Spatial representativeness matters for Climate-Driven Dengue Forecasting

Accurate forecasting of dengue incidence requires statistical models that explicitly accommodate overdispersion, temporal dependence, and delayed environmental forcing. We develop a Bayesian negative binomial dynamic regression model to generate monthly forecasts of dengue hemorrhagic fever (DHF) incidence in Kanchanaburi Province, Thailand. Transmission persistence is captured through lagged dengue incidence, while delayed climatic effects are represented using locally observed maximum temperature and relative humidity. Model adequacy and predictive performance are assessed using posterior predictive checks and leave-one-out cross-validation (LOO-CV). The negative binomial specification consistently outperforms Poisson-based alternatives under substantial overdispersion. Importantly, forecasting performance is not determined solely by the strength of marginal climate-dengue associations. Instead, it depends critically on the spatial representativeness of climatic inputs relative to the population at risk. Models informed by climatically representative observations yield more stable and robust out-of-sample forecasts, even when marginal associations are comparatively weaker. These findings underscore the distinction between explanatory association and predictive utility in climate-driven infectious disease models and provide practical guidance for the development of climate-informed dengue early warning systems in endemic settings.Author summary: Dengue fever is a mosquito-borne disease that affects hundreds of millions of people each year, particularly in tropical regions. Because dengue transmission is strongly influenced by weather conditions such as temperature and humidity, climate data are widely used to build forecasting models that help public health authorities anticipate outbreaks. However, an important practical question is often overlooked: does the location of the meteorological station used to represent climate conditions affect the reliability of dengue forecasts? In this study, we compared dengue forecasting models that used climate data from two meteorological stations located in different environments within the same province in Thailand. One station showed stronger statistical correlations with dengue incidence, while the other was located closer to the main population center where most dengue cases occur. Using Bayesian predictive model comparison, we found that the station located nearer to the population center produced forecasting performance comparable to that of the station with the stronger marginal climate-dengue correlation. These findings suggest that strong historical correlations between climate and disease do not necessarily lead to better forecasts. Instead, the spatial representativeness of climate measurements relative to the population at risk may play a key role. Our results provide practical guidance for designing climate-informed dengue early warning systems in endemic regions.