Med-Read.org
Evidence Audit: · CONSORT 2025 / GRADE
Oncology Methodology Dossier · Surrogate Validation & Multi-Endpoint Appraisal

Oncology Surrogate Endpoints, Confirmatory Failures, and Net Treatment Benefit (GPC)

A cancer trial can meet its primary statistical endpoint, earn accelerated regulatory approval, and still fail to help patients live longer or feel better. Understanding why radiographic Progression-Free Survival (PFS), Objective Response Rate (ORR), and circulating tumor DNA (ctDNA) frequently diverge from Overall Survival (OS) requires auditing both confirmatory trial cohorts and modern multi-endpoint statistics such as Generalized Pairwise Comparisons (GPC) and Net Treatment Benefit (NTB).

Mentor's Plain-Language Takeaway & Absolute Risk Verdict

Across 791 modern phase 3 oncology trials, 63% used surrogate primary endpoints, yet only 43% of positive surrogate trials demonstrated Overall Survival improvement and 15% improved global Quality of Life. Evaluating composite endpoints with Generalized Pairwise Comparisons (Net Treatment Benefit) and per-1,000 NNT versus NNH prevents toxic, non-survival regimens from masquerading as breakthroughs.

43%
Positive Surrogates Confirming OS
vs
R² = 0.13
Immunotherapy PFS-to-OS Correlation
Editorial studio photograph of a brass counterweight balance scale comparing ivory and navy survival blocks beside a terracotta glass prism and proportional calipers
SURROGATE RESPONSE VS. OVERALL SURVIVAL & NET BENEFIT FDA / EMA AUDIT · GPC / NNT

Why Surrogate Endpoints Frequently Fail Confirmatory Survival Verification

Regulatory pathways such as the US Food and Drug Administration (FDA) Accelerated Approval program permit oncology drugs to enter clinical practice based on intermediate or surrogate biomarkers judged reasonably likely to predict true clinical benefit. In principle, postmarketing confirmatory trials then verify whether tumor shrinkage or delayed radiographic progression translates into longer life or superior Health-Related Quality of Life (QoL). In practice, multiple independent audits show that this verification step breaks down more often than clinicians or patients realize.

A 2024 cohort study published in JAMA (2024, Liu et al.) tracked all 129 cancer drug-indication pairs granted FDA accelerated approval between 2013 and 2023. Among the 46 indications with more than 5 years of postapproval follow-up (approved 2013 to 2017), 63% (29 of 46) were converted to regular approval, 22% (10 of 46) were withdrawn, and 15% (7 of 46) remained ongoing after a median of 6.3 years. Crucially, fewer than half (43%, 20 of 46) demonstrated an improvement in Overall Survival or Quality of Life in confirmatory trials. Even among the 48 drug-indication pairs converted to regular approval, only 40% (19 of 48) converted on Overall Survival; 44% (21 of 48) converted on another surrogate (PFS), 14% converted on response rate metrics, and 2% (1 of 48) converted despite a negative confirmatory trial.

These modern findings mirror earlier regulatory audits. An investigation in JAMA Internal Medicine (2019, Gyawali et al.) covering 93 FDA oncology accelerated approvals from 1992 to 2017 found that confirmatory trials demonstrated an OS improvement in only 20% (19 of 93) of indications, while 41% relied on the same or another surrogate endpoint. Similarly, a European Medicines Agency (EMA) retrospective review in The BMJ (2017, Davis et al.) examining 48 cancer drugs across 68 indications approved between 2009 and 2013 reported that only 35% (24 of 68) showed OS prolongation at market entry (median gain 2.7 months, range 1.0 to 5.8 months) and only 10% (7 of 68) showed QoL improvement; after a median 5.4 years of follow-up, just 51% (35 of 68) had demonstrated any survival or QoL gain.

The surrogate gap is not restricted to accelerated single-arm approvals; it pervades modern phase 3 randomized trials as well. A 2025 meta-epidemiological analysis in JAMA Oncology (2025, 791 Phase 3 Oncology RCTs) screened 791 superiority-design trials enrolling 555,580 patients from 2002 to 2024. Alternative endpoints such as PFS served as the primary endpoint in 63% (495 of 791) of trials. Although 55% (434 of 791) achieved statistical superiority on an alternative endpoint, only 28% (221 of 791) achieved Overall Survival superiority, and only 11% (84 of 791) demonstrated global QoL superiority (with global QoL published in just 34% of trials, and 82% of those analyses unadjusted for baseline scores). Only 6% (48 of 791) proved superiority in both OS and global QoL, and among the 434 trials with a positive surrogate endpoint, a minority achieved OS superiority (43%, 185 of 434) or global QoL superiority (15%, 67 of 434).

