Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions

summary

The gist

Sample size calculations in two-arm clinical trials with survival endpoint often utilize the proportional hazards (PH) assumption…

In short

This work extends sample size calculations for two-arm clinical trials with survival endpoints to handle nonproportional hazards (NPH) and nonproportional time (NPT). It uses Weighted Relative Time (WRT) comparing two Weibull distributions to determine the required sample size, offering an interpretable metric based on improvement in longevity rather than instantaneous risk.

Key concepts

Nonproportional Hazards (NPH)
This occurs when the ratio of survival probabilities between two groups changes over time. Traditional methods fail here. This method addresses NPH by comparing survival curves at different percentiles to define a suitable effect size.
Weighted Relative Time (WRT)
WRT is a metric used as an effect size measure that changes across different percentiles (p). It compares the time-ratio between the control and treatment arms using two different Weibull distributions, providing a comprehensive assessment of longevity improvement.
Weibull Distribution
The Weibull distribution is a flexible statistical model used to describe survival times. This study employs two different Weibull distributions for the control and treatment groups to model their respective survival curves under nonproportional hazards.

Terminology used across episodes

This episode discusses

The paper

Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions · Read on arXiv

Milind A. Phadnis, Md Robiul Islam Talukder, Geethanjalee Mudunkotuwa, Matthew S. Mayo

Department of Biostatistics & Data Science, University of Kansas Medical Center

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions".

Jane: Here's the gist of the paper. Sample size calculations in two-arm clinical trials with survival endpoint often utilize the proportional hazards (PH) assumption, but this work extends a parametric approach that simultaneously allows both nonproportional hazards (NPH) and nonproportional time (NPT) by utilizing the concept of Weighted Relative Time (WRT) with two different Weibull distributions The biggest advantage of our proposed method is the ease of interpretability for both clinicians and patients who participate in the trial because it is based on the concept of improvement in longevity rather than reduction in instantaneous risk of an event Traditional sample size calculations often rely on the proportional hazards (PH) assumption or assuming an exponential distribution When analyzing survival data, adjustments accounting for NPH can be made using methods like piecewise exponential models or parametric accelerated failure time models that assume a constant time ratio (PT) assumption A key conceptual idea in designing a trial with NPH is deciding the most appropriate definition of effect size, advocating for the metric of Relative Time denoted by RT(p) – time-ratio that changes over percentile p – as a measure of the effect size in designing clinical trials This concept allows both NPH and NPT simultaneously The method involves setting up hypotheses based on user-defined percentiles p1, p2, etc., where the control arm survival time follows a Weibull distribution with shape parameter β0 and scale parameter θ0, and the treatment arm follows Weibull distribution with shape parameter β1 and scale parameter θ1 The test statistic of interest is defined as Qw = ∑wj(p j)k j=1 log(RT(p j)) (5) where the weights wj(p j) are chosen based on clinical considerations The variance of Qw is given by a complex expression involving covariance terms Under the null hypotheses of (3) or (4), we have Zw = Qw−muw√Var(Qw) N(0,1) Sample Size Calculation Steps The calculation proceeds through five steps: 1.

Tom: Elicit information about trial inputs like α, ω, pI, RT(pI), and RT(pII) 2.

Jane: Choose a weighting scheme wj(pj) based on clinical considerations 3.

Tom: Calculate d0j and d1j 4.

Title and authors: Tom: It’s about taking that old assumption—that the hazard ratio stays constant over time—and ditching it. They use this Weighted Relative Time concept to measure the effect based on how much longevity improves, instead of just looking at a single point in time risk.

Lu: Exactly, Tom. The core idea is that you can use two different Weibull distributions for your control arm and your treatment arm, even when they are nonproportional hazards and nonproportional time simultaneously. It lets the calculation account for how the risk changes throughout the whole study period in one go.

Jane: So this method aims to be flexible enough to handle those complicated curves without you having to build a separate model for every single shape of non-proportionality you might see in a trial. Is that what it does?

