KEYNOTE-158: Complete Statistical Analysis of Pembrolizumab in Advanced Solid Tumors

1. Trial at a Glance

KEYNOTE-158 is a phase 2, non-randomized, parallel-design study evaluating pembrolizumab in participants with advanced solid tumors and a range of specified cancer conditions. The registry lists an enrollment of 1,609 participants, two study arms, no masking, and a primary endpoint of objective response rate (ORR).

2. Clinical Question

The central statistical question is how frequently participants in this broad advanced-solid-tumor population experience an objective response to pembrolizumab according to the registry-defined response framework.

3. Trial Design

The design has an important statistical consequence: because allocation is non-randomized, the primary ORR is naturally interpreted as a response proportion within the study population rather than as a randomized estimate of a treatment-versus-control causal contrast. The registry also lists pembrolizumab as the intervention in both arms, so the available information does not define a conventional treatment-control comparison.

4. Population and Disease Scope

The registry describes a heterogeneous advanced-cancer population rather than a single disease entity. The listed conditions include advanced cancer and advanced solid tumors together with several specific tumor types.

This breadth matters statistically because an overall response proportion combines participants with different diseases and potentially different underlying response probabilities. A single pooled ORR therefore describes the enrolled study population as a whole; it does not automatically represent the response probability for every individual tumor type listed in the eligibility scope.

5. Primary Endpoint

What exactly counts as a response?

The endpoint combines two response categories. A complete response requires disappearance of all target lesions. A partial response requires at least a 30% decrease in the sum of diameters of target lesions. Either category qualifies as an objective response.

The response assessment uses RECIST 1.1 with a protocol-specific modification limiting the number of target lesions to a maximum of 10 overall and a maximum of 5 target lesions per organ. The registry specifies independent central radiologic review and confirmatory assessment when required by RECIST 1.1.

6. Why the Endpoint Definition Matters

ORR is a binary endpoint at the participant level: each participant is ultimately classified as having an objective response or not having an objective response under the specified assessment rules. That makes the endpoint statistically different from a continuous tumor-measurement outcome or a time-to-event endpoint.

7. Planned Analysis

The registry identifies ORR as the primary endpoint and specifies that the percentage of participants experiencing a complete response or partial response will be presented. The endpoint is assessed through independent central radiologic review under modified RECIST 1.1, with confirmatory assessment as required.

For a binary response endpoint of this type, a typical statistical analysis would estimate the response proportion and an appropriate binomial confidence interval. The response count and analysis denominator should follow the prespecified response-evaluation population, including the registry rule that unknown or missing response information is treated as non-response.

Because the study is non-randomized and the registry information does not specify a separate comparator treatment, the central inferential quantity is the response proportion within the studied population rather than a randomized between-treatment effect estimate. A comparison with an external benchmark, if one were prespecified in a statistical analysis plan, would require that benchmark and its assumptions to be explicitly defined.

8. Statistical Methodology

Binary endpoint analysis

ORR is fundamentally a binary outcome at the participant level. Each participant contributes either a response classification or a non-response classification. The simplest descriptive estimator is the observed response proportion.

The statistical uncertainty around a proportion is commonly expressed with a binomial confidence interval. The interval reflects sampling uncertainty in the estimated response proportion; it is not a range containing the response probability of individual patients.

Why binomial methods fit ORR

Unlike a continuous endpoint, ORR does not require calculating a mean and standard deviation. The relevant outcome is whether a participant satisfies a categorical response definition. Binomial methods are therefore natural for estimating the proportion of participants meeting that definition.

Role of independent central review

The registry states that responses will be determined by independent central radiologic review, with confirmatory assessment as required under RECIST 1.1. Statistically, this creates a standardized adjudication process for the binary endpoint and helps define consistently which participants enter the response numerator.

Missing response information

The registry explicitly specifies that participants with unknown or missing response information will be treated as non-responders. This is an endpoint-definition rule rather than an instruction to impute a continuous measurement. It means that missing response status does not contribute to the response numerator.

9. Statistical Methods Explained

What is an objective response rate?

ORR is the percentage of participants who achieve either a complete response or a partial response according to the specified response criteria. In this study, the complete-response and partial-response definitions are based on modified RECIST 1.1.

Why is ORR a binary endpoint?

Although tumor measurements are continuous during radiologic assessment, the primary endpoint converts those measurements into a categorical clinical outcome: response versus non-response. This allows the study to summarize the percentage of participants meeting the predefined threshold for CR or PR.

Why does the denominator matter?

A response rate cannot be interpreted without knowing the population to which the percentage applies. The registry specifies that unknown or missing response information is treated as non-response, so those participants affect the denominator and do not enter the response numerator.

What does a confidence interval around ORR mean?

A confidence interval quantifies statistical uncertainty around the estimated response proportion under the specified sampling framework. A narrower interval indicates greater statistical precision; a wider interval indicates less precision. It does not mean that individual participants have responses occurring within the interval.

Why does non-randomization matter?

Randomization is the mechanism that balances measured and unmeasured prognostic factors probabilistically between treatment groups. Because this study is listed as non-randomized, an observed ORR should not automatically be interpreted as a causal treatment effect relative to another therapy.

Why can ORR not answer every clinical question?

A response endpoint captures tumor shrinkage according to a predefined threshold. It does not by itself describe how long a response lasts, how long participants remain alive, or what happens to participants whose disease does not meet the CR or PR criteria.

10. Interpretation of the Primary Endpoint

11. Analysis Population and Missing Data

The registry identifies an overall enrollment of 1,609 participants but does not provide a separate posted analysis population for the primary endpoint in the information summarized here. The endpoint definition itself does specify an important handling rule: participants with unknown or missing response information are treated as non-responders.

