Abstract
Mediation analysis allows one to use observational data to estimate the importance of each potential mediating pathway involved in the causal effect of an exposure on an outcome. However, current approaches to mediation analysis with multiple mediators either involve assumptions not verifiable by experiments, or estimate the effect when mediators are manipulated jointly which precludes the practical design of experiments due to curse of dimensionality, or are difficult to interpret when arbitrary causal dependencies are present. We propose a method for mediation analysis for multiple manipulable mediators with arbitrary causal dependencies. The proposed method is clinically relevant because the decomposition of the total effect does not involve effects under cross-world assumptions and focuses on the effects after manipulating (i.e. treating) one single mediator, which is more relevant in a clinical scenario. We illustrate the approach using simulated data, the "framing" dataset from political science, and the HIV-Brain Age dataset from a clinical retrospective cohort study. Our results provide potential guidance for clinical practitioners to make justified choices to manipulate one of the mediators to optimize the outcome.
Introduction: The Challenge of Multiple Pathways
In clinical medicine, understanding how an exposure (like a disease) leads to an outcome is rarely a straight line. Instead, it often involves a complex web of intermediate factors known as mediators. For instance, HIV infection might accelerate brain aging through various pathways such as chronic inflammation, sleep disorders, or metabolic changes.
Mediation analysis is the statistical framework used to untangle these pathways. However, traditional methods struggle when multiple, interconnected mediators are present. They often rely on "cross-world" assumptions—mathematical comparisons of a patient's outcome under different exposure levels simultaneously—which are impossible to verify in real-world clinical settings. Furthermore, existing models often attempt to manipulate all mediators at once, leading to a "curse of dimensionality" where the data requirements become prohibitively large.
The authors address these issues by proposing a "clinically relevant" approach. Instead of asking abstract statistical questions, they ask a practical one: If we intervene on just this one specific mediator, how much will the final outcome improve?
Methodology: A Practical Decomposition
The core innovation is a new way to decompose the Total Effect (TE) of an exposure without relying on "natural" effects and their associated cross-world assumptions. The framework utilizes two primary metrics:
- Controlled Direct Effect (CDE): This measures the effect of the exposure on the outcome when a specific mediator is held at a fixed, "controlled" value. For example, it estimates the effect of HIV on brain aging if we could ensure a patient does not suffer from insomnia.
- Scaled Controlled Indirect Effect (sCIE): This new metric quantifies the impact of the pathway through a specific mediator. It is derived from the Controlled Indirect Effect (CIE), which represents the expected change in outcome if that mediator were manipulated or "treated."
To estimate these effects from observational data, the researchers employed Doubly Robust Estimation. This technique combines an outcome model with an exposure propensity model, providing accurate estimates even if one model is slightly misspecified. They implemented this using advanced machine learning algorithms like XGBoost and Random Forests to capture complex, non-linear relationships.
Datasets and Validation
The framework was validated across three distinct scenarios:
- Simulated Data: A synthetic environment with known causal ground truths to verify the mathematical accuracy of the decomposition.
- The "Framing" Dataset: A political science study examining how news framing of immigration affects public opinion through mediators like negative emotions and perceived harm.
- HIV-Brain Age (HIV-BA) Dataset: A clinical study of 3,091 subjects investigating how HIV infection influences the "Brain Age Index" (derived from sleep EEG) via mediators like hyperlipidemia, heart valve disorders, and insomnia.
Key Insights: Clinical Realism and Dependency
A major strength of this approach is its clinical realism. Physicians typically prioritize the most impactful intervention rather than treating all comorbidities simultaneously with equal weight. This framework allows for the ranking of mediators based on their CIE, providing a roadmap for clinical prioritization.
Furthermore, the method handles arbitrary causal dependencies. In human biology, mediators often cause each other (e.g., high cholesterol leading to cardiovascular issues). Unlike many previous models, this framework remains valid even when such dependencies exist. It also highlights that a mediator might have a small "natural" effect but a large "controlled" effect, meaning that while the disease might not strongly trigger that pathway, intervening on it could still yield significant clinical benefits.
Results and Discussion
In the HIV-Brain Age study, the analysis identified hyperlipidemia (high cholesterol/fats) as a dominant mediator. The CIE for hyperlipidemia was significantly higher than for other factors, suggesting that treating high cholesterol in HIV+ patients could be a highly effective strategy to mitigate accelerated brain aging.
In the Framing dataset, the analysis provided a nuanced view of public opinion. While "negative emotion" is often cited as a key mediator, the results showed that "perceived harm" actually had a higher CIE.
This suggests that interventions aimed at reducing the perceived harm of a situation might be more effective at shifting public opinion than those targeting emotions alone.
Limitations and Future Directions
Despite its advantages, the current model has limitations. It is currently constrained to binary exposures and mediators, whereas real-world clinical data often involve continuous variables (like blood pressure or drug dosage). Additionally, mediation analysis is inherently data-intensive; large cohorts are required to achieve narrow confidence intervals, as evidenced by the results in the HIV+ subgroup. Finally, like all observational causal inference, the results depend on the assumption of no unmeasured confounding.
Future research will focus on extending the mathematical framework to handle continuous variables and developing "path-specific" analysis to explore even more granular causal chains.
Summary updated