Algorithmic Bias in Recidivism Prediction: A Causal Perspective

COMPAS, a widely-used criminal justice tool, demonstrates measurable racial bias against African American defendants when analyzed through causal fairness methods, confirming earlier concerns raised by ProPublica about the software's predictions.

Traditional fairness metrics may miss important bias patterns — this research uses a causal approach (FACT) that can detect bias that standard statistical methods might overlook, suggesting the need for more sophisticated evaluation techniques in algorithmic auditing.

The bias findings remain robust even when accounting for unmeasured confounding variables, meaning the racial disparities in COMPAS predictions are unlikely to be explained away by missing data or hidden factors, strengthening confidence in the bias conclusion.

Causal reasoning frameworks offer a more rigorous foundation for understanding algorithmic fairness than conventional approaches, potentially providing better tools for regulators and organizations to identify and address bias in high-stakes decision systems like criminal justice algorithms.

This research demonstrates the importance of independent auditing of algorithms used in consequential domains — the COMPAS case shows that bias can persist in deployed systems even when developed by established companies, underscoring the need for ongoing external scrutiny.