Psychological and Normative Theories of Causal Power and the Probabilities of Causes

Clark GlymourUAI 1998

TL;DR

This paper bridges human intuition about causation with formal causal models. It shows that people's judgments of "causal power" align with a specific type of probabilistic graphical model (Bayes nets) that can be generalized from simple "noisy OR/AND" gates to complex causal networks. This work offers a unified framework for understanding how humans infer cause and effect, and how this relates to established statistical and philosophical theories of causation.

What Problem Does This Paper Address?

How do humans intuitively assess the "causal power" or influence of one factor on another, especially when faced with incomplete information, unobserved causes, and complex causal relationships? Existing psychological theories, like Cheng's (1997), propose models for how people judge causal power, but the paper aims to generalize these models and connect them to formal graphical causal models widely used in statistics and computer science. Specifically, it seeks to:

  • Generalize Cheng's model: Extend her "noisy OR/AND" gate models to arbitrary directed acyclic graphs (DAGs) with both facilitating and preventing causes.
  • Connect to formal models: Show how these psychological models are specific parameterizations of Bayes nets and relate them to frameworks like Rubin's potential outcomes.
  • Address complex scenarios: Investigate how to estimate causal power in the presence of unobserved causes, indirect influences, and multiple pathways.
  • Distinguish causal influence: Differentiate between direct and total causal influence and provide conditions for their estimation.

Methods

The paper's core method is to demonstrate that psychological models of causal judgment, particularly Cheng's theory, can be understood as specific instances of Bayes nets. The authors achieve this by:

  • Formalizing Cheng's models: Representing Cheng's "noisy OR" and "noisy AND" gates as Boolean equations and then translating these into probabilistic relationships within a DAG framework.
  • Generalizing to DAGs: Extending the "noisy OR/AND" gate concept to arbitrary directed acyclic graphs, allowing for more complex causal structures with multiple variables and pathways.
  • Intervention-based reasoning: Utilizing the concept of interventions (manipulating variables to fixed values) to derive probabilities and estimate causal parameters, mirroring how formal causal inference is often performed.
  • Mathematical Derivations: Using probability theory and algebraic manipulation to derive conditions under which causal parameters can be estimated from observational data, even in the presence of unobserved confounders.

Datasets

Not discussed in the paper. The paper focuses on theoretical derivations and conceptual connections, using hypothetical examples and data structures rather than empirical datasets.

Key Insights and Ideas

  • Cheng's models as Bayes Nets: The most significant insight is that Cheng's psychological theory of causal power is not an ad-hoc explanation but a specific parameterization of Bayes nets, particularly those involving "noisy OR" and "noisy AND" gates. This provides a formal, computational grounding for her theory.
  • Generalizability of Causal Judgment: The paper shows that the principles behind Cheng's models can be extended to complex causal networks, suggesting a unified psychological and normative theory of causal inference applicable to a wide range of scenarios.
  • Distinguishing Direct vs. Total Influence: The work highlights the importance of distinguishing between direct causal influence (along a single edge) and total causal influence (summing across all pathways). It provides conditions for estimating both, which is crucial for understanding complex causal chains.
  • Interventions as a Key Tool: The paper implicitly and explicitly uses the concept of interventions to understand causal power. This aligns with modern causal inference methodologies and suggests that human reasoning might implicitly leverage counterfactual thinking akin to interventions.
  • Relationship to Rubin's Framework: The paper draws connections between Cheng's models and the Rubin causal model (potential outcomes), suggesting that while they represent causation differently (properties vs. units), they can be mathematically related under certain assumptions.

Key Results

  • Cheng's model for facilitating causes is shown to be equivalent to a noisy OR gate in a Bayes net.
  • Cheng's model for preventing causes is shown to be equivalent to a noisy AND gate.
  • Sufficient conditions are provided for estimating causal parameters in arbitrary acyclic networks of noisy OR and noisy AND gates, even with unobserved causes.
  • The paper outlines how to distinguish and estimate direct causal influence from total causal influence.
  • Mathematical relationships are established between Cheng's models and the Rubin causal model under specific assumptions.

Limitations and Future Work

  • Assumptions of Independence: Many derivations rely on assumptions of independence between causal parameters (q-values) and between unobserved causes. The paper acknowledges that real-world scenarios might violate these.
  • Complexity of Real-World Judgments: While the paper formalizes Cheng's models, it notes that human judgments can be complex and may not always perfectly align with these formalisms, especially in highly intricate scenarios.
  • Experimental Validation: The paper primarily presents theoretical derivations. Further experimental work is needed to confirm whether humans consistently exhibit the predicted judgments in complex causal structures.
  • Interference and Interaction: The paper touches upon scenarios where preventing causes interfere with facilitating causes, suggesting this is an area for further investigation.
  • Practical Applications: The paper hints at applications in law and epidemiology (e.g., Population Attributable Fraction), suggesting future work could explore these practical implications more deeply.

Summary updated

BibTeX
@inproceedings{glymour1998psychological,
  author = {Clark Glymour},
  title = {{Psychological and Normative Theories of Causal Power and the Probabilities of Causes}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1998},
}
Figure