10 VOLUME 122

|
NUMBER 1
|
January 2014 Environmental Health Perspectives
Research
All EHP content is accessible to individuals with disabilities. A fully accessible (Section 508compliant)
HTML version of this article is available at [Link]
Introduction
Heat is a well-known natural hazard. During
the summer, high temperatures may result in
heat exhaustion, heat syncope, heat stroke,
and heat cramps in susceptible individuals,
leading to excess mortality at the population
level (Gasparrini and Armstrong 2011; Kovats
and Hajat 2008). Heat waves, commonly
dened as a few consecutive days with high
temperature above a certain threshold, are
the leading cause of weather-related mortality
in the United States (Davis et al. 2003). For
example, the severe heat wave in Chicago,
Illinois, in July 1995 resulted in approxi-
mately 700 deaths (Semenza et al. 1996).
In California, the estimated excess mortality
attributable to heat waves in July 2006 ranged
from 160 to 333 deaths (Ostro et al. 2009).
According to the Intergovernmental
Panel on Climate Change (IPCC), air tem-
perature is projected to rise by 1.86.4C by
2100 (IPCC 2007). The frequency, inten-
sity, and duration of heat waves will likely
increase in the future, thus aggravating heat-
related mortality unless population adapta-
tion occurs. Consequently, preparedness
for adverse outcomes of future heat waves is
necessary. However, accurately estimating
the health impact of heat waves is challenging
because of the uncertainties related to future
temperature projections, heat wave metrics,
population growth patterns, population sus-
ceptibility to heat waves, and spatial hetero-
geneity of heat waves. A recent study that
considered several different climate change
scenarios and seven global climate models
estimated that future heat waves in Chicago
would cause 1662,217 deaths/year during
20812100 (Peng et al. 2011). However,
Peng et al. (2011) projected future heat
wave mortality in only one city and did not
consider spatially resolved temperature data.
Because the health impacts of heat waves can
have signicant spatial variability, mitigation
strategies are unlikely to be universally eec-
tive. To take local needs into account, spa-
tially resolved estimates of health outcomes
due to heat waves are very important.
In the present study, we estimated county-
level excess mortality attributable to heat
waves across the eastern United States during
20572059 using spatially and temporally
resolved regional climate model simulation
results and also examined factors that con-
tributed to uncertainty. First, we used state-
of-the-art high-resolution climate projections
to estimate future heat waves at the county
level. Next, we estimated the future excess
mortality related to heat waves considering
several population growth patterns. Finally,
we attributed the uncertainties in the excess
mortality estimates to various factors using a
variance-decomposition method.
Methods
Study area. We focused on the eastern
United States (east of 95 longitude), an area
that covers approximately 1,700 counties, has
a large population (> 180 million), and has
diverse weather conditions and geography.
Cl i mat e model s i mul at i on dat a.
Developed at the National Center for Atmos-
pheric Research (NCAR), the Community
Earth System Model (CESM1.0) is a coupled
climate model that simultaneously simulates
the Earths atmosphere, ocean, land surface,
and sea ice (Gent et al. 2011). For the present
Address correspondence to Y. Liu, Department of
Environmental Health, Rollins School of Public
Health, Emory University, 1518 Clifton Rd. NE,
Atlanta, GA 30322 USA. Telephone: (404) 727-
2131. E-mail: [Link]@[Link]
Supplemental Material is available online (http://
[Link]/10.1289/ehp.1306670).
The study was supported mainly by the Climate
and Health Program at the Centers for Disease
Control and Prevention (CDC; grant 5 U01
EH000405 managed by G. Luber) and by the
National Institutes of Health (grant 1R21ES020225).
The climate model simulations were supported in
part by the National Science Foundation through
TeraGrid resources provided by National Institute
for Computational Sciences (grants TG-ATM110009
and UT-TENN0006). The data simulations used
resources of the Oak Ridge Leadership Computing
Facility at the Oak Ridge National Laboratory, which
is supported by the Office of Science of the U.S.
Department of Energy (DOE) (contract DE-AC05-
00OR22725). Y.G. was supported in part by the
Oce of Science of the U.S. Department of Energy
as part of the Regional and Global Climate Modeling
Program. Te Pacic Northwest National Laboratory
is operated for DOE by Battelle Memorial Institute
(contract DE-AC05-76RL01830).
Te authors declare they have no actual or potential
competing nancial interests.
Recei ved: 19 Febr uar y 2013; Accept ed:
1 November 2013; Advance Publication: 6 November
2013; Final Publication: 1 January 2014.
Estimation and Uncertainty Analysis of Impacts of Future Heat Waves
on Mortality in the Eastern United States
Jianyong Wu,
1
Ying Zhou,
1
Yang Gao,
2,3
Joshua S. Fu,
2
Brent A. Johnson,
4
Cheng Huang,
5
Young-Min Kim,
1

and Yang Liu
1
1
Department of Environmental Health, Rollins School of Public Health, Emory University, Atlanta, Georgia, USA;
2
Department of Civil
and Environmental Engineering, University of Tennessee, Knoxville, Tennessee, USA;
3
Atmospheric Science and Global Change
Division, Pacic Northwest National Laboratory, Richland, Washington, USA;
4
Department of Biostatistics and Bioinformatics, Rollins
School of Public Health, Emory University, Atlanta, Georgia, USA;
5
Department of Global Health, School of Public Health and Health
Services, George Washington University, Washington, DC, USA
BACKGROUND: Climate change is anticipated to influence heat-related mortality in the future.
