← Clinical Trials
Melanoma Phase 3 Recurrence-Free Survival NCT02362594

KEYNOTE-054: Complete Statistical Analysis of Pembrolizumab in High-Risk Stage III Melanoma

An independent statistical review of the randomized phase 3 KEYNOTE-054 trial evaluating pembrolizumab versus placebo after complete resection of high-risk stage III melanoma, with emphasis on recurrence-free survival and the Cox proportional-hazards analysis reported in the registry.

Trial start: 2015-07-16  ·  Primary completion: 2018-01-08  ·  Enrollment: 1019
Scope of this record

This page separates reported trial results from statistical interpretation. The numerical results and trial characteristics presented here are restricted to the ClinicalTrials.gov record data and its posted statistical analyses.

Registry note: This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.

1. Trial at a Glance

KEYNOTE-054 was a randomized, double-blind, phase 3 parallel trial evaluating pembrolizumab versus placebo after complete resection of high-risk stage III melanoma. The ClinicalTrials.gov record reports two primary time-to-event endpoints, both analyzed using Cox proportional-hazards models.

1019
Enrollment
2 treatment arms
2
Primary endpoints
Both time-to-event
0.57
Overall RFS HR
98.4% CI 0.43–0.74
0.54
PD-L1+ RFS HR
95.0% CI 0.42–0.69
FeatureKEYNOTE-054
Trial nameKEYNOTE-054
NCT IDNCT02362594
PhasePhase 3
ConditionMelanoma
PopulationParticipants after complete resection of high-risk stage III melanoma
DesignRandomized, double-blind, parallel
AllocationRandomized
Arms2
InterventionsPembrolizumab and placebo
Enrollment1019
Primary purposeTreatment
Lead sponsorMerck Sharp & Dohme LLC
Sponsor typeIndustry
StatusActive, not recruiting
Trial start2015-07-16
Primary completion2018-01-08

2. Clinical Question

The central statistical question is whether the randomized comparison of pembrolizumab versus placebo after complete resection of high-risk stage III melanoma is associated with a difference in recurrence-free survival, evaluated at the registered 6-month time frame and analyzed using a time-to-event framework.

Population

Participants with high-risk stage III melanoma following complete resection, as described by the trial's brief title.

Intervention

Pembrolizumab, classified in the registry data as a biological intervention.

Comparator

Placebo, classified in the registry data as a drug intervention.

Primary question

Does pembrolizumab produce a different recurrence-free survival time-to-event distribution than placebo?

3. Trial Design

The trial used randomized allocation, a parallel design, and double masking. With 1019 participants enrolled and 2 arms, the fundamental statistical comparison is between participants assigned to pembrolizumab and those assigned to placebo.

01
Randomize1019 participants
02
AssignPembrolizumab or placebo
03
FollowRecurrence / death / censoring
04
AnalyzeTime-to-event distribution
05
EstimateCox hazard ratio
ARM 1

Pembrolizumab

  • Pembrolizumab
  • Biological intervention
  • Randomized treatment assignment
ARM 2

Placebo

  • Placebo
  • Drug intervention
  • Randomized comparator assignment
Why randomization and masking matter statistically. Randomization creates the basis for comparing outcomes according to assigned treatment rather than according to post-randomization choices. Double masking is a design feature intended to reduce the influence of treatment knowledge on trial conduct and assessment. The ClinicalTrials.gov record does not provide additional details on the masking procedures.

4. Endpoints

The registry lists 2 primary endpoints, both classified as time-to-event outcomes and both assessed at 6 months. The endpoint names are reproduced below in their registered form.

Primary endpointTime frameEndpoint type
Part 1: Percentage of Participants With Recurrence-Free Survival (RFS) At 6 Months Among All Participants 6 months Time-to-event
Part 1: Percentage of Participants With Recurrence-Free Survival (RFS) At 6 Months Among Participants With PD-L1-positive Tumor Expression 6 months Time-to-event

Registry definition of recurrence-free survival

RFS was defined as the time between the date of randomization and the date of first melanoma recurrence — local, regional, or distant metastasis — or death from any cause, whichever occurred first. For participants who remained alive and whose disease had not recurred, RFS was censored on the date of last visit or contact with disease assessments.

