Recent estimates suggest that one in three children will experience at least one maltreatment investigation in their childhood, one in 17 will spend time in foster care, and one in 100 will be involved in a case where parental rights are terminated. As agencies around the country work to screen and respond to the over 4 million child maltreatment referrals that come in annually, they are increasingly incorporating predictive risk modeling (PRM) approaches into their decision-making processes. PRMs take large volumes of administrative data available on individuals associated with a case or referral and distill the information into a single risk score reflecting the likelihood of some near-term adverse event. From the very start, the use of predictive analytics in the child welfare system has been highly contentious. There is concern that those experiencing poverty and Black and brown communities — who are overrepresented at all points in the child welfare system — will only be further disadvantaged by deploying such tools.
In this talk, Alex Chouldechova will describe some of the actively deployed PRMs and discuss what recent studies have revealed about the impact of PRMs on decision-making and racial disparities in the system. She will also discuss several studies of affected community perspectives on the use of algorithms in child welfare. Throughout, she will highlight the emerging demands for procedural, informational, distributive, and interpersonal justice. Finally, she will reflect on the role of research and technology in systems widely perceived as unjust.
Speaker Bio:
Chouldechova is a principal researcher in the Fairness, Accountability, Transparency and Ethics (FATE) group at Microsoft Research NYC and the Estella Loomis McCandless Associate Professor of Statistics and Public Policy at Carnegie Mellon University. Her research investigates questions of algorithmic fairness and accountability in data-driven decision-making systems, with a domain focus on criminal justice, human services, and pre-trained models.
