A comparison of methods to elicit causal structure.

Area of Science:

• Cognitive Science
• Causal Inference
• Human-Computer Interaction

Background:

• Understanding causal structure is crucial for reasoning and decision-making.
• Eliciting causal knowledge from humans presents significant challenges.
• Existing methods for causal graph elicitation vary in their theoretical underpinnings and practical application.

Purpose of the Study:

• To compare the effectiveness of two distinct methods for eliciting causal graphs representing artifact structures.
• To evaluate which method better captures accurate causal relationships from human participants.

Main Methods:

• Developed and compared an 'Intervention' method based on interventional causality and counterfactuals.
• Utilized an online graph-drawing interface 'Loopy' for a global causal structure approach.
• Employed signal detection theory to analyze hit and false alarm rates for causal relations.

Main Results:

• The 'Intervention' method resulted in higher accuracy in generated causal models.
• Participants using the 'Intervention' method produced more correct causal relations.
• The 'Loopy' method, while allowing global consideration, was less precise in eliciting accurate local causal links.

Conclusions:

• The intervention-based approach, focusing on counterfactual reasoning, is superior for eliciting accurate causal graphs of artifacts.
• Methodological choices significantly impact the fidelity of human-generated causal models.
• Future research should explore hybrid approaches to leverage the strengths of both methods.