TL;DR
Fairness interventions that target a single sensitive attribute (like race) can unintentionally worsen disparities for other attributes (like gender), a phenomenon the authors call "cross-attribute externalities." By using a new framework based on Wasserstein distances, this paper proves that these side effects are common, often non-monotonic, and can be mathematically bounded even when some sensitive data is hidden from the practitioner.
What Problem Does This Paper Address?
Most algorithmic fairness tools are "single-issue" voters: they are designed to reduce disparities for one specific group at a time. However, individuals belong to multiple protected groups simultaneously (intersectionality).
The authors identify a critical gap: improving fairness for "Attribute A" does not guaranteed improvement for "Attribute B." In fact, it can often make things worse. This "Robbing Peter to Pay Paul" effect is particularly dangerous because practitioners often lack data for every sensitive characteristic (e.g., they might have data on age but not race). The paper seeks to understand how these interventions shift disparities across both observed and unobserved attributes and whether we can predict or bound these unintended consequences.
Methods
The authors move away from simple summary statistics (like comparing mean scores) and instead adopt a distributional framework using Wasserstein barycentric disparity indices.
- Wasserstein Indices: These measure how much the entire distribution of scores for one group differs from a "fair" average distribution (the barycenter). This captures more nuance than just looking at averages.
- Rank-Preserving Barycentric Correction: They analyze a specific intervention that moves individual scores toward the barycenter while maintaining their relative ranking within their group.
- Gaussian & Categorical Modeling: The paper derives closed-form trajectories for how disparities evolve in a jointly Gaussian model. This allows them to identify exactly when a "fix" for one attribute will "unmask" or exacerbate unfairness in another.
- Partial Identification: For cases where an attribute is hidden (unobserved), they develop a method to calculate "sharp bounds"—the best possible range of outcomes—based on known correlations and group proportions.
Datasets
The researchers evaluated their framework across nine diverse real-world datasets:
- Binary Classification: Adult (income), COMPAS (recidivism), Default (credit), Mep15, and Mep16 (healthcare utilization).
- Regression: Law School (GPA/Bar passage), Insurance (costs), Communities and Crime (crime rates), and ACSIncome (individual income).
Key Insights and Ideas
- Masking and Unmasking: Disparities can sometimes appear small because different correlated biases "cancel each other out" in the raw data. When you intervene to fix one bias, you disrupt this cancellation, "unmasking" a much larger disparity in another attribute.
- The Non-Monotonic Trap: Fairness isn't always a straight line. As you increase the intensity of a fairness correction, the disparity for a hidden attribute might initially decrease (get better) but then suddenly spike (get worse) as the correction continues.
- Joint vs. Marginal Parity: Achieving fairness for race and gender individually does not guarantee fairness for the intersectional group (e.g., Black women). The authors show that joint disparity indices are necessary to see the full picture.
- Attribution $\neq$ Intervention: Just because a model's unfairness is "attributed" to race (using methods like Shapley values) doesn't mean that a race-based intervention will reduce the total unfairness by that same amount.
Key Results
The empirical results confirm that cross-attribute externalities are the rule, not the exception:
- Classification: In more than half of the datasets studied, intervening on one attribute increased unfairness for the hidden attribute in the majority of experimental scenarios.
- Regression (ACSIncome): Correcting for race initially helped reduce sex-based disparities. However, subsequently correcting for age actually increased sex-based disparities, highlighting the complex, non-monotonic nature of sequential fixes.
- Visualizing the Trade-off: As seen in Figure 2, while the targeted disparity ($U_1$) always goes to zero, the non-targeted disparity ($U_2$) can increase immediately or after a brief period of improvement.

- Quantitative Unmasking: Table 1 shows a scenario where $U_2$ (the non-targeted attribute) starts at 0.0000 (perfect parity) but jumps to 0.1475 after a full correction for $U_1$. The intervention literally created a disparity where none was visible before.
Limitations and Future Work
- Criterion Focus: The study focuses primarily on Demographic Parity. Future work could explore how these externalities manifest under other definitions, such as Equalized Odds or Predictive Parity.
- Gaussian Assumptions: While the authors provide categorical extensions, some of the most precise theoretical bounds rely on Gaussian assumptions which may not hold for all real-world score distributions.
- Causal vs. Statistical: The paper treats these as distributional transformations rather than causal interventions. Integrating causal discovery could help explain why these correlations exist and how to break them more effectively.
Summary updated