Counterfactual Probabilities: Computational Methods, Bounds and Applications

Alexander Balke, Judea PearlUAI 1994

TL;DR

This paper introduces a computational framework for evaluating counterfactual probabilities, which are crucial for causal reasoning in fields like fault diagnosis and legal liability. They present methods for computing these probabilities precisely when causal mechanisms are known and derive bounds when only observational data is available, demonstrating their utility in real-world applications.

What Problem Does This Paper Address?

The paper tackles the challenge of reasoning about "what if" scenarios – counterfactuals – which are fundamental to understanding causality. For instance, "If Oswald hadn't shot Kennedy, would he still be alive?" requires reasoning about an alternative world. While Bayesian networks are powerful for probabilistic reasoning, they are insufficient for counterfactuals because they lack a model of temporal persistence and how to "force" a proposition to be true in an alternative world. The authors propose a method that interprets the counterfactual antecedent as an external action, allowing for more robust causal inference. A key problem is that often we don't have complete knowledge of the underlying causal mechanisms, only observational data. In such cases, precise probabilities are impossible, and the paper aims to provide informative bounds on these counterfactual probabilities.

Methods

The core of the paper lies in the use of functional models to represent causal relationships. These models go beyond standard Bayesian networks by specifying deterministic functions that map causes to effects, incorporating "response-function variables" that capture exogenous influences. The authors then derive a general expression for counterfactual probabilities based on these functional models.

When only probabilistic specifications (i.e., observed conditional probabilities) are available, the problem becomes finding bounds. This is framed as an optimization problem: finding the minimum and maximum possible values of the counterfactual probability over all functional models consistent with the observed data. This optimization is performed using linear programming techniques. The paper also introduces a notation for counterfactual queries and a framework for handling latent variables and their dependencies.

Datasets

Not discussed in the paper. The paper focuses on theoretical methods and demonstrates their application using hypothetical scenarios and real-world case studies, rather than specific datasets.

Key Insights and Ideas

  • Counterfactuals as External Actions: Interpreting the counterfactual antecedent as an external action that forces a proposition to be true is a crucial step for formalizing counterfactual reasoning.
  • Functional Models are Key: Standard Bayesian networks are insufficient for counterfactuals; functional models, which specify deterministic relationships, are necessary. Response-function variables are introduced to capture the influence of unobserved factors.
  • Bounds as a Practical Solution: When precise counterfactual probabilities are unattainable due to incomplete knowledge, deriving informative bounds is a valuable alternative for decision-making.
  • Optimization for Bounds: The problem of finding bounds can be effectively solved by optimizing the counterfactual probability expression over the space of possible functional models consistent with observed data.
  • Applications Beyond Theory: The paper demonstrates the practical relevance of these methods in complex real-world scenarios like clinical trials with imperfect compliance and legal liability judgments.

Key Results

The paper provides a general procedure for evaluating counterfactual probabilities using functional models and derives techniques for computing bounds when only probabilistic specifications are available. They demonstrate these methods in two applications:

  1. Treatment Efficacy in Imperfect Compliance Studies: They derive bounds on the average causal effect of a treatment, showing how these bounds can be tighter than previous methods and provide compelling evidence even with non-compliance. For example, in a hypothetical clinical trial, they establish that PeptAid consumption decreases the chance of developing ulcers by at least 15%.
  2. Liability Judgment: They apply the framework to a hypothetical product-safety litigation case, showing how counterfactual analysis can provide strong evidence for or against a defendant. In the PeptAid case, they show that the plaintiff would have had at most a 7% chance of developing ulcers had the product not been distributed or consumed, supporting the plaintiff's claim.

Limitations and Future Work

  • Local Minima/Maxima: The optimization procedure for finding bounds can converge to local minima/maxima, potentially leading to false bounds. Ensuring global optima are found is crucial.
  • Complexity of Optimization: While linear programming is used when the objective function is linear, the general case involves non-linear optimization, which can be computationally challenging.
  • Symbolic Bounds: The paper suggests that in some cases, symbolic bounds can be derived, which would be a significant advancement for theoretical understanding and practical application.
  • Further Applications: The authors imply that the developed framework can be applied to a wider range of problems requiring causal inference and counterfactual reasoning.

Summary updated

BibTeX
@inproceedings{balke1994counterfactual,
  author = {Alexander Balke and Judea Pearl},
  title = {{Counterfactual Probabilities: Computational Methods, Bounds and Applications}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1994},
}
Figure