THE ESTIMATION, ENSEMBLING AND COMPARISON OF METHODOLOGICAL TECHNIQUES USED IN MOTOR VEHICLE CRASH SEVERITY RESEARCH

Jill M. Bernard Bracy
Fontbonne University

ABSTRACT

This study compares the performance of longstanding methodological techniques of
multinomial logit and ordinal probit models with more recent methods of decision tree and
artificial neural network models, and combines individual models into ensembles to test if the
amalgamation of the multiple methodologies enhances the classification accuracy of crash injury
severity outcomes. The models are estimated using 2002 to 2012 crash data from the Missouri
State Highway Patrol, and the variables examined include driver characteristics, temporal
factors, weather conditions, road characteristics, and injury severity levels. The accuracy and
discriminatory power of explaining crash severity outcomes among all methods are compared
using classification tables and Area Under the Receiver Operating Characteristic curve values.
The Chi-square Automatic Interaction Detection decision tree model is found to have the greatest
accuracy and discriminatory power relative to all evaluated approaches.

Keywords: Crash severity research, model ensemble, road safety, motor vehicle crash