AI Glossary / Causal inference and experimentation / Advanced
Structural causal model
Definition
A structural causal model, or SCM, represents how variables are generated by causal mechanisms. It combines equations describing those mechanisms with background variables and, for probabilistic questions, a distribution over those background variables. The model supports questions about observations, interventions and counterfactual situations, subject to its assumptions.
What it means
A predictive model can say which outcomes tend to occur together. A structural causal model additionally states what would change if a mechanism were deliberately altered. This distinction matters when a business wants to act on a prediction. Customers who receive support may have more problems because problems trigger support; that association alone does not show that support causes problems. An SCM makes the proposed causal relationships explicit so that the question and the assumptions can be examined together.
How it works
Each modeled variable is assigned a function of its causal parents and relevant background factors. A graph can display the parent relationships, while equations supply additional detail. An intervention such as do(X=x) replaces the mechanism producing X with a fixed value, leaving other mechanisms unchanged as assumed. Counterfactual analysis also uses evidence about a particular case before evaluating an alternative intervention. Observational data alone may not identify the requested effect. Unmeasured common causes, feedback and the model's solution properties require care rather than automatic use of a simple acyclic diagram.
Three practical examples
- A retailer separates the effects of weather, promotions and customer traffic before asking what a promotion would change. A sales prediction alone cannot settle that question.
- A factory models how workload and maintenance influence downtime. The graph helps engineers notice that maintenance is often scheduled precisely when equipment condition worsens.
- A service team explores whether shorter response times would change satisfaction. It distinguishes response time from case difficulty, which may influence both response time and satisfaction.
Think of it this way
Think of a set of connected mechanisms rather than a collection of trend lines. Turning a control changes one mechanism's input, while merely noticing its current position gives information about how the system arrived there. The model describes the consequences of the turn only if the assumed connections and stable mechanisms are suitable for that intervention.
Worked example
Consider an original toy model with workload W, support hours S and resolved cases Y. Suppose W is equally likely to be 0 or 1, S=W, and Y=10+5S+8W. When S=1 is observed, W must also equal 1, so Y=23. When S=0 is observed, Y=10. The observed difference is 13. Under an intervention setting S=1 for everyone, average Y becomes 10+5+8(0.5)=19. Setting S=0 gives 14. The causal difference in this stipulated model is therefore 5, not 13. Because S always equals W in the observed data, those data alone lack the overlap needed to recover the intervention contrast without additional assumptions or information. The equations supply the answer here.
Illustrative case study
A fictional support company sees that teams receiving extra assistance also have more unresolved cases. Managers initially worry that assistance is ineffective. Analysts draw a causal model in which backlog severity triggers assistance and independently affects later unresolved cases. They distinguish this proposed mechanism from the association in the dashboard. The model guides a discussion about recording pre-assistance severity and designing a controlled rollout where feasible. It does not certify that all important causes have been measured. The case illustrates how an SCM can improve the question being asked before a business attaches a causal meaning to its data.
Common misconceptions and mistakes
A graph is not proof that its arrows are correct. A fitted regression equation also does not become structural merely because it contains several variables. Another mistake is equating conditioning on X with intervening on X: selection into an observed value can carry information about other causes. Analysts can additionally confuse a population intervention with a counterfactual about one particular person or case. The question, assumptions and available evidence must match. More data can reduce sampling uncertainty without resolving an unidentified causal question.
Business use
SCMs can help teams make assumptions visible when evaluating process changes, product features or operational policies. They are useful for organizing expert knowledge and identifying what additional evidence could change a decision. A practical discussion should name the intervention, outcome, relevant timescale and population. Stakeholders can then challenge missing causes, implausible mechanisms and whether the proposed action would leave other relationships stable. The business value lies in more disciplined reasoning, not in treating a diagram as an automatic decision engine.
People also ask
Must the equations be linear?
No. Structural mechanisms can be nonlinear.
Is every graphical model causal?
No. A graph can describe statistical dependence without intervention semantics.
Can an SCM answer every causal question?
No. Some questions remain unidentified or depend on assumptions that are not supported well enough for the intended use.
Check your understanding
Why was the observed difference 13 while the stipulated causal difference was 5?
Answer: Workload affected both support hours and resolved cases.
What did the intervention replace?
Answer: The equation S=W.
Why does the example not prove that observational analysis recovered the effect?
Answer: The toy mechanisms were supplied in advance, and the observed assignments lacked overlap.
Related terms
- Causal graphs
- Confounding
- Causal graph
- structural equation
- intervention
- do-operator
- counterfactual
Learn next
Draw a small graph for a familiar workflow and write what each arrow claims. Compare observing a variable with setting it. Then study identification and adjustment criteria before estimating an effect from business records.
More in Causal inference and experimentation
- Silhouette Coefficient - The Silhouette Coefficient measures how well an observation fits its assigned cluster compared with the nearest competing cluster.
- Similarity measure - A similarity measure is a rule that assigns a numerical score to how alike two objects are in a chosen representation.
- Squared hinge loss - Squared hinge loss is a classification loss that squares the positive shortfall from a target margin.
- Super learner - Super learner is an ensemble-learning approach that uses cross-validation to select or combine predictions from a library of candidate algorithms according...
- Test and Evaluation, Verification and Validation - Test and Evaluation, Verification and Validation, commonly abbreviated TEVV, is a collection of activities that builds evidence about an AI system's...
- Test loss - Test loss is the value of a chosen loss function on a dataset reserved for evaluating a trained and selected model.
Sources
Judea Pearl, An Introduction to Causal Inference: https://ftp.cs.ucla.edu/pub/stat_ser/r354-corrected-reprint.pdf Pearl, Glymour and Jewell, Causal Inference in Statistics: A Primer: https://web.cs.ucla.edu/~kaoru/BIB5/pearl-etal-2016-primer-text.pdf The support equations and organizational scenarios are original teaching examples with explicitly stipulated assumptions.
Previous: Structural break Next: Structural EM