Abstract
In clinical trials, principal stratification analysis is commonly employed to address the issue of truncation by death, where a subject dies before the outcome can be measured. However, in practice, many survivor outcomes may remain uncollected or be missing not at random, posing a challenge to standard principal stratification analyses. In this paper, we explore the identification, estimation, and bounds of the average treatment effect within a subpopulation of individuals who would potentially survive under both treatment and control conditions. We show that the causal parameter of interest can be identified by introducing a proxy variable that affects the outcome only through the principal strata, while requiring that the treatment variable does not directly affect the missingness mechanism. Subsequently, we propose an approach for estimating causal parameters and derive nonparametric bounds in cases where identification assumptions are violated. We illustrate the performance of the proposed method through simulation studies and a real dataset obtained from a Human Immunodeficiency Virus (HIV) study.
Introduction
Long-term clinical trials, particularly those involving severe conditions like HIV or cancer, frequently encounter two significant data challenges: truncation by death and missingness not at random (MNAR). Truncation by death occurs when a subject dies before follow-up, rendering their outcome undefined. This creates a selection bias, as comparing survivors across treatment arms is often an "apples-to-oranges" comparison—the treatment may have kept sicker individuals alive who would have died in the control group.
Furthermore, even among survivors, data may be missing because patients feel too ill to attend follow-up appointments. This is a classic case of MNAR, where the missingness mechanism is directly related to the unobserved outcome. While existing methods often address one of these issues, this paper provides a unified framework to handle both, specifically targeting binary outcomes.
Methodology: Principal Stratification and Identification
The core of the proposed approach is Principal Stratification, which focuses on the "always-survivors" ($G = AT$)—the subpopulation of individuals who would survive regardless of their treatment assignment ($Z$). To identify the Survivor Average Causal Effect (SACE), the authors introduce two critical requirements:
- Proxy Variable ($A$): A baseline measurement (e.g., initial CD4 count) that predicts a patient's principal stratum but does not directly influence the change in outcome post-treatment.
- Treatment-Independent Missingness: An assumption that while health status may cause dropout (MNAR), the treatment assignment itself does not directly cause missingness once health and other covariates are accounted for.
As illustrated in Figure 1, the observed data patterns (treatment $Z$, survival $S$, and response $R$) map to different latent principal strata ($G$). Under the monotonicity assumption ($S(1) \ge S(0)$), which posits that treatment does not decrease the probability of survival, the "Harmed" or "Defier" (DE) stratum is eliminated. This simplification allows for clearer identification of the "Always-survivor" (AT) group from the observed data.
Results and Empirical Application
Simulation Studies
Extensive simulations with varying sample sizes (500 to 5000) demonstrated that the proposed estimator remains unbiased and achieves target 95% coverage probabilities. In contrast, "naive" estimators that ignore either death or MNAR exhibited significant bias.
HIV Study: ACTG 175
The method was applied to the ACTG Study 175, which compared zidovudine (ZDV) monotherapy to combination therapies in 2,139 HIV-I infected patients. The outcome was the improvement of CD4 T-cell counts after 96 weeks.
- SACE Estimate: The proposed method yielded an effect of 0.151 (95% CI: [-0.086, 0.275]). Since the interval includes zero, the results suggest that for the "always-survivor" group, combination therapies did not provide a statistically significant advantage over ZDV monotherapy.
- Bias Correction: Methods ignoring MNAR produced a higher effect estimate of 0.226. The authors argue this reflects an upward bias caused by failing to account for the fact that sicker patients (with poorer outcomes) were more likely to drop out.
Conclusion
This research provides a robust toolkit for clinical trialists facing complex data attrition. By utilizing proxy variables and establishing nonparametric sharp bounds, researchers can obtain more reliable estimates of treatment effects even when standard assumptions are challenged. Future work will look to extend this framework from binary to continuous outcomes and integrate machine learning techniques for high-dimensional covariate adjustment.
Summary updated