Reading REMAP-CAP Results by Treatment Domain

Why Read REMAP-CAP by Treatment Domain?

REMAP-CAP is a platform for studying treatments for severe community-acquired pneumonia, including COVID-19. Its platform structure allows several treatment questions to be studied within an ongoing trial. It does not make those questions interchangeable. A patient eligible for an immune-modulating treatment comparison is not automatically eligible for an anticoagulation comparison; the interventions, clinical rationale, and sometimes the outcome definitions differ.

The practical task when reading a result is to recover the question that was randomized: which patients were eligible, which interventions were compared, against what control, and on which outcome? A platform-wide number that leaves those features unstated can blend clinically distinct questions. Even if a numerical summary can be calculated, it may not estimate an effect that is meaningful for any particular treatment decision.

Definition: A treatment domain is a component of the platform that evaluates one broad type of treatment question, such as immune modulation or anticoagulation. Within a domain, eligible participants may be randomized among specified interventions and a control. A domain-specific treatment effect applies to that domain's eligible population, interventions, comparator, outcome, and analysis period; it is not automatically an effect for every platform participant.

Keep three levels separate. The platform is the overall trial structure. A domain identifies a treatment question. The intervention arms within that domain specify what participants can actually receive. Eligibility rules define who contributes to that comparison. A patient can be eligible for more than one domain, and assignment in one domain need not determine assignment in another.

Read the Question Before the Estimate

For each result, first identify the population and treatment contrast. In REMAP-CAP's COVID-19 work, examples include an immune-modulation domain comparing interleukin-6 receptor antagonists such as tocilizumab or sarilumab with control, and an anticoagulation domain comparing therapeutic-dose anticoagulation with usual-care thromboprophylaxis in critically ill patients. These are distinct interventions for distinct biological questions, even when the participants share a broad description such as “critically ill with COVID-19.”

The published REMAP-CAP findings illustrate why the distinction matters. The interleukin-6 receptor antagonist domain reported benefit on its primary outcome among critically ill patients with COVID-19. In contrast, the therapeutic anticoagulation comparison in critically ill patients did not show improvement in organ-support-free days and was stopped for futility. These conclusions are not contradictory: they concern different interventions and randomized contrasts. Nor do they establish that every immune-modulating treatment is beneficial or that anticoagulation is ineffective in every population or clinical setting.

The earlier material on corticosteroid comparisons in REMAP-CAP concerns another domain-specific question. When reading that domain, keep its eligible patients and its particular hydrocortisone strategies distinct from the comparisons of interleukin-6 receptor antagonists or anticoagulation. A result from one domain is not a substitute control group or a direct head-to-head comparison for another.

Key reading rule: A treatment effect is conditional on its randomized comparison. Before interpreting an estimate, state the eligible population, intervention, control, outcome, and follow-up period. If any of these differ across results, do not describe the estimates as if they answered the same question.

Outcomes Are Part of the Treatment Question

An important COVID-19 outcome used in REMAP-CAP was organ-support-free days through day 21. It combines survival and time without ICU-level respiratory or cardiovascular support on an ordinal scale. A participant who dies in the period is assigned −1; a participant who survives is assigned the number of days, from 0 to 21, alive and free of the specified organ support. Thus, higher values are generally more favorable, but the endpoint is not simply a count of days for every participant: death is deliberately ranked below survival.

The analysis can use the ordering of outcomes rather than treating the scale as an ordinary continuous measurement. In reported analyses, an odds ratio above 1 for a model parameterized in the favorable direction indicates a shift toward better outcomes with the intervention. Always check the article's definition of the effect measure and direction before interpreting its sign. A model-based estimate for an ordinal endpoint is not the same as a difference in mean days, a risk ratio for death, or a comparison of medians.

Outcome definitions also limit across-domain comparisons. A mortality outcome at a specified follow-up and an ordinal organ-support outcome do not measure the same thing, even if both are clinically important. If a report changes the outcome or follow-up, a platform-wide summary cannot erase that difference. Report what was measured, not a broader construct such as “overall improvement” unless the analysis actually defines and supports it.

Worked Example: Coding an Organ-Support-Free-Day Outcome

Consider two hypothetical critically ill patients followed through day 21. Patient A survives and is alive without the specified organ support on 14 days. Patient B dies on day 12. Apply the outcome definition.

For Patient A, the score is 14: the number of days alive and free of the specified organ support during the period. For Patient B, the score is −1 because death receives that value, regardless of the number of preceding days without support.

The ordering places Patient A above Patient B. It would be incorrect to score Patient B as, for example, 9 free days and compare that value as though it were an ordinary count on the same scale. The death assignment is part of the outcome definition. This example explains the coding, not a treatment effect; comparing two individual scores cannot establish that a treatment works.

Domain Effects and Platform-Wide Summaries

Suppose a domain-specific outcome is a binary indicator of a favorable response. Let \(p_{I,d}\) be the response proportion under the intervention in domain \(d\), and \(p_{C,d}\) the response proportion under its control. The domain risk difference is

\Delta_d = p_{I,d} - p_{C,d}.

This simple measure is useful for illustrating how populations affect summaries. A platform-wide average of domain effects would require weights \(w_d\) that specify the target population or decision context:

\Delta_{\text{weighted}} = \sum_d w_d \Delta_d, \qquad \sum_d w_d = 1.

The weights are not a cosmetic detail. Equal weights answer a different question from weights proportional to the number of eligible patients, and either can differ from weights that represent a target clinical population. Moreover, this arithmetic does not make the interventions interchangeable: it only summarizes the stated domain effects under the stated weights.

