Comparison of temperature-mortality associations estimated with different exposure metrics

Abstract

Studies of the short-term association between ambient temperature and mortality often use temperature observations from a single monitoring station, frequently located at the nearest airport, to represent the exposure of individuals living across large areas. Population-weighted temperature estimates constructed from gridded meteorological data may offer an opportunity to improve exposure assessment in locations where station observations do not fully capture the average exposure of the population of interest.

Methods: We compared the association between daily mean temperature and mortality in each of 113 United States counties using (1) temperature observations from a single weather station and (2) population-weighted temperature estimates constructed from a gridded meteorological dataset. We used distributed lag nonlinear models to estimate the 21-day cumulative association between temperature and mortality in each county, 1987-2006, adjusting for seasonal and long-term trends, day of week, and holidays.

Results: In the majority (73.4%) of counties, the relative risk of death on extremely hot days (99th percentile of weather station temperature) versus the minimum mortality temperature was larger when generated from the population-weighted estimates. In contrast, relative risks on extremely cold days (first percentile of weather station temperature) were often larger when generated from the weather station observations. In most counties, the difference in associations estimated from the two temperature metrics was small.

Conclusions: In a large, multi-site analysis, temperature-mortality associations were largely similar when estimated from weather station observations versus population-weighted temperature estimates. However, spatially refined exposure data may be more appropriate for analyses seeking to elucidate local health effects.

Figure 1.

Location of the 113 study counties within the contiguous United States.

Figure 2.

Exposure–response curves showing relative risks (RR) for the 21-day cumulative association between daily mean temperature and all-ages mortality in three counties modeled using temperature observations from weather stations (red) and population-weighted temperature estimates from PRISM (blue), 1987–2006. Vertical dashed lines are placed at the 1st and 99th percentile of the county-specific temperature distribution as observed at the weather station. RR, relative risk.

Figure 3.

Scatterplot of RRs for the 21-day cumulative association between daily mean temperature and mortality modeled using temperature observations from weather stations (x-axis) versus population-weighted temperature estimates from PRISM (y-axis) in each of 113 study counties. For each county, the value of temperature for which associations are plotted is held constant across datasets (i.e., at the 99th percentile of each county’s weather station temperature distribution). RR, relative risk.

Figure 4.

Distribution of the difference between the log(RR) for the 21-day cumulative association between temperature and mortality as estimated using PRISM versus station temperature in each of 113 counties. In each county, differences are calculated for the log(RR) at four values of temperature (i.e., the 1st, 2.5th, 97.5th, and 99th percentile of that county’s weather station temperature distribution) versus the minimum mortality temperature. Differences are calculated such that counties where the PRISM curve yields a larger estimate of the temperature-mortality association receive a positive value.