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
CheckMate 915 was a randomized, double-blind, parallel-group phase 3 trial evaluating nivolumab plus ipilimumab versus nivolumab alone in participants with completely resected stage IIIb/c/d or stage IV melanoma with no evidence of disease. The two registered primary endpoints were recurrence-free survival (RFS) in all randomized participants and RFS in the subgroup with PD-L1 expression level <1%.
| Feature | CheckMate 915 |
|---|---|
| Phase | Phase 3 |
| Condition | Melanoma |
| Population | Completely resected stage IIIb/c/d or stage IV melanoma with no evidence of disease |
| Design | Randomized, double-blind, parallel |
| Allocation | Randomized |
| Primary purpose | Treatment |
| Primary endpoints | Recurrence-free survival in all randomized participants; recurrence-free survival in all randomized participants with PD-L1 expression level <1% |
| Primary endpoint type | Time-to-event |
| Primary analysis | Stratified Cox proportional-hazards model |
| Effect measure | Hazard ratio |
| Hypothesis type | Superiority |
| Status | Completed |
| ClinicalTrials.gov | NCT03068455 |
2. Clinical Question
The central statistical question was whether treatment with nivolumab plus ipilimumab produced a different recurrence-free survival experience from nivolumab alone after complete surgical removal of stage IIIb/c/d or stage IV melanoma with no evidence of disease.
Population
Participants with completely resected stage IIIb/c/d or stage IV no evidence of disease (NED) melanoma.
Intervention
Nivolumab combined with ipilimumab.
Comparator
Nivolumab by itself.
Primary question
Does nivolumab plus ipilimumab improve recurrence-free survival relative to nivolumab alone?
3. Trial Design
Nivolumab + ipilimumab
- Nivolumab
- Ipilimumab
- Compared with nivolumab alone for the registered efficacy endpoints
Nivolumab
- Nivolumab alone
- Comparator for the registered efficacy analyses
The registry describes the study as randomized, double-blind, and parallel. Those features are important statistically because randomization establishes the primary comparison, masking can reduce differential assessment or treatment-related bias, and the parallel structure means participants remain associated with their randomized treatment group rather than crossing between randomized interventions as part of the design.
4. Trial Timeline
Study start
The registry lists April 11, 2017 as the study start date.
Primary completion
The registry lists June 12, 2020 as the primary completion date.
Registry status
The study is listed as completed, with results posted and 9 statistical analyses reported in the registry.
5. Primary Endpoints
| Endpoint | Time frame | Definition |
|---|---|---|
| Recurrence-free Survival (RFS) - All Randomized Participants | From randomization to Primary Completion Date (up to approximately 3 years) | Time between randomization and the date of first recurrence (local, regional or distant metastasis), new primary melanoma (including melanoma in situ), or death from any cause, whichever occurred first. Median values based on Kaplan-Meier estimates. |
| Recurrence-free Survival (RFS) - All Randomized Participants With PD-L1 Expression Level < 1% | From randomization to Primary Completion Date (up to approximately 3 years) | Time between randomization and the date of first recurrence (local, regional or distant metastasis), new primary melanoma (including melanoma in situ), or death from any cause, whichever occurred first. Median based on Kaplan-Meier estimates. |
Both primary endpoints are time-to-event endpoints. The event definition is composite: the first qualifying recurrence, a new primary melanoma, or death determines the RFS event time. Participants who do not experience an event during the relevant observation period contribute censored follow-up rather than an observed event time.
6. Secondary Endpoints Reported in the Registry
The registry also reports overall survival and recurrence-free survival analyses involving baseline tumor PD-L1 expression. These analyses provide additional time-to-event comparisons beyond the two primary RFS analyses.
| Secondary endpoint | Time frame | Effect measure |
|---|---|---|
| Overall Survival (OS) - All Randomized Participants | From randomization to date of death (up to approximately 45 months) | Hazard ratio |
| Overall Survival (OS) - All Randomized Participants With PD-L1 Expression Level < 1% | From randomization to date of death (up to approximately 45 months) | Hazard ratio |
| Recurrence-free Survival by Baseline Tumor PD-L1 Expression | From randomization to Study Completion Date (up to approximately 45 months) | Hazard ratio |
7. Statistical Methodology
Kaplan-Meier estimation
The registry defines RFS using time from randomization to the first qualifying event and states that median values are based on Kaplan-Meier estimates. Kaplan-Meier estimation is appropriate for time-to-event data because it uses the observed event times while retaining information from participants whose follow-up ends before an event occurs.
