← Clinical Trials
Atrial Fibrillation Phase 3 Stroke Prevention NCT00412984

ARISTOTLE: Complete Statistical Analysis of Apixaban in Atrial Fibrillation

An independent statistical analysis of the randomized phase 3 ARISTOTLE trial comparing apixaban with warfarin for the prevention of stroke and systemic embolism in participants with atrial fibrillation and atrial flutter.

NCT00412984 · COMPLETED · RANDOMIZED · DOUBLE-MASKED · PARALLEL DESIGN
Independent analysis: 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

ARISTOTLE was a randomized, double-masked, parallel-group phase 3 prevention trial enrolling 20,976 participants. The registered comparison was apixaban versus warfarin, and the posted primary analysis used an intention-to-treat population and a Cox proportional-hazards model.

20,976
Enrollment
Randomized participants
2
Arms
Apixaban and warfarin
0.79
Primary HR
95% CI .66–.95
0.0114
Primary P-value
Two-sided
Trial
ARISTOTLE — Apixaban for the Prevention of Stroke in Subjects With Atrial Fibrillation
Phase
3
Therapeutic area
Cardiology
Conditions
Atrial Fibrillation; Atrial Flutter
Allocation
Randomized
Masking
Double
Design model
Parallel
Primary purpose
Prevention

2. Clinical Question

Population

The trial enrolled participants with atrial fibrillation or atrial flutter. The registry reports an enrollment of 20,976 participants.

Intervention

The intervention was apixaban.

Comparator

The comparator was warfarin.

Primary statistical question

The principal registered efficacy question concerned the time to the first event of ischemic/unspecified stroke, hemorrhagic stroke, or systemic embolism during the intended treatment period, together with the corresponding rate of adjudicated stroke or systemic embolism.

Statistically, the posted primary comparison estimated the relative hazard of the adjudicated stroke or systemic embolism rate for apixaban versus warfarin.

3. Trial Design

Step 1
Randomization
Step 2
Apixaban
Step 3
Warfarin
Step 4
Follow-up
Step 5
Time-to-event analysis

ARISTOTLE used a randomized, double-masked, parallel design. Its primary purpose was prevention. The registry reports two intervention arms and 20,976 enrolled participants.

Apixaban

INTERVENTION ARM
  • Drug intervention
  • Compared with warfarin
  • Included in the intention-to-treat primary efficacy analysis

Warfarin

COMPARATOR ARM
  • Drug comparator
  • Compared with apixaban
  • Included in the intention-to-treat primary efficacy analysis

Trial timeline

2006-12-31

Trial start

The registry reports the study start date as December 31, 2006.

2011-05-25

Primary completion

The registry reports primary completion on May 25, 2011.

4. Endpoints

Primary endpoint 1: first efficacy event

Registered endpoint

Number of Participants With First Event of Ischemic/Unspecified Stroke, Hemorrhagic Stroke, or Systemic Embolism (SE) During the Intended Treatment Period

Time frame: Time to first event in "Intended Treatment Period": started on day of randomization, ended at efficacy cut-off date.

The registry states that all suspected efficacy events were adjudicated by the Central Events Committee (CEC). Its stroke definition is a nontraumatic focal neurological deficit lasting at least 24 hours and includes ischemic stroke, hemorrhagic stroke, ischemic stroke with hemorrhagic conversion, stroke of uncertain type, and retinal ischemic event (embolism, infarction). The registry defines systemic embolism using a clinical history consistent with an acute loss of blood flow to an arterial site.

Primary endpoint 2: event rate

Registered endpoint

Rate of Adjudicated Stroke or Systemic Embolism (SE) During the Intended Treatment Period

Time frame: "Intended Treatment Period" started on the day of randomization and ended at the efficacy cut-off date.

Unit: Number of adjudicated stroke or SE events per 100 patient years.

The registry therefore describes both an event-based formulation and a rate formulation of the principal efficacy outcome. The posted formal statistical analysis is for the rate of adjudicated stroke or systemic embolism and treats it as a time-to-event endpoint.

