Structural Causal Model

The Causal Graph or Directed Acyclic Graph (DAG) is the qualitative, visual component of an SCM. It encodes the assumed causal relationships between variables using nodes and directed edges.

• Nodes represent the variables in the system.
• Directed Edges (arrows) represent direct causal relationships, where the parent variable is a direct cause of the child variable.
• The acyclic property ensures no variable can be a cause of itself, preventing causal loops.

This graph provides the structural assumptions necessary for formal causal analysis, distinguishing correlation from causation.

Structural Equations are the quantitative, functional component of an SCM. They specify how each variable is determined by its direct causes.

• Each variable X has an equation: X_i := f_i(PA_i, U_i).
• PA_i are the parent variables of X_i from the causal graph.
• f_i is a function (often deterministic) mapping parents to the child variable.
• U_i represents exogenous variables—unobserved background factors or 'noise' not caused by other variables in the model.

These equations define the data-generating process, moving from a static graph to a dynamic, computable model.

Exogenous Variables (U) represent all factors outside the model that influence the system. They are the source of randomness and unobserved confounding.

• Key Properties: Exogenous variables have no parents within the SCM; they are external inputs. They are often assumed to be independent of each other.
• Role in Inference: The distribution of U, combined with the structural equations, induces the joint probability distribution over all observed (endogenous) variables. This allows the SCM to answer interventional ('do-operator') and counterfactual queries by manipulating these equations.
• In practice, U captures measurement error, hidden confounders, and stochasticity.

The do-operator, denoted do(X=x), is a mathematical operator defined by an SCM that represents an external intervention to force a variable to a specific value, breaking its natural causal mechanisms.

• Mechanism: Applying do(X=x) modifies the structural equation for X, replacing it with the constant X := x. All other equations remain unchanged.
• Purpose: This allows the model to compute causal effects, answering 'what if' questions like 'What would happen to Y if we set X to x?'
• Contrast with Observation: P(Y | do(X=x)) is the interventional distribution, which often differs from the observational conditional probability P(Y | X=x) due to confounding.

It is the foundation for causal inference from the model.

Counterfactuals are the most advanced queries an SCM can answer. They involve reasoning about hypothetical, alternative pasts given what actually occurred.

• Structure: A counterfactual query has the form: 'What would have happened to Y if X had been x', given that we observed a specific outcome where X was not x.
• Computation: Answering counterfactuals requires three steps:
  1. Abduction: Update beliefs about the exogenous noise variables (U) based on the actual evidence.
  2. Action: Apply the do-operator to modify the model as per the hypothetical intervention.
  3. Prediction: Use the modified model with the updated U to compute the new outcome.

This 'abduction-action-prediction' cycle is the hallmark of deep causal reasoning.

SCMs provide a formal, mathematical framework for causal abduction—inferring the best causal explanation for observed data.

• Hypothesis Generation: The space of possible explanations is defined by the SCM's structure and the possible values of its exogenous (U) and latent variables.
• Hypothesis Ranking: Explanations (specific configurations of U) are ranked by their likelihood given the observed evidence and prior beliefs, often using probabilistic or Bayesian methods.
• Example in Diagnostics: Given symptoms (observed variables), an SCM of a disease system can abduce the most likely underlying fault (a value for a latent 'disease state' variable) by reasoning backwards through the structural equations.

Thus, SCMs operationalize inference to the best explanation within a rigorous causal context.

SCMs are the foundation for causal discovery algorithms that infer causal graphs from observational data. These algorithms, such as PC and FCI, test for conditional independencies to propose plausible causal structures.

• Key Use: Uncovering hidden cause-effect relationships in complex systems like genomics or econometrics where controlled experiments are impossible.
• Example: Identifying that a specific gene expression (variable X) is a direct cause of a disease marker (variable Y), not just correlated through a confounder.

Given a causal graph, SCMs enable the precise estimation of interventional effects (the do-operator). This answers "what if" questions by mathematically removing confounding bias.

