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
EOLIA was a randomized, parallel-group trial evaluating extracorporeal membrane oxygenation (ECMO) versus conventional care for patients with acute respiratory distress syndrome (ARDS). The registry identifies all-cause mortality on day 60 following randomization as the primary endpoint.
| Feature | EOLIA |
|---|---|
| Trial acronym | EOLIA |
| Clinical condition | Acute Respiratory Distress Syndrome (ARDS) |
| Design | Randomized, parallel-group |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 249 |
| Number of arms | 2 |
| Interventions | ECMO (Quadrox®, Jostra®, Maquet®) versus conventional care |
| ClinicalTrials.gov | NCT01470703 |
| Lead sponsor | Assistance Publique - Hôpitaux de Paris |
| Sponsor type | OTHER |
2. Clinical Question
The central statistical question is whether assignment to ECMO rather than conventional care changes the risk of death from any cause by day 60 following randomization in patients with severe acute respiratory distress syndrome.
Population
Patients enrolled in the trial for the condition identified in the registry as acute respiratory distress syndrome (ARDS).
Intervention
Extracorporeal membrane oxygenation (ECMO), using Quadrox®, Jostra®, and Maquet® devices as identified in the registry.
Comparator
Conventional care.
Primary question
What is the difference between the randomized treatment groups in all-cause mortality on day 60 following randomization?
3. Trial Design
ECMO
- Extracorporeal membrane oxygenation
- Device intervention
- Devices identified in the registry include Quadrox®, Jostra®, and Maquet®
Conventional care
- Conventional care
- Comparator intervention
- Classified in the registry as an "other" intervention
The parallel randomized structure is important statistically because the treatment groups are defined by assignment at randomization. The primary comparison is therefore naturally framed around differences in outcomes between the two randomized groups rather than differences between patients who independently chose or received one treatment or the other.
4. Endpoints
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Primary endpoint | All cause mortality on day 60 following randomization | 60 days |
What the primary endpoint measures
The endpoint is a binary status evaluated at a prespecified time point: whether the participant has died from any cause by day 60 following randomization. The phrase all cause mortality means that deaths are counted without restricting the endpoint to a particular cause of death.
The time origin is also explicit: follow-up begins at randomization. This is statistically important because the treatment comparison is anchored to the point at which randomized treatment assignment occurs.
5. Planned Analysis
The registry identifies all-cause mortality on day 60 following randomization as the primary endpoint, with a 60-day time frame. Results have not been posted on ClinicalTrials.gov.
Why a time-to-event analysis may also be informative
Although the registry defines the primary endpoint at a fixed 60-day point, mortality is inherently a time-to-event outcome. If the underlying follow-up times and censoring information are available, survival methods such as Kaplan-Meier estimation and a log-rank comparison can describe when deaths occur during follow-up rather than reducing the entire follow-up period to a single day-60 status.
A Cox proportional-hazards model is another common approach for a mortality time-to-event endpoint. It estimates a hazard ratio comparing the instantaneous event rates between treatment groups over the analyzed follow-up. That model requires additional assumptions and information beyond the single endpoint definition in the registry.
A fixed-time analysis answers whether mortality status differs at day 60. A time-to-event analysis uses the timing of deaths and censoring throughout follow-up and therefore answers a somewhat different statistical question.
6. Statistical Methodology
Binary endpoint analysis
The primary endpoint can be represented for each participant by an indicator taking the value 1 if death from any cause occurs by day 60 following randomization and 0 otherwise. The randomized groups can then be compared using a two-group method appropriate for binary outcomes.
The risk difference is expressed on an absolute scale. A negative value would correspond to a lower observed day-60 mortality proportion in the ECMO group, while a positive value would correspond to a higher proportion.
Relative effect measures
A risk ratio compares the day-60 mortality probability in the ECMO group with the corresponding probability in the conventional-care group. An odds ratio instead compares the odds of death between groups. These measures are related but are not interchangeable, particularly when the outcome is not rare.
Confidence intervals
A confidence interval provides information about the statistical uncertainty around an estimated treatment effect. The width of the interval is influenced by the amount of information available and the variability of the observed data.
For a binary mortality endpoint, confidence intervals should accompany the point estimate rather than relying on a p-value alone. A point estimate describes the observed magnitude of the group difference; the interval communicates how precisely that magnitude has been estimated.
Hypothesis testing
A conventional two-group hypothesis test would begin with a null hypothesis that the relevant mortality measure is equal between randomized groups. The p-value then describes how compatible the observed data are with that null hypothesis under the specified statistical model.
The p-value is not a measure of the size of the treatment effect. A very small p-value can accompany a small absolute difference in a large study, while a clinically meaningful difference can fail to reach a conventional significance threshold when the estimate is imprecise.
