Designing infection prevalence and seroprevalence surveys for learning infection dynamics

Knowledge of the true number of infections over time is valuable for accurately predicting the future course of an epidemic and planning effective interventions, but the number of cases reported offers only a noisy underestimate of the true number of infections. Disease surveillance strategies based on assessing subsets of the population for current infection (infection prevalence surveys) or antibody presence (seroprevalence surveys) yield crucial information about the number truly infected, but are expensive. To explore impact of survey design considerations—both sample size and sampling frequency—on inference of the number of incident infections over time, we coupled agent-based simulations of respiratory virus epidemics with simulations of infection prevalence and seroprevalence surveys. While returns diminish with increased sample size, we find inference generally improved by increasing survey frequency relative to participants-per-round for any given sample size. After survey rounds reach a sufficient frequency, comparable inference performance may be achieved with either more frequent rounds or more participants per round. Rolling designs with tests conducted each day tend to outperform designs in which testing is divided into discrete rounds. We also show that misspecified assumptions about seroreversion may substantially decrease the quality of inference results.