Algorithmic Fairness amid Social Determinants: Reflection, Characterization, and Approach
Zeyu Tang, Alex John London, Atoosa Kasirzadeh, Sanmi Koyejo, Peter Spirtes, Kun Zhang · 2025
Read Algorithmic Fairness amid Social Determinants: Reflection, Characterization, and Approach by Zeyu Tang, Alex John London, Atoosa Kasirzadeh, Sanmi Koyejo, Peter Spirtes, Kun Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-08337.txt.
Introduction
Zeyu Tang and co-authors argue that machine learning fairness work looks at the wrong target, in a 2025 position paper. Most fairness work checks bias against traits like race or sex, person by person. The authors say much unfairness is structural, built into housing, health, income, and other social conditions. Those conditions shape both traits and outcomes, yet models often treat them as noise to wash out.
The paper backs its claim with a theory model of college admissions, census data, and a breast cancer screening case. Fixes aimed only at single traits could create fresh structural harm, they show. Their call is to audit the social conditions first, then design fixes. New tools that measure structural injustice directly are needed, they argue.
Catalog record
| CREATOR | Zeyu Tang, Alex John London, Atoosa Kasirzadeh, Sanmi Koyejo, Peter Spirtes, Kun Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 92 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-08337
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Algorithmic Fairness amid Social Determinants: Reflection, Characterization, and Approach Zeyu Tang, Alex John London, Atoosa Kasirzadeh, Sanmi Koyejo, Peter Spirtes, Kun Zhang Abstract Social determinants are variables that, while not directly pertaining to any specific individual, capture key aspects of contexts and environments that have direct causal influences on certain attributes of an individual. Previous algorithmic fairness literature has primarily focused on sensitive attributes, often overlooking the role of social determinants. Our paper addresses this gap by introducing formal and quantitative rigor into a space that has been shaped largely by qualitative proposals regarding the use of social determinants. To demonstrate theoretical perspectives and practical applicability, we examine a concrete setting of college admissions, using region as a proxy for social determinants. Our approach leverages a region-based analysis with Gamma distribution parameterization to model how social determinants impact individual outcomes. Despite its simplicity, our method quantitatively recovers findings that resonate with nuanced insights in previous qualitative debates, that are often missed by existing algorithmic fairness approaches. Our findings suggest that mitigation strategies centering solely around sensitive attributes may introduce new structural injustice when addressing existing discrimination. Considering both sensitive attributes and social determinants facilitates a more comprehensive explication of benefits and burdens experienced by individuals from diverse demographic backgrounds as well as contextual environments, which is essential for understanding and achieving fairness effectively and transparently. 1 Introduction Structural injustice refers to circumstances in which social practices, social structures, or the environment reinforce and compound prior histories of injustice [ 20 , 114 , 137 , 128 , 107 , 6 ] . We use the term “social determinants” to refer to the specific aspects of social practices, social structures, or the environment that have a profound impact on individuals’ opportunities, behaviors, and outcomes. When members of specific demographic groups have been the subject of histories of unjust treatment, their demographic membership often correlates with circumstances in which they face significant social impediments [ 47 , 134 , 104 , 135 , 27 ] . Because social determinants are features of places, institutions, policies, or practices, they persist even if animuses that cause unjust treatment have been subject to significant reform. Their effects may not be tied directly to demographic group membership but to broader traits (such as income level or job type) or to geographic areas. As a result, individuals within the same demographic group, depending on their unique circumstances, may experience different levels of (dis)advantage due to intersecting social determinants, e.g., various environmental impacts on health in different geographic locations [ 33 , 136 , 123 ] . Conversely, individuals from different demographic groups in the same geographic neighborhood may encounter similar impediments, e.g., poverty and pollution in the neighborhood, lack of educational resource in the community [ 34 , 127 , 106 ] . Previous research on algorithmic fairness has focused on sensitive attributes, e.g., race, sex, gender, and age [ 105 , 75 , 35 , 84 , 89 , 130 , 21 , 30 , 77 , 82 , 145 , 94 , 125 ] . Various fairness metrics that are directly defined upon sensitive attributes are proposed to estimate or bound empirical violations of fairness, based on observational statistics [ 17 , 52 , 140 ] , causal properties and/or quantities [ 67 , 72 , 86 , 28 , 36 ] , and dynamic modelings [ 74 , 146 , 58 , 124 ] . In terms of auditing potential fairness violations, the focus on sensitive attributes is natural since these are the features in virtue of which individuals might be subject directly to unfair treatment or might experience disproportionate burdens. However, the goal of mitigation goes beyond auditing fairness violations by seeking to intervene in ways that will reduce burdens and promote fairer outcomes in the future. Both sensitive attributes and social determinants play important roles in the underlying causal mechanism, and therefore, need to be explicitly addressed when designing and evaluating mitigation strategies. Our contributions can be summarized as follows: • We identify a critical gap in algorithmic fairness by drawing on cross-disciplinary perspectives on social determinants. We find that existing approaches primarily focus on sensitive attributes, both in technical methods and data processing practices, largely overlooking social determinants. • Through a concrete setting of college admissions, we demonstrate how incorporating social determinants, even with geographic region as a simple surrogate, offers quantitative leverage and recovers nuanced insights that resonate with prior qualitative debates in jurisprudence. • We provide empirical evidence of real-world disparities in the interplay between sensitive attributes and social determinants. We also demonstrate how to link existing datasets to socioeconomic status indicators using address information, facilitating the development of richer fairness benchmarks. 2 Cross-Disciplinary Engagement with Social Determinants In this section, we review and reflect on the cross-disciplinary engagements with social determinants. 1 1 1 Due to space limit, we provide further discussions on related works in Appendix A . In Section 2.1 , we clarify definitions of sensitive attributes and social determinants. In Section 2.2 , we summarize discussions on social determinants from the literature of political philosophy, economics, sociology, and healthcare. In Section 2.3 , we review approaches and data processing practices in the algorithmic fairness literature, with special attention to sensitive attributes and social determinants. 2.1 Definitions: Sensitive Attributes and Social Determinants Definition 2.1 (Sensitive Attributes) . A sensitive attribute , also referred to as a protected feature or a social category , is an intrinsic attribute of the individual that is canonically recognized in law, ethics, or social norms as warranting protection from discrimination or bias. Definition 2.2 (Social Determinants) . A social determinant is a variable representing an aspect of the data generating process, in which one or more characteristics of the context (e.g., social practices, social structures, and environments) where individuals live or operate, that are not an attribute of any specific individual, have direct influence on individual’s attributes. Sensitive attributes are stable identifiers arising from longstanding legal and moral frameworks, that can be uniquely ascribed to an individual [ 105 , 75 , 35 , 84 , 89 , 130 , 21 , 30 , 77 , 57 , 82 , 145 , 94 , 125 , 109 ] . Examples include sex, race, ethnic group, disability status, religion, and so on. Social determinants refer to external conditions that influence an individual’s opportunities, behaviors, and outcomes [ 20 , 114 , 128 , 111 , 139 , 107 , 68 , 97 , 6 ] . Examples of social determinants include environmental impact on health, educational resource in the neighboring area, economic profile of the geolocation, collective values of the community, and so on. The core distinction between sensitive attributes and social determinants does not lie in whether an attribute is measured at the individual level, but rather in whether it is intrinsic to the individual. In other words, the distinction hinges on whether the attribute is determined solely by the individual or shaped by contextual influences from the surrounding environment. For instance, while the commute time is an individual-specific variable, it is not intrinsic to the person and depends on external factors such as transportation infrastructure. A person’s commute time can