Basket Trial

Basket Trial Evidence for Tumour-Agnostic Medicines and HTA

Histology-independent (tumour-agnostic) cancer medicines can be licensed for any solid tumour that carries a given molecular alteration, and the basket trial underpins the evidence for them. The resulting evidence is typically a set of small single-arm cohorts with no common comparator, supporting a licence broader than the patients studied, and health technology assessment bodies have found it difficult to appraise. Baskets can be analysed separately, pooled or linked through a hierarchical model, and NICE has had to adapt its approach to histology-independent appraisals. Basket trials also differ from umbrella and platform trials, and their evidence is often misread.

A shared biomarker as the organising principle

A basket trial is one form of master protocol, in which a single protocol governs several sub-studies run in parallel. The common element is the treatment and its molecular target, and the sub-studies are defined by tumour type. The premise comes from precision medicine, which holds that if a genetic alteration drives tumour growth, blocking it might work wherever the tumour arises.

A systematic review of master protocols by Park and colleagues, with searches run in July 2019, found 49 basket trials. Almost all were exploratory, at phase I or II (47 of 49), and most had no control group or randomisation (44 of 49). The primary endpoint is often the objective response rate within each basket, as in the vemurafenib basket study by Hyman and colleagues. The evidence reaching an appraisal is therefore usually a set of single-arm response estimates with immature survival data, and only rarely a comparative randomised controlled trial. The first histology-independent marketing authorisation in Europe was granted in 2019, on the basis of a common biomarker rather than the site of the primary tumour, and a licence of that kind covers tumour types barely represented in the trial, or not represented at all.

Analysing baskets separately, pooled or with borrowing

Each basket is usually too small to estimate a response rate precisely, so the analytical choice is how far one basket should inform another. Analysing baskets separately assumes the tumour types share nothing, which gives very imprecise estimates. Pooling every patient assumes the treatment works identically everywhere, which gives one precise figure that may be wrong for each tumour type. A hierarchical model sits between the two, in the form shown below.

$$y_j \sim \text{Binomial}(n_j, p_j), \quad \text{logit}(p_j) = \theta_j, \quad \theta_j \sim N(\mu, \tau^2)$$

where $y_j$ is the number of responders in basket $j$, $n_j$ is the number of evaluable patients in that basket, $p_j$ is the true response probability, $\theta_j$ is that probability on the log-odds scale, $\mu$ is the mean log-odds of response across baskets, and $\tau^2$ is the between-basket variance. When $\tau^2$ is close to zero the model behaves like complete pooling; when it is large the baskets are analysed almost separately. With only a handful of small baskets the data say little about $\tau^2$, so its prior largely decides how much borrowing occurs, which makes prior choice part of the evidence in this kind of Bayesian analysis.

The pull towards the common mean can be shown with a simplified normal approximation, treating the mean and between-basket variance as known and working on the probability scale, as follows.

$$\tilde{p}_j = w_j \hat{p}_j + (1 - w_j)\mu, \quad w_j = \frac{\tau^2}{\tau^2 + s_j^2}, \quad s_j^2 = \frac{\mu(1 - \mu)}{n_j}$$

where $\tilde{p}_j$ is the partially pooled response estimate for basket $j$, $\hat{p}_j = y_j / n_j$ is the observed response proportion, $w_j$ is the weight on the basket's own data, $s_j^2$ is the sampling variance of a proportion near the common mean in a basket of size $n_j$, and $\mu$ and $\tau^2$ are here a probability and its variance. Using the common mean in $s_j^2$ makes the weight depend only on basket size, so small baskets receive low weight and are pulled most strongly towards the mean. A full model estimates everything jointly on the log-odds scale and avoids this approximation.

Borrowing has known costs. In simulations of phase II designs, Berry and colleagues found that hierarchical modelling raised power in many scenarios, but that when the drug worked in only one group the power for that group fell to 65%, and that borrowing could inflate the type I error rate when the drug was promising in some groups and not others. The exchangeability-nonexchangeability (EXNEX) model of Neuenschwander and colleagues lets each basket either share a common distribution with similar baskets or stand apart, which tempers borrowing for extreme baskets.

Worked example with four illustrative baskets and an NHS case mix

The figures are illustrative. A single-arm basket trial enrols 50 evaluable patients across four tumour types. For partial pooling the common mean is fixed at $\mu = 0.50$ and the between-basket standard deviation at 0.15, so $\tau^2 = 0.0225$.

1. Pool every patient. Complete pooling gives 25 responders out of 50, a response rate of 0.50 for every tumour type.

2. Weight basket C. Its sampling variance is $s_C^2 = 0.50 \times 0.50 / 10 = 0.025$, so its weight is as follows.

$$w_C = \frac{0.0225}{0.0225 + 0.025} = \frac{0.0225}{0.0475} = 0.474$$

where $w_C$ is the share of basket C's partially pooled estimate that comes from its own data.

