Applied Geography 33 (2012) 45e52
Contents lists available at ScienceDirect
Applied Geography
journal homepage: www.elsevier.com/locate/apgeog
The spatial variability of heat-related mortality in Massachusetts
David Hattis*, Yelena Ogneva-Himmelberger, Samuel Ratick
Clark University, 950 Main St., Worcester, MA 01610, USA
a b s t r a c t
Keywords:
Heat-related mortality
GIS
Climate change
Urbanization
This study assesses heat-related mortality in Massachusetts during the months of May through
September from 1990 to 2008. Daily maximum apparent temperature was interpolated across space via
kriging, and aggregated to 29 municipality groups (MGs), a spatial unit composed of municipalities that
was designed to have minimal variation in population. Death certificate data were analyzed to determine
the spatial distribution of excess mortality on days that exceeded the 85th, 90th, and 95th percentiles of
apparent temperature. We find that the average statewide mortality anomalies were 5.11, 6.26, and 7.26
deaths on days exceeding the 85th, 90th, and 95th percentiles of apparent temperature respectively. A
linear stepwise regression showed that percent AfricaneAmerican population and percent elderly
population (those above the age of 65) were positively associated with an MG’s mortality anomaly on
days exceeding the 85th percentile of apparent temperature (p < 0.05). In spite of the urban heat island
effect, our measure of urbanization was not associated with higher rates of heat-related mortality.
Ó 2011 Elsevier Ltd. All rights reserved.
Introduction
Heat events cause significant increases in mortality. The 2003
European heat wave is believed to have caused 14,947 excess
deaths in France (Poumadére, Mays, Le Mer, & Blong, 2005). In the
United States, heat exposure is considered to be the most deadly
weather event (Changon et al., 1996). Between 1979 and 1999, 8015
deaths in the United States were classified as being caused by heat
exposure (CDC, 2001). This amount, in all likelihood, is significantly
understated because heat exposure often exacerbates pre-existing
health conditions, and is thus often not classified as a direct
cause of death. During heat events, many causes of death show
significant increases, including cardiovascular disease, respiratory
diseases, diabetes, stroke, accidents, suicide, and homicide (Ellis,
1972; McGeehin & Mirabelli, 2001).
Heat vulnerability varies significantly within populations, often
along demographic and socio-economic lines (Reid et al., 2009).
The elderly and infant age groups are at the highest risk for heatrelated mortality (Basu & Samet, 2002), due to their physiological
difficulties regulating heat, and their restricted mobility, which
decreases their ability to access fluids when needed (CDC, 1993).
Some evidence suggests that vulnerability also varies by ethnicity.
A recent case control study found that AfricaneAmericans had the
* Corresponding author. Clark University, 63 Hancock St. Somerville, MA 02144,
USA. Tel.: þ1 617 335 4540.
E-mail addresses: [email protected] (D. Hattis), [email protected]
(Y. Ogneva-Himmelberger), [email protected] (S. Ratick).
0143-6228/$ e see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apgeog.2011.07.008
highest risk of heat-related mortality, and that Caucasians had
a higher risk than Hispanics (Basu & Ostro, 2008). Income also may
be an important factor. Jones et al. (1982) found that those in lowincome census tracts were six times more likely to experience heat
stroke during a heat event in St. Louis and Kansas City than those in
upper-income census tracts. Some evidence suggests that education also might influence heat vulnerability. In a study of 50 United
States cities, those without a degree above a high school diploma
had a higher heat-related death rate than those with more education (Medina-Ramon, Zanobetti, Cavanagh, & Schwartz, 2006).
Finally, social isolation has been identified as a factor that increases
heat vulnerability. Living alone was found to be a strong risk factor
for heat-related mortality in two case control studies of Chicago
heat waves (Naughton et al., 2002; Semenza et al., 1996).
Much of the research on heat-related health impacts has focused
on urban areas (Baccini et al., 2008; Davis, Knappenberger, Michaels,
& Novicoff, 2003; Kalkstein & Davis, 1989; Smoyer, Rainham, &
Hewko, 2000), where urban heat island effects tend to increase
temperatures relative to surrounding locations. Some studies have
found that heat waves generate greater mortality responses in cities
than surrounding areas (Jones et al., 1982; Tan et al., 2010), and some
future projections of heat-related health effects have focused
primarily on urban areas. Tol (2002), for example, assumes that heatrelated mortality due to cardiovascular disease will occur mainly in
urban areas. However, a study of Ohio found that suburban and rural
counties experienced a higher proportional increase in heat-related
mortality than urban counties, although the difference was not
statistically significant (Sheridan & Dolney, 2003).
