This page separates reported trial results from statistical interpretation. Numerical results are restricted to the statistical analyses and trial characteristics provided in the ClinicalTrials.gov record.
1. Trial at a Glance
KEYNOTE-010 was a randomized, parallel, open-label phase 2/3 treatment trial enrolling 1034 participants with non-small cell lung cancer. Three treatment arms evaluated pembrolizumab at 2 mg/kg, pembrolizumab at 10 mg/kg, and docetaxel at 75 mg/m2.
| Feature | KEYNOTE-010 |
|---|---|
| Phase | Phase 2/3 |
| Status | Completed |
| Condition | Non Small Cell Lung Cancer (NSCLC) |
| Design | Randomized, parallel, open-label |
| Primary purpose | Treatment |
| Enrollment | 1034 |
| Arms | 3 |
| Primary endpoints | Overall Survival (OS); Progression-free Survival (PFS) by RECIST 1.1; Percentage of Participants Experiencing Adverse Events (AEs); Percentage of Participants Discontinuing Study Drug Due to AEs |
| Primary endpoint types | Binary; Time-to-event |
| Results posted | Yes |
| Statistical analyses posted | 12 |
| Lead sponsor | Merck Sharp & Dohme LLC |
| Sponsor type | Industry |
| Trial dates | 09 Aug 2013 to 30 Sep 2015 primary completion |
2. Clinical Question
The statistical question was whether pembrolizumab produced better time-to-event outcomes than docetaxel in previously treated participants with non-small cell lung cancer, with the registry reporting superiority comparisons for both overall survival and progression-free survival.
Population
Previously treated participants with non-small cell lung cancer, with the reported efficacy analyses focused on participants with strongly PD-L1 positive tumors and PD-L1 positive tumors.
Intervention
Pembrolizumab at 2 mg/kg or 10 mg/kg, according to the randomized treatment arm.
Comparator
Docetaxel 75 mg/m2.
Primary question
Does pembrolizumab improve OS and PFS relative to docetaxel under the prespecified superiority analyses?
3. Trial Design
Pembrolizumab 2 mg/kg First Course
- Pembrolizumab 2 mg/kg
- Randomized treatment arm
- Serious AEs: 124/339
Pembrolizumab 10 mg/kg First Course
- Pembrolizumab 10 mg/kg
- Randomized treatment arm
- Serious AEs: 133/343
Docetaxel 75 mg/m2
- Docetaxel 75 mg/m2
- Randomized treatment arm
- Serious AEs: 107/309
4. Endpoints
| Registered endpoint | Type | Time frame | Registry definition |
|---|---|---|---|
| Overall Survival (OS) | Time-to-event | Through pre-specified database cutoff date of 30 Sep 2015 (Up to approximately 24 months) | OS was defined as the time from randomization to death due to any cause. OS was analyzed using the Kaplan-Meier method and is reported in months. Per protocol, final analysis for this primary outcome measure was performed for the first pembrolizumab course and docetaxel treatment arms, with a protocol-specified analysis data cutoff date of 30 September 2015. |
| Progression-free Survival (PFS) by RECIST 1.1 | Time-to-event | Through pre-specified database cutoff date of 30 Sep 2015 (Up to approximately 24 months) | PFS was defined as the time from the first day of study treatment to the first documented disease progression per RECIST 1.1 based on blinded independent central radiologists' review or death due to any cause, whichever occurred first. |
| Percentage of Participants Experiencing Adverse Events (AEs) | Binary | Through pre-specified database cutoff date of 30 Sep 2015 (Up to approximately 24 months) | An AE was defined as any untoward medical occurrence in a participant administered a pharmaceutical product and which did not necessarily have to have a causal relationship with this treatment. |
| Percentage of Participants Discontinuing Study Drug Due to AEs | Binary | Through pre-specified database cutoff date of 30 Sep 2015 (Up to approximately 24 months) | The endpoint concerns discontinuation of study drug due to adverse events under the registry's AE definition. |
5. Statistical Methodology
Kaplan-Meier estimation
The registry defines OS as a time-to-event endpoint and states that it was analyzed using the Kaplan-Meier method. Kaplan-Meier estimation is designed for follow-up data in which some participants may not have experienced the event by the analysis cutoff. Rather than treating those participants as if they had survived indefinitely, the method uses the information available up to their censoring time.
