← Clinical Trials
Atrial Fibrillation Phase 3 Time-to-Event NCT00403767

ROCKET AF: Complete Statistical Analysis of Rivaroxaban in Non-Valvular Atrial Fibrillation

An independent statistical analysis of the randomized phase 3 ROCKET AF trial evaluating rivaroxaban versus warfarin for the prevention of stroke and non-CNS systemic embolism in patients with non-valvular atrial fibrillation.

Completed trial  ·  Enrollment 14,269  ·  Randomized, parallel, quadruple-masked
Scope of this record

This page separates reported trial results from statistical interpretation. Numerical results are restricted to the information posted for NCT00403767 on ClinicalTrials.gov in the ClinicalTrials.gov record. Where the registry does not provide additional details, this page does not infer them from outside publications.

Registry note: 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

ROCKET AF was a randomized, parallel-group, quadruple-masked phase 3 trial in cardiology. The trial compared rivaroxaban with warfarin for prevention in patients with non-valvular atrial fibrillation, with stroke and non-CNS systemic embolism forming the primary efficacy endpoint and major/non-major clinically relevant bleeding forming the primary safety endpoint.

14,269
Enrollment
Randomized patients
2
Arms
Rivaroxaban vs Warfarin
3
Primary endpoints
2 efficacy analyses + safety
0.79
Primary efficacy HR
95% CI 0.66–0.96
FeatureROCKET AF
Trial nameROCKET AF
PhasePhase 3
StatusCompleted
Start2006-12
Primary completion2010-09
PopulationPatients with non-valvular atrial fibrillation; registered conditions include atrial fibrillation, stroke, and embolism
AllocationRandomized
Design modelParallel
MaskingQuadruple
Primary purposePrevention
Enrollment14,269
Lead sponsorJohnson & Johnson Pharmaceutical Research & Development, L.L.C.
Sponsor typeIndustry
ClinicalTrials.govNCT00403767

2. Clinical Question

The central statistical question was whether rivaroxaban was non-inferior to warfarin with respect to the time to first occurrence of stroke or non-CNS systemic embolism while on treatment. A separate superiority analysis was also posted for the same primary efficacy composite, while the primary safety analysis evaluated the time to first major or non-major clinically relevant bleeding event.

Population

Patients with non-valvular atrial fibrillation. The registry lists atrial fibrillation, stroke, and embolism among the trial conditions.

Intervention

Rivaroxaban, with matching placebo for the warfarin component used to maintain masking.

Comparator

Warfarin, with matching placebo for the rivaroxaban component used to maintain masking.

Primary question

How does rivaroxaban compare with warfarin for the time to stroke/non-CNS systemic embolism, and what is the comparative hazard of clinically relevant bleeding?

3. Trial Design

01
Randomize14,269 patients
02
Parallel armsRivaroxaban vs warfarin
03
Quadruple maskMatching placebos
04
FollowUp to 4 years
05
AnalyzeTime-to-event Cox models
ARM A · RIVAROXABAN

Rivaroxaban

  • Rivaroxaban was the active drug intervention.
  • Matching placebo for the warfarin component was used to support masking.
  • Primary efficacy and safety comparisons were made against the warfarin arm.
ARM B · WARFARIN

Warfarin

  • Warfarin was the active comparator intervention.
  • Matching placebo for the rivaroxaban component was used to support masking.
  • Primary efficacy and safety comparisons were made against the rivaroxaban arm.
Why the masking is statistically relevant: the registry describes the trial as quadruple-masked and lists matching placebos for both treatment components. Masking is a design feature intended to reduce the influence of treatment knowledge on trial conduct and assessment. The ClinicalTrials.gov record does not provide a more detailed description of the masking procedures.

4. Endpoints

ClinicalTrials.gov lists three primary endpoints. All three are time-to-event outcomes in the posted statistical analyses. The registry defines the efficacy endpoint using time from randomization to the first event while on treatment, whereas the safety endpoint uses time from the first dose to the first event while on treatment.

