A hybrid statistical-machine learning methodology for addressing endogeneity and temporal instability in...

Area of Science:

• Road Safety Engineering
• Traffic Behavior Analysis
• Statistical Modeling

Background:

• Speeding is a major contributor to road crashes, particularly on horizontal curves.
• Accurately estimating speeding's impact is challenging due to data errors and complex interrelationships between driver behavior, road geometry, and crash risk.

Purpose of the Study:

• To develop a new methodology combining improved data collection and a hybrid statistical-machine learning model.
• To accurately identify speeding and estimate its effect on crash frequency on horizontal curves.

Main Methods:

• A hybrid model integrating negative binomial regression with gradient boosting and Shapley values was developed.
• The model incorporated random parameters and mixed spline indicators to address unobserved heterogeneity and temporal instability.
• The methodology was tested on 179 km of horizontal curves on rural roads in Iran.

Main Results:

• The machine learning model demonstrated high predictive power for speeding using exogenous variables.
• Shapley values and feature importance provided intuitive insights into variable contributions.
• The proposed model significantly outperformed existing state-of-the-art techniques in statistical fit.
• Curve geometry and traffic characteristics were identified as strong predictors of speeding.

Conclusions:

• Driving over 20% above the speed limit substantially increases crash frequency.
• The effects of passenger and heavy vehicle traffic on crashes exhibit temporal variations.
• The developed hybrid model offers a superior approach for analyzing speeding's impact on road safety.