Endpoint-versus-analysis distinction: the registered endpoint is expressed as the percentage of participants with RFS at 6 months, but the posted formal analysis compares the RFS time-to-event distributions using a Cox regression model and reports a hazard ratio. These are related descriptions of the same time-to-event outcome, but they are not mathematically interchangeable quantities.

5. Statistical Methodology

Cox proportional-hazards model

The registry reports Regression, Cox as the method for both primary analyses. The statistical method is a Cox proportional-hazards model, placing the primary comparison in the survival-analysis family.

Model interpretation
HR = estimated hazard in pembrolizumab group ÷ estimated hazard in placebo group

An HR below 1 indicates a lower estimated instantaneous event rate in the pembrolizumab group relative to placebo under the fitted model. The HR is a relative time-to-event measure; it is not a percentage of participants who remain recurrence-free.

Covariate adjustment and stratification

The statistical analysis text specifies that treatment was included as a covariate in the Cox regression model and that the model was stratified by stage. The registry description gives the stage strata as IIIA (>1 mm metastasis), IIIB, IIIC 1–3 nodes, and IIIC ≥4 nodes, as indicated at randomization.

Stratification allows the baseline hazard function to differ across the specified stage strata while estimating a common treatment hazard ratio across those strata. This is different from simply putting stage into the model as an ordinary numerical covariate.

Analysis populations

For the overall primary endpoint, the analysis population was all randomized participants in Part 1. For the PD-L1-positive primary endpoint, the analysis population was all randomized participants in Part 1 with PD-L1-positive tumors.

Primary analysisAnalysis populationComparisonMethod
Overall RFS All randomized participants in Part 1 Pembrolizumab vs Placebo Cox proportional-hazards model, stratified by stage
PD-L1-positive RFS All randomized participants in Part 1 with PD-L1-positive tumors Pembrolizumab vs Placebo Cox proportional-hazards model, stratified by stage

Hypothesis type

The registry identifies the hypothesis type as superiority for the posted primary analyses. The reported effect measure is the hazard ratio, and the registry-reported confidence intervals are two-sided.

6. Results: Recurrence-Free Survival Among All Participants

The first primary analysis evaluated recurrence-free survival at 6 months among all randomized participants in Part 1. The registry reports a Cox proportional-hazards comparison between pembrolizumab and placebo.

Hazard ratio for recurrence or death

0.57

98.4% two-sided CI: 0.43–0.74   ·   P < 0.0001

Analysis: all randomized participants in Part 1

EndpointAnalysis populationEffect estimateConfidence intervalP-value
RFS at 6 months among all participants All randomized participants in Part 1 HR 0.57 98.4% CI 0.43–0.74 <0.0001
Clinical Biostats interpretation

What the estimate means: An HR of 0.57 means that, under the fitted Cox model, the estimated instantaneous rate of recurrence or death was 0.57 times that in the placebo group. Equivalently, this corresponds to a 43% lower estimated hazard because 1 − 0.57 = 0.43.

What it does not mean: HR 0.57 does not mean that 57% of participants remained recurrence-free, that 43% of participants avoided recurrence, or that every participant experienced a 43% reduction in risk. It is a model-based relative measure of the event rate over time.

Confidence interval: The 98.4% two-sided CI of 0.43–0.74 describes statistical uncertainty around the estimated hazard ratio under the specified analysis framework. It does not describe the range of outcomes that individual participants could experience.

P-value: The reported P < 0.0001 addresses the evidence against the null hypothesis under the trial's statistical testing framework. It does not measure the size of the treatment effect. Effect size is communicated by the HR and its confidence interval.

