Causal analysis of ordinal treatments and binary outcomes under truncation by death

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

Abstract

It is common that in multiarm 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.

Introduction

In clinical and social science research, "truncation by death" occurs when an outcome of interest can only be measured if a subject survives an intermediate event. A classic example is an HIV vaccine trial where viral load is only measurable in participants who become infected. If the vaccine affects the probability of infection, a simple comparison of viral loads between treatment and control groups suffers from selection bias. To address this, researchers typically focus on the Survivor Average Causal Effect (SACE)—the treatment effect within the subpopulation that would have "survived" (or experienced the intermediate event) regardless of their treatment assignment.

While SACE is well-defined for two-arm trials, extending it to multiarm trials (e.g., Placebo vs. Low Dose vs. High Dose) introduces significant complexities. These include non-identifiability, where causal effects can only be bounded rather than uniquely determined, and transitivity paradoxes, where pairwise comparisons lead to logically inconsistent conclusions.

Methodology: Principal Stratification and Monotonicity

The proposed framework utilizes Principal Stratification to categorize subjects into latent groups based on their potential survival responses across all treatment levels. To make the problem tractable, the authors invoke a monotonicity assumption, positing that increasing the treatment level does not decrease the probability of survival. This assumption reduces the number of possible principal strata.

Graphical representation of the relationship between latent principal strata parameters and observed conditional expectations.

As shown in the figure above, the observed conditional expectations $E[Y | Z=z, S=1]$ are modeled as mixtures of latent stratum-specific means ($\mu$). Because multiple latent parameters contribute to a single observed value, the causal effects are generally non-identifiable. The authors address this by using linear programming to derive sharp lower bounds for the maximum treatment effect ($\Delta_{max}$).

Simultaneous Testing and Bayesian Inference

A key contribution of this work is a "step-down" simultaneous testing procedure. Unlike traditional marginal tests that compare arms pairwise, this procedure tests the global null hypothesis by checking the compatibility of observed data with the null across all strata sequentially. To account for sampling uncertainty, a Bayesian approach using Dirichlet priors is employed, allowing researchers to calculate the posterior probability of a non-zero treatment effect.

Empirical Applications

The HVTN 503 (Phambili) Study

The framework was applied to the HVTN 503 HIV vaccine trial, where the treatment was the number of injections (0, 1, or 2+). While the trial was halted for lack of efficacy in preventing infection, the authors' simultaneous analysis found a 99.6% posterior probability of a treatment effect on CD4 counts (at the <200 cells/mm³ threshold). This suggests the vaccine altered the immune response in infected individuals, a finding that marginal pairwise tests were less equipped to confirm with certainty.

Simulation Results

In hypothetical three-arm vaccine simulations, the simultaneous test demonstrated vastly superior power. In one scenario with a small sample size ($n_1 = 36$), the simultaneous test achieved a 98.8% rejection rate of the null hypothesis, compared to only 4.0% for marginal testing. This underscores the information gain achieved by considering the joint constraints of all treatment arms simultaneously.

Conclusion

This research provides a robust framework for causal inference in multiarm trials with truncated outcomes. By shifting from marginal to simultaneous inference and leveraging the geometry of non-identifiable parameters, the authors offer a method that avoids transitivity traps and maximizes statistical power. Future work aims to extend this binary outcome framework to continuous and categorical data, as well as integrating baseline covariates to further sharpen causal bounds.

Summary updated

Abstract

It is common that in multiarm 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},
}
Figure