Measuring disparate impact in human and machine decisions

Area of Science:

• Quantitative Criminology
• Algorithmic Fairness
• Statistical Discrimination Analysis

Background:

• Empirical analyses are crucial in discrimination litigation, often using regression models to adjust for covariates.
• Traditional methods struggle to audit algorithms that exclude protected characteristics like race.
• Existing approaches may understate discrimination when protected factors are unobserved or excluded from models.

Purpose of the Study:

• To develop a novel method for measuring "unjustified" disparities in both human and algorithmic decision-making.
• To address the limitations of traditional covariate-adjusted regression in auditing modern decision systems.
• To quantify racial disparities in a real-world context, accounting for unobserved risk factors.

Main Methods:

• Introduced "risk-adjusted regression" to measure unjustified disparities.
• Step 1: Used machine learning to estimate individual "risk" based on all available data.
• Step 2: Measured decision disparities adjusted solely for estimated risk; Step 3: Assessed sensitivity to risk mismeasurement.

Main Results:

• Demonstrated the approach on 2.2 million New York City police stops.
• Showed that traditional statistical tests can significantly understate racial disparities.
• Risk-adjusted regression identified substantial, previously understated, racial disparities in police stops.

Conclusions:

• Risk-adjusted regression offers a more accurate measure of "unjustified" disparities than traditional methods.
• The approach is vital for auditing algorithmic and human decision-making for fairness.
• Findings highlight the need for advanced statistical methods to detect subtle discrimination.