Analysis of vehicle accident-injury severities: A comparison of segment- versus accident-based latent class ordered probit models with class-probability functions

Using information from 1990 single-vehicle accidents that occurred between 2011 and 2013 in the state of Washington, the injury severity level of the most severely injured vehicle occupant is studied using two latent class modeling approaches: segment-based and accident-based latent class ordered probit model with class-probability functions. The segment-based latent class ordered probit framework allows explanatory parameters to vary across unobserved groups (classes) of the highway segment population, while the modeling structure treats all segment-specific injury observations homogeneously (grouped). The accident-based latent class ordered probit framework allows for the explanatory parameters to vary across unobserved groups of the accident population, and the modeling structure treats all accident injury-severity observations individually (ungrouped). To further address heterogeneity arising from the probabilistic assignment of the highway segments or accident observations in the latent classes, the class probabilities are allowed to vary as a function of explanatory parameters, respectively. For both modeling approaches, the proposed latent class approach with class-probability functions is compared to its latent class counterpart with fixed class probabilities, and the results support the statistical superiority of the former, in terms of statistical fit and explanatory power. The empirical findings show the potential of both modeling approaches to unmask driver-, vehicle-, collision- and weather-specific sources of heterogeneity and, specifically, the capability of the segment-based approach to account for segment-specific heterogeneity. The comparative evaluation between the two modeling approaches shows that the segment-based approach provides better overall statistical fit. Furthermore, the forecasting accuracy of both approaches is explored through probability- and error-based measures demonstrating the forecasting accuracy benefits of the segment-based approach.