This page separates reported trial results from statistical interpretation. Numerical trial results are restricted to the information contained in the ClinicalTrials.gov record. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
KEYNOTE-716 is a randomized, double-blind, parallel-group phase 3 trial evaluating pembrolizumab versus placebo in participants with resected high-risk stage II melanoma. The registered primary endpoint is recurrence-free survival (RFS), a time-to-event endpoint assessed from randomization.
| Feature | KEYNOTE-716 |
|---|---|
| Phase | Phase 3 |
| Condition | Melanoma |
| Brief title | Safety and Efficacy of Pembrolizumab Compared to Placebo in Resected High-risk Stage II Melanoma |
| Design | Randomized, double-blind, parallel-group |
| Allocation | Randomized |
| Primary purpose | Treatment |
| Enrollment | 976 |
| Primary endpoint | Recurrence-free survival (RFS) |
| Primary endpoint type | Time-to-event |
| Primary analysis method | Log-rank test; Cox regression model with stratified analysis |
| Lead sponsor | Merck Sharp & Dohme LLC |
| Sponsor type | Industry |
| ClinicalTrials.gov | NCT03553836 |
2. Clinical Question
The central statistical question was whether participants randomized to pembrolizumab experienced a different recurrence-free survival experience than participants randomized to placebo. The registered primary endpoint defines recurrence as a melanoma recurrence at any site or death from any cause, whichever occurs first.
Population
Participants with resected high-risk stage II melanoma, as described by the trial's registered brief title.
Intervention
Pembrolizumab.
Comparator
Placebo.
Primary question
Does pembrolizumab improve recurrence-free survival relative to placebo?
3. Trial Design
Pembrolizumab
- Biological intervention
- Compared with placebo
Placebo
- Other intervention category in the registry
- Comparator for the pembrolizumab group
The ClinicalTrials.gov record identifies the total enrollment as 976 and the study as having two arms, but do not provide randomized sample sizes by arm in the data used for this page. The serious-adverse-event denominator data are reported separately below and should not be substituted for randomized arm sizes.
4. Trial Timeline
Study start
The registered study start date is September 12, 2018.
Primary completion
The registered primary completion date is June 21, 2021.
Active, not recruiting
The ClinicalTrials.gov record classifies KEYNOTE-716 as active, not recruiting.
5. Primary Endpoint
| Endpoint | Registered definition | Time frame |
|---|---|---|
| Recurrence-free Survival (RFS) | RFS was defined as the time from randomization to any of the following events: recurrence of melanoma at any site (local, in-transit or regional lymph nodes or distant recurrence) per Response Evaluation Criteria in Solid Tumors Version 1.1 (RECIST 1.1) or death due to any cause, whichever occurs first. | Up to ~32.7 months |
This endpoint combines two clinically distinct event pathways into one time-to-event outcome: recurrence of melanoma and death from any cause. The event that occurs first determines the RFS event time. Participants who have not experienced either event by the relevant end of observation are handled as censored observations in a conventional survival-analysis framework.
6. Results
The ClinicalTrials.gov record contains one formal primary-endpoint statistical analysis and two secondary safety analyses. The primary RFS analysis compares all randomized participants according to their randomized groups.
Recurrence-free Survival
Hazard ratio for recurrence or death
95% CI: 0.45–0.82 · P = 0.00046
Analysis population: all randomized participants
| Endpoint | Pembrolizumab vs placebo | Statistical analysis |
|---|---|---|
| Recurrence-free Survival | HR 0.61 (95% CI 0.45–0.82) | Log-rank test; Cox regression with stratification |
A hazard ratio of 0.61 means that, under the fitted time-to-event model, the estimated instantaneous rate of experiencing an RFS event in the pembrolizumab group was approximately 61% of the corresponding rate in the placebo group over the analyzed follow-up. Equivalently, the estimate corresponds to an approximately 39% lower estimated hazard of recurrence or death.
The HR does not mean that 39% of participants avoided recurrence, that 39% of participants were cured, or that every individual participant had exactly a 39% reduction in risk. It is a relative time-to-event measure, not an absolute probability.
The two-sided 95% confidence interval of 0.45–0.82 describes statistical uncertainty around the estimated hazard ratio under the analysis framework. It does not describe the range of outcomes that individual patients might experience.
The P = 0.00046 result addresses evidence against the null hypothesis under the specified testing framework. A p-value is not a measure of effect size, clinical importance, or the probability that the treatment is effective. The magnitude and precision of the treatment effect are better represented by the HR and its confidence interval.
