This page separates reported trial results from statistical interpretation. The numerical results presented here are restricted to the trial data posted on ClinicalTrials.gov for CheckMate-649. Where the ClinicalTrials.gov record does not provide a statistic, subgroup result, or additional design detail, that information is not added from outside sources.
1. Trial at a Glance
CheckMate-649 was a randomized, parallel-group, open-label phase 3 trial with 2031.0 participants evaluating nivolumab plus chemotherapy or nivolumab plus ipilimumab against chemotherapy in stomach cancer or stomach/esophagus junction cancer. The two registered primary endpoints analyzed in the ClinicalTrials.gov record was overall survival and progression-free survival in participants with PD-L1 CPS ≥ 5.
| Feature | CheckMate-649 |
|---|---|
| Phase | Phase 3 |
| Conditions | Gastric Cancer; Gastroesophageal Junction Cancer; Esophageal Adenocarcinoma |
| Design | Randomized, parallel |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 2031.0 |
| Arms | 5 |
| Primary endpoints | Overall survival and progression-free survival |
| Primary endpoint type | Time-to-event |
| Hypothesis type | Superiority |
| Primary analysis method | Log-rank test |
| Primary effect measure | Hazard ratio |
| ClinicalTrials.gov | NCT02872116 |
| Lead sponsor | Bristol-Myers Squibb |
| Sponsor type | INDUSTRY |
2. Clinical Question
The primary efficacy question in the analyses posted on ClinicalTrials.gov was whether nivolumab plus chemotherapy improved overall survival and progression-free survival compared with chemotherapy among randomized participants with PD-L1 combined positive score (CPS) ≥ 5.
Population
Participants enrolled in the phase 3 CheckMate-649 trial with gastric cancer, gastroesophageal junction cancer, or esophageal adenocarcinoma. The primary analyses reported here use the randomized population with PD-L1 CPS ≥ 5.
Intervention
Nivolumab plus chemotherapy. The registry-reported primary analysis identifies the chemotherapy regimens as XELOX or FOLFOX.
Comparator
Chemotherapy, identified in the primary analysis as XELOX or FOLFOX.
Primary question
Does nivolumab plus chemotherapy produce a superior time-to-event outcome compared with chemotherapy for overall survival and progression-free survival?
3. Trial Design
Nivolumab + chemotherapy
- Nivolumab
- Chemotherapy
- XELOX or FOLFOX in the registry-reported primary analysis
Chemotherapy
- Chemotherapy alone
- XELOX or FOLFOX in the registry-reported primary analysis
4. Endpoints
The trial has two registered primary endpoints, both time-to-event outcomes. Their analysis populations and definitions are important because the reported hazard ratios describe treatment comparisons within these particular endpoint frameworks.
| Endpoint | Registered time frame | Definition in the ClinicalTrials.gov record |
|---|---|---|
| Overall Survival (OS) in Participants Treated With Nivolumab Plus Chemotherapy vs Chemotherapy With PD-L1 CPS ≥ 5 | From the date of randomization up to the date of death, up to approximately 17 months | Overall survival is defined as the time from randomization to the time of death. The comparison is nivolumab plus chemotherapy versus chemotherapy in participants with PD-L1 CPS ≥ 5. |
| Progression Free Survival (PFS) in Participants Treated With Nivolumab Plus Chemotherapy vs Chemotherapy With PD-L1 CPS ≥ 5 | From randomization to the date of the first documented progressive disease (PD) per BICR or death due to any cause (up to approximately 10 months) | Progression-free survival is defined as the time from randomization to the date of the first documented PD or death due to any cause. PD is determined by blinded independent committee review (BICR) per RECIST1.1 criteria. |
PD-L1 CPS in the OS definition
The registry definition specifies that CPS is calculated from the number of PD-L1 staining cells—tumor cells, lymphocytes, and macrophages—divided by the total number of viable tumor cells, multiplied by 100.
