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
RECOVERY is a randomized phase 3 factorial trial evaluating multiple therapeutic interventions in patients with pneumonia. The registry describes a 23-arm design with an enrollment target of 70000 participants and three registered primary endpoints focused on all-cause mortality and time to discharge alive.
| Feature | RECOVERY |
|---|---|
| Acronym | RECOVERY |
| Official title | Randomised Evaluation of COVID-19 Therapy |
| Phase | Phase 3 |
| Status | Recruiting |
| Condition | Pneumonia |
| Allocation | Randomized |
| Design model | Factorial |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 70000 |
| Number of arms | 23 |
| Start date | 2020-03-19 |
| Primary completion date | 2028-09-30 |
| Lead sponsor | University of Oxford |
| Sponsor type | OTHER |
| ClinicalTrials.gov | NCT04381936 |
2. Clinical Question
The registry describes a broad treatment platform for pneumonia in which multiple interventions are evaluated within a randomized factorial framework. The primary statistical questions are framed around whether each relevant study treatment affects all-cause mortality and, for the influenza co-primary outcome, whether treatment affects time to discharge alive from hospital.
Population
Participants with pneumonia enrolled in the RECOVERY trial.
Interventions
The registry lists multiple drug and biological interventions within the factorial treatment platform.
Comparator framework
For the community-acquired pneumonia mortality endpoint, the registry states that each pairwise comparison is made with the “no additional treatment” arm.
Primary questions
Do study treatments affect all-cause mortality within 28 days, and for the influenza co-primary outcome, do they affect mortality and time to discharge alive?
3. Trial Design
A factorial design allows more than one treatment question to be addressed within a common randomized trial structure. Statistically, the key feature is that treatment factors can be evaluated within the same overall platform rather than requiring a separate conventional two-arm trial for every intervention.
4. Interventions and the Factorial Platform
The registry describes a broad treatment platform. It lists the following interventions:
| Intervention | Type |
|---|---|
| Lopinavir-Ritonavir | Drug |
| Corticosteroid | Drug |
| Hydroxychloroquine | Drug |
| Azithromycin | Drug |
| Convalescent plasma | Biological |
| Tocilizumab | Drug |
| Immunoglobulin | Biological |
| Synthetic neutralising antibodies | Drug |
| Aspirin | Drug |
| Colchicine | Drug |
| Baricitinib | Drug |
| Anakinra | Drug |
| Dimethyl fumarate | Drug |
| High Dose Corticosteroid | Drug |
| Empagliflozin | Drug |
| Sotrovimab | Drug |
| Molnupiravir | Drug |
| Paxlovid | Drug |
| Baloxavir Marboxil | Drug |
| Oseltamivir | Drug |
| Corticosteroids (dexamethasone) | Drug |
| Corticosteroids (dexamethasone) | Drug |
The registry identifies the overall design as having 23 arms. The intervention list contains two entries for corticosteroids (dexamethasone), as recorded in the registry.
5. Primary Endpoints
The registry defines three primary endpoints. Two concern all-cause mortality, while the influenza co-primary outcome also includes time to discharge alive from hospital.
| Primary endpoint | Time frame | Registry description |
|---|---|---|
| Community-acquired pneumonia: All-cause mortality (with subsidiary analyses of cause of death and of death at various timepoints following discharge) | Within 28 days after randomisation | For each pairwise comparison with the “no additional treatment” arm, the primary objective is to provide reliable estimates of the effect of study treatments on all-cause mortality. |
| Influenza co-primary outcome: All-cause mortality (with subsidiary analysis of cause of death and death at various timepoints following discharge) | Within 28 days after randomisation | Primary outcome specified in the registry. |
| Influenza co-primary outcome: Time to discharge alive from hospital | Within the first 28-days | Primary outcome specified in the registry. |
Why the endpoint definitions matter
All-cause mortality is a binary status at a defined follow-up horizon when expressed as “within 28 days,” although it can also be analyzed as a time-to-event outcome when the timing of death is retained. Time to discharge alive is explicitly a time-to-event endpoint and has a competing-event feature: a participant must remain alive to experience discharge alive.
The distinction is important because an analysis of mortality within 28 days answers a different question from an analysis of the time until discharge. A treatment could affect the speed of hospital discharge without producing the same magnitude of effect on mortality, and the two endpoints should therefore remain statistically distinct.
6. Planned Analysis
The registry has not posted formal statistical analyses or numerical endpoint results for these primary outcomes. The appropriate analysis therefore begins with the prespecified endpoint definitions and the structure of the factorial randomized design.
Primary statistical target
For the community-acquired pneumonia endpoint, the registry specifically frames the primary objective as reliable estimates of treatment effects on all-cause mortality for each pairwise comparison with the “no additional treatment” arm.