Emerging molecular biomarkers face the same evidentiary hurdle. In the 2026 randomized phase 3 ALTAIR trial (ClinicalTrials.gov NCT04457297) published in Nature Medicine (2026, ALTAIR Phase 3 Trial), 243 patients with resected colorectal cancer who turned circulating tumor DNA (ctDNA) positive without radiological recurrence were randomized 1:1 to trifluridine/tipiracil (n = 122) or placebo (n = 121) for 6 months. Median disease-free survival was 9.30 months versus 5.55 months (HR 0.79, 95% CI 0.60 to 1.05, P = 0.107), failing its primary endpoint while Grade 3 or higher hematologic toxicity jumped from 3.3% to 73.0% (an absolute harm increase of 69.7 percentage points, or NNH = 2). Even biologically compelling biomarkers require randomized confirmation before routine escalation.

Surrogate-to-OS Trial-Level Correlation (R²): Where Surrogacy Holds and Where It Collapses

A surrogate endpoint is clinically valid only when a treatment's effect on the surrogate reliably predicts its effect on Overall Survival across randomized trials (trial-level R² ≥ 0.70), not merely because individual responders live longer than non-responders.

Clinicians often confuse patient-level prognostic correlation (patients whose tumors shrink live longer than patients whose tumors grow) with trial-level surrogacy (whether a drug that raises ORR or prolongs PFS by 3 months actually extends median OS compared with the control arm). In two-stage meta-analytic validation, trial-level R² measures how tightly the log Hazard Ratio on the surrogate predicts the log Hazard Ratio on Overall Survival across multiple trials.

Trial-Level Surrogate-to-Overall-Survival Correlation (R²) Across Oncology Contexts

Weighted R² of treatment effects (log-HR surrogate vs. log-HR OS) across verified meta-analyses. R² >= 0.70 is the conventional validation threshold.

Weak / Invalid Surrogate (R² < 0.50) Validated Surrogate (R² >= 0.70) R² = 0.70 Surrogacy Threshold Prostate: Local Failure vs. OS (75 trials)0.085 (0.00-0.37)Prostate: Biochem. FFS vs. OS (75 trials)0.12 (0.003-0.33)Immunotherapy: ORR OR vs. OS HR (13 RCTs)0.13Immunotherapy: PFS HR vs. OS HR (13 RCTs)0.13Prostate: PFS vs. OS (75 trials)0.46 (0.22-0.67)Prostate: MFS vs. OS (75 trials, Lancet Oncol)0.78 (0.59-0.89)First-Line DLBCL: PFS vs. OS (13 RCTs, SEAL)0.83 (WLS)NSCLC Adjuvant Chemo: DFS vs. OS (60 RCTs)0.92 (0.88-0.95) 0.00.20.40.60.81.0 Trial-Level Coefficient of Determination (R², 0.00 = zero prediction to 1.00 = perfect prediction) Sources: Lancet Oncol 2021 (Prostate, 75 trials); Clin Cancer Res 2018 (Immunotherapy, 13 RCTs); JCO 2017 & 2018 (ICECaP/SEAL); Lancet Oncol 2013 (NSCLC, 60 RCTs).

As the correlation forest plot illustrates, surrogacy is highly context-specific and mechanism-dependent:

  • Strong validated surrogacy (R² ≥ 0.78): In cytotoxic chemotherapy and radiotherapy for operable non-small-cell lung cancer (The Lancet Oncology 2013, 60 RCTs, 15,071 patients), adjuvant DFS correlated strongly with OS (R² = 0.92, 95% CI 0.88 to 0.95 without radiotherapy), though the authors explicitly warned against extrapolating this validation to targeted therapies. Similarly, in first-line diffuse large B-cell lymphoma (Journal of Clinical Oncology 2018, SEAL IPD analysis of 7,507 patients), PFS showed R² = 0.83, and in localized prostate cancer (The Lancet Oncology 2021, 75 trials, 53,631 patients and JCO 2017 ICECaP), Metastasis-Free Survival (MFS) reached R² = 0.78 to 0.92 while biochemical PSA failure (R² = 0.12) and local failure (R² = 0.085) failed completely.
  • Weak or collapsed surrogacy in immunotherapy (R² = 0.13): Across 13 randomized active-control immunotherapy trials submitted to the FDA (Clinical Cancer Research 2018, Mushti et al.), trial-level correlation between PFS Hazard Ratio and OS Hazard Ratio collapsed to R² = 0.1303, and between ORR Odds Ratio and OS Hazard Ratio to R² = 0.1277. Immune checkpoint inhibitors can produce delayed survival tails without early PFS separation, or conversely cause pseudoprogression and late toxicity that decouple radiographic scans from lifespan.

