📊 statistics

Recovering the Target Hazard Ratio Under Nonproportional Hazards Induced by an Omitted Covariate: Simulation-based Approach

This paper proposes and validates a simulation-based method that recovers the target hazard ratio from a correctly specified model by identifying the value that generates the tightest band of survival curves enclosing observed data, effectively addressing bias caused by omitted covariates and nonproportional hazards under minimal assumptions.

Original authors: Jong-Hyeon Jeong

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jong-Hyeon Jeong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Mystery of the Vanishing Effect

Imagine you are a detective trying to solve a crime, but you only have half the clues. In the world of medical research, scientists often run "clinical trials" to see if a new medicine works better than a fake one (a placebo). They use a special mathematical tool called the Cox model to measure how much the medicine slows down a bad event, like a tumor growing back. This tool relies on a rule called proportional hazards, which is a fancy way of saying: "The medicine works at a steady, constant rate from day one to the last day of the study." If the medicine cuts the risk of failure in half, it should cut it in half every single day, forever.

But real life is messy. Sometimes, a crucial piece of information is missing from the detective's file. Maybe there is a hidden biological marker in the patients' blood that the researchers didn't measure, or maybe the patients didn't take their pills exactly as told. When this happens, the "steady rate" rule breaks. The medicine might look super effective at the start, but then its power seems to fade away over time, or it might look like it's working when it's actually not. This is called nonproportional hazards. When this happens, the standard math tools get confused and give a wrong answer about how good the medicine really is. The big question for scientists is: How can we find the true strength of the medicine when our data is missing a key ingredient and the rules have changed?

The "Tightest Band" Detective Work

In this paper, Dr. Jong-Hyeon Jeong proposes a clever, simulation-based detective method to recover the true answer, even when a key piece of data is missing. The paper suggests that when we forget to include an important factor (like a hidden biomarker), the math gets messy and the estimated benefit of the treatment becomes biased. Previous attempts to fix this relied on complex formulas that assumed we knew exactly how the missing data was distributed, which is rarely true in the real world. Dr. Jeong argues that these old, rigid formulas are often impractical and can lead us astray.

Instead of trying to solve a complicated equation with missing pieces, the author suggests a "guess-and-check" game played on a computer. The core idea is to create a Tightest Survival Band (TSB). Imagine you have a set of survival curves from a real clinical trial (the "observed" data), which look a bit wobbly and don't follow a perfect straight line. Now, imagine you have a computer that can generate thousands of fake trials based on different guesses about how strong the medicine really is and how the missing data behaves.

The goal is to find the specific "true" numbers that make the computer's fake survival curves form a tight, neat band that perfectly wraps around the wobbly real-life curves. If your guess for the medicine's strength is too high or too low, the real curves will poke out of the band. But if you hit the jackpot—the exact "target hazard ratio"—the real data will sit snugly inside the simulated band, like a secret message hidden inside a box.

The paper runs extensive simulations to test this idea. It finds that while trying to guess the distribution of the missing data (using models like "gamma" or "log-normal" frailty) helps a little bit, it isn't a perfect fix. However, the new Tightest Survival Band method works remarkably well. In the simulations, when the researchers used this method, they could recover the true strength of the treatment with high accuracy, even when the missing data followed different, unknown patterns. The method is robust, meaning it doesn't matter if the hidden factor was distributed one way or another; the "tightest band" still finds the right answer.

To prove it works in the real world, the author applied this method to data from a famous breast cancer trial (the NSABP B-14 study). In this trial, the standard math suggested the treatment effect faded over time. By using the TSB method, the researchers were able to "recover" a target hazard ratio that made sense, suggesting the treatment had a consistent, strong effect that was just obscured by the missing data. The paper concludes that this simulation-based approach is a practical, powerful tool for untangling the mess caused by missing information, offering a way to see the true picture when the standard rules of the game have been broken.

Technical Summary: Recovering the Target Hazard Ratio Under Nonproportional Hazards Induced by an Omitted Covariate

Problem Statement
In survival analysis, the Cox proportional hazards (PH) model relies on the assumption that hazard ratios remain constant over time. However, this assumption is frequently violated in practice due to factors such as unpredictable treatment onset, patient noncompliance, or the omission of a covariate that, if included, would restore proportionality. When a relevant covariate $Z$ is omitted from a PH model, the estimated treatment effect $\beta$ becomes biased. This bias arises from two sources: (1) marginalization over the distribution of the omitted covariate and (2) model misspecification resulting from fitting a PH model to data that inherently exhibits nonproportional hazards.

The paper focuses on the scenario where the omitted covariate represents an unobservable biomarker, a stratification factor, or a strong prognostic factor. The goal is to recover the "target hazard ratio," defined as the hazard ratio that would be obtained from a correctly specified PH model including the omitted covariate. While theoretical frameworks exist to correct this bias, they often rely on strong assumptions (e.g., known censoring mechanisms or specific frailty distributions) that are impractical for real-world data applications.