3. Shrink basket C.

$$\tilde{p}_C = 0.474 \times 0.10 + 0.526 \times 0.50 = 0.0474 + 0.263 = 0.310$$

where $\tilde{p}_C$ is the partially pooled response for basket C. Basket C moves furthest, from 0.10 to 0.31, because its result lies far from the common mean and it carries less than half the weight. Basket D, with 5 patients, has the lowest weight of all (0.310) and moves from 0.60 to 0.53, further than the larger baskets A and B.

4. Reweight to the population expected in practice. Suppose the alteration is found in NHS practice mainly in tumour type C, giving an illustrative mix of 10% A, 10% B, 60% C and 20% D. With the separate basket estimates the expected response is $R_{sep} = 0.1 \times 0.60 + 0.1 \times 0.60 + 0.6 \times 0.10 + 0.2 \times 0.60 = 0.30$. With the partially pooled estimates it is $R_{pp} = 0.1 \times 0.56 + 0.1 \times 0.56 + 0.6 \times 0.31 + 0.2 \times 0.53 = 0.404$.

The pooled trial figure implies 50 responders per 100 patients treated; the estimates that respect tumour type imply about 30 to 40. A response-based economic model fed the pooled figure would overstate benefit, because the trial over-represents tumour types that respond well. The partially pooled result is the better central estimate only if the baskets are similar enough to justify borrowing, since basket C rises from 0.10 to 0.31 purely because of the other baskets.

Evidence problems for economic evaluation

The trial answers a regulatory question about activity in a molecular subgroup, whereas a cost-effectiveness analysis needs comparative effects on survival and quality of life in the population that will be treated. NICE's appraisal of larotrectinib for NTRK fusion-positive solid tumours (TA630, 2020) shows each gap.

The evidence was a pooled analysis of 102 patients from 3 single-arm trials, one of them a basket trial whose responding baskets were expanded. With no control arm, the company built a comparator from last-line treatment arms in earlier appraisals and the literature, weighted by the trial's tumour mix, and the committee judged this naive indirect comparison open to bias. Convenience sampling had over-represented tumour types with a high prevalence of the fusion. Survival data were very immature, with 14 deaths among 102 patients, and the NIHR-funded review by Murphy and colleagues concluded that response endpoints may not be reliable surrogate endpoints for survival. The committee also included diagnostic testing costs, and the same review flagged the apportioning of genomic testing costs across multiple targets as a research need.

Heterogeneity ran through all of these problems. The committee found that assuming equal response across tumour and fusion types was inappropriate, given differences such as 84% response for the most common fusion partner. The evidence review group's Bayesian hierarchical model reduced the overall response from 72% to 57%. The 72% came from the main efficacy analysis, which excluded the 9 patients with primary central nervous system tumours, whereas the 57% included them, so part of the fall reflects the wider population as well as the borrowing. The committee considered this lower estimate, with its wide credible intervals, more appropriate for decision making.

How NICE has appraised histology-independent technologies

NICE's manual (PMG36) notes that histology-independent cancer treatments may suit basket trials with a heterogeneous population. It states that basket trials should be appropriately designed and analysed, assess heterogeneity and allow borrowing between baskets, and should include relevant comparators, random allocation, endpoints with a validated relationship to overall survival and quality of life, and all relevant patient groups. Few basket trials meet the whole list, since most in the Park review were not randomised.

The appraisals so far have relied on managed access. Larotrectinib (TA630) was recommended within the Cancer Drugs Fund while more data were collected, and its managed access review (ID6292) was paused in September 2026 so that the company could supply further analyses requested by the committee. Entrectinib for the same indication (TA644, 2020) was also available through the Cancer Drugs Fund; the review that replaced it (TA1118) was terminated in January 2026 because the company did not provide a complete evidence submission. A 2020 BMJ paper co-authored by NICE staff had said, as NICE reported at the time, that additional data collection, for example through the Cancer Drugs Fund, was likely to be needed.

Pembrolizumab for 5 named tumour sites with high microsatellite instability or mismatch repair deficiency (TA914, 2023) had a narrower licence, and the committee considered the sites individually rather than as one solid tumour population. The company modelled overall survival and progression-free survival with a Bayesian hierarchical model. The committee found it useful but noted substantial heterogeneity between sites and that the method had not been applied to time-to-event data in a NICE appraisal before; it accepted both this and separate survival modelling as plausible because the choice barely changed cost-effectiveness. The external assessment group recommended sharing information only between tumour sites shown by clinical arguments and evidence to be comparable.

In January 2020 NICE said that an NIHR-funded research project would recommend methods for appraising these medicines and that its results would be considered in NICE's methods review. Separately, an NIHR-funded report by teams at York and Sheffield on modelling approaches for histology-independent cancer drugs to inform NICE appraisals was published in the Health Technology Assessment journal in 2021.

Decision options when effects may differ by tumour type

Heterogeneity turns one approve or reject decision into a set of possible recommendations. The Murphy review recommended that, where there is evidence of heterogeneity in treatment effects and cost-effectiveness, consideration be given to optimised recommendations. It suggested that routine presentation of the scale of the consequences of heterogeneity and decision uncertainty could add to the assessments in NICE's methods guide. Its decision framework explored approval policies involving further data collection, pricing schemes and stratified decisions.

