This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record. Numerical results on this page are restricted to the trial data reported in that registry record.
1. Trial at a Glance
ENGAGE AF-TIMI 48 was a randomized, parallel-group, quadruple-masked phase 3 trial evaluating two edoxaban dosing regimens against warfarin in patients with atrial fibrillation, with prevention of stroke and embolism as the primary purpose.
| Feature | ENGAGE AF-TIMI 48 |
|---|---|
| Phase | Phase 3 |
| Condition | Stroke; Atrial Fibrillation; Embolism |
| Design | Randomized, parallel-group, quadruple-masked |
| Allocation | Randomized |
| Primary purpose | Prevention |
| Enrollment | 21105 |
| Primary endpoint | Composite of stroke and systemic embolic events (SEE) |
| Primary analysis framework | Non-inferiority, followed by superiority analysis |
| ClinicalTrials.gov | NCT00781391 |
| Lead sponsor | Daiichi Sankyo |
2. Clinical Question
The registered clinical question was whether edoxaban could be compared with warfarin for the time to first occurrence of the composite of stroke and systemic embolic events (SEE), first using a non-inferiority framework and, for the registered superiority endpoint, an intention-to-treat superiority analysis.
Population
Patients enrolled in the ENGAGE AF-TIMI 48 phase 3 trial for the conditions registered as stroke, atrial fibrillation, and embolism.
Intervention
Edoxaban tablets using a high-dose 60 mg regimen or a low-dose 30 mg regimen, with placebo warfarin used as part of the masked design.
Comparator
Warfarin tablets with placebo edoxaban.
Primary question
How do the two edoxaban regimens compare with warfarin for time to first stroke or systemic embolic event?
3. Trial Design
Warfarin / placebo edoxaban
- Warfarin tablets
- Placebo edoxaban
- Reference group for the reported pairwise comparisons
High-dose edoxaban
- Edoxaban tablets
- High-dose regimen: 60 mg
- Placebo warfarin
Low-dose edoxaban
- Edoxaban tablets
- Low-dose regimen: 30 mg
- Placebo warfarin
Three-group comparison
- Two prespecified edoxaban-versus-warfarin comparisons
- Primary efficacy endpoint: stroke/SEE
- Time-to-event statistical framework
4. Endpoints
Primary efficacy endpoint
The registered primary endpoint was the composite of stroke and Systemic Embolic Events (SEE). The registry specifies an on-treatment period and an overall study period, each described as 2.5 years of median study drug exposure and 2.8 years of median follow-up.
| Endpoint | Analysis period | Analysis population | Purpose |
|---|---|---|---|
| Composite of stroke and SEE | On-treatment | mITT and PP | Non-inferiority; mITT on-treatment considered primary |
| Composite of stroke and SEE | Overall study period | mITT and PP | Non-inferiority |
| Composite of stroke and SEE | Overall study period | ITT | Superiority analysis |
Secondary endpoints reported in the statistical analyses
- Composite of stroke, systemic embolic event (SEE), and cardiovascular (CV) mortality.
- Major adverse cardiac event (MACE): composite of non-fatal MI, non-fatal stroke, non-fatal SEE, and death due to CV cause or bleeding.
- Composite of stroke, SEE, and all-cause mortality.
- Adjudicated bleeding events.
- Major or clinically relevant non-major bleeding events.
5. Statistical Methodology
The registry identifies the Cox proportional-hazards model and the log-rank test as the principal statistical methods. Because the principal outcomes are time-to-event endpoints, the analysis is concerned not only with whether an event occurred, but also with the time until the first event while accounting for censoring.
Primary non-inferiority framework
The primary efficacy endpoint was first compared between each edoxaban regimen and warfarin using the mITT analysis set during the on-treatment period. The registry states that the non-inferiority analysis also included PP analysis sets and the overall study period.
For the reported non-inferiority criterion, an upper confidence-limit below 1.38 was considered sufficient to establish non-inferiority for the edoxaban treatment group.
Multiplicity control
The registry analysis notes that two edoxaban-versus-warfarin non-inferiority comparisons were performed concurrently. To control the study-wise type-I error rate of two-sided α=0.05 for non-inferiority, each comparison was performed at a two-sided significance level of α=0.025.
This matters because two simultaneous primary comparisons create a multiplicity problem if each is evaluated at the ordinary two-sided 0.05 level. The reported analysis instead allocates the error rate across the two comparisons.