All-cause mortality within 28 days
For a mortality endpoint defined as death from any cause within 28 days after randomisation, a natural analysis can compare the proportion of participants who die by day 28 between randomized treatment groups. Depending on the prespecified statistical analysis plan, the effect can be expressed using an absolute risk difference, a risk ratio, an odds ratio, or a time-to-event measure such as a hazard ratio.
Because the registry's stated objective is to provide reliable treatment-effect estimates, the effect estimate should be accompanied by a confidence interval. A confidence interval communicates the statistical precision of the estimate; it does not describe the range of outcomes that individual participants might experience.
Time to discharge alive
Time to discharge alive is naturally represented using time-to-event methods when the precise discharge time is available. Kaplan-Meier methods can estimate the distribution of time to discharge, while a Cox model can provide a hazard ratio if the model assumptions are appropriate. However, death prevents subsequent discharge alive, so death must be handled explicitly rather than treating it as ordinary right-censoring.
The hazard describes the instantaneous event rate among participants who remain event-free immediately before time t. It is not the same quantity as the probability of experiencing the event by day 28.
Factorial treatment comparisons
The factorial design requires the analysis to respect the treatment-factor structure. If a treatment is compared with a no-additional-treatment control, the primary contrast should preserve the randomized comparison rather than being reconstructed from observed treatment uptake or post-randomization treatment.
When several treatment questions are evaluated within the same platform, multiplicity becomes a central design consideration. The relevant issue is not simply the number of arms; it is the number and dependence structure of statistical hypotheses being tested and the error-control strategy specified for those hypotheses.
7. Statistical Methodology
Randomization
Randomization is the central design mechanism that supports causal comparison between treatment groups. Under successful randomization, baseline differences between groups arise from chance rather than from deliberate selection of treatment. This allows an appropriately analyzed randomized comparison to estimate the causal effect associated with treatment assignment.
Factorial analysis
Factorial designs evaluate multiple treatment factors within a shared randomized structure. The statistical model can represent the treatment factors simultaneously, allowing treatment effects to be estimated efficiently when the design and assumptions support that approach.
Here, A and B represent treatment factors. The interaction term assesses whether the effect of one factor depends on the level of the other factor.
The usefulness of a factorial model depends on the scientific question and the assumptions underlying the design. An interaction is a different question from the main treatment effect: it asks whether the effect of one treatment changes according to exposure to another treatment factor.
Binary mortality analysis
When mortality is summarized as survival versus death by a fixed time point, the endpoint can be analyzed as a binary outcome. A risk difference compares absolute event probabilities, while a risk ratio compares them multiplicatively.
A negative risk difference for mortality would correspond to a lower observed probability of death in the treatment group. The sign convention must always be interpreted with the endpoint coding explicitly defined.
Kaplan-Meier estimation
If the timing of death or discharge is retained, Kaplan-Meier estimation can describe the time-to-event distribution. The method accommodates right-censored observations by allowing participants to contribute information up to their last known event-free time.
Here, di is the number of events at time ti, and ni is the number at risk immediately before that time.
Cox proportional-hazards model
A Cox model can estimate a hazard ratio for a time-to-event endpoint. The hazard ratio compares modeled instantaneous event rates between treatment groups, conditional on the model structure.
The exponentiated treatment coefficient, exp(β), is the hazard ratio for the treatment contrast represented by X.
The proportional-hazards assumption deserves particular attention. A Cox hazard ratio is most straightforward to interpret as a common relative hazard over time when the proportional-hazards assumption is reasonable. If hazards vary substantially over time, alternative summaries such as restricted mean survival time or time-specific risks may provide additional information.
Competing events for discharge alive
For time to discharge alive, death is not simply another censoring event because a participant who dies can no longer experience the event “discharge alive.” Treating death as ordinary non-informative censoring can therefore change the estimand. A competing-risks framework can instead recognize death as an event that precludes discharge alive.
8. Statistical Methods Explained
Why is a factorial design useful here?
A factorial design allows multiple treatment questions to be addressed within a common randomized platform. Its statistical advantage is efficiency: when the design assumptions hold, information from a shared population can contribute to more than one treatment comparison.
What is the difference between an arm and a treatment factor?
An arm is a specific randomized treatment configuration. A treatment factor is a component of the factorial structure. With several factors, multiple arms can represent different combinations of factor levels. The analysis should therefore be based on the prespecified factor structure rather than treating every arm as an unrelated experiment.
Why is all-cause mortality a useful endpoint?
