A structural causal model (SCM) is a mathematical representation of a system in which each modeled variable is determined by a specified mechanism involving other variables and background factors. Unlike a model that describes only statistical associations, an SCM specifies how the system would change if some mechanisms were externally replaced. It provides a framework for causal inference, including the analysis of interventions and counterfactual questions.
The defining feature is not that the equations are linear, deterministic at the population level, or directly observable. It is that the equations receive a causal interpretation: each represents a mechanism assumed to remain unchanged when other mechanisms are modified, except where the model explicitly specifies otherwise.
Formal definition
An SCM is commonly written as a tuple
where:
- is a collection of exogenous variables, whose causes are not modeled within the system.
- is a collection of endogenous variables, whose values are determined within the model.
- is a collection of structural functions.
- is a joint probability distribution over the exogenous variables.
Each endogenous variable satisfies an equation of the form
where denotes its endogenous parents and denotes the exogenous inputs used by its mechanism. The assignment symbol emphasizes that the equation specifies how is generated, rather than merely asserting an algebraic equality.
The functions are deterministic once their inputs are fixed. Randomness arises through the exogenous random variables. This does not restrict SCMs to simple deterministic relationships between observed quantities: sufficiently rich exogenous variables can represent stochastic mechanisms.
“Exogenous” does not mean independent, randomized, or necessarily unobservable. It means that the variable's generating mechanism lies outside the model. Dependence between exogenous variables can represent unmodeled common causes.
Solutions and induced distributions
If the causal structure is acyclic, the equations can be evaluated in an order in which parents precede children. Under suitable measurability assumptions, every assignment of exogenous variables then determines a unique assignment of endogenous variables, inducing an observational distribution .
Models containing feedback require additional care. Their equations can have no solution, one solution, or multiple solutions for a given exogenous state. Existence, uniqueness, and behavior under interventions must therefore be specified or established; they do not follow merely from writing down a system of equations.
Causal graphs and statistical structure
An SCM has an associated directed graph with an arrow when the structural function for depends on . For an acyclic model, this is a directed acyclic graph.
The graph summarizes which variables directly enter which mechanisms, but it does not ordinarily determine the numerical strengths, functional forms, or exogenous distribution. Many distinct SCMs can therefore share the same graph.
When exogenous inputs are mutually independent and the endogenous graph is acyclic, the induced distribution factorizes as
with densities or probability masses used as appropriate. This resembles the factorization of a Bayesian network. The causal interpretation adds information absent from a purely statistical network: it specifies what happens when mechanisms are changed.
Graphical separation criteria imply conditional independence relations under the relevant Markov assumptions. The reverse inference generally requires further assumptions. In particular, faithfulness rules out conditional independences arising solely from special cancellations of causal effects.
If some common causes are unobserved, a graph over observed variables may use bidirected edges to represent latent common causes. Such edges are a graphical summary of latent structure, not reciprocal causal arrows.
Interventions and the do-operator
A central operation is an ideal intervention, written
It replaces the structural equation for with
while leaving the other equations and unchanged. The resulting model is often denoted . Its outcome distribution is written
This is fundamentally different from the conditional probability . Conditioning selects cases in which happened to equal ; intervention changes how is generated.
Example: association without a causal effect
Consider
where the exogenous variables are mutually independent, have mean zero, and have finite variances.
Here is a common cause of and . Their covariance is
so they are associated whenever has positive variance.
Under , however, only the equation for changes. The equations determining and remain untouched, so
The model therefore attributes no causal effect on to intervening on , despite their observational association. This is a basic example of confounding.
Other interventions
Not every intervention fixes a variable to a constant. A stochastic intervention replaces a mechanism with a specified random assignment rule. A policy intervention can assign values as a function of observed information. A mechanism intervention, sometimes called a soft intervention, changes a generating function or its noise distribution without necessarily removing all dependence on existing causes.
These interventions must be defined explicitly. Different ways of changing a variable can correspond to different causal questions.
Counterfactuals
An SCM also defines counterfactual quantities such as : the value that would take under for a particular exogenous state .
The same exogenous state is retained across factual and counterfactual scenarios. This connects statements about alternative outcomes for the same modeled unit, rather than merely comparing unrelated populations.
Counterfactual inference is commonly described in three steps:
- Abduction: update the distribution of exogenous variables using observed evidence , obtaining .
- Action: replace the relevant structural equations to form the intervened model.
- Prediction: evaluate the outcome in that modified model using the updated exogenous distribution.
Abduction can be understood as Bayesian inference over the background state, although the model's functions and parameters need not themselves be assigned Bayesian priors.
For example, suppose
For a unit observed with and , the equation implies . Under an intervention setting , the model gives
This unit-level conclusion depends on the specified structural equation and the assumption that the same applies in both scenarios. It is not generally determined by observational data alone.
