Conditional Quantile Analysis for Crash Count Data

Abstract

Crashes are important evidence for identifying deficiencies existing in highway systems, but they are random and rare. The investigation of the nature of the problem normally draws on crashes collected over a multiyear period and from different locations to obtain a sizable sample. Hence, the issue of data heterogeneity arises because the pooled data originated from different sources. Data heterogeneity has to be addressed to obtain stable and meaningful estimates for variable coefficients. A desirable method of handling heterogeneous data is quantile regression (QR) because it focuses on depicting the relationship between a family of conditional quantiles of the crash distribution and the covariates. The QR method is appealing because it offers a complete view of how the covariates affect the response variable from the full range of the distribution, which is of particular use for distributions without symmetric or normal forms (i.e., heavy tails, heteroscedasticity, multimodality, etc.). Crash data possess some of the properties that quantile analysis can handle, as demonstrated in an intersection crash study. The compelling results illustrate that conditional quantile estimates are more informative than conditional means. The findings provide information relative to the effect of traffic volume, intersection layout, and traffic control on crash occurrence.