Primary endpointTime frameRegistry definition
The Composite Event of Stroke/Non-CNS Systemic Embolism: Primary Efficacy (Non-Inferiority) Up to 4 years The number of patients with the first occurrence of a stroke or non-CNS systemic embolism while on treatment, defined as the time interval from the first dose to the last dose of study drug plus 2 days. The statistical analysis is based on time from randomization to the first occurrence of the event while on treatment.
The Composite of Event of Stroke/Non-CNS Systemic Embolism: Primary Efficacy (Superiority) Up to 4 years The number of patients with the first occurrence of a stroke or non-CNS systemic embolism while on treatment, defined as the time interval from the first dose to the last dose of study drug plus 2 days. The statistical analysis is based on time from randomization to the first occurrence of the event while on treatment.
The Composite Event of Major/Non-major Clinically Relevant Bleeding Events: Primary Safety Up to 4 years The number of patients with the first occurrence of a major or non-major clinically relevant bleeding event while on treatment. The statistical analysis is based on time from the first dose of study drug to the first occurrence of the event while on treatment.
Important endpoint distinction: the non-inferiority and superiority analyses concern the same stroke/non-CNS systemic embolism composite but represent different statistical questions. The registry explicitly identifies the first analysis as non-inferiority and the second as superiority.

5. Analysis Populations

The registry identifies different analysis populations for the non-inferiority efficacy analysis and the other posted primary analyses. This distinction is central to interpreting the results because the non-inferiority analysis was based on the per-protocol population, whereas the superiority efficacy and primary safety analyses were based on the safety population.

PopulationRegistry description / role
Per-protocol population For the primary non-inferiority efficacy analysis, the PP population consisted of all randomized unique patients excluding those who had specific pre-defined major protocol deviations that occurred by the time of enrollment into the study. The registry text is truncated after this definition.
Safety population For the superiority efficacy and primary safety analyses, the safety population consisted of all randomized unique patients who took at least 1 dose of study medication after randomization during the double-blind treatment period. The registry provides additional site-specific exclusion wording for some efficacy analyses, but the registry-reported text is truncated.
Why population choice matters: non-inferiority trials require particular attention to analysis populations. The ClinicalTrials.gov record explicitly state that the primary non-inferiority analysis used the per-protocol population. The registry also states that the superiority analysis used the safety population. These are not interchangeable definitions, so their estimates should not be treated as though they came from an identical analysis set.

6. Statistical Methodology

Cox proportional-hazards model

All 10 posted statistical analyses use a Cox proportional-hazards model. The effect measure is a hazard ratio. The analysis notes describe a non-stratified Cox proportional-hazards model with treatment as a covariate. The registry also identifies covariate adjustment and stratified analysis among the other concepts associated with the analyses.

Model interpretation
HR = estimated hazard in the rivaroxaban group ÷ estimated hazard in the warfarin group

For these analyses, an HR below 1 indicates a lower estimated instantaneous event hazard for rivaroxaban relative to warfarin, while an HR above 1 indicates a higher estimated instantaneous event hazard. The HR is a relative time-to-event measure, not a direct probability or absolute risk difference.

Time-to-event analysis

The endpoints are defined around the timing of a first event. This structure makes survival-analysis methods appropriate because participants can contribute different amounts of follow-up and may remain event-free through the end of their applicable observation period.

Covariate adjustment

The registry analysis notes identify treatment as a covariate in the non-stratified Cox proportional-hazards model and also identify covariate adjustment as an analysis concept. Thus, the reported HR is model-based rather than a simple ratio of two cumulative event proportions.

Stratified analysis

The registry identifies stratified analysis as an associated analysis concept. However, the ClinicalTrials.gov record does not specify the individual stratification factors or their exact implementation. No additional stratification variables are therefore inferred here.

Kaplan-Meier estimation

Kaplan-Meier estimation is a standard descriptive method for time-to-event endpoints and would normally be used to estimate event-free survival over time. However, the ClinicalTrials.gov record identifies the Cox proportional-hazards model as the posted method and do not provide a ClinicalTrials.gov statistical-analysis record specifically identifying Kaplan-Meier estimates. The registry therefore does not report Kaplan-Meier numerical results in the ClinicalTrials.gov record.

Educational note: a Kaplan-Meier curve cannot be reconstructed from the reported hazard ratios and confidence intervals alone. Event times and censoring information, or suitably detailed underlying data, would be required.

7. Non-Inferiority Framework

The primary efficacy non-inferiority analysis was based on the per-protocol population and used a non-inferiority margin of 1.46 in hazard ratio. The registry states that the alternative hypothesis was non-inferiority by this margin and that the required number of primary efficacy endpoint events was determined using a 1-sided alpha of 0.025.

