Enhanced Aggregate Framework to Model Crash Frequency by Accommodating Zero Crashes by Crash Type

Abstract

In recent years, joint count and fractional split model structure based approaches have emerged as a credible alternative for multivariate crash frequency dependent variables. However, current approaches in the fractional split theme have a limitation. The fractional split component of these frameworks allocates a proportion to all crash configurations. It is possible that, across spatial units, several crash configurations might have a large share of zero crashes. In the traditional multivariate context, in the presence of a high share of zeros, researchers employ zero-inflated or hurdle variants such as the zero inflated negative binomial model. The current research effort improves the fractional split based multivariate model systems with an explicit consideration for the potential presence of zeros by crash configuration. The newly included binary component can be employed to identify safer (or riskier) zones by crash configuration. The framework also accommodates unobserved heterogeneity across the components of the model system. The proposed model structure is estimated using zonal data from the Central Florida region, U.S., for 2016. The model considered six crash types—rear-end, angle, sideswipe, single-vehicle, multi-vehicle (three or more), and non-motorized crashes. The model estimation is conducted using an exhaustive set of independent variables. The model results clearly highlight the importance of accommodating zero crashes by crash type in the analysis. The model exercise is further augmented with a validation analysis.

A nested grouped random parameter negative binomial model for modeling segment-level crash counts

“…Another recent study developed an improved mixture model framework called joint negative binomial-multinomial logit fractional split to model crash counts by crash type for traffic analysis zones. The model accommodated zero crashes and incorporated random parameters to account for heterogeneity across traffic analysis zones [ 9 ]. Clearly, the random-parameter model is widely used in different applications.…”

Predicting the Type of Road Traffic Accident for Test Scenario Generation

“…Additionally, models regarding crash type, typically describing how road users hit each other, 1 seem to be part of the research focus [37], [38], [39], [40], [41]. In addition, police accident data have been used to predict types of victims involved [42] or to estimate the number of road traffic accidents expected to occur [43].…”