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
PALLAS was a randomized, triple-masked, phase 3 parallel-group trial evaluating dronedarone versus placebo in patients with atrial fibrillation. The registry identifies two co-primary time-to-event outcomes and reports hazard ratios estimated using Cox regression alongside two-sided log-rank tests.
| Feature | PALLAS |
|---|---|
| Trial name | PALLAS |
| Brief title | Permanent Atrial fibriLLAtion Outcome Study Using Dronedarone on Top of Standard Therapy |
| Phase | Phase 3 |
| Condition | Atrial Fibrillation |
| Design | Randomized, triple-masked, parallel-group |
| Allocation | Randomized |
| Primary purpose | Treatment |
| Enrollment | 3236 |
| Arms | 2 |
| Interventions | Dronedarone; placebo for dronedarone |
| Lead sponsor | Sanofi |
| Sponsor type | Industry |
| Status | Terminated |
| Start | 2010-07 |
| Primary completion | 2011-09 |
2. Clinical Question
The statistical question was whether patients randomized to dronedarone differed from patients randomized to placebo in the time to either of two prespecified co-primary outcomes.
Population
Patients enrolled in the PALLAS phase 3 trial with atrial fibrillation.
Intervention
Dronedarone on top of standard therapy.
Comparator
Placebo for dronedarone on top of standard therapy.
Primary question
How did dronedarone compare with placebo for the two co-primary time-to-event outcomes?
3. Trial Design
Dronedarone arm
- Dronedarone
- Administered on top of standard therapy
- Included in the randomized intention-to-treat efficacy population
Placebo arm
- Placebo for dronedarone
- Administered on top of standard therapy
- Included in the randomized intention-to-treat efficacy population
The randomized treatment groups were closely balanced in size, with 1614 participants assigned to dronedarone and 1609 assigned to placebo in the serious-adverse-event denominators reported by the registry.
4. Endpoints
The trial registered three primary-endpoint entries, consisting of an overview of the two co-primary outcomes and the formal time-to-event definitions for each outcome.
| Endpoint | Definition | Time frame |
|---|---|---|
| First co-primary outcome | Time from randomization to the first event among stroke, systemic arterial embolism, myocardial infarction (MI), or cardiovascular death. | From randomization up to the CSED at study termination; maximum follow-up of 1 year. |
| Second co-primary outcome | Time from randomization to the first event among unscheduled cardiovascular hospitalization or death from any cause. | From randomization up to the CSED at study termination; maximum follow-up of 1 year. |
Both co-primary outcomes were determined through central review and adjudication by a blinded Adjudication Committee of reported deaths, MI, systemic arterial embolisms, strokes, transient ischemic attacks, heart-failure hospitalization, and unplanned cardiovascular hospitalizations.
5. Statistical Methodology
Intention-to-treat analysis
The registry identifies the intention-to-treat population as the analysis population for both primary statistical analyses. In an intention-to-treat framework, participants remain associated with their randomized treatment group when efficacy outcomes are compared. This preserves the treatment comparison created by randomization.
Kaplan-Meier estimation
For each co-primary endpoint, the registry states that the cumulative incidence function in each treatment group was calculated using a non-parametric Kaplan-Meier estimate. The registry also states that 95% confidence intervals were computed at each time point using Greenwood's variance estimation.
The Kaplan-Meier estimator updates the estimated event-free probability whenever events occur while allowing participants who are censored to contribute information up to their censoring time.
Log-rank test
The formal comparison for both co-primary endpoints used a two-sided asymptotic log-rank test. The log-rank test compares the observed pattern of events between randomized treatment groups across follow-up rather than comparing only a single fixed time point.
Cox regression and hazard ratios
For each primary endpoint, the hazard ratio and confidence interval were estimated using a Cox regression model with treatment group as the only factor. The reported effect measure was the hazard ratio for dronedarone versus placebo.
Because the comparison is explicitly dronedarone versus placebo, a hazard ratio above 1 indicates a higher estimated event hazard in the dronedarone group under the fitted model. The hazard ratio is a relative time-to-event measure; it is not an absolute risk difference or a probability that an individual participant will experience the event.
