Modeling Multilevel Data In Traffic Safety PDF Download

The socioeconomic burden of road traffic crashes is immense. Safer roads and vehicular mechanisms to reduce distracted driving help reduce collisions. Additionally, computational models can be used to understand the reasons for crashes and devise interventions. We study models predicting the severity of a crash based on the data reported at the crash scene. Many U.S. states have developed traffic safety programs to make the anonymized crash data publicly available. These datasets aid researchers in the creation of predictive models for crashes. While many states make data from collisions publicly available, each state reports data differently. There is a lack of standardization. As a result, it is difficult for researchers to develop machine learning algorithms to process data from multiple states without adequate preprocessing. Currently, the vast majority of projects in this field of study utilize a dataset of a single city, road, or state. This limits the use of the developed model to a region. This project aims to create a large crash database that will allow researchers to develop algorithms that utilize data from across the country. Additionally, we want to examine if the use of data from multiple states is effective in increasing the accuracy of machine learning models. In order to achieve these goals, we develop software to find common data categories from state reports and combine them into one large dataset. The data categories were selected based on reports from previous projects that identified variables having a large impact on model accuracy. In order to test the effectiveness of the new multi-state dataset, we used two models (neural network-based and decision tree-based) to predict crash injury severity. We trained and tested these models on datasets from a single state, combined two-state datasets, and a combined multi-state dataset. The results of this research reveal that there is a drop in accuracy when data from multiple states are combined. This trend is present in both the models tested, with the trend being more pronounced in the decision tree. There are some cases in the neural network model where multi-state data lead to a higher accuracy compared to the single-state experiments. We also observe a downward trend between neural network accuracy and the distance between the states present in the dataset. This implies that the closer the states are together geographically, the better the accuracy will be using the neural network model. In the decision tree model, there is a positive correlation between overall accuracy and the number of features present in the dataset. This observation means that the more features states have in common, the better the accuracy will be for a decision tree classifier. The software artifacts from this project are open-sourced.