Leveraging Directed Causal Discovery to Detect Latent Common Causes in Cause-Effect Pairs

Area of Science:

• Causal Inference
• Machine Learning
• Statistical Modeling

Background:

• Causal discovery from observational data is crucial in science and medicine.
• Existing methods often fail to detect latent common causes, limiting their applicability.
• There is a need for algorithms that can identify both directed relationships and unobserved common causes.

Purpose of the Study:

• To develop a general heuristic for modifying existing causal discovery algorithms.
• To enable these algorithms to detect latent common causes in addition to directed causal relationships.
• To validate the enhanced algorithms on synthetic and real-world datasets.

Main Methods:

• A general heuristic was devised to adapt causal discovery algorithms.
• The heuristic was applied to Information Geometric Causal Inference (IGCI) and Kernel Conditional Deviance.
• Extensive testing was performed on synthetic data across various noise models and on real-world data.

Main Results:

• The modified algorithms successfully detected latent common causes in synthetic data under different noise regimes.
• Known common causes were identified in real-world datasets using the enhanced methods.
• The performance of the original algorithms in distinguishing directed causal relations was preserved.

Conclusions:

• The proposed heuristic effectively enhances causal discovery algorithms to identify latent common causes.
• The modified IGCI and Kernel Conditional Deviance algorithms offer improved causal inference capabilities.
• This approach advances the field of causal discovery by addressing the limitation of unobserved confounders.