This page separates reported trial results from statistical interpretation. The numerical results presented here are limited to the statistical analysis reported in the ClinicalTrials.gov record.
1. Trial at a Glance
CheckMate-017 was a randomized, parallel, open-label phase 3 treatment trial comparing nivolumab with docetaxel in participants with previously treated advanced or metastatic squamous cell non-small cell lung cancer. The ClinicalTrials.gov record reports 272 enrolled participants and one formal statistical analysis for the primary overall-survival time endpoint.
| Feature | CheckMate-017 |
|---|---|
| Trial name | CheckMate-017 |
| Phase | Phase 3 |
| Condition | Squamous Cell Non-small Cell Lung Cancer |
| Population | Previously treated advanced or metastatic squamous cell non-small cell lung cancer |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 272 |
| Arms | 2 |
| Interventions | Nivolumab; Docetaxel |
| Lead sponsor | Bristol-Myers Squibb |
| Sponsor type | Industry |
| Status | Completed |
| Start date | October 16, 2012 |
| Primary completion | November 17, 2014 |
| ClinicalTrials.gov | NCT01642004 |
2. Clinical Question
The trial addressed whether nivolumab should be compared with docetaxel for treatment in participants with previously treated advanced or metastatic squamous cell non-small cell lung cancer. The registered hypothesis type for the formal analysis was superiority.
Population
Participants with squamous cell non-small cell lung cancer who had previously been treated and had advanced or metastatic disease.
Intervention
Nivolumab, identified in the registry data as a biological intervention.
Comparator
Docetaxel, identified in the registry data as a drug intervention.
Primary question
Does nivolumab produce superior overall-survival outcomes compared with docetaxel under the prespecified randomized comparison?
3. Trial Design
Nivolumab
- Intervention type: biological
- Compared against docetaxel
Docetaxel
- Intervention type: drug
- Comparator for the randomized analysis
The ClinicalTrials.gov record does not provide treatment dosing schedules, detailed eligibility criteria, baseline characteristics, randomization ratio, or additional design features such as factorial allocation or crossover.
4. Endpoints
The registry lists three primary endpoints. Two are explicitly time-to-event measures and one is listed as binary in the registry-reported normalized endpoint profile. The formal statistical analysis posted in the ClinicalTrials.gov record concerns the overall-survival time endpoint.
| Primary endpoint | Time frame | Registry definition / statistical role |
|---|---|---|
| Overall Survival (OS) Time in Months for All Randomized Participants at Primary Endpoint | Randomization until 199 deaths, up to November 2014, approximately 25 months | OS was defined as the time between the date of randomization and the date of death from any cause. Participants were censored at the date they were last known to be alive. Median OS time was calculated using Kaplan-Meier (KM) method. Hazard ratio and the corresponding confidence interval were estimated in a stratified Cox proportional hazards model. |
| Overall Survival (OS) Rate in All Randomized Participants | Randomization to 18 months post-randomization, up to June 2015 | The overall survival rate is the probability that a participant will be alive at 6, 12, and 18 months following randomization. Survival rates were determined via Kaplan-Meier estimates. |
| Number of Deaths From Any Cause in All Randomized Participants at Primary Endpoint | Randomization until 199 deaths, up to November 2014, approximately 25 months | The number of participants who died from any cause was reported for each arm. Interim analysis was planned to occur after at least 196 deaths, with the actual analysis occurring at 199 deaths. |
5. Statistical Methodology
Kaplan-Meier estimation
The registry definition specifies that overall survival is measured from randomization to death from any cause, with participants censored at the date they were last known to be alive. The Kaplan-Meier method is appropriate for this type of right-censored time-to-event data because it estimates the survival function while retaining information from participants whose event time is not observed during follow-up.
Here, di represents events at time ti, while ni represents participants at risk immediately before that time.
Log-rank test
The posted statistical analysis identifies the method as a Log Rank test. The log-rank test compares the observed and expected numbers of events between randomized groups over the observed follow-up and is designed specifically for time-to-event comparisons.