Secondary endpoints represented in the posted analyses

EndpointTime framePopulation
Rate of Adjudicated Major (ISTH) Bleed Events During Treatment PeriodTreatment Period; first dose of blinded study drug through 2 days after last doseTreated participants
Rate of Adjudicated All-Cause Death During the Intended Treatment PeriodIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Ischemic or unspecified strokeIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Hemorrhagic strokeIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Systemic embolismIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Myocardial infarctionIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Composite stroke / systemic embolism / major bleeding in warfarin/VKA-naive participantsIntended Treatment PeriodIntention-to-treat analysis, randomized participants
Major or clinically relevant non-major bleedTreatment PeriodTreated participants
All bleeding eventsTreatment PeriodTreated participants
GUSTO bleeding endpointsTreatment PeriodTreated participants
TIMI bleeding endpointsTreatment PeriodTreated participants
Net-clinical benefitTreatment PeriodTreated participants

5. Statistical Methodology

Primary analysis

The posted primary analysis used a Cox proportional-hazards model for the rate of adjudicated stroke or systemic embolism during the intended treatment period. The analysis population was the intention-to-treat population of randomized participants.

Primary statistical model

The model compared apixaban versus warfarin and reported a hazard ratio (HR).

HR = instantaneous event rate under apixaban / instantaneous event rate under warfarin

The posted estimate was 0.79, with a two-sided 95% confidence interval of .66 to .95 and P = 0.0114.

Covariate adjustment and stratification

The analysis text identifies covariate adjustment and stratified analysis as concepts in the primary analysis. For the posted major-bleeding analysis, the registry explicitly states that the model included treatment group as a covariate and was stratified by investigative site and prior warfarin/vitamin K antagonist status. The registry also identifies covariate adjustment and stratified analysis in the primary efficacy analysis.

Intention-to-treat analysis

The primary efficacy analysis was based on randomized participants. Participants who did not experience an efficacy endpoint event were censored according to the registry's stated censoring rules, including the earlier of death date when death was not part of the endpoint, last contact, or the applicable efficacy cut-off.

This distinction matters because an intention-to-treat analysis preserves the randomized comparison. Censoring, meanwhile, determines which portions of follow-up contribute information to the time-to-event model.

Analysis of bleeding endpoints

The registry used the same general survival-analysis framework for several bleeding outcomes. The major ISTH bleeding analysis used treated participants and censored participants without a bleeding endpoint at the earlier of 2 days after discontinuation of study drug, death date, or last-contact date, according to the registry description.

Multiplicity and hierarchical testing

Closed testing matters: The registry states that four key objectives were tested using a closed testing procedure. Non-inferiority for the primary efficacy was tested first. If non-inferiority was demonstrated, superiority for the primary efficacy was tested; subsequent superiority testing depended on the preceding result.

The important statistical point is that the subsequent superiority P-values should not be interpreted in isolation from the prespecified testing sequence. The registry explicitly describes a hierarchical closed-testing framework.

6. Primary Result: Stroke or Systemic Embolism

Apixaban vs Warfarin

HR 0.79

Two-sided 95% CI: .66–.95  ·  P = 0.0114

Analysis: Cox proportional-hazards model  ·  Population: intention-to-treat randomized participants

The primary posted analysis estimated a hazard ratio of 0.79 for adjudicated stroke or systemic embolism for apixaban relative to warfarin. In relative terms, an HR of 0.79 corresponds to a 21% lower estimated hazard under the fitted model, because 1 − 0.79 = 0.21.

Clinical Biostats interpretation

The HR of 0.79 describes a relative difference in the instantaneous rate of the modeled time-to-event outcome. It does not mean that 21% of participants avoided an event, that every individual had exactly a 21% reduction in risk, or that cumulative event probability was reduced by exactly 21% at every time point.

The two-sided 95% CI of .66–.95 describes statistical uncertainty around the estimated hazard ratio under the model and analysis framework. It does not describe the range of treatment effects experienced by individual participants.

