Identifiability of Non-Gaussian Structural VAR Models for Subsampled and Mixed Frequency Time Series

Causal inference in multivariate time series is confounded by subsampling in time between the true causal scale and the observed data sampling rate. In practice, this presents challenges for inferring causal interaction between time series due to differences in sampling rates across time series and generally low sampling rates due to technological limitations. To determine instantaneous and lagged effects between time series at the true causal scale, we take a model based approach based on structural vector autoregressive (SVAR) models. We show that when the underlying noise, or shocks, to the system are non-Gaussian, both the parameters of the true model and its causal structure are identifiable from subsampled and mixed frequency data. Our work builds on the recent work of Gong et al. [1], who established the identifiability of VAR models from subsampled time series with non-Gaussian noise, but with no instantaneous interactions. Here, we generalize their work to the SVAR case to handle instantaneous interactions and, additionally, the multifrequency setting. The resulting approach provides a complete picture of identifiability in non-Gaussian SVAR models under arbitrary mixed frequency subsampling.