Modeling and Predicting Geospatial Teen Crash Frequency

This research project 1) evaluates the effect of road network, demographic, and land use characteristics on road crashes involving teen drivers, and, 2) develops and compares the predictability of local and global regression models in estimating teen crash frequency. The team considered data for 201 spatially distributed road segments in Mecklenburg County, North Carolina, USA for the evaluation and obtained data related to teen crashes from the Highway Safety Information System (HSIS) database. The team extracted demographic and land use characteristics using two different buffer widths (0.25 miles and 0.5 miles) at each selected road segment, with the number of crashes on each road segment used as the dependent variable. The generalized linear models with negative binomial distribution (GLM-based NB model) as well as the geographically weighted negative binomial regression (GWNBR) and geographically weighted negative binomial regression model with global dispersion (GWNBRg) were developed and compared. This research relied on data for 147 geographically distributed road segments for modeling and data for 49 segments for validation. The annual average daily traffic (AADT), light commercial land use, light industrial land use, number of household units, and number of pupils enrolled in public or private high schools are significant explanatory variables influencing the teen crash frequency. Both methods have good predictive capabilities and can be used to estimate the teen crash frequency. However, the GWNBR and GWNBRg better capture the spatial dependency and spatial heterogeneity among road teen crashes and the associated risk factors.