D-Separation and Markov Equivalence in Layered DAGs

Why d-Separation Matters

A directed acyclic graph (DAG) represents a set of variables and directed edges, with no directed cycle. In a multi-omics setting, nodes may represent molecular measurements from different layers. A graph’s arrows specify a factorization of a joint distribution, but a particular set of observed conditional independencies does not necessarily identify every arrow direction. D-separation provides a precise way to read conditional-independence implications from a DAG; Markov equivalence describes when different DAGs make the same such implications.

The graph-based conclusions require care. If a distribution is Markov with respect to a DAG, every d-separation in that graph implies the corresponding conditional independence. The reverse implication—that all conditional independencies in the distribution are represented by d-separations—requires the additional assumption of faithfulness. Neither assumption, by itself, makes an edge a causal effect.

Paths, Colliders, and Active Paths

Definition: A path between two nodes is a sequence of adjacent nodes in the graph, followed without regard to arrow direction. At an interior node on a path, the node is a collider on that path if both path edges point into it (\(A \to C \leftarrow B\)); otherwise it is a noncollider. A node \(D\) is a descendant of \(C\) if there is a directed path from \(C\) to \(D\); \(C\) counts as its own descendant for the purpose of the rule below.

Let \(Z\) be a set of nodes on which we condition, excluding the two endpoints of the path. A path is active given \(Z\) if every noncollider on it is outside \(Z\), and every collider on it either belongs to \(Z\) or has a descendant in \(Z\). A path is blocked otherwise. The apparently opposite rules for colliders and noncolliders are essential: conditioning on a noncollider blocks a path, while conditioning on a collider, or on one of its descendants, can activate it.

Definition: Two disjoint node sets \(X\) and \(Y\) are d-separated by \(Z\) in a DAG if every path between any node in \(X\) and any node in \(Y\) is blocked given \(Z\). Write \(X \perp_d Y \mid Z\) for this graphical relation. If a distribution is Markov with respect to the DAG, then \(X \perp_d Y \mid Z\) implies \(X \perp Y \mid Z\) in that distribution.

The local Markov property says that, given a node’s parents, the node is independent of its non-descendants that are not its parents. A distribution satisfying the DAG Markov property obeys these local independencies, and hence the d-separation implications. For a positive joint density or mass function, the corresponding factorization is

p(x_1,\ldots,x_m)=\prod_{i=1}^{m}p(x_i\mid x_{\operatorname{Pa}(i)}),

where \(\operatorname{Pa}(i)\) is the set of parents of node \(i\). This factorization is a distributional assumption, not a consequence of drawing arrows. Faithfulness is stronger: it rules out conditional independencies that arise from parameter cancellations even though the graph contains an active path.

Worked Example: Blocking a Chain in a Hypothetical Gaussian Model

Suppose three standardized-layer measurements \(X\), \(M\), and \(Y\) have the hypothetical structural model \(X \sim N(0,1)\), \(M=0.8X+\varepsilon_M\), and \(Y=1.5M+\varepsilon_Y\). Let \(\varepsilon_M\sim N(0,0.36)\) and \(\varepsilon_Y\sim N(0,1)\), independently of each other and of \(X\). The DAG is \(X \to M \to Y\). The independent errors and equations give a distribution Markov to this DAG.

The path from \(X\) to \(Y\) has \(M\) as a noncollider. Conditioning on \(M\) therefore blocks the path, so the graph implies \(X \perp Y\mid M\). The numerical covariance check agrees. First, \(\operatorname{Var}(M)=0.8^2(1)+0.36=1\), \(\operatorname{Cov}(X,M)=0.8\), and \(\operatorname{Cov}(M,Y)=1.5\). Also, \(\operatorname{Cov}(X,Y)=1.5(0.8)=1.2\). For jointly Gaussian variables,

\operatorname{Cov}(X,Y\mid M) =\operatorname{Cov}(X,Y) -\frac{\operatorname{Cov}(X,M)\operatorname{Cov}(M,Y)} {\operatorname{Var}(M)} =1.2-\frac{(0.8)(1.5)}{1}=0.

In this Gaussian model, zero conditional covariance gives conditional independence. Thus the calculation is consistent with the graph. It does not establish that the same relation holds in arbitrary data or under a different model.