Confidence intervals
Both primary analyses report two-sided 95% confidence intervals. A confidence interval describes uncertainty around the estimated hazard ratio under the statistical model and sampling framework. It does not describe the range of effects experienced by individual patients.
Hypothesis type
The registry identifies the hypothesis type for both co-primary analyses as superiority. Thus, the statistical comparison was framed around whether the treatment groups differed in the prespecified direction of the superiority hypothesis rather than around a non-inferiority margin.
6. Primary Results: First Co-primary Outcome
The first co-primary outcome was time to the first event among stroke, systemic arterial embolism, MI, or cardiovascular death. The analysis used the intention-to-treat population, a two-sided log-rank test, and a Cox regression model with treatment group as the only factor for the hazard ratio.
Hazard ratio for dronedarone vs placebo
95% CI: 1.337–3.936 · P = 0.0019
Two-sided log-rank test; Cox regression estimate for the treatment-group hazard ratio.
| Statistical element | Reported result |
|---|---|
| Analysis population | Intent-to-treat population as previously defined |
| Groups compared | Placebo vs dronedarone |
| Method | Log-rank |
| Effect measure | Hazard ratio |
| Estimate | 2.294 |
| 95% CI | 1.337–3.936 |
| P-value | 0.0019 |
| Hypothesis | Superiority |
The reported hazard ratio of 2.294 means that, under the Cox model used for this analysis, the estimated instantaneous rate of the first co-primary event was higher in the dronedarone group than in the placebo group. Expressed descriptively, the estimate corresponds to an approximately 129.4% higher estimated hazard, although that relative-hazard interpretation should not be confused with a 129.4% higher cumulative probability of experiencing the event.
The 95% CI of 1.337–3.936 indicates substantial uncertainty around the point estimate, but the entire reported interval is above 1. The confidence interval is about the treatment-effect parameter estimated by the model; it does not mean that individual participants' hazards necessarily fell within that numerical range.
The P-value of 0.0019 is evidence against the null hypothesis used in the two-sided log-rank comparison under the specified analysis. It is not a measure of the size or clinical importance of the effect. The hazard ratio and its confidence interval provide that effect-size information.
Interpretation also requires attention to the registry's stated caveat: the study was terminated early, and the analysis was performed for information only without adjustment. Consequently, the nominal p-value should not be interpreted as though it arose from an analysis with a newly controlled multiplicity budget after early termination.
7. Primary Results: Second Co-primary Outcome
The second co-primary outcome was time to the first event among unscheduled cardiovascular hospitalization or death from any cause. As with the first co-primary endpoint, the analysis used the intention-to-treat population and a two-sided asymptotic log-rank test, with the hazard ratio estimated from a Cox model containing treatment group as the only factor.
Hazard ratio for dronedarone vs placebo
95% CI: 1.448–2.617 · P < 0.0001
Two-sided log-rank test; Cox regression estimate for the treatment-group hazard ratio.
| Statistical element | Reported result |
|---|---|
| Analysis population | Intent-to-treat population as previously defined |
| Groups compared | Placebo vs dronedarone |
| Method | Log-rank |
| Effect measure | Hazard ratio |
| Estimate | 1.947 |
| 95% CI | 1.448–2.617 |
| P-value | <0.0001 |
| Hypothesis | Superiority |
The hazard ratio of 1.947 indicates a higher estimated instantaneous rate of the second co-primary event in the dronedarone group relative to placebo under the fitted Cox model. As a simple descriptive translation, the estimate is approximately a 94.7% higher estimated hazard. This is not equivalent to saying that 94.7% more participants experienced the outcome.
The 95% CI of 1.448–2.617 quantifies uncertainty around the hazard-ratio estimate. Every value in the reported interval is above 1, so the estimated direction of the treatment effect is consistent across the interval, although the magnitude remains uncertain.
