Predicting severe wildlife vehicle crashes (WVCs) on New Hampshire roads using a hybrid generalized additive model

Abstract

Across the United States, wildlife vehicle crashes (WVCs) are increasing and remain consistently deadly to drivers, despite a downward trend in fatal automobile accidents overall. That said, the factors related to severe WVCs are unclear. With this in mind, we pursued a statistical model to reveal factors associated with WVCs that result in severe injury or death to drivers. We hypothesize that there are statistically significant interactions and non-linear relationships between these factors and severity occurrence. We developed a generalized additive model (GAM) with linear terms, additive terms, and a binary response for severity. We surmise that our fitted model results will quantify the relationship between significant variables and severity occurrence, and ultimately help to develop countermeasures to mitigate serious injury. The model was fitted to WVC records occurring between 2002 and 2019 in the state of New Hampshire. Fitted linear terms revealed: 1) in inclement weather, there is about a 22% increase in the odds of severity for slick surface conditions compared to dry surface conditions; 2) for the warmer months (spring/summer), there is a 42% decrease in the odds of severity for straight roads compared to those with curvature/incline; 3) for highways, the odds of severity decreases by 48% for accidents occurring on NH’s two major intestates highways, and 4) for spring/summer (as compared to the fall/winter), there is more than a 3-fold increase in the odds of severity for two-way traffic. Fitted additive terms revealed: 1) the odds of severity increased in the early hours, between midnight and 6AM, and after 5PM; 2) speeds between 45 and 60 mph are associated with an increase in the odds of a severe accident, while both lower and higher speeds (those below 45 and above 60 mph) are associated with a decrease in the odds of a severe accident; and 3) low, mid-range, and high human population densities are associated with decreases, increases, and decreases in odds of severity, respectively. Cross validation and resulting ROC curves gave evidence that our model is well specified and an effective predictor. Results could be used to inform drivers of potentially dangerous roadways/conditions/times.

Keywords

• Generalized additive model
• logistic regression
• wildlife vehicle crash
• accident severity
• interaction