On Defense of the Hazard Ratio

Andrew Ying, Ronghui XuarXiv 2023 · DOI · arXiv

Abstract

In this short communication, we describe the recent debate on whether the hazard function should be used for causal inference in time-to-event studies and consider three different potential outcomes frameworks (by Rubin, Robins, and Pearl, respectively) as well as use the single-world intervention graph to show mathematically that the hazard function has causal interpretations under all three frameworks. In addition, we argue that the hazard ratio over time can provide a useful interpretation in practical settings.

Introduction: The Hazard Ratio Controversy

In survival analysis, the Hazard Ratio (HR), typically estimated via the Cox proportional hazards model, has long been the standard for quantifying treatment effects. However, recent critiques from prominent researchers have suggested that the HR is fundamentally unsuitable for causal inference. The core of this "whistleblowing" argument is that the HR suffers from built-in selection bias, often characterized as collider bias.

The argument posits that because the "risk set"—the group of individuals who have not yet experienced the event—changes over time, the treated and control groups are no longer comparable after time zero, even in a perfectly randomized trial. This perceived bias is often discussed in the context of standard Directed Acyclic Graphs (DAGs) where the survival status at time $t$ is seen as a potential collider between treatment and unobserved covariates.

Standard DAG of survival analysis showing the relationships between covariates (L), treatment (A), and survival status over time.

Methods: Unifying Causal Frameworks

This paper provides a rigorous mathematical defense of the HR by re-evaluating it through three dominant causal inference frameworks: Rubin’s Potential Outcomes, Robins’ Framework, and Pearl’s Structural Causal Models.

  1. Rubin’s Potential Outcomes: The authors show that the hazard ratio can be expressed as a contrast between potential outcomes ($T_1$ and $T_0$), fitting Rubin's definition of a causal effect.
  2. Robins’ Framework: The hazard function is demonstrated to be a functional of the marginal distribution of counterfactual outcomes, qualifying it as a population causal effect.
  3. Pearl’s Structural Causal Models: Using $do$-calculus, the counterfactual hazard function is shown to be a valid causal quantity.

The most significant technical contribution is the application of Single-World Intervention Graphs (SWIGs). By splitting the treatment variable $A$ into a random component and an intervention component $a$, the authors prove that the potential at-risk process is not a collider, thereby refuting the claim of inherent selection bias.

Single-World Intervention Graph (SWIG) illustrating the intervention $a$ and the resulting potential outcomes $T_a$, demonstrating the lack of inherent selection bias.

Key Insights: The Collider Bias is an Illusion

The authors argue that the "collider bias" identified by critics is an artifact of looking at observed variables rather than potential outcomes. When shifting to the counterfactual perspective—considering what would happen to a subject under a specific intervention—the independence between treatment assignment and baseline characteristics remains intact. Specifically, they prove that $A \perp L | Y_a(t) = 1$, meaning treatment is independent of covariates given the potential at-risk status.

Furthermore, the paper defends the hazard as a measure of "velocity." Just as velocity describes a dynamic physical process, the hazard function describes the instantaneous rate of failure. Even if this rate changes over time, it remains a valid and interpretable measure of treatment impact.

Results and Practical Implications

The paper validates the use of time-varying hazard ratios $\beta(t)$. Rather than being misleading averages, these are essential for understanding treatments with evolving risk profiles, such as organ transplantations or vaccine effectiveness over time. For instance, in COVID-19 vaccine studies, time-varying HRs are crucial for estimating current risk as immunity wanes or variants emerge.

The authors also critique proposed alternatives like the "causal hazard ratio," noting that they often rely on joint distributions of potential outcomes ($T_1, T_0$) which are unobservable and thus practically unusable. In contrast, the standard HR remains a parsimonious and sensitive tool, often outperforming metrics like Restricted Mean Survival Time (RMST) in detecting early treatment effects.

Conclusion

By reconciling the hazard ratio with the major frameworks of causal inference, this work restores the HR's status as a valid tool for causal discovery. While acknowledging that unobserved heterogeneity (frailty) is a real phenomenon, the authors argue it should be treated as a data feature rather than a reason to abandon the metric. Future research may extend this defense to multi-state systems and more complex censoring mechanisms.

Summary updated

Abstract

In this short communication, we describe the recent debate on whether the hazard function should be used for causal inference in time-to-event studies and consider three different potential outcomes frameworks (by Rubin, Robins, and Pearl, respectively) as well as use the single-world intervention graph to show mathematically that the hazard function has causal interpretations under all three frameworks. In addition, we argue that the hazard ratio over time can provide a useful interpretation in practical settings.
BibTeX
@misc{ying2023defense,
  author = {Andrew Ying and Ronghui Xu},
  title = {{On Defense of the Hazard Ratio}},
  year = {2023},
  eprint = {2307.11971},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.11971},
}
Figure