An empirical assessment of fixed and random parameter logit models using crash- and non-crash-specific injury data

• Published in: Accident analysis and prevention 2011-05, Vol.43 (3), p.1140-1147
• Main Authors: Anastasopoulos, Panagiotis Ch, Mannering, Fred L.
• Format: Article
• Language: English
• Subjects: Accidents, Traffic - classification; Accidents, Traffic - mortality; Accidents, Traffic - statistics & numerical data; Assessments; Crash-injury severities; Crashes; Data Collection - statistics & numerical data; Environment Design; Forensic engineering; Highways; Humans; Indiana; Injuries; Logistic Models; Mathematical models; Military technology; Mixed logit model; Outcome Assessment (Health Care) - statistics & numerical data; Pavement condition; Random parameters logit model; Reproducibility of Results; Roadway geometrics; Safety - statistics & numerical data; Survival Analysis; Vehicles; Wounds and Injuries - epidemiology; Wounds and Injuries - mortality
• ISSN: 0001-4575; 1879-2057
• DOI: 10.1016/j.aap.2010.12.024

▶ We compare fixed & random parameter models using crash- & non-crash-specific injury data. ▶ The random-parameter logit is statistically superior to the fixed-parameter model. ▶ The individual crash-data models provide better fit compared to the proportions models. ▶ Using proportions data may be practically very close to using individual-crash data. ▶ Both approaches identify nearly the same segments with the top 5% fatality proportions. Traditional crash-severity modeling uses detailed data gathered after a crash has occurred (number of vehicles involved, age of occupants, weather conditions at the time of the crash, types of vehicles involved, crash type, occupant restraint use, airbag deployment, etc.) to predict the level of occupant injury. However, for prediction purposes, the use of such detailed data makes assessing the impact of alternate safety countermeasures exceedingly difficult due to the large number of variables that need to be known. Using 5-year data from interstate highways in Indiana, this study explores fixed and random parameter statistical models using detailed crash-specific data and data that include the injury outcome of the crash but not other detailed crash-specific data (only more general data are used such as roadway geometrics, pavement condition and general weather and traffic characteristics). The analysis shows that, while models that do not use detailed crash-specific data do not perform as well as those that do, random parameter models using less detailed data still can provide a reasonable level of accuracy.