CAUSAL INFERENCE WITH A GRAPHICAL HIERARCHY OF INTERVENTIONS.

Area of Science:

• Causal inference
• Statistical modeling
• Observational data analysis

Background:

• Identifying causal parameters from observational data is challenging due to selection bias and confounding.
• Complex causal effects (e.g., treatment on the treated, mediated effects) may not be identifiable across different causal models.
• The diversity of causal models leads to a fragmented understanding of parameter identification.

Purpose of the Study:

• To provide a unifying framework for identifying a broad class of causal effects, including novel ones.
• To simplify identification theory by reducing it to the identification of random variables under a hierarchy of interventions.
• To connect existing causal models (Robins' FFRCISTG and Pearl's NPSEM-IE) to this unified framework.

Main Methods:

• Introduced a hierarchy of interventions to unify causal effect identification.
• Demonstrated that identification theory for a large class of causal effects simplifies to identifying random variables under interventions from this hierarchy.
• Linked specific interventions in the hierarchy to Robins' extended g-formula and Pearl's NPSEM-IE via the edge g-formula.

Main Results:

• A unified view of a large class of causal effects, including novel ones, is presented.
• Identification theory is shown to reduce to identifying random variables under interventions within the proposed hierarchy.
• The edge g-formula is identified as a key tool arising in mediation analysis and settings with unobserved causes of treatment.

Conclusions:

• The proposed hierarchy of interventions offers a unified approach to causal effect identification.
• The edge g-formula is crucial for identifying causal effects in complex scenarios, motivating further research in its estimation theory.
• This work simplifies the determination of parameter identifiability and the development of estimation and sensitivity analysis procedures.