← Clinical Trials
Advanced Urothelial Cancer Phase 3 Completed NCT02256436

KEYNOTE-045: Complete Statistical Analysis of Pembrolizumab in Advanced Urothelial Cancer

An independent statistical analysis of the randomized phase 3 KEYNOTE-045 trial comparing pembrolizumab with paclitaxel, docetaxel, or vinflunine in participants with advanced urothelial cancer.

Trial start: 22-Oct-2014  ·  Primary completion: 07-Sep-2016  ·  Enrollment: 542  ·  Lead sponsor: Merck Sharp & Dohme LLC
Scope of this record

This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record. Numerical results on this page are restricted to the ClinicalTrials.gov record.

1. Trial at a Glance

KEYNOTE-045 was a randomized, parallel-group, phase 3 trial in advanced urothelial cancer. The trial compared pembrolizumab with a control consisting of paclitaxel, docetaxel, or vinflunine, with overall survival and progression-free survival assessed as primary endpoints in all participants and in PD-L1-defined populations.

542
Enrolled
Randomized phase 3 trial
2
Arms
Parallel allocation
0.73
All-participant OS HR
95% CI 0.59–0.91
0.00224
OS P-value
All participants
FeatureKEYNOTE-045
Trial nameKEYNOTE-045
ClinicalTrials.gov identifierNCT02256436
PhasePhase 3
ConditionUrothelial Cancer
AllocationRandomized
Design modelParallel
MaskingNone
Primary purposeTreatment
Enrollment542
Arms2
Primary endpoint typesBinary; Time-to-event
Results postedYes
Outcome measures posted20
Statistical analyses posted15
Primary endpoint analyses6
Primary analyses with estimate + CI6

2. Clinical Question

The primary statistical question was whether pembrolizumab differed from the control treatment with respect to progression-free survival and overall survival in participants with advanced urothelial cancer. The registered primary endpoints included all participants as well as participants with PD-L1-positive and strongly PD-L1-positive tumors.

Population

Participants with advanced urothelial cancer enrolled in the phase 3 KEYNOTE-045 trial. The ClinicalTrials.gov record reports 542 enrolled participants.

Intervention

Pembrolizumab, classified in the registry data as a biological intervention.

Comparator

The control intervention consisted of paclitaxel, docetaxel, or vinflunine.

Primary question

How does pembrolizumab compare with control for PFS and OS, including the prespecified PD-L1-defined populations?

3. Trial Design

01
Enroll542 participants
02
Randomize2 parallel arms
03
TreatPembrolizumab or control
04
AssessPFS, OS, response
05
AnalyzeRegistry-posted results
INTERVENTION ARM

Pembrolizumab

  • Pembrolizumab was the intervention assigned in one randomized arm.
  • The ClinicalTrials.gov record classifies pembrolizumab as a biological intervention.
CONTROL ARM

Paclitaxel, docetaxel, or vinflunine

  • The control intervention consisted of paclitaxel, docetaxel, or vinflunine.
  • The ClinicalTrials.gov record identifies the comparison as Control vs Pembrolizumab.
Allocation
Randomized allocation was used.
Model
Parallel-group trial design.
Masking
None.
Primary purpose
Treatment.

The trial began on 22-Oct-2014 and reached primary completion on 07-Sep-2016. The primary time-to-event analyses used a database cutoff of 07-Sep-2016, corresponding to up to approximately 20 months. Several secondary analyses used a later final-analysis database cutoff of 26-Oct-2017, corresponding to up to approximately 34 months.

4. Endpoints

The registry lists six primary endpoints. Four concern all participants or participants with PD-L1-positive tumors, and two concern participants with strongly PD-L1-positive tumors. The primary endpoint family therefore combines time-to-event outcomes with biomarker-defined analysis populations.

