Survival extrapolation for HTA: from NICE DSU TSD 14 to flexible and cure models

Why extrapolation decides so many appraisals

Trials end long before patients' lives do. When a technology affects survival, the economic model has to estimate mean survival over a lifetime horizon, and most of the area under the curve often lies beyond the last observed event. NICE DSU Technical Support Document (TSD) 14 states the problem directly: exponential, Weibull, Gompertz, log-logistic and log-normal models can all be fitted, they can give very different survival estimates, and it is more important to justify the plausibility of the extrapolated portion than the fit to the observed data.^[1]

The evidence that this was not happening is well documented. Latimer reviewed the survival analysis in 45 NICE appraisals in cancer and found that the chosen survival models were not systematically justified in any of them. The range of models considered was usually insufficient, and the plausibility of the extrapolated portion was very rarely considered explicitly.^[2] Much of the methods guidance since then has aimed to close that gap.

NICE's manual now sets the expectation in plain terms. The appropriateness of survival models and the validity of their extrapolations should routinely be considered, the clinical plausibility of their underlying hazard functions should routinely be assessed, and uncertainty in the extrapolated portion should be explored.^[3]

The TSD 14 approach: standard parametric models

TSD 14 was published by the NICE Decision Support Unit in June 2011 and last updated in March 2013.^[1] It recommends that exponential, Weibull, Gompertz, log-logistic, log-normal and generalised gamma models should all be considered, and explains the hazard shapes each can represent:

• Exponential: constant hazard; a proportional hazards model.
• Weibull and Gompertz: monotonic hazards that increase or decrease but cannot change direction. The Gompertz log-hazard is linear in time, the Weibull's in log time.
• Log-logistic and log-normal: accelerated failure time models whose hazards can rise then fall; log-logistic models often produce long tails in the survivor function.
• Generalised gamma: a three-parameter model that includes the Weibull, exponential and log-normal as special cases.

The TSD 14 Survival Model Selection Process algorithm compares candidate models on internal validity, meaning how well they fit the observed data, and external validity, meaning how plausible the extrapolated portions are.^[1] The tools include log-cumulative hazard plots to judge hazard shape and proportionality, AIC and BIC, visual inspection, and clinical validity and external data. With patient-level data, the proportional hazards assumption should be tested rather than relied upon; the result indicates whether to fit separate models to each arm or allow a time-varying hazard ratio.^[1] NICE's manual adds that the proportional hazards assumption should always be assessed, with reference to TSD 14, and that alternative methods described in TSD 21 should be considered when it fails.^[3]

A common misreading

The lowest AIC tells you which model fits the observed data best, with a penalty for parameters. It says nothing directly about the part of the curve you have not observed. In our experience, submissions that lead with the information criterion and end with clinical plausibility invite more challenge than those that do the reverse.

Start with the hazard, not the survival curve

A small difference in the slope of a survival curve can hide a large difference in hazard. TSD 21 recommends that fitted and extrapolated hazard and survival functions should always be presented, and stresses the hazard scale because it is the scale on which models are estimated and on which treatment effects are often assumed to act. It also recommends plotting general population survival and hazard alongside the modelled curves.^[4]

Three plots answer most of an assessor's first questions:

1. Smoothed empirical hazards with each model's fitted hazard, per arm, over the observed period. This shows whether the model can represent the shape at all.
2. Extrapolated hazards per arm against general population hazards for the same age and sex mix. A modelled hazard that falls below background mortality in later years needs an explicit justification.
3. The implied treatment effect over time, as a hazard ratio or difference, when arms are fitted separately. TSD 21 recommends plotting this whether or not a proportional effect is assumed.^[4]

NICE's manual also warns that functions with stable or decreasing variance over time are likely to underestimate the uncertainty in the extrapolation.^[3] Plotting the probabilistic spread of extrapolated survival, not only the central estimate, is worth the extra page.