Analysis populations
- mITT: randomized subjects who received 1 or more dose of study drug; the on-treatment period included time while taking study drug and up to 3 days.
- PP: randomized subjects who received at least 1 dose of randomized study drug and did not have major protocol violations.
- ITT: all randomized subjects, whether or not they received a single dose of randomized study drug.
- Safety set: used for the reported on-treatment bleeding analyses.
6. Results: Primary Non-Inferiority Analyses
The registry reports nine primary statistical analyses: eight non-inferiority analyses covering mITT and PP populations across on-treatment and overall study periods, plus one ITT superiority analysis.
6.1 mITT on-treatment: high-dose edoxaban vs warfarin
Primary mITT on-treatment comparison
97.5% CI .632 to .985 · P <.0001
High Dose Edoxaban/Placebo Warfarin vs Warfarin/Placebo Edoxaban
The estimated hazard ratio of 0.79 indicates an estimated hazard in the high-dose edoxaban group that was 21% lower than the warfarin group's estimated hazard during the specified on-treatment period.
What the estimate means: HR 0.79 is a relative time-to-event measure. Under the model, it represents the estimated ratio of the instantaneous event hazards for high-dose edoxaban versus warfarin.
What it does not mean: It does not mean that 21% fewer participants necessarily experienced an event, nor does it mean that an individual participant's risk was reduced by exactly 21%.
Precision: The 97.5% confidence interval ranges from .632 to .985. For the registry's non-inferiority framework, the upper limit is below the stated 1.38 criterion.
P-value: P <.0001 indicates strong statistical evidence under the specified testing framework; it is not a measure of the size or clinical importance of the hazard ratio.
Caution: This result is based on the mITT on-treatment population, which the registry identifies as the primary non-inferiority analysis. Time-to-event inference also depends on the handling of censoring and the proportional-hazards model.
6.2 mITT on-treatment: low-dose edoxaban vs warfarin
Primary mITT on-treatment comparison
97.5% CI .874 to 1.314 · P .0055
Low Dose Edoxaban/Placebo Warfarin vs Warfarin/Placebo Edoxaban
An HR of 1.07 corresponds to an estimated hazard 7% higher in the low-dose edoxaban group than in the warfarin group under the Cox model. That direction should not be confused with the non-inferiority question: the relevant comparison for non-inferiority is the upper confidence limit relative to the prespecified margin.
What the estimate means: HR 1.07 is the estimated ratio of the event hazards for low-dose edoxaban versus warfarin.
What it does not mean: It does not imply that the probability of an event was exactly 7% higher for an individual patient, because a hazard ratio is not a simple risk ratio.
Precision: The 97.5% CI extends from .874 to 1.314. The registry's non-inferiority framework evaluates the upper confidence limit against the stated non-inferiority criterion rather than asking whether the interval excludes 1.
P-value: P .0055 reflects the specified statistical test and should not be interpreted as the magnitude of the treatment effect.
Caution: The result is an on-treatment mITT analysis, and interpretation should remain tied to the prespecified non-inferiority framework rather than a conventional superiority interpretation.
6.3 Overall study period: high-dose edoxaban vs warfarin
mITT overall-study comparison
97.5% CI .719 to 1.029 · P <.0001
High Dose Edoxaban/Placebo Warfarin vs Warfarin/Placebo Edoxaban
For the overall study period, the high-dose regimen had an estimated hazard ratio of .86 relative to warfarin. The registry explicitly states that an upper confidence limit below 1.38 established non-inferiority for the edoxaban treatment group.
Estimate: HR .86 corresponds to an estimated 14% lower hazard under the Cox model for high-dose edoxaban relative to warfarin during the overall study period.
Precision: The 97.5% CI is .719 to 1.029. Although the interval includes 1, its upper limit remains below the registry's stated non-inferiority margin of 1.38.
P-value: P <.0001 supplies evidence under the specified non-inferiority testing procedure; it does not quantify how large or clinically important the effect is.
Caution: This is an overall-study-period analysis rather than the primary mITT on-treatment analysis. The estimand therefore encompasses the registry's broader study-period definition.
6.4 Overall study period: low-dose edoxaban vs warfarin
mITT overall-study comparison
97.5% CI .955 to 1.336 · P .0074
Low Dose Edoxaban/Placebo Warfarin vs Warfarin/Placebo Edoxaban
The low-dose regimen produced an estimated hazard ratio of 1.13 over the overall study period. The registry's non-inferiority criterion focuses on whether the upper confidence limit exceeds 1.38; the reported upper limit was 1.336.