However, estimates of excess mortality attributable to future heat waves are subject to large uncer-
tainties and have not been projected under the latest greenhouse gas emission scenarios.
OBJECTIVES: We estimated future heat wave mortality in the eastern United States (approximately
1,700 counties) under two Representative Concentration Pathways (RCPs) and investigated sources
of uncertainty.
METHODS: Using dynamically downscaled hourly temperature projections for 20572059, we pro-
jected heat wave days that were dened using four heat wave metrics and estimated the excess mor-
tality attributable to them. We apportioned the sources of uncertainty in excess mortality estimates
using a variance-decomposition method.
RESULTS: Estimates suggest that excess mortality attributable to heat waves in the eastern
United States would result in 2007,807 deaths/year (mean 2,379 deaths/year) in 20572059.
Average excess mortality projections under RCP4.5 and RCP8.5 scenarios were 1,403 and 3,556
deaths/year, respectively. Excess mortality would be relatively high in the southern states and
eastern coastal areas (excluding Maine). Te major sources of uncertainty were the relative risk esti-
mates for mortality on heat wave versus nonheat wave days, the RCP scenarios, and the heat wave
denitions.
CONCLUSIONS: Mortality risks from future heat waves may be an order of magnitude higher than
the mortality risks reported in 20022004, with thousands of heat waverelated deaths per year in
the study area projected under the RCP8.5 scenario. Substantial spatial variability in county-level
heat mortality estimates suggests that eective mitigation and adaptation measures should be devel-
oped based on spatially resolved data.
CITATION: Wu J, Zhou Y, Gao Y, Fu JS, Johnson BA, Huang C, Kim YM, Liu Y. 2014. Estimation
and uncertainty analysis of impacts of future heat waves on mortality in the eastern United States.
Environ Health Perspect 122:1016; [Link]
Uncertainties in heat wave mortality projections
Environmental Health Perspectives VOLUME 122
|
NUMBER 1
|
January 2014 11
study, we used CESM-projected coarse-
resolution temperature data for 20572059
under two Representative Concentration
Pathway (RCP) emissions scenarios. RCPs,
the most recent approach to emissions
trajectories used by the IPCC (Inman 2011;
Moss et al. 2010), include four emissions
scenarios (RCP8.5, RCP6.0, RCP4.5, and
RCP2.6), representing the radia tive forcing
levels of greenhouse gases (GHGs) and other
forcing agents in 2100. To capture the range
of the possible future climate conditions for
the present analysis, we used the RCP4.5
scenario, a low-medium scenario of climate
change that assumes moderate emissions and
the use of a range of technologies and strategies
for reducing GHG emissions (Thomson
et al. 2011), and the more extreme RCP8.5
scenario, which assumes fossil fuelintensive
energy consumption, with increasing GHG
emissions (van Vuuren et al. 2011). Te 2050s
are generally regarded as the mid-range of
climate projection and can potentially be used
to capture enough climate change signal while
avoiding excessive uncertainties associated
with projecting too far into the future (Nolte
et al. 2008). We selected the last 3 years of
the 2050s (20572059) in order to capture
signals and heat wave impacts for far enough
into the decade.
We used the Weather Research and
Forecasting (WRF) model to dynamically
downscale the CESM projections. Daily
temperature data in 20012004, including
daily average temperature (T
avg
), daily
maximum temperature (T
max
), daily minimum
temperature (T
min
), and dew point, were
generated by WRF on a 4-km 4-km grid
in the study area. A detailed description
of the CESM/WRF modeling was reported
previously by Gao et al. (2012). We aggregated
251,262 4-km 4-km grid cells to 1,703
counties to match population projection data
and reduce computational demands. We
determined the population- weighted centroid
for each county and then averaged the data
in the nine grid cells closest to the centroid.
To reduce the bias of the WRF temperature
simulation results, we used the weather station
observations collected in the Meteorological
Assimilation Data Ingest System (MADIS)
(see Supplemental Material, Figure S1) as the
reference to calibrate the WRF temperature
data. Specifically, we first calculated the
dierence between the MADIS and WRF data
at each station each day, and then excluded the
data points beyond the 99th percentile of the
daily temperature data in the MADIS stations.
For counties with more than one MADIS
station, we averaged the MADIS and WRF
data and then obtained the ratios of WRF to
MADIS data in each county. We interpolated
the calibration ratios from 625 counties with
MADIS stations to the 1,703-county study
area using either a fixed search radius or
exible search radii. For estimates based on a
fixed search radius, we derived county-level
calibration ratios based on the average of all
calibration ratios in a 150-km radius centered
on the population-weighted centroid of
each county. The 150-km search radius was
selected to ensure that there would be at least
one calibration ratio in every search radius.
For the second method, we identied the ve
calibration ratios closest to the county centroid
and used the average of these ratios as the
calibration ratio for each county.
Heat wave definitions. Because there is
no universally accepted definition of a heat
wave, we used slightly modified versions of
four frequently used heat wave definitions
(HWDs) (Table 1) to project the frequency
and duration (in days) of heat waves from
1 May through 31 September of each year
at baseline (20012004) and in the future
(20572059). Specic criteria vary among the
definitions, but in brief, the first heat wave
denition (HWD_HI) is based on daily low
and high heat index values (Robinson 2001),
the second (HWD_Tavg) is based on the
number of days with a T
avg
above the 95th
percentile (Anderson and Bell 2011), the third
(HWD_Tmax) is based on T
max
values (Meehl
and Tebaldi 2004; Peng et al. 2011), and the
fourth (HWD_Tmin) is based on the number
of days with a T
min
above the 95th percentile.
