Causal Identification in Linear Non-Gaussian Models with Latent Confounders

This direction is merged into another direction: non-Gaussian causal identifiability. Its last version stays readable and citable here.

Version 1 of 112 papers16 conceptsfrozen 2026-10-04

In classical statistics, the "Gaussian world" is a place of frustrating symmetry. If you observe two correlated variables, $X$ and $Y$, and they both follow a normal distribution, it is mathematically impossible to determine whether $X$ causes $Y$ or $Y$ causes $X$ based on data alone. To a Gaussian model, the two scenarios look identical.

This problem becomes even more difficult when we suspect latent confounders—unobserved variables that influence both $X$ and $Y$. In a Gaussian system, these hidden variables create "ghost" correlations that make it impossible to distinguish a true causal link from a shared external influence.

However, a growing body of research shows that if we step outside the Gaussian world, the symmetry breaks. By leveraging non-Gaussian noise, researchers can now identify causal structures and strengths even in the presence of complex, unobserved confounding.

The Symmetry of the Bell Curve

To understand why this works, we first have to understand why Gaussian models fail. A Gaussian distribution is entirely defined by its first two moments: the mean and the variance. When we look at the relationship between two Gaussian variables, we are essentially only looking at their covariance.

Because covariance is symmetric, the statistical "footprint" of $X \to Y$ is indistinguishable from $Y \to X$. This is known as Markov equivalence. In this paradigm, we can only identify a set of possible graphs (an equivalence class) rather than the one true causal direction.

Breaking Symmetry with Non-Gaussian Noise

The breakthrough in this field, often centered around the LiNGAM (Linear Non-Gaussian Acyclic Model) framework, is the realization that non-Gaussian distributions contain "higher-order" information. While Gaussian distributions are defined by two moments, non-Gaussian distributions have unique signatures in their third-order (skewness), fourth-order (kurtosis), and higher moments.

As Kun Zhang, Dominik Janzing, and Bernhard Schölkopf demonstrated in Causal Inference by Identification of Vector Autoregressive Processes with Hidden Components, non-Gaussian noise signals do not "mix" the same way Gaussian signals do. When a non-Gaussian cause influences an effect, the resulting distribution of the effect variable carries a specific statistical distortion that cannot be replicated in the reverse direction. This higher-order information acts as a "breadcrumb trail" that points back to the original source of the signal.

Identifying the Unseen

The most powerful application of this idea is in handling latent confounders—variables we know exist but cannot measure.

In a linear system, a latent confounder usually acts as a "third party" that injects correlated noise into two observed variables. If all variables are Gaussian, this injected noise is indistinguishable from the variables' own internal noise. But if the noise is non-Gaussian, the "shape" of the confounding becomes visible in the higher-order moments of the observed data.

Jalal Etesami showed in Parameter identification in linear non-Gaussian causal models under general confounding that this allows for the identification of causal effects even under arbitrary non-linear confounding. Similarly, Daniele Tramontano and Negar Kiyavash developed GRICA, an algorithm that uses these non-Gaussian signatures to calculate the exact strength of causal links that would be ambiguous in any other setting.

From Geometry to Time Series

This research direction has expanded into several specialized domains:

  • Algebraic Geometry: Researchers like Carlos Améndola and Elina Robeva have proven that the relationships between second and third-order moments in these models form specific geometric structures called "toric ideals." In Third-Order Moment Varieties of Linear Non-Gaussian Graphical Models, they show that these geometric properties provide a rigorous mathematical proof for why non-Gaussian models break Markov equivalence.
  • Time Series: In real-world data, we often suffer from "subsampling"—measuring a system slower than its natural frequency. Alex Tank and Ali Shojaie demonstrated that non-Gaussianity allows us to recover instantaneous causal relationships that appear simultaneous due to slow sensors but are actually sequential.
  • Hierarchical Discovery: Sometimes the causal structure isn't just between observed variables, but between layers of hidden variables. Feng Xie and Biwei Huang introduced LaHME to uncover these "latent-to-latent" hierarchies, proving that non-Gaussianity can map out the "dark matter" of a causal system.

Why It Works: The Intuition of Independence

The fundamental reason non-Gaussianity works is the Darboux-Skitovich theorem, which essentially states that if a linear combination of independent random variables is Gaussian, then each of those variables must have been Gaussian to begin with.

In causal inference, we assume that the "exogenous noise" (the random spark that starts a causal chain) is independent for each variable. When we observe the system, we are seeing mixtures of these independent sparks. If the sparks are non-Gaussian, the way they mix is unique to the direction of the flow. By analyzing the "non-Gaussianity" of the residuals, we are essentially looking for the only orientation of the graph where the underlying noise sources are truly independent. If we pick the wrong direction, the "noise" will appear suspiciously correlated—a tell-tale sign that we've got the causality backward.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_identification_in_linear_non_gaus_3b6089eb,
  title = {Causal Identification in Linear Non-Gaussian Models with Latent Confounders},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-identification-in-linear-non-gaussian-models-with-latent-confounders/rdr_3b6089eb2cfc472eb1d44b5cb30835db},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_identification_in_linear_non_gaus_3b6089eb_v1,
  title = {Causal Identification in Linear Non-Gaussian Models with Latent Confounders},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-identification-in-linear-non-gaussian-models-with-latent-confounders/rdr_3b6089eb2cfc472eb1d44b5cb30835db/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}