Stratified Cox proportional-hazards model
The registry analysis notes a stratified Cox proportional hazard model. The hazard ratio was defined as Nivolumab over Docetaxel. Thus, the reported HR is a relative comparison of the estimated event hazards, with values below 1 favoring a lower estimated hazard in the nivolumab group under the fitted model.
An HR below 1 indicates a lower estimated instantaneous event rate in the numerator group, here nivolumab, relative to the comparator under the model. It is not an absolute survival probability and is not the percentage of participants who benefit.
All-randomized analysis population
The posted formal analysis identifies its analysis population as all randomized participants. This is important because the treatment comparison remains anchored to the randomized assignment rather than redefining the efficacy population according to treatment exposure or later clinical events.
What the registry does not establish
The ClinicalTrials.gov record does not identify a randomization ratio, specific stratification variables, missing-data or imputation procedures, a non-inferiority margin, Bayesian methods, a factorial design, crossover rules, or a multiplicity-adjustment strategy. These features are therefore not assumed in this analysis.
6. Results: Overall Survival Time
The ClinicalTrials.gov record contains one formal statistical analysis. It evaluates Overall Survival (OS) Time in Months for All Randomized Participants at Primary Endpoint, from randomization until 199 deaths, up to November 2014, approximately 25 months.
Hazard ratio for overall survival
96.85% two-sided CI: 0.43–0.81 · P = 0.0002
Analysis population: all randomized participants · Nivolumab vs Docetaxel
Method: log-rank test with a stratified Cox proportional-hazards model for the hazard ratio
| Analysis element | Reported result |
|---|---|
| Endpoint | Overall Survival (OS) Time in Months for All Randomized Participants at Primary Endpoint |
| Analysis population | All randomized participants |
| Groups compared | Nivolumab vs Docetaxel |
| Time frame | Randomization until 199 deaths, up to November 2014, approximately 25 months |
| Method | Log Rank |
| Model for effect estimate | Stratified Cox proportional hazard model |
| Effect measure | Hazard Ratio |
| Hazard ratio | 0.59 |
| Confidence interval | 96.85% two-sided CI: 0.43–0.81 |
| P-value | 0.0002 |
| Hypothesis type | Superiority |
| Hazard-ratio direction | Nivolumab over Docetaxel |
The reported HR of 0.59 means that, under the stratified Cox model used for this analysis, the estimated instantaneous rate of death in the nivolumab group was approximately 59% of the corresponding estimated rate in the docetaxel group. Expressed as a simple relative interpretation, 1 − 0.59 = 0.41, so the estimate corresponds to an approximately 41% lower estimated hazard for nivolumab relative to docetaxel.
This does not mean that 41% of participants avoided death, that each individual participant had exactly a 41% reduction in risk, or that the absolute probability of survival differed by 41 percentage points. The hazard ratio is a model-based relative measure of event rates over time.
The reported 96.85% two-sided confidence interval of 0.43–0.81 quantifies uncertainty around the estimated hazard ratio under the specified analysis framework. It describes uncertainty in the estimated treatment effect; it does not describe the range of effects that must occur in individual patients.
The P-value of 0.0002 measures evidence against the null hypothesis within the statistical testing framework. It does not measure the size, clinical importance, or probability of the treatment effect. The effect size is described by the hazard ratio, while its statistical precision is described by the confidence interval.
Because the analysis uses a Cox proportional-hazards model, interpretation of a single hazard ratio is most straightforward when the proportional-hazards assumption is a reasonable description of the hazard relationship over follow-up. The ClinicalTrials.gov record does not provide a diagnostic assessment of that assumption.
Finally, the analysis is based on all randomized participants and involves right-censored survival observations. The hazard ratio therefore summarizes the randomized time-to-event comparison rather than simply comparing crude percentages of deaths.
