Effect of Climate Change on Heat- and Cold-Related Mortality
Estimation of mortality associated with temperature consisted of three phases: a) time-series regression analyses of daily counts of all-cause mortality and weather factors for UK regions and Australian cities; b) assessment of health impacts attributable to heat and cold under current and future climatic conditions; c) modeling the contribution of demographic changes on future mortality burdens.
All deaths in England and Wales from 1993 through 2006 and in the five largest Australian cities with populations > 1 million (Sydney, Melbourne, Brisbane, Adelaide, and Perth) from 1990 through 2006 were obtained from the UK Office for National Statistics (ONS) and the Australian Bureau of Statistics (ABS) respectively. Individual deaths were aggregated to create time series of the daily number of deaths in the study periods. Because the weather can contribute to mortality from many causes, we created series for all-cause mortality (including external causes). This was done by age group (0–64, 65–74, 75–84, ≥ 85 years), for the 10 government regions of England and Wales and the 5 largest Australian cities.
Ambient measurements of temperature and relative humidity were downloaded from the British Atmospheric Data Centre (2014) and the Australian BoM (2014). We created daily mean temperature and relative humidity series that were representative of each region or city, and included only stations with data capture of at least 75% of days during the study periods. Where such data sets from more than one station were available, we calculated daily mean values from these data using weights equal to the populations residing closest to each station. We followed the averaging process and handling of missing values described by Armstrong et al. (2011).
Time-series regression analyses were carried out using Stata (version 12; StataCorp, College Station, TX, USA) to estimate the short-term relationships between daily average temperatures and daily mortality. Poisson variation with scale overdispersion was assumed in all statistical models. Heat and cold effects were estimated separately. For heat, models were restricted to the summer months (i.e., June–September in the United Kingdom and December–March in Australia). Possible long-term trends in mortality were modeled using linear and quadratic terms for time. Intra-annual seasonal patterns in the mortality series were controlled more flexibly using natural cubic splines (NCS) of time with 4 degrees of freedom (df) per summer. NCS were also used to adjust for the estimated effects of daily relative humidity, and day-of-week effects were modeled using six indicator terms. For cold-related mortality assessment, based on all-year models to capture impacts occurring in nonwinter periods also, confounder control was the same as for the heat-related mortality models, but with additional seasonal control (8 df/year) and the inclusion of daily influenza deaths obtained from the UK ONS (comparable influenza data were not available for Australia). Although adjustment for air pollution was not carried out, previous work showed that control for daily particulate matter ≤ 10 μm (PM10) and ozone (O3) in the London region only changed the estimated relative risk (RR) for heat by a very small amount, whereas the estimated RR for cold remained unchanged (Hajat et al. 2014). Control for daily PM10, nitrogen dioxide (NO2), or O3 in Brisbane did not change the association between temperature and mortality in a previous study (Huang et al. 2012).
Relationships between daily mean temperature and daily mortality, assessed graphically using NCSs of temperature, indicated thresholds in both heat and cold models at which risk increased (data not shown). Therefore, for quantification, we used the 93rd percentile (average of lags 0–1) and the 60th percentile (average of lags 0–27) of the daily mean temperature distribution within each UK region or Australian city as the heat and cold thresholds respectively, based on evidence from earlier studies (Armstrong et al. 2011; Hajat et al. 2014). As a sensitivity test, heat and cold thresholds equivalent to the 90th percentile (average of lags 0–1) and the 65th percentile (average of lags 0–27) of the daily mean temperature distribution within each UK region or Australian city were also used to capture any mortality effects occurring at more moderate temperatures. For the daily mortality and temperature distributions, and thresholds for heat and cold effects in UK regions and Australian cities, see Supplemental Material, Figures S1 and S2 http://ehp.niehs.nih.gov/wp-content/uploads/122/12/ehp.1307524.s001.508.pdf. Because heat-related deaths occur soon after exposure (Basu and Samet 2002), they were modeled using the average of same day and previous day temperatures. Cold impacts can be delayed by weeks (Bhaskaran et al. 2010) and therefore were modeled using temperatures averaged over 28 days.
In all models, heat or cold effects are presented as the estimated RR of death for every 1°C increase or decrease in temperature above or below the threshold. RRs are estimated separately by UK region or Australian city, for all ages and separate age groups, as well as a mean RR for each country using a DerSimonian and Laird procedure for a random effects meta-analysis (DerSimonian and Laird 1986).
Projected monthly mean temperatures for the periods 2020–2029 (2020s), 2050–2059 (2050s), and 2080–2089 (2080s) were obtained from the UK Climate Impacts Programme (UKCIP) Climate Projections (UKCP09 2009) and the Australian Climate Change Scenarios Generator (OzClim 2011). These databases provide output from the Met Office Hadley Centre Climate Model (HadCM3) (Gordon et al. 2000) for the Special Report on Emissions Scenarios (SRES) (Nakicenovic et al. 2000) at a horizontal resolution of approximately 25 km. In this study, three SRES scenarios were used: low (B1), medium (A1B), and high (A1FI) emissions.
