In this study, we found significant estimated effects of heat on mortality when using different exposure methods. The heat related mortality odds ratio was highest when using the monitoring station dataset among the four exposure methods linked to individual residence, and highest when using the modeled PRISM dataset among the four exposure methods linked to residential county. Also, the heat-mortality risk was different by level of urbancity. Rural region had the higher heat-mortality risk than urban areas or urban cluster when using either the modeled, gridded exposure or the monitoring station dataset. When using the modeled temperature dataset, the heat-mortality risk was lower using county-aggregated exposure compared to individual level for urban cluster and rural region. However, the heat-related mortality risk was higher in urban cluster region using the individual-level exposure data compared to the county-aggregated monitoring station exposure dataset. These study results indicate that using a different method to estimate exposure, such as a different dataset and different level of aggregation, while producing similar overall results, could result in different estimates of the relative health burden related to temperature.

Even though the mean temperature value among the four exposure methods were similar, the heat-related mortality risk was different. Exposure based at the individual residence resulted in a higher heat-mortality risk compared to use of exposure based on residential county using either the monitoring station temperature datasets or modeled exposure dataset. Overall, the high-resolution modeled, gridded temperature dataset resulted in lower heat-mortality risk compared to estimates based on the monitoring station temperature dataset. To date, most studies used weather station datasets when assessing temperature-related health effects. Since the monitoring stations are usually placed near populated areas, they may be close to the true average temperature exposure of the study population in ecological studies [4], however, they may not provide as good representation of temperatures in rural regions. In studies using modeled exposure, populations far from monitors are either excluded or have more uncertain exposure estimates. These populations can have different population characteristics as well (e.g., race/ethnicity, urbanicity). Our study results found different health effects when using different exposure methods, indicating that the choice of exposure method matters. Although all methods indicate increased risk of mortality under heat, the level of that risk differed by exposure method. This can have important implications for decision-makers, such as the development of heat action plans, and estimates of the health impacts of temperature. Further, estimates of the health consequences of climate change would be affected by the different risk estimates.

In our study, we found different heat-mortality risk by level of urbanicity. Importantly, whether rural or urban areas have the highest heat-mortality risk depended on the type of exposure method. This suggests that use of exposure method can also influence results in studies of environmental justice and other research on vulnerability or susceptibility. When examining the heat-mortality risk using all individuals, rural areas had the highest heat-mortality risk under most exposure approaches, but urban areas had higher heat-mortality risks than rural areas when using modeled temperature data based on residential county. This could be explained by the different individual characteristics included in the populations under different subgroup analysis. The people included in the monitoring station exposure analysis had a higher percentage of urban cluster regions and lower percentage of urban area and rural regions compared to the gridded modeled exposure analysis. This is due to the location of the monitoring stations [4], which are mostly located in more urban settings. Therefore, it is important for future studies to consider where the study site is located and what type of temperature exposure methods is being used. Often the type of exposure method selected is based on the availability of the underlying health data (e.g., aggregated health data vs. geocoded location of study participants). If the study area is urbanized, there may be sufficient monitoring stations to estimate temperature-related exposures from measurements, however, studies in rural regions may have a sparser monitoring network that might lead to exposure misclassification. In this case, modeled gridded temperature datasets may be considered, although this brings different limitations as the exposure data would be modeled and exposure uncertainty may differ by participant. The population included in the analysis based on monitoring data was more urban and had a higher percentage of non-Hispanic Black participants than the population based on the modeled temperature data. Also, when using the subset of individuals with exposure estimates under all four methods, results were different regarding risk across levels of urbancity. This could indicate that when estimating the temperature-related health burden in rural regions, using the monitoring station dataset can yield different health results. Most temperature-mortality studies have focused on cities, and most studies including rural regions relied on weather monitoring stations, which are often sparse in rural areas. Thus, the populations studied, and the estimated variation in risks across populations, could differ.

Some studies showed different results from our study. Guo et al. [12] estimated the association between temperature and mortality in Brisbane, Australia using three exposure methods: a time series temperature exposure from a single weather station, a time series averaging across three weather stations, and spatially interpolated (via kriging) temperature estimates for each administrative areas within Brisbane. These three exposures yielded similar effect sizes for both hot and cold temperatures. A study conducted in Paris found little difference in temperature-mortality relationships among different exposure definitions (single monitoring station and population-weighted daily estimates from multiple monitoring stations) [27]. However, while these studies compared results across different exposure methods, they considered only methods based on monitoring data. They also focused on a single city, whereas we considered both rural and urban areas.