Prespecified non-inferiority margin

HR 1.46

Non-inferiority margin for the primary efficacy analysis

Analysis population: per-protocol  ·  1-sided alpha: 0.025

The logic of a hazard-ratio non-inferiority analysis is different from simply asking whether a conventional two-sided confidence interval includes 1. The relevant question is whether the observed treatment effect is sufficiently far from the unfavorable boundary that the prespecified non-inferiority criterion is met.

Non-inferiority logic
Observed HR < NI margin → evidence can support non-inferiority, subject to the prespecified testing framework

Here, the registry specifies an unfavorable hazard-ratio boundary of 1.46. The reported estimate of 0.79 is below that boundary. The formal conclusion should be tied to the prespecified one-sided testing framework rather than to an isolated comparison with HR = 1.

Why this is different from superiority: a non-inferiority analysis asks whether the experimental treatment has not lost more than a prespecified amount of efficacy. A superiority analysis instead tests whether the treatment groups differ in the specified favorable direction. ROCKET AF posted both analyses for the primary efficacy composite.

8. Primary Results

Primary Efficacy — Non-Inferiority

The primary non-inferiority analysis compared rivaroxaban with warfarin for the first occurrence of stroke or non-CNS systemic embolism while on treatment, with a time frame of up to 4 years. The analysis used the per-protocol population and a Cox proportional-hazards model.

Stroke / non-CNS systemic embolism

HR 0.79

95% CI: 0.66–0.96   ·   P < 0.001

Rivaroxaban vs warfarin · per-protocol population · non-inferiority analysis

Clinical Biostats interpretation

What the estimate means: the reported HR of 0.79 means that, under the fitted Cox model, the estimated instantaneous hazard of the first on-treatment stroke/non-CNS systemic embolism event was 0.79 times that in the warfarin group. Expressed as a simple relative interpretation, this corresponds to an estimated hazard that was 21% lower for rivaroxaban under the model.

What it does not mean: it does not mean that 21% of patients avoided an event, that each individual patient's risk fell by exactly 21%, or that the cumulative event probability was 21% lower.

Confidence interval: the 95% CI of 0.66–0.96 describes uncertainty around the estimated hazard ratio under the analysis framework. It does not describe the range of individual treatment effects.

P-value: P < 0.001 describes evidence against the relevant null hypothesis in the posted analysis; it does not measure the size or clinical importance of the treatment effect.

Non-inferiority caution: the key comparison is with the prespecified non-inferiority margin of 1.46, not simply with HR = 1. The registry specifies a one-sided alpha of 0.025 for this non-inferiority framework.

Population and censoring: the analysis was based on the per-protocol population and on-treatment event definitions. Consequently, the result should be interpreted within that analysis framework rather than as a generic statement about every randomized patient regardless of protocol deviations or treatment exposure.

Proportional-hazards assumption: the Cox model expresses the treatment comparison through a hazard ratio. The ClinicalTrials.gov record does not report a diagnostic assessment of proportional hazards, so the assumption cannot be evaluated from this record alone.

Primary Efficacy — Superiority

The same composite of stroke/non-CNS systemic embolism was also analyzed under a superiority hypothesis. The posted analysis used the safety population and a Cox proportional-hazards model with treatment as a covariate.

Stroke / non-CNS systemic embolism

HR 0.79

95% CI: 0.65–0.95   ·   P = 0.015

Rivaroxaban vs warfarin · safety population · superiority analysis

Clinical Biostats interpretation

What the estimate means: the HR of 0.79 represents an estimated instantaneous event hazard approximately 0.79 times that of warfarin in the posted superiority analysis. In relative terms, that corresponds to an estimated hazard 21% lower for rivaroxaban under the model.

What it does not mean: the HR is not a 21% absolute reduction in the number of patients with events, and it does not indicate that every patient experiences the same proportional change.

Confidence interval: the 95% CI of 0.65–0.95 quantifies uncertainty around the HR estimate. Its width reflects statistical precision; it does not provide a range of outcomes that individual patients can be expected to experience.

P-value: P = 0.015 is evidence from the specified superiority test against its null hypothesis. A p-value is not an effect-size measure and should not be used as a substitute for the HR or its confidence interval.