Series of daily mean temperatures were calculated separately for each emissions scenario by taking the mean of the monthly temperature for each grid square that fell within a particular regional boundary in England and Wales or by taking the most central grid square for each Australian city, and then estimating the difference from the corresponding average monthly temperature in the same region or city over the baseline period (1993–2006), and applying this difference to the daily mean temperatures over the same period. This produced three daily mean temperature time series (one for each emission scenario) per England and Wales region or Australian city.
Temperature-related mortality was initially calculated for the baseline period (1993–2006) and then for three future decades, 2020s, 2050s, and 2080s, for 10 UK regions and 5 Australian cities using the relationship:
where Mij represents the estimated temperature-related deaths per year in region i during decade j, Pij the population in region i during period j, Dik is the daily mortality rate for all-cause deaths in region i on day k, RRijk is the calculated relative risk for heat or cold effects in region i on day k of decade j, and
and
are the slopes of the temperature–mortality relationship for heat ( H) and cold ( C) respectively in region i reflecting the change in all-cause mortality per 1°C change in daily mean temperature above or below the regional threshold for heat or cold derived from the statistical time-series analysis over the baseline period. Finally,
and
are the actual or projected excursions of daily mean temperatures above or below the regional threshold for heat or cold. Mortality burdens were estimated for all ages and for four age groups (0–64, 65–74, 75–84, ≥ 85 years) separately.
The temperature–mortality relationships (i.e., slopes and threshold temperatures) and daily mortality rates were held constant over time, because our main focus was to estimate the effect of changes in temperature on mortality under the three selected emissions scenarios.
Regional population data based on the 2001 census for the United Kingdom and Australia were initially used in the health impact assessment. We subsequently repeated the calculations for future decades using regional population projections. Population data were extracted from the 2010-based principal projections for the United Kingdom (ONS 2011) and from the 2006-based mid-range (Series B) projections for Australia (ABS 2008b), and aggregated into the four age groups for the three decadal periods. The population projections are based on assumptions about future levels of fertility, mortality, internal and overseas migration used by the UK ONS (2011) and the ABS (2008b), which are broadly similar, though not identical, between the countries.
Methods
Estimation of mortality associated with temperature consisted of three phases: a) time-series regression analyses of daily counts of all-cause mortality and weather factors for UK regions and Australian cities; b) assessment of health impacts attributable to heat and cold under current and future climatic conditions; c) modeling the contribution of demographic changes on future mortality burdens.
Statistical Time-series Analyses
All deaths in England and Wales from 1993 through 2006 and in the five largest Australian cities with populations > 1 million (Sydney, Melbourne, Brisbane, Adelaide, and Perth) from 1990 through 2006 were obtained from the UK Office for National Statistics (ONS) and the Australian Bureau of Statistics (ABS) respectively. Individual deaths were aggregated to create time series of the daily number of deaths in the study periods. Because the weather can contribute to mortality from many causes, we created series for all-cause mortality (including external causes). This was done by age group (0–64, 65–74, 75–84, ≥ 85 years), for the 10 government regions of England and Wales and the 5 largest Australian cities.
Ambient measurements of temperature and relative humidity were downloaded from the British Atmospheric Data Centre (2014) and the Australian BoM (2014). We created daily mean temperature and relative humidity series that were representative of each region or city, and included only stations with data capture of at least 75% of days during the study periods. Where such data sets from more than one station were available, we calculated daily mean values from these data using weights equal to the populations residing closest to each station. We followed the averaging process and handling of missing values described by Armstrong et al. (2011).
Time-series regression analyses were carried out using Stata (version 12; StataCorp, College Station, TX, USA) to estimate the short-term relationships between daily average temperatures and daily mortality. Poisson variation with scale overdispersion was assumed in all statistical models. Heat and cold effects were estimated separately. For heat, models were restricted to the summer months (i.e., June–September in the United Kingdom and December–March in Australia). Possible long-term trends in mortality were modeled using linear and quadratic terms for time. Intra-annual seasonal patterns in the mortality series were controlled more flexibly using natural cubic splines (NCS) of time with 4 degrees of freedom (df) per summer. NCS were also used to adjust for the estimated effects of daily relative humidity, and day-of-week effects were modeled using six indicator terms. For cold-related mortality assessment, based on all-year models to capture impacts occurring in nonwinter periods also, confounder control was the same as for the heat-related mortality models, but with additional seasonal control (8 df/year) and the inclusion of daily influenza deaths obtained from the UK ONS (comparable influenza data were not available for Australia). Although adjustment for air pollution was not carried out, previous work showed that control for daily particulate matter ≤ 10 μm (PM10) and ozone (O3) in the London region only changed the estimated relative risk (RR) for heat by a very small amount, whereas the estimated RR for cold remained unchanged (Hajat et al. 2014). Control for daily PM10, nitrogen dioxide (NO2), or O3 in Brisbane did not change the association between temperature and mortality in a previous study (Huang et al. 2012).
