This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
KEYNOTE-164 was a completed phase 2, non-randomized, single-group treatment study evaluating pembrolizumab monotherapy in participants with previously treated locally advanced unresectable or metastatic colorectal cancer. The registry lists 124 participants and two cohorts, with objective response rate as the registered primary endpoint.
| Feature | KEYNOTE-164 |
|---|---|
| Phase | Phase 2 |
| Condition | Colorectal Carcinoma |
| Population | Participants with previously treated locally advanced unresectable or metastatic colorectal cancer |
| Design | Non-randomized, single-group, unmasked |
| Primary purpose | Treatment |
| Enrollment | 124 |
| Number of arms | 2 |
| Intervention | Pembrolizumab (biological) |
| Status | Completed |
| Start date | 2015-08-25 |
| Primary completion date | 2019-09-09 |
| Lead sponsor | Merck Sharp & Dohme LLC |
| Sponsor type | Industry |
| ClinicalTrials.gov | NCT02460198 |
2. Clinical Question
The registered clinical question concerns the activity of pembrolizumab monotherapy in participants with previously treated locally advanced unresectable or metastatic colorectal cancer. The primary statistical endpoint was objective response rate, defined using RECIST 1.1 and assessed by a central imaging vendor.
Population
Participants with previously treated locally advanced unresectable or metastatic colorectal cancer.
Intervention
Pembrolizumab monotherapy.
Comparator
There was no separate randomized comparator group. The design is registered as non-randomized and single-group.
Primary question
What proportion of participants in the analysis population achieved a complete response or partial response according to RECIST 1.1?
3. Trial Design
The registry describes KEYNOTE-164 as a phase 2, non-randomized, single-group study with no masking and a primary purpose of treatment. Although the registry lists two arms, the intervention field identifies pembrolizumab as the biological intervention, and the registered design is not a randomized comparative design.
First and second course
- First Course: 61 participants at risk for the reported serious-adverse-event measure.
- Second Course: 6 participants at risk for the reported serious-adverse-event measure.
- Intervention: pembrolizumab.
First and second course
- First Course: 63 participants at risk for the reported serious-adverse-event measure.
- Second Course: 3 participants at risk for the reported serious-adverse-event measure.
- Intervention: pembrolizumab.
4. Primary Endpoint
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Objective Response Rate (ORR) | Objective response rate was defined as the percentage of the participants in the analysis population who had a complete response (CR) or partial response (PR). Complete Response: disappearance of all target lesions. Partial Response: at least a 30% decrease in the sum of diameters of target lesions, taking as reference the baseline sum diameters. Responses were based upon blinded central imaging vendor per RECIST 1.1. | Up to approximately 48 months |
The registry specifies that participants without response data were counted as nonresponders. This rule is statistically important because it defines how missing response assessments contribute to the numerator and denominator of the reported response analysis.
The endpoint is therefore a binary participant-level outcome for the purpose of the primary analysis: a participant either meets the registry-defined response criterion or is counted as a nonresponder.
5. Results
The registry indicates that results have been posted, but no formal statistical analyses are posted in the ClinicalTrials.gov record. The registry specifies the primary endpoint and its planned statistical procedure, but it does not provide a posted ORR estimate, confidence interval, or p-value in the statistical-analysis information available for this record.
How the primary endpoint would normally be analysed
For a single-group objective response endpoint, the natural analysis is a one-sample binomial analysis. Each participant contributes a binary response outcome, and the observed response proportion is calculated from the number of responders divided by the analysis-population size. The Clopper-Pearson method then provides an exact confidence interval for the underlying response probability.
This approach is different from a two-arm comparison. Because the registered design is single-group and non-randomized, the primary response analysis is fundamentally an estimation of a proportion rather than an estimation of a hazard ratio, risk ratio, or between-arm treatment difference.
6. Statistical Methodology
Exact binomial distribution
The registry specifies an exact binomial distribution for the ORR point estimate and its 95% confidence interval. If the analysis population contains n participants and x participants meet the response definition, the observed ORR is x/n.
Here, X represents the number of participants with a complete or partial response, n is the analysis-population size, and p is the underlying response probability being estimated.
Clopper-Pearson confidence interval
The registry specifies the Clopper-Pearson method for the 95% confidence interval. This is an exact binomial interval rather than a normal-approximation interval. The distinction is useful for response-rate analyses because the binomial proportion is bounded between 0 and 1 and may have substantial uncertainty when the number of responses is limited.
The interval should be interpreted as statistical uncertainty around the underlying response probability under the specified binomial framework. It is not a range containing the response probability for individual patients, and it does not describe variation in tumor shrinkage among participants.
Participants without response data
The registry explicitly states that participants without response data were counted as nonresponders. This rule has a direct effect on the primary endpoint: absence of an evaluable response assessment does not remove the participant from the response denominator under the stated analysis convention.
