Why Organize a Platform by Treatment Domains?
A critically ill patient may be eligible for more than one therapeutic question at the same time. For example, a trial may evaluate an immune-modulating intervention and an antithrombotic intervention, each against an appropriate control. Treating these as separate conventional trials would require separate recruitment and control groups, even when the same patient could safely and appropriately enter both comparisons. REMAP-CAP addresses this structure by organizing treatment questions into domains.
A domain is not simply a list of drugs. It defines a clinical question, the interventions that answer it, the participants eligible for that question, and its comparator. This structure is especially relevant in critical care, where treatment decisions may concern different biological pathways and have distinct contraindications. The platform can therefore make several comparisons available without assuming that every intervention is appropriate for every patient.
In REMAP-CAP, the domain structure accommodates questions about critical-care treatments for community-acquired pneumonia, including questions involving immune modulation, anticoagulation, antiplatelet treatment, or antiviral therapy. The exact interventions and eligibility rules depend on the protocol version and clinical context. The important design point is that eligibility and comparison are defined at the domain level rather than by a single platform-wide list of treatments.
Eligibility and the Meaning of a Domain Comparison
Let \(E_d=1\) indicate that a patient meets the eligibility criteria for domain \(d\). Eligibility can depend on clinical characteristics, treatment contraindications, and whether the patient has already received a treatment that makes the comparison unsuitable. Randomization within domain \(d\) occurs among patients with \(E_d=1\), not among all participants in the platform.
For an eligible population, a treatment-control contrast estimates the effect of assignment to that treatment versus assignment to that domain’s control, under the trial’s randomization and analysis plan. Its target population is therefore the patients eligible for that particular domain. Two domains may have different eligible populations, even when they recruit through the same platform.
Being enrolled in one domain does not itself mean a patient is eligible for another. Nor does entry into two domains mean that the patient receives only one randomized assignment. If a participant qualifies for both, the trial can make a separate assignment in each. A patient may consequently receive, for example, an active immune-modulating assignment and a control assignment in an antithrombotic domain, subject to the protocol’s eligibility and safety rules.
Worked Example: A Patient Eligible for Two Domains
Suppose, hypothetically, a critically ill patient meets the eligibility criteria for both an immune-treatment domain and an anticoagulation domain. For illustration only, assume the immune domain has two active options and a control, with equal assignment probabilities, while the anticoagulation domain has one active option and a control, also with equal probabilities. Assume the two domain assignments are independent.
In the immune domain, the probability of assignment to a particular active option is \(1/3\). In the anticoagulation domain, the probability of assignment to active treatment is \(1/2\). By independence, the probability that this patient receives both specified active assignments is
The probability of receiving immune option A and the anticoagulation-domain control is likewise \(1/3\times 1/2=1/6\). These are joint assignment probabilities under the stated hypothetical scheme, not probabilities that treatment will succeed. Actual assignment probabilities are determined by the protocol and may not be equal or independent in every circumstance.
This patient contributes to the immune-domain comparison according to the immune assignment and to the anticoagulation-domain comparison according to the anticoagulation assignment. The two contributions must be interpreted as belonging to their respective randomized questions. Treating the patient as if they had received only one overall assignment would discard information about one of the comparisons.
How Shared Controls Work
Suppose a domain compares two active interventions, \(A\) and \(B\), with a common control, \(C\). A participant randomized to \(C\) can provide control information for both the \(A\)-versus-\(C\) and \(B\)-versus-\(C\) contrasts, if that is how the protocol defines the domain. This can avoid splitting a limited control group into separate groups for each active comparison. It does not make the two estimated treatment effects statistically independent: both contrasts include the same estimated control outcome.
The following derivation shows the consequence for a simple continuous outcome. Assume participants in the three randomized groups are independent, outcomes within each group have the same variance \(\sigma^2\), and the sample means are calculated from \(n_A\), \(n_B\), and \(n_C\) participants. Define the estimated contrasts as \(\hat{\Delta}_A=\bar{x}_A-\bar{x}_C\) and \(\hat{\Delta}_B=\bar{x}_B-\bar{x}_C\). Since the treatment groups are independent of each other and of the control group, the only term shared by the two contrasts is \(\bar{x}_C\):
The covariance is positive: if the observed control mean happens to be low, both treatment-minus-control estimates tend to be higher. This dependence matters when making joint statements about several interventions. A method that treats the contrasts as independent would not correctly represent their sampling uncertainty.
Worked Example: Precision and Dependence With a Shared Control
Consider a hypothetical domain with a continuous clinical outcome measured in the same units for all participants. Assume a common standard deviation of 1.0 outcome unit in each randomized group. Suppose 50 participants receive intervention \(A\), 50 receive intervention \(B\), and 100 receive the shared control. The observed means are 6.2, 5.8, and 5.0 units, respectively.
The estimated contrasts are
For either contrast, the estimated variance under the stated common-variance assumptions is \(1/n_{\text{active}}+1/n_C\), in squared outcome units. Thus,
The same calculation gives \(SE(\hat{\Delta}_B)=0.1732\) units. The covariance between the two contrasts is \(1/100=0.010\) squared units, so their correlation is \(0.010/\sqrt{0.030\times0.030}=0.333\). The estimates are positively correlated because they share the control mean.
For comparison, if the total of 100 control participants were instead split into two independent control groups of 50, each contrast would have estimated variance \(1/50+1/50=0.040\), and standard error \(\sqrt{0.040}=0.200\) units. In this illustrative equal-variance calculation, sharing 100 controls gives each contrast a smaller standard error than allocating only 50 controls to each comparison. The gain comes with correlated contrasts; the shared-control design does not make the comparisons independent.
