Injury Severity analysis based on police crash report

Objective. The current research aims to develop a statistical method to analyze police-reported Injury Severity data looking for the most relevant variable, and which condition are more prone to increase the severity level as a crash output. This study tries to contribute to the knowledge about crash data analysis in literature, thanks to the fact of the large amount of data available in the Maryland Crash Database (two years with more than 300K crashes per year). Methods. Since Injury Severity (IS) level is the dependent variable in the model and being this consisting of discrete ordered value, a Discrete Choice Modelling approach was considered as the most suitable option in this scenario. IS contains discrete and ordered hierarchically value, so the model selected for this study is an Ordered Logit model. Computations were carried out through the Python language using two libraries: Biogeme and Pandas (package PandasBiogeme). Results. Several independent variables play a role in the IS magnitude. Although the result was found to be consistent with literature. In few cases the sign contrasted with expectations and literature, but some considerations were carried out to explain such discrepancies. Conclusion. The calibrated model provided a reasonable statistical fit. The strong point of this study is the huge availability of data that allow a solid statistical consistency. This could be a starting point to many different insights studies regarding the variables here analyzed, in order to limit the loss of life on roads and improve road safety.