Estimate: HR 1.13 represents an estimated 13% higher hazard for low-dose edoxaban relative to warfarin under the fitted Cox model.
Precision: The 97.5% CI ranges from .955 to 1.336. Its upper limit remains below the stated 1.38 non-inferiority criterion.
P-value: P .0074 belongs to the prespecified statistical comparison and should not be used as a substitute for examining the hazard ratio and its confidence interval.
Caution: Non-inferiority is a different question from superiority. An interval that includes 1 can still satisfy a non-inferiority criterion if the entire relevant confidence boundary remains within the prespecified margin.
7. Results: Per-Protocol Sensitivity Analyses
The registry reports PP analyses alongside the mITT analyses. This is particularly relevant for non-inferiority studies because departures from assigned treatment or major protocol violations can affect the interpretability of the treatment contrast.
| Population | Period | Comparison | HR | 97.5% CI | P-value |
|---|---|---|---|---|---|
| PP | On-treatment | High-dose edoxaban vs warfarin | 0.79 | .634–.989 | <.0001 |
| PP | On-treatment | Low-dose edoxaban vs warfarin | 1.08 | .878–1.32 | .0064 |
| PP | Overall study | High-dose edoxaban vs warfarin | .86 | .720–1.032 | <.0001 |
| PP | Overall study | Low-dose edoxaban vs warfarin | 1.13 | .958–1.34 | .0084 |
The PP results are not a second, unrelated clinical trial. They are a sensitivity perspective on the same randomized comparison after restricting the analysis set to participants who received at least 1 dose and did not have major protocol violations. In a non-inferiority analysis, examining both mITT and PP populations can be informative because the two populations can be affected differently by treatment adherence and protocol deviations.
The numerical pattern is similar across the corresponding mITT and PP analyses: the high-dose estimates are below 1, while the low-dose estimates are above 1. The confidence limits for the reported PP analyses remain below the stated 1.38 non-inferiority criterion.
8. Results: Registered Superiority Analysis
High-dose edoxaban vs warfarin for stroke and SEE
ITT superiority analysis
99% CI .709 to 1.068 · P .081
High Dose Edoxaban/Placebo Warfarin vs Warfarin/Placebo Edoxaban
The registered superiority analysis used the ITT analysis set and a log-rank test. The reported hazard ratio was 0.87 with a 99% confidence interval of .709 to 1.068 and P .081.
Estimate: HR 0.87 corresponds to an estimated 13% lower hazard for high-dose edoxaban relative to warfarin under the reported time-to-event model.
What the interval says: The 99% confidence interval extends from .709 to 1.068. Because it includes 1, the interval is compatible with no hazard-ratio difference as well as with hazard ratios on either side of 1 within that interval.
P-value: P .081 is the reported superiority-test result. A p-value is evidence against a specified null hypothesis under the statistical model; it is not a probability that the treatment effect is zero and it does not measure effect size.
Important distinction: The superiority analysis addresses a different question from the earlier non-inferiority analyses. Satisfying a non-inferiority criterion does not automatically establish superiority.
9. Secondary Endpoint Results
Stroke, SEE, and cardiovascular mortality
High-dose edoxaban vs warfarin
95% CI .786 to .959 · P .0053
This secondary endpoint used the ITT population over the overall study period. The registry states that time to first event was estimated by a Kaplan-Meier estimate and compared between the high-dose edoxaban and warfarin groups using a log-rank test at a pairwise comparison significance level of α=0.01.
MACE
High-dose edoxaban vs warfarin
95% CI .806 to .972 · P .0109
MACE was defined in the registry as a composite of non-fatal MI, non-fatal stroke, non-fatal SEE, and death due to CV cause or bleeding. The analysis used the ITT population and the overall study period, with Kaplan-Meier estimation and a log-rank comparison at a pairwise significance level of α=0.01.
Stroke, SEE, and all-cause mortality
High-dose edoxaban vs warfarin
95% CI .823 to .981 · P .0168
This secondary time-to-event endpoint was also analyzed in the ITT population over the overall study period using Kaplan-Meier estimation and a log-rank test with a pairwise significance level of α=0.01.
All three reported secondary efficacy estimates are below 1, but each represents a different composite endpoint. Composite endpoints can be statistically efficient because several types of events contribute to the analysis, but the resulting hazard ratio applies to the composite rather than automatically to every component individually. The ClinicalTrials.gov record does not provide component-specific estimates, so no component-level conclusion should be inferred from these composite HRs.