The P-value of 0.0114 is evidence against the relevant null hypothesis within the prespecified testing framework; it is not a measure of effect size. The magnitude of the estimated relative effect is communicated by the HR, while the confidence interval communicates its precision.

Because this is a Cox proportional-hazards analysis, interpretation of a single HR also depends on the proportional-hazards model being an appropriate summary of the relative event hazards over follow-up. The registry's use of censoring means that participants contribute information only until the specified censoring time or event.

7. Secondary Results: Major Bleeding

Adjudicated Major (ISTH) Bleed Events

HR 0.69

Two-sided 95% CI: 0.60–0.80  ·  P < .0001

Analysis: Cox proportional-hazards model  ·  Population: treated participants

The posted analysis estimated a hazard ratio of 0.69 for adjudicated major ISTH bleed events during the treatment period, comparing apixaban with warfarin. The corresponding two-sided 95% CI was 0.60–0.80, with P < .0001.

Clinical Biostats interpretation

An HR of 0.69 corresponds to a 31% lower estimated hazard of the analyzed major-bleeding endpoint under the fitted model. This is a relative hazard interpretation, not a statement that 31% of patients avoided bleeding or that absolute bleeding probability changed by 31 percentage points.

The 95% CI of 0.60–0.80 gives the statistical uncertainty around the estimated HR. The interval is narrower than the primary efficacy interval in absolute HR units, but precision should always be considered in relation to the endpoint, follow-up, and analysis population.

The P-value of < .0001 does not itself quantify the magnitude of the difference. The HR and confidence interval provide that information. This endpoint was analyzed among treated participants rather than the randomized ITT population, which is an important distinction when comparing the statistical analyses across endpoints.

8. Secondary Results: All-Cause Death

All-Cause Death

HR 0.89

Two-sided 95% CI: 0.80–1.00  ·  P = 0.0465

Analysis: Cox proportional-hazards model  ·  Population: intention-to-treat randomized participants

The estimated hazard ratio for adjudicated all-cause death during the intended treatment period was 0.89. The two-sided 95% CI was 0.80–1.00, and the posted P-value was 0.0465.

Clinical Biostats interpretation

The HR of 0.89 corresponds to an estimated 11% lower hazard under the fitted model. The confidence interval reaches 1.00, illustrating why the exact interval should be considered alongside the point estimate rather than focusing only on the P-value.

The P-value of 0.0465 is a hypothesis-testing quantity. It does not say that there is a 4.65% probability that the treatment effect is due to chance, nor does it measure clinical importance. The HR and its confidence interval remain the primary descriptive measures of the estimated relative effect.

This endpoint was analyzed within the same broad intended-treatment time-to-event framework as the primary efficacy analysis, but it was a secondary endpoint. The registry also describes a closed-testing sequence for the key objectives, so the multiplicity framework matters when interpreting secondary superiority results.

9. Additional Secondary Efficacy Results

The registry contains several additional time-to-event analyses. The estimates below are reported exactly as posted and should be interpreted as hazard ratios for apixaban versus warfarin.

OutcomeHR95% CIP-value
Ischemic or unspecified stroke0.920.74–1.130.4220
Hemorrhagic stroke0.510.35–0.750.0006
Systemic embolism0.870.44–1.750.7020
Myocardial infarction0.880.66–1.170.3720
Stroke / systemic embolism / major bleeding0.770.69–0.86< .0001
Stroke / systemic embolism / all-cause death0.890.81–0.980.0192
Stroke / systemic embolism / major bleeding / all-cause death0.850.78–0.920.0002
Stroke / systemic embolism / MI / all-cause death0.880.80–0.970.0107
Ischemic or unspecified stroke / all-cause death0.900.82–1.000.0432
Hemorrhagic stroke / all-cause death0.880.79–0.980.0167
Systemic embolism / all-cause death0.890.80–1.000.0464
Myocardial infarction / all-cause death0.890.80–0.990.0253
Stroke / systemic embolism / major bleeding in warfarin/VKA-naive participants0.800.67–0.950.0098
How to read the secondary table

These estimates are not interchangeable. Each HR refers to a different endpoint, and several endpoints are composites containing different clinical events. An HR of 0.51 for hemorrhagic stroke, for example, describes a different outcome from an HR of 0.92 for ischemic or unspecified stroke.

