Development of crash prediction models by assessing the role of perpetrators and victims: a comparison of ANN & logistic model using historical crash data

Road traffic injuries cost countries 3% of their annual GDP. In developing countries like India, every year around 150,000 people die on roads. The type of vehicles involved in a crash contribute majorly to the outcome of casualty (injury/death). Barring few studies, literature are less regarding the role of vehicle as perpetrator and victim on road crash fatalities. Historical crash data has been used in the present study to examine the role of vehicles (both as perpetrator & victim). The study reveals that victim's effect is more as compared to perpetrator/accused for determining the outcome of crash. Heavy vehicles as perpetrator, and self-hitting vehicles along with pedestrians as victims have higher fatality rates. Binary logistic regression and artificial neural network (ANN) has been utilized for developing prediction models. Binary logistic model predicted around 75% of outcomes correctly with default cut-off value (0.5). However, based on reported crash data, where 19% of total crashes lead to deaths, 0.19 has been proposed as cut-off value which increases the accuracy of the predictions. Accuracy of ANN technique directly depends on the number of crashes reported for a definite pair of perpetrator and victim and the type of validation technique used (Holdback/K-Fold) along with the type of hidden layer chosen for the study based on different types of sigmoid activation function. ROC curves in ANN suggest that the analysis can predict 75% of the outcomes which can be increased by deleting the pairs of vehicles which are present/have occurred in very less number. A comparison has been made between the two techniques based on their advantages and limitations. The developed models can be used as safety indicators based on composition of traffic flow on urban roads.