Preprints.org logo
Preprint
Article

This version is not peer-reviewed.

Two Survival Cuts, One Silent Substitution: The Lurbinectedin Joint Clinical Assessment

Submitted:

21 August 2026

Posted:

25 August 2026

You are already at the latest version

Abstract
This methodological commentary examines the Joint Clinical Assessment (JCA) report of lurbinectedin combined with atezolizumab as maintenance therapy for extensive-stage small-cell lung cancer, based on the phase 3 IMforte trial. The JCA's evaluation rests on two overall survival data cuts: a confirmatory, hypothesis testing, statistically significant interim cut (hazard ratio 0.73, p=0.0174) and a mature, hypothesis generation, non-significant cut (hazard ratio 0.81). We identify two main critical methodological issues in the JCA's handling of these results. First, the JCA prefers the mature cut under a treatment-policy strategy that embeds subsequent crossover to lurbinectedin in the control arm, which is unavailable in routine European practice and dilutes the estimated treatment effect. The report lacks a decision-relevant, hypothetical-strategy analysis adjusting for this treatment switching. Second, the JCA discounts the positive interim result by assigning an "Unclear" risk-of-bias rating based on a late alpha-spending protocol amendment. This suspicion was left unresolved despite available data, the power to seek developer clarification, and regulatory findings from the European Medicines Agency (EMA) clearing the amendment as operational. Instead of a reasoned synthesis, the JCA performs a silent substitution of the mature cut for the interim. The report ignores the information both cut bring and questionably assesses the second cut as low risk of bias. To address these concerns, we recommend: complementing treatment-policy estimates with hypothetical-strategy analyses when trial crossover differs from European practice; resolving risk-of-bias concerns rather than leaving them "unclear"; engaging with prior regulatory evaluations; and providing graded certainty statements to reconcile divergent clinical results.
Keywords: 
evidence-based medicine
;  
health technology assessment
;  
Joint Clinical Assessment
;  
lurbinectedin
;  
overall survival
;  
risk of bias
Subject: 
Public Health and Healthcare  -   Public Health and Health Services
The Joint Clinical Assessment (JCA) of lurbinectedin in combination with atezolizumab, as maintenance treatment for extensive-stage small-cell lung cancer, rests on a single randomised trial, IMforte [1], with one comparator, one target population, and overall survival as a co-primary outcome. Overall survival was analysed at two prespecified data cuts: an interim analysis (29 July 2024) and a later analysis (12 February 2025). The report treats the later cut as its principal data source on grounds of maturity, preferring the more mature analysis as the applicable default. Our concern is what the report does with the two overall-survival results once that choice is made. At the mature cut, overall survival is not statistically significant (hazard ratio 0.81; 95% confidence interval 0.65–1.01), while it is significant at the interim cut (0.73; 0.57–0.95; p = 0.0174), where the result was prespecified, hypothesis testing, controlled for multiplicity, and confirmatory (Table I). The report leads with the former and discounts the latter, through two decisions that are individually questionable and jointly decisive: an estimand governing a mature estimate and a risk-of-bias judgement used to discount the interim estimate. Our concerns are that the assessors did not really consider whether the estimand governing the mature estimate matched the decision the assessment must inform, and that the confirmatory interim is set aside on a suspicion of unclear biases (insufficient information) the assessors never resolved while the European Medicines Agency (EMA) assessors investigated and cleared.