Tom: It does. They set up five clear steps for the calculation, starting with you giving them your desired confidence level and some initial relative time measurements. Then they choose how much weight to give different parts of the survival curve—early failures versus late failures—based on clinical needs.

Meng: Five steps sounds like a lot of manual work when I’m trying to deploy something that runs automatically on real patient data streams. What's the biggest practical benefit for someone actually running a trial?

Lu: The big thing is interpretability, Meng. It shifts the focus from just "reducing risk" to "improvement in longevity." That makes it much clearer for clinicians and patients what they are actually looking at when they see the sample size numbers.

Jane: That’s a really important shift. So instead of a number that says, "You need five hundred people to reduce the instantaneous risk by ten percent," it’s more like, "You need five hundred people to improve longevity by this much."

Tom: Right. And they did compare their Weighted Relative Time method against another test called the Max-Combo test, and they found that both of them hit their target error rate and power level quite reliably under different weighting schemes.

Meng: If the results are close to the target when you run simulations, does that mean it’s reliable enough for me to consider using it in my production pipeline?

Lu: It suggests the underlying math is sound for these specific scenarios, but we have to remember they are basing this on a specific assumption—that there is a monotonic trend in the hazard ratio. They aren't claiming it works perfectly for every single curve imaginable.

Jane: So, it’s a very powerful tool when you know that the hazard ratio is trending in one specific direction over time, which is where its strength lies.

Tom: So this paper gives us a concrete mathematical recipe for sample size calculation in complex survival settings without needing those messy piecewise models or making assumptions about constant time ratios. It’s a way to get precise numbers when the hazard isn't steady across the timeline.

Lu: It opens up a more flexible parametric approach to trial design, moving beyond just the proportional hazards assumption in a very concrete way. This is where we need to look next—how they extend this idea for more general forms of nonproportionality beyond that specific monotonic trend.

The paper's summary: Tom: So, what's the real value here beyond just calculating a number? The paper suggests improvements by making the weighting scheme flexible based on raw scores u j of p j.

Jane: That means researchers can actually tune how much they focus on early failures versus late ones in their sample size calculation, which directly lets them define the effect size as an improvement in longevity instead of just a reduction in risk.

Lu: That’s the point—it allows for a more targeted design. Instead of one fixed calculation, you can create different sample sizes depending on what clinical question you want to answer, by adjusting those raw scores u j.

Meng: So it gives researchers control over the sensitivity of their sample size based on their specific clinical priorities? That sounds useful if I’m trying to manage my computational load at scale.

Tom: It gives them control, yes. And when you look at the performance comparison, they showed that their Weighted Relative Time method performs very close to its target error rate and power level even across different weighting schemes they tested.

Jane: So it’s not just a theoretical construct; the empirical results show it’s empirically robust when tuned correctly, which is what we need to see before we trust a new calculation tool.

Lu: They are showing that this framework is competitive with other tests like the Max-Combo test, which gives confidence in its capability to handle nonproportional hazards and time simultaneously.

Meng: But what about the cost? If I have to run these complex weighting schemes five different ways, does that make the total calculation slow or computationally expensive for production?

Tom: The complexity is there because of those covariance terms, but they’ve done the work to define a clear five-step process, which should help streamline it. They are focused on making the concept practical for use in clinical settings.

Jane: It changes how we think about trial design by giving us a way to measure effect that aligns better with how patients experience treatment over time, not just some abstract risk metric.

Lu: The implication is that we can move away from relying entirely on constant hazard ratio assumptions and build more realistic trial designs directly into the sample size planning phase.

Tom: So, the takeaway is a more flexible, interpretable calculation method that lets you customize your sample size based on what clinical outcome matters most.

Jane: And with the future work mentioned—extending this to handle more general forms of nonproportionality—it suggests this WRT concept is really foundational for tackling these tricky survival problems.