The analysis notes specify a Cox regression model using Efron's method of tie handling, with treatment as a covariate and stratification by melanoma T Stage (T3b, T4a, T4b). As with other Cox-model hazard ratios, interpretation depends on the model being an appropriate representation of the treatment effect over time; a single HR can be less descriptive if hazards are substantially non-proportional.
7. How the Primary Analysis Works
Kaplan-Meier estimation
RFS is a time-to-event endpoint, so the underlying survival experience can be represented using the Kaplan-Meier estimator. It accounts for right censoring by allowing participants who have not yet experienced recurrence or death to contribute information until their last relevant observation.
Here, di represents the number of events at event time ti, while ni is the number of participants at risk immediately before that time.
The registry-reported statistical-analyses data do not provide a complete set of event times and censoring times, so this page does not attempt to reconstruct a Kaplan-Meier curve. The reported HR and confidence interval are presented exactly as reported in the registry rather than being reverse-engineered from an unavailable survival curve.
Log-rank test
The registry reports a log-rank method for the primary RFS comparison. The log-rank test compares the observed and expected numbers of events between randomized treatment groups over the follow-up period. It is particularly suited to randomized time-to-event comparisons because it uses information across the observed event times rather than reducing follow-up to a single fixed time point.
Stratified Cox regression
The formal analysis notes state that the hazard ratio was based on a Cox regression model with Efron's method of tie handling, with treatment as a covariate and stratification by melanoma T Stage (T3b, T4a, T4b).
In a stratified Cox model, the baseline hazard is allowed to differ across the specified strata while the treatment effect is estimated across the stratified analysis. This can account for clinically relevant baseline differences associated with the stratification variable without requiring a single common baseline hazard across all strata.
For KEYNOTE-716, the reported HR of 0.61 is a relative treatment-effect estimate for the RFS endpoint. It should not be substituted for an absolute recurrence probability, median RFS, or risk difference.
Efron's method for ties
The registry-reported analysis notes explicitly identify Efron's method of tie handling. Ties occur when multiple participants have event times recorded at the same observed time. Cox regression requires a convention for handling such tied event times, and Efron's method provides one such approximation to the partial likelihood calculation.
8. Secondary Safety Results
The registry data contain two formal secondary statistical analyses concerning adverse events. Both compare pembrolizumab with placebo among all randomized participants who received at least one dose of study treatment. The reported effect measure is a difference in percentage, normalized here as a risk difference.
Participants With at Least One Adverse Event
Difference in percentage
95% CI: 1.0–7.3 · Two-sided
Time frame: up to ~19.3 months
The reported estimate is a 4.1 percentage-point difference between the randomized treatment groups for participants experiencing at least one adverse event. The ClinicalTrials.gov record identifies the effect measure as a difference in percentage and state that the calculation was based on the Miettinen & Nurminen method.
The 95% confidence interval of 1.0–7.3 percentage points quantifies uncertainty around the estimated between-group difference. Because the interval is expressed in percentage points, it should not be interpreted as a 4.1% relative increase or as a hazard ratio.
The analysis population is not the full randomized population: it includes randomized participants who received at least one dose of study treatment. This distinction matters when comparing safety results with the primary efficacy analysis, which used all randomized participants.
Participants Who Discontinued Study Treatment Due to an Adverse Event
Difference in percentage
95% CI: 9.1–16.9 · Two-sided
Time frame: up to ~19.3 months
The reported estimate is a 12.9 percentage-point difference between pembrolizumab and placebo for discontinuation of study treatment due to an adverse event. The 95% confidence interval is 9.1–16.9 percentage points.
This is an absolute between-group difference, not a relative risk and not a hazard ratio. The confidence interval describes uncertainty around the percentage-point difference; it does not describe the probability that an individual participant will discontinue treatment.
The Miettinen & Nurminen method is appropriate to the registry's reported comparison of two binary proportions and provides a confidence interval for the difference between the groups. The ClinicalTrials.gov record does not report a separate formal p-value for this secondary analysis.
Serious Adverse Events by Arm
| Group | Participants affected / at risk |
|---|---|
| Pembrolizumab | 101 / 483 |
| Placebo | 91 / 486 |
| Placebo switched over to pembrolizumab | 6 / 45 |
The registry-reported serious-adverse-event data are reported as affected participants divided by participants at risk. These figures should be kept distinct from the two formal secondary analyses above: the registry does not identify a formal statistical comparison for these serious-adverse-event counts in the ClinicalTrials.gov record.
