Evaluating alternate discrete outcome frameworks for modeling crash injury severity

Area of Science:

• Traffic Safety Research
• Statistical Modeling
• Transportation Engineering

Background:

• Driver injury severity in traffic crashes is a critical area of research.
• Accurate modeling frameworks are essential for understanding injury determinants and informing safety interventions.
• Existing models often face challenges in handling the complexities of injury severity data, including potential underreporting.

Purpose of the Study:

• To empirically compare the performance of various discrete outcome models for predicting driver injury severity.
• To evaluate ordered and unordered response frameworks, including specific models like generalized ordered logit (GOL) and mixed multinomial logit (MMNL).
• To investigate the impact of potential data underreporting on the performance of these modeling frameworks.

Main Methods:

• Utilized the 2010 General Estimates System (GES) data, a national sample of US road crashes.
• Compared ordered response models (ordered logit, generalized ordered logit, mixed generalized ordered logit) and unordered response models (multinomial logit, nested logit, ordered generalized extreme value logit, mixed multinomial logit).
• Assessed model performance using various metrics, including estimation, validation at aggregate and disaggregate levels, and analysis of underreporting effects with and without corrections.

Main Results:

• The generalized ordered logit (GOL) framework, particularly the mixed generalized ordered logit (MGOL) variant, demonstrated strong performance.
• GOL (MGOL) emerged as a competitive alternative to the mixed multinomial logit (MMNL) model in modeling driver injury severity.
• The study highlighted the importance of considering underreporting and its impact on model selection and accuracy.

Conclusions:

• The generalized ordered logit (GOL) framework offers a robust and competitive approach for modeling driver injury severity.
• Researchers should carefully consider the choice of discrete outcome model, especially when dealing with crash data that may be subject to underreporting.
• The findings provide valuable insights for improving the accuracy and reliability of driver injury severity analyses in traffic safety research.