Here, di denotes the number of events at event time ti, while ni is the number at risk immediately before that time.
Stratified log-rank testing
The registry method field reports a log-rank test for each registry-reported OS and PFS comparison. The analysis notes further state that the p-value was based on a stratified log-rank test, in accordance with the statistical analysis plan.
Stratified Cox regression
The ClinicalTrials.gov record states that the hazard ratio was based on a stratified Cox regression model with treatment as a covariate. This combines a treatment-effect estimate with a survival-analysis framework that accounts for the timing of events rather than simply counting how many events occurred.
An HR below 1 indicates a lower estimated instantaneous event rate in the numerator treatment group under the fitted model. It is not a probability, a relative risk, or a statement that every participant experiences the same proportional reduction.
Score-based confidence intervals for response
For the secondary ORR analyses, the registry reports the Miettinen & Nurminen method. This places the analysis among score-based confidence-interval methods for proportions, a group that includes the Miettinen-Nurminen, Newcombe, and Wilson methods. The reported effect measure is a risk difference, expressed as a difference in percentages.
Per-protocol analysis
The registry analysis records explicitly describe the OS, PFS, and ORR analyses as per-protocol. This is important because the interpretation of a treatment comparison depends partly on which participants are included in the analysis population. A per-protocol analysis can differ conceptually from an intention-to-treat analysis because it may restrict the analysis to participants meeting protocol-defined analysis criteria.
6. Statistical Methods Explained
Why was the log-rank test used?
OS and PFS are time-to-event endpoints. The log-rank test is designed to compare survival distributions between treatment groups while using the timing of events and accommodating censoring. In this trial, the registry reports the p-values as based on a stratified log-rank test.
What does an OS hazard ratio of 0.54 mean?
For the comparison of pembrolizumab 2 mg/kg with docetaxel in participants with strongly PD-L1 positive tumors, the reported HR was 0.54. In the context of the fitted stratified Cox model, this represents an estimated instantaneous death rate approximately 54% as large in the pembrolizumab group as in the docetaxel group. Equivalently, 1 − 0.54 = 0.46, so it corresponds to an approximately 46% lower estimated hazard under that model.
Does an HR of 0.50 mean half the patients survived?
No. A hazard ratio is a relative time-to-event measure. An HR of 0.50 does not mean that 50% of participants survived, that exactly half of participants benefited, or that each participant's individual risk was reduced by exactly 50%.
Why does the confidence interval matter?
The 95% confidence interval describes statistical uncertainty around the estimated treatment effect under the analysis framework. For the strongly PD-L1 positive OS comparison of pembrolizumab 10 mg/kg versus docetaxel, the HR was 0.50 with a 95% CI of 0.36–0.70. The interval therefore communicates considerably more information about precision than the point estimate alone.
What does a risk difference of 22.2 mean?
For strongly PD-L1 positive participants, the secondary ORR comparison of pembrolizumab 10 mg/kg versus docetaxel produced an adjusted risk difference of 22.2 percentage points, with a 95% CI of 14.0–30.7. This is a difference in response percentages, not a hazard ratio and not a statement about survival duration.
Why is the p-value not an effect-size measure?
A p-value addresses evidence against the statistical null hypothesis under the specified testing framework. It does not tell us the magnitude of the treatment effect. For example, the OS comparison of pembrolizumab 2 mg/kg versus docetaxel in strongly PD-L1 positive participants had an HR of 0.54 and P = 0.00024; the HR communicates the estimated relative effect, while the p-value addresses the strength of evidence against the null under the test used.