According to HWD_Tavg, some days
with relatively mild temperatures might be
counted as heat wave days in counties with
a low T
avg
during May through September.
Terefore, we modied the original denition
used by Anderson and Bell (2011) by setting
26.7C (80F) as the minimum threshold of
T
avg
used to dene a heat wave. For the same
reason, we set 21.3C (70.8F) as the T
min

threshold, and 32.7C (90.9F) as the T
max

threshold, when dening heat waves based on
HWD_Tmin or HWD_Tmax, respectively.
These thresholds were selected because the
average T
min
and T
max
dier from the average
T
avg
from 1 May31 September by 5.4C
and 6C in our data set, respectively (see
Supplemental Material, Figure S2).
Future population projections. We based
our estimates of county-level populations for
20572059 on state-level projections con-
ducted by the U.S. Census Bureau using the
cohort-component method (U.S. Census
Bureau 2009), in which the components of
population change (births, deaths, and net
international migration) were projected from
the 2000 base population to the year 2050
for each birth cohort (Preston et al. 2001). To
conduct this projection, the Census Bureau
used multiple data sources to generate infor-
mation about fertility, mortality, and migra-
tion. In addition, the Census 2009 projection
incorporated four net international migration
assumptions: a) high migration, b) constant/
medium migration, c) low migration, and
d) zero migration. For the present analy-
sis, we projected future populations at the
county level using the constant ratio method
(Shryock et al. 1973). Specifically, for each
county, we multiplied the census-projected
state population in 2050 by the ratio of the
county population to the state population at
baseline in 2000. Te approach assumes that
each countys share of the state population
remains constant over time.
Estimation of heat wave mortality. The
expected number of excess deaths attributable
to heat waves (ED
hw
) in each county was cal-
culated using Equation 1 (Peng et al. 2011):
ED
hw
= N (RR 1) L, [1]
where N is the expected number of deaths on
nonheat wave days in each county, which
is equal to the expected daily mortality rate
on nonheat wave days multiplied by county
population, and RR is the relative risk of
Table1. Heat wave denitions used as the basis for the present analysis.
Type Name Heat wave criteria References
HWD_HI Heat indexbased denition Days of which the low and the high daily heat index
(Hi) are no less than the NWS thresholds of 26.7C
(80F) and 40.5C (105F), respectively.
Robinson 2001
HWD_Tavg Daily average temperature
based denition
At least 2 consecutive days with daily mean
temperature (T
avg
) >95th percentile of T
avg
during
20012004 in summer.
a
Anderson and
Bell 2011
HWD_Tmax Daily maximum temperature
based denition
A heat wave meets three criteria: a)daily maximum
temperature (T
max
) >97.5th percentile of T
max
of
summer days during 20012004 for at least 3 days,
b)the average of T
max
greater than this threshold
for the entire period, and c)T
max
>81.5th percentile
of T
max
of summer days during 20012004 for every
day during the entire period.
a
Meehl and
Tebaldi 2004;
Peng etal.
2011
HWD_Tmin Daily minimum temperature
based denition
At least 2 consecutive days with daily minimum
temperature (T
min
) >95th percentile of T
min
during
20012004 in summer.
a
Zhang etal.
2012
NWS, National Weather Service.
a
The denitions of heat waves were modied for the present analysis because we used temperature data collected
during 20012004 (from 1May through 31September of each year) as the baseline data.
Wu et al.
12 VOLUME 122
|
NUMBER 1
|
January 2014 Environmental Health Perspectives
death on heat wave days compared with non
heat wave days. RR 1 is the attributable risk
(AR) to heat waves (i.e., the increase in the
risk of non accidental mortality on heat wave
days compared with nonheat wave days),
and L is the length of the heat wave in days.
Because nearly 99% of counties had no
heat waves in 2001 by our definitions, we
used the county-level non accidental daily
mortality rate in 2001 as the expected daily
mortality rate on nonheat wave days during
the 20022004 baseline period. Similar to
Peng et al. (2011), we assumed that the non-
accidental mortality rate on nonheat wave
days in 20572059 was unchanged from
the baseline rate. We used the 95% confi-
dence interval (CI) estimated by Anderson
and Bell (2011) for the percentage increase
in non accidental mortality on heat wave days
compared with nonheat waves days dur-
ing 19872005 to dene the ranges of pos-
sible AR values for counties in the Northeast
(1.7911.98%), the Midwest (3.367.93%),
and the South (0.11 to 3.84%). These
estimates reflect estimated excess risks of
mortality on heat wave days defined using
HWD_Tavg (Table 1). Based on the geo-
graphic location of each state, the Northeast
regi on i ncl udes Connecti cut, Mai ne,
Massachusetts, New Hampshire, New Jersey,
New York, Pennsylvania, Rhode Island,
and Vermont. Te Midwest region includes
Illinois, Indiana, Iowa, Kansas, Michigan,
Minnesota, Missouri, Nebraska, Ohio,
and Wisconsin. The South region includes
Alabama, Arkansas, Delaware, Florida,
Georgia, Kentucky, Louisiana, Maryland,
Mississippi, Oklahoma, North Carolina,
South Carolina, Tennessee, Texas, Virginia,
Washington, DC, and West Virginia. We
generated 100 random samples of AR values
from uniform distributions based on the
ranges for each region using Latin hypercube
sampling, a modified Monte Carlo simula-
tion method (Helton and Davis 2003) and
randomly selected nine of the sampled sets
of values (each of which included three AR
values, one for each regionthe counties in
each region have the same AR values) to use
in our nal calculations.
