Dengue fever is a legally notifiable infectious disease in China. Monthly notified dengue fever cases in the metropolitan area of Guangzhou City from 2001 to 2006 were retrieved from the Notifiable Infectious Disease Report System in China Centre for Disease Control and Prevention (China CDC). Monthly weather data, including minimum temperature (T min), maximum temperature (T max), total rainfall, minimum relative humidity (H ) and wind velocity, were retrieved from China Meteorological Data Sharing Service System for the years 2001-2006.
We performed Spearman rank correlation tests to examine the relationship between monthly dengue incidence and weather variables with a lag of zero to three months. The monthly dengue incidence was modeled using a generalized estimating equations (GEE) approach, with a Poisson distribution. This model enables both specification of an over-dispersion term and a first-order autoregressive structure that accounts for the autocorrelation of monthly numbers of dengue cases. A basic multivariate Poisson regression model can be written as:
The model that adjusts for first-order autocorrelation can be written as:
where T , T , Rain, Wind and H stand for monthly minimum and maximum temperatures, total rainfall, minimum relative humidity and wind velocity, respectively.
As GEE are not a full likelihood-modeling method, the Akaike information criterion (AIC) cannot be used for model selection. We therefore computed the quasi-likelihood based information criterion (QICu) developed by Pan to select the most parsimonious model. Highly correlated explanatory variables were included in separate models to avoid multicollinearity. When using QICu to compare two models, the model with the smaller statistic was preferred. We considered models with ΔQICu ≤ 2 to be equivalent and preferred the model with fewest parameters. All analyses were performed using SAS version 9 for Windows (SAS Institute, Inc., Cary, North Carolina).