Abstract
This paper expands the toolkit for identifying direct causal effects in linear Structural Equation Models (SEMs). By using "edge removal" to transform complex graphs, the authors show that descendants of cause or effect variables can serve as "path-specific instrumental variables," allowing researchers to calculate causal strengths in scenarios where traditional graphical criteria fail.
Introduction
In the world of causal inference, a central challenge is identification: determining if a specific causal effect (like the impact of a drug on a recovery rate) can be uniquely calculated from observed data and a known model structure. For linear SEMs, researchers often rely on graphical criteria like the "single-door criterion" or "instrumental variables" (IVs). An IV is a third variable that affects the cause but only affects the outcome through that cause. However, these standard tools are not "complete"—there are many models where a causal effect is algebraically identifiable, but existing graphical rules simply can't see it. The authors aim to bridge this gap by providing new graphical rules that handle more complex "contaminated" relationships.
Methodology: Edge Removal and Path Analysis
The authors introduce a technique based on edge removal and Wright’s method of path analysis.
- Edge Removal: They define a modified graph ($G'$) where certain directed edges are deleted. For example, if a variable $W$ is a descendant of the cause $X$, it normally can't be an IV because $X$ influences $W$ directly. By "removing" the edges from $X$ to $W$ in the model's logic, they create a scenario where $W$ would satisfy the IV criteria.
- Path-Specific IVs: They treat these descendants as "path-specific" instruments.
- Algebraic Re-integration: Since we can't actually delete edges in real-world data, the authors use Wright’s path analysis to relate the covariance of the original graph to the modified one. They derive new formulas that "subtract out" the influence of the unwanted paths using other identifiable total effects.
Theorem 1: Descendants of the Cause
Theorem 1 addresses the use of a descendant of the cause variable $X$ as an instrument. Traditionally, descendants are avoided because they carry "extra" correlation that muddies the causal signal. This paper proves that if the total effect of $X$ on its descendant $W$ is identifiable, that "extra" correlation can be mathematically removed.
Theorem 2: Descendants of the Effect
Theorem 2 extends this logic to descendants of the effect variable $Y$. By identifying the total effect of $Y$ on a descendant $W$, researchers can leverage $W$ as a path-specific instrument to calculate the direct effect of $X$ on $Y$.
Key Insights and Results
The core "aha" moment is the realization that the direct causal effect $\alpha_{yx}$ can be expressed as a ratio of "adjusted" covariances. By subtracting the product of known total effects and covariances from the raw data, you isolate the specific causal link of interest.
The authors successfully derived three new identification formulas. For example, in a scenario where $W$ is a descendant of cause $X$, the direct effect $\alpha_{yx}$ is given by:
$$\alpha_{yx} = \frac{\sigma_{wy\cdot z} - \tau_{wx}\sigma_{xy\cdot z}}{\sigma_{wx\cdot z} - \tau_{wx}\sigma_{xx\cdot z}}$$
Where $\tau_{wx}$ is the total effect of $X$ on $W$. This formula allows for identification even when the standard IV formula ($\sigma_{wy} / \sigma_{wx}$) would yield a biased or incorrect result. These results provide a more "robust" way to compute effects; if two different graphical paths lead to the same result, it increases confidence in the model's correctness.
Limitations and Future Work
- Linearity Constraint: The entire framework is built on linear SEMs. These specific algebraic derivations do not directly translate to non-linear models or models with categorical variables without further significant research.
- Requirement of Prior Knowledge: To use these new criteria, you must first be able to identify certain "total effects" ($\tau_{wx}$ or $\tau_{wy}$). If those aren't identifiable via standard methods, this new method can't proceed.
- Completeness: The authors note it remains to be seen if these new criteria, combined with existing ones, are "complete" (i.e., whether they cover every possible identifiable case in a linear SEM).
Conclusion
This work expands the toolkit for causal identification by demonstrating that descendants of cause or effect variables, previously considered "forbidden" as instruments, can be utilized through path-specific analysis. This provides researchers with more robust methods for calculating causal strengths in complex linear structural equation models.
Summary updated