7. Results: Overall Survival
The registry contains four primary OS analyses: two pembrolizumab-versus-docetaxel comparisons in strongly PD-L1 positive participants and two corresponding comparisons in PD-L1 positive participants. All use a two-sided 95% confidence interval and a superiority hypothesis.
Strongly PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Docetaxel
Hazard ratio for death
95% CI: 0.38–0.77 · P = 0.00024
Comparison: Pembrolizumab 2 mg/kg vs Docetaxel 75 mg/m2
The reported HR of 0.54 indicates that the estimated instantaneous rate of death was approximately 46% lower in the pembrolizumab 2 mg/kg group than in the docetaxel group under the reported stratified Cox model. The estimate describes a relative hazard, not an absolute survival probability or the percentage of participants who benefited.
The 95% CI of 0.38–0.77 describes uncertainty around the estimated hazard ratio. It does not mean that individual participants experienced effects ranging only from 0.38 to 0.77. The p-value of 0.00024 addresses evidence against the null hypothesis under the reported stratified log-rank test; it does not measure the size of the treatment effect.
Because this is a time-to-event analysis, censoring and the assumptions underlying the Cox model remain relevant. The registry also describes the analysis population as per protocol and specifies the strongly PD-L1 positive population for this comparison.
Strongly PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Hazard ratio for death
95% CI: 0.36–0.70 · P = 0.00002
Comparison: Pembrolizumab 10 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.50 means that the estimated instantaneous rate of death was approximately 50% lower in the pembrolizumab 10 mg/kg group than in the docetaxel group under the reported model. This is a relative hazard statement; it does not mean that exactly half of participants survived or that each participant's mortality risk was reduced by exactly half.
The 95% CI of 0.36–0.70 provides the statistical uncertainty around the estimated HR. The interval is entirely below 1, while the two-sided p-value is 0.00002. The p-value should be interpreted as evidence against the null under the prespecified statistical test, rather than as a measure of how large or clinically important the effect is.
The registry specifies a stratified Cox model for the HR and a stratified log-rank test for the p-value. The analysis is also described as per protocol, so the analysis population is part of the interpretation.
PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Docetaxel
Hazard ratio for death
95% CI: 0.58–0.88 · P = 0.00076
Comparison: Pembrolizumab 2 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.71 corresponds to an approximately 29% lower estimated instantaneous rate of death in the pembrolizumab 2 mg/kg group under the reported model. It does not represent a 29% absolute reduction in mortality, nor does it imply that every participant experiences that same reduction.
The 95% CI of 0.58–0.88 quantifies uncertainty around the estimate. The p-value of 0.00076 is evidence against the null under the reported stratified log-rank test, but it does not itself describe the magnitude or clinical importance of the treatment effect.
The comparison is specifically for the registry-defined PD-L1 positive analysis population. It should not be silently generalized to populations that were not part of the ClinicalTrials.gov record.
PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Hazard ratio for death
95% CI: 0.49–0.75 · P < 0.00001
Comparison: Pembrolizumab 10 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.61 represents an approximately 39% lower estimated instantaneous rate of death in the pembrolizumab 10 mg/kg group relative to docetaxel under the reported stratified Cox model. Again, this is not an absolute mortality reduction and does not mean that 39% of participants were spared death.
The 95% CI of 0.49–0.75 describes the precision of the estimated HR. The p-value of <0.00001 indicates strong evidence against the null under the reported two-sided stratified log-rank test, but a very small p-value should not be mistaken for a very large effect.
The analysis population is described as per protocol and the comparison is specifically within the registry-reported PD-L1 positive analysis population.
8. Results: Progression-Free Survival
The registry supplies four primary PFS comparisons using RECIST 1.1: two within strongly PD-L1 positive participants and two within PD-L1 positive participants. PFS was defined from the first day of study treatment to the first documented progression by blinded independent central radiologists' review or death, whichever occurred first.
