Statistical Analysis

This domain employed an adaptive 2-stage design with an initial evaluation period given limited experience with the study treatments in critically ill patients (eFigure 1 in Supplement 2). During the evaluation period, interventions were required to demonstrate an acceptable safety profile as judged by the data and safety monitoring board and an intermediate probability of efficacy, defined as at least 50% posterior probability of at least 20% improvement in the proportional odds ratio (OR) for organ support–free days for ACE inhibitor and ARB initiation compared with control or at least 30% for the combined ARB and DMX-200 intervention compared with both ARB and control to proceed to stage 2. Stage 1 was planned up to maximum sample sizes of 300 patients in each of the ACE inhibitor and ARB groups and 200 patients in the combined ARB and DMX-200 intervention group. Graduation rules were prespecified and would be implemented in a blinded fashion. Interventions that satisfied graduation criteria would continue to the uncapped evaluative period, which would enroll until platform-level adaptive stopping triggers for efficacy (posterior probability >99% of OR >1.0 compared with control) or futility (posterior probability >95% of OR <1.2 compared with control) were reached. The futility trigger could be reached at any adaptive analysis. Interventions failing to graduate would be withdrawn in stage 1. Enrollment was closed in stage 1 for safety concerns prior to an adaptive analysis being performed.

The primary analysis was an intention-to-treat analysis and included all consenting patients with suspected or proven COVID-19 with available primary outcome. The primary analysis was a bayesian cumulative logistic model adjusted for age, sex, site, and enrollment period (in 2-week intervals), and included covariates reflecting intervention and domain eligibility. Treatment effects were estimated only from patients randomized in the domain. Patients with COVID-19 enrolled in REMAP-CAP but outside of this domain did not contribute to estimates of RAS inhibitor effects, but did contribute to overall model covariate coefficient estimation.

The primary model was fit using a Markov chain Monte Carlo algorithm with 20 000 samples from the joint posterior distribution. The model calculated posterior distributions for the proportional OR, including medians and 95% credible intervals (CrIs), and the posterior probabilities of efficacy for each intervention compared with control. The probability of harm is the complement of the probability of efficacy (ie, posterior probability OR <1.0). Distinct treatment effects were estimated in critically ill and non–critically ill patients by nesting intervention effects in a hierarchical prior distribution centered on an overall intervention effect estimated with a neutral prior; the posterior distributions for these effects were shrunk toward the overall estimate to an extent reflective of their similarity (dynamic borrowing).²⁷

Secondary analyses were performed using bayesian logistic regression models for ordinal and dichotomous outcomes, bayesian linear models for continuous outcomes, and bayesian piecewise exponential models for time-to-event outcomes. No formal hypothesis tests were performed on secondary outcomes, and summaries of posterior distributions are provided for descriptive purposes only. Sensitivity and other secondary analyses were performed by investigators blinded to ongoing interventions and did not include adjustment for treatment assignment in ongoing domains.

Prespecified subgroup analyses assessed treatment effect by age (<50, 50-70, or >70 y), sex, baseline invasive mechanical ventilation status, estimated glomerular filtration rate (<90 mL/min/1.73 m², ≥90 mL/min/1.73 m², or unknown), and baseline vasopressor receipt. Machine learning with causal forests^28,29 estimated subgroup- and individual-level heterogeneity of treatment effects by considering all available baseline covariates in separate and pooled treatment analyses. Expected absolute risk differences were estimated for conditional average treatment effects at the levels of the individual and the subgroup (see eAppendix 1 in Supplement 2).

Analysis details are provided in the statistical analysis plan in Supplement 1. The primary and key secondary analyses were performed using R, version 4.1.3 (R Foundation). The causal forests heterogeneity of treatment effect analyses were conducted using R, version 4.0.5, with the R package grf, version 2.1.0.