Detecting and quantifying causal associations in large nonlinear time series datasets.

Area of Science:

• Complex dynamical systems
• Causal inference
• Time series analysis

Background:

• Identifying causal relationships in observational time series data is crucial for understanding complex systems.
• Challenges include high dimensionality, nonlinearity, and limited sample sizes in real-world datasets.
• Existing data-driven causal inference methods struggle with these complexities.

Purpose of the Study:

• To develop a novel method for estimating causal networks from large-scale time series data.
• To flexibly combine conditional independence tests with causal discovery algorithms.
• To improve the accuracy and power of causal discovery in complex dynamical systems.

Main Methods:

• A novel causal discovery algorithm integrating linear or nonlinear conditional independence tests.
• Application to time series data from the Earth system and human body.
• Validation using well-understood physical mechanisms and large-scale synthetic datasets.

Main Results:

• The proposed method demonstrates superior detection power compared to state-of-the-art techniques.
• Successful validation on both real-world (climate, cardiac) and synthetic time series data.
• Effective estimation of causal networks from high-dimensional, nonlinear time series with limited samples.

Conclusions:

• The novel method offers enhanced capabilities for discovering and quantifying causal networks from time series data.
• This advancement has broad applicability across various research fields studying complex systems.
• Opens new possibilities for data-driven causal inference in Earth science, biomedical research, and beyond.