Algorithmic fairness in child protection decision-making

Description

The field of child protection is faced with two major challenges: One is to identify children who need to be protected from child abuse and neglect, the other is to select the most promising intervention once a case of child maltreatment has been identified. In recent years, new approaches to statistical prediction exploiting the advantages of machine learning have been applied to the first task, the identification of risks. In a previous study funded by the Swiss National Science Foundation’s Spark program, David Lätsch and his team used machine learning, among other novel approaches, for the second task, the selection of interventions. In the current project, which is funded through a DIZH Fellowship granted to David Lätsch, the objectives are to continue this work by developing, testing, and comparing different approaches to prediction with regard to the trajectories of child protection interventions. In the second part, the team will work out conceptual clarifications and corresponding statistical metrics for different criteria of algorithmic fairness, such as calibration, predictive parity, and error-rate balance. The predictive approaches developed in the previous part of the study will then be examined with regard to how well they satisfy fairness criteria. In the final part, the team is going to present and discuss the implications of their findings in several workshops with stakeholders from the field partner and selected researchers.