Advantages of a Two-Stage Randomized Trial Design to Evaluate Antimicrobial Treatment Strategies: a Simulation Study

BackgroundAntimicrobial prescribing policies affect not only treated patients but also their contacts. Two-stage randomized (2SR) designs can be used to estimate these spillover effects, yet this study design has not been widely applied to evaluate antimicrobial strategies. MethodsWe developed a stochastic agent-based model that simulates a hospital ward with two competing bacterial strains (drug-A-susceptible and drug-A-resistant). We used the simulation to emulate a 2SR trial: six hospital ward clusters were randomized 1:1 to either a 90/10 (90% Drug A, 10% Drug B, drug B was assumed to have no known resistance) or 50/50 treatment allocation strategy; individuals within clusters were then randomized to Drug A or Drug B following the assigned cluster-level allocation strategy. We estimated direct, indirect, total, and overall causal effects on incident infection and mortality. Sensitivity analyses varied the treatment effect, transmission rate, mortality structure, and number of clusters. ResultsThe direct effect of drug choice showed that Drug A recipients had higher mortality (due to non-concordant treatment of resistant infections). This effect varied over time as the wards strain ecology diverged between strategies. There was also an indirect effect for Drug A recipients--reflecting spillover from higher resistant-strain prevalence under 90/10--but it was approximately null for Drug B recipients, whose broad-spectrum coverage insulated them from changes in the ward strain distribution. The overall effect--the policy-level comparison--showed that the 50/50 strategy reduced total mortality, but this net benefit concealed a redistribution: resistant-strain deaths decreased while susceptible-strain deaths increased, a consequence captured by the overall effect but invisible to the direct effect. These findings were qualitatively consistent across all sensitivity scenarios. ConclusionsWe demonstrate that antimicrobial prescribing produces spillover effects not captured by conventional individually randomized trials. These effects can substantially alter treatment outcomes in a population. We propose that the 2SR design, grounded in a formal causal framework for interference, is better suited for evaluating population-level effects of antimicrobial strategies--whether implemented as a randomized trial or emulated with observational data.