|
Abstract:
|
Prediction of infectious disease dynamics is important to public health officials planning resource allocation and interventions. We develop methods for predicting disease incidence using kernel conditional density estimation (KCDE). We introduce several novel ideas in our formulation of KCDE. First, we use a discretized multivariate Gaussian kernel function which allows us to estimate the distribution of count data while using a fully parameterized bandwidth matrix. Second, we use low-pass filtered observations of lagged incidence as conditioning variables to mitigate the effects of noise that obscures the short-term trend in incidence. Third, we use periodic transformations of the observation time as conditioning variables to capture seasonality. We estimate the bandwidth and filtering parameters using cross-validation. We apply the method to prediction of influenza in the United States and Dengue fever in San Juan, Puerto Rico, and demonstrate that our contributions yield improvements in the log score of the predictive distribution relative to a naive application of KCDE and to a baseline seasonal ARIMA model.
|
