Rapid build-up of greenhouse gases is expected to increase Earth’s mean surface temperature, with unclear effects on temperature variability1, 2, 3. This makes understanding the direct effects of a changing climate on human health more urgent. However, the effects of prolonged exposures to variable temperatures, which are important for understanding the public health burden, are unclear. Here we demonstrate that long-term survival was significantly associated with both seasonal mean values and standard deviations of temperature among the Medicare population (aged 65+) in New England, and break that down into long-term contrasts between ZIP codes and annual anomalies. A rise in summer mean temperature of 1 °C was associated with a 1.0% higher death rate, whereas an increase in winter mean temperature corresponded to a 0.6% decrease in mortality. Increases in standard deviations of temperature for both summer and winter were harmful. The increased mortality in warmer summers was entirely due to anomalies, whereas it was long-term average differences in the standard deviation of summer temperatures across ZIP codes that drove the increased risk. For future climate scenarios, seasonal mean temperatures may in part account for the public health burden, but the excess public health risk of climate change may also stem from changes of within-season temperature variability.
Emissions of greenhouse gases will change the Earth’s climate, most notably by changing temperature and temperature variability1. The Intergovernmental Panel on Climate Change (IPCC), in its newly released Fifth Assessment Report, forecasts a rise in world average temperature ranging from 0.2–5.5 °C by 2100 (ref. 1). This has increased interest in the impact of temperature on health.
Although many studies have reported associations between short-term temperature changes and increased daily deaths4, 5, 6, 7, 8, 9, 10, 11, evidence on the association between annual mortality and changes in seasonal temperature averages is scarce12, 13. As such impacts by definition cannot be short-term mortality displacement (we are looking at annual averages and displacement of deaths by a few weeks will not influence these), these results can imply important public health effects of the changing climate. In addition to the seasonal mean temperature, its variability may also play a significant role in raising the risk of mortality12, 14. There is evidence that people adapt to the usual temperature in their city14, 15. For these reasons, this study focused on both mean temperature, and temperature variability for summer and winter.
Conventionally, ambient temperature (Ta) was obtained from monitors near airports, and analysed at the city or county level16. Overlooking the temporal and spatial variation of Ta may introduce an exposure measurement error and may bias health effect estimates17. We recently presented new hybrid models for assessing high-resolution spatio-temporal Ta for epidemiological studies, based on surface temperature (Ts) measured by satellite. This approach allows prediction of daily Ta at 1 × 1 km spatial resolution throughout the New England area18. We also showed that using high-resolution Ta predicted from the hybrid models better captured the associations between Ta and adverse health outcomes than temperature data from existing monitoring stations19.
Using temporally and spatially resolved Ta estimates, the present study aims to tease apart the associations of annual all-cause mortality with seasonal mean temperature and temperature variability for both summer and winter, among the Medicare population in New England during 2000–2008, and to further separate these into long-term spatial differences and annual anomalies. Temperature variability is represented by the standard deviation of daily mean temperature within season.
The results, presented as percentage increases in mortality, are shown in Fig. 1 for spatial contrasts, annual anomalies, and the overall effect estimates. Both seasonal mean temperatures and their standard deviations were significantly associated with all-cause mortality for summer and winter (p < 0.05). A rise of 1 °C in summer mean temperature corresponded to an overall estimated 1.0% increase in mortality (95% CI: 0.6, 1.5%). This was entirely due to yearly anomalies; living in a location with long-term warmer summer temperatures in New England was associated with lower mortality rates, suggesting acclimatization. A 1 °C increase in winter mean temperature was observed to lower mortality by 0.6% (95% CI: 0.3, 0.9%). Here, it was the spatial contrast that was associated with lower mortality, and the anomalies with higher mortality, suggesting little acclimatization to cold. For each 1 °C increase in standard deviation of temperature, 1.3% (95% CI: 0.2, 2.4%) and 4.1% (95% CI: 3.0, 5.2%) increases in annual deaths were seen in summer and winter, respectively. The observed associations in summer and winter were respectively attributable to spatial contrasts and yearly anomalies.
Figure 1: Changes in mortality due to seasonal mean temperature and temperature variability across New England, 2003–2008.
Percentage increase in mortality for per 1 °C (a) and per interquartile range (IQR) (b). A sensitivity analysis controlling for the number of heat waves in each ZIP code in each year found essentially identical results. Error bars stand for 95% confidence intervals. s.d. is standard deviation. ∗p < 0.05.