Population matters: this analysis used the safety population, rather than the per-protocol population used for the primary non-inferiority analysis. The two primary efficacy estimates therefore arise from different analysis populations and should be labeled accordingly.

Model caution: the result comes from a Cox proportional-hazards model. The ClinicalTrials.gov record does not report a formal assessment of the proportional-hazards assumption.

Primary Safety

The primary safety endpoint was the first occurrence of a major or non-major clinically relevant bleeding event while on treatment, assessed over up to 4 years. The posted analysis used the safety population and a Cox proportional-hazards model.

Major / non-major clinically relevant bleeding

HR 1.03

95% CI: 0.96–1.11   ·   P = 0.442

Rivaroxaban vs warfarin · safety population · superiority analysis

Clinical Biostats interpretation

What the estimate means: an HR of 1.03 indicates that the estimated instantaneous hazard of the first major or non-major clinically relevant bleeding event was approximately 1.03 times that in the warfarin group under the posted Cox model.

What it does not mean: it does not establish that the two treatments have identical bleeding risks. It also does not translate into a 3% absolute increase or decrease in the proportion of patients experiencing bleeding.

Confidence interval: the 95% CI of 0.96–1.11 describes uncertainty around the HR. It includes HR = 1, so the estimate is compatible with modestly lower or modestly higher instantaneous hazard within this interval.

P-value: P = 0.442 is not a measure of the magnitude of the observed HR. A nonsignificant p-value should not be interpreted as proof that the treatment effects are identical.

Safety interpretation: this was a superiority analysis of a composite safety endpoint. The registry separately reports serious adverse events by arm, but those serious-adverse-event figures should not be substituted for the primary bleeding endpoint.

Model and exposure caution: the analysis was based on the safety population and on-treatment timing. The Cox proportional-hazards assumption is also relevant here, and the ClinicalTrials.gov record does not provide a formal diagnostic assessment.

9. Primary Results Side by Side

EndpointPopulationHR95% CIP-valueHypothesis
Stroke/non-CNS systemic embolism Per-protocol 0.79 0.66–0.96 <0.001 Non-inferiority
Stroke/non-CNS systemic embolism Safety 0.79 0.65–0.95 0.015 Superiority
Major/non-major clinically relevant bleeding Safety 1.03 0.96–1.11 0.442 Superiority

The two efficacy analyses produce the same point estimate, 0.79, but different confidence intervals and p-values because they represent distinct analyses with different populations and hypothesis frameworks. The safety endpoint has a point estimate close to 1 and a confidence interval spanning both sides of 1.

10. Secondary Endpoint Results

The registry posts seven secondary time-to-event analyses in addition to the three primary analyses. Each used a Cox proportional-hazards model, compared rivaroxaban with warfarin, and was analyzed under a superiority hypothesis.

Secondary endpointHR95% CIP-value
Stroke/non-CNS systemic embolism/vascular death 0.86 0.74–0.99 0.034
Stroke/non-CNS systemic embolism/myocardial infarction/vascular death 0.85 0.74–0.96 0.010
Stroke 0.85 0.70–1.03 0.092
Non-CNS systemic embolism 0.23 0.09–0.61 0.003
Myocardial infarction 0.81 0.63–1.06 0.121
Vascular death 0.89 0.73–1.10 0.289
All-cause mortality 0.85 0.70–1.02 0.073

These results are all relative hazard estimates rather than absolute event probabilities. For example, the reported HR of 0.23 for non-CNS systemic embolism indicates a substantially lower estimated instantaneous hazard under the Cox model, but the ClinicalTrials.gov record does not include the underlying event counts or cumulative incidence needed to translate that estimate into absolute risk.

Multiplicity matters: the registry identifies these seven outcomes as secondary superiority analyses, but the ClinicalTrials.gov record does not provide a multiplicity-adjustment procedure, endpoint hierarchy, or alpha-allocation scheme for the secondary endpoints. The individual p-values should therefore be interpreted as reported statistical results rather than automatically treating every secondary test as an independent confirmatory claim.

11. Reading the Secondary Hazard Ratios

HR 0.86

For the composite of stroke/non-CNS systemic embolism/vascular death, the estimated instantaneous hazard was 0.86 times the warfarin-group hazard under the posted model.

HR 0.85

For the broader stroke/non-CNS systemic embolism/myocardial infarction/vascular death composite, the estimated hazard ratio was 0.85.