Kaplan-Meier estimation
If participant-level event times are available, Kaplan-Meier estimation can describe the probability of remaining alive over time. This approach accommodates right-censored observations, such as participants whose event status is not observed for the entire follow-up period.
Here, di is the number of deaths at time ti and ni is the number of participants at risk immediately before that time.
Log-rank comparison
For a time-to-death analysis, a log-rank test can compare survival distributions between the randomized groups. It evaluates the observed and expected numbers of events over follow-up rather than comparing only the final number of deaths.
Cox proportional-hazards model
A Cox model can estimate a hazard ratio for death between ECMO and conventional care. The hazard ratio is a relative measure of the instantaneous event rate, conditional on the model. It is not equivalent to a risk difference or a risk ratio at day 60.
The hazard ratio should be interpreted over the analyzed follow-up and under the proportional-hazards model. It does not mean that the same percentage of participants experienced a corresponding reduction in their individual probability of death.
7. Statistical Methods Explained
Why is day-60 mortality a binary endpoint?
The registry defines the primary endpoint at a fixed time point: day 60 following randomization. Each participant can therefore be classified according to whether death from any cause occurred by that point. This produces a directly interpretable proportion for each randomized group.
Why is randomization important for the statistical comparison?
Randomization establishes the treatment groups through an allocation mechanism rather than through post-randomization selection by investigators or participants. Under the assumptions of the randomized design, differences in outcomes between the assigned groups can therefore be attributed to the randomized treatment strategy with substantially less concern about measured and unmeasured baseline confounding than in an observational comparison.
What does an absolute risk difference tell us?
The risk difference describes the separation between the two treatment-group mortality probabilities on an absolute scale. For example, if one group had a mortality probability of 40% and the other 30%, the absolute risk difference would be 10 percentage points. That is distinct from saying that mortality was reduced by 10% in relative terms.
Why might Kaplan-Meier analysis be useful even though the endpoint is defined at day 60?
The day-60 endpoint reduces the outcome to a prespecified status at one time point. Kaplan-Meier analysis retains information about when deaths occur throughout follow-up. If two groups have the same day-60 mortality but very different timing of deaths, their survival curves can still provide important statistical information that a single binary endpoint does not capture.
What does a hazard ratio actually measure?
A hazard ratio compares the modeled instantaneous rates of the event between two groups. An HR below 1 indicates a lower estimated instantaneous event rate in the numerator group under the model. It is not a direct statement that the same proportion of patients has a corresponding percentage reduction in absolute mortality.
Why are confidence intervals important?
A point estimate alone does not reveal how precisely the treatment effect has been estimated. The confidence interval supplies a range of values compatible with the statistical uncertainty of the analysis under its assumptions. Wider intervals indicate greater uncertainty, while narrower intervals indicate greater precision.
8. Interpreting the Primary Endpoint
The day-60 endpoint should be interpreted first as an absolute mortality comparison. The key quantity is the proportion of randomized participants who died from any cause by the specified 60-day time point in each treatment group.
A relative measure such as a risk ratio can describe how the mortality probability in the ECMO group compares with that in the conventional-care group. Relative measures are useful for comparing proportional differences, but they do not replace the absolute mortality probabilities.
If a hazard ratio is estimated from participant-level follow-up, it describes the relative instantaneous event rate over time. It does not directly provide the probability that an individual participant will survive to day 60.
A p-value does not measure clinical importance, the magnitude of the treatment effect, or the probability that one treatment is superior. It is a measure of compatibility between the observed data and the specified null hypothesis under the chosen statistical framework.
9. Randomization and Causal Interpretation
The randomized allocation is central to the causal interpretation of the EOLIA comparison. The trial enrolled 249 participants and assigned them to 2 parallel arms: ECMO or conventional care.
For a randomized trial, the estimand is naturally tied to the treatment strategy assigned at randomization. This means the primary comparison should preserve the randomized groups rather than selectively excluding participants after randomization because of subsequent treatment exposure, withdrawal, or outcome availability.
Strength of the design
Randomized allocation provides the framework for comparing outcomes between treatment assignments while reducing confounding by factors that influence both treatment selection and mortality.
Open-label structure
The registry identifies the trial as having no masking. Lack of masking can affect aspects of treatment delivery, clinical management, outcome assessment, and post-randomization care, depending on the endpoint and study procedures.
10. Fixed-Time Endpoints and Censoring
Day-60 mortality and time-to-event survival analysis use related but different information. A fixed-time endpoint focuses on whether the event has occurred by a prespecified date. A survival analysis also uses the timing of the event and can account for censoring before the end of follow-up.
