The Pursuit of Algorithmic Fairness: On "Correcting" Algorithmic Unfairness in a Child Welfare Reunification Success Classifier

TL;DR

This paper introduces a new algorithm for predicting reunification success in child welfare, addressing fairness issues through a novel post-processing method that balances fairness and accuracy.

Contribution

It presents a novel classification algorithm for child welfare reunification prediction and a unique fairness mitigation procedure using penalized optimization and subsampling.

Findings

The algorithm improves prediction of stable reunifications.

The fairness mitigation method effectively balances fairness and accuracy.

The approach is adaptable to various fairness definitions and contexts.

Abstract

The algorithmic fairness of predictive analytic tools in the public sector has increasingly become a topic of rigorous exploration. While instruments pertaining to criminal recidivism and academic admissions, for example, have garnered much attention, the predictive instruments of Child Welfare jurisdictions have received considerably less attention. This is in part because comparatively few such instruments exist and because even fewer have been scrutinized through the lens of algorithmic fairness. In this work, we seek to address both of these gaps. To this end, a novel classification algorithm for predicting reunification success within Oregon Child Welfare is presented, including all of the relevant details associated with building such an instrument. The purpose of this tool is to maximize the number of stable reunifications and identify potentially unstable reunifications which may require additional resources and scrutiny. Additionally, because the algorithmic fairness of the resulting tool, if left unaltered, is unquestionably lacking, the utilized procedure for mitigating such unfairness is presented, along with the rationale behind each difficult and unavoidable choice. This procedure, though similar to other post-processing group-specific thresholding methods, is novel in its use of a penalized optimizer and contextually requisite subsampling. These novel methodological components yield a rich and informative empirical understanding of the trade-off continuum between fairness and accuracy. As the developed procedure is generalizable across a variety of group-level definitions of algorithmic fairness, as well as across an arbitrary number of protected attribute levels and risk thresholds, the approach is broadly applicable both within and beyond Child Welfare.