Causal Effect Identifiability in the Presence of Latent Confounders Without Auxiliary Variables

Xiu-Chuan Li, James Kwok, Jiaxian Guo, Tongliang LiuICML 2026 · OpenReview

Introduction

In causal inference, the Randomized Controlled Trial (RCT) remains the gold standard. However, when RCTs are unfeasible, researchers must rely on observational data, which is frequently compromised by latent confounders—unobserved variables that influence both the treatment ($T$) and the outcome ($O$). These hidden factors create spurious correlations that masquerade as causal relationships.

Traditionally, researchers untangle these spurious correlations by leveraging auxiliary variables, such as instrumental variables, mediators, or proxies. But what happens when these auxiliary variables are entirely unavailable?

This paper investigates the irreducible statistical basis for causal identification: determining whether a causal effect can be uniquely identified using only the joint distribution of the treatment-outcome pair. Focusing on linear Structural Causal Models (SCMs), the authors prove that identifiability depends entirely on the distributional "shape"—specifically the Gaussianity—of the exogenous noise.

Methodology: Linear SCMs and Gaussian-Separability

The authors ground their theoretical framework in Linear SCMs, where variables are expressed as linear combinations of their direct causes plus an independent "exogenous noise" term (randomness unique to that specific variable).

To determine identifiability without auxiliary variables, the paper utilizes higher-order cumulants (statistical properties of a distribution beyond its mean and variance) and introduces the concept of Gaussian-separability. A distribution is considered Gaussian-separable if it can be decomposed into a non-Gaussian component and an independent Gaussian component.

By analyzing the contrast between the noise distribution of the treatment and the noise distribution of the latent confounders, the authors establish a comprehensive framework for identifiability.

Key Insights: The "Contrast" Rule

In linear systems, Gaussian noise tends to blend together, making it difficult to isolate individual effects. Non-Gaussian noise, however, leaves a distinguishable statistical footprint. The central insight of this paper is that identifiability relies on the contrast between the treatment's noise and the confounders' noise.

If the treatment exhibits non-Gaussian noise while the confounders are Gaussian, the non-Gaussian footprint allows researchers to "see" the true causal effect through the confounding noise.

The authors formalize this into four mutually exclusive and collectively exhaustive conditions:

Identifiable Conditions

  1. Condition 1: The treatment ($T$) is non-Gaussian, and the latent confounder ($L$) is Gaussian. The causal effect is uniquely identifiable.
  2. Condition 2: The treatment ($T$) is Gaussian, and the latent confounder ($L$) is not Gaussian-separable. The causal effect is uniquely identifiable.

Unidentifiable Conditions

When the causal effect is unidentifiable, the authors mathematically quantify the exact degree of ambiguity:

  1. Condition 3 (Finite Solutions): Both $T$ and at least one $L$ are non-Gaussian. The causal effect cannot be uniquely identified, but there is a precise, finite number of possible solutions (exactly $1 + \text{number of non-Gaussian confounders}$). This tightens previous theoretical bounds found in prior literature.
  2. Condition 4 (Infinite Solutions): $T$ is Gaussian and at least one $L$ is Gaussian-separable. In this scenario, there are infinitely many possible solutions for the causal effect.

Empirical Validation

To validate these theoretical proofs, the authors conducted experiments using synthetic datasets (SCMs $M_1$–$M_6$) simulating various confounding scenarios. They tested standard Gaussian noise against non-Gaussian distributions like Uniform, Exponential, and Laplace, scaling sample sizes from 500 to 500,000 observations.

The empirical results confirmed the theory: as data volume increases, the estimation error for the identifiable cases (Conditions 1 and 2) reliably drops toward zero, demonstrating that the mathematical conditions translate to real-world learnability.

Limitations and Future Work

While the paper establishes a rigorous foundation for causal identifiability without auxiliary variables, it notes several limitations:

  • Linearity Assumption: The primary proofs apply strictly to linear models. Although the authors provide a preliminary extension to non-linear models in the appendix, a comprehensive non-linear theory remains an open challenge.
  • Sample Complexity: While Condition 2 is theoretically identifiable, estimating it in practice is statistically challenging. The convergence rate is merely logarithmic, requiring exceptionally large datasets to achieve accuracy.
  • Testability: Proving whether a distribution is "truly" Gaussian or non-Gaussian from a finite dataset is notoriously difficult. In practice, researchers will likely need to rely on domain knowledge to justify which of the four conditions applies to their specific observational data.

Summary updated

Abstract

It is a fundamental challenge to ascertain whether the causal effect of a treatment on an outcome is identifiable in the presence of latent confounders, which serves as the logical prerequisite for recovering the causal effect in a partially observed system. While prior literature demonstrates that the causal effect is identifiable when there exist auxiliary variables subject to stringent structural constraints, this paper investigates identifiability of the causal effect without such variables. This means that we ground identifiability solely in the joint distribution of the treatment-outcome pair, which constitutes the irreducible statistical basis for causal effect identification. Focusing on linear structural causal models (SCMs), we provide a nuanced and complete characterization of identifiability of the causal effect contingent on the distributional properties of exogenous noises. Specifically, we formulate a set of mutually exclusive and collectively exhaustive conditions regarding the Gaussianity of exogenous noises, ascertain under which conditions the causal effect is identifiable and under which it is not, while also quantifying the cardinality of the feasible solution set for the unidentifiable cases. Finally, we empirically validate our theoretical findings.
BibTeX
@inproceedings{li2026causal,
  author = {Xiu-Chuan Li and James Kwok and Jiaxian Guo and Tongliang Liu},
  title = {{Causal Effect Identifiability in the Presence of Latent Confounders Without Auxiliary Variables}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=8ewf5I4shW},
}
Figure