Composite Endpoints, Win Ratio Fallacies, and Net Treatment Benefit (GPC)

Because single surrogate endpoints overlook toxicity and traditional survival analyses ignore quality of life, clinical trialists increasingly test multi-endpoint composite frameworks. However, conventional time-to-first-event composite endpoints suffer from a fatal flaw: they treat the earliest event as equally important to death. If a composite combines death, symptomatic bone metastasis, and asymptomatic radiographic progression, the statistical test is dominated by whichever event occurs earliest and most frequently, usually the scan measurement.

How Generalized Pairwise Comparisons (GPC) Compute Net Treatment Benefit

Generalized Pairwise Comparisons (GPC), introduced by Marc Buyse and extended to censored oncology outcomes (Statistical Methods in Medical Research 2018, Péron et al.), solve this problem by establishing an explicit clinical hierarchy of outcomes. Every patient in the experimental arm is compared pairwise against every patient in the control arm across prioritized criteria:

  1. Priority 1 (Most Important, e.g., Overall Survival): Determine whether one patient lived longer by at least a clinically meaningful threshold (for example, >2 months). If yes, that pair is scored a win or loss and comparison stops.
  2. Priority 2 (Tie-Breaker, e.g., Grade 3/4 Toxicity or Quality of Life): Only if the pair is tied or neutral on Priority 1 does the comparison proceed to Priority 2, scoring whether one patient avoided severe toxicity.
  3. Net Treatment Benefit (Δ = P_win - P_loss): Subtract the probability that a treated patient fares worse from the probability that a treated patient fares better. An NTB of +0.20 (+20%) means a randomly chosen patient on the new regimen has a 20 percentage-point higher net chance of a better overall outcome than a control patient.

In a landmark oncology application (Oncotarget 2016, Péron et al., 342 metastatic pancreatic cancer patients), GPC compared FOLFIRINOX against gemcitabine by prioritizing OS differences exceeding 2 months first and Grade 3/4 toxicity second. Despite FOLFIRINOX's substantial toxicity, the Net Treatment Benefit strongly favored FOLFIRINOX at +24.7% (P < 0.001) across all sensitivity thresholds. Crucially, methodological work in the Journal of Clinical Epidemiology (2021, Péron & Buyse) demonstrates that analyzing marginal efficacy and marginal toxicity separately misleads clinicians whenever response and toxicity are correlated within the same patient; GPC evaluates the joint patient-level benefit-risk distribution directly. Modern de-escalation trials, such as a 5-endpoint acute promyelocytic leukemia protocol (Clinical Trials 2024), now use GPC to power studies for an NTB of 0.19 with 260 to 300 patients where traditional non-inferiority designs would be prohibitively large.

Win Ratio vs. Net Treatment Benefit: Avoiding Tie and Censoring Fallacies

The related Win Ratio (P_win / P_loss), popularized by Pocock, uses the same pairwise logic as GPC but expresses the result as a ratio rather than a difference. As detailed in JACC: Basic to Translational Science (2023) and Journal of Biopharmaceutical Statistics (2020, Dong et al.), an unadjusted Win Ratio can look impressively large (e.g., 1.80) even when 75% of patient pairs are completely tied (15% wins vs. 8.3% losses yields a Win Ratio of 1.80, whereas Net Treatment Benefit truthfully reports +6.7%). Furthermore, unequal follow-up or heavy censoring distorts unadjusted win proportions unless corrected with Inverse-Probability-of-Censoring Weighting (IPCW).

Endpoint Bias & Multi-Endpoint Verdict Ledger

How to classify oncology endpoint designs, censoring artifacts, and confirmatory failures before translating trial claims to bedside care.