For the primary analyses reported here, the analysis population is restricted to participants with PD-L1 CPS ≥ 5.
5. Statistical Methodology
Time-to-event analysis
Both primary endpoints are time-to-event outcomes. Rather than simply asking whether an event occurred, the analysis incorporates when the event occurred after randomization. For OS, the event is death. For PFS, the event is the first documented progressive disease or death due to any cause.
Log-rank test
The registry method for both primary analyses is the log-rank test. The log-rank test compares the survival experience of randomized groups over follow-up by considering the observed and expected numbers of events at successive event times.
The resulting test addresses whether the observed time-to-event patterns differ between the treatment groups. It is a hypothesis test, not a direct measure of how large the treatment effect is.
Hazard ratio
The primary effect measure reported in both analyses is the hazard ratio (HR). An HR below 1 indicates a lower estimated instantaneous event rate in the nivolumab-plus-chemotherapy group relative to chemotherapy, under the time-to-event model represented by the estimate.
The hazard ratio is a relative time-to-event measure. It is not an absolute survival probability, a median survival time, or the proportion of patients who benefit.
Confidence intervals
The OS analysis reports a 98.4% two-sided confidence interval, while the PFS analysis reports a 98% two-sided confidence interval. These intervals quantify uncertainty around the respective HR estimates within the statistical framework of the reported analyses.
Superiority testing
Both registry-reported primary analyses are classified as superiority hypotheses. The statistical question is therefore whether the treatment comparison provides evidence of a difference favoring the nivolumab-plus-chemotherapy group rather than whether it merely meets a non-inferiority requirement.
6. Results: Overall Survival
The first primary endpoint is overall survival in all randomized participants with PD-L1 CPS ≥ 5. The registry-reported analysis compares Arm 1, nivolumab + chemotherapy, with Arm 2a, chemotherapy.
Hazard ratio for death
98.4% two-sided CI: 0.59–0.86 · P < 0.0001
Analysis population: all randomized participants with PD-L1 CPS ≥ 5
| OS analysis element | Reported result |
|---|---|
| Comparison | Nivolumab + chemotherapy vs chemotherapy |
| Endpoint | Overall survival |
| Analysis population | All randomized participants with PD-L1 CPS ≥ 5 |
| Method | Log-rank test |
| Effect measure | Hazard ratio |
| Estimate | 0.71 |
| Confidence interval | 98.4% two-sided CI 0.59–0.86 |
| P-value | <0.0001 |
| Hypothesis | Superiority |
| Time frame | From the date of randomization up to the date of death, up to approximately 17 months |
An OS hazard ratio of 0.71 means that the estimated instantaneous rate of death in the nivolumab-plus-chemotherapy group was approximately 71% of the corresponding rate in the chemotherapy group, under the analysis represented by the reported HR. Equivalently, 1 − 0.71 = 0.29, so the estimate corresponds to an approximately 29% lower estimated hazard of death.
The HR does not mean that 29% of patients were saved, that each patient experienced exactly a 29% reduction in mortality risk, or that the absolute probability of death was 29% lower. It is a relative time-to-event measure.
The 98.4% confidence interval of 0.59–0.86 describes the statistical uncertainty surrounding the HR estimate. It is narrower than a very imprecise interval but still represents uncertainty about the underlying treatment effect. It does not describe the range of outcomes that individual patients might experience.
The P-value < 0.0001 addresses the compatibility of the observed data with the null hypothesis used for the superiority test. It does not measure the magnitude of the treatment effect. The magnitude is better conveyed by the HR and its confidence interval.
Because this is a time-to-event analysis, interpretation also depends on the event and censoring rules used in the endpoint definition. A single HR is most straightforward when the relative hazards are reasonably represented by the model over follow-up; it should not be translated into an absolute risk reduction without appropriate time-specific survival estimates.
7. Results: Progression-Free Survival
The second primary endpoint is progression-free survival in all randomized participants with PD-L1 CPS ≥ 5. The registry-reported definition counts the first documented progressive disease per blinded independent committee review or death due to any cause as the event.