The Mature Cut: A Defensible Estimand for an Ill-Fitting Question

The report assesses the trial within a treatment-policy strategy [5], and says so explicitly. Under that strategy, therapy received after discontinuation of the study drug, including receipt of the experimental agent by control patients, is part of the effect being estimated rather than a source of bias, because the full time to event is the quantity of interest. This is a legitimate and defensible position, and the report is right that, on that reading, subsequent therapy is not a risk-of-bias problem and treatment-switching adjustments are not required.
The issue is that the estimand may answer a question the Member States are not asking. A treatment-policy overall-survival estimate is only as decision-relevant as the subsequent therapy embedded within it. In IMforte, the subsequent therapy of the control arm included lurbinectedin itself: at the mature cut, 22 control patients (9.1%) received lurbinectedin at some point after the study drug (Table I).1 The report itself notes that lurbinectedin is not authorised by the EMA for second-line use at the time of the assessment. The care that the treatment-policy estimate builds into the control arm therefore does not correspond to routine practice in the Member States, where control-arm patients would not have access to second-line lurbinectedin. This matters directionally: receipt of the experimental agent by control patients dilutes an intention-to-treat overall-survival comparison towards the null, so removing that specific crossover would move the estimate away from unity. The effect relevant to European practice, in which that crossover does not occur, is thus plausibly more favourable than the reported 0.81, and at a cut whose upper confidence limit already stands at 1.01, the difference between a diluted and an undiluted estimate is not academic.
The initial cut has a higher internal validity. The second cut might still be considered more relevant if it had higher external validity, but that is not explicitly justified. In fact, two points challenge this view: first as specified above lurbinectedin is available in the US but not in Europe; second, the trial enrolled patients in three geographies, US, Asia and Europe, the last being heterogeneous in availability of therapies and practice. So, the assumed higher external validity at the second cut may not apply unless justified with relevant evidence assessing the post progression therapies. Considering the second cut at low risk of bias in this situation is incomprehensible.
The report acknowledges that “with respect to other [estimand] strategies, the assessment… may differ,” but it neither presents nor requests an analysis under any other strategy. A hypothetical-strategy estimate, removing the effect of the control-arm crossover to lurbinectedin, is exactly the complementary analysis that would inform Member States whose practice differs from the trial’s. Established methods exist for it (the rank-preserving structural failure time model; inverse-probability-of-censoring weighting) [6], and neither was performed or requested at either cut. The developer did supply a “crude” survival-discounting sensitivity analysis for the mature cut, but the assessors set it aside in a single line as “not considered relevant,” and that analysis is in any case non-specific: it discounts survival after any subsequent therapy, symmetrically across arms, and so cannot isolate the crossover of interest (Table II).
We are careful not to overstate the magnitude. The one available proxy points to a modest effect: the developer’s non-specific discount moved the interim hazard ratio only from 0.73 to approximately 0.71–0.72 (Table I).2 A crossover-specific adjustment might be larger or smaller, and its direction away from the null, is more certain than its size. The point is not that a hypothetical-strategy analysis would necessarily restore significance, but that it was decision-relevant, that its omission runs in the direction that disadvantages the intervention, and that a treatment-policy estimate embedding a therapy unavailable in European practice should not have been presented as the decision-relevant result without it.

The Interim: A Confirmatory Result Set Aside on an Unresolved Suspicion

The interim analysis is the only overall-survival result that is statistically significant, and it is the stronger result in many respects: prespecified, hypothesis testing, part of the sequential testing strategy, controlled for multiplicity, and significant at the boundary in force. The report contests none of this. It displaces the interim in two steps. First, the mature cut is made the main data source on grounds of maturity, a defensible default, and not our quarrel. But maturity alone would not justify disregarding a significant, prespecified, multiplicity-controlled result that is in any case required to be reported; a mature estimate sits alongside a confirmatory one, it does not overwrite it. What licenses setting the interim aside rather than weighing it is a single risk-of-bias judgement: in the assessment of the first data cut, overall survival is rated “Unclear” overall, on the “other bias” domain, and the sole reason given is the late amendment of the alpha-spending approach.3 The concern stated by assessors is twofold: that the timing of the change “does not exclude” influence from emerging data, and that it “potentially” shortened the trial and so risks overestimating the effect.
Three features of this judgement warrant scrutiny.
First, the objection is not resolved to the standard the report applies elsewhere. A rating of “Unclear,” in the tool the assessors use [7], denotes insufficient information to decide, not a moderate risk. Yet for an analogous suspicion at the same trial (that the open-label design could bias investigator-assessed progression), the assessors did resolve the question, rating the domain High but the overall risk Low because the investigators’ estimates closely matched those of the blinded independent review facility (Table I). There the suspicion was tested against data; here it was recorded and left standing. The asymmetry is the more striking because overall survival is the outcome the report itself describes as objective and “not susceptible to detection bias… as death is an objective event.” The outcome least vulnerable to bias receives the less favourable rating, without any attempt to assess the potential risk of bias it rested on.
Second, the “overestimation” concern is answerable on the report’s own terms. The report invokes the literature on trials stopped early for benefit [8]. That literature concerns trials that stop at an interim and are therefore never checked against longer follow-up. IMforte did not stop; it continued to its planned final analysis, and the check accordingly exists: with more than 320 deaths the estimate moved only from 0.73 to 0.81, each estimate lying within the other’s confidence interval (Table I). A concern about unchecked early estimates has limited purchase against an estimate that was, in fact, checked.
Third, the assessors had the power to seek clarification but did not use it on this question. Under Article 11(2) they issued two requests, concerning the first subsequent therapy per arm and the reasons for discontinuation during induction; the developer’s response was recorded as fully addressing them.4 Neither request touched the amendment. The report thus left “Unclear”, a rating that by definition reflects missing information, on the single judgement through which the confirmatory result is discounted, while seeking and obtaining information on other questions.
To this the regulatory record adds two facts the JCA does not mention. The interim result met not only the amended boundary but also the initial stricter O’Brien–Fleming boundary the amendment had replaced (Table II); and the regulator, examining the same amendment, concluded that the protocol changes were operational rather than data-driven.5 If both hold, the JCA’s “does not exclude” suspicion had already been examined and closed by the regulator on the same evidence, and the JCA neither engages that finding nor explains why the question remains open despite it.

