Counterfactual Simulation

Definition

Counterfactual simulation is the computational task of drawing samples from a counterfactual distribution, usually in a fully specified structural causal model, after conditioning on evidence from the factual world and applying a hypothetical intervention in the counterfactual world.

Why It Matters

Many counterfactual questions are distributional rather than yes-or-no. Fairness audits, treatment-choice studies, and causal decision systems may need to estimate how outcomes would change under an intervention for a particular observed case. When evidence includes continuous variables, direct analytic conditioning can be intractable, so simulation becomes the practical route.

Formalism / Key Objects

• A recursive SCM M=(V,U,F,p(u)) with endogenous variables V, background variables U, structural functions F, and exogenous distribution p(u).
• Evidence C=c, intervention do(X=x), and target variables W.
• Counterfactual distribution:

$$p(W_{do(X=x)}=w\mid C=c).$$

• Pearl-style workflow: update p(U) by factual evidence, construct M_do(X=x), then simulate target variables using the updated background-variable sample.
• For continuous evidence, Karvanen2024 - Simulating Counterfactuals assumes dedicated error terms and u-monotonic structural equations so root finding can adjust background errors to satisfy C=c.
• Counterfactual Particle Filtering records the sequential Monte Carlo interpretation and convergence anchor.

Connections

• Structural Equation Models provide the equations and intervention semantics that make the counterfactual sample meaningful.
• Causal Decision Making uses counterfactual estimates when evaluating unchosen policies, treatments, or fairness interventions.
• Actual Causality also asks counterfactual questions, but usually about whether an event caused a particular outcome rather than estimating a full distribution.
• Explainable AI intersects when counterfactual simulation is used for fairness auditing; this is stronger than ordinary contrastive explanation because it depends on a causal model.

Common Confusions

• Counterfactual simulation is not the same as simulating an interventional distribution. The factual evidence updates the shared background variables before the intervention is applied.
• A predictive model alone is not enough; the method needs causal structure and assumptions about unobserved background variables.
• Counterfactual fairness can fail even if a model omits sensitive variables directly, because proxies or descendants can still transmit sensitive-variable effects.

Key Sources

• Karvanen2024 - Simulating Counterfactuals