This page separates reported trial results from statistical interpretation. The numerical results presented here are restricted to the statistical analyses posted for NCT04003636 and the ClinicalTrials.gov record.
1. Trial at a Glance
KEYNOTE-966 was a randomized, triple-masked, parallel-group phase 3 treatment trial in biliary tract carcinoma. It compared pembrolizumab plus chemotherapy with placebo plus chemotherapy and used overall survival as its registered primary endpoint.
| Feature | KEYNOTE-966 |
|---|---|
| Trial name | KEYNOTE-966 |
| NCT identifier | NCT04003636 |
| Phase | Phase 3 |
| Status | COMPLETED |
| Condition | Biliary Tract Carcinoma |
| Enrollment | 1069 |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Triple |
| Primary purpose | Treatment |
| Primary endpoint type | Time-to-event |
| Registered primary endpoints | 1 |
| Outcome measures posted | 6 |
| Statistical analyses posted | 3 |
| Lead sponsor | Merck Sharp & Dohme LLC |
| Sponsor type | Industry |
2. Clinical Question
The primary statistical question was whether first-line pembrolizumab added to gemcitabine/cisplatin would improve overall survival compared with placebo plus gemcitabine/cisplatin in participants with advanced and/or unresectable biliary tract carcinoma.
Population
Participants with biliary tract carcinoma in a phase 3 treatment trial evaluating first-line therapy for advanced and/or unresectable disease.
Intervention
Pembrolizumab plus gemcitabine and cisplatin.
Comparator
Placebo plus gemcitabine and cisplatin.
Primary question
Does the pembrolizumab-containing regimen improve overall survival relative to placebo plus gemcitabine/cisplatin?
3. Trial Design
Pembrolizumab + Chemotherapy
- Pembrolizumab
- Gemcitabine
- Cisplatin
Placebo + Chemotherapy
- Placebo
- Gemcitabine
- Cisplatin
The trial began on 2019-09-24 and reached primary completion on 2022-12-15. The registry classifies the allocation as randomized, the design model as parallel, and masking as triple.
4. Endpoints
| Endpoint | Registry definition / assessment | Time frame | Type |
|---|---|---|---|
| Overall Survival (OS) | Overall survival was defined as the time from randomization to death due to any cause. | Up to approximately 38 months | Time-to-event |
| Progression-free Survival (PFS) Per RECIST 1.1 as Assessed by BICR | Progression-free Survival per Response Evaluation Criteria in Solid Tumors Version 1.1 (RECIST 1.1) as assessed by Response Evaluation Criteria in Solid Tumors Version 1.1 (BICR). | Up to approximately 26 months | Time-to-event |
| Objective Response Rate (ORR) Per RECIST 1.1 as Assessed by BICR | Objective Response Rate per Response Evaluation Criteria in Solid Tumors Version 1.1 (RECIST 1.1) as assessed by Blinded Independent Central Review (BICR). | Up to approximately 26 months | Binary |
The registry identifies overall survival as the single registered primary endpoint. The posted statistical analyses include one primary endpoint analysis and two secondary endpoint analyses.
5. Analysis Populations and Statistical Questions
The posted analyses use all randomized participants as the analysis population for overall survival, progression-free survival, and objective response rate. This is important because the efficacy comparison is anchored to randomized treatment assignment rather than being restricted to participants who remained on treatment.
| Outcome | Analysis population | Groups compared | Effect measure |
|---|---|---|---|
| Overall Survival | All randomized participants | Pembrolizumab + Chemotherapy vs Placebo + Chemotherapy | Hazard ratio |
| Progression-free Survival | All randomized participants | Pembrolizumab + Chemotherapy vs Placebo + Chemotherapy | Hazard ratio |
| Objective Response Rate | All randomized participants | Pembrolizumab + Chemotherapy vs Placebo + Chemotherapy | Risk difference |
This population choice means that the treatment comparison retains the assignment created by randomization. It does not mean that every randomized participant necessarily had identical follow-up or exposure to treatment; those issues are handled differently in time-to-event and safety analyses.
6. Statistical Methodology
Log-rank test for time-to-event outcomes
The registry reports the log-rank test for the overall survival and progression-free survival analyses. The log-rank test compares the observed and expected numbers of events between randomized groups over follow-up and is designed for time-to-event data with censoring.