Hazard ratio for progression or death
98% two-sided CI: 0.56–0.81 · P < 0.0001
Analysis population: all randomized participants with PD-L1 CPS ≥ 5
| PFS analysis element | Reported result |
|---|---|
| Comparison | Nivolumab + chemotherapy vs chemotherapy |
| Endpoint | Progression-free survival |
| Analysis population | All randomized participants with PD-L1 CPS ≥ 5 |
| Method | Log-rank test |
| Effect measure | Hazard ratio |
| Estimate | 0.68 |
| Confidence interval | 98% two-sided CI 0.56–0.81 |
| P-value | <0.0001 |
| Hypothesis | Superiority |
| Time frame | From randomization to the date of the first documented progressive disease (PD) per BICR or death due to any cause (up to approximately 10 months) |
A PFS hazard ratio of 0.68 means that the estimated instantaneous rate of experiencing the PFS event—first documented progression or death—was approximately 68% of that in the chemotherapy group, under the reported time-to-event analysis. The corresponding relative interpretation is an approximately 32% lower estimated hazard because 1 − 0.68 = 0.32.
This does not mean that 32% of patients avoided progression, nor does it mean that progression was delayed by exactly 32% for every participant. PFS is defined using an event-time framework, and its interpretation depends on both progression and death as specified by the endpoint.
The 98% confidence interval of 0.56–0.81 expresses uncertainty around the estimated HR. It gives a range of plausible values for the relative treatment effect under the specified inferential framework; it does not give a range of individual patient outcomes.
The P-value < 0.0001 provides evidence against the null hypothesis for the reported superiority test, but it is not an effect-size measure. A very small P-value can occur with different effect magnitudes depending on the amount of information in the trial.
The PFS endpoint also incorporates blinded independent committee review for progression according to RECIST1.1 in the registry-reported definition. This is relevant because progression is not simply a subjective clinical impression; it is defined through the trial's specified assessment framework.
8. Comparing the Two Primary Hazard Ratios
The two primary estimates address related but distinct clinical events. OS uses death as the event, whereas PFS uses first documented progression or death. The estimates should therefore not be treated as two measurements of exactly the same outcome.
| Primary endpoint | HR | Confidence interval | P-value | Event represented |
|---|---|---|---|---|
| Overall survival | 0.71 | 98.4% two-sided CI 0.59–0.86 | <0.0001 | Death |
| Progression-free survival | 0.68 | 98% two-sided CI 0.56–0.81 | <0.0001 | First documented PD or death |
What is consistent
Both reported HRs are below 1, and both primary analyses report P < 0.0001 under superiority hypotheses.
What differs
The endpoints have different event definitions and different confidence levels: 98.4% for OS and 98% for PFS.
Why both matter
PFS captures progression or death, while OS captures death. They answer different questions about the course of disease after randomization.
What cannot be inferred
The ClinicalTrials.gov record does not provide median OS, median PFS, Kaplan-Meier survival probabilities, event counts, or subgroup estimates, so those quantities should not be reconstructed from the HRs.
9. Safety Results
The ClinicalTrials.gov record reports serious adverse events as affected participants divided by the number at risk for three named treatment groups. These counts should be interpreted separately from the efficacy analyses because safety summaries concern treatment exposure rather than the randomized time-to-event comparison itself.
| Safety group | Serious adverse events affected / at risk |
|---|---|
| Nivo + Chemo | 529 / 782 |
| Chemo | 474 / 806 |
| Nivo + Ipi | 307 / 403 |
The ClinicalTrials.gov record does not identify these serious adverse-event figures as a formal hypothesis test, provide confidence intervals, or specify a comparative statistical model. Accordingly, they are presented as reported affected/at-risk counts rather than converted into an independently calculated inferential comparison.
10. Statistical Methods Explained
Why was a log-rank test used?