Cautions: Interpretation of a Cox HR depends on the proportional-hazards structure of the model. The analysis was also stratified by stage, so the HR should be understood as the treatment effect estimated within that stratified modeling framework rather than as an unadjusted ratio of simple event proportions.

Why the 6-month endpoint should not be treated as a simple proportion

The registered endpoint is phrased as a percentage with RFS at 6 months, but the formal analysis is explicitly a time-to-event analysis. Participants can experience recurrence or death before 6 months, while others can remain event-free and contribute follow-up until the relevant censoring point. The Cox model uses the ordering and timing of events rather than reducing every participant to a single binary indicator without regard to follow-up time.

7. Results: Recurrence-Free Survival Among Participants With PD-L1-positive Tumors

The second primary analysis restricted the analysis population to randomized Part 1 participants with PD-L1-positive tumors. The same general Cox modeling strategy was used, with stage as the stratification factor.

Hazard ratio for recurrence or death in PD-L1-positive tumors

0.54

95.0% two-sided CI: 0.42–0.69   ·   P < 0.0001

Analysis: all randomized Part 1 participants with PD-L1-positive tumors

EndpointAnalysis populationEffect estimateConfidence intervalP-value
RFS at 6 months among participants with PD-L1-positive tumor expression All randomized participants in Part 1 with PD-L1-positive tumors HR 0.54 95.0% CI 0.42–0.69 <0.0001
Clinical Biostats interpretation

What the estimate means: An HR of 0.54 means that, under the fitted Cox model for this analysis population, the estimated instantaneous rate of recurrence or death was 0.54 times that in the placebo group. Equivalently, this corresponds to a 46% lower estimated hazard.

What it does not mean: HR 0.54 does not mean that 54% of participants were recurrence-free or that 46% of individual patients necessarily benefited by exactly that amount. It is a relative model-based measure of the event hazard.

Confidence interval: The 95.0% two-sided CI of 0.42–0.69 expresses uncertainty around the estimated HR. A narrower interval would indicate greater statistical precision, while a wider interval would indicate more uncertainty; the interval itself is not a prediction interval for individual outcomes.

P-value: The reported P < 0.0001 is evidence against the relevant null hypothesis under the analysis framework. It is not a measure of clinical importance, effect size, or the probability that the null hypothesis is true.

Cautions: This is a restricted analysis population defined by PD-L1-positive tumor expression. The result should therefore not be treated as if it were numerically identical to the overall randomized population analysis. The Cox model was also stratified by stage, preserving the specified stage structure in the comparison.

8. Comparing the Two Primary Hazard Ratios

The two posted primary analyses produce HR estimates of 0.57 in the overall randomized population and 0.54 among randomized participants with PD-L1-positive tumors. These estimates are numerically close, but comparing their point estimates alone is not a formal test of whether the treatment effect differs between the two populations.

AnalysisHRConfidence intervalP-valuePopulation
All participants 0.57 98.4% CI 0.43–0.74 <0.0001 All randomized participants in Part 1
PD-L1-positive 0.54 95.0% CI 0.42–0.69 <0.0001 Randomized Part 1 participants with PD-L1-positive tumors
Do not infer interaction from two subgroup estimates. A difference between two estimated hazard ratios does not by itself demonstrate treatment-effect modification. A formal claim that the effect differs by PD-L1 status would require an appropriate interaction or heterogeneity analysis. The ClinicalTrials.gov record does not report such an interaction test.

9. Safety

The ClinicalTrials.gov record reports serious adverse events by randomized treatment arm using affected participants over the corresponding at-risk populations.

Safety measurePembrolizumabPlacebo
Serious adverse events, affected / at risk 128 / 509 82 / 502

These figures describe serious adverse events by arm as reported in the ClinicalTrials.gov record. They should not be substituted for the recurrence-free survival analysis because safety and efficacy are different outcome domains with different statistical interpretations.