For this trial, the reported hypothesis type was superiority. The formal OS analysis produced a two-sided P-value of 0.0034.
Cox regression for hazard ratios
For overall survival, the registry states that the hazard ratio and associated 95% confidence interval were calculated using a Cox regression model with Efron's method of tie handling with treatment as a covariate. The same Cox-regression approach with Efron's tie handling and treatment as a covariate is reported for PFS.
An HR below 1 indicates a lower estimated instantaneous event rate in Arm A relative to Arm B under the fitted model. It is not an absolute risk difference and does not mean that every participant experiences the same proportional change in risk.
Efron's method for tied event times
Survival datasets can contain multiple events at the same recorded time. The registry specifically identifies Efron's method for handling tied event times in the Cox regression. This is a detail of how the partial likelihood is constructed and matters because the treatment-effect estimate comes from the fitted Cox model.
Score-based confidence interval for the response difference
For objective response rate, the registry reports the Miettinen & Nurminen method for calculating the difference in percentages. The normalized methodology classification identifies this as a score-based confidence interval approach for proportions, including Miettinen-Nurminen, Newcombe, and Wilson methods.
A positive value means that the observed response proportion was higher in Arm A. The reported ORR estimate is 0.2, with a two-sided 95% CI of -5.2 to 5.6.
Kaplan-Meier estimation
Kaplan-Meier estimation is the standard descriptive framework for displaying and estimating time-to-event distributions such as overall survival and progression-free survival. It accommodates right-censoring by allowing participants without an observed event to contribute information until their censoring time. The registry's posted formal comparison for the time-to-event analyses is the log-rank test, with Cox regression used for the reported hazard ratios.
7. Results
The registry contains statistical analyses for the primary overall survival endpoint and two secondary endpoints: progression-free survival and objective response rate. All three analyses compare Arm A, pembrolizumab plus chemotherapy, with Arm B, placebo plus chemotherapy.
Overall Survival — Primary Endpoint
Hazard ratio for death
95% CI: 0.72–0.95 · P = 0.0034
Two-sided 95% confidence interval · Superiority hypothesis
| Primary endpoint | Arm A | Arm B | Reported analysis |
|---|---|---|---|
| Overall Survival | Pembrolizumab + Chemotherapy | Placebo + Chemotherapy | HR 0.83 (95% CI 0.72–0.95); P = 0.0034 |
The analysis used all randomized participants and a log-rank test. The registry notes that the HR and associated 95% CI were calculated using Cox regression with Efron's method of tie handling and treatment as a covariate.
An OS hazard ratio of 0.83 means that, under the fitted Cox model, the estimated instantaneous rate of death in the pembrolizumab-plus-chemotherapy group was 0.83 times that in the placebo-plus-chemotherapy group. Expressed as a simple relative interpretation, this corresponds to an estimated 17% lower hazard in Arm A relative to Arm B.
The HR does not mean that 17% of participants avoided death, that survival time increased by 17%, or that each participant had exactly a 17% reduction in risk. A hazard ratio is a relative time-to-event measure derived from a statistical model.
The 95% CI of 0.72–0.95 describes uncertainty around the estimated hazard ratio under the analysis framework. It does not describe the range of effects experienced by individual participants. Because the entire reported interval is below 1, the estimated treatment effect is consistently below the null value within this confidence interval.
The P-value of 0.0034 addresses the statistical evidence against the null hypothesis used for the superiority comparison. It does not measure the size or clinical importance of the treatment effect. Effect size is better described by the HR and its confidence interval, alongside absolute survival measures when those are available.
As with other Cox-model hazard ratios, interpretation also depends on the model and the time-to-event structure. A single HR summarizes a relative event-rate comparison and should not automatically be interpreted as a constant percentage reduction in individual risk at every point in time.
Progression-free Survival — Secondary Endpoint
Hazard ratio for progression-free survival
95% CI: 0.75–1.00 · P = 0.0225
Two-sided 95% confidence interval · Superiority hypothesis
| Secondary endpoint | Time frame | Effect estimate | Analysis |
|---|---|---|---|
| Progression-free Survival per RECIST 1.1 as assessed by BICR | Up to approximately 26 months | HR 0.86 (95% CI 0.75–1.00) | Log-rank test; P = 0.0225 |
The analysis population was all randomized participants. The registry states that the HR and 95% CI were calculated using Cox regression with Efron's method of tie handling and treatment as a covariate.
