Exploring Spatial Variations in Factors Influencing Cyclist Injury Severity in Traffic Crashes: A Comparison of Geographically Weighted Regression and Spatial Machine Learning
TL;DR
Results show that older cyclist age, higher vehicle speeds, intersection locations, multi-lane roadways, and dark conditions are associated with higher probabilities of evident or more severe injuries, highlighting the predictive capability of spatial machine learning while retaining interpretable geographic variation, supporting targeted bicycle safety interventions.
Abstract
Cyclists are vulnerable road users, and factors associated with cyclist injury severity may vary across geographic contexts. This study compares a spatial statistical model, Geographically Weighted Logistic Regression (GWLR), with a spatial machine learning model, Geographically Weighted Neural Network (GWNN). Eight years of North Carolina motor vehicle–bicycle crash data (2015–2022), comprising 6373 crashes, were analyzed, with marginal effects used to compare spatially varying relationships between the two models. Results show that older cyclist age, higher vehicle speeds, intersection locations, multi-lane roadways, and dark conditions are associated with higher probabilities of evident or more severe injuries. Cyclists over 65 years showed mean marginal effects of 15.0% in GWLR and 5.7% in GWNN, indicating a strong association with injury severity. Vehicle speed, intersection location, and unlit dark conditions exhibit strong spatial heterogeneity, with elevated effects concentrated mainly in western and central North Carolina. GWNN achieved 73.7% accuracy on the test set, while GWLR achieved 64.7% accuracy on the full estimation sample. Despite different modeling frameworks, both models identified similar spatial patterns of local effects for most risk factors. These findings highlight the predictive capability of spatial machine learning while retaining interpretable geographic variation, supporting targeted bicycle safety interventions.