Dengue fever is a major public health risk in Brazil, with millions of reported cases per year. The dengue virus is spread by the mosquito vector Aedes aegypti, and as such, prevention of dengue fever is accomplished through mosquito control programs. While these programs have been successful in preventing dengue cases (Pepin et al. 2013), we still lack a detailed understanding of how mosquito abundance maps to human disease risk. Previous studies have found that mosquito surveillance data do not substantially improve our ability to predict dengue cases compared to models that include lagged case data alone (Pepin et al. 2015). However, this approach ignores the fact that mosquito abundance drives human disease through a nonlinear dynamical system, and thus case reports already contain information about mosquito abundance (Sugihara et al. 2012). So, rather than use predictive performance to test the causal relationship between mosquito abundance and dengue cases, we instead ask whether and how we can reconstruct the mosquito surveillance time series from a time series of human case reports. We do this through the use of a mechanistic differential equation model embedded in a Bayesian statistical framework and fit to data from the city of Vitoria, Brazil.
Results/Conclusions
We model dengue dynamics using a deterministic SEIRS model. Infection dynamics are driven through a latent mosquito abundance process that we estimate from five years of weekly case reports. Though the model produces close fits to the observed case reports, the mosquito abundance time series estimated from these case reports does not reproduce the features of the observed mosquito surveillance data. Though some of the seasonal fluctuations of the surveillance data are reproduced, especially early in the annual outbreak, posterior predictive checks indicate that the estimated trajectory is overall a poor fit to the observations. In particular, the estimated trajectory is generally flatter than the observed trajectory, and the autocorrelation structures differ substantially. This suggests that the causal influence of mosquito abundance on human disease risk is not well captured by the surveillance data. This could be due to discrepancies in our simulated measurement process (e.g. assuming a constant mosquito capture probability when it should vary with weather) or in the mechanistic model of mosquito dynamics (e.g. failure to capture fluctuations in mosquito age structure). Future work will seek to better elucidate which mechanisms are necessary to effectively link mosquito surveillance data and human disease risk.