Here, di represents the number of events at event time ti, while ni represents the number at risk immediately before that time.
Stratified Cox proportional-hazards model
The two primary RFS analyses were conducted using a stratified Cox proportional-hazards model. The same model family was reported for the two secondary overall-survival analyses. A Cox model estimates a relative hazard between treatment groups while accounting for the time-to-event structure of the outcome.
For the primary analyses, the registry specifies that the hazard ratio is Nivolumab + Ipilimumab over Nivolumab. Thus an HR below 1 corresponds to a lower estimated instantaneous event hazard in the combination arm under the fitted model.
Stratified analysis
The primary analysis text identifies a stratified Cox proportional-hazards model. Stratification allows the baseline hazard to differ across the specified strata while estimating the treatment comparison within the Cox framework. The registry does not identify the specific stratification variables in the normalized analysis description.
Superiority framework
The primary analyses are identified as superiority hypotheses. This means the statistical question is whether the randomized treatment comparison demonstrates a difference in the prespecified direction of interest rather than whether two treatments are sufficiently similar to establish non-inferiority.
Analysis populations
The primary RFS population was all randomized participants with completely resected stage IIIb/c/d or stage IV no evidence of disease melanoma. The PD-L1 <1% primary analysis was restricted to randomized participants with PD-L1 expression level <1% and the same melanoma/NED eligibility characteristics.
8. Primary Results: Overall Recurrence-Free Survival
The first primary analysis compares recurrence-free survival among all randomized participants with completely resected stage IIIb/c/d or stage IV no evidence of disease melanoma.
Nivolumab + ipilimumab vs nivolumab
95% CI: 0.78–1.04 · Two-sided 95% confidence interval
Stratified Cox proportional-hazards model
| Primary endpoint | Combination | Nivolumab | Effect estimate |
|---|---|---|---|
| RFS — all randomized participants | Nivolumab + ipilimumab | Nivolumab | HR 0.90 (95% CI 0.78–1.04) |
An HR of 0.90 means that the fitted Cox model estimated the instantaneous recurrence/death hazard in the nivolumab-plus-ipilimumab group at approximately 90% of the corresponding hazard in the nivolumab group. Expressed as a relative model-based quantity, this is consistent with an estimated hazard approximately 10% lower in the combination group.
The HR does not mean that 10% fewer participants experienced recurrence, that individual participants had a 10% reduction in their personal probability of recurrence, or that the absolute difference in recurrence-free survival was 10 percentage points.
The 95% CI of 0.78–1.04 describes uncertainty around the estimated hazard ratio under the model and sampling framework. Because the interval includes 1, the registry's reported estimate does not provide a confidence interval that excludes equal hazards.
The p-value is not reported in the statistical analysis information reported by the registry record. It therefore should not be substituted with a reconstructed or calculated value. In a Cox analysis, a p-value would address a hypothesis about the treatment effect under the prespecified testing framework; it would not measure the size or clinical importance of the HR.
Because this is a Cox proportional-hazards analysis, interpretation also depends on the model's proportional-hazards structure. A single HR summarizes the relative hazard over the modeled follow-up and does not by itself describe how the treatment contrast changes at every point in time.
9. Primary Results: RFS in Participants With PD-L1 Expression Level <1%
The second primary analysis focuses on randomized participants with PD-L1 expression level <1%, with PD-L1 expression levels based on Interactive Response Technology (IRT), and with completely resected stage IIIb/c/d or stage IV no evidence of disease melanoma.
RFS in the PD-L1 <1% population
95% CI: 0.75–1.14 · Two-sided 95% confidence interval
Stratified Cox proportional-hazards model
| Primary endpoint | Combination | Nivolumab | Effect estimate |
|---|---|---|---|
| RFS — PD-L1 expression level <1% | Nivolumab + ipilimumab | Nivolumab | HR 0.92 (95% CI 0.75–1.14) |
An HR of 0.92 corresponds to an estimated instantaneous recurrence/death hazard approximately 92% as large in the combination group as in the nivolumab group. As a relative hazard interpretation, this corresponds to an estimated hazard approximately 8% lower under the fitted model.