Primary endpointTime frameEndpoint type
Progression-Free Survival (PFS) Per RECIST 1.1 - All Participants Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event
Overall Survival (OS) - All Participants Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event
PFS Per RECIST 1.1 - Participants With PD-L1 Positive Tumors Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event
OS - Participants With PD-L1 Positive Tumors Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event
PFS Per RECIST 1.1 - Participants With Strongly PD-L1 Positive Tumors Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event
OS - Participants With Strongly PD-L1 Positive Tumors Through primary analysis database cut-off date of 07-Sep-2016 (Up to approximately 20 months) Time-to-event

Progression-Free Survival

PFS was defined as the time from randomization to the first documented disease progression, or death due to any cause, whichever occurred first. Per RECIST 1.1, progressive disease was defined as at least a 20% increase in the sum of diameters of target lesions. In addition to the relative increase of 20%, the sum must also demonstrate an absolute increase of at least 5 mm.

Overall Survival

OS was defined as the time from randomization to death due to any cause. The all-participant OS endpoint was assessed through the primary analysis database cutoff date of 07-Sep-2016.

PD-L1-defined populations

For the purposes of the study, participants with PD-L1 CPS ≥1% were considered to have PD-L1-positive tumor status. Participants with PD-L1 CPS ≥10% were considered to have strongly PD-L1-positive tumor status. These definitions apply to the corresponding primary OS and PFS analyses.

5. Statistical Methodology

Analysis population

The primary analysis population consisted of all randomized participants, regardless of whether or not they received study treatment. Participants were included in the treatment group to which they were randomized. Corresponding PD-L1 analyses used all randomized participants within the applicable PD-L1-defined population.

Stratified Cox proportional-hazards model

The registry reports Cox regression as the statistical method for all six primary endpoint analyses. The statistical method is a Cox proportional-hazards model. The analysis notes specify treatment as a covariate, stratified by ECOG Performance Status, presence or absence of liver metastases, hemoglobin, and time from completion of the most recent chemotherapy.

Primary time-to-event model
h(t|X) = h0(t) exp(βX)

The exponentiated treatment coefficient, exp(β), is interpreted as a hazard ratio under the Cox model. The model compares the estimated instantaneous event rate between treatment groups while incorporating the prespecified stratification structure.

Risk difference for response

The secondary objective response analyses used the Miettinen & Nurminen method. This is a score-based confidence-interval approach for differences in proportions, related to the Newcombe and Wilson methods. The effect measure was a difference in percentages, that is, a risk difference.

Risk-difference interpretation
Risk difference = P(response | pembrolizumab) − P(response | control)

A positive risk difference means the observed response proportion was higher in the pembrolizumab group. Unlike a hazard ratio, a risk difference is expressed directly in percentage points.

Covariate adjustment and stratification

The primary and secondary analyses identify covariate adjustment and stratified analysis as additional concepts. The Cox analyses explicitly state that treatment was used as a covariate with stratification by ECOG Performance Status, liver metastasis status, hemoglobin, and time from completion of the most recent chemotherapy.

6. Primary Results: All Participants

Progression-Free Survival

PFS hazard ratio

0.98

95% CI: 0.81–1.19   ·   P = 0.41648

Primary analysis cutoff: 07-Sep-2016; up to approximately 20 months

EndpointEffect estimate95% CIP-value
PFS — all participantsHR 0.980.81–1.190.41648
Clinical Biostats interpretation

The estimated hazard ratio of 0.98 is very close to 1. Under the fitted Cox model, this corresponds to an estimated instantaneous hazard of progression or death that is approximately 2% lower in the pembrolizumab group than in the control group.

The estimate does not mean that 2% fewer participants progressed or died, and it is not an absolute risk difference. It is a relative time-to-event measure from the Cox model.

The two-sided 95% confidence interval of 0.81–1.19 spans 1. This indicates substantial uncertainty around the point estimate and includes values compatible with either a lower or higher estimated hazard under the model. The p-value of 0.41648 is a measure of evidence against the specified null hypothesis; it is not a measure of the magnitude of the treatment effect.

Because this is a Cox-model result, interpretation also depends on the proportional-hazards framework and on censoring and follow-up patterns. The ClinicalTrials.gov record does not provide median PFS, event counts, or Kaplan-Meier estimates, so those quantities are not inferred here.