The P-value < 0.0001 describes the statistical evidence against the null hypothesis under the specified two-sided log-rank procedure. It does not indicate that the effect is four times as large as an analysis with P = 0.0019, nor does it quantify clinical importance.
The same early-termination qualification applies here. The registry states that the analysis was performed for information only without adjustment following premature discontinuation. That design feature should remain part of the interpretation rather than treating the nominal p-value as a standalone measure of evidentiary strength.
8. Secondary Result: Time to Cardiovascular Death
The registry also posts a formal statistical analysis for time to cardiovascular death. This was a secondary endpoint and used the same general time-to-event framework: intention-to-treat analysis, two-sided log-rank testing, and a Cox regression model with treatment group as the only factor.
Hazard ratio for cardiovascular death
95% CI: 0.996–4.490 · P = 0.0460
Two-sided log-rank test; Cox regression estimate for dronedarone vs placebo.
| Statistical element | Reported result |
|---|---|
| Endpoint | Time to cardiovascular death |
| Analysis population | Intent-to-treat population as previously defined |
| Groups compared | Placebo vs dronedarone |
| Method | Log-rank |
| Effect measure | Hazard ratio |
| Estimate | 2.115 |
| 95% CI | 0.996–4.490 |
| P-value | 0.0460 |
| Hypothesis | Superiority |
The point estimate of 2.115 corresponds to a higher estimated cardiovascular-death hazard in the dronedarone group relative to placebo under the Cox model. As a descriptive transformation, it is approximately a 111.5% higher estimated hazard.
The confidence interval is particularly important here: 0.996–4.490. Unlike the two co-primary hazard-ratio intervals, this interval extends to just below 1. The point estimate therefore suggests a higher hazard, but the interval reflects considerable uncertainty about the precise magnitude and includes values very close to the null value of 1.
The P-value of 0.0460 is close to the conventional 0.05 threshold, but a p-value near a threshold should not be interpreted as a graded measure of treatment effect. The effect estimate and confidence interval are more informative about magnitude and precision.
This endpoint was secondary, and the registry explicitly states that analyses following early termination were performed for information only without adjustment. The cardiovascular-death result therefore requires the same multiplicity and early-termination caution as the other posted analyses.
9. Results in One Statistical View
| Endpoint | HR, dronedarone vs placebo | 95% CI | P-value | Role |
|---|---|---|---|---|
| Stroke, systemic arterial embolism, MI, or cardiovascular death | 2.294 | 1.337–3.936 | 0.0019 | Co-primary |
| Unscheduled cardiovascular hospitalization or death from any cause | 1.947 | 1.448–2.617 | <0.0001 | Co-primary |
| Cardiovascular death | 2.115 | 0.996–4.490 | 0.0460 | Secondary |
The three posted hazard ratios all exceed 1, meaning that the estimated event hazard was higher in the dronedarone group for each of these outcomes. The confidence intervals differ substantially in precision, with the cardiovascular-death interval being wider and extending to a value just below 1.
10. Safety: Serious Adverse Events
The registry reports serious adverse events by randomized arm. These data provide an additional safety perspective but answer a different question from the time-to-event efficacy analyses.
| Arm | Serious adverse events | At risk |
|---|---|---|
| Placebo | 77 | 1609 |
| Dronedarone | 113 | 1614 |
The reported serious-adverse-event counts are 77 among 1609 participants in the placebo arm and 113 among 1614 participants in the dronedarone arm. These are counts of serious adverse events among the reported at-risk populations, not hazard ratios and not estimates of the co-primary efficacy outcomes.
11. Statistical Methods Explained
Why was a Kaplan-Meier estimator used?
The two co-primary outcomes were defined by the time from randomization until the first qualifying event. Time-to-event data require methods that can account for participants who have not experienced an event by the end of their available follow-up. Kaplan-Meier estimation provides an estimate of the event-free distribution over time while incorporating right-censored observations.
What does a hazard ratio of 2.294 mean?