Oncology Endpoint Designs, Confirmatory Evidence & Bias Verdict Ledger
Endpoint Design / Trial Paradigm Verified Empirical Benchmark Per-1,000 Survival vs. Toxicity Reality Methodological Verdict
FDA Accelerated Approval (Surrogate Entry)
JAMA 2024 (2013-2023 Cohort, n=129)
46 approvals with >5 yr follow-up: 63% converted, 22% withdrawn, 43% proved OS or QoL gain. Over half of approvals lack confirmed survival/QoL gain at 5+ yr while patients face full drug toxicity and cost. SURROGATE ONLY
Check whether confirmatory OS trial finished and whether indication shifted.
Phase 3 Alternative Primary Endpoint
JAMA Oncol 2025 (791 RCTs, 555,580 pts)
55% positive on alternative endpoint; only 43% of positive trials proved OS and 15% proved global QoL. In 1,000 trial enrollees across modern phase 3 oncology, only 6% of trials proved both longer life and better QoL. MODERATE · VERIFY OS/QOL
Verify baseline-adjusted QoL scores; 82% of published QOL analyses omit adjustment.
ctDNA MRD Post-Adjuvant Escalation
ALTAIR Phase 3, NCT04457297 (Nat Med 2026)
FTD/TPI vs. placebo (n=243): DFS 9.30 vs 5.55 mo, HR 0.79 (P=0.107, missed primary endpoint). 0 confirmed DFS/OS gain vs. +697 per 1,000 Grade ≥3 hematologic adverse events (73.0% vs 3.3%, NNH = 2). NEGATIVE · HIGH HARM
Do not escalate toxic chemotherapy on ctDNA positivity outside randomized trials.
PFS with Informative Censoring / BICR Dropouts
JCO 2008 (Dodd et al.) & Toxicity Dropouts
Early toxicity withdrawals or unconfirmed local progressions censored instead of counted as events. Artificially lowers experimental failure hazard, inflating PFS median separation with zero OS difference. HIGH BIAS RISK
Inspect number-at-risk and censoring tick marks before month 3 on KM plots.
Unweighted Composite or Unadjusted Win Ratio
JACC BTS 2023 & J Biopharm Stat 2020
Equal weight for minor/major events or omission of high tie rates and differential follow-up censoring. A Win Ratio of 1.80 with 75% ties equals only +67 net wins per 1,000 pairs; censoring shifts ratio direction. AUDIT TIES & IPCW
Require absolute Net Treatment Benefit (Δ), tie rate, and IPCW adjustment.
Hierarchical GPC Net Treatment Benefit
Oncotarget 2016 (FOLFIRINOX) & Clin Trials 2024
Prioritizes OS (>2 mo) above Grade 3/4 toxicity across all patient pairs (FOLFIRINOX NTB = +24.7%, P<0.001). Integrates survival extension and toxicity within the same individual: +247 net better-outcome pairs per 1,000. HIGH UTILITY · GRADE A/B
Gold-standard multi-endpoint synthesis when OS holds Priority 1.

Notice a critical failure mode in row 4: informative censoring. Even when trials deploy Blinded Independent Central Review (BICR) to reduce investigator assessment bias, Journal of Clinical Oncology (2008, Dodd et al.) showed that censoring locally determined progressions that central radiologists do not immediately confirm creates systematic informative censoring bias. Always inspect Kaplan-Meier "number at risk" tables below the x-axis: if censoring drops steeply in the experimental arm during the first two scan cycles while OS curves overlap, the PFS advantage is likely an attrition artifact.

Bedside Appraisal Protocol: Converting Oncology Claims into Per-1,000 NNT and NNH

When reviewing an oncology trial report or discussing a regimen in clinic, apply this five-step protocol alongside our 1,000-Person NNT Workbench and 10-Minute Study Appraisal Guide:

