Causal Discovery from Temporally Aggregated Time Series

Area of Science:

• Dynamical systems analysis
• Time series analysis
• Causal inference

Background:

• Discovering causal structure in dynamical systems from time series is crucial but challenging.
• Temporal aggregation of data (averaging or summing consecutive observations) further complicates causal discovery.
• Existing methods struggle with aggregated time series data.

Purpose of the Study:

• To investigate the recovery of causal relations at the original frequency from temporally aggregated data.
• To determine conditions under which causal structure is identifiable from aggregated time series.
• To develop and evaluate a method for causal discovery from aggregated data.

Main Methods:

• Assuming a vector autoregressive (VAR) model for the original time series.
• Leveraging independent and non-Gaussian noise assumptions.
• Employing non-Gaussian state-space modeling for estimation.
• Evaluating performance on synthetic and real-world datasets.

Main Results:

• Identifiability of the causal structure at the original frequency from aggregated time series is demonstrated under specific conditions.
• The proposed non-Gaussian state-space modeling approach is effective for causal discovery.
• The method's performance is validated through empirical evaluations.

Conclusions:

• Causal structure can be recovered from temporally aggregated time series data.
• Non-Gaussian state-space modeling provides a viable method for this challenging causal discovery task.
• The findings advance the ability to infer causality from aggregated observational data.