What the 0.59 estimate tells us statistically
The direction of the effect is clear from the reported hazard-ratio scale: because the numerator is nivolumab and the HR is below 1, the fitted model estimates a lower death hazard for nivolumab than for docetaxel. The confidence interval also lies below 1, from 0.43 to 0.81.
However, the hazard ratio should not be read as a substitute for the Kaplan-Meier survival curve. Survival analysis contains information about when events occur and how censoring enters the risk set. Two studies can have similar hazard ratios while having different absolute survival probabilities at specific time points.
7. Other Primary Endpoint Results
The registry data state that results were posted for all three registered primary endpoints, but the ClinicalTrials.gov record contains a formal statistical analysis only for the primary endpoint measuring overall-survival time.
| Primary endpoint | Results posted | Formal statistical analysis reported |
|---|---|---|
| Overall Survival (OS) Time in Months for All Randomized Participants at Primary Endpoint | Yes | Yes — log-rank test and stratified Cox model; HR 0.59, 96.85% CI 0.43–0.81, P = 0.0002 |
| Overall Survival (OS) Rate in All Randomized Participants | Yes | No formal statistical comparison reported in the posted statistical analyses |
| Number of Deaths From Any Cause in All Randomized Participants at Primary Endpoint | Yes | No formal statistical comparison reported in the posted statistical analyses |
How the OS-rate endpoint would normally be analyzed
The registered OS-rate endpoint defines survival as the probability that a participant is alive at 6, 12, and 18 months following randomization, with survival rates determined by Kaplan-Meier estimates. For a time-to-event trial, those estimates would ordinarily be accompanied by uncertainty measures and, where a between-group comparison is planned, an appropriate survival-comparison framework.
The ClinicalTrials.gov record does not provide a formal comparative estimate, confidence interval, or P-value for this endpoint.
How the number-of-deaths endpoint differs
The third primary endpoint reports the number of participants who died from any cause in each arm. The ClinicalTrials.gov record states that the interim analysis was planned after at least 196 deaths and that the actual analysis occurred at 199 deaths. A simple event count does not itself account for differences in follow-up or censoring, which is why the time-to-event analysis is the more informative framework for the primary OS comparison reported here.
8. Planned Analysis and Interim Monitoring
The registered primary endpoint definition states that an interim analysis was planned to occur after at least 196 deaths, with the actual analysis occurring at 199 deaths. The primary endpoint time frame was randomization until 199 deaths, up to November 2014, approximately 25 months.
Event-driven analysis
The analysis was tied to the accumulation of deaths rather than simply to a fixed calendar date. The actual primary analysis occurred when 199 deaths had accumulated.
Why event timing matters
In survival trials, the amount of information available for estimating treatment differences depends strongly on the number of observed events and the censoring pattern.
The ClinicalTrials.gov record does not specify an alpha-spending function, stopping boundary, or other formal group-sequential error-control method. The fact that an interim analysis occurred should therefore not be used to infer a particular multiplicity-adjustment procedure.
9. Statistical Methods Explained
Why use a log-rank test?
The primary OS endpoint is a time-to-event outcome with right censoring. The log-rank test is designed to compare survival distributions between groups while incorporating the timing of observed deaths and the changing number of participants at risk. It therefore uses more information than a simple comparison of the total number of deaths.
What does an HR of 0.59 mean?
Because the registry defines the hazard ratio as nivolumab over docetaxel, an HR of 0.59 indicates a lower estimated hazard of death in the nivolumab group under the fitted Cox model. The complementary quantity, 1 − 0.59 = 0.41, corresponds to an approximately 41% lower estimated hazard. That is a relative hazard interpretation, not an absolute survival difference.
Why is the confidence interval 96.85% rather than 95%?
The registry-reported statistical analysis explicitly reports a 96.85% two-sided confidence interval. Confidence intervals are linked to the statistical testing framework used for an analysis, and an interval need not always be 95%. The page therefore preserves the registry's reported 96.85% level rather than substituting a conventional 95% interval.
