ParceLiNGAM: a causal ordering method robust against latent confounders.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Learning causal order from observational data is crucial but sensitive to model assumptions.
• Existing methods for linear nongaussian acyclic models (LiNGAM) can fail when assumptions like no latent confounders are violated.
• Latent confounders can distort the estimated causal relationships.

Purpose of the Study:

• To develop a novel algorithm for learning causal orders that is robust to the presence of latent confounders in LiNGAM.
• To improve the reliability of causal discovery in the presence of unobserved common causes.

Main Methods:

• Propose a new algorithm for causal order learning in LiNGAM.
• Detect latent confounders by testing independence between estimated external influences.
• Identify variable subsets (parcels) unaffected by latent confounders to enhance robustness.

Main Results:

• The proposed algorithm demonstrates robustness against latent confounders, a common violation in LiNGAM.
• Effectiveness validated using both artificial datasets and simulated brain imaging data.
• Successful detection of latent confounders and identification of unaffected variable subsets.

Conclusions:

• The new algorithm provides a more reliable approach to causal discovery in LiNGAM when latent confounders are present.
• This method enhances the accuracy of causal ordering by mitigating the impact of unobserved variables.
• The findings have implications for fields relying on causal inference from observational data, such as neuroscience.