Overall Survival

OS hazard ratio

0.73

95% CI: 0.59–0.91   ·   P = 0.00224

Primary analysis cutoff: 07-Sep-2016; up to approximately 20 months

EndpointEffect estimate95% CIP-value
OS — all participantsHR 0.730.59–0.910.00224
Clinical Biostats interpretation

The OS hazard ratio of 0.73 means that, under the fitted Cox model and over the analyzed follow-up, the estimated instantaneous hazard of death was approximately 27% lower in the pembrolizumab group than in the control group.

This does not mean that 27% of participants benefited, that 27% of participants were prevented from dying, or that each individual participant experienced a 27% reduction in risk. The hazard ratio is a model-based relative measure of the event rate over time.

The 95% confidence interval of 0.59–0.91 quantifies uncertainty around the estimated hazard ratio. Because the entire interval is below 1, the interval is consistent with a lower estimated hazard under the fitted model. It does not describe the range of individual treatment effects.

The p-value of 0.00224 quantifies statistical evidence under the specified testing framework; it does not measure the size or clinical importance of the effect. The analysis was a superiority analysis, and the Cox model incorporated the registry-specified stratification structure.

7. Primary Results: PD-L1-Positive Tumors

The registry prespecified PFS and OS analyses among randomized participants with PD-L1-positive tumors, defined in the study as PD-L1 CPS ≥1%.

Primary endpointHazard ratio95% CIP-value
PFS — PD-L1-positive tumors0.910.68–1.240.26443
OS — PD-L1-positive tumors0.610.43–0.860.00239

PD-L1-Positive PFS

PFS hazard ratio

0.91

95% CI: 0.68–1.24   ·   P = 0.26443

Clinical Biostats interpretation

The estimated PFS hazard ratio of 0.91 corresponds to an estimated 9% lower instantaneous hazard of progression or death in the pembrolizumab group under the fitted model.

The 95% CI of 0.68–1.24 is relatively broad and crosses 1. Thus, the point estimate alone should not be treated as evidence of a definitive treatment effect in this population. The p-value of 0.26443 provides evidence against the superiority null hypothesis according to the reported analysis framework, but it does not quantify effect size or establish that the two treatment effects are identical.

As with the all-participant PFS analysis, the hazard ratio is not a median survival difference or a percentage of participants who remained progression-free.

PD-L1-Positive OS

OS hazard ratio

0.61

95% CI: 0.43–0.86   ·   P = 0.00239

Clinical Biostats interpretation

The OS hazard ratio of 0.61 corresponds to an estimated instantaneous hazard of death approximately 39% lower in the pembrolizumab group under the fitted Cox model.

The 95% CI of 0.43–0.86 remains below 1, indicating that the uncertainty interval for the model-based estimate is entirely on the lower-hazard side of the null value. The p-value of 0.00239 is evidence against the superiority null within the reported analysis; it is not a measure of the clinical magnitude of benefit.

The population is restricted to PD-L1-positive participants, so this estimate should not be silently substituted for the all-participant OS estimate of 0.73. These are different analysis populations and answer related but distinct statistical questions.

8. Primary Results: Strongly PD-L1-Positive Tumors

The strongly PD-L1-positive population was defined by PD-L1 CPS ≥10%. The registry reports separate PFS and OS primary analyses for this population.

Primary endpointHazard ratio95% CIP-value
PFS — strongly PD-L1-positive tumors0.890.61–1.280.23958
OS — strongly PD-L1-positive tumors0.570.37–0.880.00483

Strongly PD-L1-Positive PFS

PFS hazard ratio

0.89

95% CI: 0.61–1.28   ·   P = 0.23958

Clinical Biostats interpretation

The estimated hazard ratio of 0.89 corresponds to an approximately 11% lower estimated instantaneous hazard of progression or death in the pembrolizumab group under the fitted model.

The 95% CI of 0.61–1.28 crosses 1, so the interval is compatible with both a lower and a higher hazard relative to control. The p-value of 0.23958 does not measure the size of the observed HR and should not be interpreted as the probability that the treatment has no effect.