Denominator matters. The registry safety figures are presented as affected participants divided by participants at risk: 128/509 for pembrolizumab and 82/502 for placebo. The ClinicalTrials.gov record does not provide additional definitions, event-level detail, severity distributions, exposure-adjusted rates, or formal between-arm safety hypothesis tests.

10. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

RFS is a time-to-event endpoint because the outcome depends not only on whether recurrence or death occurs, but also on when it occurs. The Cox model is designed to compare event hazards between treatment groups while accommodating censored observations and allowing specified covariates or strata to enter the analysis.

What does an HR of 0.57 mean?

An HR of 0.57 means the estimated hazard in the pembrolizumab group is 57% of the estimated hazard in the placebo group under the fitted model. It can also be expressed as a 43% lower estimated hazard. It is not a statement that 43% of patients avoided recurrence.

Why was stage used for stratification?

The posted analysis notes specify stratification by stage at randomization: IIIA (>1 mm metastasis), IIIB, IIIC 1–3 nodes, and IIIC ≥4 nodes. Stratified Cox modeling permits the baseline hazard to differ across these stage strata while estimating the treatment effect within the stratified framework.

Why is a confidence interval reported with the hazard ratio?

A point estimate alone does not communicate statistical precision. The confidence interval provides a range of values representing uncertainty around the estimated HR under the specified statistical framework. For the overall analysis, the interval was 0.43–0.74 at the 98.4% confidence level; for the PD-L1-positive analysis, it was 0.42–0.69 at the 95.0% confidence level.

Why doesn't the P-value measure the treatment effect?

The P-value addresses evidence against a null hypothesis under the statistical model and testing framework. It is affected by both the observed data and the amount of information available. The magnitude of the estimated effect is communicated by the HR, while the confidence interval communicates uncertainty around that estimate.

Why does censoring matter for RFS?

Some participants can remain alive without a documented recurrence when follow-up ends. The registry definition specifies that RFS is censored at the date of the last visit or contact with disease assessments for participants who remain alive and whose disease has not recurred. A survival analysis can therefore use their available event-free follow-up without treating the observation as if recurrence had occurred.

Why should the overall and PD-L1-positive analyses be kept separate?

The analyses use different populations. The first includes all randomized participants in Part 1; the second includes only randomized participants with PD-L1-positive tumors. A restricted analysis can answer a more specific question, but its result should not automatically be interpreted as the treatment effect in the entire randomized population.

11. Understanding the Hazard Ratio More Deeply

Relative effect

An HR below 1 indicates a lower estimated instantaneous event rate in the pembrolizumab group. For KEYNOTE-054, the reported overall HR of 0.57 and PD-L1-positive HR of 0.54 are both below 1.

Absolute versus relative information

The statistical analyses posted on ClinicalTrials.gov report hazard ratios rather than absolute RFS percentages at specific time points. Consequently, the HR should not be converted into an absolute probability of remaining recurrence-free. An absolute 6-month RFS percentage would answer a different question and cannot be derived from the HR alone.

Timing matters

A hazard ratio summarizes relative event rates within a time-to-event model. It does not identify the exact number of recurrences avoided, the median RFS, or the probability that a particular individual will remain recurrence-free. Those quantities require corresponding survival estimates or additional patient-level information.

Proportional-hazards assumption

The Cox proportional-hazards model is based on a proportional-hazards structure for the treatment effect. A single HR is therefore most naturally interpreted as a model-based summary of relative hazards rather than as a literal constant percentage difference in risk at every moment for every patient.

12. Stratified Analysis in This Trial

The posted analysis notes identify stage as the stratification variable in both primary Cox models. The stage categories were specified at randomization and were used to structure the survival comparison.

Stage stratumRegistry-specified category
Stage IIIA>1 mm metastasis
Stage IIIBIIIB
Stage IIIC1–3 nodes
Stage IIIC≥4 nodes

Stratification is useful when an important baseline characteristic is expected to influence the underlying event rate. Rather than forcing a single baseline hazard shape across all stage groups, a stratified Cox model can allow the baseline hazard to vary between strata while estimating the treatment coefficient across the stratified analysis.

