Incorporating driver behavior in crash prediction modeling
Majka, Kevin M.
MetadataShow full item record
Motor vehicle crashes continue to be a top ten reason for all deaths in the US and ranks 3rd in terms of the years of life lost only behind cancer and heart disease. Intersection crashes in particular, accounted for over 9,600 fatalities and over 100,000 serious injuries in 2008. In these crashes, driver error is the number one causal factor yet most traditional crash prediction models only include coarse aggregate or surrogate measures of driver behavior. The main hypothesis of this dissertation is that the inclusion of driver behavior in motor vehicle crash prediction models will increase their accuracy and allow traffic safety engineers to make more informed decisions on how to increase safety at intersections. In order to test the hypothesis, this research focuses on identifying locations in Erie County, NY that have had multiple fatal crashes, modeling those intersections with the Highway Safety Manual (HSM) Crash Prediction Method (CPM), developing two local modification factors (LMF) based on driver behavior, one utilizing a negative binomial (NB) model and the second utilizing a neural network (NN). This study is unique in that it considers the actions or errors made by drivers rather than just the physical makeup of intersections. The base HSM CPM performed rather poorly for the 50 intersections modeled with a mean absolute deviation (MAD) of 23.78. The NB models built for each driver type show average performance in predicting crashes by type and when used to modify the HSM CPM has a MAD of 22.91. The NN also predicts crashes slightly better when used to modify the HSM CPM with a MAD of 21.76. In the context of the poorly performing HSM CPM model, this research shows that including driver behavior is an important aspect of crash prediction modeling which can lead to more accurate and better crash prediction models through the use of a NN behavioral model. The limitations encountered with the models can be largely attributed to the input data, the formulation of the HSM CPM and the development of the LMF. The number of crashes varies greatly form one geographic area to another, however the HSM CPM was validated and calibrated for only a few selected geographic regions, not including WNY. In addition, the exact factors that were found to be significant in developing the HSM CPM model do not adequately describe the crashes in this particular area and careful consideration is needed to identify new significant factors. Traditional transportation research has typically disregarded driver behavior or only provided it via aggregated surrogate measures due to its complexity and difficulty to model. Although the results of the models presented in this research performed below expectations, they do show that including driver behavior is an important aspect of crash prediction modeling and that they do have the potential to improve crash prediction models. This research provides only a demonstration of what types of behavioral information it may be possible to include in transportation research in order to better understand drivers actions and errors.