Strongly PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Docetaxel
Hazard ratio for progression or death
95% CI: 0.43–0.77 · P = 0.00009
Comparison: Pembrolizumab 2 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.58 corresponds to an approximately 42% lower estimated instantaneous rate of progression or death in the pembrolizumab 2 mg/kg group under the reported model. The event is a composite endpoint: the first documented RECIST 1.1 progression or death, whichever occurred first.
The 95% CI of 0.43–0.77 communicates uncertainty around the HR estimate. It does not represent the range of PFS outcomes experienced by individual participants. The p-value of 0.00009 addresses the statistical test and is not an effect-size measure.
As with OS, interpretation depends on the censoring process, the analysis population, and the assumptions associated with the Cox model. The registry specifies a per-protocol analysis and a stratified log-rank test.
Strongly PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Hazard ratio for progression or death
95% CI: 0.45–0.78 · P = 0.00007
Comparison: Pembrolizumab 10 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.59 indicates an approximately 41% lower estimated instantaneous rate of progression or death in the pembrolizumab 10 mg/kg group under the reported model. It does not mean that 41% of participants avoided progression or death, and it is not equivalent to an absolute percentage-point difference in PFS.
The 95% CI of 0.45–0.78 describes uncertainty around the estimate. The p-value of 0.00007 provides evidence against the null under the reported stratified log-rank test, but it does not quantify treatment benefit.
The analysis uses the registry-defined per-protocol population and the RECIST 1.1 progression-or-death endpoint.
PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Docetaxel
Hazard ratio for progression or death
95% CI: 0.73–1.04 · P = 0.06758
Comparison: Pembrolizumab 2 mg/kg vs Docetaxel 75 mg/m2
The estimated HR of 0.88 corresponds to an approximately 12% lower estimated instantaneous rate of progression or death under the reported model. The confidence interval of 0.73–1.04 spans 1, indicating that the registry-reported confidence interval includes the null hazard ratio.
The reported p-value is 0.06758. A p-value above a conventional 0.05 threshold should not be converted into a claim that the treatment has no effect; it indicates that this particular statistical test did not provide evidence crossing that conventional threshold. Conversely, the estimate should not be described as proof of benefit simply because it is below 1.
The result illustrates why the point estimate, confidence interval, and p-value should be interpreted together. The analysis is also specific to the registry-reported PD-L1 positive population and per-protocol analysis.
PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Hazard ratio for progression or death
95% CI: 0.66–0.94 · P = 0.00462
Comparison: Pembrolizumab 10 mg/kg vs Docetaxel 75 mg/m2
The HR of 0.79 corresponds to an approximately 21% lower estimated instantaneous rate of progression or death in the pembrolizumab 10 mg/kg group under the reported model. The endpoint remains a composite of progression by RECIST 1.1 or death.
The 95% CI of 0.66–0.94 describes uncertainty around the estimated HR and lies below 1. The p-value of 0.00462 is evidence against the null under the reported stratified log-rank test, but it does not measure effect magnitude or establish an absolute benefit.
The registry describes the analysis as per protocol. As with all time-to-event analyses, censoring and model assumptions remain relevant to interpretation.
9. Primary Results Summary
| Endpoint | Population | Comparison | Effect estimate | 95% CI | P-value |
|---|---|---|---|---|---|
| OS | Strongly PD-L1 positive | Pembro 2 mg/kg vs docetaxel | HR 0.54 | 0.38–0.77 | 0.00024 |
| OS | Strongly PD-L1 positive | Pembro 10 mg/kg vs docetaxel | HR 0.50 | 0.36–0.70 | 0.00002 |
| OS | PD-L1 positive | Pembro 2 mg/kg vs docetaxel | HR 0.71 | 0.58–0.88 | 0.00076 |
| OS | PD-L1 positive | Pembro 10 mg/kg vs docetaxel | HR 0.61 | 0.49–0.75 | <0.00001 |
| PFS | Strongly PD-L1 positive | Pembro 2 mg/kg vs docetaxel | HR 0.58 | 0.43–0.77 | 0.00009 |
| PFS | Strongly PD-L1 positive | Pembro 10 mg/kg vs docetaxel | HR 0.59 | 0.45–0.78 | 0.00007 |
| PFS | PD-L1 positive | Pembro 2 mg/kg vs docetaxel | HR 0.88 | 0.73–1.04 | 0.06758 |
| PFS | PD-L1 positive | Pembro 10 mg/kg vs docetaxel | HR 0.79 | 0.66–0.94 | 0.00462 |
10. Secondary Results: Overall Response Rate
The statistical analyses posted on ClinicalTrials.gov include four secondary ORR comparisons. Each uses the Miettinen & Nurminen method and reports an adjusted risk difference estimated by a stratified Miettinen & Nurminen method.