HR 0.23

For non-CNS systemic embolism alone, the estimated hazard ratio was 0.23, with a 95% CI of 0.09–0.61.

HR 0.85

For all-cause mortality, the estimated hazard ratio was 0.85, with a 95% CI of 0.70–1.02.

The variation among estimates illustrates why a trial should not be summarized using a single hazard ratio. Each endpoint defines a different event process. Composite endpoints also combine multiple clinical events, so their interpretation depends on which components contribute events and how those components relate to the clinical question.

12. Safety: Serious Adverse Events

The ClinicalTrials.gov record reports serious adverse events by treatment arm as affected patients divided by patients at risk.

ArmAffected / at riskRegistry measure
Rivaroxaban2649 / 7111Serious adverse events
Warfarin2720 / 7125Serious adverse events
Serious adverse events: affected patients
Rivaroxaban
2649
Warfarin
2720

The serious-adverse-event figures are descriptive affected/at-risk counts. They are not the same endpoint as the primary composite of major/non-major clinically relevant bleeding events, and the ClinicalTrials.gov record does not provide a statistical-analysis record for the serious-adverse-event measure itself.

13. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

The primary and secondary endpoints are time-to-event outcomes. A Cox model is designed to compare event hazards while accounting for the timing of events and censoring rather than reducing follow-up to a single yes/no outcome at a fixed time. The registry explicitly identifies the Cox proportional-hazards model as the method used for all 10 posted statistical analyses.

What does an HR of 0.79 mean?

An HR of 0.79 means that the fitted model estimates the instantaneous event hazard in the rivaroxaban group at 79% of the corresponding hazard in the warfarin group. A simple relative interpretation is a 21% lower estimated hazard. It is not equivalent to saying that 21% fewer patients experienced the event.

Why is the non-inferiority margin 1.46 rather than 1?

Superiority asks whether the treatment effect differs from the null value of HR = 1 in the favorable direction. Non-inferiority asks a different question: whether the experimental treatment is not worse than the comparator by more than a prespecified clinically acceptable amount. The registry specifies that amount as an HR margin of 1.46.

Why was the per-protocol population important for non-inferiority?

The registry states that the primary non-inferiority analysis used the per-protocol population. This is important because protocol deviations can make an experimental treatment and comparator appear more similar, potentially making it easier to conclude non-inferiority. The ClinicalTrials.gov record does not provide the complete protocol-deviation definition, so the analysis should be described using the registry's stated population rather than reconstructing additional criteria.

Why are there two analyses of the same efficacy endpoint?

The registry explicitly distinguishes a primary efficacy non-inferiority analysis from a primary efficacy superiority analysis. The first uses the per-protocol population and a non-inferiority hypothesis; the second uses the safety population and a superiority hypothesis. Because the hypotheses and populations differ, the two results answer related but non-identical statistical questions.

What does a 95% confidence interval tell us?

A 95% confidence interval communicates statistical uncertainty around the estimated hazard ratio under the specified analysis framework. A narrower interval generally indicates greater precision than a wider interval, but the interval does not describe individual patients' possible treatment responses. For the primary non-inferiority result, the interval is also not the sole criterion because the registry specifies a one-sided non-inferiority framework with a prespecified margin.

Why doesn't the p-value measure treatment effect size?

A p-value is a measure of compatibility between the observed data and a specified null hypothesis under the statistical test. It depends on both the observed effect and the amount of information in the analysis. The hazard ratio describes the estimated relative effect, while its confidence interval describes uncertainty around that estimate. These quantities should be considered together.

14. Confidence Intervals and Precision

The three primary analyses illustrate why a point estimate should never be interpreted without its confidence interval.

EndpointEstimate95% CIPrecision interpretation
Primary efficacy, non-inferiority 0.79 0.66–0.96 The interval spans a range of plausible model-based hazard ratios below 1 and remains below the non-inferiority margin of 1.46.
Primary efficacy, superiority 0.79 0.65–0.95 The interval is relatively concentrated around the point estimate and remains below 1.
Primary safety 1.03 0.96–1.11 The interval is centered close to 1 and includes both values below and above 1.

The intervals should be interpreted in the context of their respective hypotheses and analysis populations. In particular, the primary efficacy non-inferiority interval should not be treated as though it were a conventional superiority confidence interval answering the same question.

