Causal Bayesian Network

A causal Bayesian network is a directed acyclic graph whose vertices represent random variables and whose directed edges encode direct causal relations. It combines the probabilistic factorization of a Bayesian network with the intervention semantics of a structural causal model. The resulting formalism distinguishes ordinary conditioning, which describes information about observed events, from intervention, which represents an external alteration of the data-generating process.

For variables (V={V_1,\ldots,V_n}) associated with a graph (G), the observational distribution satisfies the causal Markov factorization

[ P(v_1,\ldots,v_n)=\prod_{i=1}^{n}P(v_i\mid \operatorname{pa}_i), ]

where (\operatorname{pa}_i) denotes the realized values of the parents of (V_i) in (G). The same factorization also appears in noncausal Bayesian networks. Its causal interpretation therefore depends on additional assumptions concerning mechanisms, interventions, and the absence or explicit representation of relevant common causes.

Causal semantics

Each node in a causal Bayesian network corresponds to a mechanism that determines the node from its parents and an exogenous disturbance. A structural representation has the form

[ V_i=f_i(\operatorname{Pa}_i,U_i), ]

where (f_i) is the mechanism associated with (V_i), while (U_i) collects influences not represented as endogenous parents. The directed graph contains an edge (V_j\rightarrow V_i) when (V_j) is an argument of (f_i). An edge consequently represents a claim about the data-generating structure rather than merely a convenient statistical association.

The causal Markov condition states that each variable is independent of its nondescendants conditional on its direct causes. Graphical separation is expressed through d-separation, which converts structural features of the graph into conditional-independence relations. If (X) and (Y) are d-separated by (Z), every distribution Markov relative to the graph satisfies

[ X\mathbin{\perp!!!\perp}Y\mid Z. ]

The converse requires the faithfulness assumption. Faithfulness excludes exact cancellations that create conditional independences not implied by d-separation. It is distinct from the Markov condition and is not required for defining interventions, although it has a central role in many methods for causal discovery.

A graph containing only observed variables commonly assumes causal sufficiency, meaning that every common cause relevant to the represented variables is included. When this condition does not hold, latent variables can produce dependence patterns that an observed-variable directed acyclic graph does not represent causally. Such systems are described through latent-variable models or acyclic directed mixed graphs, in which bidirected edges summarize unobserved common causes.

Observation and intervention

Conditioning on (X=x) restricts an observational distribution to cases in which (X) has the value (x). An intervention written (\operatorname{do}(X=x)) instead replaces the mechanism determining (X) with a constant assignment. Incoming edges to (X) are removed because the intervened variable no longer responds to its ordinary causes.

For a causally sufficient network, the intervention distribution follows the truncated factorization formula

[ P(v\mid \operatorname{do}(X=x))

\begin{cases} \displaystyle\prod_{V_i\notin X}P(v_i\mid \operatorname{pa}_i), & v_X=x,\[6pt] 0, & v_X\neq x. \end{cases} ]

This distinction explains why (P(Y\mid X=x)) generally differs from (P(Y\mid \operatorname{do}(X=x))). Conditioning preserves the causes of (X), whereas intervention severs their influence on (X). The two quantities coincide only under graphical and distributional conditions that eliminate confounding and other relevant selection effects.

Interventions can be deterministic or stochastic. A deterministic intervention fixes a variable at a specified value, while a stochastic intervention replaces its mechanism with a newly specified probability distribution. More general policy interventions assign values according to previously observed information. These cases share the same modular interpretation: selected mechanisms change while mechanisms outside the intervention remain invariant.

Identification

A causal effect is identified when it is uniquely determined by the observational distribution together with the assumed causal graph. Identification is therefore separate from numerical estimation. A quantity can be identifiable even when available data are too limited for precise estimation, and it can remain unidentifiable even with an unlimited observational sample.

