Survival Analysis Beyond Proportional Hazards

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2180 papers227 conceptsfrozen 2026-10-04

In clinical research and engineering, we often want to know not just if an event will happen, but when. This is the domain of survival analysis. For decades, the field has been dominated by the Cox Proportional Hazards model and its central metric: the Hazard Ratio (HR).

The Hazard Ratio is elegantly simple. It suggests that if Treatment A reduces the risk of death by 50% compared to Treatment B, that 0.5 multiplier applies today, tomorrow, and three years from now. This is the Proportional Hazards (PH) assumption. It assumes that the relative "speed" of the event occurring for two groups remains constant over time.

However, as Ziren Jiang, Satrajit Roychoudhury, and colleagues point out in The hazards of using hazard ratios from proportional hazard models in indirect treatment comparisons, this assumption is a mathematical convenience that frequently collapses when confronted with real-world data. When the assumption fails, the Hazard Ratio becomes a misleading average that can even lead to "non-transitive" conclusions—where the math suggests Treatment A is better than B, and B is better than C, but a direct trial might show C is better than A.

The Failure of Constant Risk

Why does the proportional hazards assumption fail? Intuition suggests that risks are rarely static. Consider a high-risk surgery: the risk of complications is massive in the first 24 hours but drops significantly after a week. Conversely, a drug might have a "delayed effect," showing no benefit for six months before suddenly diverging from the control group.

When these dynamics occur, survival curves can intersect. As Chirag Nagpal notes in Deep Cox Mixtures for Survival Regression, standard models are mathematically incapable of representing these "crossing" curves. If the relative hazard is not monotonic—meaning the risk ratio flips or fluctuates—a single HR number fails to capture the true clinical benefit.

Modeling the Entire Survival Curve

To move beyond static ratios, researchers are turning to more flexible frameworks that model the survival function itself rather than a fixed multiplier.

One approach is to use Restricted Mean Survival Time (RMST). As Kaiyuan Hua and Xiaofei Wang demonstrate in their work on Biomarker-Guided Adaptive Enrichment Design, RMST measures the area under the survival curve up to a specific time point. This provides a "life expectancy" interpretation that remains valid even when hazards are non-proportional or effects are delayed.

Another direction involves leveraging deep learning to capture time-varying interactions. Chirag Nagpal and Artur Dubrawski introduced Deep Survival Machines, which represent survival functions as a mixture of individual parametric distributions. This allows the model to learn complex, non-linear relationships between a patient's features and their risk over time without being shackled to the PH assumption.

Continuous Time and Neural Dynamics

If risk changes continuously, perhaps our models should too. Weijing Tang and Ji Zhu proposed SODEN, which uses Ordinary Differential Equation (ODE) networks to model the cumulative hazard function. By treating survival as a differential equation, the model can naturally handle non-proportional hazards where the relative risk between subjects evolves or crosses over time.

Similarly, Gabriele Campanella and Lucas Kook have explored Deep Conditional Transformation Models. These models use neural networks to parameterize transformation functions (often using Bernstein polynomials), allowing the impact of a feature to "evolve" throughout the observation period.

Finding Islands of Interpretability

While deep learning offers flexibility, it often sacrifices the transparency that clinicians value in the original Cox model. This has led to a "middle ground" research direction: identifying where the simple models actually work.

In Subgroup Discovery with the Cox Model, Zachary Izzo and Iain Melvin introduce a framework to find "islands of interpretability." Instead of forcing a global PH assumption on a messy dataset, they identify specific subgroups where the Cox model is exceptionally accurate. This allows researchers to use simple, interpretable hazard ratios where they are valid, while flagging more complex cases for flexible machine learning models.

Summary

The shift in survival analysis is a move from averages to dynamics. By moving beyond the proportional hazards assumption, researchers like Ted Westling and Marco Carone are developing tools that can handle the messy reality of clinical data—where treatments have delayed starts, risks fluctuate, and the "ratio" of survival is anything but constant. Whether through RMST, Neural ODEs, or mixture models, the goal is the same: a more faithful representation of time-to-event outcomes.

Papers in this version

Showing 60 of 180 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 150 papers
  2. Version 1 · 2026-10-04 · 180 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_proportional_hazards_assumption_c284e715,
  title = {proportional hazards assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proportional-hazards-assumption/rdr_c284e715c415439d8d689b2db735acf3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_survival_analysis_beyond_proportional_ha_c284e715_v1,
  title = {Survival Analysis Beyond Proportional Hazards},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/survival-analysis-beyond-proportional-hazards/rdr_c284e715c415439d8d689b2db735acf3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}