Development of Safety Performance Functions: Incorporating Unobserved Heterogeneity and Functional Form Analysis

TL;DR

This study enhances transportation safety models by incorporating unobserved heterogeneity and testing various functional forms, leading to more accurate crash prediction models for Tennessee roadways.

Contribution

It introduces Tennessee-specific random parameter Poisson Safety Performance Functions that outperform existing models in crash prediction accuracy.

Findings

Statewide calibration factor is 2.48, indicating crashes are underestimated by HSM predictions.

Regional calibration factors range from 2.02 to 2.77, showing significant heterogeneity.

Random parameter SPF with specific variables best predicts crash frequency on Tennessee roads.

Abstract

To improve transportation safety, this study applies Highway Safety Manual (HSM) procedures to roadways while accounting for unobserved heterogeneity and exploring alternative functional forms for Safety Performance Functions (SPFs). Specifically, several functional forms are considered in Poisson and Poisson-gamma modeling frameworks. Using five years (2011-2015) of crash, traffic, and road inventory data for two-way, two-lane roads in Tennessee, fixed- and random-parameter count data models are calibrated. The models account for important methodological concerns of unobserved heterogeneity and omitted variable bias. With a validation dataset, the calibrated and uncalibrated HSM SPFs and eight new Tennessee-specific SPFs are compared for prediction accuracy. The results show that the statewide calibration factor is 2.48, suggesting rural two-lane, two-way road segment crashes are at least 1.48 times greater than what HSM SPF predicts. Significant variation in four different regions in Tennessee is observed with calibration factors ranging between 2.02 and 2.77. Among all the SPFs considered, fully specified Tennessee-specific random parameter Poisson SPF outperformed all competing SPFs in predicting out-of-sample crashes on these road segments. The best-fit random parameter SPF specification for crash frequency includes the following variables: annual average daily traffic, segment length, shoulder width, lane width, speed limit, and the presence of passing lanes. Significant heterogeneity is observed in the effects of traffic exposure-related variables on crash frequency. The study shows how heterogeneity-based models can be specified and used by practitioners for obtaining accurate crash predictions.