The indirect comparison was too uncertain. NICE still said yes.

Dostarlimab, sold as Jemperli, is a PD-1 immunotherapy. The Medicines and Healthcare products Regulatory Agency, or MHRA, authorised it for use in Great Britain in 2021.[1]

The medicine can be used on its own for adults with a specific type of recurrent or advanced endometrial cancer that has progressed after platinum chemotherapy. The cancer must have mismatch repair deficiency or high microsatellite instability, two related abnormalities that can make a tumour more responsive to immunotherapy.[2,4]

In August 2026, NICE recommended dostarlimab for routine NHS use in this population.[2]

There was no trial comparing it directly with pembrolizumab, the treatment patients would usually receive. The submitted indirect comparison was highly uncertain and appeared to lack face validity. In plain English, its results did not form a clinically convincing picture.[2,3]

NICE still said yes.

How can an HTA body recommend a medicine when the indirect comparison does not reliably establish how well it works against the existing option?

The comparison joined two single-arm trials without a common control

The main evidence for dostarlimab came from GARNET. This was a single-arm trial, meaning that everybody received dostarlimab. There was no control group.[4]

The pembrolizumab evidence came from KEYNOTE-158. This was also a single-arm trial.[5]

The company used a matching-adjusted indirect comparison, or MAIC, to compare them. A MAIC gives different weights to patients in the trial for which patient-level data are available. The aim is to make their average characteristics resemble those reported in the other trial.

This MAIC was unanchored. There was no shared placebo or control group connecting the two studies. The comparison therefore relied on absolute outcomes observed in two separately recruited populations.[2,3]

We explain the distinction between anchored and unanchored methods in our field guide to indirect treatment comparisons.

An unanchored MAIC makes a demanding assumption

A MAIC can adjust for differences that are measured and reported in both studies. In TA1189, the company matched the GARNET population to KEYNOTE-158 using age, previous treatment lines, performance status and metastatic status.[2]

But an unanchored MAIC needs much more than a few balanced averages. It assumes that every important prognostic factor and treatment-effect modifier has been measured and adjusted for.[6]

A prognostic factor predicts a patient’s outcome regardless of treatment. A treatment-effect modifier changes how well one treatment works relative to another.

NICE’s technical guidance describes the assumption that all these factors have been captured as very strong and largely impossible to satisfy fully. The EU HTA Coordination Group similarly says unanchored MAICs are highly problematic because important variables will often be unavailable.[6,7]

That problem was visible here. Some important characteristics, including histology, were not available from KEYNOTE-158 and could not be included in the adjustment.[2,3]

The adjustment left less information than the trial recruited

Weighting also reduced the effective sample size of GARNET substantially. The exact value is confidential.[2,3]

Effective sample size is not the number of people recruited. It estimates how much information remains after some patients are given much more weight than others. A large reduction suggests limited overlap between the study populations and a result that depends heavily on a smaller group of patients.[6]

NICE reported that none of the MAIC estimates showed a statistically significant difference and described their 95% confidence intervals as wide. The numerical hazard ratios and confidence limits were marked confidential by the company, so the published guidance does not report them. It does report that progression-free survival numerically favoured dostarlimab, while overall survival numerically favoured pembrolizumab.[2]

The analysis also assumed that the relative rate of progression or death for the two treatments stayed constant throughout follow-up. NICE reported that this proportional hazards assumption did not hold for progression-free survival. In other words, the relative event rates changed over time, so one hazard ratio may conceal how the comparison changed during follow-up.[2]

The committee concluded that the MAIC was highly uncertain and appeared to lack face validity. These are recurring reasons why indirect comparisons struggle to influence HTA decisions. Across 195 comparisons submitted to the French Transparency Committee, common problems included differences between studies, risk of bias, missing data, statistical limitations and small samples.[8]

No significant difference does not mean equal efficacy

The company argued that the MAIC found no statistically significant difference between the treatments. It set the hazard ratios for progression-free and overall survival to 1 in its economic model.[3]

A hazard ratio of 1 means that the model assumes no difference between treatments for the outcome being modelled.

