The performance of accident severity multiclass classification

Abstract

One way to monitor accidents on highway is to analyze the accident characteristic to predict the accident severity. This study applied multinomial logistic regression model to predict accident severity. Predicted accident severities are compared with actual accident severities to evaluate the prediction performances of the model. The aim of this study is to determine the performance of accident severity classifications by multinomial logistic regression model. The predicted accident severities could be used to estimate potential effect of changes in factors contributing to accidents. Data was obtained from the Malaysian Highway Authority for the year 2013 and 2014. The accident severity was grouped into four categories of death, serious injury, minor injury and damage. Based on the results, the model correctly classified accident severities by 63.52% using training data and 61.45% using validation data. The Hosmer- Lemeshow test indicated the model has a good fit between the actual accident severities and predicted accident severities and the ROC results indicted the model able to distinguish between the classifications. The classifier of the model inclined more toward the damages compared to other accident severities resulted in classifying accident severity classes with more samples better and remains weak on the accident severity classes with lesser samples.