Evaluating the role of observational uncertainty in climate impact assessments: Temperature-driven yellow fever risk in South America.

Global gridded temperature data sets (GGTDs) differ in data sources, quality control, generation methods, and spatial-temporal resolution, introducing observational uncertainty. This uncertainty is critical not only for studies on current climate conditions but also for future climate change projections, where observational data sets are used for bias correction and downscaling of global climate model (GCM) outputs. It is hence essential to ensure that reference data sets accurately represent the true climate state and span a sufficiently long period to filter out internal variability. The selection of appropriate GGTDs is hence a crucial yet often overlooked factor in research that examines the impact of climate variability and change on vector-borne diseases such as yellow fever (YF), a climate-sensitive arboviral disease endemic to tropical regions of Africa and South America. In this study, we evaluated four GGTDs, namely the Berkeley Earth Surface Temperatures (BEST), the Climatic Research Unit Time-Series (CRUTS), the fifth-generation atmospheric reanalysis of the global climate from the European Centre for Medium-Range Weather Forecasts (ECMWF), ERA5, and its land-focused derivative, ERA5Land, for health-related impact research, specifically examining YF transmission in South America. Each data set was evaluated via grid-based analysis and validated against national weather station data, focusing on Brazil and Colombia, where YF out-break risk remains. While reanalysis generally outperformed lower-resolution products, ERA5 demonstrated a slight advantage over ERA5Land despite the latter's higher spatial resolution. Most importantly, our findings show that substantial differences among GGTDs affected the spatial representation of climate change indices, bioclimatic variables, and spatially aggregated temperature estimates at the administrative (AD) unit level, with substantial variations in the latter translating into markedly different estimates of key disease transmission parameters. In Colombia, admin-level temperature inputs differing by more than 6°C led to differences of about 0.2 in simulated reproduction numbers generated within a dynamic compartmental YF modeling framework.