But a confidence interval that crosses 1 does not establish equal efficacy. It means that the analysis did not establish a statistically significant difference at the chosen threshold. If the interval is wide, the results may still be compatible with a meaningful benefit or a meaningful disadvantage.[9]

Formal non-inferiority or equivalence needs a clinical threshold defined in advance. This could be informed by a minimal clinically important difference, meaning the smallest difference that patients or clinicians would consider important. For a treatment comparison, the threshold is often expressed as a non-inferiority margin: the greatest loss of benefit that would still be considered acceptable.[10,11]

Consider a hypothetical survival comparison in which a hazard ratio above 1 favours the existing treatment. Before looking at the results, the team decides that a hazard ratio of 1.20 is the largest acceptable disadvantage for the new treatment.

• An estimated hazard ratio of 0.98 with a 95% confidence interval from 0.84 to 1.12 stays below 1.20. If the study and analysis are otherwise valid, this could support a conclusion of non-inferiority against that prespecified margin.
• An estimated hazard ratio of 0.98 with a 95% confidence interval from 0.72 to 1.35 crosses 1.20. It includes effects that would be clinically unacceptable, so non-inferiority has not been established, even though the estimate is not statistically different from 1.

The numbers are illustrative. The dostarlimab MAIC was not designed or powered to establish non-inferiority, and its exact hazard ratios are confidential.[2,3]

There is an equally important reverse point. Failure to establish comparable efficacy is not evidence that efficacy differs.

The dostarlimab MAIC did not establish equality. It also did not provide reliable evidence that one treatment was better or worse.

It was inconclusive.

NICE used a pragmatic modelling assumption, not proof of equality

An economic model still needs a treatment-effect input when the clinical evidence is inconclusive.

The company used hazard ratios of 1. The External Assessment Group did not agree. It preferred the MAIC point estimates and allowed their uncertainty to flow through the economic model.[3]

The committee agreed that the MAIC did not provide enough evidence to support equal efficacy. But it also considered the MAIC estimates too uncertain and clinically unconvincing to be its preferred model inputs.[2]

After reviewing all the evidence and hearing from clinical experts, NICE concluded that:

“for the purpose of modelling in this evaluation, assuming equal efficacy between treatments is preferred over the MAIC for decision-making.”[2]

That wording is precise. NICE accepted equal efficacy as a pragmatic modelling assumption for this appraisal. It did not conclude that equal efficacy had been demonstrated.

Wider evidence can support an assumption without proving the comparison

The assumption was not based on the non-significant MAIC alone.

Clinical experts advised that PD-1 inhibitors have a class effect. Dostarlimab and pembrolizumab act on the same immune pathway. The company also submitted PERLA, a randomised phase 2 trial that compared the two drugs directly, each combined with chemotherapy.[2,3,12]

PERLA found similar response rates. Progression-free survival statistically favoured dostarlimab, while the overall-survival estimate also favoured dostarlimab but remained uncertain.[12]

However, PERLA studied previously untreated metastatic lung cancer, not previously treated endometrial cancer. Both drugs were combined with chemotherapy. The paper also explains that its sample size and statistical design could not establish true non-inferiority.[12]

PERLA formed part of the wider evidence considered. It did not prove equal progression-free or overall survival in the population considered by TA1189.

This is the appropriate role of wider evidence. It can support or challenge an assumption. It does not automatically establish the missing comparison.

The economic conclusion followed the committee’s chosen assumptions

The committee preferred three assumptions in the model:

• equal progression-free and overall survival for dostarlimab and pembrolizumab
• a maximum treatment duration of three years for dostarlimab
• treatment-effect waning beginning between seven and nine years for both treatments.[2]

NICE explicitly stated that all three assumptions had a high level of uncertainty.[2]

With those assumptions, the model predicted equal quality-adjusted life years for dostarlimab and pembrolizumab. Dostarlimab then had similar or lower costs after the confidential commercial arrangement. NICE concluded that this was consistent with its cost-comparison principles and that dostarlimab was a cost-effective use of NHS resources.[2,13]

This was not a finding that uncertainty had disappeared. It was the economic consequence of the committee’s preferred assumptions, which NICE stated openly.

