Multivariate Poisson Lognormal Modeling of Weather-Related Crashes on Freeways

Abstract

Adverse weather conditions are one of the primary causes of motor vehicle crashes. To identify the factors contributing to crashes during adverse weather conditions and recommend cost-effective countermeasures, it is necessary to develop reliable crash prediction models to estimate weather-related crash frequencies. To account for the variations in crash count among different adverse weather conditions, crash types, and crash severities for both rain- and snow-related crashes, crash data on freeways was collected from the State of Connecticut, and crash prediction models were developed to estimate crash counts by crash type and severity for each weather condition. To account for the potential correlations among crash type and severity counts due to the common unobserved factors, integrated nested Laplace approximation (INLA) multivariate Poisson lognormal (MVPLN) models were developed to estimate weather-related crashes counts by crash type and severity simultaneously (four MVPLN models were estimated in total). To verify the model prediction ability, univariate Poisson lognormal (UPLN) models were estimated and compared with the MVPLN models. The results show that the effects of factors contributing to crashes, including median width, horizontal curve, lane width, and shoulder width, vary not only among different adverse weather conditions, but also among different crash types and severities. The crash types and severities are shown to be highly correlated and the model comparison verifies that the MVPLN models significantly improve the model prediction accuracy compared with the UPLN models. Therefore, the MVPLN model is recommended to provide more unbiased parameter estimates when estimating weather-related crashes by crash type and severity.