Uncertainty and sensitivity analysis. We
considered the factors contributing to the
uncertainties in projected heat wave health
impacts in each analytical step. These fac-
tors include the two RCP scenarios (RCP4.5
and RCP8.5), the two methods used to cal-
culate the temperature calibration ratios for
each county (xed search radius and exible
search radii), the four heat wave definitions
(Table 1), the four population projections
(assuming high, constant/medium, low, or
zero net migration), and the nine sets of
sampled AR values. In addition, we estimated
excess mortality due to heat waves for 1,703
individual counties in each of 3 future years
(20572059). After calculating the total excess
mortality attributable to heat waves consid-
ering all the sources of variation mentioned
above, we derived probability distributions,
mean values, SDs, and 95% CIs for an overall
estimate and for subsets of estimates according
to different assumptions (e.g., according to
RCP scenarios and heat wave denitions).
We used the variance-decomposition
method in our sensitivity analysis of the inu-
ence of different factors on heat wave mor-
tality estimates because it not only identi es
influential factors but also apportions the
sources of uncertainty (Chan et al. 1997;
Saltelli et al. 2008). For a generic model,
y = f(x1, x2,xn), the total variance of y can
be decomposed into the partial variance attrib-
utable to each factor (x1, x2,xn) and their
interactions. In our analysis, y is the excess
mortality estimate; x1, x2,xn are n fac-
tors related to y; and the total variance was
decomposed using the following equations:
V(y) = V(y|x1) + V(y|x1, x2)
+V(y|x1, x2,xn) [2]
and
Si = V(y|xi)/V(y), [3]
where V(y) is the total variance of y, V(y|x1)
is the variance of y attributable to x1, and
V(y|x1,x2,xn) is the variance of y attribut-
able to the interactions of x1, x2,xn. Si, the
rst-order sensitivity index for factor i, reects
the main eect of each factor on the estimate
of the heat wave mortality. A large value of S
for a given factor indicates that the estimate
of excess mortality is more sensitive to this
factor than factors with a smaller S. We identi-
ed factors that have larger inuences on the
total excess mortality of the whole study area
(which are the main sources of uncertainty in
the estimation of heat wave mortality).
Results
Characteristics of heat waves. After calibrating
the WRF temperature simulations to observed
temperatures in MADIS using the calibra-
tion method based on a fixed search radius
(the distribution of calibration ratios is shown
in the Supplemental Material, Figure S3),
the estimated T
avg
( SD) over the entire
study area during MaySeptember over both
RCP scenarios was 22.74C ( 4.67) dur-
ing 20022004 and 24.75C ( 4.46) dur-
ing 20572059 (Table 2). Average estimated
numbers of heat waves per county per year
were 0.38 0.91 during 20022004 and
1.88 3.11 during 20572059. Temperatures
projected for 20572059 based on the
RCP8.5 scenario were nearly 1C higher than
projections based on the RCP4.5 scenario,
resulting in almost twice as many projected
heat waves. Te projected average duration of
heat waves in 20572059 based on both RCP
scenarios (4.53 3.09 days) is nearly 1 day
longer than the estimated average duration in
20022004 (Table 2).
When estimated using only one heat wave
definition at a time, HWD_HI resulted in
the smallest estimated numbers of future heat
waves, and HWD_Tmin resulted in the larg-
est numbers (Table 3). HWD_Tmax pro-
duced the longest estimate of average heat
wave duration (4.88 2.98 days), whereas
projections based on the other three heat
wave definitions were between 3 and 4
days. Average heat wave durations estimated
under the RCP8.5 scenario were almost
1 day longer than under the RCP4.5 sce-
nario, regardless of the heat wave denition
used (Table 3).
Table 2. Estimated current (20022004) and future (20572059) heat waves and related excess mortality
(meanSD).
Period
Daily average
temperature (C)
a
Heat wave frequency
(episodes/year)
b
Heat wave duration
(days)
b
Total excess deaths/year
(95% CI)
c
20022004 22.744.67 0.380.91 3.441.78 187173 (2, 614)
20572059
d
24.754.46 1.883.11 4.533.09 2,3792,008 (200, 7,808)
20572059 (RCP4.5) 24.114.30 1.312.57 4.063.31 1,4031,015 (137, 3,788)
20572059 (RCP8.5) 25.394.51 2.443.49 4.853.49 3,5562,265 (300, 8,577)
a
During May through September in 1,703 eastern U.S. counties.
b
Integrated averages obtained using four heat wave
denitions.
c
Integrated averages obtained using four heat wave metrics and four population projections.
d
Values shown
are integrated over the two RCP scenarios, and the total excess deaths account for all possible excess relative risk
values.
Table 3. Projected average heat wave days and episodes per year during 20572059 in each county
according to emission scenario (RCP4.5 or RCP8.5) and heat wave denition (meanSD).
