Is every DAG an ancestral graph?

Yes. DAGs are included as a special case using directed edges only.

Does a bidirected edge identify a particular hidden variable?

No. It can summarize latent structure without uniquely naming or reconstructing the hidden causes.

Does maximal mean that no edges are missing?

No. It means every missing adjacency corresponds to some separating set in the graph's independence model.

AI Glossary / Statistics and probability / Advanced

Ancestral graph

Definition

An ancestral graph is a mixed graph used to represent conditional independence relationships, allowing directed, bidirected and undirected edges under specific structural restrictions. Maximal ancestral graphs can represent independence information arising from directed acyclic graph models after some variables are marginalized out and others are conditioned on.

What it means

A directed acyclic graph, or DAG, may describe a system containing variables that are not observed. Simply deleting those variables and their edges can lose dependencies among the variables that remain. Ancestral graphs provide a richer representation for the resulting independence structure without explicitly retaining every hidden or selection variable.

The word ancestral refers to restrictions involving directed ancestry and arrowheads. It does not mean a family tree. A graph is a formal representation under assumptions, and drawing an edge does not by itself establish a causal effect from observational data.

How it works

An ancestral graph has no directed cycles. If an edge has an arrowhead at a vertex, that vertex cannot be a directed ancestor of the other endpoint. A vertex incident to an undirected edge cannot also have an incoming arrowhead. These restrictions distinguish ancestral graphs from arbitrary mixtures of edge types.

Read conditional independence using m-separation, which extends d-separation to mixed graphs. Along a path, noncolliders must remain outside the conditioning set for the path to stay open, while colliders must be ancestors of conditioned variables, including themselves. If no open path remains, the graph implies the corresponding conditional independence under its Markov property.

A maximal ancestral graph has the additional property that every nonadjacent pair can be separated by some conditioning set. Maximal does not mean complete: missing edges remain informative. In latent-variable interpretations, bidirected and undirected edges can summarize features associated with hidden causes and selection, but the reduced graph generally does not uniquely reconstruct the underlying full DAG.

Three practical examples

1. A causal-modeling researcher represents observed relationships while allowing for unmeasured variables instead of assuming that every relevant cause was recorded.

2. An analyst studies how conditioning on a selection process changes the independence structure among the remaining measured variables.

3. A graphical-model developer checks whether a proposed mixed graph satisfies ancestral restrictions and which conditional independencies it implies.

Think of it this way

Think of a simplified transport map that removes some intermediate stations but preserves important connections among the stations still shown. Different connection symbols retain information that would be lost by erasing intermediate stations without updating the map. The simplified map need not reveal the exact original route structure.

Worked example

Consider an original Gaussian example with mutually independent, mean-zero, variance-one variables U, eX, eY and eZ. Define X = U + eX, Y = U + eY and Z = 2Y + eZ. The full causal construction has U pointing to X and Y, and Y pointing to Z.

When U is unobserved, the observed structure can be represented as X bidirected Y directed Z. The bidirected edge summarizes the hidden common-source relationship in this construction; it is not two opposing causal arrows. The path from X to Z passes through Y as a noncollider, so conditioning on Y blocks it.

Check this numerically: Var(Y) = 2, Cov(X,Y) = 1, Cov(Y,Z) = 4 and Cov(X,Z) = 2. The conditional covariance is 2 - (1 times 4)/2 = 0. Joint Gaussianity makes this zero conditional covariance an independence statement: X and Z are conditionally independent given Y. Without conditioning, their covariance is 2. The example verifies one graph implication under its specified model, not a method for discovering the graph from covariance alone.

Illustrative case study

Fictional case study: Latentpath models customer activity and recognizes that an important background preference is unmeasured. Its analysts avoid forcing all observed associations into a fully observed DAG. They use a suitable ancestral-graph representation to discuss independence assumptions and possible hidden influences.

The team distinguishes relationships supported by domain knowledge from orientations that remain uncertain across compatible models. Before recommending interventions, it checks whether the intended causal effect is identifiable under the assumptions rather than reading every directed edge as a complete intervention recipe.

This fictional case illustrates more cautious representation of incomplete information. It does not establish that using an ancestral graph automatically resolves unmeasured confounding or identifies a unique causal system.

Common misconceptions and mistakes

A bidirected edge is not a feedback loop with two causal directions. Also, an ancestral graph is not any graph containing arrows: its restrictions exclude certain directed and mixed-edge configurations. A maximal ancestral graph is not a graph with every possible edge.

Conditional independence information may leave several graphs observationally equivalent. A partial ancestral graph represents features shared across an equivalence class and is a different object from one fully specified ancestral graph. Finally, absence of m-separation does not force dependence in every distribution satisfying the Markov property; additional faithfulness assumptions are often used when reasoning in the reverse direction.

Business use

Ancestral graphs help technical teams articulate limitations caused by missing variables and selection when modeling complex systems. They can support better questions about what observational evidence does and does not determine.

Document the graph semantics, measured variables and assumptions about latent and selection processes. Distinguish a chosen graph from an equivalence class returned by a discovery method. Evaluate the plausibility of conditional-independence assumptions and sensitivity to measurement error or distributional changes. For AI decision systems, a graph that represents dependencies is not enough to justify an intervention: causal identification and estimation require additional analysis tied to the proposed action.

Check your understanding

In X bidirected Y directed Z, is Y a collider on the X-to-Z path?

Answer: No. The edge from Y to Z has a tail at Y, so the two incident path edges do not both have arrowheads at Y. Conditioning on Y blocks this path.

Related terms

Directed acyclic graph
Latent variable
Selection bias
M-separation
maximal ancestral graph
partial ancestral graph
Markov equivalence and faithfulness
independence semantics and causal assumptions

Learn next

compare a full DAG containing a hidden common cause with its observed ancestral representation. List the implied marginal and conditional independencies, then identify which causal details cannot be recovered from that independence information alone.

Common cause

Sources

Primary sources: Richardson and Spirtes, Ancestral graph Markov models, 2002: https://www.cmu.edu/dietrich/philosophy/docs/spirtes/annals2002.pdf

Ali, Richardson and Spirtes, Markov equivalence for ancestral graphs: https://arxiv.org/abs/0908.3605

Previous: Anaphoricity detector Next: Anchor modeling