A companion analysis by the same group showed the stakes. Using published larotrectinib response data, a hierarchical model estimated a response probability of 60.9% (95% credible interval 16.0% to 91.8%), against 74.5% from naive pooling, and predicted 56.9% (0.2% to 99.9%) for a histology not represented in the trial. In a hypothetical response-based model, the probability of cost-effectiveness was 78% under identical response, but ranged from 93% to 11% across approval decisions once heterogeneity was allowed. Uncertainty of that size, which more data could reduce, is why value of information analysis and managed access fit this evidence.

Basket, umbrella and platform trials compared

The three master protocol designs share a single overarching protocol, and basket and umbrella trials also use molecular screening to decide which sub-study a patient enters. They differ in what varies across sub-studies, which shapes the economic evidence each can supply.

In the Park review, 8 of 18 umbrella trials and 15 of 16 platform trials were randomised. Platform trials therefore come closest to randomised evidence, although their shared control may not be the comparator relevant to a NICE appraisal, with an adaptive trial design governing when arms are dropped or added.

Where basket trial evidence is misread

The most common error is to read a histology-independent licence as evidence of a histology-independent effect. The vemurafenib basket study concluded that histologic context is an important determinant of response in BRAF V600-mutated cancers. A second error is to treat the pooled response rate as a population estimate, when it reflects whoever was easiest to recruit. A third is to treat borrowing as free precision, when a hierarchical model tightens basket estimates only by assuming the baskets are exchangeable, so it reduces the subgroup analysis problem of small, noisy groups without removing it. Finally, a single-arm response rate carries no information on relative effect by itself, so every model built on it depends on external comparator data whose population and biomarker status rarely match, a source of structural uncertainty that sensitivity analysis on the trial data alone cannot capture.

Sources

• Berry SM, Broglio KR, Groshen S, Berry DA. Bayesian hierarchical modeling of patient subpopulations: efficient designs of Phase II oncology clinical trials. Clinical Trials. 2013;10(5):720-734. doi:10.1177/1740774513497539
• Hyman DM, Puzanov I, Subbiah V, et al. Vemurafenib in multiple nonmelanoma cancers with BRAF V600 mutations. New England Journal of Medicine. 2015;373(8):726-736. doi:10.1056/NEJMoa1502309
• Murphy P, Claxton L, Hodgson R, Glynn D, Beresford L, Walton M, Llewellyn A, Palmer S, Dias S. Exploring heterogeneity in histology-independent technologies and the implications for cost-effectiveness. Medical Decision Making. 2021;41(2):165-178. doi:10.1177/0272989X20980327
• Murphy P, Glynn D, Dias S, Hodgson R, Claxton L, Beresford L, Cooper K, Tappenden P, Ennis K, Grosso A, Wright K, Cantrell A, Stevenson M, Palmer S. Modelling approaches for histology-independent cancer drugs to inform NICE appraisals: a systematic review and decision-framework. Health Technology Assessment. 2021;25(76). doi:10.3310/hta25760
• National Institute for Health and Care Excellence. Additional evidence needed to assess histology-independent cancer drugs. News article, 2 January 2020. Describes Cooper S, Bouvy JC, Baker L, et al., BMJ 2020;368:l6435, which was not read directly.
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). London: NICE; 2022, last updated 13 October 2026. Sections 3.3.7 and 3.3.8.
• National Institute for Health and Care Excellence. Larotrectinib for treating NTRK fusion-positive solid tumours (TA630). London: NICE; 2020. Sections 1 and 3.
• National Institute for Health and Care Excellence. Larotrectinib for treating NTRK fusion-positive advanced solid tumours (MA review of TA630) [ID6292]. Project information page, accessed October 2026.
• National Institute for Health and Care Excellence. Entrectinib for treating NTRK fusion-positive solid tumours (TA644). London: NICE; 2020. Replaced by TA1118.
• National Institute for Health and Care Excellence. Entrectinib for treating NTRK fusion-positive solid tumours in people 12 years and over (terminated appraisal) (TA1118). London: NICE; 2026.
• National Institute for Health and Care Excellence. Pembrolizumab for previously treated endometrial, biliary, colorectal, gastric or small intestine cancer with high microsatellite instability or mismatch repair deficiency (TA914). London: NICE; 2023. Sections 1 and 3.
• Neuenschwander B, Wandel S, Roychoudhury S, Bailey S. Robust exchangeability designs for early phase clinical trials with multiple strata. Pharmaceutical Statistics. 2016;15(2):123-134. doi:10.1002/pst.1730
• Park JJH, Siden E, Zoratti MJ, Dron L, Harari O, Singer J, Lester RT, Thorlund K, Mills EJ. Systematic review of basket trials, umbrella trials, and platform trials: a landscape analysis of master protocols. Trials. 2019;20:572. doi:10.1186/s13063-019-3664-1
• Sugden BM, Grimm SE, Wolff R, et al. Bayesian hierarchical modelling for histology-independent time-to-event outcomes in the NICE single technology appraisal of pembrolizumab for solid tumours with MSI-H/dMMR: External Assessment Group perspective. PharmacoEconomics. 2024;42:615-618. doi:10.1007/s40273-024-01381-z