An PFS hazard ratio of 0.86 means that the estimated instantaneous rate of the PFS event in Arm A was 0.86 times that in Arm B under the fitted Cox model. A simple relative interpretation is an estimated 14% lower hazard for the PFS event in Arm A.
This does not mean that 14% of participants had delayed progression, nor does it provide the difference in median PFS or the probability of remaining progression-free at a particular time. Those quantities require corresponding survival estimates, which are not part of the ClinicalTrials.gov record.
The 95% CI extends from 0.75 to 1.00. This indicates that the reported estimate has more uncertainty than a very narrow confidence interval would, and the upper endpoint reaches the conventional null value of 1.00.
The P-value of 0.0225 is evidence against the specified null hypothesis in the reported superiority analysis; it is not a measure of the magnitude of the PFS effect. The HR and its confidence interval provide the effect-size information.
Objective Response Rate — Secondary Endpoint
Difference in response percentages
95% CI: -5.2–5.6 · P = 0.4735
Miettinen & Nurminen method · Arm A minus Arm B
| Secondary endpoint | Time frame | Effect estimate | Analysis |
|---|---|---|---|
| Objective Response Rate per RECIST 1.1 as assessed by BICR | Up to approximately 26 months | Risk difference 0.2 (95% CI -5.2–5.6) | Miettinen & Nurminen; P = 0.4735 |
The reported difference in percentage was calculated as Arm A minus Arm B. The analysis population was all randomized participants.
The reported risk difference of 0.2 represents the response percentage in Arm A minus the response percentage in Arm B. A positive value therefore indicates a higher response percentage in Arm A by the amount represented by the estimate.
The estimate should not be confused with a hazard ratio. ORR is a binary endpoint: each participant is classified according to whether the prespecified response criterion was met. The Miettinen-Nurminen method is used here to quantify uncertainty around the difference between the two proportions.
The 95% CI of -5.2 to 5.6 spans zero, the null value for a risk difference. This interval indicates substantial uncertainty about the direction and magnitude of the between-arm response difference based on this analysis.
The P-value of 0.4735 does not describe the size of the response difference. It quantifies the evidence against the specified null hypothesis under the reported testing framework. The estimate and confidence interval are needed to understand the possible magnitude of the difference.
8. Comparing the Three Posted Analyses
| Endpoint | Type | Method | Effect measure | Estimate | 95% CI | P-value |
|---|---|---|---|---|---|---|
| Overall Survival | Primary time-to-event | Log-rank; Cox regression | Hazard ratio | 0.83 | 0.72–0.95 | 0.0034 |
| Progression-free Survival | Secondary time-to-event | Log-rank; Cox regression | Hazard ratio | 0.86 | 0.75–1.00 | 0.0225 |
| Objective Response Rate | Secondary binary | Miettinen & Nurminen | Risk difference | 0.2 | -5.2–5.6 | 0.4735 |
The statistical structure is instructive because the endpoints require different effect measures. OS and PFS are time-to-event outcomes, so the registry reports hazard ratios and log-rank tests. ORR is binary, so the registry reports a difference in percentages with a score-based confidence interval.
9. Safety Results
The ClinicalTrials.gov record reports serious adverse events by treatment arm for the first course and also provide a second-course Arm A safety denominator.
| Safety period | Arm | Serious adverse events affected | At risk |
|---|---|---|---|
| First Course | Arm A: Pembrolizumab + Che | 280 | 529 |
| First Course | Arm B: Placebo + Chemother | 264 | 534 |
| Second Course | Arm A: Pembrolizumab + Ch | 0 | 10 |
These are reported affected/at-risk counts rather than efficacy effect measures. They should therefore be interpreted separately from the OS, PFS, and ORR analyses. In particular, a safety count does not establish causality for an individual adverse event or provide a direct measure of comparative treatment efficacy.
10. Statistical Methods Explained
Why was a log-rank test used?
Overall survival and progression-free survival are time-to-event outcomes. Participants can have different follow-up times, and some participants may be censored before the event is observed. The log-rank test is designed to compare the event-time distributions of randomized groups while accounting for this censoring structure.
What does an OS hazard ratio of 0.83 mean?
