Why the Corticosteroid Domain Is a Distinct Comparison
In severe COVID-19, corticosteroids became a treatment question because inflammation can contribute to critical illness, while corticosteroids can also have important harms. The relevant question was not simply whether “steroids work.” It was which treatment strategy to use, for which patients, and compared with what alternative. REMAP-CAP addressed that question within an international platform trial among critically ill patients receiving respiratory or cardiovascular organ support in an intensive care unit.
The corticosteroid domain compared hydrocortisone strategies. It did not directly randomize patients to dexamethasone, and its comparison groups should not be confused with those in RECOVERY. This distinction matters when reading the results: evidence about one corticosteroid regimen can inform the clinical context without establishing that another regimen has the same effect.
The Three Strategies and the Patient Setting
The domain enrolled adults with suspected or confirmed COVID-19 who were critically ill and receiving intensive care organ support. The two active strategies used intravenous hydrocortisone, but the rules governing its use differed. The fixed-dose strategy specified hydrocortisone 50 mg every six hours for seven days. The shock-dependent strategy used hydrocortisone 50 mg every six hours while the patient had shock, according to the protocol’s clinical criteria. The third strategy assigned no hydrocortisone as the domain intervention.
“No hydrocortisone” describes the assigned domain strategy, not a guarantee that every participant received no corticosteroid of any kind under all circumstances. Clinical care, rescue decisions, and other platform interventions form part of the practical setting in which an assigned strategy is implemented. When interpreting a randomized comparison, the primary question is what the assigned strategies imply, not whether every participant followed the protocol perfectly.
The primary outcome was organ support-free days through day 21, with death assigned the worst value. This outcome combines survival and time without respiratory or cardiovascular organ support. A strategy can therefore look unfavorable because more patients die, because survivors need support longer, or both. It should not be interpreted as a simple average number of hospital-free days.
Worked Example: What the Assignment Means
Consider a hypothetical eligible patient who has severe COVID-19 and is receiving respiratory support in intensive care. Suppose the domain assigns this patient to the shock-dependent hydrocortisone strategy. The protocol calls for hydrocortisone while shock criteria are met; it does not call for a fixed seven-day course regardless of shock status. If the patient is assigned to no hydrocortisone, that means the patient is not assigned hydrocortisone as the domain treatment. These descriptions distinguish treatment policies; neither assignment by itself guarantees a particular clinical outcome.
For a hypothetical outcome illustration, suppose two patients survive to day 21. Patient A is free of organ support for 10 days in the outcome window, and Patient B is free for 4 days. Their organ support-free-day values differ by 6 days. If a patient dies before day 21, death receives the prespecified worst value rather than being treated as an ordinary survivor with zero free days. Thus, an arm’s outcome distribution reflects both survival and organ-support duration.
What the Interim Comparison Could Establish
REMAP-CAP used Bayesian adaptive methods. In this setting, posterior probabilities summarize uncertainty about treatment effects conditional on the observed data, the statistical model, and its assumptions. They are not p-values. A posterior probability that an intervention is superior to control is also not the probability that a particular patient will benefit.
In the corticosteroid domain, the reported interim analysis gave a 93% posterior probability of superiority for fixed-dose hydrocortisone over no hydrocortisone on the primary outcome, and an 80% probability for shock-dependent hydrocortisone over no hydrocortisone. These probabilities indicated evidence leaning toward benefit, especially for the fixed-dose strategy, but they did not mean that either regimen had been proved superior under every decision rule. Nor did the two probabilities directly establish which hydrocortisone strategy was better than the other.
A useful way to see the distinction is to consider a simplified Bayesian calculation. This calculation is illustrative only; it is not a reconstruction of REMAP-CAP’s model or results. Suppose an estimated log odds ratio for a favorable outcome is \(\hat{\theta}\), with standard error \(SE\), and suppose a normal prior is assigned to \(\theta\). If larger values favor treatment, superiority corresponds to \(\theta>0\). Under a normal prior and normal likelihood, the posterior distribution is also normal. Its probability above zero depends on the posterior mean relative to its posterior standard deviation.
For prior variance \(1\), the posterior variance is \(v=(1+1/SE^2)^{-1}\), and the posterior mean is \(m=v\hat{\theta}/SE^2\). Therefore, \(P(\theta>0\mid \hat{\theta})=\Phi(m/\sqrt{v})\), where \(\Phi\) is the standard normal cumulative distribution function. This shows why a favorable point estimate alone is insufficient: uncertainty, as reflected in \(SE\), affects the probability of superiority.
Worked Example: A Simplified Posterior Probability
Suppose, purely hypothetically, an intervention has estimated log odds ratio \(\hat{\theta}=0.25\) with \(SE=0.20\), under the prior \(\theta\sim N(0,1)\). A positive log odds ratio favors the intervention. The posterior variance is
The posterior mean is
The posterior standard deviation is \(\sqrt{0.03846}=0.1961\). Thus,
Under these hypothetical assumptions, the posterior probability that the intervention’s log odds ratio is positive is about 89.0%. This is not an 89% probability that an individual patient will improve. It is also not a REMAP-CAP result: the real domain used its specified model and outcome, not this simplified calculation.
Why RECOVERY Changed the Interpretation
While the REMAP-CAP corticosteroid domain was enrolling, RECOVERY reported that dexamethasone reduced mortality among hospitalized patients with COVID-19 who required respiratory support, with benefit varying by level of respiratory support. This was important external evidence: it supported the use of a corticosteroid in a clinically overlapping population and changed the standard-of-care context. REMAP-CAP stopped enrollment in its corticosteroid domain after the RECOVERY result became available.