Methodology
The paper proposes a simulation-based graphical approach known as the Tightest Survival Band (TSB) method to recover the target hazard ratio without relying on direct parameter estimates from the observed data or complex analytic bias corrections.

Model Framework: The study considers three models:
- Model I: The correctly specified PH model including the omitted covariate $Z$ .
- Model II: A misspecified PH model where the omitted covariate is absorbed into the baseline via a univariate frailty term $W = \exp(\theta Z)$ .
- Model III: A marginal model (univariate frailty model) that explicitly accounts for the distribution of the omitted covariate.
The TSB Procedure:
The core of the proposed method is an optimization procedure based on minimal assumptions: a Weibull baseline event-time distribution and a unit variance for the omitted covariate.
- Step 1: Compute Kaplan-Meier estimates for the observed data, stratified by treatment group.
- Step 2: For a grid of candidate parameters (treatment effect $\beta$ , omitted covariate effect $\theta$ , and Weibull parameters $\sigma, \rho$ ), simulate $N$ datasets. The simulation generates event times based on the Weibull baseline and the candidate parameters, incorporating a frailty term (either Gamma or Log-Normal) and a censoring mechanism (uniform or mixture of uniforms).
- Step 3: For each candidate parameter set, compute the Simulated Pointwise Survival Range (SPSR) for each treatment group. This band represents the minimum and maximum survival curves observed across the $N$ simulated replicates.
- Step 4: Calculate a Coverage Coefficient (CC), defined as the proportion of time points where the observed survival curves fall within the simulated SPSR bands for both treatment groups.
- Step 5: Identify the parameter values that maximize the CC. The resulting band is the "Tightest Survival Band." If the maximum CC is achieved over a flat region (a plateau), the midpoint of that region is used as the recovered parameter estimate.

Key Contributions and Results

Evaluation of Frailty Models: Simulation studies compared the performance of univariate Gamma and Log-Normal frailty models. The results indicated that while modeling the exponentiated omitted covariate via a frailty term can recover the target hazard ratio to some extent, the performance is highly dependent on the true underlying distribution. Specifically, assuming a Log-Normal distribution for the frailty term proved more robust across different scenarios than the Gamma assumption.
Performance of TSB: The simulation studies demonstrated that the TSB method recovers the target hazard ratio reasonably well under practical scenarios involving various true hazard ratios, Weibull baseline parameters, and censoring mechanisms.
- The method performs well for moderate to large sample sizes (e.g., $n \ge 500$ per group).
- As sample size decreases, the optimization surface (CC) becomes flatter, leading to wider intervals of plausible parameter values, though the midpoint of these intervals remains close to the true value.
- The method is robust to misspecification of the frailty distribution (i.e., using a Gamma working model when the truth is Log-Normal, or vice versa).
Application to Real Data: The method was applied to the NSABP B-14 breast cancer clinical trial, where the proportional hazards assumption was violated ( $p=0.00018$ $p = 0.00018$ ).
- Naive analysis (ignoring the omitted covariate) yielded a hazard ratio of 0.75.
- Univariate frailty analyses yielded varying estimates (0.48 for Gamma, 0.74 for Log-Normal).
- The TSB procedure, using an exhaustive grid search, recovered a target hazard ratio of approximately 0.68–0.69 (log-hazard ratio $\approx -0.37$ to $-0.39$) under both Gamma and Log-Normal working models. This result was consistent across both frailty assumptions, suggesting the method successfully identified a stable target parameter despite the unknown true distribution of the omitted covariate.
- In contrast, an existing analytic correction formula (Henderson and Oman) produced a hazard ratio of 0.21, which the paper notes was far from the TSB results.

Significance and Claims
The paper claims that the proposed TSB approach offers a practical, simulation-based alternative to analytic bias correction methods that often fail in real-data applications due to reliance on unknown population parameters. By identifying the target parameters that generate a survival band enclosing the observed data as tightly as possible, the method provides a way to recover the target hazard ratio under minimal assumptions.

The authors emphasize that while the method relies on assumptions (unit variance of the omitted covariate, Weibull baseline, and specific censoring structures), these are practically motivated. The study concludes that the TSB method is a viable tool for reducing bias in treatment effect estimates when nonproportional hazards are induced by omitted covariates, particularly in settings with moderate to large sample sizes where analytic solutions are intractable or sensitive to sampling variability. The authors modestly note that the optimization metric (CC) may produce flat regions rather than unique maximizers, necessitating the use of midpoints or medians as summary estimates, and suggest that developing alternative optimization criteria with stronger theoretical justification is a direction for future research.

The Mystery of the Vanishing Effect

The "Tightest Band" Detective Work

Technical Summary: Recovering the Target Hazard Ratio Under Nonproportional Hazards Induced by an Omitted Covariate

More like this