Relationship to potential outcomes
The potential-outcomes framework represents outcomes under alternative assignments, such as . An SCM generates such quantities through its intervention semantics.
For a binary exposure, the average causal effect can be expressed as
where denotes expectation. An SCM additionally specifies mechanisms linking variables and can express joint counterfactual statements involving several hypothetical worlds.
Neither framework eliminates the need for assumptions. In particular, the joint distribution of and is generally not identified merely by knowing their separate distributions.
Identification and estimation
An SCM may define a causal quantity without making it recoverable from available data. Identification asks whether all models compatible with the stated assumptions and available distributions agree on that quantity. This is distinct from estimating it accurately from a finite sample.
Adjustment for confounding
If an observed set satisfies an appropriate back-door adjustment criterion relative to and , then, for discrete ,
For continuous variables, an integral replaces the sum.
The graphical criterion requires that contain no descendants of and block all paths from to that begin with an arrow into . Data-based estimation also requires sufficient support: the relevant exposure values must occur within the covariate strata being used.
Conditioning on more variables is not automatically safer. Adjusting for a common effect can introduce an association between its causes, while adjusting for a mediator can remove part of the total effect being studied.
Beyond adjustment
Some causal effects can be identified even when ordinary covariate adjustment is unavailable. Relevant approaches include front-door identification, instrumental variables, and do-calculus, a system of graphical rules for transforming expressions involving interventions.
These methods have different requirements. Instrumental-variable methods, for example, need assumptions about how the instrument affects treatment and outcome; the existence of a variable labeled an “instrument” does not by itself identify a desired effect.
Randomization can remove dependence between treatment assignment and background causes under the experimental assignment mechanism. Nevertheless, interpreting an experiment still requires specifying what intervention was implemented and how outcomes were measured.
Once identification is established, statistical methods estimate the identified expression. A large sample or a highly flexible predictive model cannot, by itself, resolve a failure of identification.
Linear and nonlinear models
A linear SCM can be written as
where the entry is the coefficient of the direct effect from to . If is invertible,
The matrix inverse combines contributions transmitted through the system. In an acyclic model, an appropriate ordering makes strictly triangular, ensuring that is invertible.
Direct coefficients need not equal total causal effects: effects can also travel through intermediate variables. Nor do coefficients from linear regression automatically have structural interpretations. Such interpretations depend on the causal model and assumptions about omitted causes, measurement, and disturbance dependence.
Nonlinear SCMs allow thresholds, interactions, discrete outcomes, and arbitrary measurable mechanisms. This flexibility broadens the class of systems represented, but does not automatically make their mechanisms identifiable or estimable.
For cyclic linear systems, invertibility of establishes a unique solution for the original equations, but not necessarily for every intervened system. Dynamic stability is also a separate property: a unique equilibrium need not be stable under a chosen time-evolution process.
Historical development
Structural causal modeling draws on several traditions. Sewall Wright's early twentieth-century path analysis represented relationships through directed diagrams and equations. Structural-equation methods in econometrics, including work by Trygve Haavelmo, distinguished structural relationships from the observational distributions generated by a system.
Later work connected structural equations with graphical models, formal interventions, and counterfactual semantics. Judea Pearl's research was central to the development and dissemination of the modern SCM framework and do-calculus.
The modern term encompasses more than traditional linear structural-equation modeling: structural functions may be nonlinear, variables may be discrete or continuous, and latent causes can be represented explicitly.
Applications and limitations
SCMs are used in economics, the social sciences, engineering, and machine learning to formalize causal assumptions, evaluate policy changes, study mediation, and examine how predictions might behave when a system changes.
Their principal limitation is that the causal content is not supplied by observational fit alone. Different models can generate the same observational distribution while disagreeing about interventions. Even models agreeing on all single-world interventional distributions may disagree about some joint counterfactual quantities.
Important modeling issues include:
- Mechanism invariance: unchanged equations must remain applicable under the intervention being considered.
- Model boundaries: omitted common causes, feedback, interference between units, or measurement processes can invalidate a proposed representation.
- Well-defined interventions: different procedures producing the same recorded exposure value may have different consequences.
- Measurement and selection: measurement error, missing observations, and selection bias can alter the relationship between the modeled system and the available data.
- Assumption sensitivity: conclusions may depend on untestable restrictions concerning hidden causes or functional form; sensitivity analysis examines how results change when such restrictions are varied.
An observational goodness-of-fit test can reject some SCMs, but cannot generally establish that a surviving model has the correct causal interpretation. Establishing that interpretation requires substantive knowledge, experimental evidence where available, and explicit justification of the assumptions connecting the equations to the system.