Strongly PD-L1-Positive OS

OS hazard ratio

0.57

95% CI: 0.37–0.88   ·   P = 0.00483

Clinical Biostats interpretation

The OS hazard ratio of 0.57 corresponds to an approximately 43% lower estimated instantaneous hazard of death in the pembrolizumab group under the fitted model.

The 95% CI of 0.37–0.88 is entirely below 1, while the p-value of 0.00483 provides statistical evidence against the superiority null under the reported analysis framework.

The estimate should nevertheless be interpreted as a model-based relative treatment effect in the strongly PD-L1-positive population, not as an absolute survival probability or a guarantee of benefit for an individual participant.

9. Primary Endpoint Results Side by Side

PopulationPFS HRPFS 95% CIPFS P-valueOS HROS 95% CIOS P-value
All participants 0.980.81–1.190.41648 0.730.59–0.910.00224
PD-L1-positive, CPS ≥1% 0.910.68–1.240.26443 0.610.43–0.860.00239
Strongly PD-L1-positive, CPS ≥10% 0.890.61–1.280.23958 0.570.37–0.880.00483

This table illustrates why endpoint definition and analysis population matter. The six primary analyses are not six measurements of exactly the same quantity. They represent two different time-to-event outcomes evaluated in three nested or differently restricted populations.

Important statistical distinction: the registry data provide separate PFS and OS analyses for increasingly restricted PD-L1-defined populations. Differences between hazard ratios across these populations should not automatically be described as evidence that PD-L1 modifies treatment effect. A formal interaction analysis is required to test effect modification, and the ClinicalTrials.gov record does not provide such an interaction result.

10. Secondary Endpoint Results

The registry contains nine secondary statistical analyses. These include objective response rate under RECIST 1.1 and modified RECIST, as well as modified-RECIST PFS. The response analyses use risk differences, while the PFS analyses use Cox hazard ratios.

Objective Response Rate — RECIST 1.1

PopulationRisk difference95% CIP-valueAnalysis cutoff
All participants10.03.9–16.20.0006826-Oct-2017
PD-L1-positive tumors15.66.5–25.70.0004926-Oct-2017
Strongly PD-L1-positive tumors17.26.8–29.40.0006126-Oct-2017

The registry reports these effects as differences in percentages. Thus, the all-participant estimate of 10.0 is a 10.0-percentage-point difference in objective response rate under the specified analysis, rather than a 10.0% relative increase.

Clinical Biostats interpretation

The 95% confidence interval for the all-participant RECIST 1.1 risk difference is 3.9–16.2. The interval therefore quantifies uncertainty around a positive percentage-point difference, rather than uncertainty around a hazard ratio.

The p-value of 0.00068 addresses statistical evidence under the reported superiority analysis. It does not tell us that the response difference is clinically important, nor does it establish how long responses lasted.

The PD-L1-positive and strongly PD-L1-positive estimates are 15.6 and 17.2, respectively, but their larger numerical values should not by themselves be interpreted as proof of effect modification. Formal comparison of treatment effects across populations requires an appropriate interaction analysis.

Progression-Free Survival — Modified RECIST

PopulationHazard ratio95% CIP-valueAnalysis cutoff
All participants0.860.71–1.040.0532826-Oct-2017
PD-L1-positive tumors0.820.60–1.100.0874526-Oct-2017
Strongly PD-L1-positive tumors0.770.53–1.110.0706626-Oct-2017
Clinical Biostats interpretation

The modified-RECIST PFS estimates are all below 1: 0.86 for all participants, 0.82 for PD-L1-positive participants, and 0.77 for strongly PD-L1-positive participants.

However, each corresponding 95% confidence interval includes 1. The all-participant interval is 0.71–1.04, the PD-L1-positive interval is 0.60–1.10, and the strongly PD-L1-positive interval is 0.53–1.11. These intervals communicate more uncertainty than the point estimates alone.

The associated p-values—0.05328, 0.08745, and 0.07066—should not be converted into claims about effect size. In particular, a p-value just above or below a conventional threshold is not a scientifically meaningful discontinuity in the underlying evidence.

