In causal inference, we often want to know more than just whether a treatment works; we want to know how it works. While standard mediation analysis splits a total effect into "direct" and "indirect" components, real-world systems are rarely that simple. A single treatment might trigger multiple intermediate processes—mediators—that interact, conflict, or confound one another.
Path-specific effect (PSE) estimation provides a surgical tool for this complexity. It allows researchers to isolate the causal influence flowing along one specific trajectory in a network while "freezing" the influence along all others.
The Geometry of Causal Trajectories
Imagine a clinical trial for a new antiretroviral therapy (ART). The drug is intended to suppress viral load by improving patient adherence. However, the drug also has side effects (toxicity), which might decrease adherence or independently affect health. As Kanki et al. demonstrate, a simple "indirect effect" through adherence is hard to measure because the treatment itself induces a new confounder: toxicity.
To solve this, researchers use counterfactual graphical models. As formalized by Shpitser, a path-specific effect is defined by an intervention that changes the treatment's value only for a specific subset of causal edges. In the HIV example, this means calculating what the viral load would be if the drug's effect on toxicity were held at a baseline level, while its effect on adherence was allowed to vary. This "path-based" intervention allows for the isolation of mechanisms even in longitudinal settings with unobserved confounding.
Decomposing Complex Networks
When multiple mediators are involved, the causal web becomes a dense thicket. If a lockdown reduces COVID-19 spread, does it work by reducing inter-city travel, intra-city gathering, or both? And how do these mediators interact?
The ANOCE framework, proposed by Cai et al., introduces an additive causal decomposition. The authors show that the total indirect effect can be broken down into a sum of individual mediator contributions. By using constrained structural learning, they can attribute causal sources even when mediators are not parallel but influence one another.
This additive property is a recurring theme in the literature. Kawakami and Tian extend this logic to the "Probability of Causation"—the likelihood that a treatment was necessary for an outcome to occur. They establish that the total probability of necessity and sufficiency can be partitioned into mutually exclusive direct and indirect components. This ensures that the importance of different pathways can be expressed as intuitive proportions that sum to 100%, avoiding the interpretability issues of negative effect sizes.
Overcoming Exposure-Induced Confounding
One of the most difficult challenges in mediation is "exposure-induced confounding," where the treatment affects a variable that then acts as a confounder between a later mediator and the outcome. Traditional methods fail here because you cannot simply "control" for the intermediate variable without blocking the very causal path you want to study.
To address this, Miles et al. developed semiparametric estimators that are "multiply robust." These estimators provide researchers with multiple chances to correctly specify their statistical models. If at least one of the models for the treatment, the mediator, or the outcome is correct, the resulting path-specific effect estimate remains consistent. This robustness is critical when dealing with high-dimensional biological or social data where the exact relationship between variables is unknown.
From Averages to Necessity
Most causal analysis focuses on population averages: "On average, how much does the mediator help?" However, path-specific analysis can also be retrospective. Kawakami and Tian provide a framework to decompose the probability of necessity and sufficiency into four distinct paths when two mediators are present.
This allows us to ask: "In this specific patient, was the outcome caused by the direct path, the simple indirect path, or a nested sequence of both?" By identifying which specific chains of events are essential for an outcome, researchers can pinpoint the most critical causal links to target in future interventions.
Applications in Fairness and Healthcare
The utility of path-specific effects extends into algorithmic fairness and diagnostic bias. For instance, Zhang et al. investigated why pulse oximeters—which are less accurate for dark-skinned patients—might lead to disparate outcomes in critical care. By isolating the path-specific effect of oximetry errors, the authors found that while the errors didn't change the decision to start ventilation, they causally led to significantly longer ventilation durations for Black patients.
Ultimately, as Zhang and Bareinboim argue, this approach generalizes the century-old rules of path analysis to modern, non-parametric structural causal models. By providing a mathematical operator to isolate influence along specific trajectories, PSE estimation transforms the "black box" of a total effect into a transparent map of causal mechanisms.
Key Foundational Work
Counterfactual Graphical Models for Longitudinal Mediation Analysis With Unobserved Confounding (2013) — contributed 2 concept elements to this direction.
Quantifying an Adherence Path-Specific Effect of Antiretroviral Therapy in the Nigeria PEPFAR Program (2017) — contributed 1 concept elements to this direction.