Heat wave metrics
Heat wave frequency
(episodes/year/county) Heat wave duration (days)
RCP4.5 RCP8.5 RCP4.5 RCP8.5
HWD_HI 0.491.80 0.792.46 3.291.83 4.003.63
HWD_Tmax 0.500.88 1.501.75 4.882.98 5.804.31
HWD_Tavg 1.842.54 3.403.37 3.962.36 4.693.57
HWD_Tmin 2.433.61 4.063.58 3.751.47 4.33 1.97
Uncertainties in heat wave mortality projections
Environmental Health Perspectives VOLUME 122
|
NUMBER 1
|
January 2014 13
Projected average values for the duration
and frequency of heat waves in 20572059
(integrated over the four heat wave defini-
tions) suggest that there will be consider-
able spatial variability at the county level
(Figure 1). In particular, most counties in the
northern half of the study region would not
experience any heat waves under the RCP4.5
scenario (Figure 1B), whereas most coun-
ties along the southern coast would average
2/year, and Florida counties would average
4 heat waves/year. Under the RCP8.5 sce-
nario, the number of nonheat wave counties
in the Northern region would decrease, and
the average frequency of heat waves in most
of the counties in the southern states includ-
ing Alabama, Florida, Georgia, Louisiana,
and South Carolina would increase to
4/year (Figure 1C). Counties in southern
Florida have the longest average heat wave
durations projected for the study area under
both RCP scenarios (Figure 1E,F).
Future heat wave mortality. The esti-
mated average number of heat wave deaths
in the eastern United States during 2002
2004 was 187 173 deaths/year, in contrast
with 2,379 2,008 deaths/year projected
for 20572059 (integrated over four popu-
lation projections, four heat wave defini-
tions, and two RCP scenarios) (Table 2).
Projected numbers of deaths differ substan-
tially between the two RCP scenarios, with a
mean of 1,403 1,015 deaths/year projected
under RCP4.5, compared with 3,556 2,265
deaths/year under RCP8.5. Projections also
differ according to the heat wave definition
used (Table 4), with almost 2.5 times more
deaths projected based on HWD_Tavg and
HWD_Tmin than estimated using the other
two denitions, for both RCP scenarios. As
expected, estimated numbers of heat wave
deaths increase as projected populations
increase in size due to migration (Table 4).
Our excess mortality estimates have sub-
stantial spatial variability (Figure 2). Under
the RCP4.5 scenario (Figure 2A,C), the
highest county-level excess mortality esti-
mates (> 10 deaths/year) were projected for
counties in Florida, New York, and Illinois.
Under the RCP8.5 scenario (Figure 2B,D),
> 10 deaths/year are projected for a larger
number of counties in Florida in addition to
some counties along the East Coast (e.g., in
Massachusetts, New York, and New Jersey)
and a few counties in inland states includ-
ing Illinois and Michigan. Te high numbers
of projected deaths in metro politan areas of
Illinois, Massachusetts, Michigan, New Jersey,
and New York might be associated with the
high population density in these areas (see
Supplemental Material, Figure S4).
Uncertainty and sensitivity analysis.
Figure 3 shows the probability distributions
of annual total excess mortality estimates in
the eastern United States for 20572059
under RCP4.5, RCP8.5, and the two sce-
narios combined. All the factors that con-
tribute to the uncertainty in estimated excess
deaths were accounted for, including two
temperature calibration methods, four heat
wave definitions, four population projec-
tions, three model years, and nine sets of AR
values. Under the RCP4.5 scenario, the esti-
mated number of deaths is 1,403 deaths/year
(95% CI: 137, 3,788). Te distribution curve
under this scenario peaks at around 500
deaths/year, then drops rapidly. Te cumula-
tive probabilities of < 1,000 excess deaths/year
and < 2,000 excess deaths/year are 43% and
76%, respectively; we saw no probability of
> 5,000 excess deaths/year. In contrast, the
Figure1. Baseline (20022004; A,D) and future (20572059; B,C,E,F) estimates of heat wave frequency and duration in study area counties. Heat wave frequency
and duration estimates are integrated averages based on four heat wave metrics.
Wu et al.
14 VOLUME 122
|
NUMBER 1
|
January 2014 Environmental Health Perspectives
probability distribution of estimated excess
deaths under RCP8.5 is relatively at, with
a cumulative probability of 42% for having
1,0003,000 excess deaths/year, and 15%
for having 5,0007,000 excess deaths/year,
and a very small probability for having
> 10,000 excess deaths/year. When estimates
are integrated over both RCP scenarios, the
probabilities of different excess mortality
estimates fall between those of the scenario-
specic estimates.
Usi ng the vari ance-decomposi ti on
method, we attri buted 23. 7% of the
un certainty to the two RCP scenarios, 22.2%
to the four heat wave denitions, and 32.2%
to the dierent values for the relative risk of
mortality on heat wave days compared with
nonheat wave days [including nine randomly
sampled sets of three area-specic ARs based
on previously published estimates (Anderson
and Bell 2011)]. In contrast, the two methods
used for the WRF data calibration, the four
migration scenarios used in the population
projections, and the inter annual variability of
temperature projections during 20572059
were not major sources of uncertainty, as
indicated by their low Si values (Figure 4).
Discussion
We estimated that heat waves will be 3.56.4
times more frequent in 20572059 than in
20022004 and that excess mortality attribut-
able to heat waves would be 7.519.0 times
higher. Te major sources of uncertainty in our
heat wave mortality projections were the RCP
scenarios, the ARs of mortality on heat wave
days compared with nonheat wave days, and
the criteria used to dene heat waves.
In contrast with previous studies that only
focused on one or a few cities (Peng et al.