TSD 21: flexible methods and what they assume

TSD 21, by Rutherford and colleagues, was written because complex hazard functions, particularly with immunotherapies, had led to increasingly sophisticated survival models in NICE appraisals that were not used or interpreted consistently.^[4] It describes flexible parametric (spline) models, mixture models, landmark models, piecewise models, cure models and excess mortality (relative survival) models, tests them in simulation, and makes model-specific recommendations. It does not recommend a single type of model; its simulations showed that none always extrapolates best.^[4]

Source: TSD 21 recommendations and extrapolation sections.^[4] One cross-cutting recommendation stands out. TSD 21 recommends incorporating background mortality into survival models, because it helps avoid extremely implausible projections, and describes it as essential for cure models. Even imperfect background rates are likely to give a better extrapolation than ignoring background mortality altogether.^[4]

TSD 21 also records a telling example of misinterpretation: in one appraisal, a two-knot spline model was interpreted by the company, the review group and the committee as implying three heterogeneous patient subgroups. A spline is simply a way of modelling a complex hazard function and makes no such assumption.^[4] Describing what each model assumes, in plain language, is part of the submission.

Cure models: when a plateau is plausible

Treatments such as immune checkpoint inhibitors and CAR-T therapies can produce survival curves that appear to flatten. A cure model formalises that by estimating a fraction of patients whose mortality returns to (or close to) general population levels. Latimer and Rutherford's 2024 tutorial notes that cure models are often not included in HTA dossiers and are relatively rarely relied on in decision-making, which they attribute largely to limited understanding of how the models work and what they assume. They highlight flexible parametric non-mixture cure models as well suited to HTA when a cure assumption is valid but follow-up is limited.^[5]

TSD 21 is cautious. It notes that the benefit comes mainly from incorporating background mortality rather than from the label "cure", that flexible parametric models with background mortality may be more appropriate when cure is uncertain, and that a cure fraction should be supported by evidence such as response rates, earlier-phase trials with longer follow-up and the mechanism of action. Hazard and survival functions for cured and uncured patients should both be presented.^[4]

Validation studies show why this matters. Revisiting NICE's assessment of ipilimumab in melanoma with 5-year data, Bullement, Latimer and Bell Gorrod found that standard parametric, spline and non-cure mixture models fitted to 3-year data substantially underestimated 5-year survival. Only the model used in the appraisal and a mixture cure model came close, and the two diverged considerably after 5 years.^[6] In a CAR-T example using four data cuts of ZUMA-1, cure-based models fitted best from the earliest cut and stayed consistent as data matured, while standard parametric and spline models fitted poorly.^[7] Both studies used specific products and should not be generalised, but they illustrate a consistent point: models that bring in external information, through background mortality or registry data, tended to extrapolate better than flexible models that did not.^[6]

External data and clinical plausibility

NICE asks for the external validity of an extrapolation to be assessed by considering clinical and biological plausibility and coherence with external sources such as historical cohorts or other studies. Where plausibility is hard to define, it suggests expert elicitation to identify a plausible distribution of values.^[3] TSD 21 adds that registry data may be useful but that research is ongoing; relevant populations should be compared, external sources should be considered early, and analyses using them should be pre-specified.^[4]

The methods for bringing external data into extrapolation differ in what they assume about how hazards relate between trial and external populations, and over time. Jackson and colleagues categorised the approaches used in UK HTA reports on exactly that basis.^[8] In practice we find it useful to state those assumptions in a sentence each: for example, "after year 5, the hazard in the comparator arm is assumed to equal that of the registry cohort, adjusted for age and sex".

Our approach

Before fitting anything, write down what clinicians expect long-term survival to look like for each arm, and why, then source the external data that could confirm or refute it. Fitting first and asking clinicians to choose between curves afterwards invites the post hoc subjectivity that TSD 21 warns about.

Treatment effect waning

The assumed duration of benefit is often more influential than the choice of distribution. NICE's manual asks for scenarios in which the treatment effect stops or diminishes gradually, is sustained while people continue treatment, or persists to some degree after discontinuation.^[3]

How committees have handled this has been reviewed. Trigg and colleagues examined NICE appraisals published between October 2021 and September 2023 and found waning assumptions in 28 appraisals. Five years was the most common time point, used in 30 of 59 methods applied (50.8%). The most common reason given for including waning was precedent from earlier appraisals, and the authors found considerable heterogeneity in methods and justifications.^[9] A 2026 comparison of immune checkpoint inhibitor appraisals in lung cancer by NICE and Canada's Drug Agency reported that waning approaches evolved from abrupt loss of effect towards gradual decline, with persistence assumptions varying with data maturity.^[10] Outside oncology, a review of NICE appraisals of disease-modifying therapies in multiple sclerosis found substantial variation in how waning was implemented, with a recent preference for all-cause discontinuation as a proxy.^[11]

Our reading is that precedent can explain a committee's starting point, but evidence is what moves it. In our view, the most persuasive case presents long-term data on the same mechanism, off-treatment follow-up, or plotted hazard ratios over time from the trial, and shows the ICER across a range of waning start points and durations.