The estimate is not an 8% absolute reduction in recurrence, an 8% improvement in the chance of remaining recurrence-free, or an estimate of the proportion of patients who benefit.
The 95% CI of 0.75–1.14 is wider than the point estimate alone and includes 1. This indicates uncertainty that encompasses both a lower hazard and a possible higher hazard under the model. The confidence interval is therefore essential for understanding the precision of the treatment estimate.
The registry identifies the analysis as a superiority analysis, but no p-value is reported in the posted statistical-analysis information. A p-value should not be inferred from the confidence interval by reconstructing an unreported test.
This analysis is also a subgroup-defined primary endpoint. Restricting the analysis to PD-L1 expression level <1% changes the population being compared, so its interpretation should remain specific to that population rather than being generalized automatically to every participant in the trial.
10. Secondary Overall Survival Results
Overall Survival — All Randomized Participants
OS hazard ratio
95% CI: 0.80–1.32 · Two-sided 95% confidence interval
Stratified Cox proportional-hazards model
The OS HR of 1.03 places the estimated hazard of death in the nivolumab-plus-ipilimumab group at approximately 103% of that in the nivolumab group under the fitted model. An HR close to 1 indicates little relative separation in the estimated instantaneous hazards at the model level.
The 95% CI of 0.80–1.32 is compatible with a range of possible treatment effects, including lower and higher hazards. The interval is therefore more informative than the point estimate alone.
This secondary analysis is distinct from the primary RFS analyses. Overall survival measures time to death from any cause, whereas RFS counts recurrence, a new primary melanoma, or death, whichever occurs first.
Overall Survival — PD-L1 Expression Level <1%
OS hazard ratio in PD-L1 <1%
95% CI: 0.85–1.73 · Two-sided 95% confidence interval
Stratified Cox proportional-hazards model
Within the randomized participants with PD-L1 expression level <1%, the reported OS HR of 1.22 corresponds to an estimated instantaneous hazard of death approximately 122% of that in the comparator group under the fitted model.
The 95% CI of 0.85–1.73 is broad and includes 1. It therefore encompasses both a possible lower and a possible higher hazard for the combination group. This uncertainty is important when interpreting a subgroup-defined secondary endpoint.
11. Secondary RFS Analyses by Baseline Tumor PD-L1 Expression
The registry reports four additional RFS analyses by baseline tumor PD-L1 expression. These analyses use all randomized participants with variable PD-L1 expression levels, with PD-L1 expression levels based on the clinical database. The registry does not report the statistical method for these four analyses.
| PD-L1 category | HR | 95% CI | Hypothesis |
|---|---|---|---|
| <1% tumor PD-L1 expression | 0.91 | 0.73–1.14 | Superiority |
| ≥1% tumor PD-L1 expression | 0.95 | 0.76–1.18 | Superiority |
| ≥5% tumor PD-L1 expression | 0.98 | 0.71–1.34 | Superiority |
| <5% tumor PD-L1 expression | 0.92 | 0.77–1.10 | Superiority |
| Non-quantifiable tumor PD-L1 expression | 0.76 | 0.38–1.51 | Superiority |
The point estimates range from 0.76 to 0.98, but the corresponding confidence intervals are important because each interval includes 1. The widths of the intervals also vary substantially, particularly for the non-quantifiable PD-L1 category.
These analyses describe treatment-effect estimates within defined PD-L1 categories. They do not, by themselves, establish that the treatment effect differs between PD-L1 categories. A formal claim of effect modification would require an appropriate interaction or heterogeneity analysis.
12. Safety: Serious Adverse Events by Arm
The registry reports serious adverse events by randomized treatment arm. In the nivolumab-plus-ipilimumab arm, 374 participants were affected among 916 at risk. In the nivolumab arm, 242 participants were affected among 917 at risk.
| Safety measure | Nivolumab + ipilimumab | Nivolumab |
|---|---|---|
| Participants with serious adverse events | 374 / 916 | 242 / 917 |
These are safety counts rather than time-to-event efficacy estimates. They should therefore be interpreted separately from the RFS and OS hazard ratios. The registry reports the affected and at-risk counts directly; no additional safety effect measure is reported here.