The width of the confidence interval also varies substantially. The systemic-embolism estimate of 0.87 has a 95% CI of 0.44–1.75, illustrating considerably greater uncertainty than some of the composite estimates. A wide interval can occur when relatively fewer events contribute information to the analysis.

Because multiple secondary outcomes were evaluated, the individual P-values should not automatically be treated as though each were an isolated primary hypothesis test. The registry specifically documents a closed testing procedure for four key objectives.

10. Bleeding and Net-Clinical-Benefit Analyses

EndpointHR95% CIP-value
Major or clinically relevant non-major (CRNM) bleed0.680.61–0.75< .0001
All bleeding events0.710.68–0.75< .0001
Severe GUSTO bleeding events0.460.35–0.60< .0001
Severe or moderate GUSTO bleeding events0.600.50–0.71< .0001
Major TIMI bleeding event0.570.46–0.70< .0001
Major or minor TIMI bleeding criteria0.630.54–0.75< .0001
Net-clinical benefit0.740.65–0.83< .0001

The registry identifies these analyses as time-to-event analyses using Cox proportional-hazards models. The GUSTO and TIMI outcomes are separately defined bleeding classifications, so their HRs should be interpreted within their respective endpoint definitions.

The net-clinical-benefit analysis is likewise a composite endpoint rather than a generic measure of overall benefit. Its HR of 0.74 summarizes the specific events included in the registry's definition; it should not be interpreted as a universal treatment-effect measure across all clinical outcomes.

11. Safety: Serious Adverse Events

Serious adverse events by arm
Warfarin
3182
Apixaban
3302
ArmAffectedAt risk
Warfarin31829088
Apixaban33029052

The registry reports serious adverse events by arm as 3182/9088 for warfarin and 3302/9052 for apixaban. These are affected-participant counts and at-risk denominators as reported in the ClinicalTrials.gov record.

Safety interpretation: Serious adverse-event counts should not be interpreted as equivalent to the time-to-event bleeding analyses. The serious-adverse-event summary and the Cox analyses answer different statistical questions and use different endpoint definitions and, for the bleeding analyses, different analysis populations.

12. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

The registered outcomes are time-to-event endpoints. Rather than reducing follow-up to a simple yes/no event indicator, a survival model uses information about when events occur and accounts for censoring. The Cox model expresses the treatment comparison through a hazard ratio.

What does an HR of 0.79 mean?

An HR of 0.79 means that the fitted model estimates the instantaneous event hazard under apixaban at 79% of the corresponding hazard under warfarin, giving a relative hazard reduction of 21%. It does not mean a 21% absolute reduction in the probability of stroke or systemic embolism.

Why is the confidence interval important?

The 95% CI of .66–.95 shows the statistical uncertainty surrounding the primary HR estimate. The point estimate alone can make an effect appear more precise than the data justify. The interval gives readers a range of values compatible with the statistical estimation framework used for the analysis.

Why does the P-value not measure effect size?

A P-value addresses compatibility with a null hypothesis under the specified statistical framework. It is affected by the amount of information in the study as well as by the magnitude of the observed effect. The HR communicates relative effect size; the confidence interval communicates both the estimate and its precision.

Why does intention-to-treat analysis matter?

The primary efficacy analysis included randomized participants. Maintaining participants in their randomized groups preserves the treatment comparison established by randomization. This is particularly important in a randomized trial because changing the analysis population based on post-randomization events can compromise the comparability created at randomization.

Why are censoring rules important?

