Counterfactual Fairness

Counterfactual fairness is fundamentally grounded in causal inference, specifically the Structural Causal Model (SCM) framework. An SCM represents the data-generating process using:

• Endogenous Variables (V): Observable variables, including the protected attribute (A), other features (X), and the outcome (Y).
• Exogenous Variables (U): Unobserved background variables that represent latent factors.
• Structural Equations (F): Functions that define how each endogenous variable is determined by its parents (other variables) and exogenous noise.

This model allows for the formal definition of counterfactual queries: "What would the prediction be for this individual if, possibly contrary to fact, their protected attribute A were set to value a'?"

A predictor Ŷ is considered counterfactually fair if, for any individual with observed attributes, the prediction remains identical under all counterfactual manipulations of the protected attribute. Formally:

P(Ŷ_{A←a}(U) = y | X = x, A = a) = P(Ŷ_{A←a'}(U) = y | X = x, A = a)

Where:

• Ŷ_{A←a}(U) is the prediction in the real world.
• Ŷ_{A←a'}(U) is the prediction in the counterfactual world where A is set to a different value a'.
• U represents the individual's unobserved background variables.

The criterion must hold for all possible values of the protected attribute (a, a') and all individuals. This ensures fairness is evaluated at the individual level, not just on average across groups.

A critical practical rule derived from the core definition is that a counterfactually fair predictor should only depend on variables that are non-descendants of the protected attribute A in the causal graph. If a variable is causally influenced by A (a descendant), using it may propagate bias.

Admissible variables are those that are not caused by A. For example, in a hiring model:

• Admissible: An applicant's innate skill level (assuming it is not shaped by discrimination).
• Inadmissible: Educational pedigree, if access to elite schools is causally influenced by race (A).

Using only admissible variables blocks all causal paths from A to the prediction Ŷ, preventing the model from using proxy variables that encode the protected attribute.

Counterfactual fairness addresses key limitations of group fairness metrics like demographic parity or equalized odds:

• Individual vs. Group Focus: Group fairness ensures statistical parity across populations but can justify unfairness to specific individuals (e.g., rejecting a qualified candidate from a high-achieving group to meet a quota). Counterfactual fairness is defined per individual.
• Causality vs. Correlation: Group metrics assess observed correlations. Counterfactual fairness requires modeling the underlying causal structure to determine if differences are justified by non-discriminatory factors.
• Handling of Proxy Variables: A model can satisfy demographic parity while being blatantly unfair if it uses a perfect proxy for the protected attribute. Counterfactual fairness explicitly forbids using variables causally downstream of A, neutralizing such proxies.

Implementing counterfactual fairness is non-trivial and relies on strong assumptions:

• Causal Graph Specification: Requires domain expertise to build a credible causal DAG (Directed Acyclic Graph) showing relationships between all variables. Incorrect graphs lead to incorrect fairness assessments.
• Identification of Exogenous Variables: The unobserved background variables (U) for each individual must be inferred or modeled, often requiring strong parametric assumptions.
• Computational Complexity: Performing counterfactual inference, especially with complex, high-dimensional data and non-linear models, is computationally intensive.
• No Unmeasured Confounding: The SCM assumes all relevant common causes of variables are observed and included. Violations (hidden confounding) can invalidate the analysis. These challenges make it a rigorous but often aspirational standard in production systems.

Consider a model predicting loan default (Y) using features like income, credit score (CS), and zip code (Z), with race (R) as the protected attribute.

Causal Assumptions (Graph):

• Race (R) → Zip Code (Z) (Due to historical residential segregation).
• Race (R) → Income (I) (Due to societal discrimination).
• Income (I) → Credit Score (CS).
• Income (I), Credit Score (CS) → Loan Default (Y).

Analysis:

• Inadmissible Variables: Zip code and income are descendants of race. Using them directly would violate counterfactual fairness.
• Fair Predictor: A counterfactually fair predictor must base its decision on an individual's exogenous background (U)—their innate financial responsibility—and possibly credit score only to the extent it is not caused by race. This requires explicitly modeling and removing the causal influence of R on CS and I.
• Counterfactual Query: For a specific individual, the fair prediction should not change when answering, "What would the prediction be if this person were of a different race, but all their non-discriminatory latent factors (U) remained the same?"

The core requirement for counterfactual fairness is a correct causal model (a Directed Acyclic Graph or DAG) that encodes the assumed relationships between protected attributes (A), other observed variables (X), and the outcome (Y).

• Key Challenge: There is rarely a single, universally agreed-upon causal graph for complex social phenomena. Different domain experts may propose conflicting structures.
• Consequence: Fairness assessments become graph-dependent. A model deemed fair under one causal assumption may be unfair under another, undermining the objectivity of the audit.
• Mitigation: Requires extensive domain expertise and sensitivity analysis to test conclusions across a set of plausible graphs.

A confounder is a variable that influences both the protected attribute and other features/outcomes. If such a confounder is not measured and included in the causal model, the counterfactual inference will be biased.

• Example: Socioeconomic status (SES) may influence both an individual's race (due to systemic factors) and their educational history (a feature X). If SES is unobserved, the estimated effect of race is confounded.
• Impact: This violates the ignorability or unconfoundedness assumption required for valid counterfactual inference. The resulting fairness assessment is unreliable, potentially masking real bias or indicating bias where none exists causally.
• This is often the most fundamental limitation in real-world applications where data collection is incomplete.

Performing counterfactual inference for every individual in a large dataset is computationally intensive, especially with complex, non-linear models like deep neural networks.