  1. Check the Primary Endpoint Hierarchy: Ask whether the trial was powered for Overall Survival or a radiographic/molecular surrogate (PFS, DFS, ORR, pCR, ctDNA). If it is a surrogate, verify its context-specific trial-level R² (for example, MFS in localized prostate cancer has R² ≥ 0.78, whereas immunotherapy PFS has R² = 0.13).
  2. Audit Confirmatory Status for Accelerated Approvals: Verify the ClinicalTrials.gov NCT identifier to see if a randomized confirmatory OS trial is completed, delayed, or negative.
  3. Convert Survival Probabilities into Absolute Risk Reduction (ARR) and NNT: If 2-year survival rises from 58% to 65%, ARR = 7.0% (70 more patients alive per 1,000 treated, NNT = 15). Conversely, if median PFS improves by 2.1 months with identical 2-year OS (40% vs. 40%), OS NNT is infinite.
  4. Compute NNH from the Grade 3/4 Adverse Event Table: Subtract control Grade ≥3 toxicity from experimental Grade ≥3 toxicity. In the ALTAIR ctDNA trial, 73.0% minus 3.3% = 69.7% absolute harm increase (NNH = 2). Never rely on summary prose calling toxicity "manageable."
  5. Check Multi-Endpoint Priority and Baseline-Adjusted QoL: Where trials report composite endpoints, Win Ratios, or GPC Net Treatment Benefit, confirm that death/OS takes Priority 1, ties are disclosed, and Quality of Life curves adjust for baseline symptom burden.

Continue exploring our Oncology Endpoints (OS vs. PFS) Primer, Cardiometabolic NNT Ledger, and 2026 Evidence & Reproducibility Updates.

Primary peer-reviewed sources: JAMA Oncology 2025 (791 Phase 3 RCTs) · JAMA 2024 (FDA Accelerated Approval Outcomes) · JAMA Internal Medicine 2019 (Gyawali et al.) · BMJ 2017 (EMA Approvals) · Nature Medicine 2026 (ALTAIR NCT04457297) · Clinical Cancer Research 2018 (Immunotherapy R²) · The Lancet Oncology 2021 (Prostate R²) · JCO 2018 (SEAL DLBCL) · The Lancet Oncology 2013 (NSCLC R²) · Oncotarget 2016 (FOLFIRINOX GPC) · J Clin Epidemiol 2021 (Net Benefit Correlation) · Clinical Trials 2024 (GPC Design) · JACC BTS 2023 (Win Ratio Fallacies) · J Biopharm Stat 2020 (IPCW Win Ratio) · JCO 2008 (BICR Informative Censoring).

Common Questions from Patients, Clinicians & Researchers

Candid, plain-language answers grounded in CONSORT 2025, GRADE, WHO guidelines, and per-1,000 outcome arithmetic.

What are the FDA guidelines and confirmatory evidence gaps for surrogate endpoints in oncology clinical trials?

Under FDA Accelerated Approval rules, cancer drugs may be authorized on surrogate endpoints such as Objective Response Rate or Progression-Free Survival if reasonably likely to predict clinical benefit, subject to mandatory confirmatory trials. However, a 2024 JAMA analysis of 46 accelerated oncology approvals with over 5 years of follow-up found that only 43% demonstrated Overall Survival or Quality of Life benefit in confirmatory trials, while 22% were withdrawn and 44% of conversions still relied on PFS rather than OS.

Source: JAMA 2024 (FDA Accelerated Approval Outcomes) · JAMA Internal Medicine 2019

How does Net Treatment Benefit (Generalized Pairwise Comparisons) differ from the Win Ratio and traditional composite endpoints?

Traditional composite endpoints give equal weight to the earliest event, allowing minor scan progressions to overshadow survival. Both the Win Ratio and Generalized Pairwise Comparisons (GPC) rank outcomes by clinical priority (such as Overall Survival first and Grade 3/4 toxicity second), comparing every treated patient with every control patient. Unlike the Win Ratio (P_win divided by P_loss), Net Treatment Benefit (P_win minus P_loss) keeps tied pairs in the denominator, giving an absolute net probability difference.

Source: Oncotarget 2016 (FOLFIRINOX GPC Analysis) · JACC: Basic to Translational Science 2023

How do you calculate Number Needed to Treat (NNT) and Number Needed to Harm (NNH) in oncology trials reporting Hazard Ratios?

You cannot compute NNT from a relative Hazard Ratio alone. Extract the fixed-time survival or event percentages in both arms (for example, 2-year mortality or Grade 3/4 toxicity rate) to calculate Absolute Risk Reduction (ARR = Control Event Rate minus Experimental Event Rate), then divide 100 by ARR(%). For harm, NNH = 100 divided by the percentage-point increase in Grade 3/4 adverse events; in the ALTAIR trial, a 69.7 percentage-point rise in hematologic Grade >=3 toxicity yields an NNH of 2.

Source: Med-Read.org Interactive NNT Workbench · Nature Medicine 2026 (ALTAIR NCT04457297)