Exploring risky driving behaviors in severe taxi crashes : Accounting for unobserved heterogeneities using latent class clustering and partial proportional odds model

Objective: Taxis play an important role in public travel, yet limited concerns are devoted to taxi-involved safety compared with other travel modes. In particular, few studies have examined the heterogeneities of risky driving behaviors in taxi-involved crashes. This study quantifies the heterogeneity and characterizes the injury distribution of risky driving behaviors based on police-reported taxi crashes. Methods: Latent class clustering (LCC) is applied to identify the optimal number of clusters by maximizing heterogeneities across clusters. The number is determined by the Akaike information criterion (AIC), Bayesian information criterion (BIC), and entropy-based values of an LCC test. Partial proportion odds (PPO) models as well as marginal effects are tailored to quantify the unobserved heterogeneities based on different recognized sub-data sets. Results: The model outcomes of the combined approach highlight that submodel outcomes perform better than the whole data set. It is easy to determine the significant heterogeneities in risky driving behaviors across the characterized clusters. The top 3 contributing factors are dangerous overtaking, running red lights, and sudden acceleration/deceleration (maximum marginal effects exceeding +31%). There are significant heterogeneities in the top 3 factors across different clusters. Notably, aggressive driving behaviors display great heterogeneities, especially dangerous overtaking (moderate: +16.43%, severe: +36.88%). Conclusions: These results evidence that the submodels have better fitness than the whole model. These findings reveal the necessity of the combined approach and the remarkable heterogeneities across various clusters, providing insightful guidance to make cluster-specific countermeasures for road safety improvement.