• Process: For each individual, the model must generate a prediction in the actual world and the counterfactual world where their protected attribute is changed.
• Bottlenecks: This requires sampling from the posterior distribution of latent variables or running the model through modified input pipelines thousands or millions of times.
• Practical Limit: This complexity often restricts the use of counterfactual fairness to small-scale audits or research settings, rather than as a runtime constraint for high-throughput production systems.

The conceptual act of "changing" a protected attribute like race or gender in isolation is philosophically and practically fraught. These attributes are often deeply entangled with an individual's lived experience and other features.

• The "What If" Problem: What does it mean to ask, "What would this person's credit score be if they were a different race, but everything else 'the same'?" Many other features (name, neighborhood, historical treatment) are direct consequences of race.
• Feature Manipulation: Simply flipping a race variable from 'A' to 'B' while holding all other X constant may create an implausible or contradictory data point, leading to nonsensical model queries.
• Solution Approach: The causal model must explicitly define which features are descendants of the protected attribute and should therefore also change in the counterfactual world, a non-trivial modeling decision.

Unlike predictive accuracy, there is no ground truth for counterfactual outcomes. We can never observe what would have happened to the same individual under different protected attributes.

• Audit Consequence: It is impossible to empirically verify that a model satisfying counterfactual fairness on observed data is truly fair in a metaphysical sense. The audit validates consistency with a model of the world, not the world itself.
• Reliance on Simulation: Validation often relies on synthetic data generated from known causal models or carefully constructed semi-synthetic benchmarks where counterfactuals are known by design.
• This shifts the burden of proof to the robustness of the causal assumptions, as the final fairness measure cannot be directly tested against reality.

Enforcing strict counterfactual fairness can constrain model capacity, potentially reducing its overall predictive accuracy or business utility.

• Mechanism: The fairness criterion removes the model's ability to use any information that is a causal descendant of the protected attribute, even if that information is statistically predictive of the outcome.
• Business Trade-off: A lender, for example, may be prohibited from using an individual's zip code (a descendant of historical racial segregation) even if it correlates with default risk. This can lead to less accurate risk assessments overall.
• Decision Point: Organizations must explicitly decide if the fairness guarantee is worth a potential decrease in aggregate performance—a value-laden policy choice, not just a technical one.

A protected attribute is a personal characteristic—such as race, gender, age, or religion—that is legally or ethically prohibited from being used as a basis for discriminatory treatment. In counterfactual fairness, this is the attribute whose value is hypothetically changed in the causal model to test for equitable outcomes.

• Key Role: Serves as the central variable in the counterfactual query: "What would the prediction be if this attribute were different?"
• Exclusion vs. Causal Modeling: Simply removing a protected attribute from training data is insufficient, as proxy variables (e.g., zip code for race) can still enable discrimination. Counterfactual fairness requires explicitly modeling its causal influence.

A causal graph (or causal DAG) is a visual and mathematical model representing the assumed cause-and-effect relationships between variables, including protected attributes, legitimate features, and the model's prediction. It is the foundational scaffold for counterfactual fairness analysis.

• Structural Requirement: Defines which variables are confounders, mediators, or descendants of the protected attribute.
• Audit Dependency: The fairness conclusion is only valid under the assumed causal graph. Incorrect graph specification (e.g., missing a confounder) invalidates the audit. Graph construction often requires domain expertise.

A proxy variable is a feature in the dataset that is statistically correlated with a protected attribute (e.g., occupation with gender, zip code with race). Even if the protected attribute is excluded, a model can use proxies to replicate discriminatory patterns, violating fairness goals.

• Core Challenge for Audits: A key step in preparing for a counterfactual fairness audit is identifying potential proxies in the data.
• Causal Handling: In a well-specified causal graph, proxies are typically modeled as descendants of the protected attribute. The counterfactual inference accounts for how changing the protected attribute would also change the proxy's value.

Individual fairness is the principle that "similar individuals should receive similar predictions." Counterfactual fairness is a specific, causal instantiation of this principle, where similarity is defined through a causal model.

• Contrast with Group Fairness: Unlike demographic parity or equalized odds, which assess statistical parity across groups, individual fairness focuses on consistency at the level of single data points.
• Causal Similarity: Two individuals are considered similar for counterfactual fairness if they share the same values for all non-descendant variables in the causal graph, differing only in the protected attribute.

Adversarial debiasing is an in-processing bias mitigation technique where a primary model is trained to make accurate predictions while an adversarial component tries to predict the protected attribute from the primary model's internal representations. This removes information about the protected attribute from the features used for the main task.

• Connection to Counterfactual Fairness: Both approaches aim to make predictions independent of the protected attribute. Adversarial debiasing achieves this through an optimization constraint, while counterfactual fairness provides a formal causal test for whether it has been achieved.
• Audit Use: A model debiased with adversarial training should, in theory, pass a counterfactual fairness audit if the causal assumptions are correct.

An Algorithmic Impact Assessment (AIA) is a structured, often policy-guided process to evaluate the potential risks, benefits, and societal impacts of deploying an automated decision system. It encompasses technical audits, stakeholder consultation, and documentation.

• Audit Framework: A counterfactual fairness analysis is a rigorous technical component that can be included within a broader AIA to address individual fairness and causal non-discrimination.
• Holistic Context: While counterfactual fairness provides a mathematical criterion, an AIA ensures the audit's assumptions (causal graph, variable definitions) are scrutinized and its findings are communicated alongside other ethical and operational risks.