This distinction matters because a participant who is alive at day 60 contributes a favorable binary outcome to the fixed-time analysis, regardless of whether that participant died shortly after day 60 or remained alive much longer. A time-to-event analysis retains more information about the timing of events beyond that binary classification.
Conversely, a fixed day-60 endpoint can be attractive when the clinical question is explicitly about mortality status at a clinically meaningful landmark. It creates a simple and directly interpretable estimand: the probability of death from any cause by day 60 after randomization.
11. Open-Label Treatment and Endpoint Objectivity
The registry identifies masking as none. This is particularly relevant when assessing endpoints because knowledge of treatment assignment can influence clinical decisions and the collection or assessment of some outcomes.
For all-cause mortality, however, the event definition is relatively direct: the endpoint is death from any cause by a prespecified time. This reduces some of the subjectivity that can affect endpoints based on symptom scores, clinician judgments, or imaging interpretation.
The distinction illustrates an important statistical principle: the consequences of open-label treatment depend partly on the objectivity of the endpoint. A highly objective endpoint can be less susceptible to certain forms of assessment bias than a subjective endpoint, although open-label care can still influence other aspects of the trial pathway.
12. Sample Size and Statistical Precision
The registry reports an enrollment of 249 participants across 2 randomized arms. Sample size affects the precision with which a treatment effect can be estimated.
For a mortality endpoint, statistical information depends not only on the total number enrolled but also on the number of deaths, the allocation between treatment groups, and the completeness of follow-up. More events generally provide more information for estimating a mortality difference or time-to-event effect.
| Design quantity | Registry value | Statistical relevance |
|---|---|---|
| Total enrollment | 249 | Defines the overall randomized sample available for the trial. |
| Number of arms | 2 | Creates the two treatment groups for the primary comparison. |
| Primary endpoint | All cause mortality on day 60 following randomization | Defines the principal treatment-effect comparison. |
| Primary time frame | 60 days | Defines the landmark at which the primary mortality endpoint is evaluated. |
The enrollment number alone does not determine the eventual precision of the treatment effect. For mortality outcomes, the number and timing of deaths are especially important because they determine the amount of event information available for the comparison.
13. Trial Timeline
Trial start
The registry lists 2011-12-08 as the study start date.
Primary completion
The registry lists 2017-07 as the primary completion date.
Current registry status
The ClinicalTrials.gov record identifies the study status as completed.
14. What Can Be Learned From the Endpoint Definition Alone?
Even without a posted numerical treatment effect, the endpoint specification reveals several important features of the statistical question.
Time origin
Randomization is the starting point for the 60-day follow-up period. This anchors outcome measurement to the randomized treatment assignment.
Event definition
The event is death from any cause rather than a disease-specific cause of death.
Landmark
The primary endpoint is evaluated at day 60, creating a prespecified fixed-time mortality question.
Comparative structure
The outcome is compared between 2 randomized parallel treatment arms.
These features define the estimand more clearly than a generic statement such as "survival was compared." The statistical question is specifically about all-cause mortality at a defined time following randomized assignment.
15. Limitations
- No posted statistical analyses: the ClinicalTrials.gov record does not post formal statistical analyses or numerical primary-endpoint results for the day-60 mortality endpoint.
- Limited endpoint information: the registry identifies the primary endpoint and its time frame but does not provide additional endpoint estimates in the record represented here.
- Analysis method: the endpoint definition alone does not establish whether a particular statistical test, regression model, or survival-analysis procedure was used.
- No numerical effect estimate: without a posted treatment-effect estimate and confidence interval, the magnitude and precision of the difference between treatment groups cannot be summarized from the registry record.
- No subgroup interpretation: the available registry information does not provide subgroup-specific mortality estimates.
- No safety comparison: the available record does not provide serious adverse-event counts by randomized arm for analysis here.
- No baseline comparison: the available registry information does not provide a numerical baseline characteristics table for comparison between arms.
16. Why This Trial Matters Statistically
EOLIA provides a useful example of how a randomized critical-care trial can be framed around a clinically concrete mortality endpoint while also illustrating the distinction between a fixed-time endpoint and a full time-to-event analysis.
| Statistical concept | How it appears in EOLIA |
|---|---|
| Randomization | The trial is randomized and compares 2 parallel arms. |
| Comparative effectiveness | ECMO is compared with conventional care. |
| Binary mortality endpoint | All-cause mortality is evaluated at day 60 following randomization. |
| Time-to-event analysis | Mortality can also be studied using survival methods when event times and censoring information are available. |
| Risk difference | Provides an absolute comparison of day-60 mortality probabilities. |
| Risk ratio | Provides a relative comparison of day-60 mortality probabilities. |
| Kaplan-Meier estimation | Can describe survival over time rather than only day-60 status. |
| Log-rank testing | Can compare survival distributions between randomized groups. |
| Cox modeling | Can estimate a hazard ratio for mortality when appropriate time-to-event data are available. |
| Censoring | Becomes relevant when analyzing the timing of death rather than only day-60 status. |
| Open-label design | The registry identifies masking as none, making endpoint objectivity relevant to interpretation. |
17. Fixed-Time Mortality vs Hazard Ratio
One of the most important statistical distinctions in this trial is the difference between a day-60 mortality comparison and a hazard ratio.
