Position: Beyond Sensitive Attributes, ML Fairness Should Quantify Structural Injustice via Social Determinants

Technical Summary: Beyond Sensitive Attributes, ML Fairness Should Quantify Structural Injustice via Social Determinants

Problem Statement
Current algorithmic fairness research predominantly operationalizes unfairness as discrimination along sensitive attributes (e.g., race, sex, gender). While this approach addresses individual-level bias, the authors argue it fails to capture structural injustice: systemic inequities instantiated through social determinants. Social determinants are contextual variables (e.g., neighborhood deprivation, school funding, environmental exposure) that shape opportunities and outcomes but are defined at the level of places, institutions, or policies rather than individuals. The prevailing technical paradigms often treat these contextual factors as noise to be normalized or drop them entirely during data processing, thereby rendering structural injustice invisible. Consequently, mitigation strategies focused solely on sensitive attributes may inadvertently introduce new forms of structural injustice or fail to address the root causes of disparity.

Methodology
The paper employs a multi-faceted approach combining conceptual analysis, theoretical modeling, and empirical validation:

1. Conceptual Framework: The authors define Sensitive Attributes as intrinsic, protected individual characteristics and Social Determinants as context-level variables satisfying three criteria: (1) defined at the context level (e.g., neighborhood), (2) characterized by social-structural content (shaped by policy or resource allocation), and (3) involving exogenous stratification (groupings not derived from individual traits).
2. Theoretical Modeling: A stylized model of college admissions is constructed to analyze the tension between sensitive-attribute-centered mitigation (e.g., quotas) and structural justice. The model incorporates region-specific demographic makeup and stochastic dominance in academic preparedness to derive closed-form conditions under which mitigation strategies create unintended burdens.
3. Empirical Analysis (Census Data): The authors link U.S. Census Public Use Microdata Areas (PUMAs) with the Area Deprivation Index (ADI). They visualize income distributions for specific demographic groups (e.g., African American women) across varying levels of area deprivation to demonstrate how social determinants create heterogeneity within sensitive groups.
4. Empirical Analysis (Healthcare): Using de-identified records from OSF HealthCare (approx. 54,000 mammography records), the authors analyze breast cancer screening patterns. They compare screening ages across "rich" and "poor" regions (defined by ADI) and conduct a semi-synthetic simulation to evaluate how policy interventions targeting social determinants (e.g., improving screening patterns in poor regions) affect early detection rates.
5. Technical Critique: The paper reviews existing causal fairness frameworks, demonstrating their mismatch with community-level structures. It argues that individual-level causal graphs cannot adequately represent the bidirectional influences between individuals and their communities or the aggregate nature of social determinants.

Key Contributions

• Reframing Fairness: The paper proposes a shift from auditing unfairness solely as discrimination along sensitive attributes to quantifying structural injustice via social determinants.
• Theoretical Paradox: Through the college admissions model, the authors demonstrate a counter-intuitive finding: sensitive-attribute-centered mitigation (like quotas) is less likely to disadvantage non-protected individuals in severely unjust contexts but becomes increasingly likely to generate new forms of structural injustice as conditions improve. Specifically, such strategies can impose higher competitive thresholds on non-protected applicants from disadvantaged regions who face similar contextual barriers as protected groups.
• Empirical Evidence of Heterogeneity: The analysis of Census and healthcare data reveals that individuals with identical sensitive attributes experience vastly different outcomes based on their social determinants (e.g., median income for African American women varies significantly by ADI; screening ages for White women vary by over 3 years between rich and poor regions).
• Critique of Current Paradigms: The authors show that standard benchmarks (e.g., Adult, Folktables) often drop address-related variables, and causal models typically operate at the individual level, failing to capture community-level structural forces.
• Actionable Pillars: The paper outlines three pillars for future work: (1) Data Governance (privacy-aware use of contextual aggregates), (2) Metric Design (dynamic, context-adaptable fairness metrics like "Social Determinant Parity"), and (3) Causal Modeling (multi-layer models treating social determinants as explicit intervention levers).

Results

• Theoretical: Theorem 4.5 establishes that quota-based admissions can disadvantage non-Underrepresented Minority (URM) applicants from poor regions more than URM applicants from rich regions, particularly as structural injustice improves and the gap in academic preparedness narrows.
• Empirical (Census): Income distributions for African American women are heavily skewed toward lower incomes in high-ADI (high deprivation) PUMAs compared to low-ADI PUMAs, despite identical sensitive attributes.
• Empirical (Healthcare): Even within a unified healthcare system with uniform guidelines, the median age of first breast cancer screening for White women differs by nearly 5 years between rich and poor regions. Semi-synthetic simulations show that adopting "improved" screening patterns (mimicking rich regions) in poor regions increases early detections from ~1,367 to ~1,461 per 10,000 screenings, demonstrating that structural interventions yield non-trivial gains.
• Technical: Enforcing Demographic Parity (DP) while including social determinants as features does not consistently reduce disparity and can sometimes worsen observed violations when outcomes are stratified by social determinants.

Significance and Claims
The paper claims that the field of ML fairness must move beyond "non-discrimination" (focusing on sensitive attributes) to address "structural justice." The authors argue that without auditing and quantifying structural injustice via social determinants, mitigation strategies risk being superficial or even counterproductive. They contend that social determinants are not merely "another" sensitive attribute or a proxy for one, but represent fundamental structural levers that require distinct technical approaches, including new data governance frameworks, dynamic metrics, and causal models capable of handling community-level interactions. The ultimate goal is to enable policy interventions that are better informed by the systemic drivers of inequality rather than just the demographic outcomes.