Abstract
Identifying causal parameters from observational data is fraught with subtleties due to the issues of selection bias and confounding. In addition, more complex questions of interest, such as effects of treatment on the treated and mediated effects, may not always be identified even in data where treatment assignment is known and under investigator control, or may be identified under one causal model but not another. Increasingly complex effects of interest, coupled with a diversity of causal models in use, resulted in a fragmented view of identification. In this paper, we give a unifying view of a large class of causal effects of interest in terms of a hierarchy of interventions and show that identification theory for this large class reduces to an identification theory of random variables under interventions from this hierarchy. Moreover, we show that one type of intervention in the hierarchy is naturally associated with queries identified under the Single World Model (via the extended g-formula), and another is naturally associated with queries identified under the Multi-World Model via a more general functional we call the edge g-formula.
Introduction
Causal inference is currently a fragmented field. Researchers often find themselves choosing between competing frameworks—such as Pearl’s Structural Equation Models (NPSEM-IE) or Robins’ Structure Tree Graphs (FFRCISTG)—which can lead to conflicting conclusions regarding the identifiability of complex causal queries. While these frameworks generally agree on simple "total effect" questions, they diverge significantly when addressing mediation or effects on the treated. This paper introduces a rigorous mathematical framework to unify these perspectives by establishing a graphical hierarchy of interventions.
The Hierarchy of Interventions
The core of this framework is the classification of interventions into three nested levels of specificity:
- Node Interventions: The standard "do-operator," where a variable (node) is set to a specific value.
- Edge Interventions: A more surgical strike that intervenes on a specific relationship (arrow) between variables. For example, setting a drug's effect on a mediator to one value while allowing its direct effect on the outcome to remain at another.
- Path Interventions: The most general form, intervening on specific causal pathways through the graph.
As illustrated in the hierarchy above, these intervention types map directly to specific causal models and mathematical functionals. Node interventions are identified under Single World Models (SWM) using the extended g-formula, while edge and path interventions require the stricter assumptions of Multi-World Models (MWM) and are identified via the newly defined edge g-formula.
Key Insights and the Edge G-Formula
A major contribution of this work is the derivation of the edge g-formula. The authors demonstrate that this functional provides a single mathematical language for mediation analysis, longitudinal studies, and hidden variable problems.
One of the most striking "aha moments" in the paper is the realization that Pearl’s famous "front-door criterion" (used to identify effects in the presence of unobserved confounding) and the standard mediation formula are actually the same mathematical object—the edge g-formula—just applied to different graph structures. This unification simplifies the theoretical landscape significantly.
Results and Robustness
The paper proves that the edge g-formula is the correct tool for identifying any edge-consistent intervention. Furthermore, for specific applications, the authors present a "triply robust" estimator. This ensures that causal estimates remain valid even if two out of the three underlying statistical models (treatment, mediator, or outcome models) are incorrectly specified, providing a safeguard against model misspecification in practical applications.
Limitations and Future Directions
While the paper solves the mathematical problem of identification, several challenges remain for future research:
- Estimation Complexity: The actual computation of the general edge g-formula with noisy data is computationally intensive and requires further development in semi-parametric theory.
- Hidden Variables: Extending this hierarchy to Acyclic Directed Mixed Graphs (ADMGs), which explicitly model unobserved confounding, is a primary direction for future work.
- Principal Stratification: Some causal targets, such as principal stratification effects, remain outside this hierarchy as they involve conditioning on variables with conflicting values across different causal "worlds."
Summary updated