SD2S: Multiscale Granger causality analysis based on serial decomposition state space models

Abstract

Granger causality, widely recognized for its ease of use, intuitive interpretability, and applicability to complex multivariate systems, facilitates the inference of causal connections between variables through observational data and elucidates their dynamic interactions. With the advancement of the significant data era, an increasing multiscale characteristic of data is evident, presenting dynamics across multiple temporal scales. Current research in this domain typically relies on vector autoregressive models and wavelet transformations, which are susceptible to noise and dependent on substantial prior knowledge for selecting basis functions. To more accurately interpret and analyze Granger causality at multiple scales, this paper employs state space models and Empirical Mode Decomposition, introducing a novel multiscale Granger causality analysis approach based on serial decomposition state space models (SD2S). Experiments on simulated and real datasets confirm that (1) the integration of Empirical Mode Decomposition with state space models enhances the analysis of multiscale Granger causality; (2) serial decomposition state space models can improve the accuracy of multiscale analysis and effectively reduce computation time; (3) the proposed method successfully identifies dynamic causal relationships that vary with time scales in real-world data.