Conceptual Cox model
h(t) = h0,stratum(t) × exp(β × Treatment)

Here the baseline hazard can differ by stage stratum, while the treatment coefficient β determines the estimated treatment hazard ratio through exp(β).

13. What the Confidence Intervals Tell Us

The overall primary analysis reports a 98.4% two-sided confidence interval of 0.43–0.74. The PD-L1-positive analysis reports a 95.0% two-sided confidence interval of 0.42–0.69.

Overall population

HR 0.57 with a 98.4% two-sided CI of 0.43–0.74. The interval quantifies uncertainty around the estimated relative hazard under the reported Cox analysis.

PD-L1-positive population

HR 0.54 with a 95.0% two-sided CI of 0.42–0.69. This is a separate estimate for the prespecified analysis population reported in the registry.

The different confidence levels should be retained exactly as reported. A 98.4% interval and a 95.0% interval are not interchangeable, and their widths should not be compared without considering both the confidence level and the underlying information.

14. P-Values and Superiority Testing

Both posted primary analyses have a reported P-value of <0.0001, and the registry identifies the hypothesis type as superiority.

Primary analysisHypothesis typeP-valueEffect measure
RFS at 6 months, all participants Superiority <0.0001 Hazard ratio
RFS at 6 months, PD-L1-positive participants Superiority <0.0001 Hazard ratio

A small P-value does not tell us whether the estimated treatment effect is large or clinically important. That question requires examination of the effect estimate, its confidence interval, the endpoint definition, the analysis population, and the clinical context. Statistical significance and effect magnitude are related but distinct concepts.

15. Multiplicity and the Two Primary Endpoints

The ClinicalTrials.gov record identifies 2 primary endpoints and 2 posted primary statistical analyses. Both use the superiority hypothesis type and both report hazard ratios from Cox models.

Primary endpointFormal analysisEffect measureP-value
RFS at 6 months among all participants Cox regression stratified by stage HR 0.57 <0.0001
RFS at 6 months among PD-L1-positive participants Cox regression stratified by stage HR 0.54 <0.0001
Multiplicity limitation: The ClinicalTrials.gov record identifies two primary endpoints and provide their formal analyses, but they do not provide the full multiplicity-control procedure, alpha allocation, testing hierarchy, or interim-analysis plan. Those details should not be reconstructed from the reported P-values.

16. Missing Data, Censoring, and What the Registry Supports

The RFS definition explicitly describes censoring for participants who remain alive and whose disease has not recurred: their RFS is censored at the date of the last visit or contact with disease assessments.

This is an important distinction from ordinary missing-data handling. In a time-to-event analysis, a participant without an observed event by the end of available follow-up is not automatically treated as having a missing outcome. Instead, the participant can contribute information up to the censoring time.

What is not specified in the ClinicalTrials.gov record: the registry extract does not provide a detailed missing-data or imputation strategy, a specific sensitivity-analysis framework for missing assessments, or a detailed censoring algorithm beyond the RFS definition quoted above. No such procedures are added here.

17. Interim Analysis, Crossover, and Bayesian Methods

Design topicWhat the ClinicalTrials.gov record supports
Interim analysisNo interim-analysis procedure is provided in the ClinicalTrials.gov record.
CrossoverNo crossover procedure is provided in the ClinicalTrials.gov record.
Factorial designThe design model is parallel; no factorial structure is reported.
Non-inferiority marginNot applicable to the reported superiority hypothesis; no non-inferiority margin is provided.
Bayesian methodsNo Bayesian method is reported in the statistical analyses posted on ClinicalTrials.gov.

The absence of a procedure in the ClinicalTrials.gov record should not be interpreted as proof that a protocol or statistical analysis plan contained no additional methodological detail. It means only that those details are not available in the ClinicalTrials.gov record.

