Inferring causal relations from multivariate data using Large-Scale Augmented Granger Causality (lsAGC)

Area of Science:

• Neuroscience
• Climate Science
• Economics
• Complex Systems

Background:

• Causal inference from high-dimensional and short time-series data is vital for scientific discovery.
• Standard causal inference methods often fail under these challenging data constraints (T

Purpose of the Study:

• To introduce Large-scale Augmented Granger Causality (lsAGC), a novel method for causal inference in large-scale, high-dimensional, and short time-series data.
• To demonstrate the superior performance and efficiency of lsAGC compared to existing state-of-the-art methods.

Main Methods:

• lsAGC integrates dimension reduction, a Granger-based predictive framework, and data augmentation.
• The method was evaluated using extensive simulations on synthetic and semi-realistic fMRI data (linear and nonlinear).
• Validation was performed on real clinical fMRI data from 40 subjects (118 brain regions).

Main Results:

• lsAGC demonstrated high efficiency in handling high-dimensional data, confirmed by simulations.
• On real clinical fMRI data, lsAGC achieved an Area Under the Curve (AUC) of 0.83, significantly outperforming baselines (AUC 0.50-0.62).
• lsAGC maintained an AUROC above 0.70 on a 34-node network with only 50 samples, where other methods fell below 0.60.

Conclusions:

• lsAGC is computationally efficient (e.g., 8.3s for 118-region networks) and robust to noise, nonlinearities, and short time spans.
• The method's speed and accuracy make it practical for real-world applications in neuroscience, climate science, and economics.
• lsAGC addresses a critical gap in causal inference for prevalent short, large-scale time-series data.