Several strengths of this study are that, to the best of our knowledge, it is the first study to compare heat-mortality effect estimates based on different exposure approaches consider both the individual and county aggregated temperature based on both a modeled, gridded dataset and monitoring station dataset. We used the case-crossover study design, which inherently adjusts for all time-invariant confounders. This study included rural populations that are often excluded from epidemiological research based on monitoring networks and assessed effect modification by urbanicity. However, further studies are needed, such as exploring various potential effect modification on heat-mortality relationship when using different exposure methods, such as by access to greenspace. Also, our study results may not be generalizable to different study areas, as the populations, monitoring networks, and accuracy of modeled data may differ.

There are some limitations to this study. We did not account for how the relationship between heat and mortality could differ by air conditioning, including central air conditioning and window units. Nearly 90% of the US household have access to any air conditioning equipment (i.e., central air conditioning or window units) [28], and this rate is similar in North Carolina, where 84% of households have a central air conditioning system [28]. Data are limited on air conditioning, especially for the use of air conditioning, as opposed to its prevalence, as well as window units versus central air conditioning, and other forms of cooling such as electric or hand fans. Having access to air conditioning system does not mean each individual is actually using the air conditioning system. According to the 2015 US Census, 35% or more of households in North Carolina experience energy poverty. Individuals experiencing energy poverty may not be able to pay for the energy needed to use an air conditioning system, even if such a system is installed [29]. Studies investigating the role of air conditioning on heat-mortality relationship are largely based on air conditioning data on average across a country or at the city level [25, 30, 31]. These studies also generally used data on prevalence rather than use of air conditioning. Existing temperature-mortality studies have noted limitations in accurately reflecting the individuals’ exposed indoor temperature, even when considering air conditioning [32,33,34]. More research and data are needed on how air conditioning influences the heat-mortality relationship including types of units (e.g., central, window units), the use of air conditioning versus its prevalence, and a multi-city scale including urban and rural areas.

Another limitation of this study is in the use of ambient temperature for exposure. This does not account for differences across populations due to indoor/outdoor activity patterns and the corresponding temperatures, which would affect the human health. Some studies reported the relationship between indoor temperature and health, with findings consistent with our study results. High indoor temperature was associated with poor self-rated health in England [35]. Also, a study in Texas found indoor heat was associated with adverse health effects, especially mortality [36]. Many studies have investigated the correlation between indoor and outdoor temperature, and the results vary depending on the region and households. A study in Germany showed poor correlation between outdoor and indoor temperature [37], whereas, a study in Boston found a strong correlation of 0.91 at warm outdoor temperatures [38]. Also, a study conducted for Seoul, Korea showed that outdoor temperature and apparent temperature are sufficient indicators for indoor conditions [35]. Indoor conditions vary between households, but correspond to outdoor conditions [39, 40]. Although the ambient temperature is particularly relevant for policy, as decision-makers often develop policies for ambient levels (e.g., heat warning systems), further work is needed to assess how the heat-mortality relationship differs by subpopulation in relation to indoor temperatures and indoor/outdoor activity patterns, which can relate to different personal exposures.

There are several topics that warrant future research. In this study we analyzed the outdoor temperature as the exposure, although other studies explored a range of metrics including the heat index, which considers the temperature and humidity. Many studies found association between heat index and mortality [41,42,43]. However, the heat index used in previous studies were generated from different methods and algorithms, which could result in inconsistent results [44]. Future studies could consider more complex aspects of exposure to high temperatures such as the heat index, indoor temperatures, and indoor/outdoor activity patterns, including how different methods of assessing exposure impact estimated health risks. Furthermore, measurement error is a limitation in environmental health studies [45]. Future studies could investigate the impact of measurement error on temperature-mortality relationship, including how this differs by exposure method. Existing studies showed different results where one study found Berkson-type error that reduced the logistic regression results less than 1% [46], whereas other studies stated that measurement error could result in biased estimates and contribute to uncertainty in the results [47, 48]. Also, future studies could evaluate different types of exposure methods according to the study region’s characteristics and data availability. A multi-city multi-country study suggested that the climate reanalysis dataset well represents the monitoring station temperature dataset, expect for the tropical regions where it showed a low performance [49], however that study did not include rural areas. A spatially refined exposure dataset was found to be more appropriate for locations far from the weather stations [11].

This study showed that using different temperature exposure methods can result in different heat-mortality risk. The heat-related mortality was higher when basing exposure on monitoring station data. These findings offer useful information to researchers, communities, and policy makers, on efforts to reduce the health burden from heat by highlighting the importance of exposure assessment methods in estimating risk and in comparing risks across populations (e.g., rural versus urban), which is important for the present day and estimates of risks under climate change.