Abstract: Fairness Analysis of a Risk Stratification Model Developed to Support Child Welfare Supervisors (Society for Social Work and Research 26th Annual Conference - Social Work Science for Racial, Social, and Political Justice)

Fairness Analysis of a Risk Stratification Model Developed to Support Child Welfare Supervisors

Sunday, January 16, 2022
Marquis BR Salon 12, ML 2 (Marriott Marquis Washington, DC)
* noted as presenting author
Eunhye AHN, MSW, PhD Student, University of Southern California, Los Angeles, CA
John Prindle, PhD, Research Faculty, University of Southern California, Los Angeles, CA
Rhema Vaithianathan, PhD, Professor, Auckland University of Technology, Auckland, New Zealand
Emily Putnam-Hornstein, PhD, John A. Tate Distinguished Professor for Children in Need, University of North Carolina at Chapel Hill, Chapel Hill, NC
Background and Purposes: As child welfare agencies consider the use of predictive risk models to support decision-making, a fundamental question is whether a data model can be applied “fairly” to different subpopulations of children and their families. Of specific concern and interest is the extent to which these models can accurately assess risk for children of different racial / ethnic groups given structural forces that have led to significant economic inequalities and biases that have affected who comes to the attention of the system. In the current analysis, we adopt different definitions of fairness to assess the classification accuracy of a predictive risk model developed in Los Angeles County, California by race and ethnicity.

Methods: Fairness was examined using three commonly metrics: (a) calibration, examines whether predicted risk scores reflect the same outcome rates regardless of an individual’s racial group membership; (b) predictive parity, examines whether the rates at which the outcome occurred at a given risk threshold were the same independent of racial and ethnic membership, and (c) error rate balance, assesses whether the false positive and false negative error rates were equal between racial and ethnic groups for a given risk classification threshold.

Results: Focusing on the referrals assessed to be at the top 10%, or the threshold established by Los Angeles County for enhanced services, the model was well calibrated with similar removal rates emerging across racial / ethnic groups. Predictive parity was also satisfied as referrals at the top 10% of risk had similar true positive rates across racial groups. Yet, high-risk referrals did not satisfy the error rate balance definition of fairness. The false positive rate of referrals included Black children (5.4%) was almost two folds of those included White children (2.6%) and Hispanic children (2.9%). When it comes to the false negative rates, referrals included Hispanic children showed the highest numbers (58.4%) followed by those included White (54.9%) and Black (50.4%).

Conclusions and Implications: The results of the current fairness analysis must be interpreted in the context of observed removal (or foster care placement) rates that differ between racial and ethnic groups. Specifically, Black children have historically had the highest removal rate (22%) followed by Hispanic (16%) and White (15%) children. Racial disparities of error rates can be attributed, in part, to this varying removal rate. The implications of different definitions of fairness must be understood in the context of the interventions associated with classification designations.