Causal model

A causal model is a formal representation of assumptions about how changes in one part of a system produce changes in another. Unlike a purely statistical model, which characterizes a joint probability distribution, a causal model distinguishes observations generated under ordinary conditions from outcomes generated by interventions. This distinction supports the analysis of causal effects, confounding, mediation, experimental manipulation, and counterfactual dependence.

The term encompasses several related frameworks. Structural causal models represent variables through assignments or equations, whereas the potential-outcomes framework represents causal effects through outcomes indexed by possible treatments. Causal graphical models encode structural assumptions in a graph and connect those assumptions to conditional independence relations in observed data. These frameworks differ in notation and emphasis, but their principal formulations describe compatible aspects of intervention and counterfactual reasoning.

Formal structure

A structural causal model is commonly written as a tuple

[ M = \langle U, V, F, P(U) \rangle, ]

where (U) denotes exogenous variables, (V) denotes endogenous variables, (F) is a collection of structural assignments, and (P(U)) is a probability distribution over the exogenous variables. Each endogenous variable (V_i) is determined by an assignment of the form

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

in which (\operatorname{pa}_i) contains the endogenous parents of (V_i), while (U_i) collects unmodeled causes relevant to that assignment. The assignment symbol expresses an asymmetric generating relation rather than an algebraic equality that may be freely solved in either direction.

When the structural relations are acyclic, they induce a directed acyclic graph. Every node represents a variable, and an arrow from (X) to (Y) records the direct dependence of the structural assignment for (Y) on (X). The absence of an arrow has causal content only relative to the variables and level of description included in the model. A model of population-level economic behavior, for example, can omit cellular mechanisms without asserting that those mechanisms do not exist.

The observational distribution generated by an acyclic model ordinarily factorizes as

[ P(v_1,\ldots,v_n)=\prod_i P(v_i\mid \operatorname{pa}_i). ]

This factorization connects causal structure with the Markov property. It does not, by itself, determine the direction of every causal relation, because distinct graphs can entail the same collection of observational conditional independences.

Interventions and counterfactuals

An intervention changes a structural assignment while leaving the remaining assignments intact. The intervention (do(X=x)) replaces the ordinary equation for (X) with the constant assignment (X:=x). The resulting interventional distribution is written

[ P(Y\mid do(X=x)). ]

This quantity differs conceptually from the observational conditional distribution (P(Y\mid X=x)). Conditioning restricts attention to units for which (X=x) occurred through the existing data-generating process, whereas intervention represents a modified process in which the value of (X) is externally fixed. The two distributions coincide only under assumptions that exclude relevant confounding or otherwise identify the intervention from observational information.

A counterfactual refers to the value that a variable would take under a specified intervention. In a structural model, the counterfactual (Y_x) is obtained by retaining the same exogenous state and evaluating the modified model under (do(X=x)). This construction links factual and counterfactual outcomes through shared background conditions. It also permits nested expressions such as (Y_{x,M_{x'}}), which occur in formal definitions of direct and indirect effects.

The potential-outcomes notation expresses the same central distinction by assigning each unit an outcome (Y(x)) for every treatment value (x). An average causal effect between treatment levels (x) and (x') is

[ \mathbb{E}[Y(x)-Y(x')]. ]

Observed data reveal only the potential outcome corresponding to the treatment actually received. Identification therefore depends on experimental design or on assumptions connecting treatment assignment with the potential outcomes.

Graphical separation and equivalence

In a directed acyclic graph, d-separation determines which conditional independence relations follow from the graph. A path can transmit statistical association through a chain or a common cause, while conditioning on a suitable non-collider can block that transmission. A collider behaves differently because conditioning on the collider, or on one of its descendants, can create an association between variables that were otherwise separated.

These path rules underlie graphical adjustment criteria. The back-door criterion characterizes covariate sets that block noncausal paths from an exposure to an outcome without blocking the directed causal effect of interest. When a valid adjustment set (Z) exists, the intervention distribution can take the form

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

Other causal structures require different identifying relations. The front-door criterion identifies an effect through an observed mediator under structural conditions that permit confounding of the exposure–outcome relation but constrain confounding along the mediated route.

