Modeling of incident type and incident duration using data from multiple years

Abstract

The paper presents a model system that recognizes the distinct traffic incident duration profiles based on incident types. Specifically, a copula-based joint framework has been estimated with a scaled multinomial logit model system for incident type and a grouped generalized ordered logit model system for incident duration to accommodate for the impact of observed and unobserved effects on incident type and incident duration. The model system is estimated using traffic incident data from 2012 through 2017 for the Greater Orlando region, employing a comprehensive set of exogenous variables, including incident characteristics, roadway characteristics, traffic condition, weather condition, built environment and socio-demographic characteristics. In the presence of multiple years of data, the copula-based methodology is also customized to accommodate for observed and unobserved temporal effects (including heteroscedasticity) on incident duration. Based on a rigorous comparison across different copula models, parameterized Frank-Clayton-Frank specification is found to offer the best data fit for crash, debris, and other types of incident. The value of the proposed model system is illustrated by comparing predictive performance of the proposed model relative to the traditional single duration model on a holdout sample.