TL;DR
This paper introduces a robust framework for decomposing treatment effects on terminal events (like death) into direct and indirect pathways when intermediate events (like disease relapse) are present. By shifting the focus to "cumulative incidence" and offering a new hazard-based decomposition strategy, the authors provide researchers with more interpretable and clinically relevant tools for causal mediation analysis in survival data.
What Problem Does This Paper Address?
In clinical trials, we often care about more than just the final outcome (the "terminal event," such as death). We also want to understand how a treatment affects intermediate milestones (the "non-terminal event," such as cancer progression or relapse). This is known as semi-competing risks: death can "censor" (prevent us from seeing) a relapse, but a relapse doesn't stop death from occurring.
Traditional mediation analysis helps us understand if a drug works directly on the outcome or indirectly through an intermediate variable. However, applying this to survival data is tricky. Previous methods often relied on "cumulative hazards," which are mathematically convenient but lack intuitive meaning and are "non-collapsible"—meaning you can't easily average effects across different groups. Furthermore, existing models often assumed that the prevalence of the intermediate event was the key factor to control, which doesn't always align with biological reality.
Methods
The authors propose two distinct strategies to decompose the total treatment effect into a Natural Direct Effect (NDE) and a Natural Indirect Effect (NIE) using the potential outcomes framework.
- The Estimand: Instead of hazards, they use counterfactual cumulative incidence. This represents the probability of an event occurring before time $t$, which is much easier for clinicians to interpret than abstract hazard rates.
- Decomposition 1 (Prevalence-based): This follows earlier work by adjusting for the prevalence of the non-terminal event. It asks: "What would the effect be if we kept the proportion of patients who relapsed the same as it was under the control group?"
- Decomposition 2 (Hazard-based): This is the paper’s novel contribution. It adjusts for the hazard (the instantaneous risk) of the non-terminal event. This is often more realistic in medicine; for example, a therapy might change the rate at which patients relapse, rather than just the total number who have relapsed by a certain date.
The researchers established the mathematical "identifiability" of these effects (proving they can be calculated from observable data) and used Martingale theory to prove that their estimators are consistent and follow a normal distribution as sample sizes grow.
Datasets
The authors validated their approach using:
- Simulated Data: Three different scenarios designed to test how the two decomposition strategies react when only specific risks (direct, intermediate, or indirect) are varied.
- Hepatitis B Dataset: Data from 4,954 male participants used to study how Hepatitis B infection (treatment) leads to death (terminal event) through the development of liver cancer (non-terminal event).
- Leukemia Dataset: Data from 303 patients comparing two types of bone marrow transplants (MSDT vs. Haplo-SCT) to see how they affect mortality through the risk of leukemia relapse.
Key Insights and Ideas
- Collapsibility Matters: By using cumulative incidence, the authors ensure that the treatment effect measured in the whole population is a meaningful average of the effects in subpopulations.
- The "Hazard vs. Prevalence" Choice: The paper highlights a subtle but vital distinction. Controlling prevalence (Decomposition 1) can sometimes "leak" direct effects into the indirect effect category. This happens because the number of people at risk for relapse changes if the treatment successfully prevents early death.
- Clinical Alignment: In the leukemia study, the hazard-based approach (Decomposition 2) aligned much better with clinical experience. It correctly identified that certain transplants reduce mortality primarily by lowering the risk rate of relapse (the "Graft-versus-Leukemia" effect), whereas the prevalence-based approach failed to find a significant result.
Key Results
- Mathematical Proofs: The authors successfully provided the asymptotic properties (large-sample behavior) for their estimators, allowing for the calculation of 95% confidence intervals.
- Hepatitis B: Both decomposition methods yielded similar results, showing that Hepatitis B increases mortality almost entirely by increasing the risk of liver cancer (a strong indirect effect).
- Leukemia: Decomposition 2 revealed that Haplo-SCT transplants significantly lower mortality by reducing the risk of relapse. Specifically, the natural indirect effect became significant as time progressed, a finding that was obscured when using the older prevalence-based method.
Limitations and Future Work
The primary limitation, common in causal inference, is the reliance on untestable assumptions. Specifically, "sequential ignorability" assumes there are no unmeasured confounders affecting the intermediate and terminal events. While the authors provide a framework for randomized experiments, real-world observational data often contains hidden biases.
The authors suggest that future work could involve:
- Developing methods to handle unmeasured confounding more robustly.
- Applying these decompositions to multi-state models with more than two events.
- Improving variance estimation for small sample sizes, as the current asymptotic formulas can be slightly biased when data is sparse.
Summary updated