These measures can point in the same general direction while describing different aspects of the data. A hazard ratio does not substitute for an absolute day-60 mortality estimate, and an absolute mortality difference does not describe the timing of events throughout follow-up.
For clinical interpretation, presenting absolute and relative measures together is generally more informative than relying on one summary statistic alone.
18. Statistical Interpretation Without a Posted Result
The absence of a numerical result changes the appropriate statistical interpretation. The design tells us what comparison the trial was constructed to make, but the endpoint definition by itself does not tell us the magnitude of the observed difference between ECMO and conventional care.
The trial is randomized, has 2 parallel arms, enrolls 249 participants, compares ECMO with conventional care, and identifies all-cause mortality on day 60 following randomization as the primary endpoint.
The primary analysis would quantify mortality by day 60 in each randomized group and compare those probabilities using an appropriate inferential framework.
A confidence interval around an absolute or relative treatment effect would show the statistical precision of that estimate. Without the numerical estimate and interval, the magnitude and precision of the treatment comparison cannot be described.
A p-value from a prespecified hypothesis test would describe the compatibility of the observed data with the null hypothesis under that test. It would not itself quantify the size or clinical importance of the treatment effect.
19. Interpretation of the Randomized Treatment Effect
Because treatment assignment is randomized, the primary comparison should be interpreted at the level of the randomized treatment groups. This preserves the causal structure established at allocation and avoids turning the analysis into an observational comparison of patients who happened to receive different forms of care.
For a day-60 mortality endpoint, an appropriate treatment-effect statement would normally include the mortality proportion in each arm, the absolute difference, a relative measure such as a risk ratio when appropriate, and corresponding confidence intervals. If time-to-event methods are used, the analysis could additionally report survival estimates and a hazard ratio with its confidence interval.
The choice among these summaries should be driven by the prespecified estimand and analysis plan. They should not be treated as interchangeable descriptions of the same statistical quantity.
20. Design Considerations for a Mortality Trial
Endpoint timing
Specifying day 60 before examining outcomes establishes a common landmark for the primary mortality comparison.
Competing causes of death
Because the endpoint is all-cause mortality, deaths are counted regardless of their specific cause.
Event timing
When full follow-up times are available, survival analysis can use information about when deaths occur rather than only whether they occurred by day 60.
Open-label care
The absence of masking makes objective outcome definitions particularly valuable when interpreting the primary endpoint.
21. Statistical Lessons From EOLIA
The trial is a useful teaching example because the clinical question can be translated directly into a statistical estimand:
- Start with the randomized treatment assignment.
- Define the event precisely: death from any cause.
- Define the time origin: randomization.
- Define the primary landmark: day 60.
- Compare mortality between the randomized groups on an absolute scale.
- Use relative measures when they add clinically meaningful information.
- Use time-to-event methods when the timing of events and censoring are part of the intended estimand.
- Report confidence intervals so that effect magnitude and precision can be evaluated together.
- Interpret p-values as hypothesis-testing quantities rather than measures of effect size.
This sequence helps separate the clinical question from the statistical representation of that question. The endpoint definition comes first; the statistical model should then be chosen to match that estimand and the available data.
22. Sources
- ClinicalTrials.gov: EOLIA, NCT01470703.
Continue through the Clinical Biostats trial library
Explore additional clinical trial analyses and statistical resources covering randomized designs, survival analysis, treatment-effect estimation, and clinical-trial methodology.
23. Record Summary
EOLIA is a completed randomized, parallel-group trial of ECMO versus conventional care in acute respiratory distress syndrome, with 249 participants and 2 treatment arms. The registry identifies all-cause mortality on day 60 following randomization as the primary endpoint with a 60-day time frame.
Statistically, the endpoint is naturally expressed as a fixed-time mortality comparison between randomized groups. Absolute mortality proportions and their difference directly address the day-60 question, while relative measures such as a risk ratio provide a proportional comparison. If complete event-time information is available, Kaplan-Meier estimation, log-rank testing, and Cox modeling can provide additional information about the timing of death and the relative event rate over follow-up.