Immature data in oncology

Immature overall survival is the norm in oncology submissions. A review of 11 NICE appraisals of immune checkpoint inhibitors found that later data cuts added a mean of 18 months of follow-up and that the initial extrapolations generally predicted overall survival reasonably well, but on average underestimated it, by 0.4% to 2.7% depending on the method and whose preferred extrapolation was used.^[12] A study using four pre-planned data cuts of the JAVELIN Merkel 200 trial found that increasingly flexible models were favoured by goodness of fit as data matured, and that every method fitted to the earliest cut underestimated longer-term survival.^[13]

Three practical implications follow. First, show how model rankings and mean survival estimates change across available data cuts; stability across cuts is evidence in its own right. Second, pre-specify the analyses for the next data cut, because committees will see them. Third, where uncertainty cannot yet be resolved, plan for managed access. In England, under the arrangements in place since July 2016, the Cancer Drugs Fund supports NICE managed access: patients receive promising cancer treatments while further data are collected to address clinical uncertainty and inform NICE's decision on routine funding.^[14]

Treatment switching often compounds immaturity. NICE advises against simple adjustments such as censoring or excluding patients who switch, and statistical methods that adjust for switching can be presented.^[3] TSD 21 notes that switching adjustment must be done before extrapolation.^[4]

The EU joint clinical assessment perspective

Under the EU HTA Regulation, joint clinical assessment is limited to relative clinical effectiveness and safety; it does not cover economic evaluation, pricing or reimbursement, which remain national.^[15] The HTA Coordination Group's guidance on outcomes, adopted in June 2024, discusses mortality as time-to-event data summarised, for example, by Kaplan-Meier overall survival, with effect measures such as hazard ratios.^[16] Lifetime extrapolation for cost-effectiveness is therefore a national task.

The two are still linked. The time-to-event analyses submitted for a JCA, including any assessment of proportional hazards and the data cuts used, will be visible to national agencies that later receive an economic model built on the same data. In our view, a single, consistent survival analysis plan across the JCA dossier and national economic submissions avoids awkward questions about why the same trial was analysed differently.

What this means for evidence planning

1. Write a survival analysis plan before the data cut, covering candidate models from TSD 14 and TSD 21, the plausibility criteria, the external data sources and the treatment effect scenarios.
2. Collect clinical expectations early, ideally through structured elicitation, so plausibility is judged against priors rather than chosen from fitted curves.
3. Present hazards, not only survival curves, with general population hazards and the implied treatment effect over time.
4. Incorporate background mortality in every model, and use cure models only where a cure fraction is supported by clinical and biological evidence.
5. Quantify waning explicitly, with evidence for its start and duration rather than precedent alone.
6. Use every data cut to test the stability of extrapolations, and plan the next one with the managed access route in mind.
7. Align the JCA and national analyses so that the same trial tells one story.

Evidaxia builds and critiques survival analyses for NICE, SMC, AWMSG and NCPE submissions, including flexible, cure and relative survival models, and supports evidence plans for joint clinical assessment. See health economics and evidence, HTA submissions in the UK and model adaptation.