Strongly PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Pembrolizumab 10 mg/kg
Adjusted difference in response percentages
95% CI: −12.7–8.2 · P = 0.66608
Effect measure: adjusted risk difference in ORR
The reported risk difference of −2.3 percentage points compares the two pembrolizumab dose groups, with the numerator defined by the registry as pembrolizumab 2 mg/kg relative to pembrolizumab 10 mg/kg. The estimate therefore describes a difference in response percentages rather than a time-to-event effect.
The 95% CI of −12.7 to 8.2 includes zero, and the p-value is 0.66608. This means the registry-reported analysis does not provide evidence of a statistically detectable difference between these two dose groups under the reported test. It does not establish that the two doses are clinically identical.
Strongly PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Adjusted difference in response percentages
95% CI: 14.0–30.7 · P < 0.00001
Effect measure: adjusted risk difference in ORR
The adjusted risk difference of 22.2 percentage points indicates that the reported ORR percentage was estimated to be 22.2 percentage points higher for pembrolizumab 10 mg/kg than for docetaxel in the strongly PD-L1 positive population.
The 95% CI of 14.0–30.7 provides uncertainty around that difference. The p-value of <0.00001 addresses the statistical test and does not mean that the treatment effect is exactly 22.2 percentage points in every population or participant.
PD-L1 positive tumors: Pembrolizumab 2 mg/kg vs Docetaxel
Adjusted difference in response percentages
95% CI: 3.6–13.9 · P = 0.00045
Effect measure: adjusted risk difference in ORR
The risk difference of 8.7 percentage points represents the adjusted difference in ORR between pembrolizumab 2 mg/kg and docetaxel in the registry-reported PD-L1 positive analysis population.
The 95% CI of 3.6–13.9 gives the uncertainty around the estimate. Because this is a binary endpoint, the effect measure is directly expressed in percentage points rather than as a hazard ratio. The p-value of 0.00045 is evidence against the null under the reported statistical test, not a measure of the size of the response difference.
PD-L1 positive tumors: Pembrolizumab 10 mg/kg vs Docetaxel
Adjusted difference in response percentages
95% CI: 4.1–14.3 · P = 0.00024
Effect measure: adjusted risk difference in ORR
The adjusted ORR risk difference of 9.1 percentage points describes the difference between pembrolizumab 10 mg/kg and docetaxel in the registry-reported PD-L1 positive population.
The 95% CI of 4.1–14.3 quantifies uncertainty around the estimated difference. The p-value of 0.00024 provides evidence against the null under the reported stratified Miettinen & Nurminen analysis, but it should not be interpreted as an estimate of clinical magnitude.
11. ORR Results Summary
| Population | Comparison | Adjusted risk difference | 95% CI | P-value |
|---|---|---|---|---|
| Strongly PD-L1 positive | Pembro 2 mg/kg vs pembro 10 mg/kg | −2.3 | −12.7–8.2 | 0.66608 |
| Strongly PD-L1 positive | Pembro 10 mg/kg vs docetaxel | 22.2 | 14.0–30.7 | <0.00001 |
| PD-L1 positive | Pembro 2 mg/kg vs docetaxel | 8.7 | 3.6–13.9 | 0.00045 |
| PD-L1 positive | Pembro 10 mg/kg vs docetaxel | 9.1 | 4.1–14.3 | 0.00024 |
12. Safety Results
The ClinicalTrials.gov record contains serious adverse-event counts by arm. These are reported as affected participants divided by participants at risk. The registry also lists percentage of participants experiencing AEs and percentage discontinuing study drug due to AEs as primary endpoints, but no formal statistical-analysis estimates for those endpoints were included in the ClinicalTrials.gov record.