2011; Sheridan et al. 2012), we estimated the
impacts of future heat waves over a wide geo-
graphic area. Our estimates suggest that there
will be high spatial variability in future heat
wave mortality and that estimates of climate
change health impacts based on a single geo-
graphic area will have limited value for estimat-
ing health impacts in other areas. In addition,
we performed a comprehensive examination of
six factors that may contribute to uncertainty
in heat wave mortality projections, whereas
previous studies explored a relatively limited
set of factors [e.g., GHG emission scenarios
(Peng et al. 2011)]. Furthermore, we used
recently developed RCPs (Inman 2011; Moss
et al. 2010) for temperature projections at a
high spatial resolution, whereas most previous
studies (Kolstad and Johansson 2011; Peng
et al. 2011) were based on older emission sce-
narios described in the IPCC Special Report on
Emissions Scenarios (IPCC 2000).
We estimated that there would be an
average of 1.88 (95% CI: 0, 11) heat wave
episodes per year per county in 20572059.
Te projected geographic distribution of heat
waves has considerable spatial variability, with
35.7% of counties expected not to experience
any heat waves, whereas 10.4% of counties
would experience > 4 heat waves/year under
the RCP4.5 scenario, and corresponding esti-
mates of 11.6% and 26.5% under the RCP8.5
scenario, respectively. Heat waves would be
expected to occur most often in the south-
ern coastal states, including Florida, Georgia,
and Louisiana, with Florida counties having
the highest numbers of deaths attributable to
heat waves. In addition, our projections sug-
gest that despite relatively low frequencies of
heat waves, heat wave mortality will be high in
densely populated counties located in several
northeastern coastal states (e.g., Massachusetts,
New Jersey, and New York).
Increases in future temperatures will
depend largely on GHG emissions. Terefore,
GHG emission scenarios may have a substan-
tial inuence on estimates of the frequency
and duration of future heat waves. We esti-
mated future heat waves under two scenarios
representing moderate and high GHG emis-
sions. Under the high-emission RCP8.5 sce-
nario, excess mortality was estimated to be
almost 2.5 times higher than that estimated
under the moderate RCP4.5 scenario, sug-
gesting that curtailing GHG emission will
have a great impact on the reduction of heat
wave mortality in the future.
Our ndings also indicate that heat wave
definitions have a strong influence on heat
wave mortality projections, with the fewest
days per year being classied as heat wave days
Table4. Projected excess mortality per year attributable to heat waves during 20572059 under different
heat wave denitions, population projections, and emission scenarios (meanSD).
Factor Category RCP4.5 RCP8.5
Heat wave denitions HWD_HI 898509 1,668936
HWD_Tmax 707378 2,4961,515
HWD_Tavg 1,9531,033 4,4992,274
HWD_Tmin 2,0531,091 4,7592,315
Population projection Extreme high 1,5681,118 3,7462,484
High 1,4721,048 3,5172,334
Low 1,4121,006 3,3772,242
Extreme low 1,159826 2,7831,854
Figure 2. Projected county-level excess deaths [means (A,B) and SDs (C,D)] due to heat waves in 2057
2059 under RCP4.5 (A,C) and RCP8.5 (B,D) scenarios. The excess deaths for each scenario are average
estimates obtained using four heat wave metrics and four population projections.
Uncertainties in heat wave mortality projections
Environmental Health Perspectives VOLUME 122
|
NUMBER 1
|
January 2014 15
when the HDW_HI denition was used, and
the largest numbers of heat wave days when
heat waves were dened using HDW_Tavg or
HDW_Tmin. Our results are consistent with
recent studies that reported substantial incon-
sistencies among heat wave mortality estimates
when using different heat wave definitions
(Hajat et al. 2010; Zhang et al. 2012).
Estimates of the relative risk of mortality
on heat wave days compared with nonheat
wave days have varied among previous studies
and may at least partly reect regional dier-
ences in susceptibility to adverse eects of heat
at the population level as well as individual-
level dierences in susceptibility according to
race, age, occupation, or other factors (ONeill
et al. 2005). Currently, there is little informa-
tion about regional variation in the relative
risk of heat wave mortality; therefore, we used
a Monte Carlo simulation to sample AR from
values previously reported by Anderson and
Bell (2011) for dierent regions of the United
States. Although this allowed us to estimate
excess mortality based on an appropriately
wide range of excess relative risk estimates,
it also contribu ted substantial uncertainty to
our estimates. In addition, we considered AR
estimates reported by Bobb et al. (2011) for
heat waves dened based on a T
max
, in con-
trast with estimates reported by Anderson and
Bell (2011), who defined heat waves based
on a T
avg
. In general, AR values estimated by
the two groups were similar except for the
Midwest region, where the estimates reported
by Anderson and Bell (2011) were slightly
higher. We sampled the values of excess rela-
tive risks for the three regions of our study
area from values spanning the 95% CIs of
corresponding ARs reported by Anderson and
Bell (2011) because their estimates had wider
ranges, thus representing more conservative
uncertainty estimates in the calculated excess
deaths due to heat waves.
Our study has several limitations. Our
estimates of heat wave impacts on mortality
are primarily based on temperatures in urban
areas, where the majority of eastern U.S.
residents live. However, heat stress responses
may dier between urban and rural popula-
tions (Fischer et al. 2012). Urban tempera-
tures are generally higher than temperatures
in surrounding rural areas, especially at
night, because of the urban heat island eect.
