Causal Inference With Outcomes Truncated by Death and Missing Not at Random

Wei Li, Yuan Liu, Shanshan Luo, Zhi Geng

TL;DR

This paper tackles a complex causal inference problem in clinical trials where outcomes are affected by both death (truncation) and missingness that depends on the unobserved outcome (missing not at random). The authors develop a novel method to identify, estimate, and bound the average treatment effect in a specific subpopulation (always-survivors) under these challenging conditions, using a proxy variable and a treatment-independent missingness assumption.

What Problem Does This Paper Address?

In many clinical trials, especially those with long follow-up periods, subjects may die before their outcomes can be measured. This is known as "truncation by death." Even for those who survive, their outcomes might not be recorded, and this missingness can be "not at random" (MNAR), meaning it depends on the unobserved outcome itself. This is common in studies of diseases like HIV, where patients with poorer health might be less likely to complete follow-up surveys, and their health status is directly related to the outcome being measured (e.g., CD4 cell count).

Standard causal inference methods struggle with these combined issues. Principal stratification is a useful framework for dealing with truncation by death, but it becomes complicated when outcomes are also MNAR. The challenge is to disentangle the effects of treatment from the effects of death and non-random missingness to get a reliable estimate of the treatment's true impact on those who would survive regardless of treatment.

Methods

The core of the paper lies in developing a method to identify, estimate, and bound the "survivor average causal effect" (SACE) – the average treatment effect among individuals who would survive under both treatment and control conditions.

  1. Identification: The authors introduce a key assumption: treatment-independent missingness. This means that while the outcome might be missing not at random, the treatment assignment itself does not directly influence why the outcome is missing, given the survival status and covariates. They also leverage a proxy variable (e.g., baseline CD4 count in the HIV study) that is related to the outcome only through the principal strata (i.e., the latent groups defined by survival potential). This proxy helps to bridge the gap in identifying the causal parameter.
  2. Estimation: They propose a three-step parametric estimation procedure that models:
  3. The proportions of different principal strata (always-survivors, never-survivors, compliers, defiers).
  4. The missingness mechanism.
  5. The outcome model.
  6. Bounds: When the identification assumptions are violated, the paper derives nonparametric bounds for the SACE. This provides a range within which the true causal effect must lie, even without strong assumptions.

Datasets

The paper uses the AIDS Clinical Trials Group (ACTG) Study 175 dataset for its empirical application. This is a real-world dataset from a randomized, double-blind trial investigating different HIV treatments. No new benchmarks are introduced; the focus is on applying the developed methodology to this existing dataset.

Key Insights and Ideas

  • The "always-survivor" subpopulation is crucial: Focusing on individuals who would survive regardless of treatment is key to obtaining a causally interpretable treatment effect when death is a factor.
  • Treatment-independent missingness is a powerful assumption: This assumption, combined with a proxy variable, allows for identification of the causal effect even when outcomes are MNAR. It essentially says that while the outcome might be missing due to unobserved factors, the choice of treatment doesn't directly influence whether those unobserved factors lead to missingness.
  • Proxy variables can unlock identification: A variable that influences the outcome only through latent groups (like principal strata) can act as a bridge to identify causal effects that would otherwise be unidentifiable.
  • Bounds provide robustness: When strong assumptions for identification are questionable, deriving bounds offers a valuable alternative, providing a range for the causal effect that is guaranteed to contain the true value.

Key Results

  • Identifiability Theorem: The paper establishes conditions under which the survivor average causal effect can be identified, relying on assumptions about treatment-independent missingness and the existence of a suitable proxy variable.
  • Estimation Procedure: A practical, three-step parametric estimation method is proposed and evaluated.
  • Nonparametric Bounds: Sharp upper and lower bounds for the SACE are derived, which are shown to be narrower than unadjusted bounds when covariates are incorporated.
  • Simulation Studies: The proposed estimator demonstrates negligible bias and reduced RMSE with increasing sample size, outperforming naive methods and a method that ignores MNAR issues. The coverage probabilities are close to the nominal level.
  • HIV Data Application: Applying the method to the ACTG Study 175 data suggests that while new HIV therapies showed some improvement in CD4 counts, the effect was not statistically significant for the always-survivor group, possibly due to the effectiveness of the ZDV-only regimen. The proposed method provided a more conservative estimate compared to methods that ignored missing data.

Limitations and Future Work

  • Assumption Sensitivity: The identification of the causal effect relies on assumptions like treatment-independent missingness and the existence of a proxy variable. The paper explores sensitivity to violations of these assumptions.
  • Parametric Models: While the estimation uses parametric models, the authors acknowledge that misspecification can lead to biased results. They binarize covariates in some analyses to mitigate this risk for bounds.
  • Extension to Continuous Outcomes: The current work focuses on binary outcomes. Extending the methods to continuous outcomes is suggested.
  • SUTVA Violations: The no-interference assumption (part of SUTVA) might not always hold in medical studies, especially for communicable diseases. Exploring violations of SUTVA is a future research direction.
  • Unmeasured Confounders: Assumption 1 assumes no unmeasured confounders between treatment and potential outcomes. Relaxing this assumption is also noted as a future research avenue.
  • Machine Learning Integration: For large numbers of covariates, exploring LASSO-based variable selection or other machine learning techniques to improve model flexibility is suggested.

Summary updated

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 analysis. In this article, 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 study.
BibTeX
@inproceedings{li2025causal,
  author = {Wei Li and Yuan Liu and Shanshan Luo and Zhi Geng},
  title = {{Causal Inference With Outcomes Truncated by Death and Missing Not at Random}},
  booktitle = {Statistics},
  year = {2025},
  doi = {10.1002/sim.70126},
}
Figure