Associating street crime incidences with geographical environment in space using a zero-inflated negative binomial regression model

Urban geographical environments play an important role in inducing the occurrences of street crimes, so they have raised attention regarding the relationships between geographical environments and street crimes. Limited by the high data acquisition costs of traditional environmental auditing methods, current related studies still lack specific investigations on how street crime risks are quantitatively associated with diverse geographical environment variables, especially the multiple inherent streetscapes elements. Moreover, it is also necessary to discover the effects of pairwise variable interactions on street crimes by considering the existing zero inflation phenomenon. Therefore, this study introduces street view images to automatically characterize built urban environments with the help of deep learning models. By combining built environment variables with socioeconomic variables, we construct a zero-inflated negative binomial regression model to quantify both single-factor effects and the interaction effects among pairwise variables on street crime instances in space. By performing experiments on a real-life dataset, we find that the discovered interaction effects and seasonal variations in regression relationships yield the most comprehensive and reasonable explanations for the spatial distributions of street crime instances. These findings can guide the optimization of urban environmental element layouts and patrol policies to enhance the level of public security. • This study aims to quantitatively associate street crime incidences with diverse geographical environment variables. • A two-phase strategy is designed to reveal the interaction effects of geographical environment variable on street crimes. • The obtained interaction effects can provide scientific policy recommendations about place-based crime prevention.