NICE does not automatically replace uncertain estimates with no difference

TA1189 is not permission to set every inconclusive treatment effect to 1.

In its 2025 appraisal of nemolizumab for atopic dermatitis, NICE considered a network meta-analysis of EASI 75 response. This outcome measures whether eczema severity has improved by at least 75% after 16 weeks.[14]

The company assumed equal efficacy whenever the network meta-analysis did not find a statistically significant difference between nemolizumab and another treatment. NICE rejected that assumption. The relative-effect estimates were not close to 1, and their wide credible intervals allowed substantial differences between treatments. The committee instead used the External Assessment Group’s estimated odds ratios for EASI 75 response.[14]

NICE ultimately recommended nemolizumab, but only when a biological medicine would otherwise be offered. It did not recommend it as an alternative to JAK inhibitors because the cost-effectiveness estimates were not acceptable for that comparison.[14]

Both appraisals produced a positive recommendation. But the acceptable model inputs and the final place in treatment differed. Context matters.

Plan the reimbursement decision, not just the analysis

The lesson from TA1189 is not that a weak indirect comparison is good enough. It is that teams should plan how comparative evidence, wider evidence and economic modelling will work together.

1. Define what the comparison must establish

Superiority, non-inferiority, clinical similarity and a treatment-effect input for a model are different objectives.

If the claim is non-inferiority or clinical similarity, define the clinically acceptable margin before seeing the answer. Then assess the estimate and its uncertainty against that margin, not simply against no difference.[10,11]

2. Check whether the indirect comparison is feasible

Map the full comparator evidence first. Check which patient characteristics, outcomes and follow-up times are available. Assess whether time zero and outcome definitions align, whether the populations overlap and how weighting is likely to affect the effective sample size.[6,7]

If the analysis is unanchored, identify every known prognostic factor and treatment-effect modifier. Then state which are unavailable and what that means for residual bias.[6,7]

3. Build the wider evidence before submission

Relevant evidence may include randomised evidence from another setting, real-world comparative data, longer follow-up, mechanistic evidence and structured clinical advice.[11,13]

This evidence can test whether a model assumption is plausible. It cannot create a missing variable or a comparable patient population after the trials are complete.

4. Test whether the reimbursement conclusion changes

Model the indirect estimate with its uncertainty. If a no-difference assumption is clinically defensible, test that too. Explore plausible benefit and harm scenarios, alternative survival models, treatment stopping and treatment-effect waning.[3,13]

The practical question is not only which estimate looks most likely. It is whether the reimbursement conclusion survives other defensible interpretations of the evidence.

This is why literature synthesis, indirect-comparison design and economic interpretation belong in one evidence plan. Evidax TRACE maps the comparator evidence and reported prognostic factors. Evidax Comparison Blueprints establish which comparisons are feasible, which assumptions the data can support and how uncertainty should enter the reimbursement model.

Take home

• NICE recommended dostarlimab even though its unanchored MAIC did not reliably establish efficacy relative to pembrolizumab.
• A non-significant indirect comparison does not establish equal efficacy. It may simply be inconclusive.
• NICE used equal efficacy as a pragmatic modelling assumption for TA1189. It did not conclude that equality had been demonstrated.
• Wider evidence can support an assumption, but it cannot repair missing variables, poor population overlap or a small effective sample size.
• Plan the comparison, its interpretation and its economic consequences as one reimbursement problem.