Both primary endpoints are time-to-event outcomes, so a method that uses event timing is appropriate. The log-rank test compares the survival experience of the randomized groups across follow-up rather than reducing the endpoint to a simple yes/no outcome. In this trial, it was the method reported for both primary analyses.
What does an OS HR of 0.71 mean?
An HR of 0.71 indicates a lower estimated instantaneous rate of death in the nivolumab-plus-chemotherapy group relative to chemotherapy under the reported analysis. It corresponds to an approximately 29% lower estimated hazard, but it does not mean a 29% lower absolute probability of death or a 29% increase in the number of survivors.
Why is the PFS HR different from the OS HR?
PFS and OS have different event definitions. PFS counts first documented progression or death, whereas OS counts death. A treatment can therefore have different relative effects on the two endpoints without the difference implying a statistical inconsistency.
What does the confidence interval add beyond the hazard ratio?
The HR gives a point estimate, while the confidence interval communicates uncertainty around that estimate. For OS, the reported 98.4% two-sided interval is 0.59–0.86. For PFS, the reported 98% two-sided interval is 0.56–0.81. The intervals help distinguish a precise estimate from one with substantial statistical uncertainty.
Why doesn't the P-value measure treatment effect size?
The P-value addresses evidence against the null hypothesis under the specified testing framework. It depends not only on the size of an observed difference but also on the amount of information available. The HR and its confidence interval are therefore needed to describe the magnitude and uncertainty of the treatment effect.
Why does the analysis population matter?
The reported primary analyses are explicitly based on all randomized participants with PD-L1 CPS ≥ 5. An HR calculated in a different population would answer a different statistical question. Restricting the interpretation to the registry-reported analysis population avoids silently generalizing the estimate beyond the group for which it was reported.
Why should PFS not be treated as the same endpoint as OS?
PFS incorporates disease progression as well as death, while OS is based on death alone. The two endpoints can provide complementary information. Neither HR should be interpreted as though it were a substitute for the other endpoint.
11. Understanding Censoring in Time-to-Event Analysis
Time-to-event analysis is designed for settings in which not every participant experiences the event during the observation period. A participant can contribute information until the time at which the available follow-up ends without observing the specified event. This is commonly described as right censoring.
The statistical analysis uses the information available before censoring rather than treating a censored participant as if the event had occurred at the censoring time.
For CheckMate-649, the ClinicalTrials.gov record identifies OS and PFS as time-to-event endpoints and provide their event definitions. They do not provide the individual event and censoring records, so a patient-level Kaplan-Meier curve cannot be reconstructed from the ClinicalTrials.gov record alone.
12. Why the Analysis Uses a Hazard Ratio Rather Than a Simple Risk Ratio
A simple risk ratio compares proportions experiencing an event by a particular time. A hazard ratio instead compares the estimated instantaneous event rates over time within a time-to-event framework.
This distinction is important for CheckMate-649 because both primary endpoints explicitly measure time from randomization to a defined event. The HR therefore uses more of the temporal structure of the outcome than a single binary event indicator at an arbitrary cutoff.
Hazard ratio
Relative comparison of instantaneous event rates over follow-up, as represented by the reported survival analysis.
Absolute risk
Difference in the probability of experiencing an event by a specified time point. No such time-specific absolute risk estimates are reported here.
Median survival
The time at which the estimated survival probability reaches 50%. No median OS or PFS is reported in the ClinicalTrials.gov record.
P-value
A measure of evidence against the specified null hypothesis, not a measure of how large or clinically important an effect is.
