Development and Selection of an Appropriate Model of City Level Dengue Fever Incidence in China Using Autocorrelation, Population, Weather, Climate, Vegetation, and Land Use Predictors

摘要

Background/Aim: Dengue Fever is a vector borne disease that predominantly effects countries in tropical latitudes. In recent years the incidence rate in many countries has been increasing identifying the need for accurate predictive models. We develop and assess several models of dengue fever incidence at the 'City' administrative level in China using autocorrelation, population, weather, climate, vegetation, and land use variables. Methods: The dataset was first cleaned and processed to weekly time scales. Initial descriptive analysis included incidence rates, correlation, spatial and temporal autocorrelation analysis. Several models were then built and assessed for their suitability. These included the Zero-Inflated Negative Binomial Model, the Distributed Lag Non-Linear Model, the Generalised Additive Model with random effects, and the Random Forest Model. Results: Models were successfully fitted for all four methods. While many studies identified through a literature review used some form of Negative Binomial or Poisson model, this method was not suitable for this data because of both the variation in cases across different City categories, and the nonlinear nature of the relationship between cases and weather variables. The Distributed Lag Nonlinear Model also did not yield satisfactory results, with the model substantially overpredicting almost all case counts. The Generalised Additive Model fit the observed data very well, however, it was reliant on the random effects specified for 'City' to produce such a good fit and could not be used for predictions as a result. Finally, a Random Forest Model was investigated. This model was successful in fitting the observed cases without using the City as a covariate, it allowed non-linear effects to be considered, and achieved consistent results across numerous random seeds. Conclusion: The Random Forest algorithm is a suitable method to model case counts of Dengue Fever in China.