wiki

Simulating Counterfactuals

3 min read

Simulating Counterfactuals

Summary

This partial ingest is based on the extracted full PDF text. Karvanen, Tikka, and Vihola give a simulation algorithm for counterfactual distributions in known structural causal models when analytic conditioning is difficult, especially with mixed continuous and discrete evidence. The method updates background variables under evidence, applies a hypothetical intervention, and simulates the intervened model; the conditional-simulation step is cast as a particle filter, giving asymptotic validity under the stated assumptions.

Key Claims

  • Counterfactual distributions can be simulated in fully specified recursive SCMs even when continuous evidence makes analytic conditioning intractable.
  • For continuous conditioning variables, the method relies on dedicated error terms and u-monotonicity so a binary search can find the background-error value that makes the condition true.
  • The multi-condition simulation procedure can be interpreted as sequential Monte Carlo, so particle-filter results provide mean-square convergence and a central-limit theorem for bounded test functions.
  • Counterfactual fairness evaluation can be performed for opaque prediction models when the evaluator has an SCM and API access to model predictions.
  • The approach is useful for simulation studies and fairness auditing, but it depends on strong causal-model knowledge and can suffer particle degeneracy as conditioning grows.

Methods / Formalism

  • The source uses recursive Structural Equation Models M=(V,U,F,p(u)), with endogenous variables V, background variables U, structural functions F, and exogenous distribution p(u).
  • Counterfactual evaluation follows Pearl’s three-step pattern: update p(U=u) by evidence C=c, replace the structural functions for X to form M_do(X=x), then simulate W_do(X=x) using the updated background-variable distribution.
  • The target distribution is:
  • An ancestral-SCM theorem justifies pruning to ancestors of the target, intervention, and evidence variables before simulation.
  • Counterfactual Particle Filtering records the reusable algorithmic detail: continuous-condition root finding, discrete-condition resampling, multi-condition sequencing, counterfactual simulation, and the particle-filter convergence statement.

Evidence / Experiments

  • The authors test the simulation algorithm on benchmark SCMs where the counterfactual distribution is analytically available. Sample quality improves with larger n, while uniqueness can be low after repeated resampling.
  • In the credit-scoring fairness example, the evaluator only queries opaque prediction models. Sensitive variables are gender and ethnicity, and the outcome is default risk.
  • Three prediction models are compared: unrestricted predictors, predictors excluding sensitive variables directly, and predictors restricted to non-sensitive direct causes of default. Only the third model is evaluated as counterfactually fair under the paper’s criterion.

Connections

  • Extends Causal Decision Making by making counterfactual policy or fairness queries operational when an SCM is known but analytic counterfactual calculation is hard.
  • Strengthens Structural Equation Models with a concrete simulation workflow beyond the usual intervention-replacement schema.
  • Connects to Actual Causality through the shared need to evaluate what would have happened under altered structural conditions, though this paper targets distributions rather than single-event cause judgments.
  • Relevant to Explainable AI because the fairness use case distinguishes causal counterfactual analysis from merely contrastive model explanations.

Open Questions

  • How much SCM misspecification can the fairness conclusions tolerate before the counterfactual simulation becomes misleading?
  • Could sensitivity analysis over SCM parameters be turned into a wiki artifact for fairness and causal decision-making workflows?
  • Which existing safe-RL or alignment notes require counterfactual simulation rather than merely interventional estimates?

Citation

Karvanen, Juha, Santtu Tikka, and Matti Vihola. 2024. “Simulating Counterfactuals.” Journal of Artificial Intelligence Research 80: 835-857.

Graph View

Backlinks