Causal Frameworks for Algorithmic Fairness

191 concepts145 papers1 proposal

When we train machine learning models on historical data, they inevitably learn to replicate historical inequities. To combat this, early algorithmic fairness research focused on statistical or observational metrics. Frameworks like demographic parity or equalized odds—formalized by Moritz Hardt et al. in Equality of Opportunity in Supervised Learning—attempted to balance error rates or selection rates across demographic groups.

However, these statistical approaches suffer from a fundamental limitation: they look only at correlations within the joint distribution of predictions, attributes, and outcomes. Because correlation does not imply causation, observational metrics cannot distinguish between different underlying social mechanisms. Two entirely different data-generating processes—one socially justifiable and one deeply discriminatory—can produce identical statistical distributions. To resolve this ambiguity, a growing body of research shifts the focus from statistical parity to causal fairness.

The Failure of Pure Correlation

Why do statistical fairness metrics fall short? Consider a hiring algorithm. A statistical metric might demand that the selection rate for male and female applicants be equal (demographic parity), or that the false positive rates be balanced across groups.

However, as argued by Niki Kilbertus et al. in Avoiding Discrimination through Causal Reasoning, purely observational criteria cannot tell us why a disparity exists. A difference in outcomes might stem from a legitimate, job-relevant mediator (such as specific technical qualifications), or it might be driven by systemic bias acting through an unfair proxy variable (such as a neighborhood zip code). Because statistical metrics ignore the data provenance, forcing statistical parity can act as a blunt instrument, sometimes destroying model utility or even inadvertently worsening local discrimination, a phenomenon highlighted by Salvatore Ruggieri et al. in Can We Trust Fair-AI?.

The Counterfactual Ideal

To inject mechanistic insight into fairness, researchers introduced Structural Causal Models (SCMs). Instead of asking whether a model's predictions are statistically balanced, causal fairness asks: Did the sensitive attribute cause the decision?

The foundational concept here is Counterfactual Fairness, introduced by Matt Kusner et al. in Counterfactual Fairness. The intuition is straightforward: a decision is fair to an individual if it remains exactly the same in a counterfactual world where their protected attribute (such as race or gender) was flipped, while holding all other independent background factors constant.

$$\text{P}(\hat{Y}_{A \leftarrow a} = y \mid X=x, A=a) = \text{P}(\hat{Y}_{A \leftarrow a'} = y \mid X=x, A=a)$$

This definition forces the model to be structurally blind to the downstream effects of historical bias. However, implementing it requires deep assumptions about the world. Interestingly, the theoretical boundaries of this concept are still being mapped; Lucas Rosenblatt et al. demonstrated in Counterfactual Fairness Is Basically Demographic Parity that under standard linear assumptions, counterfactual fairness mathematically collapses into group-level demographic parity, illustrating how tightly bound these concepts become when causal structures are simplified.

Decomposing Disparity into Pathways

Causal graphs allow practitioners to perform a surgical dissection of algorithmic bias. Rather than treating all disparities as monolithic, we can trace how a sensitive attribute $A$ propagates through a network of variables to influence the final prediction $\hat{Y}$.

In Fairness in Decision-Making - The Causal Explanation Formula, Junzhe Zhang et al. formalize how total observed disparity can be decomposed into three distinct causal pathways:

  1. Direct Effects: The sensitive attribute directly alters the outcome ($A \rightarrow \hat{Y}$). This represents explicit, systemic discrimination based on the identity trait itself.
  2. Indirect Effects: The attribute influences the outcome through intermediate variables ($A \rightarrow M \rightarrow \hat{Y}$). Here, the ethics depend entirely on the nature of the mediator $M$. If $M$ is a legitimate qualification, the path may be fair; if $M$ is an unfair proxy (like a historical redlining variable), it constitutes indirect discrimination.
  3. Spurious Effects: The correlation between $A$ and $\hat{Y}$ is driven by a shared, underlying confounder ($A \leftarrow U \rightarrow \hat{Y}$).

By isolating these pathways, frameworks like the one proposed by Ilya Shpitser et al. in Fair Inference on Outcomes allow designers to block or penalize only the discriminatory pathways while preserving the predictive power of legitimate ones. This alignment is crucial for legal compliance; as the authors note in Equality of Opportunity in Classification: A Causal Approach, anti-discrimination law often requires proving specific causal mechanisms of disparate treatment or impact, which purely statistical metrics fail to satisfy.

Intervening on the Data Generation Process

Once the causal pathways of bias are identified, how do we fix the model? Rather than adjusting the final predictions post-hoc, causal fairness often focuses on pre-processing the training data itself to simulate a fairer world.

Early optimization work by Flavio Calmon et al. in Optimized Pre-Processing for Discrimination Prevention showed that data distributions could be transformed to minimize discrimination while bounding individual data distortion. In the causal domain, this manifests as mapping real-world data into an idealized, interventional distribution.

For example, Ludwig Bothmann et al. propose a data-warping framework that projects observations into a fictitious world where the causal link between the protected attribute and the outcome is broken entirely, while carefully preserving an individual's relative merit ranking. By training standard machine learning models on this causally sanitized data, practitioners can ensure that downstream predictions remain free of structural bias without sacrificing the model's operational utility.

The Causal-Observational Gap

Despite its mathematical elegance, causal fairness faces a steep practical challenge: the gap between theoretical causal definitions and available data.

As Drago Plecko et al. emphasize in Causal Fairness for Outcome Control, computing counterfactual metrics typically requires individual-level structural equations that are impossible to observe directly from standard, static datasets. Furthermore, real-world systems are plagued by unobserved confounders and latent factors that evolve dynamically over time. When models ignore how current decisions reshape future data distributions, even well-intentioned fairness interventions can backfire, a long-term risk detailed by Zeyu Tang et al. in Tier Balancing: Towards Dynamic Fairness over Underlying Causal Factors.

Ultimately, causal fairness moves the conversation away from superficial statistical matching toward a rigorous, explicit modeling of social mechanisms. It forces practitioners to state their assumptions about how bias enters the data, providing a transparent, legally justifiable foundation for building truly equitable algorithms.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 148 papers
  2. Version 1 · 2026-10-04 · 216 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_causal_fairness_ce394e88,
  title = {causal fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-fairness/rdr_ce394e885d264cdda647bfc91d6d6022},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_causal_fairness_ce394e88_v2,
  title = {causal fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-fairness/rdr_ce394e885d264cdda647bfc91d6d6022/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake