How evolving oncology trial designs are reshaping survival extrapolation for HTA 

The ISPOR Europe 2025 session, “Survival at the Margins: Navigating HTA Challenges in Modern Oncology Trials,” focused on how modern oncology studies, shaped by basket and multicohort designs, treatment switching, immature survival data, and limited comparators, are exposing limitations in the conventional survival extrapolation methods used in health technology assessment (HTA). Chaired by Hugo Pedder (University of Bristol; ConnectHEOR), with speakers Arthur Allignol (Daiichi Sankyo), Min-Hua Jen (Eli Lilly) and Ash Bullement (Petauri), the panel examined methodological issues and practical strategies for improving long-term survival estimates in increasingly complex evidence settings. 

Opening the session, Jen outlined the purpose of the ISPOR Oncology Special Interest Group and the need for continued methodological dialogue, noting that its mission is “to identify emerging trends and methodological challenges in oncology” and to translate this work into “practical impact” for decision-making. A live poll highlighted extrapolation, immature data, and treatment switching as the most pressing statistical challenges in oncology HTA. Reflecting on the results, Pedder observed that “we're seeing a lot here with extrapolation, immature data and treatment switching,” setting the focus for a session centered on the credibility of long-term survival estimates in increasingly non-traditional trial settings. 

Understanding survival extrapolation challenges in basket trials 

Allignol focused on survival extrapolation in basket trials, which evaluate targeted therapies across multiple tumor types or indications that share a biomarker. These studies are typically conducted under a single protocol and often use Phase 2, single-arm designs. While such approaches are efficient and can support tumor-agnostic development, they pose significant methodological challenges for HTA. Sample sizes within each basket are small; follow-up is often limited to 1–3 years, and there are usually no control arms. At the same time, HTA bodies expect both long-term survival estimates and tumor- or indication-specific results. 

He described the resulting bias–variance trade-off. Analyzing each tumor independently respects heterogeneity but leads to highly uncertain estimates, while pooling all tumors reduces variance at the risk of bias for particular indications. As he put it: 

“You end up between a rock and a hard place: you can either analyze the tumors independently, which leads to a lot of uncertainty and high variance, or you can pool everything, which will reduce variance but may introduce bias if certain baskets differ from the average population.” 

As a potential middle ground, Allignol pointed to Bayesian hierarchical models, which allow “partial borrowing” of information across baskets. These models assume a population-level distribution for parameters such as survival shape and scale, with basket-specific estimates treated as draws from that distribution. The extent of shrinkage toward the overall mean depends on factors such as sample size, within-basket variability, and how extreme a given basket is relative to others. A critical requirement, he stressed, is the assumption of exchangeability: 

“This exchangeability assessment assumption should be assessed… you need also to assess that, of course, clinically, whether this assumption is plausible.” 

He also noted that even when extrapolation for the experimental arm is stabilized, comparative effectiveness remains difficult in single-arm, tumor-agnostic settings. Possible approaches include comparisons between responders and non-responders, use of outcomes from prior lines of therapy, or incorporation of external controls and population-adjusted indirect comparisons, each of which introduces additional assumptions and limitations. 

Using external evidence and Bayesian methods to address immature survival data 

Returning to the podium, Min-Hua Jen presented a case study involving a trial with treatment switching built into the design and immature overall survival (OS), focusing on how her team approached extrapolation after adjusting for switching. She highlighted a core challenge in oncology HTA, noting that many trials conclude before OS data mature, making long-term projections highly uncertain and often dominant in cost-effectiveness assessments: 

“Many oncology trials conclude before the OS data mature, making long-term survival projections highly uncertain.”  

Jen cautioned that standard parametric models fitted solely to trial data risk capturing trial artefacts rather than underlying disease biology, particularly when trial populations are younger and healthier than patients seen in practice. Despite this, she noted that many HTA bodies still expect a single parametric distribution for extrapolation, which can lack credibility when OS data are immature. 

To address this, she outlined two complementary approaches: expert elicitation incorporated into Bayesian survival models, and structured use of mature external OS evidence from sources such as registries, historical trials, or real-world databases. In the case study, her team combined trial data, external registry data for both experimental and control arms, and general population mortality within a flexible Bayesian survival framework. 

Summarizing the lessons learned, Jen stressed that “external data are not a shortcut” but rather “a structured way to learn from broader evidence” when trial data are limited. She also emphasized the importance of transparency, arguing that this approach makes extrapolation assumptions explicit and aligns with HTA expectations around justification and clinical plausibility, particularly in early oncology launches, rare indications, and single-arm evidence settings. 

Comparing survival extrapolation methods across evidence sources 

Bullement broadened the discussion to consider how different forms of external evidence can be used in survival extrapolation and how available methods compare. Introducing himself as a “statistically enthusiastic health economist,” he explained that the goal is to generate “plausible estimates over a lifetime horizon” in settings where “trials are short, and so therefore we need to extrapolate for longer.” 

He defined external evidence broadly as “any data or information that's from outside of your main clinical trial” that may inform HTA decision-making, including life tables, registries, real-world data (RWD), earlier studies, and expert opinion. Bullement reviewed approaches ranging from simple piecewise models to more complex Bayesian frameworks, noting that most methods are typically assessed in isolation. 

To address this, he described a simulation-based comparison study designed to evaluate methods neutrally:  

“I don't have a horse in the race… I just want to know how they perform.”  

The analysis showed that no single method consistently outperformed others; effectiveness depended on context, data maturity, and the relevance of external sources. A central takeaway was that “garbage in, garbage out applies here,” as poor-quality external data can lead to less credible extrapolations than trial data alone. 

Bullement also defended the continued role of simpler approaches, observing that even “a really neat, simple piecewise model… gets the job done, and everybody understands that.” More complex methods, he concluded, should be adopted when they clearly add value over transparent, well-understood baselines.

Reflections on methodological complexity, evidence integration, and limits of current survival models 

The panel discussion and audience Q&A reinforced several cross-cutting themes. One audience member cautioned against searching for a single, universal solution, arguing that: 

“We should stop looking for the holy grail of the one method that does everything we need from the one data set, because that doesn't exist.”  

Instead, participants acknowledged that methodological complexity often reflects clinical reality and must be addressed accordingly. 

Other questions focused on how best to incorporate general population mortality into economic models and whether relative survival or excess hazard approaches could play a larger role. While these frameworks are compatible with some of the Bayesian methods discussed, speakers noted that they remain difficult to implement within standard HTA modelling environments. 

Across discussions of basket trials, treatment switching, and the structured use of external evidence, the session converged on a shared conclusion: there is no single best method for survival extrapolation in modern oncology HTA. Robust decision-making instead depends on transparent, context-specific approaches that integrate multiple sources of evidence, make assumptions explicit, and balance methodological sophistication with interpretability. As Bullement observed,  

“Establishing the best method is really hard when we don't know what the truth is.”