Causality indices for bivariate time series data: A comparative review of performance

Area of Science:

• Causal inference
• Time series analysis
• Complex systems

Background:

• Inferring nonlinear, asymmetric causal links in multivariate longitudinal data is challenging.
• Existing causality definitions for time series data lack a unified approach.
• Applications span clinical medicine, biology, economics, and environmental science.

Purpose of the Study:

• Evaluate ten prominent causality indices for bivariate time series.
• Assess method performance across diverse simulated dynamic systems.
• Identify robust methods for real-world causal inference.

Main Methods:

• Simulated four distinct bivariate time series models with varying coupling.
• Computed pairwise correlations between ten causality indices.
• Tested method invariance to data transformations (missing data, noise, scaling).

Main Results:

• Strong agreement observed between most causality methods (median Pearson correlation 0.719).
• Methods showed varying robustness to data imperfections like missing data and noise.
• Transfer entropy and nonlinear Granger causality demonstrated superior performance and robustness.

Conclusions:

• Transfer entropy and nonlinear Granger causality are recommended for bivariate causal inference.
• These methods are robust to common real-world data issues, including 20% missing data.
• Open-access Python code is provided for reproducibility and application.