These calculations illustrate a design principle, not an analysis prescription for every REMAP-CAP endpoint. Actual analyses must use the endpoint’s appropriate model, the trial’s allocation structure, and its prespecified inferential approach.
Overlapping Domains and the Analysis Population
When participants can enter multiple domains, the number of domain assignments is not necessarily the number of unique participants. Each domain comparison uses participants eligible for that domain and randomized within it. A participant assigned in two domains appears in the randomized data for both questions, but should not be counted as two different people in a platform-wide description of enrollment.
A domain-specific causal interpretation relies on randomization being implemented among eligible participants and on the analysis respecting the assigned treatment. Eligibility rules should be applied before the relevant randomization, and treatment assignment should not be chosen based on a participant’s prognosis after enrollment. If the analysis conditions on post-randomization events in a way that differs between groups, the randomized comparison may no longer have its intended interpretation.
A separate issue is treatment interaction. If assignments across two domains are independent, joint combinations may be randomized, making it possible in principle to investigate whether the effect in one domain differs according to assignment in the other. But an interaction is a distinct estimand from either domain’s average treatment contrast. It generally requires enough participants in the relevant assignment combinations and a prespecified model or analysis plan. The fact that patients can be randomized in multiple domains does not, by itself, guarantee that interaction estimates will be precise or clinically interpretable.
Worked Example: Counting Assignments Without Duplicating Participants
Suppose, hypothetically, 120 unique patients enter a platform. Of these, 100 meet the criteria for an immune-treatment domain, 80 meet the criteria for an anticoagulation domain, and 60 are eligible for both. The number eligible for at least one of these domains is \(100+80-60=120\), matching the number of unique patients in this example.
If each eligible patient is randomized once within every domain for which they qualify, the immune domain has 100 domain assignments and the anticoagulation domain has 80. Together there are \(100+80=180\) domain assignments, but still only 120 unique patients. The 60 patients eligible for both domains account for two assignments each.
Suppose further that each domain uses a hypothetical 1:1 active-versus-control allocation. The expected number assigned to active treatment in the immune domain is \(100\times 1/2=50\), and in the anticoagulation domain it is \(80\times 1/2=40\). These are expected counts, not guaranteed observed counts. They are domain-specific totals and should not be added to claim that 90 distinct patients received active treatment: some individuals may have received active assignments in both domains.
Assumptions, Diagnostics, and Alternatives
Domain-based randomization is useful only if the domain boundaries correspond to clinically meaningful and operationally feasible questions. Before interpreting a comparison, check that the enrolled population meets the stated domain eligibility criteria, that treatment and control definitions are clear, and that allocation probabilities and assignment records are available for each domain. Review recruitment and allocation counts by domain and arm. Unexpected imbalances can arise by chance, especially in small groups, but large or systematic discrepancies may indicate a data or implementation problem that requires investigation.
For shared-control comparisons, confirm that the control participants genuinely belong to the same domain comparison and were enrolled under compatible eligibility and outcome definitions. Check whether recruitment period, site practices, or changes in standard care could make a pooled control group unlike the patients randomized to an active intervention. Randomization supports comparability at assignment; it does not repair inconsistent outcome measurement or a control definition that changes without being represented in the analysis.
| Design choice | Strength | Important qualification |
|---|---|---|
| Separate two-arm trials | Each trial has a direct, self-contained treatment-control comparison. | Separate recruitment and controls can use more participants or resources. |
| Multiple active options with a shared domain control | One control group can inform several comparisons within that domain. | Contrasts sharing control observations are correlated and must be analyzed accordingly. |
| Randomization in multiple eligible domains | One participant can contribute randomized information to more than one treatment question. | Eligibility and assignment are domain-specific; cross-domain interactions require their own justification and data. |
The main alternative is to run separate trials, each with its own eligibility criteria and control group. That can simplify the relationship between each trial and its primary comparison, but it may duplicate screening and control recruitment. A platform with domain-based randomization can be more efficient when questions share a population and compatible control conditions. It is not automatically more efficient in every setting: different eligibility rules, incompatible controls, sparse joint assignments, or major changes in clinical practice can limit what can validly be pooled.
Common Pitfalls
- Calling the platform control universal. A control that serves several interventions within one domain is not necessarily the control for every other domain. Identify the comparator defined for the specific question.
- Assuming every participant receives one platform-wide assignment. A patient can receive separate assignments in multiple eligible domains. Describe assignments and denominators by domain.
- Treating shared-control contrasts as independent. Their estimates share the control mean. Joint uncertainty calculations must account for that covariance.
- Equating multiple-domain eligibility with evidence of interaction. A patient’s participation in two domains creates the possibility of observing joint assignments, not proof that the trial can estimate every interaction precisely.
- Generalizing beyond the eligible population. A domain’s randomized comparison directly concerns its eligible participants. It does not automatically estimate an effect in patients excluded from that domain.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- What population does a treatment-control contrast within one domain directly concern?
- In a hypothetical domain, two active arms share one control group. Why are the two estimated treatment-control contrasts correlated?
- There are 90 participants eligible for domain A, 70 eligible for domain B, and 40 eligible for both. How many unique participants are eligible for at least one domain, and how many domain assignments occur if each eligible participant is randomized once in each eligible domain?
- Does randomization in two domains automatically establish that treatment effects interact? Explain.
Solutions. (1) The contrast directly concerns participants who met that domain’s eligibility criteria and were randomized within it. (2) Both contrasts contain the same estimated control mean; variation in that mean shifts both contrasts in the same direction. (3) The number of unique participants is \(90+70-40=120\). The number of domain assignments is \(90+70=160\), because the 40 participants eligible for both each have two assignments. (4) No. Joint randomization can provide information about combinations of assignments, but estimating an interaction requires a defined interaction estimand, adequate information in the relevant combinations, and an appropriate analysis.