An efficient parallel sampling technique for Multivariate Poisson-Lognormal model: Analysis with two crash count datasets. (December 2015)

• Record Type: Journal Article
• Authors: Zhan, Xianyuan; Aziz, H.M.Abdul; Ukkusuri, Satish V.
• Is Part Of: Analytic methods in accident research. Volume 8(2016)
• Page Start: 45
• Page End: 60
• Publication Date: 2015-12
• Subjects: Accident analysis -- Pedestrian crashes -- Severity models; Accidents -- Research -- Methodology -- Periodicals; Accidents -- Prevention -- Periodicals; 363.100721
• DOI: 10.1016/j.amar.2015.10.002
• Languages: English
• ISSNs: 2213-6657

Abstract

Abstract: This study investigates the Multivariate Poisson-lognormal (MVPLN) model that jointly models crash frequency and severity accounting for correlations. The ordinary univariate count models analyze crashes of different severity level separately ignoring the correlations among severity levels. The MVPLN model is capable to incorporate the general correlation structure and also takes account of the overdispersion in the data that leads to a superior data fitting. However, the traditional estimation approach for MVPLN model is computationally expensive, which often limits the use of MVPLN model in practice. In this work, a parallel sampling scheme is introduced to improve the original Markov Chain Monte Carlo (MCMC) estimation approach of the MVPLN model, which significantly reduces the model estimation time. Two MVPLN models are developed using the pedestrian–vehicle crash data collected in New York City from 2002 to 2006, and the highway-injury data from Washington State (5-year data from 1990 to 1994) The Deviance Information Criteria (DIC) is used to evaluate the model fitting. The estimation results show that the MVPLN models provide a superior fit over univariate Poisson-lognormal (PLN), univariate Poisson, and Negative Binomial models. Further, the correlations among the latent effects of different severity levels are found significant in both datasets, that justifies the importance of jointly modeling crash frequency and severity accounting for correlations. Highlights: A MVPLN framework is used to jointly model different severity levels of crashes. The applicability of the model is shown for the bivariate case using NYC pedestrian crash data and the multivariate case using Washington State highway crash data. An efficient parallel computing MATLAB code is developed to estimate the MVPLN model. The MVPLN model shows superior fitting results compared with the univariate models. High level of correlation between fatal and severe injury pedestrian–vehicle crashes is observed.