9. Understanding the Risk Difference
The two reported safety analyses use a different effect measure from the primary RFS analysis. This distinction is fundamental when reading a clinical-trial results table.
| Measure | KEYNOTE-716 example | What it describes |
|---|---|---|
| Hazard ratio | 0.61 | Relative difference in the modeled instantaneous event rate over time |
| Risk difference | 4.1 percentage points | Absolute difference in the probability/proportion for a binary outcome |
| Risk difference | 12.9 percentage points | Absolute difference for treatment discontinuation due to an adverse event |
A hazard ratio and a risk difference cannot be compared numerically as though they were measurements on the same scale. The HR summarizes a time-to-event treatment effect, whereas a risk difference summarizes an absolute difference in a binary outcome over the specified analysis time frame.
10. Statistical Methodology
Randomization
KEYNOTE-716 is registered as a randomized trial. Randomization is central to the causal interpretation of a treatment comparison because assignment is determined by the trial design rather than by participants' or investigators' observed prognostic characteristics.
For the primary efficacy analysis, the registry specifies all randomized participants as the analysis population. This preserves the treatment assignment established at randomization when estimating the primary treatment comparison.
Stratified analysis
The primary analysis notes specify stratification by melanoma T Stage, with the listed strata T3b, T4a, and T4b. Stratification allows the Cox model to accommodate potentially different baseline recurrence hazards across these strata while estimating a common treatment effect across them.
Log-rank testing
The log-rank test evaluates whether the observed event patterns differ between treatment groups over time. Unlike a simple comparison of proportions at one fixed time point, it incorporates the timing of observed RFS events and accommodates censored follow-up.
Cox regression
The registry's analysis notes identify Cox regression as the model used to estimate the hazard ratio. Treatment is included as a covariate, while melanoma T Stage is used for stratification. Efron's method is used to handle tied event times.
Confidence intervals
The primary RFS estimate is accompanied by a two-sided 95% confidence interval. Confidence intervals are particularly important for interpreting clinical-trial estimates because the point estimate alone does not reveal how precisely the treatment effect has been estimated.
Miettinen & Nurminen method
For both reported binary safety analyses, the registry states that the difference in percentage was based on the Miettinen & Nurminen method. This provides a confidence-interval framework for comparing two binomial proportions through their difference rather than treating the difference as a simple descriptive subtraction with an unrelated standard-error calculation.
11. Statistical Methods Explained
Why was a log-rank test used for RFS?
RFS records not only whether recurrence or death occurred, but also when it occurred. Some participants may remain event-free at the end of observation and therefore be censored. The log-rank test is designed for this time-to-event structure and compares the treatment groups across the observed follow-up rather than forcing the analysis into a single binary endpoint.
What does an RFS hazard ratio of 0.61 mean?
It means that the fitted model estimates the instantaneous RFS-event rate in the pembrolizumab group to be 0.61 times that of the placebo group, under the model and follow-up used for the analysis. It does not mean that 61% of participants experienced an event, nor does it mean that every participant's individual risk was reduced by 39%.
Why was Cox regression used in addition to the log-rank test?
The log-rank test supplies a formal comparison of the time-to-event distributions, while the Cox model provides an interpretable effect estimate through the hazard ratio and its confidence interval. The two methods therefore answer related but different statistical questions.
Why stratify the Cox model by melanoma T Stage?
The analysis notes specify T Stage as a stratification factor with T3b, T4a, and T4b strata. Stratification allows the baseline hazard to vary by these strata while estimating the treatment effect across the stratified population. This can be preferable to forcing all participants to share one baseline hazard function.
What does the 95% CI of 0.45–0.82 tell us?
It describes uncertainty around the estimated HR of 0.61 under the statistical model and sampling framework. It does not provide a prediction interval for individual patients, and it does not say that the true outcome for a future participant must fall between 0.45 and 0.82.
Why are the safety analyses reported as risk differences?
The two secondary safety endpoints are binary outcomes: whether a participant experienced at least one adverse event and whether a participant discontinued study treatment because of an adverse event. A difference in percentage expresses the between-group contrast directly in percentage points, which is different from the time-to-event hazard ratio used for RFS.
12. Primary Analysis Population vs Safety Population
| Analysis | Population in the ClinicalTrials.gov record | Why the distinction matters |
|---|---|---|
| Primary RFS | All randomized participants | Preserves randomized treatment assignment for the primary efficacy comparison. |
| At least one AE | All randomized participants who received at least one dose of study treatment | Focuses the safety analysis on participants with treatment exposure. |
| Discontinuation due to AE | All randomized participants who received at least one dose of study treatment | Requires exposure to study treatment before treatment discontinuation can occur. |
This difference is not a technical footnote. Efficacy and safety answer different questions and can appropriately use different analysis populations. The important point is to avoid silently treating the two populations as interchangeable.
