Identification and estimation of causal effects with outcomes truncated by death

Linbo Wang, Xiao‐Hua Zhou, Thomas S. Richardson

TL;DR

This paper introduces a novel framework for estimating the Survivor Average Causal Effect (SACE)—the effect of a treatment on individuals who would have survived regardless of their treatment assignment. By using "substitution variables" (baseline covariates that predict survival but don't directly affect the outcome), the authors provide a way to overcome the selection bias that occurs when outcomes are "truncated by death."

What Problem Does This Paper Address?

In clinical trials and observational studies, researchers often want to measure a non-mortality outcome, such as quality of life or cognitive decline. However, if a participant dies before the follow-up, their outcome is undefined—it isn't just "missing," it's truncated.

The trap many researchers fall into is comparing only the survivors in the treatment group to the survivors in the control group. This leads to selection bias. For example, if a new drug keeps the sickest patients alive, the "survivor group" in the treatment arm will look "sicker" on average than the control group, making the drug appear less effective than it actually is.

To fix this, we look for the effect within the "always-survivor" group: the subset of people who would have survived whether they received the treatment or the control. The challenge? We can never truly know who belongs to this latent group because we only observe one reality for each person.

Methods

The authors build on the framework of Principal Stratification. Their key innovation is the introduction of a substitution variable ($A$).

  1. The Substitution Variable: Think of this as a cousin to an Instrumental Variable (IV). It is a baseline covariate that is highly predictive of a person's "survival type" (e.g., whether they are an always-survivor) but has no direct effect on the final outcome ($Y$) except through that survival status.
  2. Identification Assumptions: The paper outlines several conditions, most notably the Exclusion Restriction (the substitution variable doesn't directly change the outcome) and Substitution Relevance (the variable actually helps distinguish survival types).
  3. Relaxing Assumptions: Unlike previous work, this paper provides paths to relax the common "Monotonicity" assumption (the idea that treatment never causes death) and the "Exclusion Restriction," making the model more robust for real-world scenarios where these assumptions might be shaky.
  4. Model Parameterization: They propose specific mathematical constraints (M.1–M.4) to ensure that the estimated probabilities and effects stay within logical bounds (e.g., preventing negative probabilities), which is a common failure point in simpler models.

Datasets

  • Simulated Data: The authors used 500 simulated datasets to test their estimators against "naive" regressions and previous state-of-the-art methods (like Ding et al., 2011).
  • Southwest Oncology Group (SWOG) Trial: A real-world randomized Phase III trial comparing two treatments (docetaxel + estramustine vs. mitoxantrone + prednisone) for men with metastatic prostate cancer. The outcome of interest was the change in "health-related quality of life" one year later.

Key Insights and Ideas

  • The "Always-Survivor" focus: The paper reinforces that the only group for whom a causal effect is well-defined is the "always-survivor" group. For everyone else, the outcome under at least one treatment condition is literally non-existent.
  • Propensity Scores Aren't Enough: A major "aha" moment is the explanation of why standard propensity score methods fail here. In this context, baseline covariates can be common causes of the treatment, survival, and the outcome. Standard methods can't disentangle these overlapping influences.
  • Sensitivity Analysis: Because the SACE is notoriously difficult to identify, the authors provide a sensitivity analysis framework (using a parameter $\rho$) to show how results change if the "Monotonicity" assumption is violated.

Key Results

  • Consistency in Simulations: In scenarios with heavy confounding, the proposed "Prop-ER" (Exclusion Restriction) and "Prop-NI" (No Interaction) methods remained consistent and accurate, while naive methods showed significant bias.
  • Re-evaluating Clinical Trials: When applying their method to the SWOG prostate cancer data, the authors found that the previously reported "significant" effect of docetaxel on quality of life became much smaller and statistically insignificant once baseline covariates were properly accounted for. This suggests that earlier methods may have been overestimating the drug's benefits by failing to account for the "always-survivor" latent structure.

Limitations and Future Work

  • Choice of Variable: The success of this method hinges entirely on finding a good substitution variable. If the variable chosen actually has a direct effect on the outcome, the results will be biased.
  • Binary Focus: The current work focuses on binary exposures (Treatment A vs. Treatment B). The authors note that generalizing this to ordinal or continuous exposures (like different dosages of a drug) is a necessary next step.
  • Sensitivity to $\rho$: While they provide a sensitivity analysis, the true value of the association between potential survivals ($\rho$) remains unobservable, meaning researchers must still rely on expert judgment to interpret the bounds.

Summary updated

Abstract

It is common in medical studies that the outcome of interest is truncated by death, meaning that a subject has died before the outcome could be measured. In this case, restricted analysis among survivors may be subject to selection bias. Hence, it is of interest to estimate the survivor average causal effect, defined as the average causal effect among the subgroup consisting of subjects who would survive under either exposure. In this paper, we consider the identification and estimation problems of the survivor average causal effect. We propose to use a substitution variable in place of the latent membership in the always-survivor group. The identification conditions required for a substitution variable are conceptually similar to conditions for a conditional instrumental variable, and may apply to both randomized and observational studies. We show that the survivor average causal effect is identifiable with use of such a substitution variable, and propose novel model parameterizations for estimation of the survivor average causal effect under our identification assumptions. Our approaches are illustrated via simulation studies and a data analysis.
BibTeX
@inproceedings{wang2017identification,
  author = {Linbo Wang and Xiao‐Hua Zhou and Thomas S. Richardson},
  title = {{Identification and estimation of causal effects with outcomes truncated by death}},
  booktitle = {Biometrika},
  year = {2017},
  doi = {10.1093/biomet/asx034},
}

Loading related content...

Figure