Modelling two-vehicle crash severity by generalized estimating equations

The crash severity levels of two parties involved in a two-vehicle accident may differ markedly and may be correlated. Separately estimating the severity levels of two parties ignoring their potential correlation may lead to biased estimation; however, modelling their severity levels simultaneously by using a bivariate modelling approach requires a complex model setting. Thus, this study used generalized estimating equations (GEE) to accommodate potential correlations when estimating the crash severity levels of two parties. To investigate the performance of the GEE models, a case study on a total of 2493 crashes at 214 signalized intersections in Taipei City in 2013 is conducted. Univariate ordered probit model, bivariate ordered probit model, and GEE ordered probit model (GEE-OP) with different working matrices are respectively estimated and compared. The estimation results of GEE models showed that the GEE-OP with the exchangeable working matrix performs best and the most influential factor contributing to crash severity is vehicle type (motorcycle), followed by speeding, angle impact, and alcoholic use. Thus, to curtail motorcycle usage by increasing parking fee or reducing parking space of motorcycles, to crack down on speeding and alcoholic use, and to redesign the signal timings to avoid possible angle impact accidents are identified as key countermeasures.