GeoShapley-based interpretation of older adult pedestrian fatal vs injury crash frequency in dense urban environments

As the world's population ages, ensuring the safety of older adult pedestrians has become an urgent priority in transportation planning. However, most existing studies rely on global models that overlook spatial heterogeneity and fail to capture nonlinear, location-specific interactions between the environmental factors and crash outcomes. Moreover, subjective perceptions (e.g., how safe or walkable an area feels) may influence pedestrian behavior and crash exposure but are underexplored in traffic safety research. This study addresses these gaps by integrating subjective perception indicators extracted from Street View Images (SVI) with machine learning models to examine the severity of older adult pedestrian crashes at intersections in Taipei City. Three modeling frameworks are evaluated and compared: global Negative Binomial Regression (NBR), Geographically Weighted Negative Binomial Regression (GWNBR), and GeoShapley, a spatially interpretable extension of the SHAP framework for XGBoost. A total of 36 environmental and perceptual variables are evaluated in relation to injury and fatal crash frequencies. Among these models, GeoShapley achieved the best performance and revealed that spatial location (GEO) and its interactions with environmental factors and subjective perceptions were among the most influential predictors. In some areas, higher walkability was associated with reduced injury crash frequencies, especially in older and urban districts. In addition, the effect of convenience stores and nursing homes on the frequency of fatal crashes varied significantly across locations, reflecting the spatial clustering of pedestrian activity and older adults. Overall, the findings demonstrate the value of spatially explicit machine learning tools and subjective perceptions in understanding localized crash dynamics in aging urban populations.