KEYNOTE-756: Complete Statistical Analysis of Pembrolizumab in Early-Stage ER+/HER2- Breast Cancer

1. Trial at a Glance

KEYNOTE-756 is a randomized, parallel-group, triple-masked, phase 3 treatment trial evaluating pembrolizumab versus placebo in combination with protocol-specified neoadjuvant and adjuvant anticancer therapies for early-stage estrogen receptor-positive, human epidermal growth factor receptor 2-negative breast cancer.

2. Clinical Question

The central statistical question is whether adding pembrolizumab to the protocol-specified neoadjuvant and adjuvant anticancer therapies changes the two registered primary outcomes compared with placebo in patients with early-stage estrogen receptor-positive, human epidermal growth factor receptor 2-negative breast cancer.

Statistically, this is a useful two-endpoint framework because the outcomes capture different stages of the disease course. The pathological complete response endpoint evaluates the pathological status at surgery after neoadjuvant systemic therapy, whereas event-free survival follows participants from randomization for a substantially longer period.

3. Trial Design

The intervention list in the registry includes pembrolizumab, placebo, paclitaxel, doxorubicin, epirubicin, cyclophosphamide, endocrine therapy, radiation therapy, and surgery. The statistical interpretation therefore needs to distinguish the randomized comparison of pembrolizumab versus placebo from the background treatment components that are part of the protocol-specified treatment strategy.

4. Endpoints

Two endpoints, two statistical structures

The two primary endpoints require different statistical representations. pCR is a binary outcome assessed at a defined clinical milestone: each participant either meets the registry definition of ypT0/Tis ypN0 or does not. EFS is a time-to-event outcome, so both the occurrence of an event and the time from randomization to that event are relevant.

This distinction is important. A binary response rate compresses follow-up into one status at a specified point, whereas a time-to-event endpoint retains information about when events occur and about participants whose event status is not yet observed at the time of analysis.

5. Statistical Methodology

Analysis of pathological complete response

The registry defines pCR as the percentage of participants meeting the ypT0/Tis ypN0 definition at surgery. Because this is a binary endpoint, the fundamental analysis is a comparison of proportions between the pembrolizumab and placebo treatment strategies.

An absolute risk difference is especially intuitive for a binary endpoint because it directly describes the difference in the proportion achieving pCR between randomized groups. A relative measure provides a different perspective. These quantities should not be treated as interchangeable: a given relative effect can correspond to different absolute differences depending on the underlying control-group rate.

Analysis of event-free survival

EFS is defined as time from randomization to the first qualifying event. The registry definition combines several clinically distinct events into a single time-to-first-event endpoint: disease progression that precludes surgery, local or distant recurrence, second primary malignancy, or death from any cause.

For an endpoint of this type, a standard statistical framework would use Kaplan-Meier estimation to describe the event-free survival distribution and a time-to-event comparison such as a log-rank test and Cox proportional-hazards model to compare randomized treatment groups. These are typical methods for an endpoint whose defining feature is time from randomization to the first event.

The ClinicalTrials.gov record does not post an EFS hazard ratio, confidence interval, p-value, Kaplan-Meier estimate, median EFS, or formal statistical analysis. The appropriate interpretation at this stage is therefore methodological rather than numerical.

Why the time origin matters

The EFS definition explicitly starts the clock at randomization. This is statistically important because all randomized participants enter the time-to-event framework at the same trial-defined origin. The resulting estimate represents the experience from randomization through the first qualifying event rather than survival measured from surgery or another later milestone.

6. Statistical Methods Explained

Why is pCR treated as a binary endpoint?

The registry definition classifies participants according to whether they have no residual invasive cancer meeting the specified ypT0/Tis ypN0 definition at surgery. That produces a binary outcome. The natural summary is therefore a proportion rather than a median time or a hazard ratio.

The important statistical question is not simply whether a pCR rate can be calculated. It is whether the difference between the randomized groups is sufficiently precise to support an inference about the treatment comparison. Confidence intervals are therefore important alongside the observed proportions.

What does ypT0/Tis ypN0 mean statistically?

The registry's pCR definition combines the pathological status of the resected breast specimen with the status of all sampled regional lymph nodes. The endpoint is therefore a prespecified composite pathological classification rather than a measurement of tumor size alone.