Objective Response Rate — Modified RECIST

PopulationRisk difference95% CIP-valueAnalysis cutoff
All participants13.87.4–20.30.0000126-Oct-2017
PD-L1-positive tumors21.011.1–31.50.0000226-Oct-2017
Strongly PD-L1-positive tumors21.510.1–34.20.0000926-Oct-2017
Clinical Biostats interpretation

The modified-RECIST response analyses estimate positive differences in response percentages in all three populations. For all participants, the reported risk difference is 13.8 with a 95% CI of 7.4–20.3.

The PD-L1-positive and strongly PD-L1-positive analyses report risk differences of 21.0 and 21.5, respectively. Their confidence intervals are 11.1–31.5 and 10.1–34.2.

These are percentage-point contrasts, not hazard ratios. They also arise from a different response assessment framework than the primary RECIST 1.1 time-to-event endpoints, so the estimates should be kept conceptually separate.

11. Safety: Serious Adverse Events

The ClinicalTrials.gov record reports serious adverse events by treatment group as affected participants over participants at risk.

GroupSerious adverse events affected / at risk
Control104/255
Pembrolizumab107/266
Control switched over to pembrolizumab8/13

The ClinicalTrials.gov record does not provide a formal statistical comparison, confidence interval, or p-value for these serious-adverse-event counts. Accordingly, the counts are presented descriptively rather than converted into an unsupported inferential comparison.

Safety interpretation: the switched-over group is a distinct group reported as 8/13. It should not be merged with the randomized pembrolizumab arm or treated as though it were a randomized comparison group. The presence of treatment switching also matters when interpreting randomized-group efficacy analyses, particularly OS.

12. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

PFS and OS are time-to-event endpoints. Some participants can remain alive or progression-free at the analysis cutoff, creating right-censored observations. A Cox model can use the timing of observed events while appropriately retaining information from participants whose event has not yet occurred at their last assessment.

What does an OS hazard ratio of 0.73 mean?

A hazard ratio of 0.73 means that the estimated instantaneous hazard of death in the pembrolizumab group was approximately 27% lower than in the control group under the fitted model. It does not mean that 27% of participants survived, that survival probability increased by 27 percentage points, or that every individual experienced the same reduction.

Why does the confidence interval matter?

A point estimate is only one estimate from the observed trial data. The 95% confidence interval communicates the statistical uncertainty surrounding that estimate under the analysis framework. For example, the all-participant OS HR of 0.73 has a 95% CI of 0.59–0.91, while the all-participant PFS HR of 0.98 has a 95% CI of 0.81–1.19.

Why are some PFS confidence intervals compatible with both directions?

The PFS estimates for all participants, PD-L1-positive participants, and strongly PD-L1-positive participants have confidence intervals of 0.81–1.19, 0.68–1.24, and 0.61–1.28, respectively. Because each interval contains 1, the data represented by those intervals are compatible with both lower and higher hazards under the fitted model.

What is different about the response analyses?

Objective response is a binary endpoint rather than a time-to-event endpoint. The registry reports Miettinen & Nurminen methods and a difference in percentages as the effect measure. Therefore, the result is expressed in percentage points rather than as a hazard ratio.

Why are PD-L1 subgroup estimates not automatically interaction tests?

Comparing 0.73 in all participants with 0.61 or 0.57 in PD-L1-defined populations is descriptive. A difference between numerical estimates does not establish that the treatment effect changes according to PD-L1 status. A formal interaction test is needed to evaluate effect modification, and no such result is included in the ClinicalTrials.gov record.

Why does randomization matter?

Randomization assigns participants to treatment groups before the outcome is observed, providing the basis for a causal comparison of the randomized strategies under the trial design. The primary analysis population preserved this assignment principle by analyzing randomized participants according to their randomized group.

13. Understanding the Cox Model More Deeply

The Cox proportional-hazards model separates the baseline hazard from the relative treatment effect. Conceptually, the treatment coefficient changes the hazard multiplicatively:

Model interpretation
HR = exp(βtreatment)

An HR below 1 corresponds to a lower modeled instantaneous event rate in the treatment group; an HR above 1 corresponds to a higher modeled instantaneous event rate.

