Single World Intervention Graphs

To know if a drug worked, you need to know what would have happened without it. A Single World Intervention Graph (SWIG) is a diagram that maps out a single one of these "what if" scenarios. For example, it can draw the world where a patient did take the drug, labeling the outcome as 'Health(with drug)'. To find the drug's effect, you would conceptually compare this to a separate SWIG for the world where they didn't take it.

Try this: Think about which parts of a situation would stay the same if you changed just one decision.

A Single World Intervention Graph (SWIG) is a graphical tool used to represent a single counterfactual scenario. It starts with a standard causal diagram (like a DAG) and modifies it to show what would happen under a specific, hypothetical intervention.

To create a SWIG, you first choose the variable to intervene on, for example, setting Treatment = 'yes'. The key step is to 'split' this intervention node. This split visually separates the background factors that might influence the choice of treatment from the new, externally set value of the treatment, making the counterfactual logic clear.

After the split, all variables that are downstream consequences (descendants) of the intervention are relabeled to show they are now potential outcomes. For example, an Outcome node becomes Outcome(Treatment = 'yes').

The resulting graph represents a single, fixed world. To find a causal effect, you must conceptually compare the outcomes from two different SWIGs. For instance, to see if a new training program worked, you'd compare the potential outcome Performance(trained) from one SWIG to Performance(not_trained) from another.

Single World Intervention Graphs (SWIGs) were developed to graphically represent counterfactual quantities within the formal framework of Structural Causal Models (SCMs). While standard causal graphs (DAGs) encode structural assumptions about the world, SWIGs make the counterfactual dependencies for a specific intervention visually explicit.

• Pearl (2009) laid the groundwork with Structural Causal Models, where counterfactuals are defined as outcomes in a modified version of the model representing the hypothetical world. [1]
• Richardson & Robins (2013) formally introduced SWIGs as a graphical tool to unify Pearl's framework with the potential outcomes framework common in statistics. They demonstrated how SWIGs make it possible to read counterfactual dependencies directly from a single-intervention graph. [2]
• Hernán & Robins (2020) use SWIGs extensively in their textbook Causal Inference: What If to teach complex concepts like mediation, time-varying confounding, and selection bias, showing their utility as a powerful pedagogical and analytical tool in epidemiology and other fields. [3]

The primary limitation of SWIGs is that they require a correctly specified causal model (a DAG) as their starting point. If the underlying causal assumptions are wrong, the SWIG will also be wrong. Furthermore, for complex scenarios with many variables, interventions, or time points, SWIGs can become visually cluttered and difficult to interpret. They are a tool for formalizing questions and identifying what needs to be measured (identification), not a method for statistical estimation itself.