Exploring bicyclist injury severity in bicycle-vehicle crashes using latent class clustering analysis and partial proportional odds models.

• This paper considers unobserved or latent features when modeling injury severity. • Latent class analysis is used to segment crash data to reduce heterogeneity. • Partial proportional odds models are developed to explore bicyclist injury severity. • Some variables affect bicyclist injury severity differently between various clusters. • Policy-related recommendations and advices on traffic management are provided. Introduction: Bicyclists are more vulnerable compared to other road users. Therefore, it is critical to investigate the contributing factors to bicyclist injury severity to help provide better biking environment and improve biking safety. According to the data provided by National Highway Traffic Safety Administration (NHTSA), a total of 8,028 bicyclists were killed in bicycle-vehicle crashes from 2007 to 2017. The number of fatal bicyclists had increased rapidly by approximately 11.70% during the past 10 years (NHTSA, 2019). Methods: This paper conducts a latent class clustering analysis based on the police reported bicycle-vehicle crash data collected from 2007 to 2014 in North Carolina to identify the heterogeneity inherent in the crash data. First, the most appropriate number of clusters is determined in which each cluster has been characterized by the distribution of the featured variables. Then, partial proportional odds models are developed for each cluster to further analyze the impacts on bicyclist injury severity for specific crash patterns. Results: Marginal effects are calculated and used to evaluate and interpret the effect of each significant explanatory variable. The model results reveal that variables could have different influence on the bicyclist injury severity between clusters, and that some variables only have significant impacts on particular clusters. Conclusions: The results clearly indicate that it is essential to conduct latent class clustering analysis to investigate the impact of explanatory variables on bicyclist injury severity considering unobserved or latent features. In addition, the latent class clustering is found to be able to provide more accurate and insightful information on the bicyclist injury severity analysis. Practical Applications: In order to improve biking safety, regulations need to be established to prevent drinking and lights need to be provided since alcohol and lighting condition are significant factors in severe injuries according to the modeling results. [ABSTRACT FROM AUTHOR]