A Proposed Two-stage Bayesian Modeling for Calibration of Epidemiological Forecasts with Applications on Dengue Cases in Butuan City

Abstract

This paper presents a comparative analysis of modeling approaches for a high-frequency time series. The moving average of 567 morbidity weeks of dengue cases in Butuan City, a Highly Urbanized City (HUC), was used as the response variable, with the aggregated rainfall, mean maximum temperature, mean minimum temperature, and mean relative humidity as the set of regressors. The last 24 morbidity weeks were set to be used for validation of predictive accuracy. Some pieces of literature support the robustness of the use of Bayesian methods in drawing inferences, modeling, and predicting epidemiological data. Hence, candidate Bayesian Econometric models were applied following appropriate assumptions. The applicability of Bayesian Vector Autoregression (BVAR) for variable selection and lag inclusion purposes was empirically supported. The BVAR results show that the dependent variable was mostly sensitive only to the variabilities in both the (a) direct effects and (b) lags of the cases themselves, and rainfall. The generated lags as included regressors were used in a separate model using the Bayesian Metropolis-Hastings (BMH) Algorithm. For comparison, a Frequentist Vector Autoregression (FVAR) Model as the baseline model, and BMH Algorithm were applied, too. Predictions comparison shows that the variable and lag selection process of BVAR combined with the BMH Algorithm (BVAR-BMH) simulation resulted in promising gains in predictive accuracy against straightforwardly using FVAR, BVAR, or BMH algorithm for the original set of variables. The promising gains in predictive accuracy may be used in anticipatory actions for dengue epidemiological surveillance for the specified HUC, or other locations.