A Decision-based View of Causality

David Heckerman, Ross D. ShachterUAI 1994

TL;DR

This paper bridges the gap between decision analysis and causal modeling by defining causality in decision-analytic terms. It introduces "causal influence diagrams" as a graphical representation that captures causal relationships, enabling better prediction of action effects and facilitating the translation of insights between these two fields.

What Problem Does This Paper Address?

Traditional models of uncertainty, like belief networks, focus on associational relationships (e.g., "smoking and lung cancer are dependent"). While useful, this isn't enough for decision-making. To make rational choices, we need to predict the effects of interventions (actions). For instance, knowing smoking and lung cancer are associated doesn't tell us if starting to smoke will increase our cancer risk. Artificial Intelligence researchers have been developing causal models to address this, but decision analysts have largely avoided explicit notions of causality due to a lack of precise definitions. This paper aims to unify these two areas by providing a decision-theoretic definition of causality and showing how it can be represented graphically.

Methods

The core of the paper is the definition of causality based on the concept of a "fixed set." A fixed set for a variable $x$ with respect to a set of decisions $D$ is the set of variables that will take on the same state regardless of the choices made for $D$. Causality is then defined: a set of variables $C$ are causes for $x$ with respect to $D$ if $x$ is not in the fixed set for $D$ (meaning $D$ can affect $x$), but $x$ is in the fixed set for $D$ given $C$ (meaning $C$ "shields" $x$ from the effects of $D$).

This definition is then translated into graphical terms using influence diagrams. The authors introduce "causal influence diagrams," a special type of influence diagram where every node $x$ is in the fixed set with respect to its parents $Pa(x)$. They also discuss "Howard Canonical Form" (HCF), a transformation of influence diagrams that makes all chance nodes descendants of decisions deterministic, effectively isolating uncertainty in "causal mechanisms."

Datasets

Not discussed in the paper. The paper uses illustrative examples from decision problems (e.g., lifestyle choices, coin flips) rather than real-world datasets.

Key Insights and Ideas

  • Causality as a Decision-Analytic Concept: The paper argues that the practical utility of causality lies in predicting the effects of interventions. This leads to a definition rooted in what we can observe and what is fixed with respect to our decisions.
  • Fixed Sets are Key: The notion of a "fixed set" is central. If a variable is in the fixed set with respect to decisions $D$, it means our choices of $D$ won't change its state. This is a stronger condition than probabilistic independence.
  • Influence Diagrams Can Represent Causality (with caveats): While standard influence diagrams have limitations in representing causality, the authors propose "causal influence diagrams" and the transformation to "Howard Canonical Form" (HCF) to overcome these.
  • Causal Mechanisms: HCF introduces "causal mechanisms" as deterministic functions that encapsulate the uncertainty in causal relationships, making them observable (at least in principle) and amenable to analysis.
  • Subjectivity of Causality: The paper acknowledges that causal beliefs are subjective, and different individuals might construct different causal models (influence diagrams) for the same domain, both of which can be "correct" if consistent with their beliefs.

Key Results

  • Definition of Causal Dependence: A formal definition of causal dependence is provided in decision-analytic terms.
  • Causal Influence Diagrams: A new class of influence diagrams is introduced that can represent causal relationships more effectively.
  • Howard Canonical Form (HCF): A method for transforming influence diagrams into HCF is presented, which allows for the analysis of causal mechanisms and counterfactual reasoning.
  • Theorems on Representation: Several theorems establish conditions under which influence diagrams can represent causal relationships, particularly relating blocking in graphs to fixed set membership.

Limitations and Future Work

  • Time: The authors note that the notion of time is an important aspect of causality that they have "barely touched upon" and suggest it as a direction for future work, particularly for time-varying domains.
  • Assessment of Mechanisms: While HCF is powerful, creating the causal mechanism nodes can require significant additional probability assessments, which can be formidable in practice. The authors mention that reasonable independence assertions can help, and in some cases, no new assessments are needed.
  • Practical Construction: While the authors believe decision makers can construct causal influence diagrams, the paper focuses more on the theoretical framework than on practical tools for eliciting these causal beliefs.

Summary updated

BibTeX
@inproceedings{heckerman1994decision,
  author = {David Heckerman and Ross D. Shachter},
  title = {{A Decision-based View of Causality}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1994},
}
Figure