The use of a pairwise significance level of α=0.01 is also important. The reported p-values should be interpreted within that stated testing framework rather than against an unstated generic threshold.
10. Safety Results
The registry reports serious adverse events by treatment arm and separate statistical analyses of adjudicated bleeding events. The serious-adverse-event counts below are affected/at-risk counts as reported.
| Arm | Serious adverse events | Affected / at risk |
|---|---|---|
| Warfarin / placebo edoxaban | Serious adverse events | 3581 / 7012 |
| High-dose edoxaban / placebo warfarin | Serious adverse events | 3393 / 7012 |
| Low-dose edoxaban / placebo warfarin | Serious adverse events | 3448 / 7002 |
Adjudicated bleeding events
| Endpoint | Comparison | HR | 95% CI | P-value |
|---|---|---|---|---|
| Adjudicated bleeding events | High-dose edoxaban vs warfarin | .80 | .707–.914 | .0009 |
| Adjudicated bleeding events | Low-dose edoxaban vs warfarin | 0.47 | .406–.548 | <.0001 |
| Major or clinically relevant non-major bleeding | High-dose edoxaban vs warfarin | .86 | .800–.918 | <.0001 |
| Major or clinically relevant non-major bleeding | Low-dose edoxaban vs warfarin | .62 | .575–.666 | <.0001 |
These bleeding analyses used the safety-analysis set during the on-treatment period. The low-dose adjudicated-bleeding analysis included treatment and two stratification factors as covariates: the dichotomized CHADS2 score and the dichotomized dose-adjustment factor. The registry describes the approach as a Cox regression model with a counting-process approach for on-treatment analysis.
The bleeding HRs are time-to-event effect measures rather than simple proportions. For example, an HR of 0.47 represents an estimated hazard less than half that of the warfarin comparison group under the fitted model; it does not mean that exactly 53% fewer participants necessarily experienced bleeding.
The confidence intervals quantify statistical precision around the estimated hazard ratio. The p-values provide evidence under the specified model and comparison and should not be interpreted as measures of effect magnitude.
Because these are safety-set, on-treatment analyses, their analysis population and time window differ from the ITT superiority analysis and the mITT primary efficacy analysis. Directly comparing their HRs without considering those differences would mix different estimands.
11. Statistical Methods Explained
Why was a Cox proportional-hazards model used?
The primary outcomes were time-to-event endpoints: the analysis concerned the time until the first stroke or systemic embolic event. A Cox proportional-hazards model is designed for this setting and produces a hazard ratio comparing the event hazards between treatment groups while accommodating censored observations.
What does a hazard ratio of 0.79 mean?
An HR of 0.79 means that the fitted model estimates the event hazard in the edoxaban group at 79% of the corresponding hazard in the warfarin group. It is therefore commonly described as an estimated 21% lower hazard. It is not equivalent to saying that 21% fewer patients experienced an event.
Why is the non-inferiority margin different from 1?
Superiority and non-inferiority answer different questions. A superiority analysis asks whether the data support a difference from the null value of 1 for a hazard ratio. A non-inferiority analysis asks whether the new treatment is not unacceptably worse than the comparator according to a prespecified margin. Here, the registry states that an upper confidence limit below 1.38 was considered sufficient for non-inferiority.
Why were both mITT and per-protocol analyses used?
The mITT population preserves much of the randomized comparison while requiring receipt of at least one dose. The PP population excludes participants with major protocol violations. In a non-inferiority setting, examining both populations can help assess whether conclusions are sensitive to protocol adherence and deviations.
Why was the ITT population used for the superiority analysis?
The registry defines the ITT population as all randomized participants, whether or not they received a single dose of randomized study drug. An ITT analysis generally preserves the treatment assignment created by randomization and is therefore suited to an analysis of outcomes according to randomized treatment assignment.
What is the role of the log-rank test?
The log-rank test compares survival or event-time experience between groups over follow-up. In the ENGAGE AF-TIMI 48 registry analyses, it was used for the registered superiority analysis and the reported secondary efficacy analyses, alongside hazard-ratio estimates.
Why does multiplicity matter here?
There were two concurrent edoxaban-versus-warfarin non-inferiority comparisons. Testing multiple primary comparisons without adjustment can increase the probability of a false-positive conclusion across the family of tests. The registry states that the two comparisons were each performed at a two-sided α=0.025 level to control the study-wise two-sided α=0.05 level for non-inferiority.