Relationships between daily mean temperature and daily mortality, assessed graphically using NCSs of temperature, indicated thresholds in both heat and cold models at which risk increased (data not shown). Therefore, for quantification, we used the 93rd percentile (average of lags 0–1) and the 60th percentile (average of lags 0–27) of the daily mean temperature distribution within each UK region or Australian city as the heat and cold thresholds respectively, based on evidence from earlier studies (Armstrong et al. 2011; Hajat et al. 2014). As a sensitivity test, heat and cold thresholds equivalent to the 90th percentile (average of lags 0–1) and the 65th percentile (average of lags 0–27) of the daily mean temperature distribution within each UK region or Australian city were also used to capture any mortality effects occurring at more moderate temperatures. For the daily mortality and temperature distributions, and thresholds for heat and cold effects in UK regions and Australian cities, see Supplemental Material, Figures S1 and S2 http://ehp.niehs.nih.gov/wp-content/uploads/122/12/ehp.1307524.s001.508.pdf. Because heat-related deaths occur soon after exposure (Basu and Samet 2002), they were modeled using the average of same day and previous day temperatures. Cold impacts can be delayed by weeks (Bhaskaran et al. 2010) and therefore were modeled using temperatures averaged over 28 days.
In all models, heat or cold effects are presented as the estimated RR of death for every 1°C increase or decrease in temperature above or below the threshold. RRs are estimated separately by UK region or Australian city, for all ages and separate age groups, as well as a mean RR for each country using a DerSimonian and Laird procedure for a random effects meta-analysis (DerSimonian and Laird 1986).
Health Impact Assessment
Projected monthly mean temperatures for the periods 2020–2029 (2020s), 2050–2059 (2050s), and 2080–2089 (2080s) were obtained from the UK Climate Impacts Programme (UKCIP) Climate Projections (UKCP09 2009) and the Australian Climate Change Scenarios Generator (OzClim 2011). These databases provide output from the Met Office Hadley Centre Climate Model (HadCM3) (Gordon et al. 2000) for the Special Report on Emissions Scenarios (SRES) (Nakicenovic et al. 2000) at a horizontal resolution of approximately 25 km. In this study, three SRES scenarios were used: low (B1), medium (A1B), and high (A1FI) emissions.
Series of daily mean temperatures were calculated separately for each emissions scenario by taking the mean of the monthly temperature for each grid square that fell within a particular regional boundary in England and Wales or by taking the most central grid square for each Australian city, and then estimating the difference from the corresponding average monthly temperature in the same region or city over the baseline period (1993–2006), and applying this difference to the daily mean temperatures over the same period. This produced three daily mean temperature time series (one for each emission scenario) per England and Wales region or Australian city.
Temperature-related mortality was initially calculated for the baseline period (1993–2006) and then for three future decades, 2020s, 2050s, and 2080s, for 10 UK regions and 5 Australian cities using the relationship:
where Mij represents the estimated temperature-related deaths per year in region i during decade j, Pij the population in region i during period j, Dik is the daily mortality rate for all-cause deaths in region i on day k, RRijk is the calculated relative risk for heat or cold effects in region i on day k of decade j, and
and
are the slopes of the temperature–mortality relationship for heat ( H) and cold ( C) respectively in region i reflecting the change in all-cause mortality per 1°C change in daily mean temperature above or below the regional threshold for heat or cold derived from the statistical time-series analysis over the baseline period. Finally,
and
are the actual or projected excursions of daily mean temperatures above or below the regional threshold for heat or cold. Mortality burdens were estimated for all ages and for four age groups (0–64, 65–74, 75–84, ≥ 85 years) separately.
The temperature–mortality relationships (i.e., slopes and threshold temperatures) and daily mortality rates were held constant over time, because our main focus was to estimate the effect of changes in temperature on mortality under the three selected emissions scenarios.
Demographic Changes
Regional population data based on the 2001 census for the United Kingdom and Australia were initially used in the health impact assessment. We subsequently repeated the calculations for future decades using regional population projections. Population data were extracted from the 2010-based principal projections for the United Kingdom (ONS 2011) and from the 2006-based mid-range (Series B) projections for Australia (ABS 2008b), and aggregated into the four age groups for the three decadal periods. The population projections are based on assumptions about future levels of fertility, mortality, internal and overseas migration used by the UK ONS (2011) and the ABS (2008b), which are broadly similar, though not identical, between the countries.
Source...