.; Hadi, Mohammed; Azizinamini, Atorod
• Is Part Of: Transportation research interdisciplinary perspectives. Volume 10(2021)
• Publication Date: 2021-06
• Subjects: Work zone safety -- Crash severity -- Mixed Logit Model -- Support vector machine -- Cuckoo search optimization algorithm; Transportation -- Periodicals; 388.05
• DOI: 10.1016/j.trip.2021.100378
• Languages: English
• ISSNs: 2590-1982

Abstract

Highlights: Crash severity analysis of work zone crashes involving workers. Investigation the impacts of contributing factors from machine learning perspective. Study of unobserved heterogeneity among contributing factors. Instability check of parameter estimates for daytime and nighttime models. Assessing the capability of Cuckoo Search algorithm in SVM parameter tuning. Abstract: In the context of work zone safety, worker presence and its impact on crash severity has been less explored. Moreover, there is a lack of research on contributing factors by time-of-day. To accomplish this, first a mixed logit model was used to determine statistically significant crash severity contributing factors and their effects. Significant factors in both models included work-zone-specific characteristics and crash-specific characteristics, where environmental characteristics were only significant in the daytime model. In addition, results from parameter transferability test provided evidence that daytime and nighttime crashes need to be modeled separately. Further, to explore the nonlinear relationship between crash severity levels and time-of-day, as well as compare the effects of variables to that of the logit model and assess prediction performance, a Support Vector Machines (SVM) model trained by Cuckoo Search (CS) algorithm was utilized. Opening the SVM black-box, a variable impact analysis was also performed. In addition to the characteristics identified in the logit models, the SVM models also included the impacts of vehicle-level characteristics. The variable impact analysis illustrated that the termination area of the work zone is most critical for both daytime and nighttime crashes, as this location has the highest increase in severe injury likelihood. In summary, results of this study demonstrate that work zone crashes need to be modeled separately by time-of-day with a high level of confidence. Furthermore, results show that the CS-SVM models provide better prediction performance compared to the SVM and logit models.