Suitability of Spatially Resolved Gridded Climate Datasets for Assessing Human Health Effects of Heat

The associations between heat and human morbidity and mortality are well documented. However, most studies investigating this meteorological effect on human health rely on weather data from a single observatory per city (typically the station at the nearest airport), with the implicit assumption that the weather measured at this station applies to the entire population of interest. This common approach masks intra-urban spatial variability of key meteorological exposures, particularly where the urban heat island effect is strongest. Moreover, because of the paucity of weather observatories in rural areas, this approach systematically excludes an important proportion of the U.S. population from health assessments. Gridded climate datasets (GCDs) address these shortcomings by spatially interpolating weather data between stations, using predictive algorithms that adjust for geophysical properties.

Although GCDs present valuable opportunities both to expand the spatial extent of heat-stress studies and to potentially reduce exposure measurement error in such analyses, the use of GCDs in epidemiology has been hampered by a scarcity of data evaluating their performance in representing health-relevant weather metrics. In particular, epidemiological assessments of heat incorporate the physiological modulation of humidity on the effect of temperature, either through instrumentation of the heat index (a measure of apparent temperature that includes ambient temperature and relative humidity) or by statistically isolating the effect of temperature by adjusting for relative humidity. Thus, evaluation of the ability of GCDs to provide spatially resolved estimates of temperature, humidity, and heat index is required for future epidemiological studies of heat stress.

Here we present a comparison between two publicly available GCDs: Daymet version 3.0 from Oak Ridge National Laboratory and PRISM from the PRISM Climate Group at Oregon State University, which provide daily weather data at spatial resolutions of one and four square-kilometers, respectively. We obtained hourly weather observations from the NOAA Integrated Surface Database Lite (ISD-Lite) and calculated minimum, maximum, and mean values of temperature, vapor pressure, and heat index. We then compared these observed daily values to the corresponding calculations derived from values provided by Daymet and PRISM at over 500 points throughout the U.S. for the time period from 1 January 1981 to 31 December 2016, resulting in over five million station-days of data for validation.

We demonstrate that while Daymet and PRISM values are both highly correlated with respective observations of minimum, maximum, and mean ambient temperatures (r² values all greater than 0.96), PRISM generally provides less-biased estimates for most meteorological measures, with lower root-mean-square errors and best-fit lines generally having slopes closer to one. PRISM additionally is a substantially better estimator of observed mean daily vapor pressure: root-mean-square errors in our sample are nearly 80% lower in PRISM than in Daymet. Nonetheless, both Daymet and PRISM allow for reasonable calculations of summertime (JJA) mean daily heat indices. Importantly, PRISM, but not Daymet, provides estimates of minimum and maximum vapor pressure deficit, which facilitate direct calculations of minimum and maximum daily heat indices. Our results suggest that spatially explicit health assessments requiring only daily measurements of ambient temperature or mean daily heat index can reliably use either PRISM or Daymet, but assessments requiring minimum or maximum heat index or any daily metric of humidity should instead use PRISM.