All-cause mortality has an objective event definition: death from any cause. It avoids requiring investigators to classify every death according to a treatment-specific causal mechanism for the primary event definition. Cause of death can remain a subsidiary analysis without replacing the primary all-cause endpoint.
Why does time to discharge need special treatment?
Because the endpoint is specifically “discharge alive,” death prevents the endpoint from occurring. This creates a competing-event structure. The analysis must therefore distinguish between a participant who remains hospitalized, a participant who is discharged alive, and a participant who dies before discharge.
What does a hazard ratio mean?
A hazard ratio is a relative comparison of instantaneous event rates under a time-to-event model. A value below 1 corresponds to a lower modeled event rate in the treatment group relative to the reference group. It does not directly state the percentage of participants who will experience the event, nor does it equal a risk ratio.
Why do confidence intervals matter?
A point estimate alone does not communicate how precisely the treatment effect has been estimated. A confidence interval provides a range of parameter values compatible with the statistical model and observed information under the stated confidence procedure. Narrower intervals generally indicate greater statistical precision than wider intervals.
Why is multiplicity important in a multi-arm factorial trial?
When many hypotheses are tested, the probability of obtaining at least one apparently unusual result by chance can increase. A prespecified multiplicity strategy controls the relevant error rate across the family of hypotheses that the trial treats as confirmatory. The appropriate family depends on the trial's statistical objectives rather than simply the number of treatment arms.
9. Analysis Populations and Interpretation
The registry information identifies randomized allocation and the overall factorial design but does not provide posted definitions for detailed efficacy or safety analysis populations.
| Population concept | Statistical role |
|---|---|
| Randomized population | Provides the foundation for an intention-to-treat comparison when participants are analyzed according to randomized assignment. |
| Treated population | Can be useful for exposure-based safety summaries when treatment exposure is relevant. |
| Endpoint-specific population | Requires explicit prespecified rules for eligibility, follow-up, censoring, and competing events. |
For a randomized trial, the intention-to-treat principle is particularly important for efficacy because it preserves the treatment comparison created by randomization. Excluding participants after randomization because of subsequent events can introduce selection into the comparison.
10. Multiplicity and Multiple Comparisons
The RECOVERY registry describes a 23-arm factorial design and multiple registered primary outcomes. This makes multiplicity a central statistical issue.
Many treatment questions
Multiple interventions create multiple treatment-effect questions within the same platform.
Multiple endpoints
The registry includes all-cause mortality and time to discharge alive among the primary endpoints.
Familywise interpretation
Confirmatory claims require an appropriate prespecified strategy for controlling the relevant type I error rate.
Factorial structure
The number of possible arm-to-arm contrasts is not necessarily the same as the number of primary hypotheses.
A useful statistical distinction is between estimating treatment effects and testing hypotheses. A treatment effect estimate can be accompanied by a confidence interval even when the multiplicity framework makes an isolated nominal P-value inappropriate for confirmatory interpretation.
11. Safety
The ClinicalTrials.gov record does not include posted serious adverse-event counts by arm. No numerical safety results are reported here.
From a statistical perspective, safety in a multi-arm randomized trial should be summarized separately from efficacy. Serious adverse events are typically tabulated by treatment exposure, with the analysis population, exposure duration, coding conventions, and event definitions specified in the statistical analysis plan.
| Safety analysis question | Statistical consideration |
|---|---|
| How many participants experience a serious adverse event? | Compare counts and proportions using a clearly defined safety population. |
| How often do particular events occur? | Specify the event definition, coding level, and observation period. |
| Does exposure duration differ? | Consider whether crude proportions adequately represent exposure-adjusted event rates. |
| Are multiple safety events evaluated? | Distinguish participant-level incidence from the number of event episodes. |
12. Trial Timeline
Trial start
The registry lists 2020-03-19 as the study start date.
Randomized factorial platform
The registry classifies RECOVERY as a phase 3 randomized factorial trial with 23 arms.
Recruiting
The registry lists the study status as Recruiting.
Primary completion date
The registry lists 2028-09-30 as the primary completion date.
13. Interpreting Mortality Within 28 Days
The community-acquired pneumonia primary endpoint is all-cause mortality within 28 days after randomisation. This is a fixed follow-up horizon, so the principal quantity can be expressed as the probability of death by day 28 for each randomized comparison.
An effect estimate such as a risk difference describes the absolute difference between those probabilities. A risk ratio describes their relative relationship. Neither quantity says that every participant experiences the same treatment effect.
Relative measures can make small and large absolute differences appear similar. Reporting absolute mortality probabilities and their difference provides direct information about the magnitude of the clinical outcome at the specified 28-day horizon.
The denominator used for a 28-day mortality estimate must be defined consistently. In a randomized trial, the intention-to-treat population is generally the natural starting point for estimating the effect of treatment assignment.