13. Multiplicity and the Two Primary Endpoints
CheckMate-649 has two registered primary endpoints in the ClinicalTrials.gov record: OS and PFS. Both have formal statistical analyses posted, both use the log-rank test, and both use superiority hypotheses.
| Feature | OS | PFS |
|---|---|---|
| Primary endpoint | Yes | Yes |
| Endpoint type | Time-to-event | Time-to-event |
| Analysis population | All randomized participants with PD-L1 CPS ≥ 5 | All randomized participants with PD-L1 CPS ≥ 5 |
| Method | Log-rank test | Log-rank test |
| Effect measure | Hazard ratio | Hazard ratio |
| Hypothesis | Superiority | Superiority |
| Reported P-value | <0.0001 | <0.0001 |
The ClinicalTrials.gov record does not specify an alpha-allocation strategy, hierarchical testing sequence, alpha-spending procedure, interim-analysis schedule, or multiplicity adjustment connecting the two endpoints. Those details are therefore not inferred here. In particular, two small P-values should not automatically be interpreted as though the reader knows the exact familywise-error procedure unless that procedure is explicitly reported.
14. Design Topics Not Specified in the Supplied Data
Several statistical design features can materially affect the interpretation of a modern phase 3 trial. For this page, they are intentionally discussed only to the extent supported by the ClinicalTrials.gov record.
| Design topic | What the ClinicalTrials.gov record establishes | Interpretive consequence |
|---|---|---|
| Randomization | Allocation is RANDOMIZED | The primary efficacy comparison is based on randomized treatment assignment. |
| Parallel design | Design model is PARALLEL | The treatment groups are analyzed as parallel randomized groups. |
| Masking | Masking is NONE | The trial is classified as unmasked in the registry profile. |
| Superiority | Both primary analyses use a superiority hypothesis | The reported tests address evidence of a difference rather than a non-inferiority margin. |
| Stratification | Not reported | No stratification factors are inferred. |
| Crossover | Not reported | No crossover effect is inferred. |
| Interim analysis | Not reported | No interim efficacy boundary or alpha-spending method is inferred. |
| Missing-data or imputation strategy | Not reported | No imputation method is inferred. |
| Bayesian methods | Not reported | No Bayesian analysis is inferred. |
15. Statistical Interpretation of the Confidence Intervals
The confidence intervals deserve attention because they provide information that a P-value alone cannot provide.
The reported OS HR is 0.71, with a 98.4% two-sided CI of 0.59–0.86. The interval remains below 1, consistent with the direction of the point estimate and the reported superiority result.
The reported PFS HR is 0.68, with a 98% two-sided CI of 0.56–0.81. The interval likewise remains below 1, consistent with the reported superiority result.
Neither confidence interval establishes the magnitude of benefit for every participant. Instead, each interval describes statistical uncertainty surrounding a population-level relative effect estimate under its corresponding analysis.
16. What the Hazard Ratios Do — and Do Not — Mean
The estimate corresponds to approximately a 29% lower estimated hazard of death for nivolumab plus chemotherapy relative to chemotherapy because 1 − 0.71 = 0.29.
It does not mean that 29% of patients survived, that survival time increased by 29%, or that every patient had exactly the same proportional reduction in mortality hazard.
The estimate corresponds to approximately a 32% lower estimated hazard of progression or death for nivolumab plus chemotherapy relative to chemotherapy because 1 − 0.68 = 0.32.
It does not mean that 32% of patients avoided progression, nor does it specify how much longer an individual patient remained progression-free.
17. Limitations
- Limited result set: the statistical analyses posted on ClinicalTrials.gov contain two primary endpoint comparisons, but do not provide median OS, median PFS, event counts, Kaplan-Meier estimates, subgroup results, or other secondary efficacy estimates.
- Incomplete PFS registry time frame: the registry-reported PFS time-frame field ends with “up t”, so the page preserves the registry wording rather than supplying an inferred completion of the phrase.
- Analysis population: the reported primary HRs apply to all randomized participants with PD-L1 CPS ≥ 5, not automatically to every enrolled participant regardless of CPS.
- Hazard-ratio interpretation: an HR is a relative time-to-event measure and should not be presented as an absolute risk reduction, survival probability, or percentage of patients benefiting.