A hazard ratio of 2.294 for dronedarone versus placebo indicates that the fitted Cox model estimated the instantaneous rate of the first co-primary event to be 2.294 times as high in the dronedarone group as in the placebo group. It does not mean that 2.294 times as many participants necessarily experienced the event, because cumulative event probabilities depend on the entire event and censoring history.
Why use a log-rank test as well as a Cox model?
The log-rank test provides a formal comparison of the survival experience between treatment groups, while the Cox model provides an interpretable effect estimate through the hazard ratio and its confidence interval. They therefore serve complementary purposes within the time-to-event analysis.
What does the confidence interval add?
The point estimate alone gives one estimated hazard ratio. The confidence interval communicates how uncertain that estimate is under the model and sampling framework. For the first co-primary endpoint, the reported interval is 1.337–3.936; for the second, it is 1.448–2.617; and for cardiovascular death, it is 0.996–4.490. The different widths show that the precision of an estimate is not determined by the point estimate alone.
Why does the cardiovascular-death p-value require caution?
The cardiovascular-death analysis produced a hazard ratio of 2.115 with a P-value of 0.0460, but its 95% confidence interval was 0.996–4.490. The interval is wide and reaches just below 1, emphasizing uncertainty around the treatment effect. In addition, cardiovascular death was secondary rather than one of the two co-primary outcomes, and the registry states that the analyses were performed without multiplicity adjustment following early termination.
Why does early termination matter statistically?
The registry states that the study was prematurely discontinued after 3236 patients had been randomized, representing 30% of the initial planned number. Early stopping changes the information available for estimation and can affect the interpretation of nominal p-values and effect estimates. The registry specifically states that the reported analyses were provided for information without adjustment, so the nominal p-values should be interpreted in that context.
12. Early Termination and Multiplicity
The registry's limitation statement is an important component of the statistical interpretation: the study was prematurely discontinued after 3236 patients were randomized, representing 30% of the initial planned number, and p-values were provided for information without any adjustment for multiplicity.
Early stopping
The available randomized sample and follow-up were determined by the premature discontinuation rather than completion of the initial planned enrollment.
Nominal p-values
The reported p-values were provided for information without multiplicity adjustment, so they should not be interpreted as though a formal familywise error-control strategy had been applied to the posted analyses.
Multiplicity is particularly relevant when a study has more than one primary endpoint and additional secondary analyses. A p-value is calculated relative to a specified statistical test; it does not automatically account for the fact that multiple hypotheses may have been examined.
13. Understanding the Hazard Ratios
The HR of 2.294 means that the estimated instantaneous hazard of stroke, systemic arterial embolism, MI, or cardiovascular death was higher in the dronedarone group than in the placebo group under the fitted Cox model. A hazard ratio is not an absolute risk measure and does not specify how many additional participants experienced the event.
The HR of 1.947 means that the estimated instantaneous hazard of unscheduled cardiovascular hospitalization or death from any cause was higher in the dronedarone group than in the placebo group. The corresponding 95% CI of 1.448–2.617 quantifies uncertainty around this estimate.
The HR of 2.115 suggests a higher estimated cardiovascular-death hazard in the dronedarone group. The 95% CI of 0.996–4.490 is substantially wider and extends to a value just below 1, indicating greater uncertainty around the magnitude of this secondary endpoint estimate.
These interpretations assume the Cox model provides an appropriate summary of the treatment-group comparison over the analyzed follow-up. A single hazard ratio is most straightforward to interpret when the relative hazards are reasonably stable over time. The registry does not provide additional information that would permit a separate assessment of the proportional-hazards assumption.
14. Intention-to-Treat Analysis
Both co-primary analyses and the cardiovascular-death secondary analysis were conducted in the intention-to-treat population as previously defined. This means the treatment comparison remains anchored to randomization rather than being restricted to participants who completed a particular treatment exposure.
| Analysis feature | PALLAS |
|---|---|
| Primary efficacy population | Intent-to-treat population as previously defined |
| First co-primary comparison | Placebo vs dronedarone |
| Second co-primary comparison | Placebo vs dronedarone |
| Secondary cardiovascular-death comparison | Placebo vs dronedarone |
| Primary time-to-event method | Two-sided log-rank test |
| Hazard-ratio estimation | Cox regression with treatment group as the only factor |
The ITT framework is particularly important in randomized trials because excluding participants after randomization can introduce selection into the treatment comparison. ITT does not, however, eliminate every source of uncertainty: censoring, early termination, endpoint adjudication, and multiplicity remain relevant to interpretation.