What does the P-value of 0.0002 mean?
It quantifies the evidence against the null hypothesis within the specified superiority-testing framework. A smaller P-value represents stronger statistical evidence against that null hypothesis under the model and assumptions. It does not say that the probability the treatment works is 99.98%, and it does not quantify clinical magnitude.
Why use both Kaplan-Meier and Cox methods?
They answer complementary questions. Kaplan-Meier estimation describes the survival distribution and can produce survival probabilities at specified time points. The Cox model provides a relative hazard comparison between treatment groups. Reporting both allows the time-to-event experience to be described through survival probabilities as well as a model-based relative effect.
Why does censoring matter?
Some participants may be alive at their last known follow-up without an observed death. Those participants are censored rather than automatically treated as if they survived indefinitely. Kaplan-Meier and Cox methods are designed to incorporate such right-censored observations. The quality of the inference depends on whether the assumptions governing censoring are reasonable.
Why does randomization matter for the interpretation?
Randomization establishes the treatment groups before outcomes occur and supports comparison of outcomes according to assigned treatment. The posted OS analysis explicitly uses all randomized participants. That preserves the central randomized comparison rather than selectively analyzing only participants who remained on treatment.
10. Safety Results
The ClinicalTrials.gov record provides serious adverse-event counts by treatment group. These are reported as affected participants divided by the corresponding participants at risk. Because the data use these arm-specific affected/at-risk counts, they are reproduced exactly rather than converted into newly calculated percentages.
| Safety group | Serious adverse events, affected / at risk |
|---|---|
| Nivolumab 3 mg/kg | 85 / 131 |
| Nivolumab 480 mg | 1 / 6 |
| Docetaxel | 92 / 129 |
| Extension Phase of Docetaxel Arm: Nivolu | 4 / 6 |
| Extension Phase of Docetaxel Arm: Nivolu | 0 / 1 |
11. Understanding the Primary Statistical Result
The hazard ratio of 0.59 places the estimated death hazard for nivolumab below that of docetaxel in the fitted stratified Cox model. Since the treatment order is explicitly Nivolumab over Docetaxel, the direction of the estimate is unambiguous.
The hazard ratio does not provide an absolute probability of survival at a particular time. The registered OS-rate endpoint separately defines survival probabilities at 6, 12, and 18 months, but the registry-reported statistical analysis does not provide numerical estimates for those time points. Those estimates should therefore not be reconstructed from the HR.
The reported 96.85% confidence interval extends from 0.43 to 0.81. This interval indicates that the estimated treatment effect is uncertain, but the entire reported interval is below 1. The width of the interval is also important: a point estimate alone would provide less information about the statistical precision of the comparison.
The reported P = 0.0002 provides the formal evidence measure associated with the superiority comparison. It should be interpreted alongside the hazard ratio and confidence interval rather than treated as a standalone measure of treatment effect.
Because OS incorporates both the occurrence and timing of deaths, the analysis is fundamentally different from simply asking which arm had more deaths. Participants who were still alive at their last known follow-up contribute censored survival information rather than being counted as deaths.
12. What the Hazard Ratio Does — and Does Not — Mean
A hazard ratio is one of the most frequently misunderstood quantities in clinical-trial reporting. CheckMate-017 provides a useful example because the registry gives both the exact treatment direction and the model used to estimate the effect.
Under the fitted stratified Cox model, the estimated instantaneous event rate is lower for the nivolumab group than for the docetaxel group.
The estimate can be expressed as an approximate 41% lower estimated hazard because 1 − 0.59 = 0.41. This is a simple derived interpretation of the reported HR and does not alter the reported number.
It does not mean that 41% of participants benefited, that survival probability increased by 41 percentage points, or that every participant experienced a 41% reduction in the probability of death. Those interpretations confuse a hazard ratio with absolute risk measures.