For KEYNOTE-045, the primary analyses incorporated treatment as a covariate and stratified by ECOG Performance Status, presence or absence of liver metastases, hemoglobin, and time from completion of the most recent chemotherapy. Stratification allows the baseline hazard to differ across the specified strata while estimating the treatment contrast within the Cox framework.

This is an important distinction from simply comparing two crude event proportions. A time-to-event model uses the timing of events and censoring information, whereas a binary endpoint analysis summarizes whether an event occurred within a defined framework.

Relative measure

HR 0.73 is a relative model-based treatment effect for the instantaneous hazard of death.

Absolute measure

A risk difference such as 10.0 is a percentage-point contrast in response proportions.

Time-to-event measure

PFS and OS account for when progression, death, or censoring occurs.

Binary measure

ORR classifies participants according to whether they meet the response definition.

14. Censoring and the Proportional-Hazards Assumption

Because PFS and OS are time-to-event endpoints, participants who have not experienced the endpoint by the analysis cutoff can contribute information until the point at which their follow-up ends. This is the role of censoring in survival analysis.

The Cox model additionally relies on the proportional-hazards framework for its usual interpretation of a single hazard ratio. Under proportional hazards, the relative hazard associated with treatment is treated as stable over time after accounting for the model structure. If that assumption is inappropriate, a single HR can summarize the treatment contrast incompletely.

Registry-data limitation: the ClinicalTrials.gov record provides the Cox model, hazard ratios, confidence intervals, p-values, and stratification variables, but they do not provide diagnostic results for the proportional-hazards assumption. No such diagnostic conclusion is therefore made here.

15. Primary Endpoint Interpretation

QuestionWhat the ClinicalTrials.gov record tells usWhat it does not tell us
How large is the modeled relative effect? The HR provides a relative time-to-event estimate. It does not provide an absolute risk difference.
How precise is the estimate? The 95% CI quantifies statistical uncertainty around the estimate. It does not describe individual-level treatment-effect variability.
How much statistical evidence is present? The p-value summarizes evidence under the specified hypothesis-testing framework. It does not measure clinical importance or effect size.
Is the result generalizable? The result describes the randomized trial population and analysis framework. It does not establish effects in populations not represented by the trial.

The most direct statistical contrast in the all-participant primary analyses is between PFS and OS. The PFS estimate is 0.98 with a 95% CI of 0.81–1.19, whereas the OS estimate is 0.73 with a 95% CI of 0.59–0.91. These should not be collapsed into a single efficacy statistic because they represent different endpoints with different event definitions.

Similarly, the PD-L1-positive and strongly PD-L1-positive estimates describe restricted populations. The OS estimates are 0.61 and 0.57, respectively, while the corresponding PFS estimates are 0.91 and 0.89. A descriptive difference between OS and PFS effects does not itself identify why the endpoints differ; interpretation depends on follow-up, censoring, subsequent therapy, event definitions, and the underlying disease process.

16. Risk Difference and Response Analysis

The secondary response analyses use a different statistical language from the primary survival analyses. Instead of a hazard ratio, the registry reports a difference in percentages, normalized as a risk difference.

RECIST 1.1 risk differences
All participants
10.0
PD-L1 positive
15.6
Strongly PD-L1 positive
17.2

The bars above are visual representations of the reported point estimates and are not estimates of response rates themselves. The underlying registry data report the treatment-group difference in percentages, not the two response proportions from which the difference arose.

The same distinction applies to modified RECIST response. The reported risk differences are 13.8, 21.0, and 21.5 for all participants, PD-L1-positive participants, and strongly PD-L1-positive participants, respectively.

17. Multiplicity and the Six Primary Analyses

The trial registry identifies six primary endpoints and six corresponding formal statistical analyses. Each has a reported estimate, two-sided 95% confidence interval, p-value, and superiority hypothesis.

FeatureRegistry-supported interpretation
Primary endpoint count6
Formal primary analyses6
Primary analyses with estimate + CI6
Hypothesis typeSuperiority
CI type95%, two-sided
Primary time frameThrough 07-Sep-2016, up to approximately 20 months

The ClinicalTrials.gov record establishes that all six primary endpoint analyses were formally analyzed, but they do not provide an alpha-allocation scheme, multiplicity-adjustment procedure, or hierarchy linking the six tests. Therefore, this page does not impose a multiplicity interpretation that is not contained in the ClinicalTrials.gov record.

