Development of Crash Prediction Model Considering Characteristics of Ramp Types

Abstract

Freeway ramps have a relatively higher number of crashes per unit traffic and distance than freeway mainlines. Moreover, freeway ramps also have different crash characteristics and geometric designs than those of the mainline. As such, a crash prediction model specific to freeway ramps should be developed. The trumpet interchange, where more than 70% of ramp crashes occur, has six ramp types, based on their ramp configurations (loop/semi-direct/direct) and ramp functions (on/off ramp). Of these six types, each ramp type has a different crash rate. Therefore, the crash prediction model to be developed must consider the different ramp-type characteristics. Most previous studies have used a generalized linear model when developing a crash prediction model, and this type of model cannot take into account different ramp-type characteristics. This is because the generalized linear model assumes that every piece of data observed is independent of any other. The purpose of this study is to develop a crash prediction model based on a multilevel model to address the limitations of the conventional model. A multilevel model can consider intra-class correlations derived from several individuals included in a group. Thus, this model has the advantage of being able to take into account different ramp-type characteristics. In this study, the ramp crash data including 1,155 crashes were obtained from 2007 to 2010 for that occurred on three Korean freeway lines (Kyungbu, Yeongdong, and Seohaean lines) that had the highest number of crashes among all Korean freeway lines. Then a multilevel model is developed that considers different ramp-type characteristics and a generalized linear model that does not consider these characteristics. The annual average daily traffic (AADT), ramp length, and ramp-type dummy variables are used as the independent variables in the models. The multilevel model consists of a 1-level model for individual ramps and a 2-level model for other ramp types. The generalized linear model has the same structure as the 1-level model in the multilevel model. The parameter estimation results in the multilevel model include the different crash risk for each ramp type and the impacts of different ramp lengths on the number of crashes, according to type. In contrast, the generalized linear model estimates the same parameters for all ramp types and does not reflect the different type characteristics. To validate the models, quantitative indicators including root mean square error (RMSE), median absolute deviation (MAD), and cumulative residuals (CURE) plot are adopted. Both the RMSE and MAD values revealed the prediction accuracy of the multilevel model to be superior to that of the generalized linear model. Also, the CURE plot implies that the goodness-of-fit of the multilevel model is relatively higher than that of the generalized linear model. The multilevel model proposed in this study can improve upon the limitations of the generalized linear model. Moreover, this model would make possible the accurate investigation of crash hot-spots, due to its higher crash prediction reliability. Consequently, this model can take a role as a quantitative basis for implementation of safety improvements on each type of trumpet interchange ramp.