A Multivariate Poisson-Lognormal (MVPLN) Model for Pedestrian-Vehicle Crashes in New York City Accounting for General Correlations Among the Severity Levels

This study estimates a multivariate Poisson-lognormal (MVPLN) model using the New York City pedestrian-vehicle crash data collected from 2002 to 2006. The data is aggregated to census tract level. The MVPLN model overcomes the limitations of the ordinary univariate count models that analyze crashes of different severity level separately and ignores the correlations among different crashes severity levels. In addition, the MVPLN model can capture the general correlation structure in crashes frequency data, and takes account of the over-dispersion in the data, which provides a superior fitting result. A MATLAB code implementing parallel computing is developed to estimate the MVPLN model via a Markov Chain Monte Carlo (MCMC) approach. A comparison study is conducted to compare the model fit of MVPLN, univariate Poisson-lognormal, univariate Poisson and Negative Binomial model, and the estimation results show a better fit of the pedestrian-vehicle crash data.