References

1. Medicines and Healthcare products Regulatory Agency. Marketing authorisations granted in June 2021. Jemperli 500 mg concentrate for solution for infusion, PLGB 19494/0297. 2021. https://assets.publishing.service.gov.uk/media/60ed7f0fe90e0764c8f0a460/Monthly_new_MA_listing_June_2021.pdf

2. National Institute for Health and Care Excellence. Dostarlimab for previously treated advanced or recurrent endometrial cancer with high microsatellite instability or mismatch repair deficiency. Technology appraisal guidance TA1189. Published 26 August 2026. https://www.nice.org.uk/guidance/ta1189

3. National Institute for Health and Care Excellence. Dostarlimab for previously treated advanced or recurrent endometrial cancer with high microsatellite instability or mismatch repair deficiency: committee papers. ID6326. 2026. https://www.nice.org.uk/guidance/ta1189/evidence

4. Oaknin A, Pothuri B, Gilbert L, et al. Safety, efficacy, and biomarker analyses of dostarlimab in patients with endometrial cancer: interim results of the phase I GARNET study. Clinical Cancer Research. 2023;29(22):4564-4574. https://doi.org/10.1158/1078-0432.CCR-22-3915

5. O’Malley DM, Bariani GM, Cassier PA, et al. Pembrolizumab in microsatellite instability-high/mismatch repair deficient and non-MSI-H/non-dMMR advanced endometrial cancer: phase 2 KEYNOTE-158 study results. Gynecologic Oncology. 2025;193:130-135. https://doi.org/10.1016/j.ygyno.2024.12.020

6. Phillippo DM, Ades AE, Dias S, Palmer S, Abrams KR, Welton NJ. NICE DSU Technical Support Document 18: Methods for population-adjusted indirect comparisons in submissions to NICE. NICE Decision Support Unit; 2016. https://www.sheffield.ac.uk/nice-dsu/tsds/population-adjusted

7. Member State Coordination Group on Health Technology Assessment. Methodological Guideline for Quantitative Evidence Synthesis: Direct and Indirect Comparisons. Adopted 8 March 2024. https://health.ec.europa.eu/publications/methodological-guideline-quantitative-evidence-synthesis-direct-and-indirect-comparisons_en

8. Monnereau M, Baschet L, Jarne A, et al. Methodological advances and challenges in indirect treatment comparisons: a review of international guidelines and Haute Autorité de Santé Transparency Committee case studies. Value in Health. 2026;29(6):1025-1033. https://doi.org/10.1016/j.jval.2025.12.013

9. Altman DG, Bland JM. Absence of evidence is not evidence of absence. BMJ. 1995;311:485. https://doi.org/10.1136/bmj.311.7003.485

10. Piaggio G, Elbourne DR, Pocock SJ, Evans SJW, Altman DG, CONSORT Group. Reporting of noninferiority and equivalence randomized trials: extension of the CONSORT 2010 statement. JAMA. 2012;308(24):2594-2604. https://doi.org/10.1001/jama.2012.87802

11. Edwards SJ, Burgess BJ, Downes N, et al. Clinical similarity in cost-comparison evaluations: a systematic review of current methods in NICE appraisals and the development of a framework for the formal assessment of clinical similarity. BMJ Open. 2026;16:e112164. https://doi.org/10.1136/bmjopen-2025-112164

12. Lim SM, Peters S, Ortega Granados AL, et al. Dostarlimab or pembrolizumab plus chemotherapy in previously untreated metastatic non-squamous non-small-cell lung cancer: the randomized PERLA phase II trial. Nature Communications. 2023;14:7301. https://doi.org/10.1038/s41467-023-42900-4

13. National Institute for Health and Care Excellence. Technology appraisal and highly specialised technologies guidance: the manual. PMG36. Published 31 January 2022. Updated 31 March 2026. https://www.nice.org.uk/process/pmg36

14. National Institute for Health and Care Excellence. Nemolizumab for treating moderate to severe atopic dermatitis in people 12 years and over. Technology appraisal guidance TA1077. Published 2 July 2025. https://www.nice.org.uk/guidance/ta1077