46
D. Hattis et al. / Applied Geography 33 (2012) 45e52
This study seeks to advance this body of knowledge by assessing
the relationship between heat and mortality in Massachusetts, in
both its urban and rural areas. In doing so, this study attempts to
answer the following research questions: (1) How does heatrelated mortality vary across space in Massachusetts? (2) Can
spatial patterns of heat-related mortality be explained by sociodemographic variables? (3) Do urban areas experience a higher
level of heat-related mortality than non-urban areas?
Massachusetts was chosen as the study area because it contains
a relatively diverse mix of urban, rural, and suburban areas, and it is
located in the northeastern United States, a region where heatrelated mortality may be especially problematic (McGeehin &
Mirabelli, 2001).
Data and methodology
This study employed three types of data: meteorological data
from the National Climate Data Center (NCDC), mortality data from
the Massachusetts Department of Public Health’s Registry of Vital
Records and Statistics, and socio-demographic data from the U.S.
Census Bureau.
Apparent temperature estimation
Apparent temperature, a variable that combines temperature and
humidity, was used to measure heat exposure in this study (Steadman,
1984). Sub-hourly temperature and relative humidity data were
obtained from NCDC’s Integrated Surface Hourly Database (NCDC,
2010) for all 2908 days in the study period (May 1st e September
30th for years 1990e2008). Fig. 1 shows the 102 temperature stations
used in the analysis.
Temperature and relative humidity were used to calculate water
vapor pressure, using the following equation:
VP ¼
17:0502T
RH
6:1121e240:97 þ T
100
(1)
where VP is water vapor pressure in hectopascals, T is temperature
in C, and RH is relative humidity (a percentage) (Buck, 1981). Using
the calculated water vapor pressure, apparent temperature was
derived using:
AT ¼ 1:3 þ 0:92T þ 2:2VP
(2)
C,
C,
where AT is apparent temperature in
T is temperature in
and
VP is water vapor pressure in kilopascals (Steadman, 1984).
Maximum daily apparent temperature, a commonly used heat
exposure variable (Baccini et al., 2008; Smoyer et al., 2000), was
calculated for each available weather station on each day in the study
period. To ensure completeness, the sub-hourly data were first
aggregated to the hour by deriving the mean apparent temperature
within each hourly interval, and all stations that did not provide at
least 18 hourly observations on a particular day were excluded. This
exclusion was done to help assure that the maximum daily apparent
temperature was reasonably representative of the conditions that
occurred on that day. After removing days with an insufficient number
of hourly readings, the highest hourly apparent temperature on each
day was recorded as that day’s maximum apparent temperature.
This study used a geo-statistical approach to apparent temperature estimation. Based on the discrete point data from temperature
stations, daily maximum apparent temperature was estimated
across the entire state of Massachusetts for every day in the study
period via kriging, a method of spatial interpolation that uses the
spatial covariance structure of the datadhow similar data points are
at various distances from one anotherdto derive linear weights that
produce a continuous surface estimate in raster format (Isaaks &
Srivastava, 1989). Pixel size in the output raster surfaces was set at
330 m, which covered Massachusetts with 250,000þ pixels.
Fig. 1. The locations of the 102 meteorological stations that provided temperature and relative humidity data.
D. Hattis et al. / Applied Geography 33 (2012) 45e52
Fig. 2. The spatial distribution of the mean mortality anomaly per million population on hot days.
Fig. 3. The spatial distribution of the mean mortality anomaly per million population on very hot days.
47
48
D. Hattis et al. / Applied Geography 33 (2012) 45e52
Fig. 4. The spatial distribution of the mean mortality anomaly per million population on extremely hot days.
Kriging models the spatial covariance structure of the data using
a semivariogram (Isaaks & Srivastava,1989). Due to the large number
of days in the study period, the kriging procedure had to be automated, and to facilitate this automation, a uniform semivariogram
model was selected for each daily surface. In order to choose the
semivariogram model that would produce the most accurate estimates, a cross-validation procedure was performed to compare the
accuracy of five semivariogram models available in the ArcGIS 9.3.1
Spatial Analyst tool: Spherical, Gaussian, Exponential, Circular and
Linear. The cross-validation procedure used all data points (in this
case, all weather stations) except one to create a continuous surface
estimate. The value of the omitted data point was then compared to
the value predicted at the corresponding location by the surface
estimate to derive the error. This process was repeated for every
weather station with an apparent temperature reading on a given
day, and for a random sample of 100 days from the study period. The
exponential semivariogram produced the lowest average standard
error and was consequently chosen for use in the kriging procedure.
