Introduction
In clinical research, investigators often seek to evaluate outcomes that are only defined if a subject survives a specific period. Examples include quality of life, cognitive function, or disease recurrence. However, "truncation-by-death" presents a significant challenge: if a patient dies before the outcome is measured, the data point is not merely missing—it is undefined. You cannot measure the quality of life of a deceased individual.
Analyzing only the observed survivors (a "survivor-case analysis") typically introduces selection bias. Survivors in a treatment group may possess different baseline characteristics—such as better health or younger age—compared to survivors in a control group. To address this, researchers utilize the Principal Stratification framework to define the Survivor Average Causal Effect (SACE). The SACE focuses on the "Always-Survivor" group: those individuals who would have survived regardless of whether they received the treatment or the control.
The Observational Study Dilemma
While methods to identify SACE exist for randomized controlled trials (RCTs), observational studies are far more complex. In real-world data, the treatment assignment, survival process, and final outcomes are often confounded by unmeasured factors. Standard assumptions used in RCTs—such as the independence of treatment assignment—rarely hold. This paper proposes a novel identification strategy for SACE in observational settings where unmeasured confounding is present, making it particularly useful for post-marketing drug safety and efficiency studies using real-world data.
Methodology: The Substitutional Variable Approach
The authors' framework relies on three primary components to "untangle" confounded effects:
1. Principal Stratification and Monotonicity
The population is divided into latent groups based on potential survival:
- Always-Survivors: Survive under both treatment and control.
- Protected: Survive only under treatment.
- Harmed: Survive only under control.
- Doomed: Die under both.
By invoking the monotonicity assumption—which posits that the treatment does not make survival worse (i.e., if you would survive under control, you will definitely survive under treatment)—the "Harmed" group is eliminated, simplifying the identification process.
2. The Substitutional Variable ($V$)
The core innovation of this research is the introduction of a substitutional variable. This variable must be correlated with the principal strata (helping to distinguish Always-Survivors from Protected survivors) but must not directly affect the potential outcome itself (an exclusion restriction). By leveraging $V$, the researchers can mathematically solve for the hidden proportions of Always-Survivors even when treatment assignment is confounded.
3. Estimation and Inference
The paper provides both parametric (using logistic regression) and nonparametric (using kernel regression) methods. These allow for the estimation of the Conditional SACE, recognizing that treatment effects are rarely uniform and often vary based on individual patient covariates ($X$).
Empirical Validation
Simulation Studies
The researchers conducted extensive simulations to compare their method against naive survivor-only analysis and existing SACE models. The results demonstrated that the proposed method significantly reduced bias and mean square error, particularly in the presence of strong unmeasured confounding where traditional models failed.
Real-World Application: Leukemia Relapse
The method was applied to a retrospective study of 1,161 patients with acute lymphoblastic leukemia at Peking University People’s Hospital. The study compared two stem cell transplant types: Matched Sibling Donor (MSDT) and Haploidentical Donor (Haplo-SCT).
While a naive analysis suggested only a marginal difference between the two, the proposed method revealed a significant risk difference of 0.1501. This indicates that for the "Always-Survivor" group, Haplo-SCT—which is more readily available than the rarer MSDT—actually results in a significantly lower relapse rate. This suggests Haplo-SCT is a highly effective alternative that may have been undervalued by simpler analytical methods.
Limitations and Future Directions
As with most causal inference models, this approach relies on assumptions that are often untestable in practice, such as the exclusion restriction for the substitutional variable and the monotonicity of survival. Furthermore, the model's accuracy depends on the distribution discrepancy of the substitutional variable across strata. The authors suggest that future research could explore the use of multiple negative control variables to further relax these assumptions and enhance the robustness of causal inference in complex real-world evidence scenarios.
Summary updated