Causal Inference with Ancestral Graph Representations

205 concepts134 papers1 proposal

Causal discovery aims to map the true mechanisms of a system from observational data. However, the real world rarely presents us with a closed system where every relevant variable is measured. Instead, we are plagued by latent confounders—unobserved variables that simultaneously influence multiple observed factors, creating spurious correlations that mimic direct causal links.

Traditional causal discovery often relies on the assumption of "causal sufficiency," which posits that no unobserved common causes exist. When this assumption is violated, standard Directed Acyclic Graphs (DAGs) break down because they lack the vocabulary to express that a correlation is driven by a shadow variable. The core insight of ancestral and mixed graph representations is that we do not need to explicitly model the hidden variables' exact states or functional forms. Instead, we can expand our graphical language to encode the presence and structural implications of latent confounding directly among the observed variables.

Expanding the Graphical Vocabulary: Mixed and Ancestral Graphs

To reason about latent confounding without explicitly drawing every hidden variable, researchers introduce alternative graphical frameworks. In a standard DAG, an edge $X \rightarrow Y$ implies direct causation. When a latent confounder exists between $X$ and $Y$, Jin Tian et al. in A General Identification Condition for Causal Effects demonstrated that we can represent this spurious dependency using a bi-directed arc ($X \leftrightarrow Y$). This allows for the analysis of hidden confounding without needing to specify the state space or functional form of the latent variable.

Graphs that mix directed and bi-directed edges are known as Acyclic Directed Mixed Graphs (ADMGs) or Maximal Ancestral Graphs (MAGs). As Sofia Triantafillou explains in Score-based vs Constraint-based Causal Learning in the Presence of Confounders, MAGs capture violations of causal sufficiency by using these specialized edge types to represent unobserved common causes without altering the graph's core independence properties.

Why must these graphs be "ancestral"? In standard DAGs, acyclicity simply means you cannot follow directed arrows in a loop back to your starting point. In mixed graphs, however, a new structural trap emerges: "almost directed cycles," where a variable becomes its own ancestor through a combination of directed and bi-directed edges. Rui Chen demonstrates in Integer Programming for Causal Structure Learning in the Presence of Latent Variables that enforcing ancestrality constraints is mathematically necessary to prevent these invalid structures, ensuring that the graph accurately reflects a physically realizable causal ordering.

Encoding Uncertainty: Equivalence Classes and Partial Ancestral Graphs

Even with an expanded vocabulary, observational data alone rarely points to a single, unique causal graph. Instead, the data constraints define a Markov equivalence class—a set of multiple graphs that all entail the exact same conditional independence relationships.

To represent this entire class in a single object, researchers use Partial Ancestral Graphs (PAGs). As Juan Miguel Ogarrio et al. show in A Hybrid Causal Search Algorithm for Latent Variable Models, PAGs introduce specialized edge marks, such as circles ($\circ$), to explicitly denote structural ambiguity. For instance, an edge mark might indicate that a link is definitely causal, potentially confounded, or completely unknown in direction.

The power of this representation lies in its closure under marginalization. Daniel Malinsky et al. in Causal Structure Learning from Multivariate Time Series in Settings with Unmeasured Confounding highlight that PAGs allow us to isolate and reason about the relationships among observed variables while explicitly marking which paths are vulnerable to hidden confounding. This prevents the algorithm from inferring false direct causal links from purely associative data.

Why It Works: Causal Inference Over Equivalence Classes

Building these representations is not just an exercise in graph theory; it provides the mathematical foundation required to calculate actual causal effects in the real world.

First, they allow us to determine which variables to control for. Benito van der Zander et al. in Constructing Separators and Adjustment Sets in Ancestral Graphs proved that valid covariate adjustment sets can be efficiently identified directly from MAGs. This means an investigator can block confounding pathways and achieve an unbiased estimate of a causal effect even when parts of the system remain unobserved.

Second, they enable complete causal identification when the exact graph is unknown. Amin Jaber et al. demonstrate in Causal Identification under Markov Equivalence: Completeness Results that causal identification can be performed directly over a PAG. If a causal effect is invariant across every single candidate model within the equivalence class represented by the PAG, the algorithm is guaranteed to return a valid causal expression.

Finally, this structural framework enables robust statistical estimation. By operating on the equivalence class rather than guessing a single DAG, the authors of Estimating Identifiable Causal Effects on Markov Equivalence Class through Double Machine Learning showed how to pair PAGs with Double Machine Learning. This combination allows researchers to estimate causal effects using flexible, non-parametric machine learning models while remaining fully protected against the biases introduced by unmeasured confounding.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 142 papers
  2. Version 1 · 2026-10-04 · 169 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_latent_confounding_representation_dd52f1bc,
  title = {latent confounding representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/latent-confounding-representation/rdr_dd52f1bc9a0846aba22c25e6c47d7b8d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_latent_confounding_representation_dd52f1bc_v2,
  title = {latent confounding representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/latent-confounding-representation/rdr_dd52f1bc9a0846aba22c25e6c47d7b8d/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake