Finite mixture modeling for vehicle crash data with application to hotspot identification

Area of Science:

• Highway Safety Research
• Statistical Modeling

Background:

• Unobserved heterogeneity is a challenge in crash data analysis.
• Finite mixture models address heterogeneity by assuming data arises from unobserved components.
• Both fixed and varying weight parameter models have shown utility in explaining crash data dispersion.

Purpose of the Study:

• To compare the performance of finite mixture models and negative binomial (NB) models for identifying highway crash hotspots.
• To evaluate model reliability using observed and simulated data.
• To investigate optimal threshold values for hotspot identification.

Main Methods:

• Utilized rural multilane segment crash data from California and Texas for observed data analysis.
• Employed simulation studies to assess model performance under different conditions.
• Analyzed ranking order deviations and false positive/negative rates.

Main Results:

• Observed data showed small differences in ranking orders between finite mixture and NB models.
• Finite mixture models provided more reliable rankings due to superior model specification.
• Simulation revealed significant false positives and negatives with mis-specified models.
• A trade-off exists between false discovery and false negative rates for hotspot identification thresholds.

Conclusions:

• Finite mixture models are more reliable for highway crash hotspot identification than traditional NB models.
• Model specification is crucial for accurate hotspot identification, as mis-specification leads to errors.
• Optimal hotspot identification thresholds require balancing false discovery and false negative rates based on cost considerations.