The confidence interval is equally important. A point estimate of 0.59 without an interval would conceal how precisely the treatment effect was estimated. Here, the reported 96.85% two-sided interval is 0.43–0.81.
The P-value provides a different piece of information. It addresses statistical evidence against the null hypothesis under the testing framework; it does not measure effect magnitude. A very small P-value can accompany a small effect in a large study, while a clinically substantial estimate can have substantial uncertainty in a small study.
13. Stratification and the Cox Model
The statistical analysis notes that the Cox proportional-hazards model was stratified. Stratification in a survival model allows the baseline hazard to vary across specified strata while estimating a common treatment effect across those strata.
The ClinicalTrials.gov record identifies stratified analysis as an additional concept in the analysis text, but they do not identify the actual stratification variables. Consequently, no specific clinical characteristic is attributed to the strata here.
What stratification can accomplish
It can account for prespecified differences in baseline hazard across strata without requiring the same baseline hazard function for every stratum.
What it does not mean
Stratification does not create a new randomized comparison within each subgroup, nor does it automatically establish treatment-effect differences between strata.
The important distinction is between adjusting the primary treatment comparison and testing whether treatment effects differ. A stratified analysis can produce an overall treatment-effect estimate without implying that the treatment effect has been formally compared across the strata.
14. Primary Endpoint Analysis in Context
| Component | Role in CheckMate-017 |
|---|---|
| Randomization | Creates the randomized comparison between nivolumab and docetaxel. |
| Analysis population | All randomized participants. |
| Endpoint | Overall survival time from randomization to death from any cause. |
| Censoring | Participants censored at the date they were last known to be alive. |
| Descriptive estimator | Kaplan-Meier method. |
| Group-comparison method | Log-rank test. |
| Effect model | Stratified Cox proportional-hazards model. |
| Effect measure | Hazard ratio. |
| Estimate | 0.59. |
| Confidence interval | 96.85% two-sided CI 0.43–0.81. |
| P-value | 0.0002. |
| Hypothesis | Superiority. |
This sequence illustrates a standard survival-analysis workflow: randomization defines the comparison, Kaplan-Meier estimation describes the survival experience, the log-rank test compares the groups, and the Cox model expresses the relative treatment effect as a hazard ratio with a confidence interval.
15. Limitations and Interpretation Issues
- Limited formal results in the ClinicalTrials.gov record: although results are posted for three primary endpoints, only one formal statistical analysis with an estimate, confidence interval, and P-value is reported here.
- Hazard-ratio interpretation: the HR is a model-based relative measure and should not be interpreted as an absolute risk reduction or survival probability difference.
- Proportional-hazards assumption: the Cox model relies on a proportional-hazards framework for the usual interpretation of a single HR. The ClinicalTrials.gov record does not report a diagnostic assessment of this assumption.
- Censoring: OS includes right-censored observations. The validity of standard survival methods depends on assumptions concerning the censoring process.
- Confidence-interval level: the reported interval is 96.85% rather than the more commonly encountered 95%, and the page preserves that registry value exactly.
- Interim analysis: the registry identifies a planned interim analysis after at least 196 deaths and an actual analysis at 199 deaths, but the ClinicalTrials.gov record does not specify the complete sequential testing or alpha-spending procedure.
- Unknown stratification variables: the analysis is described as stratified, but the ClinicalTrials.gov record does not identify the variables used for stratification.
- Endpoint multiplicity: three primary endpoints are registered, but the registry-reported statistical data do not describe a complete multiplicity-control strategy for the endpoint family.
- Safety interpretation: serious-adverse-event counts are reported by arm, but the data do not provide enough detail to reconstruct a complete exposure-adjusted safety analysis.
- Generalizability: the trial population is specifically described as participants with previously treated advanced or metastatic squamous cell non-small cell lung cancer, so statistical results should be interpreted within that registered population.