14. Interpreting Time to Discharge Alive
The influenza co-primary endpoint defines time to discharge alive from hospital within the first 28-days. This is not simply the proportion discharged by day 28; it incorporates when discharge occurs during the follow-up period.
A time-to-event estimate can distinguish a participant discharged early from one discharged later, whereas a binary “discharged by day 28” endpoint would treat both as the same outcome.
Death is a competing event for discharge alive because a participant who dies cannot subsequently be discharged alive. This is a substantive part of the endpoint definition and should be incorporated into the analysis rather than hidden through ordinary censoring.
15. Why This Trial Matters Statistically
RECOVERY is a useful statistical teaching case because its design combines a very large randomized platform with multiple treatment factors and clinically distinct primary endpoints.
| Concept | How it appears in RECOVERY |
|---|---|
| Randomization | The registry classifies allocation as randomized. |
| Factorial design | The registry identifies the design model as factorial. |
| Multiple treatment questions | The registry lists 23 arms and multiple interventions. |
| Fixed-time mortality | All-cause mortality is evaluated within 28 days after randomisation. |
| Time-to-event analysis | Time to discharge alive is defined within the first 28-days. |
| Competing events | Death can prevent the event “discharge alive” from occurring. |
| Multiplicity | Multiple treatment comparisons and primary outcomes require a prespecified error-control framework. |
| Effect estimation | The registry describes the objective as providing reliable estimates of treatment effects. |
16. Important Limitations and Interpretation Issues
- Multiple comparisons: a 23-arm factorial platform creates numerous potential statistical contrasts, so confirmatory interpretation depends on the prespecified hypothesis family and multiplicity strategy.
- Factorial interactions: treatment-factor interactions are distinct from main effects. A main-effect estimate should not automatically be interpreted as applying identically across every factorial combination.
- Competing events: death is directly relevant to the interpretation of time to discharge alive because it prevents the discharge event from occurring.
- Fixed follow-up: mortality within 28 days is a time-specific estimand. It should not automatically be generalized to mortality beyond the stated time frame.
- Unmasked design: the registry specifies no masking. This can be particularly relevant for outcomes involving clinical decisions, although all-cause mortality itself is less subjective than many other endpoints.
- Endpoint distinction: mortality and time to discharge answer different clinical and statistical questions and should not be collapsed into a single measure.
- Registry scope: the ClinicalTrials.gov record identifies the design, endpoints, enrollment, interventions, and trial status but does not provide posted numerical statistical analyses for the primary endpoints.
- Analysis specification: exact choices among competing statistical models, estimands, multiplicity procedures, and missing-data methods require the applicable statistical analysis plan rather than assumptions based solely on the trial's high-level registry fields.
17. What a Complete Analysis Would Report
A statistically informative results presentation for this design would go beyond a single P-value. For each prespecified treatment comparison, it would identify the estimand, analysis population, effect measure, confidence interval, and follow-up period.
| Component | Purpose |
|---|---|
| Effect estimate | Quantifies the magnitude and direction of the treatment comparison. |
| Confidence interval | Communicates statistical precision around the estimate. |
| P-value, where applicable | Quantifies compatibility with a specified null hypothesis under the chosen statistical model; it does not measure clinical importance or effect size. |
| Analysis population | Defines which randomized participants contribute to the analysis. |
| Estimand | Defines precisely what treatment effect is being targeted. |
| Multiplicity procedure | Determines how confirmatory error is controlled across related hypotheses. |
| Time horizon | Defines the period over which the treatment effect is being estimated. |
This framework is especially important in a factorial platform because a statistically significant finding cannot be interpreted independently of which treatment contrast, endpoint, and hypothesis family produced it.
18. Statistical Concepts in This Trial
Learn more about the methods used in this trial:
19. Related Statistical Calculators
20. Sources
- ClinicalTrials.gov: RECOVERY (NCT04381936).
Continue through Clinical Biostats
Explore statistical tutorials and calculators for clinical-trial design, analysis, and interpretation.
21. Record Summary
RECOVERY is a phase 3 randomized factorial trial with a registry-listed enrollment of 70000, 23 arms, and a broad set of drug and biological interventions. Its primary statistical structure centers on all-cause mortality within 28 days after randomisation and, for the influenza co-primary outcome, time to discharge alive within the first 28-days.
The trial is particularly useful statistically because it illustrates the distinction between a conventional two-arm comparison and a factorial treatment platform. The analysis must preserve the randomized treatment contrasts, distinguish fixed-time mortality from time-to-discharge outcomes, account for death as a competing event when analyzing discharge alive, and address multiplicity when multiple treatment hypotheses are evaluated.