From Sparse Data to Smart Decisions: Region-Specific Policy Evaluation via Simulation

Effective disease outbreak response requires actionable, region-specific guidance, but most modeling tools rely on detailed surveillance or strong assumptions, such as random mixing. Agent-based models (ABMs) allow us to capture key heterogeneity in contact patterns and intervention mechanisms, but linking these models with data is often computationally intractable, particularly at the larger scales needed for decision-making (e.g., county- or state-level). We present a simulation-based framework that combines ABMs with surrogate modeling to infer key transmission and severity parameters using only routine case and hospitalization data. This enables local health agencies to evaluate candidate interventions while explicitly accounting for uncertainty. Applied to COVID-19 in Michigan counties, our method recovers core parameters (transmissibility, latent period, asymptomatic transmissibility, underreporting, hospitalization risk, and duration) that align with empirical estimates, while demonstrating regional variation linked to age and comorbidity patterns. We find that intervention effectiveness cannot be reliably predicted from simple demographic predictors such as age structure, population density, or workforce participation. While school closures aligned with child population in some settings, other interventions showed weak or counterintuitive relationships with demographics. Traditional ODE models with random mixing assumptions cannot capture how interventions target specific contact networks, making it impossible to assess whether demographic proxies predict intervention success. Our framework addresses this gap by explicitly modeling intervention mechanisms within heterogeneous contact structures. Solely using routine case and hospitalization data, our method enables practical, uncertainty-aware decision support for local health agencies facing COVID-19, influenza, RSV, or future novel pathogens. 1 Significance StatementDuring emerging outbreaks, policymakers must act before data are complete. We introduce a computational framework that bridges the gap between sparse local data and agent-based models, enabling evaluation of interventions under significant uncertainty. Our method captures how specific contact networks drive local transmission by moving beyond the random-mixing assumptions of compartmental models. We apply our framework to COVID-19 in diverse Michigan counties. Our method identifies which intervention policies are robust to the inherent uncertainty in early outbreaks for each specific community. This disease-agnostic approach empowers local agencies to make evidence-based decisions tailored to their unique population structures and contact patterns.