13. Statistical Methods Explained
Why is RFS a time-to-event endpoint?
RFS incorporates both whether a participant experiences an event and when the event occurs. In CheckMate 915, the event is the first recurrence, new primary melanoma, or death from any cause. Participants without an event during observed follow-up can contribute censored information, making conventional binary analysis less informative than a time-to-event framework.
Why was a Cox proportional-hazards model used?
The registry identifies the Cox proportional-hazards model for the primary RFS analyses and the overall-survival analyses. The model provides a relative hazard estimate while using the timing of events and censoring information. Its principal effect measure is the hazard ratio.
What does an HR of 0.90 mean?
For the primary all-randomized RFS analysis, an HR of 0.90 means that the estimated instantaneous recurrence/death hazard for nivolumab plus ipilimumab was approximately 90% of that for nivolumab under the fitted model. It is not an absolute risk difference or a statement that exactly 10% fewer participants experienced recurrence.
Why does the confidence interval matter as much as the HR?
The point estimate is only one estimate from the observed trial data. The 95% confidence interval describes statistical uncertainty around that estimate. For the primary all-randomized RFS analysis, the interval is 0.78–1.04, so the data are compatible with hazard ratios below and above 1 under the model.
Why should subgroup hazard ratios not be compared by eye?
An HR of 0.91 in one PD-L1 category and 0.98 in another does not establish that treatment works differently in those categories. Differences between subgroup estimates can arise from sampling variation. Formal effect modification requires an appropriate interaction or heterogeneity analysis.
What is the difference between RFS and OS?
RFS in this registry record counts recurrence, a new primary melanoma, or death, whichever occurs first. OS instead measures time from randomization to death. Consequently, a participant can experience an RFS event without having died, whereas death is the event defining OS.
14. Understanding the Hazard Ratio
For the primary analyses, Arm A is nivolumab plus ipilimumab and Arm B is nivolumab. An HR of 0.90 therefore corresponds to a lower estimated hazard in the combination group relative to nivolumab alone.
A hazard ratio is a relative measure of event rates over time. It should not be interpreted as a probability, a percentage of patients who benefit, or a direct measure of absolute survival improvement.
HR below 1
The estimated instantaneous event hazard is lower in the numerator group under the fitted model.
HR equal to 1
The model's relative hazard estimate indicates no difference in instantaneous hazard between the groups.
HR above 1
The estimated instantaneous event hazard is higher in the numerator group under the fitted model.
HR is not absolute risk
An HR does not tell us directly how many additional or fewer participants experience an event by a particular time.
15. Confidence Intervals and Statistical Precision
The two primary analyses provide a useful illustration of why confidence intervals belong alongside point estimates.
| Primary analysis | HR | 95% CI | Width of interval |
|---|---|---|---|
| RFS — all randomized participants | 0.90 | 0.78–1.04 | 0.26 |
| RFS — PD-L1 expression level <1% | 0.92 | 0.75–1.14 | 0.39 |
The PD-L1 <1% analysis has a wider confidence interval than the all-randomized analysis. This illustrates a basic statistical principle: restricting an analysis population can reduce the information available for estimating a treatment effect and can consequently increase uncertainty.
A 95% confidence interval around a hazard ratio is a statement about uncertainty in the estimated population-level treatment effect under the statistical model. It is not the range within which an individual patient's hazard must fall, and it does not mean that 95% of individual treatment effects lie inside the interval.
16. Primary and Secondary Analyses: Keeping the Hierarchy Clear
The registry identifies two primary endpoints and seven additional posted statistical analyses. Keeping their roles distinct is important when interpreting the evidence.
| Analysis group | Role | Key statistical feature |
|---|---|---|
| All-randomized RFS | Primary | Stratified Cox model; HR 0.90; 95% CI 0.78–1.04 |
| PD-L1 <1% RFS | Primary | Stratified Cox model; HR 0.92; 95% CI 0.75–1.14 |
| All-randomized OS | Secondary | Stratified Cox model; HR 1.03; 95% CI 0.80–1.32 |
| PD-L1 <1% OS | Secondary | Stratified Cox model; HR 1.22; 95% CI 0.85–1.73 |
| RFS by baseline tumor PD-L1 expression | Secondary | HRs reported for five PD-L1 categories; the statistical method not reported |
This hierarchy matters because a secondary or subgroup analysis should not automatically be treated as though it carried the same confirmatory role as a prespecified primary endpoint. Multiplicity can also arise when multiple endpoints, subgroups, or analyses are examined, and the appropriate interpretation depends on the prespecified testing strategy.