References

1. Latimer N. NICE DSU Technical Support Document 14: survival analysis for economic evaluations alongside clinical trials: extrapolation with patient-level data [Internet]. Sheffield: Decision Support Unit, ScHARR, University of Sheffield; 2011 Jun [updated 2013 Mar; cited 2026 Oct 30]. doi:10.15131/shef.data.33745843 Available from: https://doi.org/10.15131/shef.data.33745843
2. Latimer NR. Survival analysis for economic evaluations alongside clinical trials: extrapolation with patient-level data: inconsistencies, limitations, and a practical guide. Med Decis Making. 2013;33(6):743-54. Available from: https://pubmed.ncbi.nlm.nih.gov/23341049/
3. National Institute for Health and Care Excellence. NICE health technology evaluations: the manual (PMG36). Chapter 4: Economic evaluation [Internet]. London: National Institute for Health and Care Excellence; [date unknown] [cited 2026 Oct]. Available from: https://www.nice.org.uk/process/pmg36/chapter/economic-evaluation
4. Rutherford MJ, Lambert PC, Sweeting MJ, Pennington B, Crowther MJ, Abrams KR, et al. NICE DSU Technical Support Document 21: flexible methods for survival analysis [Internet]. Sheffield: Decision Support Unit, ScHARR, University of Sheffield; 2020 Jan [updated 2022 Mar; cited 2026 Oct 30]. doi:10.15131/shef.data.33849118 Available from: https://doi.org/10.15131/shef.data.33849118
5. Latimer NR, Rutherford MJ. Mixture and non-mixture cure models for health technology assessment: what you need to know. PharmacoEconomics. 2024;42(10):1073-90. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC11405446/
6. Bullement A, Latimer NR, Bell Gorrod H. Survival extrapolation in cancer immunotherapy: a validation-based case study. Value Health. 2019;22(3):276-83. Available from: https://pubmed.ncbi.nlm.nih.gov/30832965/
7. Vadgama S, Mann J, Bashir Z, Spooner C, Collins GP, Bullement A. Predicting survival for chimeric antigen receptor T-cell therapy: a validation of survival models using follow-up data from ZUMA-1. Value Health. 2022;25(6):1010-7. Available from: https://pubmed.ncbi.nlm.nih.gov/35667774/
8. Jackson C, Stevens J, Ren S, Latimer N, Bojke L, Manca A, Sharples L. Extrapolating survival from randomized trials using external data: a review of methods. Med Decis Making. 2017;37(4):377-90. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC5424081/
9. Trigg LA, Melendez-Torres GJ, Abdelsabour A, Lee D. Treatment effect waning assumptions: a review of National Institute of Health and Care Excellence technology appraisals. Value Health. 2024;27(8):1003-11. Available from: https://pubmed.ncbi.nlm.nih.gov/38679289/
10. Chen NC, Bullement A, Zemplenyi AT, O’Bryant CL, Kaizer AM, Anderson KE, et al. Methodological divergence in oncology HTA: implications for international harmonization and the handling of uncertainty. Pharmacoecon Open. 2026 Oct 27. doi:10.1007/s41669-026-00696-1 Available from: https://pubmed.ncbi.nlm.nih.gov/42946964/
11. Armoiry X, Wang-Steverding X, Connock M, et al. Is the assumption of waning of treatment effect applied consistently across NICE technology appraisals? A case-study focusing on disease-modifying therapies for treatment of multiple sclerosis. Int J Technol Assess Health Care. 2022;38(1):e83. Available from: https://pubmed.ncbi.nlm.nih.gov/36510406/
12. Bullement A, Meng Y, Cooper M, Lee D, Harding TL, O’Regan C, et al. A review and validation of overall survival extrapolation in health technology assessments of cancer immunotherapy by the National Institute for Health and Care Excellence: how did the initial best estimate compare to trial data subsequently made available? J Med Econ. 2019;22(3):205-14. doi:10.1080/13696998.2018.1547303 Available from: https://pubmed.ncbi.nlm.nih.gov/30422080/
13. Bullement A, Willis A, Amin A, Schlichting M, Hatswell AJ, Bharmal M. Evaluation of survival extrapolation in immuno-oncology using multiple pre-planned data cuts: learnings to aid in model selection. BMC Med Res Methodol. 2020;20(1):103. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC7204248/
14. NHS England. Cancer Drugs Fund [Internet]. London: NHS England; [date unknown] [cited 2026 Oct]. Available from: https://www.england.nhs.uk/cancer/cdf/
15. Fong H, Garau M. One Europe, one assessment? Unpacking the European joint clinical assessment [Internet]. London: Office of Health Economics; 2026 May 28 [cited 2026 Oct 30]. Available from: https://www.ohe.org/insights/one-europe-one-assessment-unpacking-the-european-joint-clinical-assessment/
16. Member State Coordination Group on HTA. Guidance on outcomes for joint clinical assessments [Internet]. Brussels: European Commission; 2024 Jun 10 [cited 2026 Oct 30]. Available from: https://health.ec.europa.eu/publications/guidance-outcomes-joint-clinical-assessments_en

Sources accessed October 2026. This article is analysis for information and is not legal, regulatory or tax advice.