The backdrop to the study is a familiar frustration in thoracic oncology. Mesothelioma patients whose disease progresses after first-line platinum-based chemotherapy have historically faced a grim outlook, and second-line options have been scarce. Immune checkpoint inhibitors, antibodies that release the molecular brakes on T cells, transformed treatment in lung cancer and mesothelioma’s first-line setting, so hopes ran high that drugs like nivolumab and pembrolizumab would rescue patients after progression. Yet the pivotal trials delivered maddeningly inconsistent answers. CONFIRM, a randomized phase II trial comparing nivolumab with placebo, initially reported an encouraging overall survival advantage. PROMISE-Meso, which tested pembrolizumab against standard chemotherapy, found no benefit whatsoever. Two well-conducted randomized studies, two irreconcilable results.

The researchers behind the new analysis suspected that the discordance might not lie in the biology of the drugs but in the statistics of the trials. Time-to-event endpoints such as overall survival and progression-free survival are estimated using the Kaplan–Meier method, which relies on censoring: when a patient’s outcome status is unknown at the time of analysis, that patient is withdrawn from the risk set rather than counted as an event. This is statistically valid only when censoring is uninformative, meaning that patients leave the analysis for reasons unrelated to their prognosis. When sicker patients are preferentially censored, the survival curve of the treated arm can be artificially flattered, and a hazard ratio calculated under such conditions can be badly misleading.

Post-progression censoring is a particularly insidious variant of this problem. In trials of immunotherapy, patients whose disease progresses are sometimes taken off study and their deaths, which may occur weeks or months later, are not captured if survival follow-up is incomplete. Because patients who progress early tend to have a worse underlying prognosis, censoring them systematically removes high-risk individuals from the treated arm’s denominator. To the Kaplan–Meier estimator, those patients appear to be alive and event-free at the moment of censoring, and the curve diverges from the comparator arm in a way that mimics a genuine treatment effect. The new study set out to determine whether this mechanism was operating in the trials that shaped second-line mesothelioma practice.

The methodological approach was elegant in its economy. Rather than requesting raw patient-level data, the team reconstructed the published Kaplan–Meier curves from CONFIRM, its later updated analysis, and PROMISE-Meso using established digitization techniques. They then applied a reverse Kaplan–Meier procedure to estimate the distribution of censoring times in each treatment arm, a trick that treats censoring itself as the event and thereby reveals when, and how asymmetrically, patients were withdrawn from observation. Differential censoring between arms was quantified over time and compared across endpoints, allowing the investigators to distinguish the noise of ordinary dropout from a pattern characteristic of post-progression censoring, in which the treatment arm loses disproportionate numbers of patients precisely after the time when progression events would be expected to occur.

What emerged was a telling contrast. In the original CONFIRM analysis, the reported overall survival advantage for nivolumab coincided with a short period of follow-up and a censoring pattern consistent with post-progression censoring in the experimental arm. The survival curves separated early, exactly when differential censoring was most pronounced, and the apparent benefit rested on data locked in prematurely. By contrast, the updated CONFIRM dataset, which incorporated longer follow-up and more complete ascertainment of deaths, showed no overall survival benefit and no fingerprint of post-progression censoring. The signal had dissolved as the missing outcomes matured, precisely what one would expect if the initial finding was an artifact of incomplete follow-up rather than a true therapeutic effect.

PROMISE-Meso told a different and equally instructive story. The reverse Kaplan–Meier analysis found no evidence of post-progression censoring at any time point in that trial, and its negative result remained stable throughout the follow-up period. In other words, the pembrolizumab trial appears to have been clean, and its negative conclusion stands on solid statistical ground. The discordance between the two trials, rather than demanding exotic biological explanations, may largely reflect the fact that one dataset was compromised by informative censoring while the other was not. When the team applied sensitivity analyses designed to correct for general differential censoring in CONFIRM, both the progression-free survival effect and the overall survival advantage lost statistical significance, reinforcing the conclusion that the initial results were fragile.

The implications extend well beyond mesothelioma. Immunotherapy trials are especially vulnerable to this bias because post-progression survival can be prolonged and variable, and because patients who progress on checkpoint inhibitors frequently move to subsequent therapies outside the trial protocol, making their outcomes harder to trace. Regulatory and journal standards increasingly demand mature survival data, but the new analysis shows that maturity alone is not enough; the pattern of censoring across arms and over time must itself be scrutinized. The authors advocate comprehensive ascertainment of survival outcomes, including linkage to death registries and proactive follow-up of patients who discontinue treatment, so that censoring remains genuinely uninformative. They also call for adequately powered biomarker programs, noting that although a modest treatment effect in a biologically defined subgroup of mesothelioma patients cannot be excluded, the current evidence base remains inconclusive.

For clinicians, the message is a caution against anchoring treatment decisions on early survival signals that may rest on statistical quicksand. For trialists and regulators, it is a roadmap: report censoring distributions alongside survival curves, test for differential censoring as routinely as one tests proportional hazards, and lock databases only after survival outcomes have been exhausted. The study is also a vindication of methods that extract new truth from published data alone, demonstrating that curve reconstruction and reverse Kaplan–Meier techniques can audit the integrity of pivotal trials without access to confidential patient records. In a field where desperate patients and optimistic investigators yearn for breakthroughs, the discipline to distinguish a real survival signal from a censoring artifact may be as consequential as any new drug, and this analysis offers a template for making that distinction with rigor.

Subject of Research: Statistical bias from post-progression censoring in second-line mesothelioma immunotherapy trials

Article Title: The impact of post-progression censoring in second-line immunotherapy studies for mesotheliomaThe impact of post-progression censoring in second-line immunotherapy studies for mesothelioma

Article References: Meirson, T., Kugler, H. D., Moskovitz, M., Cerza, F., Gray, S. G., Mutti, L., & Belli, C. (2026). The impact of post-progression censoring in second-line immunotherapy studies for mesotheliomaThe impact of post-progression censoring in second-line immunotherapy studies for mesothelioma. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08569-1

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08569-1

Keywords: malignant mesothelioma, immunotherapy, immune checkpoint inhibitors, post-progression censoring, Kaplan–Meier analysis, informative censoring, overall survival, CONFIRM trial, PROMISE-Meso, hazard ratio, biostatistics, clinical trials