| Arm | Serious AEs affected / at risk | Reported measure |
|---|---|---|
| Pembrolizumab 2 mg/kg First Course | 124/339 | Affected participants / participants at risk |
| Pembrolizumab 10 mg/kg First Course | 133/343 | Affected participants / participants at risk |
| Docetaxel 75 mg/m2 | 107/309 | Affected participants / participants at risk |
| Pembrolizumab 2 mg/kg First Course to Pe | 5/14 | Affected participants / participants at risk |
| Pembrolizumab 10 mg/kg First Course to P | 2/7 | Affected participants / participants at risk |
| Docetaxel 75 mg/m2 Switched Over to Pem | 1/8 | Affected participants / participants at risk |
13. Why Time-to-Event Analysis Matters Here
OS and PFS are fundamentally different from binary endpoints such as ORR. A binary analysis asks whether an event has occurred by a defined assessment framework. A time-to-event analysis retains information about when the event occurred and can accommodate participants whose event time is not observed during follow-up.
OS
The clock begins at randomization and ends at death from any cause. Participants without an observed death by the analysis point contribute censored follow-up.
PFS
The clock begins on the first day of study treatment and ends at documented RECIST 1.1 progression or death, whichever occurs first.
ORR
The endpoint is binary and summarizes whether participants achieved a response under the RECIST 1.1 framework.
Safety
AE endpoints are binary in the registry, while the registry-reported serious-AE data are presented as affected participants divided by participants at risk.
14. Understanding the Hazard Ratios Together
| Comparison | OS HR | PFS HR | Interpretive point |
|---|---|---|---|
| Pembro 2 mg/kg vs docetaxel, strongly PD-L1 positive | 0.54 | 0.58 | Both reported estimates are below 1, with separate confidence intervals and p-values. |
| Pembro 10 mg/kg vs docetaxel, strongly PD-L1 positive | 0.50 | 0.59 | Both reported estimates are below 1; OS and PFS remain separate endpoints. |
| Pembro 2 mg/kg vs docetaxel, PD-L1 positive | 0.71 | 0.88 | The registry-reported OS estimate is below 1, while the PFS CI includes 1. |
| Pembro 10 mg/kg vs docetaxel, PD-L1 positive | 0.61 | 0.79 | Both reported estimates are below 1, with distinct uncertainty intervals. |
The table illustrates why endpoint-specific interpretation matters. OS and PFS are related but not interchangeable. A hazard ratio for PFS describes the relative rate of progression or death, whereas an OS hazard ratio describes the relative rate of death. Neither should be substituted for the other.
15. Multiplicity and Multiple Comparisons
The ClinicalTrials.gov record contains a three-arm design and multiple formal statistical analyses: eight primary-endpoint analyses and four secondary ORR analyses. The analyses include two pembrolizumab doses and multiple PD-L1-defined analysis populations.
This is especially important because a collection of pairwise comparisons can generate more opportunities for a nominally unusual result than a single prespecified comparison. A complete statistical-analysis plan would be needed to determine precisely which hypotheses were confirmatory, how alpha was allocated, and how success on one endpoint affected testing of another.
16. Stratified Analysis
The ClinicalTrials.gov record repeatedly identify stratified analysis. For OS and PFS, the registry states that the hazard ratio was based on a stratified Cox regression model and that the p-value was based on a stratified log-rank test. For ORR, the adjusted risk difference was estimated by a stratified Miettinen & Nurminen method.
Stratification allows the analysis to account for prespecified grouping factors without necessarily estimating a separate treatment coefficient for every stratum. The exact stratification variables are not included in the ClinicalTrials.gov record, so they are not inferred here.
