Algorithmic Fairness amid Social Determinants: Reflection, Characterization, and Approach

Zeyu Tang, Alex John London, Atoosa Kasirzadeh, Sanmi Koyejo, Peter Spirtes, Kun Zhang

Introduction

Most machine learning (ML) fairness research focuses on "sensitive attributes" like race or gender. However, this narrow view often ignores the "structural injustice" embedded in our environments. This paper argues for a shift toward quantifying social determinants—contextual factors like neighborhood affluence or healthcare access—to prevent fairness interventions from inadvertently creating new forms of inequality.

The Problem: Structural Injustice vs. Individual Discrimination

Current algorithmic fairness paradigms largely frame unfairness as individual-level discrimination. If a model treats two people of different races differently, it is considered "unfair." This approach has a significant blind spot: it ignores structural injustice, where social structures and environments reinforce historical disadvantages regardless of individual traits.

In technical terms, many ML practitioners treat contextual variables, such as a user's ZIP code, as "noise" to be normalized or "proxies" to be removed to prevent redlining. We contend that these variables are actually the "signal" of structural injustice. By ignoring them, we fail to audit the very systems that produce disparate outcomes, leading to "fair" models that still operate within—and potentially exacerbate—unjust systems.

Methodology and Data

The research combines philosophical positioning with rigorous mathematical and empirical modeling. The authors introduce a stylized model of college admissions to analyze how traditional "quota-based" affirmative action interacts with regional affluence, using stochastic dominance to represent differences in academic preparedness between "rich" and "poor" regions.

To ground these theories, the paper utilizes three primary data sources:

  1. U.S. Census PUMS (Public Use Microdata Sample): Linked with the Area Deprivation Index (ADI) to visualize shifts in income distributions for the same demographic groups based on neighborhood-level deprivation.
  2. OSF HealthCare Records: A de-identified dataset of ~54,000 mammography records linked to residential ZIP codes and socio-economic indices.
  3. SEER (Surveillance, Epidemiology, and End Results): Used to provide realistic age-specific breast cancer incidence rates for semi-synthetic simulations.

Key Insights: Beyond Sensitive Attributes

Within-Group Heterogeneity

A core finding is that individuals with identical sensitive attributes face vastly different economic realities depending on their environment. For example, Black women in high-deprivation areas have significantly different income profiles than those in low-deprivation areas. Auditing only for race misses the structural burden of the environment.

Income distributions for Black women vary significantly across different neighborhood deprivation levels (ADI).

The Quota Paradox

Mathematical analysis reveals a counter-intuitive finding: traditional demographic quotas can actually harm the most disadvantaged members of "majority" groups. These quotas are more likely to generate new injustices as general social conditions improve, because the gap between "advantaged minority" and "disadvantaged majority" applicants narrows.

Key Results: Impact of Social Determinants

The paper demonstrates the power of intervening on social determinants through a breast cancer screening simulation. By simply shifting the screening patterns of a "poor" region to match those of a "rich" region, early cancer detections increased significantly.

Early detection rates improve when screening patterns in poor regions are optimized to match or exceed those in rich regions.

Specifically, the simulation showed that improving screening access in poor regions increased early detections from 1,367 to 1,461 per 10,000 screenings. Empirical validation using Census data further confirmed that the Area Deprivation Index is a more powerful predictor of income distribution shifts than intersectional demographic categories alone.

Limitations and Future Directions

The authors acknowledge several hurdles in implementing this approach:

  • Data Privacy: Using granular social determinants increases re-identification risks. The paper proposes privacy-aware data governance using differential privacy.
  • Causal Complexity: Real-world social systems involve feedback loops (individuals affect communities, and vice versa) that current acyclic causal ML tools struggle to model.
  • Metric Design: There is a need for new "Social Determinant Parity" metrics that track disparities across evolving structural conditions rather than just static demographic groups.

Ultimately, fairness cannot be "fixed" at the model level alone. Machine learning must be used to identify the structural levers—like healthcare access or educational resources—that can be pulled to create a more just society.

Summary updated

BibTeX
@misc{tang2025algorithmic,
  author = {Zeyu Tang and Alex John London and Atoosa Kasirzadeh and Sanmi Koyejo and Peter Spirtes and Kun Zhang},
  title = {{Algorithmic Fairness amid Social Determinants: Reflection, Characterization, and Approach}},
  year = {2025},
  eprint = {2508.08337},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.08337},
}
Figure