Application of GLASSO in Variable Selection and Crash Prediction at Unsignalized Intersections

Abstract

In this paper, a new promising variable screening technique is proposed to select important covariates and to improve crash prediction; the group least absolute shrinkage and selection operator (GLASSO). The GLASSO’s main power lies in its ability to deal with data sets havinga large number of categorical variables, the case in this study. Identifying the significant factors affecting the safety of unsignalized intersections was also an essential objective. Two applications of GLASSO were investigated: before fitting the negative binomial (NB) model, and before fitting the promising multivariate adaptive regression splines (MARS) technique using extensive data representing 2,475 unsignalized intersections. Regarding the NB models, GLASSO yielded close prediction capability to the backward deletion and random forest techniques. Also, MARS model fitting after using GLASSO relatively outperformed that after using random forest, with similar prediction performance. Because of its outstanding performance with categorical variables and its simplicity, GLASSO is recommended as a promising variable selection technique. Some significant predictors affecting unsignalized intersections’ safety were traffic volume on the major road, upstream and downstream distances to the nearest signalized intersection, and median type on major and minor approaches.