One important lesson is that the word "stratified" describes the analysis framework; it does not by itself identify which variables were used or prove that treatment effects are identical across all strata.
17. Per-Protocol Analysis and Interpretation
The registry-reported OS, PFS, and ORR records identify the analysis population as per protocol. This is an important feature of the reported results because the population entering an analysis can affect the estimand being described.
Randomization
Randomization creates the treatment groups and is the design mechanism supporting a causal comparison.
Per-protocol analysis
The reported efficacy analyses are described as per protocol rather than simply as all randomized participants.
Why this matters
Exclusions or restrictions after randomization can change the composition of the analyzed population.
Interpretive discipline
The exact registry-defined analysis population should be retained rather than silently relabeled as ITT.
This distinction is particularly relevant when interpreting a hazard ratio. The HR is conditional not only on the statistical model but also on the participants and events included in the analysis.
18. Clinical Biostats Interpretation of Relative Effects
A hazard ratio below 1 indicates a lower estimated instantaneous event rate in the numerator group under the fitted model. For example, an HR of 0.71 corresponds to an approximately 29% lower estimated hazard, calculated as 1 − 0.71. That statement should not be transformed into an absolute survival difference without additional survival estimates.
A 95% confidence interval describes uncertainty around the estimated parameter. For example, an HR of 0.88 with a 95% CI of 0.73–1.04 communicates both the point estimate and its uncertainty. The interval includes 1, so the conventional null value is compatible with the registry-reported interval.
A risk difference is an absolute difference in proportions. The reported ORR difference of 22.2 means 22.2 percentage points, not a 22.2% relative increase and not a hazard ratio.
A p-value quantifies the compatibility of the observed data with a null hypothesis under the specified testing procedure. It does not measure the probability that the null hypothesis is true, the probability that the treatment works, or the magnitude of the treatment effect.
19. Important Statistical Assumptions
- Proportional hazards: The Cox hazard ratio is a model-based relative measure. A single HR is most naturally interpreted when the proportional-hazards framework is reasonable over the relevant follow-up.
- Censoring: Kaplan-Meier and Cox methods depend on assumptions about censoring and the information contributed by participants whose event times are not observed.
- Analysis population: The registry-reported efficacy records specify per-protocol analyses, so the reported estimates should not be relabeled as a different analysis population.
- Endpoint definition: PFS is defined using RECIST 1.1 progression or death, whichever occurs first. The endpoint is therefore not simply tumor shrinkage.
- Stratification: The ClinicalTrials.gov record specifies stratified survival analyses and stratified response analysis, but the exact stratification variables are not included in the ClinicalTrials.gov record.
- Multiplicity: Multiple arms, endpoints, and populations create a need to distinguish individual nominal p-values from the trial's complete confirmatory testing strategy.
20. What the Results Do — and Do Not — Establish
| Reported result | What it supports | What it does not by itself establish |
|---|---|---|
| OS HR 0.54 | A lower estimated hazard of death for pembrolizumab 2 mg/kg vs docetaxel in the strongly PD-L1 positive analysis. | An absolute survival difference, cure rate, or identical individual-level risk reduction. |
| OS HR 0.50 | A lower estimated hazard of death for pembrolizumab 10 mg/kg vs docetaxel in the strongly PD-L1 positive analysis. | That exactly half of participants survived or benefited. |
| PFS HR 0.88 | The estimated progression-or-death hazard was below 1 for pembrolizumab 2 mg/kg vs docetaxel in the registry-reported PD-L1 positive analysis. | A definitive absence of any treatment effect. |
| ORR difference 22.2 | A 22.2-percentage-point adjusted difference in response percentages for pembrolizumab 10 mg/kg vs docetaxel in strongly PD-L1 positive participants. | A corresponding OS improvement or a 22.2% relative increase. |
| Serious AE 124/339 | The registry-reported number of affected participants and participants at risk in the pembrolizumab 2 mg/kg first-course category. | A randomized causal comparison with another arm without a specified safety analysis. |
21. Limitations
- Registry-level detail: The ClinicalTrials.gov record does not include the complete statistical analysis plan, so the full hypothesis hierarchy and multiplicity-control strategy cannot be reconstructed.