A participant who does not experience the endpoint does not necessarily contribute follow-up indefinitely. The registry specifies censoring rules for the primary and bleeding analyses. Cox models rely on these observed follow-up intervals, so the definition of when observation ends is part of the statistical analysis rather than an administrative detail.

Why does multiplicity matter?

ARISTOTLE's registry analysis notes describe a closed testing procedure for four key objectives, beginning with non-inferiority for the primary efficacy endpoint and then proceeding conditionally to superiority testing. This means the sequence of hypotheses is part of the inferential design. A collection of P-values should therefore not be read as though every endpoint had been designated an independent primary test.

13. Understanding the Non-Inferiority Logic

The primary analysis notes state that, with 448 subjects with confirmed strokes or systemic emboli, the study would have at least 90% power to meet both regulatory definitions of non-inferiority described in the registry.

The first stated non-inferiority criterion was that the upper bound of the two-sided 95% CI for relative risk be less than 1.38. The registry then describes superiority testing as a subsequent step after non-inferiority is demonstrated.

Key distinction
Non-inferiority: upper CI bound < 1.38

The non-inferiority margin is not the same as the null value of 1.00. The margin specifies how much relative loss of efficacy could be accepted for the non-inferiority claim, under the prespecified framework.

This distinction is central to reading the statistical design correctly. A treatment can be evaluated first against a non-inferiority margin and subsequently against the conventional superiority null. Those are different hypotheses and should not be collapsed into one generic "significant/not significant" decision.

14. What the Hazard Ratio Does—and Does Not—Tell You

What it tells you

The HR summarizes the relative event hazard estimated by the Cox model. Values below 1 indicate a lower modeled hazard for apixaban than for warfarin.

What it does not tell you

The HR is not an absolute risk difference, does not give the probability that an individual will experience an event, and does not mean the relative difference is identical at every time point.

Why the CI matters

The confidence interval indicates uncertainty around the estimated relative hazard and should be considered alongside the point estimate.

Why the P-value matters differently

The P-value addresses the statistical test; it is not a scale for clinical magnitude and should not replace examination of the HR and CI.

15. Analysis Populations: A Crucial Detail

One of the most important features of the ARISTOTLE registry results is that the analysis population is not identical for every endpoint.

Analysis typePopulation reported
Primary efficacyIntention-to-treat analysis, randomized participants
All-cause deathIntention-to-treat analysis, randomized participants
Individual stroke, systemic embolism and MI analysesIntention-to-treat analysis, randomized participants
Composite efficacy analysesIntention-to-treat analysis, randomized participants
Major ISTH bleedingTreated participants
Major or CRNM bleedingTreated participants
All bleeding eventsTreated participants
GUSTO and TIMI bleeding analysesTreated participants
Net-clinical benefitTreated participants

This difference is statistically consequential. An ITT analysis is anchored to randomization, whereas a treated-participant analysis conditions on receiving study treatment. The two approaches can answer related but not identical questions.

16. Intended Treatment Period vs Treatment Period

The registry uses two distinct time-frame labels. The Intended Treatment Period begins on the day of randomization and ends at the efficacy cut-off date. The Treatment Period begins with the first dose of blinded study drug and ends 2 days after the last dose of blinded study drug.

This distinction is important because efficacy and bleeding analyses are not necessarily observing the same time window. The primary efficacy endpoint uses the intended treatment period, whereas the bleeding endpoints reported here use the treatment period.

Statistical lesson: Endpoint definition includes the observation window. Two analyses can use the same Cox model and the same treatment contrast while still answering different questions if their endpoint definitions, censoring rules, or follow-up periods differ.

17. Interpreting Composite Endpoints

Several secondary analyses combine multiple events into composite endpoints. Examples in the registry include stroke/systemic embolism/major bleeding, stroke/systemic embolism/all-cause death, and stroke/systemic embolism/MI/all-cause death.

A composite HR summarizes the time to the first qualifying component of the composite. It should therefore be interpreted as an effect on the combined endpoint as defined, not as though the same HR separately describes every component.