It is important to distinguish an endpoint-specific missing-data rule from broader statistical imputation. The registry specifies how missing or unknown response information is classified for ORR; it does not describe a general multiple-imputation or model-based missing-data strategy in the available record.

12. Non-Randomized Design and Causal Interpretation

The study is explicitly classified as non-randomized. This is one of the most consequential statistical features of the design because the observed response rate is not protected by randomization against differences in prognosis across treatment groups or cohorts.

This distinction is central to interpreting phase 2 oncology studies. A high response proportion can be clinically informative while still being different from a randomized treatment effect. The design determines what statistical conclusions can be attached to the estimate.

13. Two Arms Without a Randomized Comparator

The registry lists two arms and identifies pembrolizumab as the biological intervention in both. Because allocation is non-randomized, the existence of two arms does not establish a conventional experimental-versus-control comparison.

For statistical interpretation, this prevents the two arms from being automatically summarized as a randomized treatment-versus-control contrast. Additional arm-specific definitions would be required before a comparative analysis could be specified accurately.

14. Response Assessment Under Modified RECIST 1.1

The response framework is one of the most important methodological details in the registry record. It uses RECIST 1.1 with a modification limiting the number of target lesions to a maximum of 10 overall and a maximum of 5 target lesions per organ.

These rules make the endpoint reproducible at the participant level. The same underlying imaging data could produce different response classifications if a different response framework or target-lesion rule were used, which is why the endpoint definition is an essential part of the statistical analysis rather than merely a clinical detail.

15. Time Frame and Ongoing Study Status

The registry lists the study start date as 2015-12-18 and the primary completion date as 2027-05-04. The current study status is listed as ACTIVE_NOT_RECRUITING.

The ORR endpoint has a time frame of up to approximately 10.5 years. Because the trial remains listed as active, not recruiting and has a future primary completion date in the registry, the endpoint time frame should not be confused with a claim that a complete 10.5-year efficacy analysis has been posted.

16. What the Registry Does and Does Not Establish Statistically

This distinction between endpoint specification and posted statistical results is important. A registry can define an endpoint precisely even when the record does not contain a formal statistical analysis of that endpoint.

17. Limitations

• Non-randomized allocation: the study is explicitly listed as non-randomized, so an observed ORR does not by itself establish a causal treatment effect against an alternative therapy.
• Heterogeneous disease population: the registry lists multiple cancer types and disease categories. A pooled response proportion can therefore conceal substantial variation among tumor types.
• No separate comparator specified: both listed interventions are pembrolizumab, and the registry information does not identify a distinct comparator treatment.
• Missing-response rule: unknown or missing response information is classified as non-response, making completeness of response assessment directly relevant to ORR.
• Response is not a time-to-event endpoint: ORR does not describe duration of response, progression-free survival, or overall survival.
• Central review does not eliminate all uncertainty: independent radiologic assessment standardizes review but does not remove statistical sampling uncertainty or the clinical complexity of tumor-response assessment.
• Broad population: an overall response rate may not represent the response experience of any one tumor type within the study.
• Posted analysis limitation: the ClinicalTrials.gov record does not contain a posted formal statistical analysis with a numerical ORR estimate, confidence interval, or p-value.

18. Why This Trial Matters Statistically

KEYNOTE-158 is useful as a statistical teaching case because it illustrates how a phase 2 response endpoint is constructed in a large, heterogeneous, non-randomized oncology study.

19. Clinical Interpretation vs Statistical Interpretation

The distinction is particularly important in a non-randomized study. Statistical estimation can describe the response experience of the enrolled population, while causal interpretation requires stronger assumptions or an appropriate comparison framework.

20. Questions a Statistical Reviewer Should Ask

Is the overall ORR enough to compare the different diseases represented?

No. A pooled ORR combines participants across multiple listed disease categories. If disease-specific response rates were available, they would answer more focused questions. Comparisons across tumor types would also require attention to sample size, eligibility criteria, response assessment, and multiplicity.

Does independent central review make the study randomized?

No. Central review addresses consistency of response assessment; it does not change the allocation mechanism. The registry still classifies the study as non-randomized.

Does treating missing responses as non-response solve missing-data concerns?

It defines how missing response information contributes to ORR, but it does not make the missingness mechanism irrelevant. If missingness is related to disease status, treatment exposure, toxicity, or follow-up, the resulting ORR can be affected.

Can a response rate establish superiority?

Not without an appropriate comparative framework. A response proportion can describe the observed study population, but superiority requires a defined comparator and a statistical hypothesis appropriate to the design.

Why is the response threshold clinically important?

Because ORR is determined by whether tumor measurements cross a prespecified threshold. The 30% decrease criterion for partial response is therefore part of the endpoint definition and cannot be separated from the resulting statistical estimate.

Why does the maximum number of target lesions matter?

The registry specifies a maximum of 10 target lesions overall and a maximum of 5 per organ. This creates a standardized measurement framework for calculating the change in the sum of diameters used in response classification.

21. Record Summary

KEYNOTE-158 is a phase 2, non-randomized, unmasked, parallel-design treatment study with 1,609 enrolled participants and two listed arms, both identifying pembrolizumab as the biological intervention. Its registered primary endpoint is objective response rate, defined as the percentage of participants achieving a complete response or partial response according to modified RECIST 1.1, with independent central radiologic review and confirmatory assessment as required.

From a statistical perspective, the key features are the binary response endpoint, explicit handling of unknown or missing response information as non-response, centralized response assessment, and non-randomized allocation. These features mean that ORR can be interpreted as a population-level response proportion, while a causal treatment comparison requires additional design or comparative information.

22. Sources

• ClinicalTrials.gov: KEYNOTE-158 — NCT02628067.