Structural Intervention Semantics
Context
Pearl2026 - Structural Equation Models and Intervention Assumptions Thread highlights the difference between using SEMs as covariance models and using them to guide interventions. The reusable formal point is that interventions replace structural assignments, rather than merely condition on observed values.
Formal Statement
A structural equation model represents variables by assignments
where Pa_i are parent variables and U_i are exogenous factors. An intervention do(X_j=x) constructs a modified model by replacing the assignment for X_j:
Downstream variables are then evaluated under the modified assignment system.
Derivation / Construction
The intervention operation is modular: it changes the mechanism for the intervened variable while leaving other structural assignments fixed unless the intervention also targets them. This modularity is the formal assumption behind using the model to predict counterfactual or interventional consequences.
Conditioning, by contrast, leaves the original assignments in place and restricts attention to units where X_j=x was observed. Thus P(Y\mid X=x) and P(Y\mid do(X=x)) can differ whenever the observed value of X is confounded with other causes of Y.
Implications
- A causal SEM needs stability or modularity assumptions, not just a graph that fits associations.
- Intervention-guiding SEMs should record which mechanisms are assumed invariant under the proposed intervention.
- The
dooperation is the technical hinge between descriptive statistical models and decision-relevant causal models.