Causal Analysis of Ordinal Treatments and Binary Outcomes Under Truncation by Death

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

TL;DR

This paper introduces a novel framework for causal inference in multiarm trials where outcomes are "truncated by death"—meaning they are only observable if a subject survives an intermediate event (like infection or mortality). By leveraging a monotonicity assumption and a "step-down" simultaneous testing procedure, the authors provide a way to detect treatment effects across multiple dosage levels even when those effects cannot be perfectly identified.

What Problem Does This Paper Address?

In many clinical and social science trials, we only care about an outcome if a specific intermediate event happens. For example, in an HIV vaccine trial, researchers want to know if a vaccine lowers viral load—but viral load can only be measured in people who actually get infected. If the vaccine itself changes the likelihood of infection, comparing the viral loads of infected people in the treatment vs. control groups introduces selection bias.

This is known as truncation by death. To solve this, researchers look at the "Survivor Average Causal Effect" (SACE): the effect of the treatment on the subpopulation that would have "survived" (or been infected) regardless of whether they received the treatment or the control.

While this is well-studied for two-arm trials (Treatment vs. Control), most real-world trials are multiarm (e.g., Placebo, Low Dose, High Dose). Multiarm trials introduce two major headaches:

  1. Non-identifiability: We can't observe what would have happened to the same person under every possible dose, so we can only calculate "bounds" for the causal effect rather than a single number.
  2. Transitivity Paradoxes: Comparing "coarsened" groups can lead to illogical conclusions (e.g., Dose 2 looks better than Dose 1, and Dose 1 looks better than Placebo, but Placebo looks better than Dose 2).

Methods

The authors propose a framework based on Principal Stratification, which categorizes subjects into groups (strata) based on how they would respond to all possible treatment levels.

Key components include:

  • Monotonicity Assumption: They assume that increasing the treatment level (e.g., higher incentive or higher dose) does not decrease the probability of survival. This simplifies the number of possible "principal strata."
  • Step-Down Algorithm (Algorithm 1): A novel procedure for testing the "global null hypothesis" (the idea that the treatment does nothing in any group). It checks for compatibility between the observed data and the null hypothesis in a sequence, starting from the most basic stratum.
  • Linear Programming: To handle the fact that parameters are non-identifiable, the authors use linear programming to find the sharp lower bound of the maximum treatment effect ($\Delta_{max}$).
  • Bayesian Inference (Algorithm 2): They use Dirichlet priors to account for sampling uncertainty. This allows them to calculate the posterior probability that a non-zero treatment effect exists.

Datasets

The paper evaluates its methods using:

  • HVTN 503 (Phambili) Study: A real-world HIV vaccine trial in South Africa. The "treatment" was the number of vaccine injections (0, 1, or 2+), the "truncation" was HIV infection, and the "outcome" was whether the median CD4 count fell below specific clinical thresholds (350 or 200 cells/mm³).
  • Survey Incentive Trials (Conceptual): The authors discuss a study by Curtin et al. (2007) where different cash incentives ($0, $5, $10) were used to increase survey response rates (survival) to measure response quality (outcome).
  • Hypothetical Simulation: A three-arm vaccine trial simulation used to demonstrate that their simultaneous testing method has significantly more power than traditional marginal (pairwise) tests.

Key Insights and Ideas

  • Simultaneous vs. Marginal Inference: The biggest "aha moment" is that testing all treatment arms together (simultaneously) is much more powerful than doing pairwise comparisons. Because the parameters are non-identifiable, the advantage of simultaneous testing doesn't disappear even if you have an infinite amount of data.
  • The Power of "Sharp" Tests: In non-identifiable settings, you can't get a "consistent" test (one that always finds the truth with enough data). Instead, the authors aim for "sharpness"—a test that rejects any hypothesis that is mathematically incompatible with the observed data distribution.
  • Avoiding the Transitivity Trap: By restricting comparisons to the "finest" principal strata (people who would survive under all treatment levels), they ensure that their causal conclusions remain logically consistent across multiple arms.

Key Results

  • HVTN 503 Analysis: The simultaneous method found a 99.6% posterior probability of a non-zero treatment effect on CD4 counts (at the <200 threshold). In contrast, marginal methods were less certain. This suggests the vaccine did indeed affect the immune systems of those who became infected, even though the trial was stopped early for lack of efficacy in preventing infection.
  • Simulation Power: In one simulation scenario ($n_1 = 36$), the simultaneous test had a 98.8% rejection rate of the null hypothesis, while the marginal testing method only rejected it 4.0% of the time. This highlights how much information is lost when you don't consider the joint constraints of all treatment arms.

Limitations and Future Work

  • Binary Outcomes Only: The current framework is restricted to binary outcomes (e.g., "Did CD4 count fall below 200? Yes/No"). The authors note they are exploring extensions for continuous and categorical data.
  • Monotonicity Sensitivity: The results rely heavily on the monotonicity assumption. While plausible in survey incentives, it can be controversial in medical trials where a high dose might increase side effects and actually decrease survival.
  • Covariates: The paper does not deeply integrate baseline covariate information (like age or health status), which the authors admit could be used to "sharpen" the bounds of the causal effects even further.

Summary updated

Abstract

It is common that in multi-arm randomized trials, the outcome of interest is "truncated by death," meaning that it is only observed or well-defined conditioning on an intermediate outcome. In this case, in addition to pairwise contrasts, the joint inference for all treatment arms is also of interest. Under a monotonicity assumption we present methods for both pairwise and joint causal analyses of ordinal treatments and binary outcomes in presence of truncation by death. We illustrate via examples the appropriateness of our assumptions in different scientific contexts.
BibTeX
@article{wang2016causal,
  author = {Linbo Wang and Thomas S. Richardson and Xiao‐Hua Zhou},
  title = {{Causal Analysis of Ordinal Treatments and Binary Outcomes Under Truncation by Death}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2016},
  doi = {10.1111/rssb.12188},
}

Loading related content...

Figure