Counterfactual

A counterfactual query is a specific type of causal question that asks about an outcome under a hypothetical, altered scenario. It is formally expressed as: What would have been the value of Y if X had been set to x', given that we observed X=x and Y=y?

• This contrasts with an intervention (a 'do-query'), which asks about the effect of a future action.
• Example: "The patient took the drug (X=1) and recovered (Y=1). Would they still have recovered if they had not taken the drug (X=0)?"

Answering counterfactual questions requires a Structural Causal Model (SCM), not just statistical data. An SCM consists of:

• Structural Equations: Functions that assign a value to each variable based on its direct causes and an independent noise term (e.g., Y := f(X, U)).
• A Causal Graph: A visual DAG representing the equations.
• A Probability Distribution over the Noise terms (U).

The model allows for "surgery"—replacing an equation while keeping the background noise fixed—to simulate the hypothetical world.

Computing a counterfactual follows a deterministic, three-step algorithm (Pearl's Abduction-Action-Prediction):

1. Abduction: Use the observed evidence (e.g., X=x, Y=y) to infer the probable state of the unobserved background variables (U). This updates the prior distribution P(U) to a posterior P(U | evidence).
2. Action: Perform a minimal intervention on the model. Modify the structural equation for the cause variable to reflect the counterfactual assumption (e.g., set X = x'), while keeping all other equations and the inferred U constant.
3. Prediction: Use the modified model to compute the new value of the outcome variable Y. The resulting probability distribution is the counterfactual answer.

Counterfactuals enable the precise definition of causal concepts like probability of necessity (PN) and probability of sufficiency (PS), which are crucial for legal, diagnostic, and attribution tasks.

• Probability of Necessity: Given that an event occurred (Y=1) after a cause (X=1), what is the probability the event would not have occurred if the cause had been absent? This addresses blame: "Would the accident have happened without the negligence?"
• Probability of Sufficiency: Given that an event did not occur (Y=0) in the absence of a cause (X=0), what is the probability it would have occurred if the cause had been present?

It is critical to distinguish counterfactuals from related causal concepts:

• vs. Prediction (Association): A predictor uses P(Y | X=x). A counterfactual uses P(Y_{X=x'} | X=x, Y=y). The latter is conditioned on specific observed facts.
• vs. Intervention (Do-Calculus): An intervention query, P(Y | do(X=x')), asks for the population-level effect of a policy change. A counterfactual is personalized, asking about a specific unit (e.g., a particular patient) based on their observed history.
• Level on the Ladder: This places counterfactuals at Level 3 (Imagining), above Interventions (Level 2, Doing) and Associations (Level 1, Seeing).

Counterfactual reasoning is not philosophical; it's a practical engineering tool for building robust, explainable AI:

• Algorithmic Recourse & Explainability: Generating actionable advice for individuals (e.g., "To get your loan approved, increase your income by $5k").
• Causal Fairness Analysis: Determining if a decision was biased by asking, "Would the outcome have been different if the individual's protected attribute (e.g., gender) were changed?"
• Robust Decision-Making: Agents use counterfactuals to evaluate past actions ("What if I had taken a different path?") to improve future policies, a form of model-based reinforcement learning.
• Diagnostic Systems: In root-cause analysis, asking, "Which component's failure was necessary for this system outage?"

A three-level framework that categorizes the types of questions a reasoning system can answer, with each level requiring more sophisticated models.

• Level 1: Association (Seeing). Observing and detecting patterns. Example: "What is the correlation between a drug and recovery?"
• Level 2: Intervention (Doing). Predicting the effect of an action. Example: "What happens to recovery if we administer the drug?"
• Level 3: Counterfactual (Imagining). Reasoning about what would have happened under different circumstances. Example: "Would this patient have recovered if they had not taken the drug?" Counterfactuals represent the highest rung on this ladder.

The formal mathematical framework that enables counterfactual reasoning. An SCM consists of:

• Structural Equations: A set of functions that assign values to each variable based on its direct causes and an independent noise term. For example, Y := f(X, U).
• Causal Graph: A directed acyclic graph (DAG) visualizing the dependencies between variables.
• The ability to compute counterfactuals by modifying equations (e.g., setting X=x), while keeping the background noise U fixed to its factual value, and propagating the changes through the model to answer 'what if' queries.

The act of forcibly setting a variable to a specific value, denoted by do(X=x). This is the engine for interventional (Level 2) reasoning and a prerequisite for counterfactuals.

• Key Difference from Conditioning: P(Y | do(X=x)) computes the effect of making X equal to x, while P(Y | X=x) computes association based on seeing X equal to x.
• Example: P(Recovery | do(Drug=1)) estimates the recovery rate if we give everyone the drug, eliminating confounding paths. Counterfactuals use this operator within a personalized SCM to simulate alternative pasts.

The property that a causal quantity, such as an Average Treatment Effect (ATE) or a counterfactual probability, can be uniquely computed from the available data and the assumed causal model.

• Fundamental Question: Can we estimate the answer to our 'what if' question from what we can observe?
• Relies on Assumptions: Identifiability often requires satisfying criteria like no unmeasured confounding (for interventions) or having a fully specified SCM (for counterfactuals). If a query is not identifiable, no statistical method can reliably answer it from the given data.

A method to decompose a total causal effect into direct and indirect (mediated) pathways. It inherently involves counterfactual comparisons.

• Natural Direct Effect (NDE): The effect of a treatment when the mediator is held at the value it would have taken without treatment.
• Natural Indirect Effect (NIE): The effect that operates through the mediator.
• Example: "Did the drug cause recovery directly (NDE), or primarily by reducing blood pressure (NIE)?" Answering this requires comparing the factual outcome to the counterfactual outcome where the treatment changes but the mediator is fixed to its no-treatment value.

A framework for assessing algorithmic bias using causal models, which relies on counterfactuals to define fair outcomes.

• Counterfactual Fairness: A decision is considered fair for an individual if it would have been the same in a counterfactual world where their protected attribute (e.g., race, gender) was different, holding all other relevant circumstances constant.
• Moves Beyond Correlation: This definition isolates the direct causal effect of the sensitive attribute, distinguishing it from spurious associations or legitimate correlations via other variables. It provides a rigorous, individual-level test for discrimination.