16. Why This Trial Matters Statistically
CheckMate-017 is a useful teaching example because the ClinicalTrials.gov record connects a randomized superiority question to a classic time-to-event analysis. The trial illustrates how several statistical concepts fit together rather than operating as isolated techniques.
| Concept | How it appears in CheckMate-017 |
|---|---|
| Randomization | The allocation is explicitly randomized in a parallel two-arm phase 3 trial. |
| Time-to-event endpoint | Overall survival is defined from randomization until death from any cause. |
| Right censoring | Participants are censored at the date they were last known to be alive. |
| Kaplan-Meier estimation | Used to calculate median OS and registered OS rates. |
| Log-rank test | Used for the posted primary OS time comparison. |
| Hazard ratio | Provides the relative treatment effect, with nivolumab as numerator. |
| Cox model | A stratified Cox proportional-hazards model was used to estimate the HR and confidence interval. |
| Confidence interval | The reported 96.85% two-sided CI is 0.43–0.81. |
| P-value | The formal reported P-value is 0.0002. |
| Event-driven analysis | The primary analysis occurred after 199 deaths, with at least 196 deaths planned for the interim analysis. |
| Multiple primary endpoints | Three primary endpoints are registered, although only one formal statistical analysis is reported in the ClinicalTrials.gov recordset. |
| Safety analysis | Serious adverse events are reported as affected participants divided by participants at risk for the listed groups. |
17. A Practical Reading Strategy for This Trial
A statistically careful reading of the CheckMate-017 result can be organized into five questions.
1. What was randomized?
The trial randomized participants to nivolumab or docetaxel in a parallel phase 3 design.
2. What was measured?
Overall survival was measured from randomization to death from any cause, with censoring at the last date known alive.
3. How was it analyzed?
Kaplan-Meier estimation, log-rank testing, and a stratified Cox proportional-hazards model were used in the registry-reported analysis.
4. What was the effect?
The reported hazard ratio was 0.59, with a 96.85% two-sided confidence interval of 0.43–0.81.
5. How strong was the statistical evidence?
The reported P-value was 0.0002 under a superiority hypothesis.
6. What remains unknown?
The registry-reported statistical data do not provide numerical results for the other registered primary endpoints or identify several additional design and analysis details.
This approach helps prevent a common statistical error: treating one reported hazard ratio as if it were a complete description of the trial. A hazard ratio summarizes one dimension of the time-to-event comparison. The endpoint definition, analysis population, censoring rules, confidence interval, hypothesis, and statistical method all contribute to its interpretation.
18. Related Tutorials
Learn more about the methods used in this trial:
19. Related Calculators
20. Sources
- ClinicalTrials.gov: NCT01642004 — CheckMate-017.
- PubMed record: PMID 41121745.
- PubMed record: PMID 36897427.
- PubMed record: PMID 33449799.
- PubMed record: PMID 32198149.
- PubMed record: PMID 30215677.
Continue with the statistical methods behind the trial
Explore survival-analysis tutorials and statistical calculators covering the concepts used to interpret randomized time-to-event evidence.
21. Record Summary
CheckMate-017 provides a clear example of randomized time-to-event analysis. The ClinicalTrials.gov record describes a phase 3, randomized, parallel, unmasked treatment trial enrolling 272 participants and comparing nivolumab with docetaxel. Its primary statistical analysis evaluated overall survival time from randomization until 199 deaths, using a log-rank test and a stratified Cox proportional-hazards model.
The reported hazard ratio was 0.59, with a 96.85% two-sided confidence interval of 0.43–0.81 and P = 0.0002 under a superiority hypothesis. The most appropriate interpretation is that the fitted model estimated a lower death hazard for nivolumab relative to docetaxel; the HR should not be translated into an absolute survival probability or an individual-level probability of benefit.
The trial also illustrates why statistical interpretation requires more than a P-value. The endpoint definition determines what is being measured, Kaplan-Meier methods describe the survival experience, censoring affects the risk sets, the log-rank test provides a time-to-event comparison, and the Cox model expresses the relative effect as a hazard ratio with an uncertainty interval.