Why Conditioning on a Collider Can Open a Path

For the path \(A \to C \leftarrow B\), the two arrows meet head-to-head at \(C\), so \(C\) is a collider. With no conditioning on \(C\) or its descendants, this path is blocked. Conditioning on \(C\) activates it. This can create a conditional association between \(A\) and \(B\) even when they are marginally independent.

Worked Example: A Collider in a Hypothetical Gaussian Model

Let \(A\) and \(B\) be independent \(N(0,1)\) variables, and let \(C=A+B+\varepsilon\), where \(\varepsilon\sim N(0,1)\) is independent of both. The graph is \(A \to C \leftarrow B\), and the specified independent errors make \(A\) and \(B\) marginally independent.

Before conditioning, \(\operatorname{Cov}(A,B)=0\). But \(\operatorname{Var}(C)=1+1+1=3\), \(\operatorname{Cov}(A,C)=1\), and \(\operatorname{Cov}(B,C)=1\). The conditional covariance is

\operatorname{Cov}(A,B\mid C) =\operatorname{Cov}(A,B) -\frac{\operatorname{Cov}(A,C)\operatorname{Cov}(B,C)} {\operatorname{Var}(C)} =0-\frac{(1)(1)}{3} =-\frac{1}{3}.

Because the variables are jointly Gaussian, the nonzero conditional covariance means \(A\) and \(B\) are conditionally dependent given \(C\). The collider rule predicts this change: conditioning on \(C\) opens the path. The calculation does not mean that conditioning always produces a negative association; its sign and magnitude depend on the distribution.

A collider’s descendant matters too. If \(C \to D\), conditioning on \(D\) activates the path through \(C\), even if \(C\) itself is not conditioned on. Conversely, conditioning on a variable that is neither a collider nor a descendant does not activate that collider path.

Markov Equivalence: Same Independencies, Different DAGs

Definition: Two DAGs are Markov equivalent if they imply exactly the same d-separation relations among their nodes. The skeleton of a DAG is its undirected graph, obtained by replacing every directed edge with an undirected one. An unshielded collider is a triple \(A \to C \leftarrow B\) in which \(A\) and \(B\) are not adjacent.

The standard criterion is that two DAGs are Markov equivalent exactly when they have the same skeleton and the same unshielded colliders. The skeleton determines which pairs are adjacent; the unshielded colliders distinguish certain arrow arrangements that change which paths are active. Other orientations can differ without changing the d-separation relations.

For instance, the three-node chain \(A \to B \to C\) and fork \(A \leftarrow B \to C\) have the same skeleton and no unshielded collider. Both imply \(A \perp C\mid B\), and neither implies that \(A\) and \(C\) are marginally independent. In contrast, the collider \(A \to B \leftarrow C\) has the same skeleton but a different unshielded-collider set. It implies marginal independence of \(A\) and \(C\) in a Markov distribution, so it is not equivalent to the chain or fork.

Worked Example: Checking Equivalence With a Covariance Calculation

Consider a hypothetical Gaussian chain with \(A\sim N(0,1)\), \(B=A+e_B\), and \(C=B+e_C\), where \(e_B\) and \(e_C\) are independent \(N(0,1)\) errors, independent of \(A\). Its graph is \(A\to B\to C\). The variances are \(\operatorname{Var}(A)=1\), \(\operatorname{Var}(B)=2\), and \(\operatorname{Var}(C)=3\); the covariances are \(\operatorname{Cov}(A,B)=1\), \(\operatorname{Cov}(A,C)=1\), and \(\operatorname{Cov}(B,C)=2\).

The chain blocks the only path from \(A\) to \(C\) when conditioning on \(B\). Directly,

\operatorname{Cov}(A,C\mid B) =1-\frac{(1)(2)}{2}=0.

Joint Gaussianity makes this zero conditional covariance equivalent to \(A\perp C\mid B\). The same joint Gaussian distribution can be factored in the order \(B,A,C\) as \(p(B)p(A\mid B)p(C\mid B)\), giving the fork \(A\leftarrow B\to C\). This fork has the same skeleton and no unshielded collider, so it has the same d-separation relations as the chain. The calculation illustrates one shared independence; the skeleton-and-collider criterion establishes equivalence of all their d-separation relations.

Covered Edges and a Justification for Equivalence

A useful way to understand some equivalent orientations is a covered-edge reversal. It applies only to an edge whose endpoint parent sets have a particular relationship. For a directed edge \(A\to B\), it is covered precisely when \(B\)’s parent set is \(A\)’s parent set together with \(A\):

\operatorname{Pa}(B)=\operatorname{Pa}(A)\cup\{A\}.