15. P-values and Hypothesis Types

AnalysisHypothesis typeP-valueWhat the test addresses
Primary efficacyNon-inferiority or equivalence<0.001Whether the treatment meets the prespecified non-inferiority framework
Primary efficacySuperiority0.015Whether the treatment shows superiority under the posted analysis
Primary safetySuperiority0.442Whether the bleeding hazard differs under the posted superiority analysis

The numerical p-values are not directly comparable as though they were measuring the same hypothesis. The first is attached to a non-inferiority framework with a specified margin and one-sided alpha; the latter two are superiority analyses.

16. Censoring and On-Treatment Definitions

Time-to-event analysis depends not only on whether an event occurs, but also on how follow-up is defined. The registry provides explicit on-treatment definitions for the primary endpoints.

Efficacy timing

The efficacy definition considers events from the first dose through the last dose of study drug plus 2 days, while the statistical analysis uses time from randomization to the first occurrence while on treatment.

Safety timing

The safety endpoint is based on time from the first dose of study drug to the first occurrence of the bleeding event while on treatment.

This distinction matters because the statistical clock and the treatment-exposure definition are not simply interchangeable concepts. A Cox model can use time from randomization while the event itself is classified according to the registry's on-treatment definition.

Missing-data limitation: the ClinicalTrials.gov record does not report a specific missing-data or imputation strategy for these time-to-event analyses. Time-to-event methods generally use censoring information rather than imputing an event time for every participant, but the exact censoring rules beyond those explicitly stated above are not provided here.

17. Stratification and Covariate Adjustment

The registry identifies both covariate adjustment and stratified analysis among the concepts associated with the posted analyses. The specific analysis notes describe a (non-stratified) Cox Proportional Hazards model with treatment as covariate.

Model
Cox proportional-hazards model
Treatment variable
Treatment entered as a covariate
Additional concept
Covariate adjustment
Additional concept
Stratified analysis

The ClinicalTrials.gov record does not specify the exact covariates or strata used beyond the treatment covariate description. A detailed reconstruction of the model matrix or stratum definitions would therefore go beyond the available registry information.

18. Interim Analysis and Multiplicity

The ClinicalTrials.gov record identifies non-inferiority and superiority hypothesis types and provide the one-sided alpha of 0.025 used for the non-inferiority design. They do not provide a documented interim-analysis schedule, alpha-spending function, stopping boundary, or multiplicity-adjustment procedure.

Design topicWhat the ClinicalTrials.gov record supports
Non-inferiority margin1.46 in hazard ratio
Non-inferiority alpha1-sided alpha of 0.025
Superiority testingPosted for the primary efficacy endpoint and primary safety endpoint
Multiplicity adjustmentNot specified in the ClinicalTrials.gov record
Interim-analysis planNot specified in the ClinicalTrials.gov record
Bayesian methodsNot reported in the statistical analyses posted on ClinicalTrials.gov
Statistical discipline: the absence of an interim or multiplicity procedure in the ClinicalTrials.gov record does not mean that no such procedure existed elsewhere in the trial documentation. It means that this page does not have sufficient registry data to describe one. No additional procedure is inferred.

19. Randomization and Blinding

ROCKET AF used randomized allocation, a parallel design, and quadruple masking. These design features address different sources of bias and support the interpretability of the treatment comparison.

Randomization

Randomized allocation helps create treatment groups whose differences can be interpreted through the assigned treatment rather than through investigator selection.

Parallel design

The two randomized groups were evaluated as parallel treatment arms rather than as sequential exposures within the same patients.

Quadruple masking

The registry identifies the study as quadruple-masked and lists matching placebos for both treatment components.

Prevention objective

The registry identifies prevention as the primary purpose, with stroke and embolism-related time-to-event outcomes forming the central efficacy framework.

20. Secondary Endpoint Interpretation

The secondary analyses show a range of hazard-ratio estimates. Reading them correctly requires keeping the event definition, confidence interval, and hypothesis test together.