- Model assumptions: time-to-event effect estimates require appropriate interpretation of the underlying survival-analysis framework. The ClinicalTrials.gov record does not provide a detailed assessment of proportional-hazards behavior.
- Multiplicity: the ClinicalTrials.gov record establishes two primary endpoints but do not provide the multiplicity-control strategy, so no particular alpha-allocation procedure is inferred.
- Unreported design details: stratification factors, crossover, interim-analysis procedures, missing-data methods, and Bayesian methods are not specified in the ClinicalTrials.gov record.
- Safety interpretation: serious adverse-event counts are posted on ClinicalTrials.gov for three groups, but the registry excerpt does not provide a formal comparative safety analysis, confidence intervals, or event timing.
- Generalizability: the primary estimates describe the registry-reported randomized analysis population with PD-L1 CPS ≥ 5 and should not automatically be generalized to populations that differ from that analysis definition.
18. Why This Trial Matters Statistically
CheckMate-649 is a useful statistical teaching case because it illustrates how a large randomized phase 3 trial can combine a disease-specific biomarker-defined analysis population with two clinically distinct time-to-event endpoints.
| Concept | How it appears in CheckMate-649 |
|---|---|
| Randomization | The trial allocation is randomized. |
| Parallel design | The registry identifies a parallel design model. |
| Open-label design | Masking is listed as none. |
| Time-to-event endpoints | OS and PFS are both registered as primary time-to-event outcomes. |
| Kaplan-Meier framework | Time-to-event endpoints are naturally interpreted through survival-distribution methods, although no Kaplan-Meier estimates are reported here. |
| Log-rank testing | The reported statistical method for both primary endpoint analyses. |
| Hazard ratio | The reported effect measure for both primary endpoints. |
| Confidence intervals | 98.4% for OS and 98% for PFS in the analyses posted on ClinicalTrials.gov. |
| Superiority testing | Both registry-reported primary analyses are classified as superiority hypotheses. |
| Biomarker-defined analysis | The primary analyses are conducted in randomized participants with PD-L1 CPS ≥ 5. |
| Independent progression assessment | The registry-reported PFS definition identifies BICR per RECIST1.1 for determining PD. |
| Safety versus efficacy populations | The efficacy analyses are defined by randomization and CPS, while registry-reported serious-AE figures are reported as affected/at-risk counts by treatment group. |
19. Related Statistical Concepts
Learn more about the methods used in this trial:
20. Related Statistical Calculators
21. Sources
- ClinicalTrials.gov: NCT02872116 — CheckMate-649.
- Linked publication: PubMed record — PMID 40632415.
- Linked publication: PubMed record — PMID 38382001.
- Linked publication: PubMed record — PMID 37713657.
- Linked publication: PubMed record — PMID 36919706.
- Linked publication: PubMed record — PMID 36121651.
Continue through the Clinical Biostats statistical pathway
Explore the statistical concepts behind randomized time-to-event trials, from confidence intervals and hazard ratios to Kaplan-Meier estimation and log-rank testing.
22. Record Summary
CheckMate-649 provides a strong example of how statistical interpretation should separate the event definition, analysis population, hypothesis test, and effect estimate. In the registry-reported primary analyses, nivolumab plus chemotherapy was compared with chemotherapy among all randomized participants with PD-L1 CPS ≥ 5. Overall survival had an HR of 0.71 with a 98.4% two-sided CI of 0.59–0.86 and P < 0.0001. Progression-free survival had an HR of 0.68 with a 98% two-sided CI of 0.56–0.81 and P < 0.0001. Both analyses used the log-rank test and were classified as superiority hypotheses.
The key statistical lesson is that these numbers answer a specific question: how the randomized treatment groups differed in the rates of defined time-to-event outcomes within the registry-reported analysis population. The hazard ratios quantify relative effects, the confidence intervals describe uncertainty, and the P-values address the corresponding hypothesis tests. None of these quantities, by itself, supplies median survival, absolute survival probabilities, individual patient benefit, or the complete safety profile.