Terefore, the relative risk of heat-related mor-
tality may be much higher in urban areas than
in surrounding rural areas (Clarke 1972). If
lower impacts of heat waves on rural popula-
tions had been accounted for in our analysis,
our estimates of annual mortality due to heat
waves would have been slightly smaller.
Our estimates did not include heat-related
deaths on hot days that were not identified
as heat wave days according to our heat wave
denitions. In addition, our estimates did not
account for higher temperatures during future
heat waves, which will be hotter than those
during heat waves in 20022004, even under
the RCP4.5 scenario (Table 2). To better
account for the temperature eect, a national-
scale epidemiological study is needed to provide
region-specific, parametric exposure- response
functions between temperature and mortality.
We also did not consider eects of human
adaptation or heat mitigation measures on
future heat wave mortality. Similar to Peng
et al. (2011), we assumed that excess mortality
from future heat waves will be the same as heat
wave mortality at baseline. However, if mitiga-
tion measures to prevent heat-related mortal-
ity are adopted as heat waves become more
frequent, excess mortality would decrease, and
our projections would overestimate future
heat-related mortality in some areas. Heat
warning systems and air conditioning have
Figure3. Probability distribution of annual excess mortality attributable to heat waves in the study domain
under RCP4.5, RCP8.5, and the two scenarios combined. Data shown were derived from pooling the total
excess deaths under all possible situations considered in this study, including two temperature calibration
methods, four heat wave denitions, four population projections, 3years, and nine sets of AR values.
0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 10,000 11,000 12,000
0
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
P
r
o
b
a
b
l
i
t
y
Excess deaths/year
RCP4.5
RCP8.5
All
Figure4. Sensitivity analysis of annual heat wave mortality affected by various factors. The factor with a
higher Si value has higher inuence on estimated the excess mortality. Abbreviations: RR, the relative risk
of heat wave days to nonheat wave days; calibration, calibration ratio interpolation methods; HWM, heat
wave metrics; RCP, RCP scenarios.
Population Year RR Calibration HWM RCP
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
S
i
Wu et al.
16 VOLUME 122
|
NUMBER 1
|
January 2014 Environmental Health Perspectives
been reported to reduce heat wave impacts
(Anderson and Bell 2009; Fouillet et al. 2008;
ONeill et al. 2009). Other measures such as
planting trees can also reduce temperatures,
thus reducing heat-related mortality (Akbari
2002). Finally, our population projections
did not consider demographic changes (e.g.,
changes according to race/ethnicity, sex, and
age) that might have considerable impacts on
heat wave mortality (Hajat and Kosatky 2010;
ONeill et al. 2005).
Conclusion
Our results suggest that numbers of heat
waverelated deaths are likely to be an order
of magnitude higher in 20572059 than in
20022004. Under the fossil-fuelintensive
RCP8.5 scenario, the probability of thousands
of heat waverelated deaths per year in the
eastern United States is much higher than
under the RCP4.5 scenario. Eective mitiga-
tion and adaptation measures will be crucial to
reduce the potential for catastrophic outcomes,
particularly in the most vulnerable geographic
regions. In addition, we found that heat wave
definitions, GHG emission scenarios, and
estimates of the relative risk of mortality on
heat wave days compared with nonheat wave
days account for a large proportion of the total
variation in projected mortality estimates and
are major sources of uncertainty.
REFERENCES
Akbari H. 2002. Shade trees reduce building energy use and CO
2

emissions from power plants. Environ Pollut 116:S119S126.
Anderson BG, Bell ML. 2009. Weather-related mortality: how
heat, cold, and heat waves affect mortality in the United
States. Epidemiology 20:205213.
Anderson GB, Bell ML. 2011. Heat waves in the United States:
mortality risk during heat waves and effect modication by
heat wave characteristics in 43 U.S. communities. Environ
Health Perspect 119:210218; doi:10.1289/ehp.1002313.
Bobb JF, Dominici F, Peng RD. 2011. A Bayesian model averag-
ing approach for estimating the relative risk of mortality
associated with heat waves in 105 U.S. cities. Biometrics
67:16051616.
Chan K, Saltelli A, Tarantola S. 1997. Sensitivity analysis of model
output: variance-based methods make the difference. In:
Proceedings of the 1997 Winter Simulation Conference,
Atlanta, GA, 710December 1997 (AndradttirS, HealyKJ,
Withers DH, Nelson BL, eds). New York:Association for
Computing Machinery, 261268.
Clarke JF. 1972. Some effects of the urban structure on heat
mortality. Environ Res 5:93104
Davis RE, Knappenberger PC, Michaels PJ, Novicoff WM.
2003. Changing heat-related mortality in the United States.
Environ Health Perspect 111:17121718; doi:10.1289/
ehp.6336.
Fischer EM, Oleson KW, Lawrence DM. 2012. Contrasting
urban and rural heat stress responses to climate change.
Geophys Res Lett 39:L03705; doi:10.1029/2011GL050576.
Fouillet A, Rey G, Wagner V, Laaidi K, Empereur-Bissonnet P,
Le Tertre A, et al. 2008. Has the impact of heat waves on
mortality changed in France since the European heat
wave of summer 2003? A study of the 2006 heat wave. Int J
Epidemiol 37:309317.
Gao Y, Fu JS, Drake JB, Liu Y, Lamarque JF. 2012. Projected
changes of extreme weather events in the eastern United
States based on a high resolution climate modeling
system. Environ Res Lett 7:044025; doi:10.1088/1748-
9326/7/4/044025.
