Citation

Rezapour M, Mehrara Molan A, Ksaibati K. Int. J. Transp. Sci. Technol. 2020; 9(2): 89-99.

Copyright

(Copyright © 2020, Elsevier Publishing)

DOI

10.1016/j.ijtst.2019.10.002

PMID

unavailable

Abstract

A review of US mortality crashes revealed that motorcycle crashes are overrepresented in fatal crashes in the United States. Unlike passenger vehicles, motorcycles do not have much protection when it comes to crash involvement. In Wyoming, fatal or severe crashes account for 4% of all crashes, while 34% of motorcycle crashes are fatal or severe. Despite the heightened severity of motorcycle crashes, not much research has been conducted by using comprehensive methods to identify the contributory factors of such crashes. Furthermore, often the previous studies conducted analyses without evaluation of prediction accuracy of the models. Thus, this study evaluated the possibility of using both parametric and nonparametric methods in prediction of motorcycle at-fault injury severity. The goodness of the fit for the included methods was evaluated by comparing the performance of the models. Emphasis is given to mountainous highways where motorcycle use is high along with high crash incidence rates. Binary logistic regression as a parametric method, and the classification tree (CT) as a nonparametric method were employed in this study. Before conducting the analyses, an optimum set of included predictors was identified based on feature reduction. Also, in order to address the biased associated with selection of the test data, k-fold cross validation was used. The binary logistic regression has been used to analyze injury severity. However, this model relies on some very specific assumptions regarding the probability distribution, and logit link function. An alternative to binary logistic could be the classification tree (CT) which predicts injury severity based upon the set of predictors. The results indicated that although both models provided a comparable error rate binary logistic perform slightly better in prediction of motorcycle at-fault injury severity. The models performed similarly in terms of identified predictors. Some of the predictors that identified similarly with the two methods are posted speed limit, age, Highway functional class, and speed compliance.

Language: en

Keywords

Injury severity; Logistic regression; Motorcycle crashes; Decision tree; Machine learning techniques