Why this matters: when multiple hypotheses are evaluated, the interpretation of individual p-values can depend on the prespecified testing strategy. A p-value should therefore be read together with the endpoint's role, population, confidence interval, and trial-level multiplicity framework rather than in isolation.

18. Missing Data, Censoring, and Analysis Populations

The primary efficacy analysis population is explicitly described as all randomized participants, regardless of whether or not they received study treatment, with participants retained in the group to which they were randomized. This is the key analysis-population principle supported by the ClinicalTrials.gov record.

The time-to-event endpoints necessarily involve censoring because the outcome is defined as occurring over time rather than at a single fixed binary assessment. However, the ClinicalTrials.gov record does not describe a specific missing-data imputation procedure, sensitivity analysis, or censoring-rule sensitivity analysis beyond the endpoint definitions themselves.

What is reported

Randomized analysis populations, Cox models, stratification variables, estimates, confidence intervals, and p-values.

What is not reported here

No specific imputation strategy or missing-data sensitivity analysis is reported in the ClinicalTrials.gov record.

For this reason, the absence of an imputation method on this page should not be interpreted as evidence that no missing-data procedures existed in the underlying protocol or statistical analysis plan. It means only that the ClinicalTrials.gov record does not specify one.

19. Crossover and Interpretation of Randomized Comparisons

The serious-adverse-event data include a distinct group labeled Control Switched Over to Pembrolizumab: 8/13. This indicates that treatment switching occurred and that a switched-over population was reported separately.

From a statistical perspective, treatment switching creates an important distinction between the effect of assignment to a randomized strategy and the effect of receiving a particular treatment exposure. An intention-to-treat comparison preserves the randomized assignment contrast, while crossover can make the observed treatment-group trajectories more similar than they would have been without switching.

Interpretation boundary: the ClinicalTrials.gov record does not provide a crossover-adjusted OS estimate. Accordingly, no quantitative correction for crossover is applied to the reported HRs.

20. Timeline of the Trial Record

22-Oct-2014

Trial start

KEYNOTE-045 began enrollment under a randomized phase 3 design.

07-Sep-2016

Primary completion and primary analysis cutoff

The primary completion date and the primary analysis database cutoff were 07-Sep-2016. The registered primary endpoint time frame extended up to approximately 20 months.

26-Oct-2017

Final-analysis secondary endpoint cutoff

The reported secondary ORR and modified-RECIST analyses used a final analysis database cutoff of 26-Oct-2017, up to approximately 34 months.

21. Statistical Interpretation of the Primary Results

All-participant PFS

The PFS HR of 0.98 indicates a modeled relative hazard close to the null value of 1. Its 95% CI of 0.81–1.19 spans 1, and the p-value is 0.41648.

All-participant OS

The OS HR of 0.73 corresponds to an approximately 27% lower estimated instantaneous hazard of death under the Cox model. Its 95% CI of 0.59–0.91 is below 1, with a p-value of 0.00224.

PD-L1-positive OS

The OS HR of 0.61 corresponds to an approximately 39% lower estimated instantaneous hazard of death under the fitted model. The 95% CI is 0.43–0.86, with a p-value of 0.00239.

Strongly PD-L1-positive OS

The OS HR of 0.57 corresponds to an approximately 43% lower estimated instantaneous hazard of death under the fitted model. The 95% CI is 0.37–0.88, with a p-value of 0.00483.

These four results illustrate a central principle of clinical-trial statistics: the point estimate, confidence interval, and p-value answer different questions. The HR describes the estimated relative treatment effect; the confidence interval describes statistical uncertainty around that estimate; and the p-value describes evidence against the specified null hypothesis.

22. Limitations

23. Why This Trial Matters Statistically

KEYNOTE-045 is a useful statistical teaching case because the registry data combine randomized treatment comparison, multiple time-to-event endpoints, biomarker-defined analysis populations, Cox regression, stratification, response-rate analysis, and treatment switching.

