Heterogeneous local dynamics revealed by classification analysis of spatially disaggregated time series data

Background: Temporal incidence patterns provide a crucial window into the dynamics of emerging infectious diseases, yet their utility is limited by the spatially aggregated form in which they are often presented. Weekly incidence data from the 2015-2016 Zika epidemic were available only at the national level for most countries in the Americas. One exception was Colombia, where data at departmental and municipal scales were made publicly available in real time, providing an opportunity to assess the degree to which national-level data are reflective of temporal patterns at local levels. Methods: To characterize differences in epidemic trajectories, our analysis centered on classifying proportional cumulative incidence curves according to six features at three levels of spatial aggregation. This analysis used the partitioning around medoids algorithm to assign departments and municipalities to groups based on these six characteristics. Examination of the features that differentiated these groups and exploration of their temporal and spatial patterns were performed. Simulations from a stochastic transmission model provided data that were used to assess the extent to which groups identified by the classification algorithm could be associated with differences in underlying drivers of transmission. Results: The timing of departmental-level epidemic peaks varied by three months, and departmental-level estimates of the time-varying reproduction number, R(t), showed patterns that were distinct from a national-level estimate. The classification algorithm identified moderate support for two to three clusters at the departmental level and somewhat stronger support for this at the municipal level. Variability in epidemic duration, the length of the tail of the epidemic, and the consistency of cumulative incidence data with a cumulative normal distribution function made the greatest contributions to distinctions across these groups. Applying the classification algorithm to simulated data showed that municipalities with basic reproduction number, R0, greater than 1 were consistently associated with a particular group. Municipalities with R0 < 1 displayed more diverse patterns, although in this case that may be due to simplifications of how the model represented spatial interaction among municipalities. Conclusions: The diversity of temporal incidence patterns at local scales uncovered by this analysis underscores the value of spatially disaggregated data and the importance of locally tailored strategies for responding to emerging infectious diseases.