This matters because the analysis population is determined by a clearly defined pathological criterion. Changing the pathological definition would change the endpoint itself and could make results from different analyses difficult to compare directly.

Why is EFS a time-to-event endpoint?

Two participants can both remain event-free at a particular analysis date but have been followed for very different lengths of time. Conversely, two participants can both experience an event but do so at different times. EFS preserves this timing information.

Time-to-event methods also accommodate censoring. A participant who has not experienced an EFS event by the last evaluable follow-up is not treated as though the participant were known to remain event-free indefinitely. Instead, the participant contributes information through the time at which follow-up ends under the analysis rules.

What would a hazard ratio mean for EFS?

If a Cox model were used, an EFS hazard ratio below 1 would represent a lower estimated instantaneous rate of the defined EFS event in the pembrolizumab group relative to the placebo group, under the model. It would not mean that the same percentage of participants avoided events, nor would it directly provide an absolute difference in event-free survival at a particular time point.

A hazard ratio also depends on the statistical model and the behavior of hazards over time. It should therefore be interpreted together with absolute event-free survival estimates and the confidence interval rather than treated as a complete description of the treatment effect.

Why is censoring important?

Censoring occurs when the analysis reaches the end of available follow-up for a participant without observing the defined event. The statistical method uses the participant's observed follow-up while accounting for the fact that the exact future event time is unknown.

The validity of survival estimates depends on assumptions about the censoring mechanism and how censoring is handled. Censoring should not be interpreted as equivalent to remaining permanently event-free.

Why does randomization matter for interpretation?

Randomization assigns participants to treatment groups according to the trial design rather than according to post-randomization outcomes. This creates the foundation for a causal comparison between the randomized strategies, subject to the trial's design, conduct, endpoint definitions, missing-data handling, and analysis assumptions.

7. Planned Analysis

No formal statistical analyses are posted to ClinicalTrials.gov for the primary endpoints in the registry information considered here. The registry does, however, clearly define what will be measured and the time frames for assessment.

What a mature pCR analysis would show

A complete pCR analysis would ordinarily present the number and percentage of participants meeting the ypT0/Tis ypN0 definition in each randomized group. The presentation should make the analysis denominator explicit because the interpretation of a percentage depends on which participants are included in the analysis population.

For an educational statistical review, the most informative effect measures would generally include the absolute difference between pCR rates and an uncertainty interval around that difference, accompanied by the underlying counts. A relative effect could supplement the absolute difference but should not replace it.

What a mature EFS analysis would show

A complete EFS analysis would ordinarily include the number of events, the number of participants remaining under observation, Kaplan-Meier estimates at clinically relevant time points, a median EFS when estimable, and a treatment-effect estimate such as a hazard ratio with a confidence interval.

The event definition is particularly important here. Because progression precluding surgery, local or distant recurrence, second primary malignancy, and death are all incorporated into EFS, the endpoint reflects the first occurrence of any qualifying event rather than one single biological process.

Why the pCR and EFS analyses should not be conflated

pCR is assessed at surgery and describes pathological response after neoadjuvant treatment. EFS begins at randomization and continues for up to approximately 12 years in the registry's stated time frame. A participant's pCR status and EFS experience therefore represent different pieces of information.

A higher pCR rate would describe a difference in pathological response at surgery. An EFS analysis would address whether the time to the first defined clinical event differs between randomized groups. The two endpoints can be related scientifically, but they are not mathematically interchangeable.

8. Interpreting the Two Primary Endpoints

Why an early endpoint and a long-term endpoint can coexist

A randomized trial can deliberately measure treatment effects at different stages of the disease course. pCR provides an early assessment around surgery, while EFS follows participants substantially longer. From a statistical perspective, this creates an opportunity to examine whether an intervention's effect is evident at an early pathological milestone and whether differences persist into clinically defined events over time.

These endpoints also have different sources of statistical variation. pCR depends on pathological classification at a defined surgical assessment. EFS depends on event ascertainment, follow-up duration, censoring, and the exact timing and classification of qualifying events.

9. Analysis Populations and the Role of Randomization

The registry information identifies the trial as randomized but does not provide detailed definitions of analysis populations such as intention-to-treat, per-protocol, or as-treated populations. Those distinctions are important in a complete statistical analysis, particularly when treatment exposure, surgery, follow-up, or endpoint assessment differs among participants.