ConceptHow it appears in KEYNOTE-045
RandomizationThe trial used randomized allocation in a parallel-group phase 3 design.
Intention-to-treat principlePrimary efficacy populations consisted of randomized participants analyzed according to randomized treatment group.
Time-to-event analysisPFS and OS were primary time-to-event endpoints.
Cox proportional-hazards modelUsed for all six primary analyses and the modified-RECIST PFS secondary analyses.
Hazard ratioUsed as the effect measure for PFS and OS.
Confidence intervalAll six primary analyses report two-sided 95% confidence intervals.
Stratified analysisPrimary analyses were stratified by ECOG Performance Status, liver metastasis status, hemoglobin, and time from completion of the most recent chemotherapy.
Risk differenceRECIST 1.1 and modified-RECIST ORR analyses report differences in percentages.
Score-based CIResponse analyses used the Miettinen & Nurminen method.
Biomarker-defined populationsPrimary endpoints were evaluated in all participants, PD-L1-positive participants, and strongly PD-L1-positive participants.
Treatment switchingA separate control-switched-to-pembrolizumab safety group is reported as 8/13.

24. A Practical Reading Strategy for KEYNOTE-045

A statistically disciplined reading of the trial starts with the endpoint rather than the p-value. First identify whether the endpoint is time-to-event or binary. Then identify the analysis population, the effect measure, the confidence interval, and the analysis method.

01
EndpointPFS, OS, or ORR
02
PopulationAll, PD-L1 ≥1%, or ≥10%
03
EffectHR or risk difference
04
Precision95% CI
05
EvidenceP-value in context

For example, the all-participant OS result is HR 0.73 with a 95% CI of 0.59–0.91 and p = 0.00224. The correct first interpretation is that the fitted model estimates a lower hazard of death in the pembrolizumab group, with the confidence interval lying below 1. Only after understanding that relative effect should the p-value be considered.

For all-participant RECIST 1.1 ORR, the effect is a risk difference of 10.0 with a 95% CI of 3.9–16.2 and p = 0.00068. This is a fundamentally different statistical object: it is a percentage-point contrast in a binary response endpoint, not a relative hazard over time.

25. Related Tutorials

Learn more about the methods used in this trial:

26. Related Calculators

27. Sources

Continue with the statistical methods

Explore the survival-analysis, confidence-interval, and clinical-trial methods that underlie the KEYNOTE-045 analysis.

28. Record Summary

KEYNOTE-045 provides a useful example of how a randomized phase 3 trial can generate several related but distinct statistical questions. The six primary analyses evaluate PFS and OS in all participants, PD-L1-positive participants, and strongly PD-L1-positive participants. Each primary analysis uses a Cox proportional-hazards model with treatment as a covariate and stratification by ECOG Performance Status, liver metastasis status, hemoglobin, and time from completion of the most recent chemotherapy.

The reported all-participant PFS hazard ratio was 0.98 with a 95% CI of 0.81–1.19 and p = 0.41648, while the all-participant OS hazard ratio was 0.73 with a 95% CI of 0.59–0.91 and p = 0.00224. The PD-L1-positive and strongly PD-L1-positive OS analyses reported hazard ratios of 0.61 and 0.57, respectively, with corresponding 95% confidence intervals of 0.43–0.86 and 0.37–0.88.

The secondary analyses demonstrate a second statistical framework. Objective response rate was analyzed using the Miettinen & Nurminen method, with risk difference as the effect measure, while modified-RECIST PFS was analyzed with Cox regression. Keeping these effect measures separate is essential: a hazard ratio describes a modeled relative event rate over time, whereas a risk difference describes an absolute percentage-point difference in a binary outcome.

Clinical Biostats methodology: The purpose of this analysis is to make the statistical structure of the trial explicit: define the endpoint, identify the analysis population, state the model and effect measure, report the confidence interval and p-value, and then explain what the result does and does not establish. Where the ClinicalTrials.gov record does not contain a number or statistical procedure, it is not inferred.