Revisiting Hit-and-Run Crashes: A Geo-Spatial Modeling Method

Abstract

Hit-and-run crashes often delay emergency response and may result in increasing/secondary harms/damages to the victims in the crash. This study revisited hit-and-run crashes using a geo-spatial modeling approach, specifically, Geographically Weighted Regression (GWR), to explore geo-referenced crash data. The data cover motor vehicle crashes ( N = 138,529) in Southeast Michigan including 20,813 hit-and-run crashes in 2015. This study presented the results from both traditional regression and GWR models. GWR model results can be mapped in space, and the maps offer visual insights about the spatially varying correlates of hit-and-run crashes that are not available from previous studies. Results from traditional binary logit model are generally consistent with findings in previous studies. For example, hit-and-run is more likely to occur on weekends or during nighttime (especially without street lights on). Driving under impairment (DUI) seems to increase the likelihood of hit-and-run. GWR models also uncovered spatially varying correlates of hit-and-run. For example, DUI crashes in the northwest of the Detroit metropolitan area are associated with an even greater hit-and-run likelihood than those in other parts in this area. In addition, the local socio-economic factors are included in the analysis. Results show that hit-and-run is more likely to occur in census tracts with a higher unemployment rate, a lower household income, a smaller portion of college-educated population, and a greater population density. The study demonstrates a way of making sense of geo-referenced traffic safety data. The geo-spatial modeling method is useful for prioritizing specific geographic regions/corridors for safety improvement countermeasures, and outperforms traditional modeling techniques.

Identifying the Local Impacts of Speed-Related Factors on Tunnel Entrance Crash Severity

“…In the conventional LR model, all observations share the same set of coefficients, and the relationship between the response and explanatory variables is globally stationary; that is, the conventional LR model works well when the data do not have a spatial nature ( 10 ). However, a primary task for the study objective is to prioritize freeway tunnel access zones for VSL implementation spatially by considering variables that contain spatial characteristics.…”

“…The bi-square kernel has a clear-cut range, whereas kernel weighting is non-zero, making it suitable for clarifying local extents for model fitting ( 29 ). Additionally, adaptive bandwidth is often employed when the response outcome is not uniformly distributed in space ( 10 , 29 ). In this study, therefore, the bi-square adaptive kernel function was used to determine weights for GWLR modeling, which is specified as follows:…”

“…To examine the extent of data explained by the model, the percent of data explained by deviance, termed pseudo R 2 for LR ( 10 ), was adopted in the current study, which is associated with the changes in deviance from a null model to the regressed model (1− D regressed model / D null model ).…”

“…By comparing the GWR model with a traditional regression model, Liu et al ( 10 , 11 ) found that the GWR model can offer visual insights about the spatially varying correlates of hit-and-run crashes or bicyclist injury severity. The GWR model outperforms traditional modeling techniques, and its results are useful in prioritizing specific geographic regions or corridors for safety improvement countermeasures.…”

Promoting Emergency Medical Service Infrastructure Equality to Reduce Road Crash Fatalities

“…Geographically weighted binary logit regression (GWBLR) was used in this study for the following reasons: the response variable (fatality vs. survival) is dichotomous; GWBLR addresses spatial nonstationary issues, allowing for local coefficients of explanatory variables that vary in space [45]; and diverse temporal factors were not included in the RF result. The GWBLR model extends the concept of the traditional binary logit regression (BLR) model to a locally estimated model, which is written by: P(Y) = e α(ui,vi)+β(ui,,vi)X 1 + e α(ui,vi)+β(ui,vi)X…”