Causal Fairness

Causal fairness is a framework for assessing and ensuring algorithmic fairness using causal models to define and measure discrimination along specific causal pathways. Unlike statistical fairness metrics, it distinguishes between direct, indirect, and spurious effects of sensitive attributes (e.g., race, gender). It answers counterfactual questions like, 'Would this individual have received a different outcome if their protected attribute were different, all else being equal?' This approach provides a principled method to audit and remove unfair causal influences from automated decision systems.

Causal models explicitly separate different pathways of influence from a sensitive attribute to an outcome.

• Direct Discrimination: The causal effect of the sensitive attribute on the outcome that does not pass through a mediator variable. This is often considered legally and ethically impermissible.
• Indirect Discrimination: The effect that flows from the sensitive attribute through a mediator (e.g., zip code influencing credit score). Assessing whether this is fair depends on the justifiability of the mediator.
• Spurious Association: A non-causal, statistical correlation caused by a confounder (a common cause of both the attribute and outcome). Causal fairness aims to isolate and remove direct and unjust indirect effects while preserving spurious associations that do not represent actual discrimination.

A leading formal definition of causal fairness. A predictor is counterfactually fair if, for any individual, the prediction is the same in the actual world and in a counterfactual world where the individual's protected attribute (e.g., race) was changed. Formally: P(Ŷ_{A←a}(U) = y | X=x, A=a) = P(Ŷ_{A←a'}(U) = y | X=x, A=a), where A is the attribute, U represents latent background variables, and the do-operator sets the attribute. This ensures decisions are based on causally relevant factors unrelated to the protected attribute, providing a strong individual-level guarantee.

The analysis is grounded in a Structural Causal Model (SCM) represented by a causal graph (a Directed Acyclic Graph).

• Nodes represent variables (sensitive attribute A, outcome Y, features X, mediators M, confounders C).
• Edges represent direct causal relationships.
• Paths from A to Y are analyzed to determine fairness.
  • Backdoor Paths: Non-causal paths opened by confounders. These are blocked by conditioning to isolate the true effect.
  • Front-door Paths: Paths through mediators. The do-calculus is used to compute effects along these paths. The graph makes assumptions explicit and enables the use of tools like the backdoor criterion and front-door criterion to identify which variables to adjust for to measure specific types of discrimination.

This family of metrics evaluates fairness from an interventional perspective (the 'do' level of the causal hierarchy). It measures the effect of intervening on the protected attribute.

• Effect of Treatment on the Treated (ETT): The average effect for those who actually have a specific attribute value.
• No Unresolved Discrimination: Requires that the protected attribute has no direct causal effect on the outcome. This is tested by checking if P(Y | do(A=a), X=x) is constant across a.
• Path-Specific Effects: Allows finer-grained analysis by quantifying the effect flowing through specific causal pathways (e.g., only through an admissible mediator like qualifications, but not through an inadmissible one like neighborhood). This enables nuanced policies that remove unfair influences while preserving legitimate ones.

Implementing causal fairness presents significant engineering and statistical challenges.

• Graph Specification: The correctness of the analysis depends on an accurately specified causal graph, which requires domain expertise.
• Unmeasured Confounding: Hidden common causes can bias estimates. Techniques like sensitivity analysis or the search for instrumental variables are used to bound possible bias.
• Estimation from Data: Once a causal quantity is identified (e.g., a path-specific effect), it must be estimated from finite data using methods like propensity score matching, inverse probability weighting, or structural equation modeling.
• Integration with ML: Methods are being developed to build fairness-aware algorithms that learn under causal constraints, such as models that enforce counterfactual fairness during training by leveraging inferred latent variables.

A Structural Causal Model (SCM) is the foundational mathematical framework for causal fairness. It represents causal relationships between variables (e.g., ZIP code, education, hiring decision) as a system of structural equations, typically visualized as a causal graph.

• Provides the formal language to define fairness (e.g., "no direct effect of gender on hiring").
• Enables the computation of counterfactual quantities (e.g., "What would this applicant's salary be if their gender were different?").
• Distinguishes between observational data (what we see) and interventional data (what happens when we act).

Counterfactual fairness is a strict, individual-level fairness criterion. An algorithm is counterfactually fair if, for any individual, its prediction is the same in the actual world and in a counterfactual world where that individual's protected attribute (e.g., race) was different, while all other circumstances remain the same.

• Asks: "Would the decision have been the same if this person were of a different race, all else being equal?"
• Requires modeling the underlying causal process to simulate these alternative worlds.
• Considered a "gold standard" but is often difficult to satisfy in practice due to data and modeling requirements.

Path-specific fairness decomposes the total effect of a sensitive attribute on an outcome into effects that travel along specific causal pathways in a graph. This allows for nuanced fairness policies.

• Direct Effect: The effect of gender on hiring that does not pass through a mediator like "years of experience."
• Indirect Effect: The effect of gender on hiring that does pass through a mediator (e.g., gender→education→hiring).
• Enables definitions like: "We allow the effect of gender through education (an indirect effect) but prohibit any direct discrimination."
• Requires specifying which pathways are considered fair or unfair.

Causal mediation analysis is the statistical technique used to implement path-specific fairness. It quantifies how much of a total effect (e.g., gender pay gap) operates through a specific intermediate variable, or mediator (e.g., job title, negotiation outcome).

• Total Effect: The overall disparity in outcomes.
• Natural Direct Effect (NDE): The portion of disparity not explained by the mediator.
• Natural Indirect Effect (NIE): The portion of disparity explained by the mediator.
• Tools include the mediation formula and methods based on the do-calculus to estimate these effects from observational data under assumptions.

Causal confounding is the primary obstacle to measuring true discrimination. It occurs when a common cause influences both a protected attribute (e.g., race) and the outcome (e.g., loan denial), creating a spurious, non-causal association.

• Example: Neighborhood (confounder) influences both racial composition and average credit score.
• A naive model might incorrectly attribute the effect of neighborhood to race.
• The backdoor criterion is used to identify a set of variables to adjust for (e.g., income, location) to block these backdoor paths and isolate the true causal effect.
• Unmeasured confounding remains a fundamental limitation.

This contrast highlights the shift from simplistic to causal approaches.

• Fairness Through Unawareness: The naive practice of simply removing a protected attribute (e.g., 'gender') from model inputs. It is ineffective because proxies (e.g., 'major,' 'hobbies') can reconstruct the attribute, leading to indirect discrimination.

• Fairness Through Causal Awareness: The causal approach. It explicitly models the relationship between the protected attribute, its proxies, other covariates, and the outcome. It uses the causal model to define what constitutes unfair discrimination (e.g., direct effects) and then debiases the model or its predictions to satisfy that definition, often by simulating interventions.