Causal Reasoning in Fairness

Causal reasoning in fairness means asking 'what would happen to this person's outcome if we changed their protected attribute in a counterfactual world, while keeping everything else about them the same?' This approach reveals hidden discrimination that statistical methods miss—like when a hiring algorithm rejects candidates from certain ZIP codes because those neighborhoods correlate with race, even though the algorithm never sees race directly. Think of it as tracing the actual pathways through which sensitive information flows, rather than just looking at surface-level correlations. Check this: Statistical fairness metrics can hide proxy discrimination, but causal models expose it.

Causal reasoning approaches fairness by building structural causal models that map how protected attributes (like race or gender) influence outcomes through different pathways. The key insight is that not all associations between protected attributes and decisions are discriminatory—some flow through legitimate intermediate variables called 'resolving variables.' For example, in college admissions, gender might affect department choice, and department choice affects admission decisions. If we consider department choice a legitimate factor (a resolving variable), then the pathway gender → department → admission is acceptable, but a direct pathway from gender to admission would be discriminatory.

This framework distinguishes between two critical types of discrimination. 'Unresolved discrimination' occurs when a protected attribute has a direct causal influence on the decision that isn't explained by legitimate resolving variables. 'Proxy discrimination' occurs when the algorithm uses variables that stand in for protected attributes—like using ZIP code, neighborhood, or name as a proxy for race or ethnicity. The causal approach lets us identify exactly which pathways in the decision-making process are problematic and which are acceptable.

In practice, this means we can design algorithms that eliminate unfair pathways while preserving legitimate ones. For hiring, we might intervene on proxy variables like name or ZIP code to remove their influence while still allowing qualifications and experience to affect decisions. For lending, we can distinguish between using credit history (a resolving variable that might correlate with race) versus using neighborhood data (a proxy that directly encodes racial information). This precision helps organizations build fairer systems without sacrificing predictive power unnecessarily.

Causal reasoning transforms fairness from selecting statistical metrics to explicitly modeling the data-generating process and justifying assumptions about which causal pathways are acceptable. Kusner et al. (2017) introduced counterfactual fairness, requiring that an individual's prediction be the same in the actual world and a counterfactual world where they belonged to a different demographic group [1]. Kilbertus et al. (2017) formalized two key causal criteria: unresolved discrimination (direct influence from protected attributes not mitigated by resolving variables) and proxy discrimination (influence through proxy variables like ZIP code or name) [2]. Russell et al. (2018) demonstrated that observational fairness criteria cannot distinguish between intuitively different discrimination scenarios, making causal assumptions necessary for any meaningful fairness analysis [3]. Pearl's structural causal model framework provides the formal foundation for these approaches, using directed acyclic graphs to represent causal relationships and the do-calculus to reason about interventions.

Real-world applications span multiple domains. In criminal justice, causal analysis has revealed that risk assessment tools may rely on proxies like arrest location or socioeconomic status that correlate with race. In hiring, algorithms using features like college prestige, zip code, or even writing style may encode demographic information. In healthcare, predictive models using healthcare spending as a proxy for health needs systematically disadvantaged Black patients because less money was historically spent on their care, even when they were equally sick.

A key technical challenge is that building accurate causal models requires domain expertise and assumptions about the data-generating process that cannot be verified from data alone. The same joint distribution of variables can correspond to multiple different causal structures that have very different fairness implications. This means fairness decisions inevitably involve normative judgments about which pathways are legitimate—not just technical fixes.

Causal approaches to fairness face several practical and conceptual limitations. First, building valid causal models requires substantial domain knowledge and assumptions about the data-generating process that often cannot be verified empirically—the same observed data can be consistent with multiple different causal graphs. Second, counterfactual reasoning about deeply rooted attributes like race or gender raises conceptual challenges, as interventions on these attributes are difficult to define meaningfully. Third, causal models can become complex quickly in real-world systems with many variables and interacting pathways, making them difficult to specify accurately and computationally intensive to work with. Fourth, determining which variables qualify as 'resolving variables' versus 'proxies' involves normative judgments that may vary across contexts and stakeholders—there is no purely technical solution to what counts as a legitimate causal pathway. Finally, causal methods typically require more data and stronger assumptions than purely statistical approaches, and they may not be feasible when key variables are unobserved or when the causal structure is poorly understood.