Decomposing Causal Pathways for Algorithmic Fairness

212 concepts111 papers2 proposals

When we train machine learning models on historical data, they risk inheriting and amplifying human biases. Traditional statistical fairness metrics—such as demographic parity or equalized odds—attempt to solve this by forcing mathematical equality across demographic groups. However, these purely statistical approaches act as a sledgehammer: they treat all correlations between a protected attribute (like race or gender) and an outcome as inherently discriminatory.

By ignoring the underlying mechanisms of how data is generated, statistical fairness can lead to severe performance degradation and even absurd real-world policies. To build models that are both fair and highly accurate, we need a scalpel, not a sledgehammer. We must look past statistical correlations and trace the specific causal pathways through which bias propagates.

The Limitations of Total Fairness

To understand why a purely statistical or global approach fails, consider how causality entered the fairness landscape. Matt Kusner et al. introduced the concept of Counterfactual Fairness, which posits that a decision is fair if it remains the same in a counterfactual world where an individual's protected attribute was different.

While counterfactual fairness was a massive leap forward, it treats the entire downstream causal footprint of a sensitive attribute as problematic. As Silvia Chiappa points out in Path-Specific Counterfactual Fairness, this global view is often overly restrictive. In many real-world scenarios, a sensitive attribute might influence an outcome through multiple distinct intermediate variables (mediators). Some of these pathways represent systemic discrimination, while others might reflect legitimate, socially acceptable, or legally permissible domain-specific relationships.

For example, if we are predicting car insurance risk, an individual's age causally influences their years of driving experience, which in turn legitimately influences their accident risk. If a fairness framework forces the model to completely ignore the entire causal chain originating from age, it destroys the model's ability to utilize valid, non-discriminatory information like driving experience.

The Scalpel: Admissible vs. Impermissible Paths

Path-specific causal fairness resolves this tension by mapping the data-generating process as a Structural Causal Model (SCM), typically visualized as a Directed Acyclic Graph (DAG). Within this graph, the total effect of a protected attribute $A$ on a decision $Y$ is broken down into a web of directed paths.

   [ Direct Path (Impermissible) ]
         /-----------------\
        v                   \
  Protected (A) --------> Outcome (Y)
        \                   ^
         v                 /
     Mediator (M) --------/
   [ Resolving Path (Admissible) ]

To separate fair variations from unfair bias, we categorize these paths into two sets:

  1. Impermissible Paths: Pathways that transmit unlawful or socially unacceptable discrimination (such as a direct link $A \to Y$, or an indirect path through an unfair proxy variable).
  2. Admissible Paths: Pathways mediated by what Niki Kilbertus et al. term resolving variables—intermediate factors that are explicitly deemed socially or legally acceptable for influencing the outcome (such as Avoiding Discrimination through Causal Reasoning).

Crucially, as Ilya Shpitser et al. emphasize in Fair Inference on Outcomes, deciding which pathways are fair or unfair is fundamentally a domain-specific policy decision rather than a purely mathematical one. The mathematics simply provides the framework to enforce these human-defined boundaries.

The Mechanism: How Path Isolation Works

How do we mathematically isolate the influence traveling along one specific path while ignoring another? The secret lies in nested counterfactuals.

Imagine a simple graph where a protected attribute $A$ influences an outcome $Y$ directly ($A \to Y$) and indirectly through a legitimate mediator $M$ ($A \to M \to Y$). We want to eliminate the direct bias but preserve the legitimate information flowing through $M$.

To do this, we evaluate the outcome under a counterfactual intervention where we split the identity of $A$. We set $A$ to a baseline counterfactual value $a'$ for the direct path, but allow it to remain at its factual value $a$ for the path leading to the mediator. This is written as:

$$Y(A = a', M(A = a))$$

This expression asks a precise question: "What would this individual's predicted outcome be if their protected attribute were changed to $a'$, but their mediator $M$ retained the exact value it naturally took under their true attribute $a$?"

By holding the mediator's causal history constant while toggling the attribute along the direct path, we isolate the direct path's effect. If the model's prediction changes during this thought experiment, it proves that the model is relying on an impermissible pathway.

To compute these complex, nested counterfactuals from real-world observational data, Daniel Malinsky et al. developed a formal graphical framework known as the potential outcomes calculus (A Potential Outcomes Calculus for Identifying Conditional Path-Specific Effects). This calculus provides the exact mathematical conditions required to identify and estimate path-specific effects without needing to run randomized controlled trials. By leveraging these tools, Junzhe Zhang et al. demonstrate in Equality of Opportunity in Classification: A Causal Approach that observed statistical disparities can be cleanly decomposed into counterfactual direct, indirect, and spurious components.

Implementing Path-Specific Interventions

Once we can isolate these pathways, we can intervene to enforce fairness. The literature implements these interventions across three distinct stages of the machine learning pipeline:

1. Pre-processing (Data Adaptation)

Instead of altering the model, we alter the training data itself to purge bias from specific channels. For instance, Drago Plecko et al. developed fairadapt: Causal Reasoning for Fair Data Preprocessing. Their framework uses the causal graph to mathematically shift and transform individual data features. It explicitly protects variations along pathways designated as resolving, while selectively erasing the demographic signature from features downstream of impermissible paths.

2. In-processing (Constrained Optimization)

Alternatively, we can penalize models during training when they exploit forbidden paths. Yuchen Ma et al. introduce counterfactual mediator regularization in Counterfactual Fairness for Predictions using Generative Adversarial Networks, adding a penalty term that minimizes the prediction discrepancy between factual data and counterfactual versions of impermissible mediators.

For complex data distributions, the authors of Achieving Causal Fairness through Generative Adversarial Networks use a GAN architecture (CFGAN) that mirrors the causal graph, generating synthetic datasets where specific discriminatory pathways are structurally neutralized while preserving legitimate predictive signals.

3. Post-processing (Output Correction)

If retraining a model is too expensive, we can adjust the outputs of a pre-trained model. Daisuke Hatano et al. show in Path-Specific Counterfactual Fairness via Dividend Correction that Harsanyi dividends from cooperative game theory are mathematically equivalent to path-specific causal effects. This allows practitioners to calculate the exact numerical contribution of an unfair causal path and surgically subtract it from the final prediction without ever modifying the underlying model weights.

Why It Works: Mitigating the Trade-off

The ultimate power of path-specific causal fairness lies in its ability to optimize the fairness-accuracy trade-off. Traditional fairness metrics often cause severe drops in utility because they force group-level parity across the board, blinding the model to valid individual differences.

By shifting the definition of fairness from statistical outcomes to data-generating mechanisms, we gain a granular understanding of performance loss. As Drago Plecko et al. demonstrate in Fairness-Accuracy Trade-Offs: A Causal Perspective, the accuracy cost of enforcing fairness is not uniform across all pathways. Blocking direct effects ($A \to Y$) typically incurs a much lower utility loss than blocking indirect paths through valid mediators.

By targeting only the truly objectionable mechanisms of bias, path-specific causal fairness allows system designers to eliminate systemic discrimination while preserving the predictive power of legitimate, domain-relevant information.

Key Foundational Work

A causal framework for discovering and removing direct and indirect discrimination (2016) — contributed 3 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 118 papers
  2. Version 1 · 2026-10-04 · 201 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_path_specific_causal_fairness_bd3ef72b,
  title = {path-specific causal fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/path-specific-causal-fairness/rdr_bd3ef72b354640939b97288e857a37b9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_path_specific_causal_fairness_bd3ef72b_v2,
  title = {path-specific causal fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/path-specific-causal-fairness/rdr_bd3ef72b354640939b97288e857a37b9/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake