Advanced Crash-Weighting Techniques for Forecasting the Future Economic Impact of Collisions

Abstract

This study evaluates the effectiveness of current methodologies used by departments of transportation (DOTs) for forecasting the future economic impact of roadway crashes and proposes new regression-based weighting models to improve predictive accuracy. Although the amount of previous equivalent property damage only (ePDO) is widely used to prioritize safety improvements, limited research validates its predictive adequacy. Utilizing over 285,000 crash records in over 125,000 road segments between 2013 and 2022, this research compares traditional forecasting methods, based on historical ePDO and past crash counts, with a novel approach that employs a custom weighting scheme derived from Lasso and Ridge regression models. Model performance was assessed using separate training and testing datasets, with cross-validation applied to enhance generalizability and prevent overfitting. Unlike traditional methods, the proposed models prioritize variables correlated with future crashes rather than relying solely on past economic costs. Findings indicate that the regression models, which address multicollinearity and emphasize severe crashes, reduce mean squared error by up to 24.8% compared with traditional approaches. Furthermore, these models perform better when ranking locations, which is crucial for project prioritization. While fatal crashes are significantly more costly, they are only marginally better predictors of future crash risk than injurious crashes, highlighting the need for a more nuanced weighting strategy. The proposed methodology enables DOTs to enhance safety prioritization programs without requiring additional data collection. It also provides a novel framework for future research on optimizing crash data weighting based on predictive risk rather than historical economic impact.