Graphical Gaussian modelling of multivariate time series with latent variables

Michael EichlerAISTATS 2010

Abstract

This paper introduces dynamic maximal ancestral graphs (dMAGs), a novel framework for identifying causal relationships in multivariate time series when some variables are unobserved (latent). By representing time series as nodes in a mixed graph, dMAGs distinguish between true Granger causality and spurious correlations caused by hidden confounders. We prove that dMAG parameters are mathematically identifiable and that the representation is closed under marginalization, ensuring model consistency even when variables are omitted.

Introduction

In time series analysis, Granger causality is a standard tool for determining predictive relationships. However, it is highly sensitive to latent variables. If an unobserved factor influences two observed variables, it can create "spurious causality," making it appear that one variable causes another when both are actually driven by a hidden third party. While graphical models have addressed latent variables in static data, applying them to time series has historically been difficult due to complex estimation procedures and lack of closure under marginalization.

Dynamic Maximal Ancestral Graphs (dMAGs)

The author proposes dynamic maximal ancestral graphs (dMAGs) to provide a unified, interpretable representation of time series dependencies. A dMAG is a mixed graph that uses three types of edges:

  1. Directed ($\to$): Direct Granger-causal links.
  2. Dashed Directed ($\dashrightarrow$): Links where the direction is known but latent variables may be involved.
  3. Dashed Undirected ($---$): Contemporaneous correlation between error terms.
Figure 5: Representation of latent variables ($L_1, L_2$) and their marginalization into dynamic maximal ancestral graphs (dMAGs).

The dMAG parametric model describes a stationary Gaussian process where the current state is a linear combination of its past (VAR component) and a structured error term (MA component). This framework effectively maps dMAGs to a specific class of Graphical ARMA models, bridging modern causal modeling with classical econometrics.

Key Theoretical Results

A major contribution of this work is the proof of parameter identifiability. Using a variation of the Yule-Walker equations, the author demonstrates that the coefficients of the VAR and the covariance of the errors can be uniquely determined from the observed data's auto-covariance function.

Furthermore, dMAGs satisfy the Granger-Markov property, meaning the paths in the graph accurately encode the conditional independence relationships in the underlying data. The ancestral constraint—which forbids an edge from a variable to its ancestor—is crucial for maintaining this identifiability and preventing mathematical breakdown from feedback loops.

Synthetic Examples and Marginalization

The most powerful feature of dMAGs is their closure under marginalization. If a variable is hidden or omitted from a large system, the relationships between the remaining variables can still be perfectly represented by a dMAG.

Figure 1: Synthetic example showing the effect of omitting a variable (node 4) on the causal structure between observed variables (1, 2, 3).

The paper uses a four-dimensional VAR(2) process to illustrate this. As shown in Figure 1, omitting a single variable (node 4) can lead to incorrect causal conclusions if not handled correctly. The dMAG framework provides a clear visual and mathematical way to identify "Type I spurious causality," where a link appears in a subset of variables but disappears when the full system is considered.

Conclusion and Future Work

While the paper establishes a strong theoretical foundation for dMAGs, several areas remain for future research. Although parameters are identifiable, high-performance estimation algorithms for noisy real-world data are still needed; "iterative conditional fitting" is suggested as a potential approach. Additionally, while the framework supports model selection via criteria like AIC or BIC, empirical demonstrations are pending. Finally, extending these results to even broader classes of graphs, such as those with "bow-free" paths, remains an open challenge.

Summary updated

BibTeX
@article{eichler2010graphical,
  author = {Michael Eichler},
  title = {{Graphical Gaussian modelling of multivariate time series with latent variables}},
  journal = {AISTATS},
  year = {2010},
}
Figure