17. Statistical Interpretation of the Secondary Results
The all-randomized OS HR was 1.03 with a 95% CI of 0.80–1.32. The point estimate is close to 1, while the confidence interval spans values below and above 1. The appropriate statistical description is therefore the estimated hazard ratio together with its uncertainty, rather than a categorical conclusion based on the point estimate alone.
The OS HR was 1.22 with a 95% CI of 0.85–1.73. The interval is relatively broad and includes 1, illustrating why the point estimate should not be interpreted without considering statistical precision.
The reported RFS HRs were 0.91, 0.95, 0.98, 0.92, and 0.76 across the listed PD-L1 categories. Their 95% confidence intervals range from relatively narrow intervals such as 0.73–1.14 to the wider 0.38–1.51 interval for non-quantifiable PD-L1 expression. Precision therefore differs materially across these analyses.
18. Proportional-Hazards Considerations
The Cox model is based on the proportional-hazards framework. Conceptually, this means that the relative hazard between treatment groups is modeled as remaining stable over time after accounting for the model structure.
When proportional hazards are reasonably appropriate, the HR provides a concise summary of the relative treatment effect. If the treatment effect changes materially over time, a single HR can obscure that pattern. This is why time-to-event interpretation is strongest when the HR is considered together with Kaplan-Meier estimates, clinically meaningful time points, and the observed follow-up structure.
19. Censoring and Time-to-Event Interpretation
RFS and OS are not simply counts of whether an event occurred. The timing of the event matters. Participants who remain event-free through the end of observed follow-up may be censored, allowing their available follow-up time to contribute information without treating the absence of an observed event as lifelong event-free survival.
Observed event
A participant experiences the endpoint event during the defined follow-up period.
Censored observation
The participant contributes information up to the censoring time without an observed endpoint event.
Kaplan-Meier
Uses event and risk-set information to estimate the survival or recurrence-free survival function.
Cox model
Uses time-to-event information to estimate a relative hazard between treatment groups.
This structure is why conventional comparisons of simple event percentages can fail to capture the full statistical information in a time-to-event endpoint.
20. PD-L1 Subgroup Analysis: What Can and Cannot Be Concluded
The registry reports several RFS analyses categorized by baseline tumor PD-L1 expression. These estimates are useful for describing the treatment comparison within specified PD-L1 categories, but subgroup analyses require particular care.
| PD-L1 category | HR | 95% CI |
|---|---|---|
| <1% | 0.91 | 0.73–1.14 |
| ≥1% | 0.95 | 0.76–1.18 |
| ≥5% | 0.98 | 0.71–1.34 |
| <5% | 0.92 | 0.77–1.10 |
| Non-quantifiable | 0.76 | 0.38–1.51 |
Two common statistical mistakes are worth avoiding. First, the magnitude of one subgroup HR should not be treated as proof that the treatment effect differs from another subgroup HR. Second, a confidence interval that includes 1 should not be described simply as proof that there is "no effect"; it indicates that the estimate is compatible with a range of effects that includes the null value under the specified model.
21. Why This Trial Matters Statistically
CheckMate 915 is a useful statistical teaching case because it combines randomized treatment comparison, double masking, time-to-event endpoints, Kaplan-Meier estimation, stratified Cox modeling, hazard ratios, confidence intervals, and prespecified population definitions within a single phase 3 trial.
| Concept | How it appears in CheckMate 915 |
|---|---|
| Randomization | Participants were randomly allocated to two parallel treatment arms. |
| Blinding | The registry identifies the study as double-blind. |
| Time-to-event endpoints | RFS and OS use time from randomization to a defined event. |
| Kaplan-Meier estimation | The registry defines RFS medians using Kaplan-Meier estimates. |
| Hazard ratio | The primary and several secondary analyses use HR as the effect measure. |
| Cox proportional-hazards model | Used for both primary RFS analyses and the posted OS analyses. |
| Stratified analysis | The primary statistical analysis text identifies a stratified Cox model. |
| Confidence intervals | All posted primary and secondary HR estimates include two-sided 95% CIs. |
| Subgroup analysis | Several analyses evaluate RFS or OS in populations defined by PD-L1 expression. |
| Safety analysis | Serious adverse events are reported by treatment arm using affected/at-risk counts. |
22. Limitations
- Incomplete numerical characterization: the registry reports selected statistical estimates, but not every descriptive feature of a time-to-event distribution. For example, the posted statistical analyses do not provide median RFS values for the primary comparisons.
