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

TL;DR

It is argued that auditing structural injustice through social determinants must precede mitigation, and called for new technical developments that move beyond sensitive-attribute-centered notions of fairness as non-discrimination.

Abstract

Algorithmic fairness research has largely framed unfairness as discrimination along sensitive attributes. However, this approach limits visibility into unfairness as structural injustice instantiated through social determinants, which are contextual variables that shape attributes and outcomes without pertaining to specific individuals. This position paper argues that the field should quantify structural injustice via social determinants, beyond sensitive attributes. Drawing on cross-disciplinary insights, we argue that prevailing technical paradigms fail to adequately capture unfairness as structural injustice, because contexts are potentially treated as noise to be normalized rather than signal to be audited. We further demonstrate the practical urgency of this shift through a theoretical model of college admissions, a demographic study using U.S. census data, and a high-stakes domain application regarding breast cancer screening within an integrated U.S. healthcare system. Our results indicate that mitigation strategies centered solely on sensitive attributes can introduce new forms of structural injustice. We contend that auditing structural injustice through social determinants must precede mitigation, and call for new technical developments that move beyond sensitive-attribute-centered notions of fairness as non-discrimination.

Figures (7)

• Figure 1: Histogram of annual income for African American women residing in different areas. Sensitive attributes (race and sex) are shared across the subfigures, whereas social determinants are not (higher ADI indicates higher area deprivation).
• Figure 2: Panels (a) and (b): Histogram of the age at the first-ever breast cancer screening for White women residing in different areas. Sensitive attributes (race and sex) are shared, whereas social determinants are not: the rich region corresponds to ADI $\in [0, 25)$ and the poor region corresponds to ADI $\in [75, 100)$. Panels (c) and (d): We conduct a semi-synthetic simulation of breast cancer onset and screening to evaluate how policy interventions targeting social determinants affect early detection outcomes. Panel (c) shows that adopting improved (rich-region) screening pattern in the poor region shifts detections earlier. Panel (d) demonstrates quantitatively that such structural improvements yield non-trivial gains in early detections across allocation scenarios (with a same total number of screenings).
• Figure 3: Fairness implications of different admission strategies. Panel (a): quota-based admission can introduce additional unfairness against non-URM applicants from the poor region. Panel (b): holistic review with plus factors tends to benefit URM applicants in the rich region more than these in the poor region. Panel (c): top-percentage plan transfer admission opportunity from the rich region to the poor region, and the redistribution is proportional to the natural region-specific demographic compositions.
• Figure 4: Age distribution in different PUMA regions in California based on US Census data.
• Figure 5: Racial composition in various PUMA regions in California based on U.S. Census data.

Theorems & Definitions (9)

• Definition 2.1: Sensitive Attributes
• Definition 2.2: Social Determinants
• Theorem 4.5: Tension Between Structural Justice and Sensitive-Attribute-Centered Mitigation
• Theorem
• proof
• Theorem 2.2: Holistic Review with Plus Factors Benefits URM in Rich Region More
• proof
• Theorem 2.3: Top-Percentage Plans Reallocate Spots from Rich Region to Poor Region
• proof