A hazard ratio of 0.83 means that the fitted model estimates the instantaneous death rate in Arm A at 0.83 times the corresponding rate in Arm B. The simple complementary interpretation is a 17% lower estimated hazard. It does not mean a 17% absolute survival improvement or that each individual has exactly a 17% lower probability of death.
Why use a Cox regression model in addition to the log-rank test?
The log-rank test provides a hypothesis test for the time-to-event comparison, while the Cox model provides an estimated hazard ratio and its confidence interval. In KEYNOTE-966, the registry specifically reports Cox regression with treatment as a covariate and Efron's method for tied event times.
What does the 95% confidence interval tell us?
The confidence interval describes statistical uncertainty around the estimated effect under the model and sampling framework. For OS, the interval is 0.72–0.95 around the HR estimate of 0.83. It is not a range containing the treatment effect for 95% of individual participants.
Why is ORR analysed differently from OS?
ORR is binary rather than time-to-event. Each participant contributes a response classification rather than a time until death or progression. The registry therefore uses the Miettinen-Nurminen method for the difference in percentages, rather than a Cox hazard ratio.
What does a risk difference of 0.2 mean?
The reported ORR risk difference is calculated as Arm A minus Arm B. Thus, 0.2 is the estimated difference in response percentages on that scale. Its 95% confidence interval is -5.2 to 5.6, showing that the estimate is uncertain enough to include zero.
Why does the P-value not measure effect size?
A P-value addresses compatibility of the observed data with a specified null hypothesis under the statistical model. It depends on both the observed effect and the amount of information available. The magnitude and precision of the treatment effect are better described by the effect estimate and its confidence interval.
11. Understanding the Primary Analysis More Deeply
The OS result illustrates an important distinction between statistical evidence and effect-size interpretation. The reported HR of 0.83, 95% CI of 0.72–0.95, and P-value of 0.0034 answer related but different questions.
Effect size
The HR of 0.83 summarizes the relative time-to-event difference estimated by the Cox model.
Precision
The 95% CI of 0.72–0.95 describes uncertainty around that estimated hazard ratio.
Evidence against the null
The P-value of 0.0034 quantifies the reported statistical evidence against the superiority null hypothesis.
What is missing
The ClinicalTrials.gov record does not provide median OS, time-specific survival probabilities, or a Kaplan-Meier event table, so those quantities cannot be inferred here.
This separation is important in clinical-trial reporting. A statistically detectable difference is not itself a complete description of the treatment effect. Conversely, the effect estimate cannot be interpreted without considering its uncertainty and the design used to obtain it.
12. Time-to-Event Analysis and Censoring
Overall survival is defined as the time from randomization to death due to any cause. Because participants can have different observation times, time-to-event analysis must account for censoring. A participant who has not experienced the event by the last available observation does not simply become equivalent to a participant with an event-free lifetime; instead, that participant contributes information up to the censoring point.
Here, di represents the number of events at time ti, while ni represents the number at risk immediately before that time. The ClinicalTrials.gov record does not provide the event and censoring history needed to reconstruct a patient-level Kaplan-Meier curve.
The same general time-to-event framework applies to PFS, although its event definition is different from the registered OS definition. The registry classifies both OS and PFS as time-to-event endpoints.
13. Confidence Intervals and the Null Value
Confidence intervals are particularly useful because their null values depend on the effect measure.
| Effect measure | Null value | KEYNOTE-966 example |
|---|---|---|
| Hazard ratio | 1 | OS HR 0.83; 95% CI 0.72–0.95 |
| Hazard ratio | 1 | PFS HR 0.86; 95% CI 0.75–1.00 |
| Risk difference | 0 | ORR difference 0.2; 95% CI -5.2–5.6 |
This distinction explains why the ORR confidence interval is evaluated relative to zero while the survival hazard-ratio intervals are evaluated relative to one. Confusing these null values can lead to incorrect interpretation of clinical-trial results.
14. Covariate Adjustment in the Posted Analyses
The registry's analysis text identifies treatment as a covariate in the Cox regression model used to calculate the OS and PFS hazard ratios and confidence intervals. The ClinicalTrials.gov record does not specify additional covariates in those model descriptions.
This wording matters. A Cox model containing treatment as a covariate estimates the treatment comparison within the model framework, while the exact interpretation of additional covariate adjustment would require those variables and their prespecified role to be documented. They should not be inferred from the presence of the phrase "covariate adjustment" alone.
