Mixed linear modeling techniques for predicting fatalities in vehicle crashes

Motor vehicle crashes cause loss of life, property, and finances. It remains a focus of traffic safety research, uncovering useful information that can be directly applied to reduce these losses. Traditionally, modeling crash fatalities has been done using machine learning techniques, considering crash level variables such as roadway characteristics, lighting conditions, weather conditions, and the prevalence of drugs or alcohol. This research moves in a new direction by incorporating heterogeneous data sources, including both police crash report data and demographic data, potentially leading to a heightened understanding of the relationship between crash location and the likelihood of a fatality. In order to utilize multiple data sources, we propose using a mixed linear modeling technique that enables data fusion in a principled way to build a better predictive model. We integrate these data sources to analyze the natural clustering of events in space by different geographic levels. Our findings indicate that using mixed effect logistic regression, incorporating both fixed and random effects, outperforms a traditional classification model in its ability to predict crash fatalities across the state of Virginia, United States of America Additionally, results from model comparison over 20 trials, using a paired-t test, indicate that adding the jurisdiction as a random effect improves the predictive performance.