Exploring spatial heterogeneity in factors associated with injury severity in speeding-related crashes: An integrated machine learning and spatial modeling approach.

Speeding, a risky act of driving a vehicle at a speed exceeding the posted limit, has consistently emerged as a leading contributor to traffic fatalities. Identifying the risk factors associated with injury severity in speeding-related crashes is essential for implementing countermeasures aimed at preventing severe injury incidents and achieving Vision Zero goals. With the wealth of traffic crash data collected by various agencies, researchers have a valuable opportunity to conduct data-driven studies and employ various modeling methods to gain insights into the correlated factors affecting injury severity in traffic crashes. Machine learning models, owing to their superior predictive power compared to statistical models, are increasingly being adopted by researchers. These models, in conjunction with interpretation techniques, can reveal potential relationships between crash injury severity and contributing factors. Traffic crashes are inherently tied to geographic locations, distributed across road networks influenced by diverse socioeconomic and geographical factors. Recognizing spatial heterogeneity in traffic safety is crucial for tailored safety measures to address speeding-related crashes, as a one-size-fits-all approach may not work effectively everywhere. However, most existing machine learning models are unable to incorporate the spatial dependency among observations, such as traffic crashes, which hinders their ability to uncover spatial heterogeneity in traffic safety. To address this gap, this study introduces the Geographically Weighted Neural Network (GWNN) model, a spatial machine-learning model that integrates neural network (NN) and geographically weighted modeling approaches to investigate spatial heterogeneity in speeding-related crashes. Unlike the traditional NN model, which trains a single set of model parameters for all observations, the GWNN trains a local NN model for each crash location using a spatially weighted subsample of nearby crashes, allowing for the quantification of corresponding local effects of features through calculating local marginal effects. To understand the spatial heterogeneity in speeding-related crashes, this study extracted two years (2020 and 2021) of speeding-related crash data from Alabama for the development of the GWNN local models. The modeling results show significant spatial variability among several factors contributing to injury severity in speeding-related crashes. These factors include driver condition, vehicle type, crash type, speed limit, weather, crash time and location, roadway alignment, and traffic volume. Based on the GWNN modeling results, this study identified three types of spatial variations in relationships between contributing factors and crash injury severity: consistent positive associations, consistent negative associations, and inverse associations (i.e., marginal effects can vary between positive and negative depending on the location). This study contributes by integrating advanced machine learning and spatial modeling approaches to uncover intricate spatial patterns and factors influencing injury severity in speeding-related crashes, thereby facilitating the development of targeted policy implementations and safety interventions.