12. How to Read the ENGAGE AF-TIMI 48 Results
The statistical structure of this trial is more informative than any single p-value. There are several layers to keep separate:
First: the estimand
The primary endpoint is time to first stroke or systemic embolic event. That makes the hazard ratio a time-to-event measure rather than a simple event-rate ratio.
Second: the analysis population
The primary non-inferiority comparison is the mITT on-treatment analysis, with PP analyses also reported. The registered superiority analysis uses ITT.
Third: the margin
Non-inferiority is judged against the stated upper confidence-limit criterion of 1.38, not simply by asking whether a confidence interval excludes 1.
Fourth: multiplicity
Two concurrent edoxaban-versus-warfarin comparisons required control of the study-wise type-I error rate, reflected in the reported two-sided α=0.025 level for each non-inferiority comparison.
13. Limitations
- The ClinicalTrials.gov record does not include a baseline-characteristics table, subgroup forest plots, or component-specific results for the composite efficacy endpoints.
- The primary endpoint combines stroke and systemic embolic events, so the reported hazard ratio applies to the composite rather than automatically to each component.
- The non-inferiority analyses use different analysis populations and time definitions, including mITT on-treatment, PP on-treatment, and overall-study analyses. These should not be treated as identical estimands.
- The Cox proportional-hazards model relies on its proportional-hazards framework. A hazard ratio should therefore not be interpreted as a constant absolute risk difference over the entire study.
- The superiority analysis uses a different confidence level and testing framework from the primary non-inferiority analyses.
- The ClinicalTrials.gov record does not provide the full underlying event-time distributions or Kaplan-Meier curves, so this page cannot independently reconstruct those distributions.
- The ClinicalTrials.gov record does not provide the full statistical-analysis-plan text. Interpretation is therefore limited to the analysis descriptions reported in the registry data.
- The safety results use the safety-analysis set and an on-treatment period, which differs from the ITT overall-study analyses.
14. Why This Trial Matters Statistically
ENGAGE AF-TIMI 48 illustrates several important principles in clinical-trial statistics.
First, it shows why non-inferiority cannot be interpreted using the conventional superiority null value alone. An HR estimate can be above 1 and still satisfy a non-inferiority criterion if the upper confidence limit remains below the prespecified margin. The low-dose overall-study estimate of 1.13 provides a direct example: its 97.5% confidence interval reaches above 1 but remains below the stated 1.38 criterion.
Second, the trial demonstrates why analysis populations matter. The mITT, PP, and ITT populations answer related but not identical statistical questions. The primary mITT on-treatment analysis is explicitly identified by the registry as the primary non-inferiority analysis, while PP analyses provide an additional perspective and the ITT analysis is used for the registered superiority comparison.
Third, the trial demonstrates the importance of multiplicity control. Two concurrent comparisons of edoxaban regimens against warfarin were part of the primary non-inferiority testing framework. The reported two-sided α=0.025 level for each comparison illustrates how a study can preserve a specified family-wise type-I error rate across multiple primary comparisons.
Fourth, the trial shows the value of separating effect size, precision, and statistical evidence. The hazard ratio describes the estimated relative hazard, the confidence interval describes uncertainty around that estimate, and the p-value describes evidence under a particular hypothesis-testing framework. None of these quantities alone provides the complete statistical interpretation.
Finally, the study illustrates why time-to-event endpoints require specialized methods. Censoring, follow-up time, analysis windows, Kaplan-Meier estimation, log-rank testing, and Cox regression all form part of the inferential structure. A single crude proportion of participants with an event would discard important information about when events occurred.
15. Related Tutorials
Learn more about the methods used in this trial:
16. Related Calculators
Practice the statistical concepts underlying this trial:
Explore the statistical methods behind clinical trials
Clinical Biostats provides tutorials and statistical calculators covering survival analysis, clinical-trial methodology, confidence intervals, and related methods.
17. Sources
- ClinicalTrials.gov: ENGAGE AF-TIMI 48 (NCT00781391) — official trial registry record and the source of the trial design, endpoints, analysis populations, statistical methods, and numerical results presented on this page.
- PubMed record: PMID 40925065
- PubMed record: PMID 39791401
- PubMed record: PMID 38985461
- PubMed record: PMID 38828563
- PubMed record: PMID 37100536