13. What the Primary Result Does — and Does Not — Establish
What the HR establishes
The reported model estimates a lower rate of the combined RFS event of melanoma recurrence or death in the pembrolizumab group relative to placebo, with HR 0.61 and 95% CI 0.45–0.82.
What the HR does not establish
It does not give an absolute recurrence probability, median RFS, number needed to treat, or the treatment effect for every individual participant.
What the p-value establishes
P = 0.00046 quantifies the statistical evidence against the relevant null hypothesis under the specified testing framework.
What the p-value does not establish
It is not a measure of effect magnitude, clinical importance, or the probability that the treatment hypothesis is true.
The most informative reading therefore combines the HR, confidence interval, p-value, endpoint definition, analysis population, and model specification rather than focusing on any one number in isolation.
14. Censoring and Time-to-Event Interpretation
RFS is inherently subject to censoring because not every participant will necessarily experience recurrence or death during the period in which their outcome is observed. A participant who remains free of an RFS event through their available follow-up can contribute information up to the point at which follow-up ends.
This is one reason a time-to-event analysis is preferable to simply calculating the proportion of participants who have experienced recurrence at an arbitrary time without accounting for unequal follow-up.
15. Proportional-Hazards Considerations
The Cox model produces a hazard ratio that is commonly interpreted as a relative comparison of the instantaneous event rates between treatment groups. A particularly important modeling consideration is whether the relative hazards are reasonably stable over time.
If the proportional-hazards assumption is substantially violated, a single HR can compress a changing treatment effect into one summary number. In that situation, Kaplan-Meier estimates, time-specific survival estimates, restricted mean survival time, or time-varying effect methods may provide additional insight.
The registry-reported KEYNOTE-716 data report the Cox HR but do not provide a proportional-hazards diagnostic. Accordingly, the HR should be interpreted as the reported model-based summary rather than as proof that the relative treatment effect was constant at every point during follow-up.
16. Multiplicity and Hypothesis Testing
The ClinicalTrials.gov record identifies the primary RFS hypothesis type as Superiority. They do not provide a detailed multiplicity hierarchy, alpha-allocation scheme, or separate interim-analysis plan in the ClinicalTrials.gov record.
| Feature | What the ClinicalTrials.gov record establishes |
|---|---|
| Primary endpoint | Recurrence-free survival |
| Hypothesis type | Superiority |
| Primary p-value | 0.00046 |
| Confidence interval | Two-sided 95% CI |
| Multiplicity adjustment details | Not specified in the ClinicalTrials.gov record |
| Interim-analysis details | Not specified in the ClinicalTrials.gov record |
That distinction matters because a p-value can only be interpreted fully within the testing framework that generated it. The ClinicalTrials.gov record supports describing the result as a superiority analysis with a two-sided p-value of 0.00046; they do not support reconstructing a more detailed alpha-spending or multiplicity procedure.
17. Secondary Endpoint Interpretation
The two secondary analyses illustrate a useful statistical distinction between efficacy time-to-event endpoints and safety binary endpoints.
| Endpoint | Type | Effect measure | Estimate | 95% CI |
|---|---|---|---|---|
| Recurrence-free Survival | Time-to-event | Hazard ratio | 0.61 | 0.45–0.82 |
| At least one adverse event | Binary | Risk difference | 4.1 | 1.0–7.3 |
| Discontinued due to an AE | Binary | Risk difference | 12.9 | 9.1–16.9 |
The estimates should not be placed on a common numerical scale. An HR of 0.61 and a 12.9-percentage-point difference are different quantities describing different endpoint structures.
18. Safety Analysis in Context
The ClinicalTrials.gov record shows serious adverse events in both randomized treatment groups and also identify a separate group of participants who switched from placebo to pembrolizumab.
Serious AEs: pembrolizumab
101 of 483 participants in the registry-reported affected/at-risk data.
Serious AEs: placebo
91 of 486 participants in the registry-reported affected/at-risk data.
Switched participants
6 of 45 participants in the placebo-switched-over-to-pembrolizumab group.
Formal safety analyses
The statistical analyses posted on ClinicalTrials.gov concern any AE and treatment discontinuation due to an AE, not a formal comparison of the serious-AE counts.