15. Superiority Testing
The posted analyses identify the hypothesis type as superiority. In a superiority framework, the treatment comparison asks whether the experimental treatment differs from the comparator in the specified favorable direction under the prespecified statistical test.
OS
The reported HR was 0.83 with a two-sided 95% CI of 0.72–0.95 and P = 0.0034.
PFS
The reported HR was 0.86 with a two-sided 95% CI of 0.75–1.00 and P = 0.0225.
ORR
The reported risk difference was 0.2 with a two-sided 95% CI of -5.2–5.6 and P = 0.4735.
Interpretive caution
The ClinicalTrials.gov record identifies superiority as the hypothesis type but do not provide a detailed multiplicity hierarchy or alpha-allocation scheme.
16. Multiplicity and What the Registry Data Do Not Establish
Three statistical analyses are posted: one for the primary OS endpoint and two for secondary endpoints, PFS and ORR. The ClinicalTrials.gov record does not specify an endpoint hierarchy, alpha-spending procedure, multiplicity adjustment, interim-analysis boundary, or other familywise-error strategy.
This is especially important when comparing the P-values for OS, PFS, and ORR. Their numerical values can be reported directly, but a formal claim about a familywise-error-controlled sequence of tests requires information about the prespecified statistical testing strategy.
17. Blinding and Risk of Bias
The trial was registered as triple-masked. Masking can reduce the potential for knowledge of treatment assignment to influence participant behavior, investigator decisions, assessment, or other aspects of trial conduct.
The PFS and ORR endpoints were assessed by BICR, or blinded independent central review. This is particularly relevant for radiologic outcomes because independent blinded review can reduce the influence of treatment knowledge on assessment of tumor response or progression.
Triple masking
The registry classifies the study as triple-masked, meaning masking was part of the registered trial design.
BICR
PFS and ORR were specified as being assessed using blinded independent central review.
18. What the Statistical Results Do and Do Not Tell Us
| Reported result | What it tells us | What it does not tell us |
|---|---|---|
| OS HR 0.83 | Relative time-to-event effect estimated by the Cox model. | It is not an absolute survival difference or an individual-level probability. |
| OS 95% CI 0.72–0.95 | Uncertainty around the estimated HR. | It is not the range of effects across individual participants. |
| OS P = 0.0034 | Evidence against the reported null hypothesis. | It is not a measure of treatment-effect magnitude. |
| PFS HR 0.86 | Relative time-to-event effect estimated by Cox regression. | It does not provide median PFS or a time-specific PFS probability. |
| ORR difference 0.2 | Difference in response percentages, Arm A minus Arm B. | It is not a hazard ratio or a time-to-response measure. |
| ORR 95% CI -5.2–5.6 | Uncertainty around the response-rate difference. | It does not describe the distribution of individual responses. |
19. Important Limitations and Interpretation Issues
- Limited reported result set: the statistical analyses posted on ClinicalTrials.gov include OS, PFS, and ORR, but do not provide median survival, time-specific survival probabilities, subgroup estimates, or a reconstructed Kaplan-Meier dataset.
- Hazard-ratio interpretation: a Cox hazard ratio is a model-based relative time-to-event measure. It should not automatically be translated into an identical percentage change in individual risk at every point in time.
- Censoring: time-to-event methods rely on appropriate handling and assumptions regarding censoring. The ClinicalTrials.gov record does not provide individual censoring histories.
- Confidence intervals: the width of an interval reflects statistical precision, while the point estimate describes the estimated effect. Neither alone describes individual patient outcomes.
- Multiplicity: the ClinicalTrials.gov record does not specify a multiplicity adjustment or endpoint hierarchy for the three posted statistical analyses.
- Subgroup interpretation: no subgroup estimates are reported, so treatment-effect heterogeneity cannot be assessed from the provided results.
- Missing-data and imputation details: the ClinicalTrials.gov record does not describe missing-data handling or imputation procedures.
- Interim analysis: the ClinicalTrials.gov record does not provide an interim-analysis schedule or alpha-spending procedure.
- Crossover: the ClinicalTrials.gov record does not identify a crossover procedure, so no crossover effect on OS should be inferred.
- Non-inferiority: the trial is identified as a superiority analysis; no non-inferiority margin is reported in the ClinicalTrials.gov record.