EndpointStatistical reading
Stroke/non-CNS systemic embolism/vascular death HR 0.86 indicates a lower estimated event hazard under the Cox model; 95% CI 0.74–0.99 and P = 0.034 are the posted uncertainty and superiority-test results.
Stroke/non-CNS systemic embolism/myocardial infarction/vascular death HR 0.85 indicates a lower estimated event hazard; 95% CI 0.74–0.96 and P = 0.010 are the posted results.
Stroke HR 0.85 indicates a lower point estimate of hazard, but the 95% CI of 0.70–1.03 crosses 1 and the posted P-value is 0.092.
Non-CNS systemic embolism HR 0.23 is substantially below 1, with a 95% CI of 0.09–0.61 and P = 0.003.
Myocardial infarction HR 0.81 is below 1, but the 95% CI of 0.63–1.06 crosses 1 and P = 0.121.
Vascular death HR 0.89 is below 1, with a 95% CI of 0.73–1.10 and P = 0.289.
All-cause mortality HR 0.85 is below 1, with a 95% CI of 0.70–1.02 and P = 0.073.

A useful statistical distinction is between direction and evidence. An HR below 1 gives the direction of the point estimate, but the confidence interval and hypothesis test determine how precisely and strongly that estimate differs from the relevant null under the specified analysis.

21. Why the Composite Endpoint Matters

The primary efficacy endpoint combines stroke and non-CNS systemic embolism. Composite endpoints can increase the number of observed events and thereby provide a broader measure of treatment effect, but the components are clinically distinct events and should also be examined individually when component analyses are available.

ROCKET AF provides component analyses for stroke and non-CNS systemic embolism. The posted HR for stroke is 0.85, whereas the HR for non-CNS systemic embolism is 0.23. These estimates demonstrate why the composite HR should not be interpreted as though every component had the same treatment effect.

Interpretive caution: the ClinicalTrials.gov record does not provide event counts for the individual components. Therefore, this page does not infer which component contributed more events to the composite or attempt to decompose the overall composite result quantitatively.

22. All-Cause Mortality as a Secondary Endpoint

All-cause mortality was analyzed as a secondary time-to-event endpoint using a Cox proportional-hazards model in the safety population.

All-cause mortality

HR 0.85

95% CI: 0.70–1.02   ·   P = 0.073

Rivaroxaban vs warfarin · superiority analysis

The point estimate is below 1, indicating a lower estimated mortality hazard under the fitted model. However, the 95% CI extends from 0.70 to 1.02, and the posted p-value is 0.073. The registry data therefore support reporting the estimate, interval, and p-value together rather than reducing the result to a binary label.

23. What the Hazard Ratio Does — and Does Not — Mean

Example: primary efficacy HR 0.79

A reported HR of 0.79 means that, under the fitted Cox model, the estimated instantaneous rate of the specified event in the rivaroxaban group was approximately 79% of the corresponding rate in the warfarin group. A simple relative interpretation is a 21% lower estimated hazard.

It does not mean that 21% fewer patients experienced the event, that each patient had exactly a 21% reduction in risk, or that the absolute difference between treatment groups was 21 percentage points.

Why the confidence interval matters

The 95% CI describes uncertainty around the estimated hazard ratio. For the primary non-inferiority analysis, the interval is 0.66–0.96. For the superiority analysis of the same endpoint, it is 0.65–0.95. The intervals are not ranges of individual patient outcomes.

Why the p-value is not effect size

The p-value quantifies evidence under a specified null hypothesis and testing framework. It does not tell us whether an effect is large, small, clinically important, or unimportant. Those questions require the hazard ratio, confidence interval, endpoint definition, and clinical context to be considered together.

24. Clinical Interpretation vs Statistical Interpretation

Statistical interpretation

The primary efficacy endpoint had an HR of 0.79 in the posted non-inferiority analysis and an HR of 0.79 in the posted superiority analysis. The primary safety endpoint had an HR of 1.03. Each estimate has its own confidence interval, p-value, hypothesis type, and analysis population.

Clinical interpretation

The statistical record describes comparative hazards for stroke/non-CNS systemic embolism and clinically relevant bleeding, along with secondary cardiovascular and mortality endpoints. The ClinicalTrials.gov record does not provide enough absolute event information to translate every HR into an absolute treatment benefit or harm.

25. Important Limitations and Interpretation Issues

26. Why This Trial Matters Statistically

ROCKET AF is a useful teaching case because it combines randomized treatment allocation, quadruple masking, a non-inferiority framework, superiority testing, time-to-event endpoints, Cox proportional-hazards modeling, different analysis populations, and component analyses of a composite endpoint.

