Identifiability and estimation of structural vector autoregressive models for subsampled and mixed-frequency time...

Area of Science:

• Statistics
• Econometrics
• Climate Science

Background:

• Causal inference in multivariate time series is complicated by subsampling and mixed frequencies.
• Observed data may not capture the true speed of causal interactions.
• Existing methods struggle with non-Gaussian noise and varying sampling rates.

Purpose of the Study:

• To develop a unifying framework for parameter identifiability and estimation in multivariate time series under subsampling and mixed frequencies.
• To enable causal inference at the true causal scale, even with limited sampling rates.
• To handle non-Gaussian noise (shocks) in time series data.

Main Methods:

• Utilizing structural vector autoregressive (SVAR) models as a foundation.
• Developing identifiability and estimation methods for causal structures with lagged and instantaneous effects.
• Deriving an exact expectation-maximization (EM) algorithm for inference.

Main Results:

• A unified framework for parameter identifiability and estimation under subsampling and mixed frequencies.
• Identifiability and estimation methods for causal structures at the desired time scale.
• An exact EM algorithm for efficient inference in challenging settings.

Conclusions:

• The proposed model-based approach effectively addresses causal inference challenges in subsampled and mixed-frequency time series.
• The methods are validated on simulated data, climate data, and econometric data.
• This work advances the ability to uncover causal relationships in complex time series data.