Book Description
In 2015, 35,092 people died in motor vehicle crashes in the United States. This was a 7.2% increase over 2014 and represented the first year-to-year fatality increase after several years of decline. In response to this, the National Highway Traffic Safety Administration (NHTSA) published their Fatality Analysis Reporting System (FARS) motor vehicle fatal crash census data three months early, and tasked students and researchers to dig into the data. The purpose of this thesis is to investigate (1) which FARS variables are associated with higher fatality risk for the 80,587 individuals who were involved in the 2015 fatal crashes, and (2) which ones are associated with those fatal crashes that involved more than one fatality (92.69% of the 32,166 fatal accidents just resulted in one death, but the other 7.31% had two or more fatalities, so this later group of accidents is particularly interesting). The individuals will be split up into four groups: drivers of motor vehicles, other occupants of motor vehicles, pedestrians, and pedalcyclists (individuals riding bicycles at the time of an accident). Analysis methods will involve logistic regressions for each of the four person groups, with survivor/fatality status as the response, and multinomial logistic regression and Poisson regression for the accidents, with fatality count as the response. Decision tree models will also be fit in tandem and will represent an alternative modeling strategy. Models will be compared with one another in terms of number of variables used, shared variables, as well as various measures of predictive ability.We find that factors such as alcohol use, number of people involved in the crash, and certain crash-specific details (some first harmful events of the crash and manners (type) of collision in the crash) are associated with increased fatality risk for individuals and increased fatality counts in accidents. We review the main findings that held across most of the models, make policy recommendations for NHTSA and the US DOT, and discuss future work that could make analysis of FARS data more sophisticated and complete.