Causal Discovery in High-Dimensional Point Process Networks with Hidden Nodes.

Area of Science:

• Causal inference
• Time series analysis
• Statistical modeling

Background:

• Technological advances provide near-continuous multivariate point process data, enabling new causal discovery opportunities.
• A significant challenge is the presence of unobserved (hidden) variables, which can lead to confounding and inaccurate causal inference.
• Existing methods often fail when faced with missing data, producing misleading results.

Purpose of the Study:

• To propose a novel deconfounding procedure for estimating high-dimensional point process networks.
• To address the challenge of unobserved variables in causal discovery from partially observed data.
• To enable flexible modeling of connections between observed and unobserved processes, even when the number of hidden processes is unknown.

Main Methods:

• Developed a deconfounding procedure for high-dimensional point process networks.
• Incorporated flexible modeling of connections between observed and unobserved processes.
• Allowed for an unknown number of unobserved processes, potentially exceeding the number of observed nodes.

Main Results:

• The proposed method effectively estimates causal interactions in the presence of unobserved confounding.
• Theoretical analyses confirm the method's validity and advantages.
• Numerical studies demonstrate the method's superior performance in identifying causal relationships.

Conclusions:

• The deconfounding procedure offers a robust solution for causal discovery in partially observed multivariate point process data.
• This method overcomes limitations of traditional approaches by accounting for hidden variables.
• It facilitates more accurate identification of causal networks in complex systems with missing data.