Reading the two Analyses Together

Each decision might pass unremarked on a first reading; on scrutiny neither holds, and together they determine the assessment’s account of overall survival: a significant, confirmatory result is set aside on an unresolved suspicion, and a non-significant result whose estimand embeds a therapy unavailable in European practice is presented in its place, with no graded statement reconciling the two. The report presents both cuts, as the guidance requires [9], but it substitutes one for the other rather than weighing them, and it offers no assessment of how they jointly inform the treatment effect. What reaches the twenty-seven national bodies that rely on the report is therefore not a reasoned judgement about the overall-survival evidence but two estimates: the report is explicit that the mature cut is preferred as the data source, yet it offers no judgement on what the two divergent survival results, taken together, establish. It foregrounds a number without doing the analytic work of saying what the evidence shows, and the same data allow that gap to be read in opposite directions.

Recommendations

Four changes would address these concerns; only the first calls for an additional analysis.
  • Where a treatment-policy estimate embeds subsequent therapy that is not available in routine Member State practice, complement it with a hypothetical-strategy analysis that removes that therapy’s effect, requesting it under Article 11(2) if the developer has not supplied it, and report its direction and, where estimable, its magnitude.
  • Apply a consistent evidentiary standard across risk-of-bias domains. Where a suspected bias is decisive for the reading of a result, either test it against available data (as was done for investigator-assessed progression) or seek the information needed to resolve it, rather than leaving it “Unclear.”
  • When discounting a prespecified, multiplicity-controlled result, engage the regulatory assessment of the same question, and state why any residual concern remains open despite it.
  • Where the same outcome is significant at one analysis and not at another, provide a graded statement of certainty reconciling the two, distinct from the risk-of-bias rating.

Conclusion

The JCA report misses a decision-relevant analysis alongside the treatment-policy estimate, a resolved risk-of-bias judgement in place of an open one at the interim, a justification of the low risk of bias attributed to the second cut, and a graded reconciliation of the two survival results in place of a silent substitution. Most of this is reinterpretation of evidence already in hand; the single analysis it calls for, a switching-adjusted overall-survival estimate, is standard, well precedented, and could be supplied by the developer. Providing these would not predetermine whether lurbinectedin's overall-survival benefit is judged established or uncertain; it would make that judgement visible, and open to argument, for the bodies that must act on it. Implicit judgement obscuring decision-making has been already pointed out as a critical issue for JCA [10].