- Missing survival summaries: The statistical analyses posted on ClinicalTrials.gov do not report median OS, median PFS, survival probabilities, or event counts for the primary efficacy comparisons.
- No participant-level data: Individual event and censoring times are not available in the ClinicalTrials.gov record, preventing independent reconstruction of Kaplan-Meier curves.
- Safety comparison: Serious-AE counts are provided by arm, but the ClinicalTrials.gov record does not include formal statistical comparisons or confidence intervals for those safety outcomes.
- Multiple comparisons: The trial contains multiple treatment contrasts and analysis populations. The ClinicalTrials.gov record does not specify enough of the complete multiplicity procedure to determine the adjusted significance level for each reported p-value.
- Hazard-ratio assumptions: A single Cox HR summarizes a relative event-rate contrast and should not automatically be interpreted as a constant treatment effect at every point in time.
- Generalizability: The ClinicalTrials.gov record identifies a specific previously treated NSCLC trial population. Results should not automatically be generalized to populations outside the studied eligibility framework.
22. Why This Trial Matters Statistically
KEYNOTE-010 is a useful statistical teaching case because the ClinicalTrials.gov record combines a three-arm randomized design with two time-to-event primary efficacy endpoints, multiple treatment contrasts, PD-L1-defined analysis populations, per-protocol analyses, stratified survival testing, Cox hazard ratios, and score-based confidence intervals for binary response outcomes.
| Concept | How it appears in KEYNOTE-010 |
|---|---|
| Randomization | Randomized allocation with three parallel treatment arms. |
| Time-to-event analysis | OS and PFS are primary time-to-event endpoints. |
| Kaplan-Meier estimation | The registry explicitly states that OS was analyzed using Kaplan-Meier methods. |
| Hazard ratio | OS and PFS treatment effects are reported as hazard ratios. |
| Log-rank test | Primary OS and PFS p-values are based on stratified log-rank testing. |
| Cox regression | HRs are based on a stratified Cox regression model with treatment as a covariate. |
| Risk difference | Secondary ORR effects are reported as adjusted differences in percentages. |
| Miettinen-Nurminen method | Secondary ORR confidence intervals and tests use the reported Miettinen & Nurminen framework. |
| Per-protocol analysis | The registry-reported efficacy analyses are identified as per protocol. |
| Multiple comparisons | Eight primary-endpoint analyses and four secondary statistical analyses are posted. |
| Subpopulation analysis | Results are reported separately for strongly PD-L1 positive and PD-L1 positive populations. |
| Safety analysis | Serious adverse-event counts are reported separately by treatment category. |
23. Related Tutorials
Learn more about the methods used in this trial:
24. Related Statistical Calculators
25. Sources
- ClinicalTrials.gov: NCT01905657.
- Linked publication: PubMed 26712084.
- Linked publication: PubMed 35101941.
- Linked publication: PubMed 33439727.
- Linked publication: PubMed 32305010.
- Linked publication: PubMed 32078391.
Continue through the Clinical Biostats statistical library
Explore the statistical methods behind randomized clinical trials, survival endpoints, confidence intervals, and categorical treatment effects.
26. Record Summary
KEYNOTE-010 provides a compact example of several important principles in clinical-trial statistics. The trial randomized 1034 participants across three parallel treatment arms and reported formal superiority analyses for OS and PFS using stratified log-rank testing and stratified Cox regression. The secondary ORR analyses used adjusted risk differences with the Miettinen & Nurminen method. The reported results therefore require more than simply reading the p-values: the treatment contrast, analysis population, endpoint definition, hazard ratio or risk difference, confidence interval, and statistical method all contribute to the interpretation.