Under this condition, reversing \(A\to B\) to \(B\to A\) produces another DAG and preserves Markov equivalence. To see the factorization reason, write \(U=\operatorname{Pa}(A)\). Before reversal, the factors involving the two nodes are \(p(A\mid U)p(B\mid A,U)\). Their product is the joint conditional distribution \(p(A,B\mid U)\). The chain rule also writes that same distribution as \(p(B\mid U)p(A\mid B,U)\), which is the pair of factors after reversal. All other parent sets and factors remain unchanged. Thus the joint factorization is preserved. The covered-edge condition also ensures the reversal creates neither a directed cycle nor a changed unshielded collider.

For example, in \(A\to B\to C\), \(\operatorname{Pa}(A)=\varnothing\) and \(\operatorname{Pa}(B)=\{A\}\), so \(A\to B\) is covered. Reversing it gives \(A\leftarrow B\to C\). This is one route between the chain and fork in the previous example. Covered reversals help explain equivalence; they are not a procedure for learning a graph from data.

The skeleton-and-unshielded-collider criterion is the general test: matching a few conditional independencies is not enough to establish Markov equivalence. In particular, an unshielded collider changes the independence implications because its path is blocked until the collider or a descendant is conditioned on.

Layer Constraints, Diagnostics, and Limits

In a layered DAG, nodes have layer labels and allowable arrow directions may be restricted by the scientific or measurement ordering. Equivalence must be interpreted relative to those constraints. Two DAGs can be Markov equivalent as unrestricted graphs, yet one may violate a prespecified layer order and therefore be inadmissible under that order. Conversely, if several orientations satisfy the allowed ordering and share a skeleton and unshielded colliders, observational conditional independencies alone may not distinguish them.

A practical check should examine every path relevant to a proposed independence, not just the shortest path or the direct edge. For each path, mark every interior node as collider or noncollider and apply the conditioning rule. One active path is enough to prevent d-separation. When comparing DAGs for equivalence, separately compare the skeleton and every unshielded triple; do not infer equivalence merely because a chosen conditioning set blocks a chosen path.

Common failure modes include:

Treating all conditioning as blocking. Conditioning on a collider or its descendant can open a path. Apply the collider rule separately from the noncollider rule.
Ignoring alternate paths. Blocking one route does not d-separate two nodes if another route remains active.
Equating d-separation with empirical independence without assumptions. The DAG Markov property gives the implication from d-separation to conditional independence; faithfulness is needed for the converse.
Reading an observationally compatible direction as a causal conclusion. Markov equivalence concerns conditional-independence implications. It does not, on its own, establish causal direction or rule out confounding.

D-separation describes implications of a specified graph, not a guarantee that the graph is correct. Unmeasured variables can create dependencies not represented among the observed nodes, and distributional cancellations can produce independencies despite active paths. In addition, practical conditional-independence tests can be unreliable in high-dimensional or small-sample settings. These issues limit what can be concluded from observed associations; they do not alter the graphical definitions.

Key takeaway: To assess a conditional independence, test whether every path is blocked under the collider and noncollider rules. To compare DAGs, check whether their skeletons and unshielded colliders match. Equivalent graphs share observational independence implications, even when some arrows differ.

Check Your Understanding

Answer each question using the path rules and equivalence criterion. Brief solutions follow.

In \(A\to B\to C\), is the path from \(A\) to \(C\) active given \(B\)?
In \(A\to C\leftarrow B\), is the path from \(A\) to \(B\) active with no conditioning? What if we condition on \(C\)?
Are \(A\to B\to C\) and \(A\to B\leftarrow C\) Markov equivalent? State the structural reason.
For \(A\to B\), with \(\operatorname{Pa}(A)=\{D\}\) and \(\operatorname{Pa}(B)=\{A,D\}\), is the edge covered?

Solutions. (1) No. \(B\) is a noncollider on the path and is conditioned on, so the path is blocked. (2) With no conditioning, the path is blocked because \(C\) is an unconditioned collider. Conditioning on \(C\) activates the path. (3) No. They have the same skeleton, but the second graph has the unshielded collider \(A\to B\leftarrow C\), whereas the first does not. (4) Yes. \(\operatorname{Pa}(B)=\{A,D\}=\operatorname{Pa}(A)\cup\{A\}\), which is exactly the covered-edge condition.