Robust causality and false attribution in data-driven earth science discoveries.

Area of Science:

• Earth and environmental sciences
• Data science
• Causal inference

Background:

• Causal and attribution studies are vital for Earth science discoveries and policy-making.
• Current data-driven methods, including transfer entropy (TE), risk spurious causal links due to complexity and estimation inaccuracies.
• Physics-informed approaches are crucial to avoid conflating correlation with causation.

Purpose of the Study:

• To address the limitations of existing causal discovery methods in Earth sciences.
• To introduce a novel framework, CAST (Causal Analysis Spuriousness Test), for quantifying the robustness of inferred causal links.
• To enhance the reliability of data-driven causal discovery, particularly TE-based methods.

Main Methods:

• Developed CAST, a subsample-based ensemble framework.
• Quantified the robustness of inferred causal links using the CAST index.
• Conducted extensive simulations across diverse system dynamics and applied to real-world climate datasets.

Main Results:

• CAST effectively filters unreliable causal connections identified by data-driven methods.
• The framework successfully preserves true causal relationships.
• Demonstrated the efficacy of CAST through simulations and real-world climate data applications.

Conclusions:

• Emphasizes the necessity of consistency-based evaluation in causal discovery.
• CAST provides a generalizable strategy to improve the reliability of causal inference in Earth sciences.
• Highlights the importance of robust methods for informing climate, ecology, and water policies.