Reconciling data-driven crime analysis with human-centered algorithms

This study combines traditional statistical methods with machine learning to better understand locally relevant, contextual models for analyzing crime in two urban American cities. Using census tracts as the units of analysis and controlling for several structural characteristics associated with crime, we find that in Milwaukee, Wisconsin, violent crime is associated with concentrated disadvantage, residential stability, ethnic heterogeneity, total population, and spatial lag of violent crime. Yet, the most important variable is the spatial lag of violent crime, followed by residential stability, ethnic heterogeneity, total population, and concentrated disadvantage. In addition, we find that in Chicago, Illinois, violent crime is associated with immigration, owner-occupied housing, proportion in professional occupations, and proportion population with college degree or higher, as well as ethnic heterogeneity, total population, and the spatial lag for violent crime. Machine learning models suggest that for Chicago's violent crime, the most important variable is the spatial lag term for violent crime, followed by total population, immigration, college education or beyond, owner occupancy, ethnic heterogeneity, and employment in professional occupations. The findings for property crime are similar: in Milwaukee, we find that disadvantage, residential stability, ethnic heterogeneity, total population and spatial lag for property crime are significant predictors in the traditional regression models. However, the most important variable for property crime in Milwaukee is the spatial lag term, followed by total population, ethnic heterogeneity, residential stability and disadvantage. The statistically significant predictors of property crime in Chicago include immigration, owner-occupied housing units, living in the same house, proportion of workforce in professional occupations, college education and beyond, total population, and the spatial lag for property crime. In Chicago, the most important variable for property crime is the spatial lag term, then the total population, the proportion of individuals in professional occupations, concentrated immigration, college education and beyond, living in the same house, and the proportion of owner-occupied housing units. Urban planners must consider policies that can effectively reduce nearby crime and violence in all cities that experience high crime levels, but also design locally responsive policies that make sense within a local context: in Milwaukee, residential stability matters more for violent crime than for property crime, while in Chicago, total population is similarly important for both violent crime and property crime. In Milwaukee, ethnic heterogeneity is similarly important for violent and property crime, while in Chicago, ethnic heterogeneity is not a very important variable for violent crime and it is not a significant predictor of property crime. Therefore, urban policy must differently approach social disorganization indicators and support the nuances of the local context for urban planning and policy.

Automated summary

This study combines traditional statistical methods with machine learning to analyze crime patterns in Milwaukee and Chicago. In Milwaukee, violent crime is associated with concentrated disadvantage, residential stability, ethnic heterogeneity, total population, and spatial lag of violent crime; while in Chicago, violent crime is linked to immigration, owner-occupied housing, professional occupations, college education, ethnic heterogeneity, total population, and spatial lag for violent crime.