Structural Causal Models

Structural causal models (SCMs) are mathematical frameworks that represent a system as a set of autonomous structural equations, each expressing a variable as a function of its direct causes and an independent noise term. Developed by Judea Pearl, SCMs provide the formal foundation for causal inference by encoding not merely what variables correlate, but how the system generates its data.

Each equation in an SCM is autonomous: it remains stable under interventions on other variables. This autonomy is what makes counterfactual reasoning possible. To ask 'what would have happened if X had been x' is to replace the equation for X with X = x and solve the modified system. The noise terms represent exogenous variables — unmodeled background factors that influence the system from outside.

SCMs are more expressive than Bayesian networks alone. While a Bayesian network encodes conditional independence constraints, an SCM encodes the full functional relationships between variables. This extra structure is what permits the three levels of causal reasoning — association, intervention, and counterfactuals — that Pearl calls the ladder of causation. The do-calculus operates on SCMs by identifying which interventional distributions can be computed from observational data given the causal structure.

Structural causal models are not a refinement of statistics. They are a different ontology entirely — one that treats the world as a machine with knobs, not as a surface of patterns waiting to be memorized.