Identifying nonlinear 1-step causal influences in presence of latent variables

Saber Salehkaleybar, Jalal Etesami, Negar Kiyavash

Introduction

In fields ranging from neuroscience to macroeconomics, the ability to discern whether one time series "causes" another is fundamental. The traditional gold standard for this task is Granger Causality: a variable $X$ is said to cause $Y$ if the past values of $X$ provide statistically significant information about the future of $Y$ beyond what is contained in the past of $Y$ and other observed variables. However, this approach faces a critical bottleneck: the presence of latent (unobserved) variables.

The Latent Variable Challenge

Standard causal analysis often fails when hidden factors influence the system. If a latent variable $Z$ influences both observed variables $X_1$ and $X_2$, a naive analysis might detect a spurious causal link between $X_1$ and $X_2$ simply because they share a common driver.

A causal graph illustrating how a latent variable $Z$ can create complex dependencies between observed variables $X_1$ and $X_2$.

Traditional methods typically require measuring every possible influencing factor in the "universe" to avoid these phantom influences—a requirement that is impossible to meet in real-world complex systems. This paper proposes a method to identify true 1-step functional dependencies among observed variables while effectively neutralizing the interference of hidden factors.

Proposed Methodology

The authors introduce a framework based on Conditional Mutual Information (CMI). The core of their approach rests on a specific system model where observed variables ($X$) and latent variables ($Z$) interact dynamically. A pivotal assumption is that while observed variables may be subject to exogenous noise, the latent variables evolve as a deterministic function of their previous states.

The "History" Trick

The most significant insight of this work is that if latent variables evolve deterministically, their state is eventually encoded within the history of the observed variables. By conditioning the analysis on a sufficiently long history of all observed data, researchers can effectively "proxy" for the hidden variables. The authors prove that $X_j$ has a 1-step causal influence on $X_i$ if and only if the mutual information between $X_i(t)$ and $X_j(t-1)$ is non-zero when conditioned on the entire history of the observed network.

For linear dynamics, this simplifies to a Vector Auto-Regressive (VAR) model of a higher order, where a Wald test can recover the causal graph. For nonlinear systems, the authors employ a K-nearest neighbor (KNN) estimator to calculate the necessary CMI values.

Experimental Validation

The theory was validated through numerical simulations of complex systems:

  1. Linear Consensus Protocols: A synthetic network consisting of 10 observed and 10 latent nodes, mimicking control systems where agents reach agreement through local interactions.
  2. Nonlinear Toy Models: A 3-variable system designed to test the information-theoretic criterion against complex, non-linear dependencies.
Average error in causal graph recovery for linear consensus protocols. As the lag length (history) increases, the error drops to zero, confirming that history accounts for latent variables.

In the linear simulations, the "Average Error" in identifying the causal graph dropped to zero as the lag length increased. This empirical result confirms that looking further into the past allows the model to account for hidden variables without observing them directly. In nonlinear tests, the CMI values successfully distinguished between true causal links and spurious correlations.

Conclusion and Future Work

This research demonstrates that it is possible to recover true causal structures without observing the "entire universe," provided the latent variables are noise-free. By leveraging the history of observed data, the method provides a robust way to identify 1-step influences.

However, the assumption of noise-free latents remains a significant limitation. If hidden variables are subject to their own random external shocks, the history of observations may no longer perfectly encode their state. Future research will likely focus on relaxing this assumption and reducing the computational complexity required for high-dimensional nonlinear estimation.

Summary updated

BibTeX
@inproceedings{salehkaleybar2017identifying,
  author = {Saber Salehkaleybar and Jalal Etesami and Negar Kiyavash},
  title = {{Identifying nonlinear 1-step causal influences in presence of latent variables}},
  booktitle = {ISIT},
  year = {2017},
  url = {https://openreview.net/forum?id=ZxnOY9eP4La},
}
Figure