The external result did not turn REMAP-CAP into a dexamethasone trial. RECOVERY compared dexamethasone with usual care; REMAP-CAP’s domain compared hydrocortisone strategies, including fixed-dose and shock-dependent use, with no hydrocortisone. The populations and outcomes also need to be considered in their respective trial contexts. Evidence from one trial can strengthen the rationale for a treatment class while leaving uncertainty about the preferred drug, dose, duration, and strategy.
Worked Example: Translating External Evidence Cautiously
Suppose, for illustration only, an external randomized trial estimates a mortality risk ratio of 0.80 for a corticosteroid in a population whose untreated mortality risk is 40%. The implied treated risk under that estimate is
That corresponds to an illustrative absolute risk reduction of \(0.40-0.32=0.08\), or 8 percentage points, in that population. It does not prove that hydrocortisone in REMAP-CAP would produce the same risk ratio or absolute reduction. To transport the estimate, one would need to assess whether the drug, dose, timing, severity distribution, outcome definition, and clinical care are sufficiently comparable. The calculation clarifies the size of the hypothetical external effect; it does not supply a hydrocortisone-specific estimate.
Deriving the Main Interpretive Lesson
The effect of a treatment strategy is always relative to a comparator and a defined population. Let \(Y(a)\) denote the outcome that would be observed for a participant under strategy \(a\), and \(Y(c)\) the outcome under comparator \(c\). A randomized domain comparison aims to learn about the contrast between outcomes under those strategies for its eligible population. It does not directly identify a contrast involving an unrandomized regimen.
This logic explains why a positive dexamethasone result can change the clinical meaning of a hydrocortisone-versus-no-hydrocortisone comparison without answering the hydrocortisone-versus-dexamethasone question. The first contrast is informed by randomization within REMAP-CAP; the second requires direct comparative evidence or additional assumptions. Treating the drugs as interchangeable would silently add an assumption that the trials themselves did not test.
An external result can also change equipoise: whether it remains ethically and scientifically appropriate to randomize participants to a strategy that omits a treatment now supported by strong evidence. This can justify stopping enrollment even when an ongoing trial has not reached its own prespecified statistical boundary. Stopping in response to external evidence is not equivalent to declaring the ongoing comparison statistically conclusive.
Diagnostics, Alternatives, and Limitations
When reading this case, first check what treatment each group actually received by assignment, which patients were eligible, and what endpoint was analyzed. Then distinguish the domain’s interim model-based probabilities from conventional hypothesis-test results. Finally, identify whether a claim is about hydrocortisone in critically ill patients, corticosteroids as a broad class, or dexamethasone in a different trial population.
- Do not equate the treatment arms. Fixed-dose and shock-dependent hydrocortisone are different policies, and neither is the same assigned regimen as dexamethasone in RECOVERY.
- Do not overread the posterior probabilities. A 93% probability of superiority is evidence favoring an effect under the model, not certainty, a 93% patient response rate, or a p-value.
- Account for the stopping context. Enrollment stopped after external evidence changed the standard-of-care context; the domain’s comparisons therefore did not answer every planned regimen question with the same completeness as a fully completed trial might.
- Keep the outcome in view. Organ support-free days incorporate death and duration of support. A result on this outcome is not automatically identical to a mortality effect.
- Be cautious about transportability. Differences in severity, timing, treatment protocols, and usual care can limit how directly results from another trial apply.
The main alternative to direct domain evidence is to use external evidence, such as RECOVERY, to guide practice. External evidence may be more relevant to a particular drug or population, but it does not provide the same randomized head-to-head comparison of REMAP-CAP’s hydrocortisone strategies. The strongest interpretation combines the sources while preserving what each design can and cannot establish.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- What distinguished the fixed-dose and shock-dependent hydrocortisone strategies?
- How should a 93% posterior probability of superiority be interpreted, and what does it not mean?
- Why did the RECOVERY dexamethasone result matter to REMAP-CAP without directly answering which hydrocortisone regimen was best?
- A hypothetical external study reports a risk ratio of 0.75 and a comparator risk of 20%. What treated risk and absolute reduction follow from those numbers? Why should they not be called REMAP-CAP hydrocortisone effects?
- Why can stopping enrollment after external evidence become available be reasonable even if an ongoing trial has not crossed its own statistical stopping boundary?
Solutions. (1) Fixed-dose hydrocortisone was specified as 50 mg every six hours for seven days; the shock-dependent strategy used that dose while shock criteria were met. (2) It is a model-conditional probability that the treatment effect is favorable, given the data and prior. It is neither a p-value nor the probability an individual patient benefits. (3) RECOVERY’s dexamethasone finding changed the standard-of-care context and equipoise, but it randomized a different corticosteroid comparison; it did not directly compare REMAP-CAP’s hydrocortisone regimens. (4) The treated risk is \(0.75\times0.20=0.15\), or 15%; the absolute reduction is \(0.20-0.15=0.05\), or 5 percentage points. Those are calculations from the hypothetical study, not estimates for REMAP-CAP hydrocortisone. (5) New evidence can alter whether continued randomization is ethically or clinically appropriate. A trial can stop for that reason without its own comparison having met a prespecified statistical conclusion rule.