Count Data Modeling for Predicting Crash Severity on Indian Highways

Abstract

This study collected data on road accidents for the years 2016-2020 for the NH-48 highway in Maharashtra, India to model their conditions. Road crash data models were developed using 70% of actual data for training and 30% for testing purposes. Negative binomial regression modeling was used to predict crash fatalities. The results showed that the factors that affected the fatality of road crashes were head-on-collision, friction, time zone, and weather conditions of the crash. The developed models were validated and tested using log-likelihood, AIC, BIC, MAD, MSE, RMSE, and MAPE values. Head-on-collision, AM, PM, light rain, mist/fog, heavy rain, fine, and cloudy were positively associated with the fatality of road crashes, while friction was negatively associated. The developed models can be used to predict the fatality/non-fatality of road crashes and implement road safety strategies on highways to reduce them.