The paper's improvements: Tom: So, to wrap up on "Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions," this paper gives us a parametric way to plan trials when hazards aren't steady over time.

Jane: It’s a solid framework for taking those tricky nonproportional hazard problems and giving us concrete numbers for sample sizes using something called Weighted Relative Time.

Lu: The main implication is that we can design more accurate trials without needing to rely on just assuming proportional hazards, especially when dealing with both nonproportional hazards and nonproportional time at the same time.

Meng: I still see a challenge in getting that weighting scheme tuned reliably across different real-world data sets. How hard is it to maintain those specific Weibull distribution assumptions when you’re just plugging in messy patient records?

Lalam: I'm confident because the simulation error they reported was tight—within zero point eight plus or minus zero point five—which suggests the method is quite stable under those conditions.

Tom: It means for people who only listen to the show, this paper is about getting a more realistic number for how big a study needs to be when treatment effects change over time.

Jane: Exactly, and it gives researchers tools like flexible weighting schemes based on raw scores u j of p j, which lets them tailor the calculation to focus on specific parts of the survival curve.

Lu: They are also looking into extending this framework to handle more general forms of nonproportionality beyond that specific monotonic trend they focused on, which is where the real ambition lies.

Meng: If we can make that extension computationally cheaper, that would be a big win for me because running those complex covariance term calculations at scale is a headache.

Tom: It’s definitely interesting work because it takes a very hard problem in survival analysis and builds a concrete, interpretable mechanism for solving it.

Jane: So moving forward, we’ll keep watching how this WRT approach compares to other methods like piecewise exponential models as we see more applications in clinical trials.

Conclusion: Tom: So we've been looking at "Sample size calculation comparing two survival curves in the presence of nonproportional hazards utilizing the concept of Weighted Relative Time using two different Weibull distributions." Basically, this paper gives us a parametric way to plan trials when your treatment effect isn't steady over time.

Jane: It’s a solid framework for taking those tricky nonproportional hazard problems and giving us concrete numbers for sample sizes using something called Weighted Relative Time.

Lu: The main implication is that we can design more accurate trials without needing to rely on just assuming proportional hazards, especially when dealing with both nonproportional hazards and nonproportional time at the same time.

Meng: I still see a challenge in getting that weighting scheme tuned reliably across different real-world data sets. How hard is it to maintain those specific Weibull distribution assumptions when you’re just plugging in messy patient records?

Lalam: I'm confident because the simulation error they reported was tight—within zero point eight plus or minus zero point five—which suggests the method is quite stable under those conditions.

Tom: It means for people who only listen to the show, this paper is about getting a more realistic number for how big a study needs to be when treatment effects change over time.

Jane: Exactly, and it gives researchers tools like flexible weighting schemes based on raw scores u j of p j, which lets them tailor the calculation to focus on specific parts of the survival curve.

Lu: They are also looking into extending this framework to handle more general forms of nonproportionality beyond that specific monotonic trend they focused on, which is where the real ambition lies.

Meng: If we can make that extension computationally cheaper, that would be a big win for me because running those complex covariance term calculations at scale is a headache.

Tom: It’s definitely interesting work because it takes a very hard problem in survival analysis and builds a concrete, interpretable mechanism for solving it. That’s what we're hearing from this paper today.

Jane: So moving forward, we’ll keep an eye on how this WRT approach compares to other methods like piecewise exponential models as we see more applications in clinical trials.

Lu: For future work, they are looking at extending this framework to handle more general forms of nonproportionality beyond the specific monotonic trend they focused on.

Meng: And for me, if we can figure out a way to make the calculation faster or less computationally expensive than those complex covariance terms suggest, that would be the next big thing.

Lalam: I’m confident because the simulation error they reported was tight—within zero point eight plus or minus zero point five—which suggests the method is quite stable under those conditions.

Tom: It’s certainly interesting work because it takes a very hard problem in survival analysis and builds a concrete, interpretable mechanism for solving it.

More episodes

← Home