Predictive analytics encompasses statistical methods that detect underlying patterns in large-scale data to predict a future outcome (Redden, 2020). In the past decade, predictive analytics in the U.S. child welfare system has focused on the frontend, child protection system. Predictive risk modeling has emerged as the more common and contentious application of predictive analytics, where “risk scores” are generated to denote children’s risks for a serious injury or fatality from child maltreatment (Church & Fairchild, 2017; Jackson & Marx, 2017) or foster care placement following a child protection investigation (Chouldechova et al., 2018). When these risk scores are used to inform or justify high-stakes decisions, from initiating an investigation to removing a child from their home, predictive analytics has received much warranted scrutiny, let alone growing concerns about prediction accuracy and systemic bias (Keddell, 2019). Coupled with failed high-profile implementation (Jackson & Marx, 2017) in the backdrop of the social justice movement (Dettlaff & Copeland, 2023) and the abolitionist vs. reformist debate about the child protection system (Dettlaff et al., 2020; Garcia et al., 2024), state and county agencies that are adopting or continuing to implement predictive analytics in child protection do so with caution and high standards for scientific rigor, prediction accuracy, and bias reduction (Grimon & Mills, 2022).
The controversial, if not negative, reputation of predictive analytics in child protection overlooks, if not stymies, the potential of predictive analytics to contribute to foster care system improvements for existing youth in care. Here we highlight two proof-of-concept research studies in Illinois that demonstrate the potential of predictive analytics to support: (1) prevention efforts, (2) caseworker decision-making, and (3) resource allocation. Using administrative data on the Illinois foster care population, two time-to-event predictive risk models were developed and validated to predict youth’s risks for running away from foster care placement (Chor et al., 2022) and for experiencing residential care placement (Chor et al., 2023), respectively, after youth’s entry to foster care.
Because youth remain in foster care in Illinois longer than any other state (Children and Family Research Center, 2023), using these models to prevent running away and divert from residential care can have downstream impacts on shortening foster care stays and promoting permanent exits. Prevention efforts during a calibratable prevention window (e.g., first 3 months in foster care) can be tailored to youth’s model-predicted risk or need scores. For higher-need youth, their individual needs can be deconstructed into contributing factors in the form of significant or important predictors derived from the models. For example, demographic factors might suggest age-appropriate programs (e.g., peer mentorship), placement factors might highlight the importance of co-sibling placement (e.g., in the same foster home), while complex emotional behavioral needs might indicate evidence-based interventions (e.g., Trauma-Focused Cognitive Behavioral Therapy) in intensive community-based care (e.g., therapeutic foster care). As such, when developed, validated, and deployed carefully, predictive risk models can shape a proactive foster care system that preempts rather than reacts to a crisis.
When developed, validated, and deployed carefully, predicative risk models can shape a proactive foster care system that preempts rather than reacts to a crisis.
Predictive risk models can quantify foster care resource capacity and allocation at the system-level. Both runaway and residential care predictive risk models provide transparent information associated with each risk score threshold on the number of youth detected, predicted to experience the outcome, and model accuracy metrics (e.g., precision and recall). Because the models are not prescriptive, a foster care system, in collaboration with stakeholders, should consider important tradeoffs and system disruption in model selection: Should prediction accuracy be prioritized at the expense of serving too few youth (i.e., minimizing false negatives)? Should benefiting the masses be prioritized at the cost of over-surveilling youth and squandering limited resources (i.e., maximizing false positives)? Should fairness and equity be prioritized (i.e., prioritizing disproportionate racial and ethnic minorities) over prediction accuracy? Well-constructed, transparent predictive risk models can illuminate these discussions that are critical to the foster care system.
Human services systems tend to be laggards in technology adoption. As our society ventures into the nascent era of artificial intelligence (AI), including large language models and generative AI, it is prudent that a foster care system establishes a robust understanding of the potential and limits of predictive analytics, a prerequisite for more advanced AI applications. The time is now.