Reconstructing regime-dependent causal relationships from observational time series.

Area of Science:

• Causal inference
• Time series analysis
• Dynamical systems

Background:

• Inferring causal relations from observational time series data is crucial when experiments are not feasible.
• Dynamical systems often exhibit regime-dependent causal relations, where underlying unobserved factors alter causal structures.
• Existing causal discovery methods face challenges in identifying these regime shifts.

Purpose of the Study:

• To develop a method for detecting regime-dependent causal relations in observational time series data.
• To address the challenge of unobserved persistent regime variables influencing causal structures.
• To combine existing causal discovery techniques with regime learning for improved accuracy.

Main Methods:

• The study combines the PCMCI (based on PC and MCI tests) method with a regime learning optimization approach.
• PCMCI is utilized for causal discovery in high-dimensional and correlated time series.
• The proposed method, Regime-PCMCI, assumes a discrete, persistent regime variable.

Main Results:

• Regime-PCMCI successfully distinguishes between regimes with different causal directions, time lags, and link signs.
• The method also identifies changes in variable autocorrelation across regimes.
• Performance was validated on numerical experiments and real-world datasets, including El Niño Southern Oscillation and Indian rainfall.

Conclusions:

• Regime-PCMCI offers a robust approach to uncovering regime-dependent causal structures in time series data.
• The method demonstrates significant utility in both simulated and complex real-world environmental systems.
• This work advances causal discovery by accounting for unobserved environmental or system states.