15. Censoring and Time-to-Event Interpretation
The PALLAS primary endpoints were measured from randomization through study termination, with a maximum follow-up of 1 year. Participants who had not experienced the relevant endpoint during their available observation period contribute information through the point at which their outcome status is no longer observed.
A time-to-event analysis does not simply classify every participant as an event or non-event. The timing of events and the amount of observed follow-up both contribute to the analysis. Kaplan-Meier estimation and Cox regression are designed to use this time information while accommodating right censoring under their respective assumptions.
For the two co-primary outcomes, the registry reports Kaplan-Meier estimation and Greenwood variance for confidence intervals at each time point. The hazard ratios were estimated from Cox regression, providing a model-based summary of the relative event hazard.
16. Primary Endpoint Adjudication
The two co-primary outcomes depended on central review and adjudication by a blinded Adjudication Committee. This is statistically relevant because composite clinical endpoints can contain several distinct event types, and consistent adjudication helps define whether an observed clinical event meets the registered endpoint definition.
First co-primary composite
Stroke, systemic arterial embolism, MI, or cardiovascular death, with the first qualifying event determining the time-to-event endpoint.
Second co-primary composite
Unscheduled cardiovascular hospitalization or death from any cause, with the first qualifying event determining the time-to-event endpoint.
Because each endpoint is a composite, the hazard ratio summarizes the time to the first qualifying component rather than providing separate effect estimates for every component. The registry's posted primary analyses are therefore appropriately interpreted at the composite-endpoint level.
17. Primary Endpoint Comparison
| Characteristic | First co-primary outcome | Second co-primary outcome |
|---|---|---|
| Endpoint structure | Composite time-to-event | Composite time-to-event |
| First event among | Stroke, systemic arterial embolism, MI, cardiovascular death | Unscheduled cardiovascular hospitalization, death from any cause |
| Analysis population | Intent-to-treat | Intent-to-treat |
| Comparison | Placebo vs dronedarone | Placebo vs dronedarone |
| Test | Two-sided log-rank | Two-sided log-rank |
| Effect measure | Hazard ratio | Hazard ratio |
| HR | 2.294 | 1.947 |
| 95% CI | 1.337–3.936 | 1.448–2.617 |
| P-value | 0.0019 | <0.0001 |
The two co-primary endpoints capture different aspects of cardiovascular clinical outcomes. The first focuses on serious thromboembolic, ischemic, and cardiovascular-death events, while the second includes unscheduled cardiovascular hospitalization and death from any cause. Their different definitions explain why they should be interpreted as distinct endpoints even though both use the same general survival-analysis framework.
18. What the P-values Do — and Do Not — Mean
For the first co-primary endpoint, P = 0.0019 represents the result of the specified two-sided log-rank comparison under its null hypothesis. It does not mean there is a 0.19% probability that the null hypothesis is true, and it does not quantify the clinical magnitude of the hazard-ratio estimate.
For the second co-primary endpoint, P < 0.0001 indicates very small tail probability under the specified null model and test. It should not be interpreted as an effect-size metric or as evidence that the second endpoint has a treatment effect exactly a particular number of times larger than the first endpoint.
For cardiovascular death, P = 0.0460 is the nominal two-sided log-rank result. Its interpretation is tempered by the wider 95% CI of 0.996–4.490, the endpoint's secondary role, and the registry's statement that the analysis was performed for information without multiplicity adjustment after early termination.
