A resampling approach to disaggregate analysis of bus-involved crashes using panel data with excessive zeros

Area of Science:

• Transportation Engineering
• Traffic Safety
• Data Science

Background:

• Public buses are Hong Kong's primary public transport, yet exhibit the highest crash rates.
• Existing bus crash risk analyses suffer from inaccurate parameter estimation due to imbalanced crash data and excessive zeros.

Purpose of the Study:

• To address the excess zero problem in disaggregate bus-involved crash analysis.
• To develop and validate a synthetic data generation technique for imbalanced panel data.

Main Methods:

• Utilized Synthetic Minority Over-Sampling Technique for panel data (SMOTE-P) to create balanced synthetic crash data.
• Compared SMOTE-P with Random Under-sampling of the Majority Class (RUMC), Cluster-Based Under-Sampling (CBUS), and mixed resampling techniques.
• Estimated Random Effect Poisson (REP) models using synthetic data and a Random Effect Zero-Inflated Poisson (REZIP) model using original data.

Main Results:

• The REP model with SMOTE-P synthetic data demonstrated superior statistical fit and prediction accuracy compared to other methods.
• Identified significant factors influencing bus crash frequency: peak hours, traffic flow, road characteristics, bus density, and bus priority lanes.
• SMOTE-P provided a more robust analysis of bus-involved crash frequency.

Conclusions:

• Synthetic data generation, particularly using SMOTE-P, offers a viable solution for analyzing imbalanced crash data in public transport safety studies.
• The identified factors provide actionable insights for implementing targeted safety interventions to reduce bus crashes.