Evaluation of Peak Over Threshold Approach for Road Safety Estimation

Abstract

Development of non-crash-based safety estimation approaches has drawn considerable research interest, and the application of extreme value theory is one of them. This study proposes the use of extreme value theory based peak over threshold (POT) approach for road safety estimation and compares the POT with the crash-based negative binomial regression approach. The traffic data is from videos recorded at 29 directional segments with time duration of approximately 3 h each. The results show that the best fitted model comes from regression model with R2 of 0.65, in which the covariate is return levels estimated from POT approach. Meanwhile, the POT estimated crashes also have relatively high R2 value but with large variances. Both of these imply that POT estimations from short observation time carry useful safety information. Moreover, the POT estimated return levels can be employed as a covariate for building better regression models.