These concepts should not be treated as interchangeable. A pCR analysis requires information about pathological response at surgery, whereas EFS requires longitudinal follow-up beginning at randomization. The statistical interpretation of each endpoint depends on the population actually included and the rules used to handle participants whose endpoint information is incomplete.

10. Missing Data and Censoring Considerations

The registry definitions identify what constitutes pCR and EFS but do not provide detailed missing-data or imputation rules in the information considered here. Those rules can materially affect the interpretation of an endpoint and therefore belong in a complete statistical analysis plan.

pCR and missing pathological assessment

For pCR, the endpoint depends on a pathological assessment of the complete resected breast specimen and all sampled regional lymph nodes. If the required pathological information is unavailable, the treatment effect can depend on how such participants are handled.

For a binary endpoint, simple exclusion of participants with missing outcomes changes the denominator. Conversely, prespecified conservative or multiple-imputation approaches can incorporate uncertainty about missing outcomes. The appropriate method depends on the protocol and statistical analysis plan rather than on the observed treatment results.

EFS and censoring

EFS naturally involves censoring because participants can reach the analysis cutoff without having experienced a qualifying event. This is fundamentally different from missingness in a binary response endpoint.

For survival analysis, the timing of the last known event-free observation matters. Treating all participants without an observed event as if they had identical follow-up would discard information that the Kaplan-Meier framework is designed to preserve.

11. Potential Interpretation Issues for EFS

Composite endpoint structure

EFS combines several event types. The first qualifying event determines the EFS time. Consequently, a treatment effect on EFS is an effect on the composite endpoint as defined, not automatically an identical effect on each component.

This is a general statistical principle for composite time-to-event outcomes: the overall endpoint can be driven by the frequency and timing of its individual components. A complete component analysis would therefore help determine which event types contribute most to an observed EFS difference.

Competing event types

The registry definition includes recurrence, progression precluding surgery, second primary malignancy, and death. These outcomes have different clinical meanings. When they are combined into a first-event endpoint, the resulting EFS measure prioritizes the occurrence of any qualifying event rather than assigning equal clinical importance to the components.

This does not make the composite endpoint invalid. It means that the statistical interpretation should remain faithful to the precise endpoint definition rather than translating EFS automatically into a single component such as recurrence-free survival or overall survival.

Hazard ratios and proportional hazards

If a Cox proportional-hazards model is used for EFS, the hazard ratio summarizes a relative event rate under the model. The usual interpretation is simplest when the relative hazards are reasonably stable over time. If hazards change materially over time, a single hazard ratio can conceal clinically important temporal patterns.

12. Multiplicity and the Two Primary Endpoints

KEYNOTE-756 has two registered primary endpoints: pCR and EFS. Whenever a randomized trial has more than one confirmatory endpoint, the statistical interpretation depends on how the testing strategy handles the possibility of multiple statistical tests.

The ClinicalTrials.gov record identifies both endpoints but does not provide an alpha-allocation scheme, hierarchical testing strategy, multiplicity-adjustment procedure, or interim-analysis boundary. Those details should therefore not be assumed.

For this reason, a future results analysis should distinguish the numerical p-value for an endpoint from the broader question of whether that p-value satisfies the trial's prespecified confirmatory testing procedure.

13. Interim Analysis and Information Timing

The registry information identifies the pCR assessment time frame as up to approximately 7 months and the EFS time frame as up to approximately 12 years. These are endpoint time frames, not evidence that an interim efficacy analysis occurred at those exact times.

An interim analysis is a separate statistical design feature in which accumulating trial information is formally examined before the planned final analysis. Such analyses can require alpha spending, group-sequential boundaries, or other error-control procedures. The registry information considered here does not report such a procedure.

14. Masking and Bias Control

The registry classifies KEYNOTE-756 as triple masked. Masking is a design feature intended to reduce the influence of knowledge of treatment assignment on aspects of trial conduct and assessment.

For the pCR endpoint, the registry specifies assessment by the local pathologist at the time of surgery. For EFS, the registry states that the endpoint is determined by the investigator. These assessment structures are therefore relevant when considering how outcome classification could be influenced by knowledge of treatment assignment.

