Effects of Mitigation and Control Policies in Realistic Epidemic Models Accounting for Household Transmission Dynamics

Abstract

Background: Compartmental infectious disease (ID) models are often used to evaluate non-pharmaceutical interventions (NPIs) and vaccines. Such models rarely separate within-household and community transmission, potentially introducing biases. We formulated an approach that incorporates household structure into ID models, extending the work of House and Keeling. Methods: We developed a multi-compartment susceptible-exposed-infectious-recovered-susceptible-vaccinated (MC-SEIRSV) modeling framework, allowing non-exponentially distributed duration in Exposed and Infectious compartments, that tracks within-household and community transmission. We simulated epidemics that varied systematically by community and household transmission rates, waning immunity rate, household size (3 or 5 members), and numbers of Exposed and Infectious compartments (1-3 each). We calibrated otherwise identical models without household structure to the early phase of each parameter combination's epidemic curve. We compared each model pair in terms of epidemic forecasts and predicted NPI and vaccine impacts on: the timing and magnitude of the epidemic peak and on its total size. Meta-analytic regressions characterized the relationship between household structure inclusion and the size and direction of biases. Results: Otherwise similar models with and without household structure produced equivalent early epidemic curves. However, forecasts from models without household structure were biased. Without intervention, they were upward-biased on peak size and total epidemic size, with biases also depending on the number of Exposed and Infectious compartments. Model-estimated NPI effects of a 60% reduction in community contacts on peak time and size were systematically overestimated without household structure. When the NPI yielded a 20% reduction, biases were smaller. Because vaccination impacted both community and household transmission, their biases were still smaller. Conclusions: ID models without household structure can produce biased outcomes. Analysts should consider incorporating household structure.