XIS-humidity: a daily spatiotemporal machine-learning model for dew point in the contiguous United States

Abstract Health effects of humidity are theoretically important but have appeared inconsistently in epidemiological studies. Better measures of humidity may help. We construct a machine-learning model of dew point covering the US from 2003 through 2024, using the same XGBoost-IDW synthesis (XIS) framework as our previously published models of temperature and particulate matter. XIS-Humidity predicts three daily variables at high spatial resolution: mean humidity; T-min humidity (the humidity observed when temperature is at its minimum); and T-max humidity (similarly, for maximum temperature). In 2024, we obtain RMSEs of 1.30 K for mean humidity, 1.72 K for T-min, and 2.07 K for T-max, compared to SDs above 10 K. For mean humidity, as well as a metric of heat index computed with the help of XIS-Temperature, XIS-Humidity is more accurate than two competing models (NLDAS-2 and PRISM) in most years. We also find a greater association of social vulnerability with heat index when using XIS. Compared to previous models of humidity or heat index, XIS-Humidity should allow for more accurate and sophisticated analyses in the epidemiology of humidity and temperature.