During the recent Zika virus (ZIKV) and chikungunya virus (CHIKV) epidemics in the Americas, transmission intensity was highly spatially heterogeneous both within and between countries. Models of transmission dynamics are routinely fit to spatially-aggregated epidemiological time series even though transmission is typically heterogeneous at fine spatial scales. Assuming well-mixed mosquito and human populations at these coarse spatial scales can limit the ability of models to replicate or predict complex epidemiological patterns. Agent-based models can incorporate spatial variation in transmission risk and may be used to estimate spatial disease dynamics even when detailed time series at finer spatial scales are not available. To determine the importance of spatial scale and heterogeneity in both model structure and the model fitting process we used a climate-forced, agent-based model to recreate the recent ZIKV and CHIKV outbreaks in several Latin American countries. Using this model, we estimated viral importation patterns via an ensemble of national (single patch) or sub-national (multi-patch) simulations fit to national-level case data. We examined whether including information on the location and timing of virus introductions within a country improved model fit, and also evaluated sub-national model outputs against sub-national level observations not used in the model-fitting process.
Results/Conclusions
Initially, we modeled the 2015-2018 ZIKV epidemics in Mexico, Guatemala, Costa Rica, Panama, Colombia, and Peru using national-level weekly case data and information on the timing of the first reported case in each first-level administrative unit. Multi-patch ZIKV models better fit national-level data than a single-patch model for all countries except Panama based on the mean squared error. Assuming that the initial ZIKV importation occurred in the same location as the first reported case improved the model fit for Guatemala, Costa Rica, and Peru; however, for the other countries model simulations with introduction into the largest administrative unit performed at least as well. Incorporating sub-national data in the model fitting process, in the form of the timing of the first reported case in each level-one administrative unit, improved model fits for Mexico, Costa Rica, and Panama, but not the remaining countries. In addition, model estimates of sub-national cumulative incidence were better correlated with observed incidence than a null model assuming spatially homogeneity. Our results highlight the importance of simulating transmission dynamics at sub-national spatial scales. Fitting these simulations to aggregated national data improved model fit and parameter estimates compared to non-spatial models and also improved estimates of sub-national incidence.