Central imaging assessment
Responses were based upon a central imaging vendor according to RECIST 1.1. Central assessment is relevant statistically because the endpoint depends on radiologic classification. Using a specified response framework and centralized assessment helps establish a consistent rule for classifying complete response, partial response, and nonresponse.
7. Statistical Methods Explained
Why use a binomial model for ORR?
ORR reduces each participant's tumor-response assessment to a binary endpoint for the primary analysis: complete response or partial response counts as response, while the remaining participants count as nonresponders under the registry rule. A binomial model is therefore a natural probability model for the number of responses among participants.
What does an ORR estimate represent?
An ORR estimate represents the proportion of participants in the analysis population who achieved a complete or partial response according to the specified RECIST 1.1 criteria. It is a proportion, not a time-to-event measure and not a measure of how long a participant survives.
Why use an exact confidence interval?
The exact Clopper-Pearson method avoids relying on a normal approximation to the binomial distribution. This is especially useful conceptually when the response count is small or when the estimated proportion is near the boundaries of 0 or 1.
Why does counting missing response data as nonresponse matter?
It makes the treatment of unavailable response assessments explicit. Rather than silently excluding participants with missing response information, the registry's rule classifies them as nonresponders. Consequently, the reported ORR reflects that prespecified handling of missing response data.
Why is this not a hazard-ratio analysis?
ORR is a binary response endpoint observed over a specified assessment period. A hazard ratio is designed for time-to-event outcomes such as death or progression. The registry-defined primary endpoint is ORR, so the specified exact binomial framework is appropriate to the endpoint's structure.
What does a 95% confidence interval tell us?
The 95% confidence interval quantifies statistical uncertainty around the estimated response probability under the exact binomial framework. A wider interval indicates less precision, while a narrower interval indicates greater precision. It does not describe the range of individual patient responses.
8. Safety Results
The registry reports serious adverse events by cohort and course. These figures are presented as affected participants divided by participants at risk for the corresponding category.
| Cohort / course | Serious adverse events | At risk |
|---|---|---|
| Cohort A First Course | 31 | 61 |
| Cohort B First Course | 25 | 63 |
| Cohort A Second Course | 1 | 6 |
| Cohort B Second Course | 1 | 3 |
First Course
The registry reports 31 affected participants among 61 at risk in Cohort A and 25 affected participants among 63 at risk in Cohort B.
Second Course
The registry reports 1 affected participant among 6 at risk in Cohort A and 1 affected participant among 3 at risk in Cohort B.
These are serious-adverse-event counts by the registry's cohort/course categories. They should not be interpreted as a comparative treatment effect because the study is registered as non-randomized and single-group, and the reported categories do not constitute randomized comparator arms.
9. Trial Timeline
Study start
The registry lists 2015-08-25 as the study start date.
Non-randomized treatment study
The registry describes the study as phase 2, non-randomized, single-group, unmasked, with a primary purpose of treatment.
Study enrollment
The registry lists an enrollment of 124 participants and two arms.
Primary completion
The registry lists 2019-09-09 as the primary completion date.
Primary endpoint data cutoff
The registry specifies 09-SEPT-2019 as the data cutoff date for the objective response rate endpoint.
10. Interpreting Objective Response Rate
Objective response rate is the percentage of participants who achieved either a complete response or a partial response under RECIST 1.1. In this study, the response classification was based on a central imaging vendor.
ORR does not mean that the responding percentage represents the percentage of participants who were cured, the percentage who survived for a particular duration, or the percentage who experienced a specified improvement in overall survival. It is specifically a radiologic tumor-response endpoint.
The denominator is important because the registry defines ORR as the percentage of participants in the analysis population with CR or PR and specifies that participants without response data were counted as nonresponders. The resulting estimate therefore depends on both the observed responses and the prespecified handling of unavailable response data.
An ORR point estimate alone does not describe how precisely the underlying response probability has been estimated. The specified 95% Clopper-Pearson confidence interval supplies that uncertainty information. A point estimate and its confidence interval should therefore be interpreted together.
11. Statistical Interpretation vs Clinical Interpretation
Statistical interpretation
The primary endpoint is a binomial proportion. The registry specifies an exact binomial analysis and a 95% Clopper-Pearson confidence interval, with participants without response data counted as nonresponders.
Clinical interpretation
The endpoint describes the proportion of participants experiencing radiologic tumor response according to RECIST 1.1. Its clinical meaning depends on the durability and broader clinical context of those responses, which are distinct from the ORR statistic itself.
The distinction is particularly important in a single-group study. An ORR estimate can characterize observed treatment activity, but without randomized allocation it does not isolate a causal treatment effect by comparing outcomes with a contemporaneous control group.
12. Important Limitations and Interpretation Issues
- No randomized comparator: the registered allocation is non-randomized and the design model is single-group. Consequently, the primary ORR endpoint is an activity estimate rather than a randomized between-treatment comparison.