References

  1. Paz-Ares, L.; Borghaei, H.; Liu, S.V.; et al. Efficacy and safety of first-line maintenance therapy with lurbinectedin plus atezolizumab in extensive-stage small-cell lung cancer (IMforte): a randomised, multicentre, open-label, phase 3 trial. Lancet 2025, 405, 2129–2143. [Google Scholar] [CrossRef]
  2. Member State Coordination Group on Health Technology Assessment, Joint Clinical Assessment Report of Lurbinectedin, Version 1.0. 2026. Available online: https://health.ec.europa.eu/publications/joint-clinical-assessment-report-lurbinectedin-zepzelca_en. (accessed on 14 August 2026).
  3. Hoffmann-La Roche, F. Lurbinectedin (Zepzelca): health technology developer submission dossier, study IMforte (GO43104). 2025–2026. Available online: https://health.ec.europa.eu/publications/joint-clinical-assessment-report-lurbinectedin-zepzelca_en. (accessed on 14 August 2026).
  4. Committee for Medicinal Products for Human Use. Zepzelca (lurbinectedin): European public assessment report. EMA/87579/2026. European Medicines Agency. 2026. Available online: https://www.ema.europa.eu/en/medicines/human/EPAR/zepzelca#assessment-history. (accessed on 14 August 2026).
  5. International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use. ICH E9(R1) addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. 2019. Available online: https://www.ich.org/page/efficacy-guidelines. (accessed on 14 August 2026).
  6. Latimer, N.R.; Abrams, K.R. Adjusting survival time estimates in the presence of treatment switching. NICE Decision Support Unit Technical Support Document 16. 2014. Available online: https://sheffield.ac.uk/nice-dsu/tsds/survival-analysis. (accessed on 14 August 2026).
  7. Higgins, J.P.T.; Altman, D.G.; Gøtzsche, P.C.; et al. The Cochrane Collaboration’s tool for assessing risk of bias in randomised trials. BMJ 2011, 343, d5928. [Google Scholar] [CrossRef] [PubMed]
  8. Bassler, D.; Briel, M.; Montori, V.M.; et al. Stopping randomized trials early for benefit and estimation of treatment effects: systematic review and meta-regression analysis. JAMA 2010, 303, 1180–1187. [Google Scholar] [CrossRef]
  9. European Commission. Guidance on reporting requirements for multiplicity issues and subgroup, sensitivity and post hoc analyses in joint clinical assessments. 2024. Available online: https://health.ec.europa.eu/publications/guidance-reporting-requirements-multiplicity-issues-and-subgroup-sensitivity-and-post-hoc-analyses_en (accessed on 14 August 2026).
  10. Toumi, M.; Falissard, B.; Jouini, A.; et al. Epistemological tensions in the EU HTA Joint Clinical Assessment: the illusion of judgment-free evaluation. J. Mark. Access Health Policy 2026, 14, 40. [Google Scholar] [CrossRef]
Table I. The two survival data cuts of IMforte.
Table I. The two survival data cuts of IMforte.
29 July 2024 — first cut 12 February 2025 — second cut Change
Status of the analysis
Place in the testing plan Planned interim; threshold met; confirmatory Outside the testing plan —
False-positive rate controlled? Yes (5% overall) No; p-values nominal lost
Called by the report Not the main data source “the main data source” —
Used by the developer as The basis of its own case Exploratory and descriptive —
Overall survival
Deaths 113 (46.7%) vs 136 (56.4%) — 249 159 (65.7%) vs 169 (70.1%) — 328 +79
Median survival, months 13.24 [11.89; 16.36] vs 10.65 [9.50; 12.16] 13.90 [12.26; 16.13] vs 11.14 [10.12; 13.01] both rose
Difference in medians 2.60 months [0.35; 5.98], p=0.0199 2.76 months [0.48; 5.32], p=0.0269 +0.16
Hazard ratio, unadjusted 0.73 [0.57; 0.95], p=0.0174 0.81 [0.65; 1.01], p=0.0605 (nominal) +0.08
Threshold passed The new threshold, 0.0313, and the old O’Brien–Fleming threshold of 0.0214 it replaced No threshold assigned —
Width of 95% CI (log scale) 0.511 0.441 −13.7%
Progression-free survival
Read by blinded review (IRF) 174 (71.9%) vs 202 (83.8%); HR 0.54 [0.44; 0.67], p<0.0001 Not analysed at this cut —
Read by investigators 178 (73.6%) vs 206 (85.5%); HR 0.55 [0.45; 0.68], p<0.0001 204 (84.3%) vs 217 (90.0%); HR 0.56 [0.46; 0.69], p<0.0001 +0.01
Follow-up and events
Median follow-up for survival 14.78 vs 15.18 months 21.59 vs 19.81 months ≈ +6 / +5