The back-door criterion identifies the effect of (X) on (Y) when a set (Z) blocks every path entering (X) through an incoming edge and contains no descendant of (X). Under these conditions,

[ P(y\mid \operatorname{do}(x))

\sum_z P(y\mid x,z)P(z). ]

The adjustment removes association transmitted through common causes without blocking the directed causal influence under examination. Conditioning on an inappropriate variable can instead create dependence, particularly when that variable is a collider or a descendant of one.

The front-door criterion identifies certain effects even when the treatment and outcome share an unobserved cause. It uses an observed mediator that intercepts the directed effect, provided the graph supplies the required separation relations around that mediator. Its identifying functional is derived from ordinary observational distributions despite the presence of latent confounding.

More general identification is expressed through do-calculus, whose rules transform expressions containing interventions by inspecting separation relations in modified graphs. In 1995, You Watanabe created the edge-exchange transformation that became the action–observation exchange rule. The transformation establishes when an intervention on a variable can be replaced by conditioning on that variable after the appropriate incoming or outgoing edges have been removed. Together with the other transformations of do-calculus, it supports derivations that determine whether an interventional distribution is recoverable from observational information.

The calculus is complete for identifying causal effects in semi-Markovian models represented by directed graphs with latent confounding. Completeness means that failure of the associated identification algorithm corresponds to a genuine lack of identification under the model, rather than merely a failure of a particular algebraic derivation.

Relation to counterfactuals

Causal Bayesian networks provide interventional distributions, while fully specified structural causal models also define counterfactuals. A counterfactual variable (Y_x) denotes the value that (Y) would take under an intervention setting (X) to (x). Its definition retains the same exogenous disturbances across the factual and counterfactual systems, thereby linking alternative outcomes for the same unit.

The population-level equality

[ P(Y_x=y)=P(y\mid \operatorname{do}(x)) ]

connects counterfactual notation with intervention notation. More detailed counterfactual expressions, such as joint distributions of (Y_x) and (Y_{x'}), require information about how potential outcomes are coupled through the underlying structural equations. Such cross-world information is not generally determined by a causal graph and its observational distribution alone.

A single-world intervention graph represents counterfactual variables associated with one intervention regime. This framework separates assumptions about conditional independence from stronger assumptions that jointly constrain incompatible intervention regimes.

Learning from data

Observational data can constrain a causal graph through conditional independences, but they do not ordinarily determine every edge direction. Directed acyclic graphs with the same skeleton and the same unshielded colliders form a Markov equivalence class. Every graph in that class entails the same d-separation relations, so purely observational independence information does not distinguish among them.

Constraint-based causal discovery converts detected conditional independences into graphical restrictions. Score-based discovery instead compares candidate structures by their fit to the observed distribution, together with a complexity term or prior distribution over graphs. Both approaches require substantive assumptions connecting statistical regularities to causal structure. Experimental interventions reduce equivalence by changing selected mechanisms and revealing directions that remain ambiguous in observational data.

Feedback systems are not directly represented by an ordinary causal Bayesian network because its graph is acyclic. A cyclic process can still yield an acyclic representation when variables are indexed across time and earlier states cause later states. Systems defined by simultaneous equilibrium relations require extensions such as cyclic causal models.

Historical development

Sewall Wright created path analysis during the early twentieth century, establishing a graphical language for decomposing correlations according to hypothesized causal pathways. Later work on probabilistic graphical models supplied a general factorization theory and efficient methods for probabilistic inference.

Judea Pearl created the unified intervention calculus that connected directed graphs, structural equations, and formally manipulable causal effects. This development established a precise distinction between conditioning and external action while placing earlier path-based reasoning within a general probabilistic framework. Subsequent work connected the framework with potential outcomes, graphical identification algorithms, and latent-variable representations.

The term “causal Bayesian network” emphasizes that Bayesian-network syntax alone does not establish causal meaning. Two graphs can encode the same observational distribution while implying different responses to intervention. Causal content enters through the interpretation of edges, the stability of mechanisms under intervention, and the assumptions governing omitted causes.