Standard causal inference often relies on "graph surgery"—the act of physically removing edges from a Directed Acyclic Graph (DAG) to represent an intervention. While the $do$-calculus provides a powerful logic for this process, it treats interventions as external operations that fundamentally change the model's structure. A growing body of research suggests a more unified approach: instead of changing the graph, we can augment it. By representing interventions as specialized nodes within the graph itself, we can treat experimental data as a structured form of observational data, unlocking new ways to discover and identify causal relationships.
Interventions as Auxiliary Nodes
The core intuition behind augmented causal graphs is to represent the mechanism of an intervention as a variable. In this framework, an intervention is not a manual override of a node's value, but rather an external influence from a "regime" or "indicator" node.
Didelez et al. formalize this by extending standard DAGs into influence diagrams using regime indicators. These indicators act as parents to the variables they target, signaling whether a variable is behaving according to its natural causal parents or is being manipulated by an external force. Similarly, Kocaoglu et al. introduce "F-nodes" to encode these effects.
This representation is powerful because it turns causal identification into a problem of conditional independence. If we can show that a target variable is d-separated from its intervention node under certain conditions, we can identify the causal effect using standard probabilistic tools.
Beyond Hard Interventions
Traditional causal models often focus on "hard" interventions, where a variable is forced to a constant value (e.g., $do(X=x)$). However, real-world experiments often involve "soft" interventions—perturbations that modify a variable’s distribution without completely severing its relationship with its parents.
Augmented graphs handle this naturally. Yang et al. demonstrate that by treating soft interventions as auxiliary nodes, researchers can learn causal structures that are otherwise indistinguishable. While observational data can only identify a graph up to its Markov equivalence class (the set of graphs with the same skeletons and v-structures), Yang et al. show that interventional data narrows this class significantly.
Winn takes this further by introducing "gates" in factor graphs. In this view, an intervention is a selector variable that switches between a variable's natural causal factor and an intervention-driven factor. This allows for "graph surgery" to be modeled within a static, unified structure, integrating causal reasoning directly into standard probabilistic inference.
Resolving Ambiguity with Context
One of the primary challenges in causal discovery is the presence of latent (unobserved) variables. When we cannot see every cause, different causal structures can produce identical data. Augmented graphs provide a way to break these ties.
The authors of A Graphical Criterion for the Identification of Causal Effects in Linear Models show that "auxiliary variables"—external variables with known structures—can serve as instruments to isolate and estimate otherwise obscured causal effects. This concept is refined by Oates et al., who propose Conditional DAGs (CDAGs). By incorporating secondary variables as known causes, CDAGs can resolve ambiguous links between primary variables that would be unidentifiable in a standard DAG.
Even when the exact target of an intervention is unknown, the augmented framework remains robust. Jaber et al. introduce auxiliary context variables that act as placeholders for unknown mechanisms. Their $\Psi$-FCI algorithm can reconstruct causal structures in complex systems with hidden confounders by detecting where these context variables have modified local distributions.
Context-Specific Independence
Finally, augmenting graphs allows researchers to exploit "Context-Specific Independence" (CSI). Standard $do$-calculus assumes that if an edge exists, it is always "active." However, Tikka et al. argue that many causal pathways only trigger under specific conditions.
By treating interventions as states within an expanded graphical model, Tikka et al. develop a "CSI-calculus." This framework can identify causal effects that are theoretically unobtainable through standard $do$-calculus by leveraging the fact that certain dependencies vanish in specific interventional contexts.
Summary of Identifiability Gains
The transition from standard DAGs to augmented graphs represents a shift from viewing interventions as changes to the system to viewing them as inputs to the system. As shown by Zhou et al. and Zhou et al., this perspective allows for:
- Refined Equivalence Classes: Distinguishing between graphs that look identical under observational data.
- Handling Latent Confounders: Using regime indicators to "triangulate" the presence of hidden variables.
- Unified Learning: Algorithms like IGSP and $\Psi$-FCI that seamlessly combine observational and interventional datasets into a single, consistent causal map.
Key Foundational Work
A Graphical Criterion for the Identification of Causal Effects in Linear Models (2002) — contributed 2 concept elements to this direction.
Direct and Indirect Effects of Sequential Treatments (2006) — contributed 1 concept elements to this direction.