Applying the heteroskedastic ordered probit model on injury severity for improved age and gender estimation

ObjectiveDriver characteristics have been linked to the frequency and severity of car crashes. Among these, age and gender have been shown to impact both the possibility and severity of a crash. Previous studies have used standard ordered probit (OP) models to analyze crash data, and some research has suggested heteroskedastic ordered probit (HETOP) could provide improved model fit. The objective of this paper is to evaluate potential improvements of the heteroskedastic ordered probit (HETOP) model compared to the standard ordered probit (OP) model in crash analysis, by examining the effect of gender across age on injury severity among drivers. This paper hypothesizes that the HETOP model can provide a better fit to crash data, by allowing heteroskedasticity in the distribution of injury severity across driver age and gender.MethodsData for 20,222 crashes were analyzed for North Carolina from 2016 to 2018, which represents the state with the highest number of fatalities per 100 million vehicle miles traveled amongst available crash data from the Highway Safety Information System.ResultsDarker lighting conditions, severe road surface conditions, and less severe weather were associated with increased injury severity. For driver demographics, the probability of severe injuries increased with age and for male drivers. Moreover, the variance of severity increased with age disproportionately within and across genders, and the HETOP was able to account for this.ConclusionsThe results of the two applied approaches revealed that HETOP model outperformed the standard OP model when measuring the effects of age and gender together in injury severity analysis, due to the heteroskedasticity in injury severity within gender and age. The HETOP statistical method presented in this paper can be more broadly applied across other contexts and combinations of independent variables for improved model prediction and accuracy of causal variables in traffic safety.