Modeling Unobserved Heterogeneity for Zonal Crash Frequencies: A Bayesian Multivariate Random-Parameters Model with Mixture Components for Spatially Correlated Data

Helai Huang,Fangrong Chang,Hanchu Zhou,Jaeyoung Lee
DOI: https://doi.org/10.1016/j.amar.2019.100105
IF: 14.556
2019-01-01
Analytic Methods in Accident Research
Abstract:This study applies mixture components in a multivariate random parameters spatial model for zonal crash counts. Three different modeling formulations are employed to demonstrate the effects of mixture components and spatial heterogeneity in the goodness-of-fit in a multivariate random parameter model. The models are built for injury (i.e., possible, non-incapacitating, incapacitating, and fatal injury) and non-injury crashes using the data from 738 traffic analysis zones (TAZs) in Hillsborough County of Florida during a three-year period. The Deviance Information Criteria (DIC) is used to evaluate the performances of these models indicate the proposed model outperforms the rests. According to the estimated results, various traffic-related, demographics, and socioeconomic factors affect the occurrences of crashes for different severity levels. With regard to the effect of mixture components, it identifies two homogeneous sub-classes labeled as "stable pattern" and "unstable pattern" to better capture the heterogeneity. The standard deviation (SD) and correlation across injury and non-injury crashes are both very high in the "stable pattern" compared with its "unstable pattern" counterpart. On the other hand, the results of model comparison reveal that: (i) adding one more mixture component has no significant influences on the spatial heterogeneity and spatial correlation of different kinds of crash frequency and (ii) the consideration of spatial effects improves the accuracy of estimate results. Moreover, the multivariate random parameters spatial model with mixture components was compared with its univariate form to highlight the validity of applying multivariate structure. (C) 2019 Elsevier Ltd. All rights reserved.
What problem does this paper attempt to address?