The component results illustrate why this distinction matters. The individual endpoint estimates reported by the registry range from 0.51 for hemorrhagic stroke to 0.92 for ischemic or unspecified stroke, with different confidence intervals and P-values. The composite results therefore cannot be used as substitutes for the individual-event analyses.

18. A Statistical Reading of the Results

Statistical featureInterpretation
Randomized comparisonThe principal efficacy comparison is anchored to randomized treatment assignment.
Time-to-event endpointTiming of events and censoring contribute to the analysis.
Cox proportional-hazards modelThe treatment effect is summarized using a hazard ratio.
Primary HR 0.79The fitted model estimates a 21% lower hazard for the primary event under apixaban relative to warfarin.
95% CI .66–.95Quantifies statistical uncertainty around the primary HR.
P = 0.0114Provides the reported hypothesis-test result and does not measure effect magnitude.
Closed testingControls the interpretation of the sequence of key objectives.
Different analysis populationsEfficacy and bleeding analyses should not be treated as though they used identical populations.
Composite endpointsEach composite HR applies to its specified combination of events.

19. Limitations and Interpretation Issues

20. Why This Trial Matters Statistically

ARISTOTLE is a useful statistical teaching case because it combines randomized treatment assignment with large-scale time-to-event analysis, adjudicated clinical endpoints, Cox modeling, intention-to-treat analysis, stratification, non-inferiority logic, hierarchical testing, and multiple safety and efficacy outcomes.

ConceptHow it appears in ARISTOTLE
Randomization20,976 participants were enrolled in a randomized two-arm parallel design.
BlindingThe registered design is double-masked.
Intention-to-treat analysisThe primary efficacy analysis used randomized participants.
Time-to-event endpointThe primary rate outcome is analyzed as time to adjudicated stroke or systemic embolism.
Cox modelThe primary analysis and multiple secondary analyses use Cox proportional-hazards models.
Hazard ratioThe primary estimate is HR 0.79 for apixaban versus warfarin.
Confidence intervalThe primary two-sided 95% CI is .66–.95.
P-valueThe primary reported P-value is 0.0114.
Covariate adjustmentThe registry identifies covariate adjustment as part of the analysis methodology.
Stratified analysisThe registry identifies stratified analysis; the major-bleeding analysis specifies stratification by investigative site and prior warfarin/VKA status.
Non-inferiorityThe analysis notes specify a relative-risk upper-bound criterion of 1.38 for one regulatory definition.
Closed testingFour key objectives were tested using a closed testing procedure.
Composite endpointsMultiple secondary analyses combine stroke, systemic embolism, bleeding, MI and/or all-cause death.
Safety analysisSerious adverse events are reported by treatment arm, alongside multiple time-to-event bleeding outcomes.

21. Related Tutorials

Learn more about the methods used in this trial:

22. Related Statistical Calculators

23. Sources

Continue through the Clinical Biostats statistical pathway

Connect the trial's endpoints and methods to deeper statistical tutorials, calculators, and clinical-trial analysis workflows.

24. Record Summary

ARISTOTLE provides a detailed example of randomized time-to-event analysis. The trial used a randomized, double-masked, parallel design with 20,976 participants and compared apixaban with warfarin. Its primary posted statistical analysis used an intention-to-treat population and a Cox proportional-hazards model, producing an HR of 0.79 with a two-sided 95% CI of .66–.95 and P = 0.0114. The registry also reports a non-inferiority framework followed by conditional superiority testing, multiple secondary efficacy and safety analyses, and distinct analysis populations for efficacy and bleeding endpoints.

The most important statistical lesson is that the headline HR is only one part of the analysis. Proper interpretation requires attention to the endpoint definition, observation period, censoring rules, analysis population, model, confidence interval, testing sequence, and distinction between individual and composite outcomes.

Clinical Biostats methodology: A trial-results page should not merely repeat the abstract. The goal is to reconstruct the statistical story of the trial while clearly separating reported evidence from educational interpretation and avoiding claims not supported by the ClinicalTrials.gov record.