- Unreported p-values: the posted primary statistical analyses provide hazard ratios and confidence intervals but do not provide p-values. Statistical significance should therefore not be inferred by reconstructing an unreported test.
- Subgroup precision: PD-L1-defined analyses generally involve narrower populations than the all-randomized analysis, which can increase uncertainty.
- Method reporting for PD-L1 strata: the registry does not report the statistical method for the five additional PD-L1-defined RFS analyses.
- Hazard-ratio assumptions: Cox HRs are model-based relative measures and depend on the proportional-hazards framework.
- Composite RFS endpoint: RFS combines recurrence, new primary melanoma, and death into a single endpoint. The HR therefore describes the composite event rather than one isolated clinical event.
- Different populations: the all-randomized and PD-L1 <1% analyses answer questions in different analysis populations and should not be treated as interchangeable estimates.
- Safety and efficacy are distinct: serious adverse-event counts describe a different dimension of the treatment comparison than RFS and OS hazard ratios.
23. Reading the Primary Results as a Statistical Story
The two primary RFS estimates are similar in magnitude: 0.90 in all randomized participants and 0.92 among participants with PD-L1 expression level <1%. Their confidence intervals are different in width, with the PD-L1 <1% analysis having the wider interval.
This distinction illustrates an important principle in clinical-trial interpretation. A point estimate summarizes the observed treatment contrast, while the confidence interval communicates the precision of that estimate. The two should be read together.
Effect estimate
The HR summarizes the relative instantaneous event hazard under the Cox model.
Precision
The 95% CI indicates uncertainty around the estimated HR.
Endpoint definition
RFS includes recurrence, new primary melanoma, or death, whichever occurs first.
Population
The estimate applies to the population defined for that particular analysis.
24. Related Tutorials
Learn more about the methods used in this trial:
25. Related Calculators
26. Sources
- ClinicalTrials.gov: NCT03068455 — CheckMate 915.
- PubMed: PMID 40645660.
- PubMed: PMID 36162037.
- PubMed: PMID 35512259.
Continue through Clinical Biostats
Explore statistical tutorials, calculators, and additional clinical trial analyses covering randomized designs, survival analysis, regression, and clinical-trial methodology.
27. Record Summary
CheckMate 915 provides a useful example of randomized time-to-event analysis in oncology. The trial enrolled 1844 participants in a randomized, double-blind, parallel-group phase 3 design comparing nivolumab plus ipilimumab with nivolumab alone after complete surgical removal of stage IIIb/c/d or stage IV melanoma with no evidence of disease.
The two primary endpoints were recurrence-free survival in all randomized participants and recurrence-free survival among participants with PD-L1 expression level <1%. Both were analyzed using stratified Cox proportional-hazards models. The reported hazard ratios were 0.90 (95% CI 0.78–1.04) and 0.92 (95% CI 0.75–1.14), respectively.
The registry also reports secondary overall-survival analyses with HRs of 1.03 (95% CI 0.80–1.32) among all randomized participants and 1.22 (95% CI 0.85–1.73) among participants with PD-L1 expression level <1%. Additional RFS analyses by baseline tumor PD-L1 expression report HRs ranging from 0.76 to 0.98, with confidence intervals that reflect varying degrees of statistical precision.
The principal statistical lesson is that these estimates should be interpreted as a connected set of endpoint definition, analysis population, model, effect measure, and uncertainty. A hazard ratio without its confidence interval is incomplete, and a subgroup estimate without attention to its population and precision can be misleading. For CheckMate 915, the combination of randomized design, time-to-event methodology, Cox modeling, Kaplan-Meier estimation, and PD-L1-defined analyses provides a practical framework for understanding how modern oncology trials translate follow-up data into statistical evidence.