18. Limitations

19. Why This Trial Matters Statistically

KEYNOTE-054 is a useful statistical teaching case because the ClinicalTrials.gov record connects a randomized oncology comparison with two related but distinct analysis populations and a common time-to-event methodology. The page also illustrates why the wording of an endpoint, the analysis population, the model, and the effect measure must all be read together.

ConceptHow it appears in KEYNOTE-054
RandomizationParticipants were randomized to pembrolizumab or placebo.
BlindingThe trial was double-blind.
Parallel designThe registry identifies the design model as parallel.
Time-to-event endpointBoth primary endpoints are classified as time-to-event outcomes.
CensoringRFS is censored at the last visit/contact with disease assessments for eligible event-free participants.
Cox modelBoth formal primary analyses use Cox regression.
Hazard ratioHR is the reported effect measure for both primary analyses.
Stratified analysisThe Cox model is stratified by stage at randomization.
Covariate adjustmentTreatment is included as a covariate in the Cox model.
Confidence intervalsBoth primary HR estimates are accompanied by two-sided confidence intervals.
Superiority testingThe registry identifies the hypothesis type as superiority.
Subpopulation analysisThe second primary endpoint restricts the population to PD-L1-positive tumors.

20. A Practical Reading of the KEYNOTE-054 Results

A disciplined statistical reading proceeds in layers.

  1. Start with the endpoint. RFS is defined as time from randomization to first melanoma recurrence or death, with specified censoring for participants remaining alive without recurrence.
  2. Identify the population. The first analysis includes all randomized Part 1 participants; the second is limited to randomized Part 1 participants with PD-L1-positive tumors.
  3. Identify the model. Both analyses use Cox regression with treatment as a covariate and stage as the stratification factor.
  4. Read the effect measure. The overall HR is 0.57; the PD-L1-positive HR is 0.54.
  5. Read the uncertainty. The overall 98.4% CI is 0.43–0.74, while the PD-L1-positive 95.0% CI is 0.42–0.69.
  6. Read the P-value separately. Both are reported as <0.0001, providing evidence against the corresponding null hypothesis but not quantifying effect magnitude.
  7. Stop where the data stop. The ClinicalTrials.gov record does not provide enough information to claim median RFS, absolute 6-month RFS percentages, subgroup heterogeneity, or additional secondary endpoint results.

This sequence prevents a common statistical error: jumping directly from a P-value to a clinical conclusion without first establishing exactly what was measured, who was analyzed, and how the treatment effect was estimated.

21. Related Tutorials

Learn more about the methods used in this trial:

22. Related Calculators

23. Sources

Continue with Clinical Biostats

Explore statistical tutorials, calculators, and additional clinical trial analyses covering the methods used across randomized clinical research.

24. Record Summary

KEYNOTE-054 provides a focused example of time-to-event analysis in a randomized phase 3 oncology trial. The ClinicalTrials.gov record identifies a randomized, double-blind, parallel design with 1019 participants, two primary recurrence-free survival endpoints, and Cox proportional-hazards analyses stratified by stage. The overall randomized analysis reports an HR of 0.57 with a 98.4% two-sided CI of 0.43–0.74 and P < 0.0001. The PD-L1-positive analysis reports an HR of 0.54 with a 95.0% two-sided CI of 0.42–0.69 and P < 0.0001.

The most important statistical lesson is that these numbers should be interpreted together with the endpoint definition, censoring rule, analysis population, stratification scheme, and Cox model. A hazard ratio summarizes a relative time-to-event effect; it does not substitute for an absolute survival probability, a median event time, or an individual-level prediction. Likewise, a small P-value provides evidence against a null hypothesis but does not measure the magnitude or clinical importance of an effect.

Clinical Biostats methodology: A trial-results page should distinguish reported evidence from statistical interpretation. Where the ClinicalTrials.gov record does not provide a numerical result, this page does not manufacture one from external publications, assumptions, or reconstructed calculations.