This is an important example of why statistical interpretation must follow the endpoint definition and analysis population. A descriptive count of serious adverse events does not automatically become a formal hypothesis test merely because two arms are displayed side by side.
19. Limitations
- Limited numerical scope: The ClinicalTrials.gov record contains one primary statistical analysis and two secondary safety analyses, so this page does not introduce median RFS, survival probabilities, subgroup estimates, or other outcomes not contained in the ClinicalTrials.gov record.
- No arm-specific randomized sample sizes reported: Total enrollment is 976, but the ClinicalTrials.gov record does not provide randomized participant counts for each arm.
- Safety populations differ from the efficacy population: The two binary safety analyses include randomized participants who received at least one dose, whereas the primary RFS analysis uses all randomized participants.
- Censoring: RFS is a time-to-event endpoint and therefore depends on appropriate handling of participants whose event status is not observed for the entire follow-up period.
- Proportional hazards: The Cox HR is a model-based summary and should not automatically be interpreted as a constant treatment effect at every time point.
- Limited multiplicity information: The ClinicalTrials.gov record identifies a superiority hypothesis but do not provide a detailed multiplicity or alpha-spending framework.
- Limited interim-analysis information: The ClinicalTrials.gov record does not specify an interim-analysis schedule or stopping boundary.
- Serious-AE comparison: The registry-reported serious-adverse-event counts are descriptive affected/at-risk data; the ClinicalTrials.gov record does not provide a formal statistical comparison for them.
- Follow-up summary: The registry-reported primary endpoint time frame is up to approximately 32.7 months, but no median follow-up or median RFS is provided in the ClinicalTrials.gov record.
20. Why This Trial Matters Statistically
KEYNOTE-716 is a useful teaching example because the ClinicalTrials.gov record connects a randomized treatment comparison to a clinically meaningful time-to-event endpoint and then contrasts that analysis with binary safety outcomes.
| Concept | How it appears in KEYNOTE-716 |
|---|---|
| Randomization | Participants were randomized to two parallel treatment arms. |
| Blinding | The trial is registered as double-blind. |
| Time-to-event endpoint | RFS is measured from randomization to recurrence or death. |
| Kaplan-Meier estimation | RFS is a survival-analysis endpoint for which Kaplan-Meier estimation is a natural descriptive framework. |
| Log-rank test | The registry reports a log-rank method for the primary RFS comparison. |
| Hazard ratio | RFS treatment effect is reported as HR 0.61. |
| Confidence interval | The HR has a two-sided 95% CI of 0.45–0.82. |
| Stratified analysis | Cox regression is stratified by melanoma T Stage: T3b, T4a, T4b. |
| Cox regression | The HR is based on a Cox model with treatment as a covariate. |
| Tie handling | Efron's method is specified for tied event times. |
| Risk difference | Secondary binary safety outcomes are reported as differences in percentage. |
| Miettinen & Nurminen method | The registry-reported safety analyses use this method for the percentage differences. |
| Analysis populations | Efficacy uses all randomized participants; safety analyses use randomized participants who received at least one dose. |
21. Statistical Concepts in This Trial
Learn more about the methods used in this trial:
22. Related Statistical Calculators
23. Sources
- ClinicalTrials.gov: NCT03553836 — KEYNOTE-716.
- PubMed: PMID 38452313.
- PubMed: PMID 41701495.
- PubMed: PMID 40198940.
- PubMed: PMID 40078077.
- PubMed: PMID 38485189.
Continue through Clinical Biostats
Connect this trial's endpoints and statistical methods with deeper tutorials, statistical calculators, and other clinical-trial analyses.
24. Record Summary
KEYNOTE-716 provides a compact example of modern randomized clinical-trial survival analysis. The ClinicalTrials.gov record describes a randomized, double-blind, parallel-group phase 3 trial with 976 participants and a primary recurrence-free survival endpoint assessed over up to approximately 32.7 months. The primary analysis used a log-rank framework and a stratified Cox regression model, with melanoma T Stage as the stratification factor and Efron's method for tied event times. The reported RFS hazard ratio was 0.61, with a two-sided 95% confidence interval of 0.45–0.82 and P = 0.00046.
The secondary safety analyses illustrate a different statistical structure. Binary adverse-event outcomes were summarized using differences in percentage, with Miettinen & Nurminen methods used for the reported confidence intervals. The difference in percentage was 4.1 for participants experiencing at least one adverse event and 12.9 for participants discontinuing study treatment due to an adverse event. These measures should not be interpreted in the same way as the RFS hazard ratio.