- Bayesian methods: no Bayesian method is identified in the statistical analyses posted on ClinicalTrials.gov.
20. Why This Trial Matters Statistically
KEYNOTE-966 is a useful teaching example because its posted analyses demonstrate how a single randomized trial can require different statistical frameworks for different endpoint types.
| Concept | How it appears in KEYNOTE-966 |
|---|---|
| Randomization | The trial uses randomized allocation with 2 parallel arms. |
| Triple masking | The registered masking structure is triple. |
| Time-to-event analysis | OS and PFS are classified as time-to-event endpoints. |
| Log-rank test | Reported for both OS and PFS. |
| Hazard ratio | Used as the effect measure for OS and PFS. |
| Cox regression | Used to calculate HRs and confidence intervals, with Efron's method for ties and treatment as a covariate. |
| Confidence intervals | Two-sided 95% CIs are reported for all three statistical analyses. |
| Binary endpoint analysis | ORR is analysed as a binary outcome. |
| Risk difference | ORR is reported as the difference in percentages, Arm A minus Arm B. |
| Score-based CI | Miettinen-Nurminen is reported for the ORR difference. |
| BICR | PFS and ORR are assessed by blinded independent central review. |
| Superiority | The posted analyses identify the hypothesis type as superiority. |
21. Statistical Interpretation of the Overall Evidence
The three posted analyses produce different forms of statistical evidence because they address different clinical endpoints. The primary OS analysis reports an HR of 0.83 with a 95% CI of 0.72–0.95 and P = 0.0034. The secondary PFS analysis reports an HR of 0.86 with a 95% CI of 0.75–1.00 and P = 0.0225. The secondary ORR analysis reports a risk difference of 0.2 with a 95% CI of -5.2–5.6 and P = 0.4735.
These results should not be collapsed into a single summary statistic. OS and PFS measure time-to-event outcomes, while ORR measures whether a response criterion was achieved. The appropriate effect measures and confidence intervals therefore differ.
The primary OS result is also not interchangeable with the PFS or ORR results. An OS hazard ratio describes the relative death hazard under the Cox model. A PFS hazard ratio describes the corresponding relative event hazard for PFS. The ORR risk difference describes a difference in response percentages. Each answers a different statistical question.
22. Statistical Questions a Reader Should Ask
What was randomized?
The treatment assignment was randomized between pembrolizumab plus chemotherapy and placebo plus chemotherapy.
What was the primary endpoint?
Overall survival, defined as time from randomization to death due to any cause, with a time frame of up to approximately 38 months.
How was OS compared?
The registry reports a log-rank test and Cox regression for the hazard ratio and confidence interval.
Why is ORR different?
ORR is binary, so the registry reports a difference in percentages using the Miettinen-Nurminen method.
These questions help separate the design, endpoint definition, analysis method, and effect estimate. That separation is central to interpreting randomized clinical-trial statistics without conflating statistical significance, effect magnitude, and clinical outcome definitions.
23. Related Tutorials
Learn more about the methods used in this trial:
24. Related Calculators
25. Sources
- ClinicalTrials.gov: NCT04003636 — KEYNOTE-966.
- PubMed: PMID 37075781.
- PubMed: PMID 42176183.
- PubMed: PMID 40268484.
- PubMed: PMID 39206260.
- PubMed: PMID 38661512.
Continue through the Clinical Biostats statistical pathway
Use the endpoint and analysis concepts from this trial to explore survival analysis, confidence intervals, categorical-data methods, and related statistical workflows.
26. Record Summary
KEYNOTE-966 provides a clear example of how randomized clinical-trial evidence is translated into endpoint-specific statistical analyses. The primary endpoint was overall survival, a time-to-event outcome analysed using a log-rank test with a Cox-regression hazard ratio. The reported OS HR was 0.83, with a two-sided 95% CI of 0.72–0.95 and P = 0.0034. Secondary analyses used the same general time-to-event framework for PFS and a score-based Miettinen-Nurminen approach for the binary ORR endpoint.
The statistical story is therefore not simply a collection of P-values. It consists of the randomized design, the endpoint definitions and time frames, the analysis populations, the choice of effect measure, the confidence intervals, and the assumptions underlying the statistical models. Reading those components together provides a more precise understanding of what the reported evidence establishes and what remains outside the ClinicalTrials.gov record.