Patients progressed / died 384 (79.5%) / 249 (51.6%) 421 (87.2%) / 328 (67.9%) +37 / +79
Treatment given after the study drug
Any further anti-cancer drug 108 (44.6%) vs 132 (54.8%) 130 (53.7%) vs 146 (60.6%) +9.1 / +5.8 pp
Control patients given lurbinectedin 22 (9.1%) 22 (9.1%) Same %, longer exposure
Developer’s discounting analysis (not switching-specific) HR 0.71–0.72, p=0.009–0.010 HR 0.76–0.79, p=0.018–0.038; supplied, “not considered relevant for the assessment” —
Switching-specific adjustment (RPSFT or IPCW) Not performed; not requested Not performed; not requested —
Risk of bias given by the assessors
Overall survival UNCLEAR — “other bias”, the amendment LOW reversed
PFS read by blinded review LOW Not rated (not analysed) —
PFS read by investigators LOW (high for blinding, settled by agreement with blinded review) LOW —
Values are as reported in the JCA report [2], the developer’s dossier [3], or the regulatory assessment report [4]; the final column is arithmetic on those values. Each cell gives lurbinectedin + atezolizumab versus atezolizumab. Abbreviations: CI, confidence interval; EPAR, European public assessment report; HR, hazard ratio; IPCW, inverse-probability-of-censoring weighting; IRF, Independent Review Facility; JCA, Joint Clinical Assessment; PFS, progression-free survival; pp, percentage points; RPSFT, rank-preserving structural failure time.
Table II. Overall-survival hazard ratio by data cut and analysis.
Table II. Overall-survival hazard ratio by data cut and analysis.
Analysis (data cut; statistical boundary) Hazard ratio 95% CI p-value Used by
First cut-off, 29 Jul 2024 — unadjusted (vs amended Hwang–Shih–DeCani boundary 0.0313) 0.73 0.57–0.95 0.0174 HTD
First cut-off, 29 Jul 2024 — same interim assessed against the original O’Brien–Fleming boundary 0.0214; still significant 0.73 0.57–0.95 0.0174 EMA (EPAR);
Second cut-off, 12 Feb 2025 — unadjusted (no boundary; nominal) 0.81 0.65–1.01 0.0605 JCA
Second cut-off, 12 Feb 2025 — crude survival-discounting (10/20/30%) 0.76–0.79 0.61–0.99 0.018–0.038 HTD (JCA: set aside)
Values are as reported in the trial manuscript [1], the JCA report [2], the developer’s dossier [3], or the regulatory assessment report [4]. Rows 1 and 2 are the same interim estimate at the 29 July 2024 cut (HR 0.73; 249 overall-survival events), shown against the two sequential-testing boundaries it cleared: the amended Hwang–Shih–DeCani boundary (0.0313) and the stricter, pre-amendment O’Brien–Fleming boundary (0.0214) it replaced. The result remained statistically significant under both, as reported by the EMA in the EPAR and in the trial manuscript. The second-cut crude discounting is the developer’s non-specific survival-discounting supplementary analysis (10%, 20%, 30% reductions), symmetric across arms and not switching-specific. Each cell is lurbinectedin + atezolizumab versus atezolizumab. Abbreviations: CI, confidence interval; EMA, European Medicines Agency; EPAR, European public assessment report; HR, hazard ratio; HTD, health technology developer; JCA, Joint Clinical Assessment.
1
Report Table 33 (all subsequent therapies, clinical cut-off 12 February 2025). As first subsequent therapy the figure is 9 patients (3.7%); report Table 18.
2
First-cut value from the developer’s dossier; the corresponding value at the mature cut is not reported in the JCA.
3
Up to protocol version 6 (23 February 2023) the sequential boundaries were determined by the O’Brien–Fleming method; version 7 (27 November 2023) replaced these with a Hwang–Shih–DeCani alpha-spending function (γ = −1.5), shifting alpha towards earlier interim looks, and added a minimum follow-up of five months after the last patient was randomised. Report Section 4.2.1.2.
4
Report Appendix D: two requests (26 February 2026), on the first subsequent therapy per arm and on induction-phase disposition; the developer’s response (dossier version 0.3, 12 March 2026) was recorded as fully addressing them.
5
Both statements derive from the EMA European public assessment report (EMA/87579/2026) and are not stated in the JCA report.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.