The fundamental challenge of causal discovery from observational data is encapsulated in the classic adage: correlation does not imply causation. If we observe a statistical dependency between two variables, $X$ and $Y$, there are three primary structural explanations: $X$ causes $Y$, $Y$ causes $X$, or an unobserved third variable $L$ simultaneously causes both.
This unobserved variable is a latent confounder. Most classical causal discovery algorithms operate under the assumption of causal sufficiency—the premise that all common causes of the observed variables are measured and included in the dataset. When this assumption fails, standard algorithms mistake the spurious correlations induced by latent confounders for direct causal links, resulting in dense, incorrect graphs.
To build reliable causal models in the real world, researchers have developed frameworks to detect, represent, and isolate the influence of hidden common causes without observing them directly.
The Geometry of Hidden Influence: MAGs and PAGs
When confounders are unobserved, we cannot uniquely determine the exact Directed Acyclic Graph (DAG) underlying the data. For instance, if a latent variable $L$ influences both $X$ and $Y$, the observed data merely shows that $X$ and $Y$ are dependent, mimicking a direct edge $X \to Y$ or $Y \to X$.
To resolve this ambiguity, researchers use generalized graph structures that explicitly represent latent confounding:
- Maximal Ancestral Graphs (MAGs): Instead of omitting hidden variables or guessing their locations, MAGs introduce alternative edge types. A bi-directed edge ($X \leftrightarrow Y$) explicitly denotes the presence of a latent confounder between $X$ and $Y$, while a directed edge ($X \to Y$) guarantees that $X$ is an ancestor of $Y$ without unmeasured confounding along that path. Sofia Triantafillou demonstrated in Score-based vs Constraint-based Causal Learning in the Presence of Confounders that score-based learning can be effectively adapted to search the space of MAGs, providing a robust alternative to constraint-based testing.
- Partial Ancestral Graphs (PAGs): Often, a single dataset cannot distinguish between several valid MAGs. A PAG represents an entire equivalence class of MAGs. It introduces a circle mark ($\circ$) when the endpoint of an edge is uncertain (e.g., $X \circ \to Y$ means the relationship could be directed or confounded). As shown in A Graphical Criterion for Effect Identification in Equivalence Classes of Causal Diagrams, grouping variables into "possibly confounded components" across a PAG allows researchers to identify valid causal effects even when the exact orientation of every edge remains unknown.
Exploiting Statistical Asymmetries
How can we infer the existence of a hidden variable we have never measured? The key lies in exploiting statistical asymmetries that appear when data deviates from a standard joint Gaussian distribution.
Non-Gaussian Noise and Independent Components
In a linear system where the error terms (noise) are non-Gaussian, the direction of causality becomes identifiable. If $X \to Y$, then regressing $Y$ on $X$ yields a residual that is completely independent of the predictor $X$. However, if we reverse the regression ($X$ on $Y$), the residual and the predictor will be statistically dependent.
When a latent confounder $L$ is introduced ($X \leftarrow L \to Y$), this independence breaks in both directions. By analyzing these patterns across subsets of variables, algorithms can detect when an association is driven entirely by an external source. For example, Tatsuya Tashiro et al. developed ParceLiNGAM in ParceLiNGAM: A causal ordering method robust against latent confounders, which leverages non-Gaussianity to correctly order observed variables even when they are corrupted by hidden common causes.
The Generalized Independent Noise (GIN) Condition
To scale this intuition to complex latent structures, researchers look for linear combinations of observed variables that isolate and cancel out the influence of latent variables.
[ Latent Confounder L ]
/ \
v v
[ Variable X ] [ Variable Y ]
If a set of variables $X$ is caused by a latent variable $L$, then $X$ will be dependent. However, if we can find a surrogate vector that purifies $X$ by removing the component co-varying with $L$, the remaining signal becomes independent of the target variable $Y$. This principle, known as the Generalized Independent Noise (GIN) condition, was formalized by Feng Xie et al. in Generalized Independent Noise Condition for Estimating Latent Variable Causal Graphs. GIN provides a mathematical condition to verify whether a specific latent variable blocks the information flow between observed clusters, enabling the deep structural identification of hidden graphs.
Higher-Order Cumulants
When dealing with non-Gaussian distributions, simple variance and covariance (second-order statistics) do not capture the full picture. Higher-order cumulants (such as skewness and kurtosis) retain information about the mixing process of hidden signals. As shown by Ruichu Cai et al. in Causal Discovery with Latent Confounders Based on Higher-Order Cumulants, these higher-order statistics allow for closed-form algebraic solutions to estimate the impact of latent variables directly, bypassing the need for computationally expensive iterative optimization.
Disentangling Time and Dynamics
Time-series data introduces a powerful physical constraint: a cause must precede its effect in time. However, temporal data is highly vulnerable to unobserved external interferences, such as shifting environmental factors or macroeconomic trends.
When analyzing time series, latent confounders manifest in two ways: delayed temporal confounding (a hidden factor at time $t-1$ affects an observed variable at time $t$) and contemporaneous confounding (a hidden factor affects multiple observed variables within the exact same time step).
Time t-1: [ Latent L_{t-1} ]
|
v
Time t: [ Observed X_t ] <---> [ Observed Y_t ]
^
|
[ Latent L_t ]
To untangle these forces, Raanan Y. Rohekar et al. proposed an iterative search ordering strategy in From Temporal to Contemporaneous Iterative Causal Discovery in the Presence of Latent Confounders. By identifying and removing long-range, delayed temporal links first, the algorithm reduces the statistical search space, making it significantly easier to isolate and identify simultaneous, contemporaneous latent confounders acting within a single time step.
Complementing this, Ruichu Cai et al. introduced a Temporal Latent Variable Structural Causal Model in Temporal latent variable structural causal model for causal discovery under external interferences. Their approach uses variational inference to explicitly separate the structural causal graph into a binary adjacency matrix (representing the network skeleton) and a continuous weight matrix, filtering out external, unobserved interferences to reveal the true underlying physical drivers.