Gasparrini A, Armstrong B. 2011. The impact of heat waves on
mortality. Epidemiology 22:6873.
Gent PR, Danabasoglu G, Donner LJ, Holland MM, Hunke EC,
Jayne SR, et al. 2011. The Community Climate System
Model Version 4. J Clim 24:49734991.
Hajat S, Kosatky T. 2010. Heat-related mortality: a review and
exploration of heterogeneity. J Epidemiol Community
Health 64:753760.
Hajat S, Sheridan SC, Allen MJ, Pascal M, Laaidi K, YagoutiA,
et al. 2010. Heat-health warning systems: a compari-
son of the predictive capacity of different approaches
to identifying dangerously hot days. Am J Public Health
100:11371144.
Helton JC, Davis FJ. 2003. Latin hypercube sampling and the
propagation of uncertainty in analyses of complex systems.
Reliab Eng Syst Safe 81:2369.
Inman M. 2011. Opening the future. Nat Clim Chang 1:79.
IPCC (Intergovernmental Panel on Climate Change). 2000. IPCC
Special Report on Emissions Scenarios. 2000. Available:
[Link]
php?idp=0 [accessed 14November 2013].
IPCC (Intergovernmental Panel on Climate Change). 2007.
Summary for Policymakers. In: Climate Change 2007:
The Physical Science Basis. Contribution of Working
Group I to the Fourth Assessment Report of the
Intergovernmental Panel on Climate Change. Cambridge,
UK:Cambridge University Press. Available: [Link]
[Link]/publications_and_data/publications_ipcc_fourth_
assessment_report_wg1_report_the_physical_science_
[Link] [accessed 14November 2013].
Kolstad EW, Johansson KA. 2011. Uncertainties associated
with quantifying climate change impacts on human
health: a case study for diarrhea. Environ Health Perspect
119:299305; doi:10.1289/ehp.1002060.
Kovats RS, Hajat S. 2008. Heat stress and public health: a critical
review. Annu Rev Public Health 29:4155.
Meehl GA, Tebaldi C. 2004. More intense, more frequent, and
longer lasting heat waves in the 21st century. Science
305:994997.
Moss RH, Edmonds JA, Hibbard KA, Manning MR, Rose SK,
van Vuuren DP, etal. 2010. The next generation of scenar-
ios for climate change research and assessment. Nature
463:747756.
Nolte CG, Gilliland AB, Hogrefe C, Mickley LJ. 2008. Linking
global to regional models to assess future climate impacts
on surface ozone levels in the United States. J Geophys
Res-Atmos 113:D14307; doi:10.1029/2007JD008497.
ONeill MS, Carter R, Kish JK, Gronlund CJ, White-NewsomeJL,
Manarolla X, et al. 2009. Preventing heat-related morbid-
ity and mortality: new approaches in a changing climate.
Maturitas 64:98103.
ONeill MS, Zanobetti A, Schwartz J. 2005. Disparities by race
in heat-related mortality in four US cities: the role of air
conditioning prevalence. J Urban Health 82:191197.
Ostro BD, Roth LA, Green RS, Basu R. 2009. Estimating the mor-
tality effect of the July 2006 California heat wave. Environ
Res 109:614619.
Peng RD, Bobb JF, Tebaldi C, McDaniel L, Bell ML, Dominici F.
2011. Toward a quantitative estimate of future heat wave
mortality under global climate change. Environ Health
Perspect 119:701706; doi:10.1289/ehp.1002430.
Preston SH, Heuveline P, Guillot M. 2001. Demography.
Measuring and Modeling Population Process. Malden,
MA:Blackwell Publishers.
Robinson PJ. 2001. On the definition of a heat wave. J Appl
Meteorol 40:762775.
Saltelli A, Ratto M, Andres T, Campolongo F, Cariboni J,
Gatelli D, et al. 2008. Global Sensitivity Analysis. The
Primer. Chichester, UK:John Wiley & Sons.
Semenza JC, Rubin CH, Falter KH, Selanikio JD, Flanders WD,
Howe HL, et al. 1996. Heat-related deaths during the July
1995 heat wave in Chicago. N Engl J Med 335:8490.
Sheridan SC, Allen MJ, Lee CC, Kalkstein LS. 2012. Future heat
vulnerability in California, part II: projecting future heat-
related mortality. Clim Change 115:311326.
Shryock HS, Siegel JS, Larmon EA. 1973. The Methods and
Materials of Demography. Washington, DC:U.S. Census
Bureau.
Thomson AM, Calvin KV, Smith SJ, Kyle GP, Volke A, Patel P,
etal. 2011. RCR4.5: a pathway for stabilization of radiative
forcing by 2100. Clim Change 109:7794.
U. S. Census Bureau. 2009. Uni ted States Popul ati on
Projections by Age, Sex, Race and Hispanic Origin: July 1,
20002050. Available: [Link]
proj ecti ons/fi l es/methodstatement. pdf [accessed
14November 2013].
van Vuuren DP, Edmonds J, Kainuma M, Riahi K, Thomson A,
Hibbard K, et al. 2011. The representative concentration
pathways: an overview. Clim Change 109:531.
Zhang K, Rood RB, Michailidis G, Oswald EM, Schwartz JD,
Zanobetti A, et al. 2012. Comparing exposure metrics for
classifying dangerous heat in heat wave and health
warning systems. Environ Int 46:2329.