Scientific understanding of how mean surface temperature responds to the increasing anthropogenic emission of GHGs has been greatly improved in recent decades1. Considering the comparable increments of both summer and winter mean temperatures predicted for the future 50 years, the benefit of warmer winters may be largely compensated by the harm of hotter summers. However, for annual anomalies, warmer winter temperature actually increases annual mortality. The degree of temperature variability, such as the within-season standard deviation, which is influenced by many factors, such as the intensity and pattern of atmospheric circulations, remains highly uncertain for future climate scenarios22, 23. For example, Petoukhov and Semenov proposed that the weakening of the polar vortex may bring about an increased frequency of very cold winters in northern America24. However, Huntingford and co-authors suggested that there is no significant change in global temperature variability in spite of changing regional patterns3. That study assessed variability across the entire year, including seasonal variability, and it is not clear whether variability changed within some seasons3. This uncertainty can be propagated into the projection of temperature-related health risk.
This study, by estimating impacts of both mean temperature and temperature variability in summer and winter simultaneously, adds considerable strength to the evidence of a significant association between mortality and prolonged exposures to temperatures, especially temperature variability. In addition to the within-season variability we also showed that between-year variability (anomalies) increased mortality risks. Hence, the variability of atmospheric temperature emerges as a key factor of the potential health impacts of climate change.
This work provides an important example of how temperature may affect human health in a temperate climate region. We would expect that the health effects of seasonal mean temperature and temperature variability can vary greatly among different climate zones12. Such variation may also exist across areas with distinct socio-economical status, which can convey vulnerability to the changing climate. A quantitative assessment of the projected risk of human health associated with future climate change can be estimated globally by incorporating more comprehensive epidemiological studies and projected climate scenario data for different climate zones. However, as an important environmental stressor emphasized in this study, the uncertainty of the projected changes in the temperature variability can induce significant errors in such estimates. Hence, better climate projections on temperature variability at multiple scales are important in assessing the risks to human health.
Study population.
We constructed a cohort using Medicare data for all residents aged 65 and older for the years 2000–2008 in New England (see Supplementary Table 1 in Supplementary Methods). This was an open cohort, including eligible persons from 1999, or at the year when they subsequently turned 65. Subjects entered the cohort for survival if they were still alive on 1 January of the year following the year that they enrolled in Medicare, and follow-up years (our time metric) were calendar years. As a national social insurance programme administered by the US federal government since 1966, Medicare guarantees Americans aged 65 and older access to health insurance25. The Medicare beneficiary denominator file from the Centers for Medicare and Medicaid services (CMS) lists all beneficiaries enrolled in the Medicare Fee-for-service (FFS), and contains information on beneficiaries’ eligibility and enrolment in Medicare, as well as the date of death.
Exposure data.
The present study uses 1 × 1 km ambient temperature (Ta) data estimated from surface temperature (Ts) measured by satellites18. Specifically, we started by calibrating the Ts − Ta relationship for each day using grid cells with both Ta and Ts measurements (model 1). This daily calibration was then used to predict Ta in grid cells in the study domain without Ta measurements, but with available Ts measurements. To fill in cells or days when no Ts measurements were available, a generalized additive mixed model was fitted with a smooth function of latitude and longitude of the grid-cell centroid (model 2). The performance of the estimated Ta was validated by tenfold cross-validation. Out-of-sample R2 was found to be very high (R2 = 0.947, yearly variation 0.933–0.958 for the years 2000–2011) for days with available satellite Ts measurements. Excellent performance was also observed, even in days with no available Ts data (R2 = 0.940, yearly variation 0.902–0.962 for the years 2000–2011). More details are published17.
By linking the ZIP code centroid to the nearest temperature grid, we assigned the grid-cell temperature exposures to each ZIP code. The predicted daily ambient temperatures allow us to calculate, for each year and each ZIP code, the mean temperature for summer (June–August), the mean temperature for winter (December–February), the standard deviation of daily mean summertime temperature, and the standard deviation of daily mean wintertime temperature. We refer to these as summer mean temperature, winter mean temperature; summer temperature variability and winter temperature variability, respectively (see Supplementary Table 2 in Supplementary Methods). Supplementary Table 2 also presents the distribution of the spatial and temporal variations for the temperature variables. Supplementary Table 3 in Supplementary Methods gives the correlation coefficients between these explanatory variables.