Worked Example: A Weighted Summary Can Change Meaning

Suppose two hypothetical domains report favorable-response risk differences. Domain A has \(p_{I,A}=0.70\) and \(p_{C,A}=0.60\), so \(\Delta_A=0.70-0.60=0.10\), or 10 percentage points. Domain B has \(p_{I,B}=0.25\) and \(p_{C,B}=0.30\), so \(\Delta_B=0.25-0.30=-0.05\), or −5 percentage points.

If a target summary gives each domain equal weight, then

\Delta_{\text{weighted}} = 0.50(0.10)+0.50(-0.05) = 0.025.

The weighted difference is 2.5 percentage points. If instead the target population is represented by 20% from Domain A and 80% from Domain B, then

\Delta_{\text{weighted}} = 0.20(0.10)+0.80(-0.05) = -0.02.

That summary is −2 percentage points. Neither calculation is automatically the “platform effect.” Each answers a different weighted question, and the domain interventions and eligible populations remain distinct. The example uses hypothetical data and is not a REMAP-CAP estimate.

A further complication is that platform membership does not guarantee a common concurrent control. Eligibility rules, available interventions, and clinical practice can change over time. Shared controls can be informative when participants are concurrently eligible for the same comparison and the analysis accounts for the design. A control drawn from another period or a different eligibility set is not automatically a valid control for a domain-specific treatment effect.

In adaptive platforms, allocation probabilities and the available arms may also change as evidence accumulates. These features are part of the design and analysis; they do not license informal comparisons of response proportions across domains. Read the reported domain-specific model and its uncertainty, and check whether the estimate is adjusted for the relevant design features. A crude pooled proportion can mix differences in prognosis, eligibility, calendar period, and treatment assignment.

Worked Reading Strategy

Name the eligible population.
Use the eligibility criteria for the domain, not the broadest description of platform participants. Identify clinically relevant restrictions and the period to which the result applies.

Write the randomized contrast.
Name the intervention arm or arms and the control actually used. Do not turn an indirect comparison across domains into a head-to-head randomized comparison.

Define the outcome and direction.
Check the outcome scale, follow-up, handling of death, and whether a larger value or an odds ratio above 1 favors treatment.

Interpret the estimate only within that question.
State the effect and uncertainty in context. Then assess whether a proposed summary across domains has a defined target population and defensible weights.

Common Pitfalls and Limitations

Calling every platform participant the study population. The denominator for a domain result is the population eligible for that domain and included in its analysis, not necessarily everyone enrolled in REMAP-CAP.
Reading a favorable outcome as a universal benefit. A domain result supports a conclusion about its studied intervention, comparator, population, and outcome. It does not automatically generalize to other treatments in the same broad class or to less severely ill patients.
Combining different effects without defining a target. A weighted average depends on the weights and can change when the target population changes. If the questions are clinically distinct, a single average may conceal rather than clarify the evidence.
Treating organ-support-free days as ordinary days for everyone. Death is assigned a specific rank on the endpoint; it is not simply a low count of support-free days.
Assuming separate domain results are direct comparisons. If two interventions were not randomized against each other in the same eligible population and comparison, their reported estimates do not establish which is superior.
Ignoring time and adaptation. Enrollment, standard care, and available arms can change during a platform trial. A result must be interpreted using the analysis and comparison actually reported, not a casually assembled pool across periods.

Domain-level reporting is valuable because it preserves the structure of the clinical questions. Its limitation is that conclusions remain bounded by the domain's eligibility, intervention choices, outcome, and available evidence. The right response to multiple different findings is not to force them into one verdict, but to state what each comparison shows and where its relevance begins and ends.

Worked Example: Interpreting a Hypothetical Ordinal-Outcome Estimate

Suppose a hypothetical domain report gives an adjusted odds ratio of 1.40, with a 95% confidence interval from 1.05 to 1.87, for an ordinal outcome parameterized so that values above 1 favor the intervention. Assume the model and interval are appropriate for the randomized comparison.

The point estimate is above 1, and the interval excludes 1. Under this model, the data are compatible with a favorable shift in the outcome distribution for the intervention compared with control in the domain's eligible population. The estimate is an odds ratio for the modeled ordinal outcome; it is not a 40-percentage-point increase in response and does not say that every participant benefits. The confidence interval describes uncertainty in the effect estimate under the model, not a 95% probability that the fixed treatment effect lies in this particular interval.

Before making a clinical conclusion, the reader would still check the outcome definition, model direction, population, comparator, and any reported posterior or frequentist decision criteria. The numbers here are hypothetical and do not represent a REMAP-CAP result.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

What five features should be identified before interpreting a REMAP-CAP domain estimate?
Why does a favorable result in the interleukin-6 receptor antagonist domain not establish benefit for every immune-modulating treatment?
In the hypothetical example, Domain A has a risk difference of 10 percentage points and Domain B has a risk difference of −5 percentage points. Calculate the weighted difference if Domain A receives 40% weight and Domain B 60%.
What does the −1 score represent in the organ-support-free-days outcome described here?
Why is it not valid to treat two interventions studied in separate domains as a randomized head-to-head comparison?

Solutions. (1) Identify the eligible population, intervention, control, outcome and its direction, and follow-up period. (2) The result applies to the studied intervention and eligible population; another treatment in the same broad class may have a different effect. (3) \(0.40(0.10)+0.60(-0.05)=0.04-0.03=0.01\), or a weighted difference of 1 percentage point. (4) It is the assigned outcome score for a participant who dies during the specified period. (5) Without randomization between those interventions in a common eligible comparison, differences may reflect populations, controls, periods, or other design features rather than treatment alone.