Masking does not eliminate all possible sources of bias, but it is an important part of the design because outcome classification can be a component of the causal chain between randomized treatment assignment and the recorded endpoint.

15. What the pCR Rate Does — and Does Not — Mean

16. What an EFS Hazard Ratio Would — and Would Not — Mean

17. Clinical Interpretation vs Statistical Interpretation

Keeping these perspectives separate is important. Statistical significance, effect size, clinical importance, and long-term patient outcomes are related but distinct concepts. A complete interpretation should therefore report the magnitude and precision of the observed effect while also describing exactly what outcome was measured.

18. Important Limitations and Interpretation Issues

• Registry status: the trial is listed as active, not recruiting, with a primary completion date of 2031-01-24.
• No posted statistical analyses: the registry information considered here contains the endpoint definitions and time frames but no posted estimates, confidence intervals, p-values, event counts, or formal statistical analysis results.
• Different endpoint structures: pCR is a binary pathological endpoint assessed at surgery, whereas EFS is a longitudinal time-to-event endpoint beginning at randomization.
• Composite EFS endpoint: EFS combines several event types, including progression precluding surgery, local or distant recurrence, second primary malignancy, and death from any cause.
• Analysis assumptions: future EFS modeling would depend on assumptions concerning censoring and, if a Cox model were used, the adequacy of the proportional-hazards representation.
• Missing pathological information: a complete pCR analysis requires clearly defined handling of participants without the pathological information necessary to classify ypT0/Tis ypN0.
• Multiplicity: two registered primary endpoints make the prespecified type I error strategy important when interpreting future confirmatory p-values.
• Assessment framework: pCR is assessed by the local pathologist, while EFS is determined by the investigator according to the registry definition.

19. Why This Trial Matters Statistically

KEYNOTE-756 is a useful teaching case because it places an early pathological endpoint and a long-term clinical time-to-event endpoint within the same randomized phase 3 framework.

20. From Pathological Response to Time-to-Event Evidence

One of the most important statistical ideas illustrated by this trial is that an early surrogate-like measure and a later clinical outcome should be interpreted on their own definitions.

pCR is assessed at the time of surgery and captures the pathological state following neoadjuvant systemic therapy. Its scale is straightforward: participants either satisfy the ypT0/Tis ypN0 definition or they do not. The resulting percentage is therefore an early snapshot of treatment response.

EFS begins at randomization and incorporates events occurring after treatment initiation. Its scale is temporal rather than binary at one fixed point. A participant can remain event-free for a short interval, a long interval, or through the end of available follow-up without an event.

These distinctions affect both study interpretation and statistical modeling. A difference in pCR rates cannot be translated directly into an EFS hazard ratio. Likewise, an EFS hazard ratio cannot be interpreted as though it were a difference in pathological response rates.

21. Reading Future Results From This Trial

When formal results are posted, a statistically complete interpretation should begin with the analysis population and endpoint definition before moving to the numerical treatment effect.

This sequence prevents a common statistical mistake: treating a single numerical measure as though it completely describes the trial. The endpoint definition, analysis population, follow-up, effect estimate, uncertainty, and multiplicity framework all contribute to the interpretation.

22. Record Summary

KEYNOTE-756 is a randomized phase 3, parallel-group, triple-masked treatment trial with 1240 planned participants and 2 treatment arms. The registered primary endpoints are pathological complete response using the definition of ypT0/Tis ypN0, assessed at surgery after neoadjuvant systemic therapy, and event-free survival, measured from randomization for up to approximately 12 years.

Statistically, the trial provides a clear example of two fundamentally different endpoint structures. pCR is a binary pathological outcome that is naturally summarized with proportions and absolute or relative measures of treatment effect. EFS is a time-to-event outcome for which Kaplan-Meier methods, time-to-event comparisons, and Cox modeling are standard analytical frameworks.

The distinction between these endpoints is central to interpreting future results. A pCR comparison describes the frequency of a defined pathological state at surgery. An EFS comparison describes the timing of a composite clinical event beginning at randomization. Neither measure should be substituted for the other.

23. Sources

• ClinicalTrials.gov: NCT03725059 — KEYNOTE-756.
• PubMed: PubMed record for PMID 39838117.
• PubMed: PubMed record for PMID 38913933.