Structural Equation Models

Definition

Structural equation models represent variables as assignments generated from parent variables and exogenous factors. In causal use, the equations are interpreted as modular mechanisms that can be replaced under interventions.

Why It Matters

SEMs are one of the core languages for moving from statistical association to intervention reasoning. They force the analyst to say which mechanisms are assumed stable, which variables are parents, and what changes under an intervention.

Formalism / Key Objects

• Structural assignment:

$$X_i:=f_i(P{a}_i,U_i),$$

where Pa_i are parents of X_i and U_i are exogenous factors.

• A causal graph summarizes parent relations, but the intervention semantics lives in the equations.
• An intervention such as do(X=x) replaces the structural assignment for X with the constant assignment X := x.
• See Structural Intervention Semantics for the compact intervention-replacement schema.
• Causal Games adapt structural causal reasoning to strategic settings by treating some mechanisms as rationality relations over agents’ decision rules.
• Temporal SEM variants index variables by time and allow update equations to refer to prior-time variables, e.g. X_{t+1} := f_X(Pa_X(t), U_X(t)); see Temporal Causal Models and Temporal Structural Equation Models.
• Dyadic Morality uses SCM-like notation for perceived agency, vulnerability, harm, suffering, and wrongness, but its psychological operators are extra constraints rather than ordinary causal mechanisms.
• Counterfactual Simulation uses a recursive SCM M=(V,U,F,p(u)) to sample from p(W_do(X=x) | C=c) after updating background variables by factual evidence.

Connections

• Central to Causality, especially interventions and counterfactual reasoning.
• Connects to Probability and Statistics because SEMs are often estimated or checked using observational and experimental data.
• Connects to Decision Theory when causal models guide which action or intervention should be chosen.
• Causal Decision Making uses SEM-style intervention semantics to distinguish policy value under action from reward observed under an existing behavior policy.
• Fox2024 - Causality and Strategic Reasoning shows why strategic agents violate the ordinary independent-causal-mechanisms picture: one agent’s decision rule can depend on other mechanisms in the game.
• Everitt2019 - Towards Safe Artificial General Intelligence uses causal graphs as a diagnostic map for agent-caused corruption of reward, observations, utility functions, and reward-learning data.
• Reward Tampering is a case where the feedback mechanism itself becomes a descendant of the agent’s actions.
• Structural Intervention Semantics records the technical distinction between intervention and conditioning.
• Actual Causality uses structural models to evaluate whether an event is a cause of an outcome in a particular actual context.
• Temporal Causal Models explore whether SEM-style equations can represent ongoing computation and non-recursive temporal feedback.
• Dyadic Morality Algebra is a cautionary case: SCM notation can organize a symbolic model of moral appraisal even when the model includes non-causal psychological back-fill.
• Karvanen2024 - Simulating Counterfactuals gives an algorithmic route from structural equations to sampled counterfactual distributions when continuous evidence makes analytic conditioning difficult.

Common Confusions

• Conditioning on X=x is not the same operation as intervening with do(X=x).
• A graph alone does not guarantee intervention validity; assumptions about stability, modularity, and omitted variables matter.
• A causal game is not an ordinary SEM with decision variables renamed; it must also model rational response and possible equilibrium multiplicity.
• Some SEM traditions emphasize covariance modeling, while causal SEMs require stronger intervention-oriented interpretation.
• Using SEM notation for perception or moral judgment does not automatically make every edge a causal edge.

Key Sources

• Fox2024 - Causality and Strategic Reasoning
• Pearl2026 - Structural Equation Models and Intervention Assumptions Thread
• Everitt2019 - Towards Safe Artificial General Intelligence
• Ge2026 - A Review of Causal Decision Making
• Varshney2026 - An Algebraic Exposition of the Theory of Dyadic Morality
• Gladyshev2025 - Temporal Causal Reasoning with Non-Recursive SEMs
• Gladyshev2026 - Temporal Causal Models as a Model of Computation
• Karvanen2024 - Simulating Counterfactuals