Covariates.
Medicare provides information on age, race and sex of all individuals. From the US Census Bureau 2000, we obtained tabulated ZIP-code-level socio-economic variables, including population density, percentage of green space, percentage of the population (age ≥65) in poverty status, and median value for owner-occupied housing units. In addition, based on ZIP-code-level primary and secondary hospital admissions for lung cancer, we calculated the long-term average hospital admission rate for lung cancer in each ZIP code as a surrogate for smoking experience. The census data were merged with individuals based on their ZIP code of residence. County-level percentage of diabetes and percentage of lack of physical activity, obtained from the CDC Behavioral Risk Factor Surveillance survey for the entire country, were adjusted as well.
Statistical methods.
In our data set, one observation is created for each Medicare participant for each year of follow-up, using the Andersen–Gill formulation of survival analysis. Survival times were calculated from enrolment date until death or 31 December 2008 (censoring), whichever came first.
We considered the following possible exposure indices: summer mean temperature, winter mean temperature, summertime standard deviation of temperature and wintertime standard deviation of temperature in each follow-up year. To separate the independent associations of mortality with mean temperature and temperature variability, all temperature-related indices were entered into the models simultaneously and treated as time-varying exposures.
We applied extended Cox proportional hazard models (Proc PHREG, SAS 9.3), which allow for time-varying covariates in survival analysis. The models were adjusted for individual risk factors, including age, sex, race, and ZIP-code-level covariates such as population density, percentage of green space, percentage of the population below poverty level, median value for owner-occupied housing units, and hospital admission rate for lung cancer (a surrogate for smoking experience), as well as county-level percentage of diabetes and percentage of lack of physical activity. To adjust for time trend, we entered an indicator for each year of follow-up. To allow for possible non-proportionality of hazard, participants were stratified by sex, five-year age groups, and race (white, black, and others), such that each sex/age/race group had its own baseline hazard.
We then assessed the association of mortality with the four exposure indices. The analyses were also repeated without mutual adjustment for seasonal mean temperatures and their standard deviation (see Supplementary Table 4 in Supplementary Methods). To separate the contribution of the spatial and temporal components of the temperature variables, we fit a model with two terms for each temperature variable. We defined each term as its mean in that ZIP code over the entire study period, plus the difference (anomaly) from that mean for the ith ZIP code in the tth follow-up year. With all eight terms put in the model, the mean for each ZIP code captures purely geographic contrasts, whereas the anomaly indicates the effects of yearly variations within each ZIP code.
The results are expressed as the percentage increase in mortality for each °C increase in an exposure index, for spatial contrasts, annual anomalies and the overall effect estimates. Percentage increases in mortality were also calculated for per IQR (Interquartile Range; IQR = third quartile − first quartile) increase, to reflect the current distribution of these variables. For example, an 8.4% (95% CI: 7.3, 9.6%) increase of mortality was observed for per IQR increase of yearly anomaly of summer mean temperature. This result can be interpreted as that, within a ZIP code, the risk of death in a relatively warm summer (third quartile in the studied period) was 8.4% higher than in a relatively cold summer (first quartile). A sensitivity analysis controlling for the number of heat waves in each ZIP code in each year was performed as well.
We also examined age (defined as ≥75 yr and 65–74 yr), sex, race (white, black, Asian, Hispanic, and other) and population density (less than 33th percentile, rural, and otherwise urban) as effect modifiers for both spatial contrasts and temperature anomalies, respectively, by adding interaction terms between such variables and each of the eight temperature indices in the model.
IPCC Climate Change 2013: The Physical Science Basis (eds Stocker, T. et al.) (Cambridge Univ. Press, 2013).
Alexander, L. & Perkins, S.Debate heating up over changes in climate variability. Environ. Res. Lett.8, 041001 (2013).
Huntingford, C., Jones, P. D., Livina, V. N., Lenton, T. M. & Cox, P. M.No increase in global temperature variability despite changing regional patterns. Nature500, 327–330 (2013).
Schwartz, J., Samet, J. M. & Patz, J. A.Hospital admissions for heart disease: The effects of temperature and humidity. Epidemiology15, 755–761 (2004).
Goodman, P. G., Dockery, D. W. & Clancy, L.Cause-specific mortality and the extended effects of particulate pollution and temperature exposure. Environ. Health Perspect.112, 179–185 (2004).
