Racial disparities in recidivism remain a persistent challenge, and the growing adoption of risk assessment algorithms has intensified the scrutiny of their sources. Past works have primarily focused on disparities in the predictions of these algorithms, viewing recidivism as a binary outcome. While sociological and criminological research has long documented non-algorithmic factors in recidivism, it remains unclear whether the risk assessments that decision-makers act on fully account for racial disparities. This work presents a multi-stage causal framework for time-to-recidivism that captures the interactions between race, the risk assessment algorithm, and contextual factors. We introduce interventional racial parity and a formal survival analysis test, conducted with observational data, of whether the algorithmic risk assessment fully accounts for racial differences in recidivism. Applied to the COMPAS dataset, the test detects no racial disparity within risk groups at short follow-up horizons. A statistically significant disparity becomes detectable after roughly nine months in the low-risk group and persists under finer score stratification, risk score perturbation, and a competing-risks analysis. This suggests that factors beyond the algorithmic scores, possibly including structural disparities in housing, employment, and social support, may shape recidivism over time, underscoring the need for policy interventions beyond algorithmic improvements, particularly for low-risk defendants.

翻译：暂无翻译