Modified Mixed Generalized Ordered Response Model to Handle Misclassification in Injury Severity Data

Abstract

Traditional crash databases that record police-reported injury severity data are prone to misclassification errors. Ignoring these errors in discrete ordered response models used for analyzing injury severity can lead to biased and inconsistent parameter estimates. In this study, a mixed generalized ordered response (MGOR) model that quantifies misclassification rates in the injury severity variable and adjusts the bias in parameter estimates associated with misclassification was developed. The proposed model does this by considering the observed injury severity outcome as a realization from a discrete random variable that depends on true latent injury severity that is unobservable to the analyst. The model was used to analyze misclassification rates in police-reported injury severity in the 2014 General Estimates System (GES) data. The model found that only 68.23% and 62.75% of possible and non-incapacitating injuries were correctly recorded in the GES data. Moreover, comparative analysis with the MGOR model that ignores misclassification not only has lower data fit but also considerable bias in both the parameter and elasticity estimates. The model developed in this study can be used to analyze misclassification errors in ordinal response variables in other empirical contexts.