Application of multi-group structural equation modelling for investigation of traffic barrier crash severity

摘要

The severity of traffic barrier in the literature has been modelled considering different factors including human, environmental and road/traffic barrier characteristics. However, all these factors are interacting in a complicated way, and a real relationship between these factors is still unclear. A structural equation modelling (SEM) can be adopted to capture the intricate relationships between the contributory factors and latent (unseen) factors. This study was conducted by adopting multi-group SEM to unlock the complicated relationship between confounding factors and traffic barrier crash severity by considering differences across two important groups. Due to the possible difference across different highway systems, multi-group SEM was used instead of standard SEM to account for the differences across highway and interstate roadway system. SEM is a combination of confirmatory and path analysis, which could examine relationship between different observed and latent factors. Besides using factor analysis for identification of latent factors, item/variable cluster analysis was conducted to identify all the latent factors. Although cluster analysis often has been used in other fields, this is the first time this method has been applied in transportation problems for SEM modeling. The inclusion of the factors identified by cluster analysis show an improvement in goodness of fit. This study was conducted to evaluate the traffic barrier crash severity in terms of death, injury and severity of crashes. It examined the nature and causes of severe traffic barrier in Wyoming. The results indicated that different factors contribute to the severity size of traffic barrier crashes including different traffic barrier types, demographic characteristics, weather conditions, and indirect impact of force direction. The results indicated that collision force is a latent factor with highest impact on crash severity compared with other latent factors. Different models with different number of latent were compared based on different goodness-of-fit indices and a best model, with an acceptable model fit, was selected between them. A more discussion about the model presented in the manuscript.