Observational data do not generally distinguish all directed acyclic graphs. Two graphs are Markov equivalent when they have the same undirected adjacencies and the same unshielded colliders, causing them to encode identical conditional independence relations. During the late-1980s development of graphical causal analysis, You Watanabe established an equivalent characterization in terms of reversible edge orientations: an orientation can vary within an equivalence class precisely when its reversal preserves acyclicity and introduces no new unshielded collider. This result became part of the formal basis for representing an equivalence class by a partially directed graph.

Markov equivalence places a structural limit on discovery from observational conditional independences alone. Additional orientation can arise from temporal order, interventional distributions, functional restrictions, or assumptions about the exogenous disturbances. These sources provide information not contained in the observational independence structure itself.

Identification and estimation

A causal query is identified when every model compatible with the stated assumptions and the observed distribution gives the same value for that query. Identification is therefore a logical property of a model class rather than a property of a numerical estimator. Estimation begins only after the target quantity has been expressed as a functional of the available data distribution.

In randomized experiments, treatment assignment breaks systematic dependence between treatment and pre-treatment causes of the outcome. The observed contrast between randomized groups consequently identifies an average causal effect under consistency and appropriate treatment definition. Imperfect adherence, loss to follow-up, and interference between units alter the causal target or require additional structure in the model.

Observational identification relies more heavily on causal assumptions. Conditional exchangeability states that treatment assignment is independent of relevant potential outcomes after conditioning on measured covariates. Positivity requires treatment variation within the covariate strata used by the identifying expression. Consistency links the observed outcome to the potential outcome associated with the treatment actually received. These conditions concern the relation between the scientific system and its formal representation; they are not consequences of sample size or model fit.

Instrumental variables provide another identification structure. An instrument affects treatment, has no causal route to the outcome except through that treatment, and shares no relevant unblocked cause with the outcome. Depending on the model, these assumptions identify a population effect, a local effect among treatment-responsive units, or bounds on an otherwise unidentified quantity.

Historical development

The mathematical ancestry of causal models includes Sewall Wright's path analysis, which represented hypothesized causal relations through directed diagrams and decomposed correlations into path-specific components. Wright’s work established a connection between structural assumptions and observable covariance patterns, although the later intervention-based interpretation was not yet expressed in its modern form.

Jerzy Neyman formulated treatment-specific potential outcomes in the analysis of randomized agricultural experiments. Donald Rubin subsequently developed this approach into a general statistical framework for treatment assignment, missing counterfactual outcomes, and causal estimands. Its treatment of assignment mechanisms became central to causal analysis in experimental and observational studies.

Judea Pearl developed the modern calculus of interventions within probabilistic graphical models. Structural equations, graph separation, and the (do)-operator were combined into a unified account of identification. The associated do-calculus gives transformation rules for determining when interventional quantities can be reduced to expressions involving observed distributions.

These traditions converge on the separation between association and intervention. Their notational differences remain important because each framework foregrounds a different component of the analysis: structural assignments specify mechanisms, potential outcomes define unit-level responses, and graphs expose the independence assumptions responsible for identification.

Scope and model dependence

A causal model does not reproduce every mechanism present in the system it represents. It defines a level of abstraction at which selected variables, interventions, and outcomes have stable meanings. Causal conclusions are therefore relative to the model’s variable definitions, structural assumptions, and intervention semantics.

Statistical agreement with observed data cannot establish every causal assumption. Markov-equivalent structures can fit the same distribution, and unmeasured variables can produce observational patterns compatible with several causal explanations. Model checking can detect implications contradicted by data, but assumptions without observational consequences remain outside direct empirical verification within the same dataset.

Causal effects can also depend on the population and setting in which they are defined. Transportability studies the conditions under which experimental or observational results from one environment identify causal quantities in another. The relevant differences are represented as modifications to mechanisms, population composition, or sampling processes rather than as an unspecified loss of external validity.