19. Why the Confidence Intervals Matter
Confidence intervals are particularly useful in this trial because they show that the precision of the estimated effect differs across endpoints.
| Endpoint | Point estimate | 95% CI | What the interval communicates |
|---|---|---|---|
| First co-primary outcome | 2.294 | 1.337–3.936 | Uncertainty around a higher estimated hazard in the dronedarone group |
| Second co-primary outcome | 1.947 | 1.448–2.617 | Uncertainty around a higher estimated hazard in the dronedarone group |
| Cardiovascular death | 2.115 | 0.996–4.490 | Wider uncertainty around the estimated cardiovascular-death hazard |
The confidence intervals should not be reduced to a simple "significant versus nonsignificant" classification. Their widths communicate statistical precision, while their location relative to 1 communicates the range of hazard-ratio values compatible with the model and data under the stated confidence framework.
20. Limitations
- Premature discontinuation: the study was terminated after 3236 participants were randomized, representing 30% of the initial planned number.
- Multiplicity: the registry states that p-values were provided for information without adjustment for multiplicity.
- Interpretive status of analyses: the registry states that the analyses were performed for information only without adjustment as a consequence of early termination.
- Follow-up duration: the registered primary endpoints were followed from randomization to study termination, with a maximum follow-up of 1 year.
- Composite endpoints: each co-primary endpoint combines multiple clinical event types, so a single hazard ratio does not identify which individual component drove the overall result.
- Hazard-ratio assumptions: Cox regression provides a model-based hazard-ratio summary; the registry does not provide additional information here for independently evaluating proportional hazards.
- Secondary endpoint uncertainty: the cardiovascular-death confidence interval is wide and extends to 0.996, emphasizing uncertainty around the point estimate.
- Serious adverse events: the reported safety counts describe serious adverse events by arm but do not by themselves provide a formal time-to-event safety comparison.
21. Why This Trial Matters Statistically
PALLAS is a useful statistical teaching case because it combines randomized treatment allocation, triple masking, composite time-to-event endpoints, Kaplan-Meier estimation, Greenwood confidence intervals, log-rank testing, Cox regression, hazard ratios, intention-to-treat analysis, and the interpretive consequences of premature trial termination.
| Statistical concept | How it appears in PALLAS |
|---|---|
| Randomization | 3236 participants randomized between dronedarone and placebo |
| Masking | Triple-masked parallel-group design |
| Intention-to-treat | Used for the posted primary and cardiovascular-death analyses |
| Kaplan-Meier estimation | Used for cumulative incidence functions for the co-primary outcomes |
| Greenwood variance | Used for 95% confidence intervals at each time point |
| Log-rank test | Two-sided asymptotic comparison between randomized groups |
| Cox regression | Used to estimate hazard ratios and confidence intervals |
| Hazard ratio | Primary effect measure for the posted time-to-event analyses |
| Composite endpoint | Both co-primary outcomes use first occurrence among multiple event types |
| Early termination | Study discontinued after 3236 randomized participants, 30% of initial planned number |
| Multiplicity | Registry states that p-values were provided without multiplicity adjustment |
22. Related Tutorials
Learn more about the methods used in this trial:
23. Related Calculators
24. Sources
- ClinicalTrials.gov: NCT01151137.
- PubMed: PMID 22082198.
- PubMed: PMID 25378467.
Continue through Clinical Biostats
Connect the statistical methods used in PALLAS with deeper tutorials and statistical calculators for clinical-trial analysis.
25. Record Summary
PALLAS provides a clear example of how randomized clinical-trial evidence can be expressed through time-to-event methodology. The two co-primary outcomes were analyzed in the intention-to-treat population using Kaplan-Meier estimation, Greenwood-based confidence intervals, two-sided log-rank testing, and Cox regression hazard ratios. The reported hazard ratios were 2.294 for the first co-primary outcome and 1.947 for the second, while the secondary cardiovascular-death analysis reported an HR of 2.115. Interpretation of all three analyses must incorporate the registry's statement that the study was prematurely discontinued after 3236 randomized participants, representing 30% of the initial planned number, and that p-values were provided without multiplicity adjustment.