- Binary endpoint: ORR records whether a participant achieved CR or PR. It does not capture the full continuous pattern of tumor-size change.
- Response is not survival: ORR and survival outcomes address different clinical questions. A response proportion should not be interpreted as a survival probability.
- Missing response data: participants without response data were counted as nonresponders. This is part of the endpoint's specified analysis rule and directly affects the estimated proportion.
- Confidence-interval uncertainty: the exact binomial interval quantifies uncertainty around the response probability; it does not indicate how individual participants will respond.
- Central imaging classification: the endpoint depends on RECIST 1.1 classification by a central imaging vendor, so the measurement framework is integral to the statistical endpoint.
- Cohort and course structure: the reported serious-adverse-event categories distinguish first and second courses and should not be combined into a randomized comparative safety analysis.
- Registry reporting: the ClinicalTrials.gov record identifies the planned statistical method for ORR but does not provide a posted numerical statistical analysis for that endpoint.
13. Why This Trial Matters Statistically
KEYNOTE-164 is a useful teaching case because it illustrates how the statistical method should follow the structure of the endpoint and the design. The primary endpoint is a binary response outcome, while the trial itself is non-randomized and single-group. That combination leads naturally to an exact one-sample binomial framework rather than a comparative survival model.
| Concept | How it appears in KEYNOTE-164 |
|---|---|
| Single-group design | The registry identifies the design model as single-group. |
| Non-randomized allocation | Allocation is registered as non-randomized. |
| Binary efficacy endpoint | ORR classifies participants according to CR or PR versus nonresponse. |
| RECIST 1.1 | Response is defined using RECIST 1.1 criteria. |
| Central assessment | Responses were based on assessment by a central imaging vendor. |
| Exact binomial analysis | The registry specifies an exact binomial distribution for the ORR point estimate and confidence interval. |
| Clopper-Pearson interval | The registry specifies the Clopper-Pearson method for the 95% confidence interval. |
| Missing-response handling | Participants without response data were counted as nonresponders. |
| Safety by cohort/course | Serious adverse events are reported separately for Cohorts A and B and for first and second courses. |
14. A Closer Look at the Exact Binomial Framework
Suppose an analysis contains x responders among n participants. The observed response rate is simply the ratio of those two quantities. The binomial model then treats the number of observed responses as arising from repeated binary outcomes with an underlying response probability p.
The point estimate is the observed proportion of participants meeting the CR or PR response definition.
The Clopper-Pearson interval is constructed from the exact binomial distribution rather than from a large-sample normal approximation. This produces an interval that respects the natural 0-to-1 boundaries of a probability.
The practical lesson is that the analysis does not require a complex regression model simply because the underlying clinical study is complex. Statistical complexity should be driven by the endpoint and design. For a binary single-group response endpoint, a carefully specified binomial analysis can be the appropriate primary method.
15. What the Registry Does and Does Not Establish
Established by the record
The registry identifies the phase, disease, population, intervention, single-group design, enrollment, primary endpoint, RECIST 1.1 response definitions, central imaging assessment, exact binomial method, Clopper-Pearson confidence interval, missing-response rule, and serious-adverse-event counts by cohort and course.
Not numerically reported in the statistical analyses
The posted statistical analyses do not provide a numerical ORR estimate, 95% confidence interval, or p-value for the primary endpoint.
This distinction matters because statistical interpretation should remain tied to the quantities actually reported. Knowing the analysis method explains how the endpoint was intended to be summarized; it does not substitute for the numerical estimate itself.
16. Sources
- ClinicalTrials.gov: NCT02460198.
- PubMed: PubMed record.
- PubMed: PubMed record.
- PubMed: PubMed record.
Continue through the Clinical Biostats knowledge graph
Explore additional Clinical Biostats resources for understanding clinical-trial endpoints, statistical models, and interpretation.
17. Record Summary
KEYNOTE-164 is a phase 2, non-randomized, single-group treatment study of pembrolizumab in participants with previously treated locally advanced unresectable or metastatic colorectal cancer. Its registered primary endpoint is objective response rate according to RECIST 1.1, assessed by a central imaging vendor over a time frame of up to approximately 48 months. The registry specifies an exact binomial distribution and the Clopper-Pearson method for the ORR point estimate and 95% confidence interval, with participants without response data counted as nonresponders.
From a statistical perspective, the most important feature of KEYNOTE-164 is the alignment between its design and its primary endpoint: a single-group binary response outcome is naturally summarized with a one-sample binomial framework. The absence of a randomized comparator means that an ORR estimate describes observed response activity within the study population rather than a randomized relative treatment effect. The serious-adverse-event data are likewise appropriately read within their cohort and course categories rather than as evidence from a conventional two-arm safety comparison.