Predicting less, understanding more: shifting the use of machine learning from individual prediction to structural insights in systems that affect children and families

Machine learning algorithms have been deeply embedded across domains, valued for their capacity to analyze large-scale data and to support a range of descriptive and predictive tasks. Despite their versatility, most applications to date have focused on individual-level prediction at the expense of broader structural insights. This paper shifts that focus by using comprehensive demographic, juvenile justice, and other key data from child-serving agencies in one southern state to examine how juvenile and family court intake structures shape case outcomes in delinquency referrals. Specifically, we combine machine learning algorithms with a doubly robust, potential outcomes-based modeling procedure to estimate the causal effect of prosecutor-controlled intake structures on youth dispositional outcomes. In doing so, we offer a novel demonstration of how these tools can support causal inquiry into the functioning of complex legal systems. Results indicate that intake structure has a material effect on youth outcomes, such that when prosecutors act without input from the juvenile justice agency, youth referrals are significantly more likely to be prosecuted rather than diverted or dismissed. Differences in prosecution rates between intake structures are most pronounced for Black girls and least pronounced for White boys, highlighting how system design choices may contribute to disparities in juvenile justice outcomes. Taken together, findings underscore the potential of using machine learning algorithms to generate insight into the policies and practices that shape them—opening new directions for structural analysis and reform in systems that affect children and families.