ConceptHow it appears in ROCKET AF
RandomizationRandomized parallel-group phase 3 design
BlindingQuadruple masking with matching placebos
Time-to-event endpointsPrimary and secondary outcomes analyze time to first event
Hazard ratioPrimary and secondary treatment effects are expressed as HRs
Cox modelPosted statistical method for all 10 analyses
Non-inferiorityPrimary efficacy analysis with HR margin 1.46
One-sided alpha1-sided alpha of 0.025 for the non-inferiority design
Per-protocol analysisUsed for the primary non-inferiority efficacy analysis
Safety populationUsed for the primary superiority efficacy and safety analyses
Composite endpointStroke/non-CNS systemic embolism is the primary efficacy composite
Component analysisStroke and non-CNS systemic embolism are separately analyzed
Secondary endpointsVascular death, myocardial infarction, and all-cause mortality are also analyzed
Serious adverse eventsReported descriptively by treatment arm

27. A Statistical Walkthrough of the Primary Efficacy Result

The primary non-inferiority result can be understood as a sequence of statistical decisions rather than as a single number.

Step 1 · Design

Define the unfavorable boundary

The registry specifies a non-inferiority margin of 1.46 in hazard ratio and a one-sided alpha of 0.025.

Step 2 · Analysis population

Use the per-protocol population

The posted primary non-inferiority analysis is based on the per-protocol population.

Step 3 · Model

Fit the Cox proportional-hazards model

Treatment is included as a covariate in the reported non-stratified Cox proportional-hazards model.

Step 4 · Estimate

Obtain the hazard ratio

The estimated HR is 0.79, with a 95% CI of 0.66–0.96.

Step 5 · Hypothesis test

Evaluate non-inferiority

The posted p-value is <0.001. The interpretation is made against the prespecified non-inferiority framework rather than treating HR = 1 as the sole decision boundary.

28. What the Registry Does Not Report Here

The ClinicalTrials.gov record is rich enough to reconstruct the main statistical architecture and all 10 posted statistical analyses, but they do not provide several details that would be necessary for a fuller reconstruction of the original statistical analysis plan.

TopicAvailable in the ClinicalTrials.gov record?
Primary HR estimatesYes
95% confidence intervalsYes
P-valuesYes
Analysis populationsYes, with some definitions truncated
Non-inferiority marginYes: 1.46
One-sided alphaYes: 0.025
Kaplan-Meier numerical estimatesNo
Median survival/event timesNo
Event counts by arm for primary endpointsNo
Baseline characteristicsNo
Subgroup estimatesNo
Interaction testsNo
Detailed missing-data strategyNo
Detailed interim-analysis planNo
Multiplicity-adjustment procedureNo
Bayesian methodsNo

These omissions are important because they define the boundary between what can be independently reconstructed from the registry and what would require additional protocol, statistical analysis plan, or underlying patient-level information.

29. Related Tutorials

Learn more about the methods used in this trial:

30. Related Statistical Calculators

31. Sources

Continue through the Clinical Biostats statistical pathway

Use the trial's endpoints and methods as a practical route into survival analysis, hazard ratios, confidence intervals, and non-inferiority trial design.

32. Record Summary

ROCKET AF provides a detailed example of how a randomized clinical trial can use a time-to-event framework to address both non-inferiority and superiority questions. The primary efficacy endpoint was analyzed first under a non-inferiority framework using a per-protocol population, a hazard-ratio margin of 1.46, and a one-sided alpha of 0.025. The posted estimate was HR 0.79 with a 95% CI of 0.66–0.96 and P < 0.001.

The same efficacy composite was subsequently analyzed for superiority in the safety population, producing HR 0.79 with a 95% CI of 0.65–0.95 and P = 0.015. The primary safety endpoint, major/non-major clinically relevant bleeding, produced HR 1.03 with a 95% CI of 0.96–1.11 and P = 0.442. Seven secondary time-to-event analyses provide additional hazard-ratio estimates for composite and individual cardiovascular outcomes.

Clinical Biostats methodology: the most informative reading of ROCKET AF is not a single p-value or hazard ratio. The statistical story depends on the endpoint definition, time origin, on-treatment framework, analysis population, Cox-model interpretation, non-inferiority margin, confidence interval, and distinction between primary and secondary hypotheses. Where the ClinicalTrials.gov record does not provide further methodological detail, this page does not infer it.