Supporting equitable and responsible highway safety improvement funding allocation strategies - Why AI prediction biases matter

The existing methodologies for allocating highway safety improvement funding closely rely on the utilization of crash prediction models. Specifically, these models produce predictions that estimate future crash hazard levels in different geographical areas, which subsequently support the future funding allocation strategies. In recent years, there is a burgeoning interest in applying artificial intelligence (AI)-based models to perform crash prediction tasks. Despite the remarkable accuracy of these AI-based crash prediction models, they have been observed to yield biased prediction outcomes across areas of different socioeconomic statuses. These biases are primarily attributed to the inherent measurement and representation biases of AI-based prediction models. More precisely, measurement bias arises from the selection of target variables to reflect crash hazard levels, while representation bias results from the issue of imbalanced number of samples representing areas with different socioeconomic statuses within the dataset. Consequently, these biased prediction outcomes have the potential to perpetuate an unfair allocation of funding resources, contributing to worsen social inequality over time. Drawing upon a real-world case study in North Carolina, this study designs an AI-based crash prediction model that utilizes previous sociodemographic and crash-related variables to predict future severe crash rate of each area to reflect the crash hazardous level. By incorporating a fair regression framework, this study endeavors to transform the crash prediction model to become both fair and accurate, aiming to support equitable and responsible safety improvement funding allocation strategies.