• Core Technique: Do-calculus provides rules for transforming interventional queries into statistical estimands using observational data.
• Enterprise Application: In marketing, estimating the true incremental sales lift of an ad campaign by controlling for seasonality and customer demographics.

SCMs are used to build ML models that generalize beyond their training distribution by learning invariant causal mechanisms. This counters the problem of spurious correlations.

• Principle: Models should predict based on causal parents of the outcome, not on unstable, correlative features.
• Impact: Creates more reliable models for deployment in changing environments, such as diagnostic AI that works across different hospital imaging devices.

SCMs provide the semantics for answering counterfactual queries—questions about past events, like "Would the patient have recovered if they had taken the drug?"

• Mechanism: Uses the structural equations and evidence about what did happen to simulate a world where an antecedent was changed.
• Critical For: Root cause analysis in system failures, legal liability assessment, and personalized treatment plans in medicine.

SCMs explicitly model confounding paths that lead to discriminatory bias. This allows for the formal definition of fairness criteria (e.g., counterfactual fairness) and the development of de-biasing techniques.

• Process: The causal graph identifies if a sensitive attribute (e.g., gender) affects the outcome through a fair mediating path or an unfair, discriminatory path.
• Outcome: Enables the audit of algorithms for unlawful discrimination and the design of fairer systems.

In abductive reasoning systems, SCMs provide the structured space of possible causal explanations. Algorithms can generate and rank causal hypotheses that best explain observed anomalies.

• Workflow: 1. Observe unexpected data. 2. Propose modifications to an existing SCM (e.g., adding a latent variable). 3. Test if the new model better explains the data.
• Domain: Used in fields like astronomy to hypothesize new celestial mechanics or in pharmacology to suggest novel drug interaction pathways.

A Causal Graph or Directed Acyclic Graph (DAG) is the graphical component of an SCM. Each node represents a variable, and a directed edge (X → Y) indicates that X is a direct cause of Y. The 'acyclic' property ensures no variable can be its own cause. This graph encodes the qualitative causal assumptions and the conditional independencies implied by the model's structure, which are essential for identifying causal effects from data.

A Structural Equation is a mathematical function that defines how a variable in an SCM is generated. For a variable Y, its structural equation is Y := f_Y(PA_Y, U_Y), where PA_Y are Y's parent variables in the causal graph (its direct causes) and U_Y is an exogenous 'noise' variable representing unmodeled factors. The assignment operator ':=' denotes a deterministic or stochastic causal relationship, not mere statistical association.

The do-operator, denoted as do(X=x), represents an external intervention that forcibly sets a variable X to a specific value x, breaking its normal structural equation. This formalizes the 'what if' question central to causal inference. In an SCM, performing do(X=x) modifies the model by replacing X's structural equation with X = x, allowing the computation of interventional distributions P(Y | do(X=x)), which differ from observational distributions P(Y | X=x).

A Counterfactual Query is the most detailed level of causal reasoning, answering 'what would have happened if...' given what actually did happen. It involves reasoning about a specific unit (e.g., a patient) under a hypothetical intervention, conditional on the observed facts. SCMs enable counterfactual computation by using the observed exogenous noise variables (U) to simulate an alternative world where the intervention was applied, keeping all other background conditions constant.

Causal Identification is the process of determining whether a causal quantity of interest (e.g., the Average Treatment Effect) can be uniquely computed from the available observational data and the assumptions encoded in the SCM's graph. It asks: can P(Y | do(X)) be expressed as a function of the observable distribution P(V)? Techniques like the backdoor criterion, front-door criterion, and instrumental variables provide graphical rules for establishing identifiability.

In an SCM, variables are partitioned into two sets:

• Endogenous Variables (V): Variables determined by other variables within the model (i.e., they have parents in the graph). Their values are given by structural equations.
• Exogenous Variables (U): Variables with no causes within the model. They represent external, background factors and are the source of all randomness. They are assumed to be mutually independent. The joint distribution P(U) over exogenous variables, combined with the structural equations, generates the full distribution over endogenous variables.