Probabilistic Evaluation of Counterfactual Queries

Alexander Balke, Judea PearlAAAI 1994

Introduction

In their seminal 1994 paper, Alexander Balke and Judea Pearl established a foundational framework for causal inference by addressing a core challenge in artificial intelligence: the formal evaluation of counterfactual queries. While standard logic and Bayesian Networks excel at handling "if-then" scenarios based on observations, they often falter when faced with "what-if" questions where the antecedent contradicts known facts. Balke and Pearl's work provides the mathematical machinery to bridge the gap between philosophical "closest world" semantics and computational probability.

The Challenge of Counterfactual Reasoning

Counterfactuals—statements like "If $A$ had been true, then $C$ would have occurred"—are essential for diagnosis, planning, and assigning responsibility. However, a counterfactual antecedent often conflicts with observed evidence. For instance, if we know Bob was not at a party, calculating the probability of a scuffle had he been there requires more than simple conditional probability.

Standard Bayesian Networks (BNs) are insufficient because they represent correlations rather than mechanisms. Two different functional models can yield identical BNs but produce divergent answers to counterfactual queries. To resolve this, the authors propose using Functional Causal Models, where each variable $x_i$ is a deterministic function of its parents and an unobserved disturbance term $\epsilon_i$: $x_i = f_i(pa_i, \epsilon_i)$.

The Three-Step Algorithm

The core contribution of the paper is a three-step algorithm for evaluating queries such as "What is the probability of $C$ if $A$ were true, given evidence $B$?"

  1. Abduction: The real-world evidence $B$ is used to update the probability distribution of the unobserved disturbance variables $\epsilon$. This step "captures" the specific background conditions of the actual world.
  2. Action (Intervention): The equation for variable $A$ is replaced with a fixed value representing the counterfactual antecedent. Graphically, this "prunes" the causal edges pointing into $A$, representing an external intervention that overrides its normal causes.
  3. Prediction: The updated $\epsilon$ distributions and the modified causal graph are used to calculate the probability of the consequent $C$ in this hypothetical world.
Graphical representation of a counterfactual query using a twin-network structure.

This process is often visualized through a "twin network," where the actual world and the counterfactual world share the same background disturbance terms but have different causal structures due to the intervention.

The Party Example

To illustrate the logic, the authors use a "Party Example" involving three friends: Ann, Bob, and Carl. The causal relationships are defined by the interpersonal dynamics: Ann's presence influences Bob and Carl, and their joint presence might lead to a scuffle.

Causal graph for the party example.

The paper demonstrates that an "indicative" query (e.g., the probability of a scuffle given we observe Bob at the party) can yield significantly different results than a "subjunctive" counterfactual query (e.g., the probability of a scuffle if Bob had been there, given he actually was not). In their quantitative walkthrough, the counterfactual probability (0.79) was nearly seven times higher than the indicative one (0.11), highlighting why "what if" reasoning is distinct from "what is" observation.

Conclusion

Balke and Pearl’s work transformed counterfactual reasoning from a philosophical abstraction into a computable engineering problem. By treating interventions as "miracles" that locally disrupt causal mechanisms without altering the past or background noise, they provided a rigorous path forward for AI systems to reason about alternative realities. This framework remains a cornerstone of modern structural causal modeling.

Summary updated

BibTeX
@inproceedings{balke1994probabilistic,
  author = {Alexander Balke and Judea Pearl},
  title = {{Probabilistic Evaluation of Counterfactual Queries}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure