Beyond Markov: Accounting for independence violations in causal reasoning

Area of Science:

• Cognitive Science
• Psychology
• Artificial Intelligence

Background:

• Causal graphical models are central to theories of causal cognition.
• Human reasoners frequently violate the Markov condition's independence relations in these models.

Purpose of the Study:

• To present three new accounts for observed independence violations in human causal cognition.
• To investigate how people reason with more complex causal graphs than previously studied.

Main Methods:

• Experimental assessment of human reasoning with extended common cause and common effect causal networks.
• Comparison of subject inferences against normative causal graphical model predictions.

Main Results:

• Subject inferences were most consistent with the beta-Q model, where consistent states are deemed more probable.
• Significant individual variability was observed, with minorities fitting dual prototype or leaky gate models.

Conclusions:

• Discrepancies arise from differing causal representations, not flawed reasoning.
• These findings are foundational for cognitive theories relying on causal graphical models, applicable to analogy, learning, and decision-making.