Unit of analysis
There is a tradeoff between geographical specificity and sample
size when selecting a unit of analysis: larger units have greater
population sample sizes and thus more statistical power per unit; on
the other hand, increasing the size of the unit of analysis reduces the
geographical specificity of the analysis, thus lessening the ability to
discern variation across space. For this study, more statistical power
per unit was desired than municipalities would provide, and more
geographical specificity was desired than counties would provide.
To resolve this tradeoff, Massachusetts was divided into 29 municipality groups (MGs), a unit created by combining municipalities in
order to maximize the geographical specificity of the analysis, while
ensuring that each unit had a 2000 Census population around
200,000, a figure that was expected to be large enough to provide
meaningful results based on the experience of a similar study
(Sheridan & Dolney, 2003). The 29 MGs, which can be seen in
Figs. 2e5, all had 2000 Census populations around 200,000, with the
exception of the Boston MG, which had a population of 589,141.
The statewide daily apparent temperature surfaces were aggregated to the MGs by deriving the mean pixel value within each MG
polygon. This task was done with the Zonal Statistics tool in ArcGIS
9.3.1.
Daily apparent temperature categories
Based on the apparent temperature estimates, each MG received
a daily classification of hot, very hot, extremely hot, or not hot. Hot
days were operationally defined as those that exceeded the 85th
percentile apparent temperature in a given MG; very hot and
extremely hot days were likewise defined as days that exceeded the
90th and 95th percentiles, and all other days were classified as not
hot. These thresholds were chosen because mortality often
increases significantly beginning at the 85th percentile apparent
temperature (Gaffen & Ross, 1998). This method determines the
daily classification based on a comparison between the maximum
apparent temperature experienced on a given day in a given MG, to
the maximum apparent temperatures experienced in that MG
throughout the entire study period. As a result, the minimum
apparent temperature at which a day was classified as hot, very hot,
or extremely hot varied between MGs. This method was used
instead of applying uniform apparent temperature thresholds
across the entire state because Massachusetts experiences some
variation in climate due to its coastal position. The mean maximum
daily apparent temperature during the study period ranged from
22.77 C in an MG including Cape Cod, to 24.89 C in an MG in
southwestern Massachusetts. With this variation in climate, some
parts of Massachusetts might be slightly better acclimated to heat
than others, and it thus was determined that what is considered
hot, very hot, and extremely hot was relative and could vary
between MGs.
D. Hattis et al. / Applied Geography 33 (2012) 45e52
49
Fig. 5. Model 2 residuals.
Calculating mortality anomaly
Death certificate data from the Massachusetts Department of
Public Health’s Registry of Vital Records and Statistics was used to
obtain the number of deaths in each MG on each day in the study
period (Massachusetts Department of Public Health, 2010). Allcause mortality was used because a wide variety of causes of
death are elevated during heat events (Ellis, 1972). A daily mortality
anomaly was then calculated using the following equation
(Sheridan & Dolney, 2003):
MA ¼ D ED
(3)
where MA is the mortality anomaly, D is the number of deaths on
a given day in a given MG, and ED is the number of deaths expected
on a given day in a given MG, which is equal to the average number
of daily deaths in the MG that year during the months of May
through September.
The average mortality anomaly per million population was then
calculated separately for hot, very hot and extremely hot days for
each MG. The Local Indicators of Spatial Association (LISA) tool in
Geoda was used to test for spatial clustering of mortality anomalies
on these days (Anselin, 1995).
There is frequently a lag between a heat event and its associated
mortality response. People may initially get sick on a hot day, and
then die as a result on a later day. Studies have found the strongest
associations between a heat event and the mortality response
ranging from the same day to 3 days following the heat event (Basu
& Samet, 2002). Results presented in this paper are for a 0-day lag,
because a 0-day lag was found to have the strongest correlation
(measured by a Pearson’s correlation coefficient) between apparent
temperature and the mortality anomaly.
Testing statistical significance of mortality anomalies
A randomization procedure was used to test whether the
mortality anomaly was statistically significantly elevated on days
classified as hot, very hot, and extremely hot (Sheridan & Dolney,
2003). This involved randomly sampling the series of mortality
anomalies in each MG 10,000 times (using a uniform distribution so
that each day was equally likely to be chosen), with each iteration
selecting a number of samples identical to the number of days that
were categorized as hot, very hot, or extremely hot. The percentage
of random samples with mean mortality anomalies exceeding the
observed mortality anomaly on the classified days determined the
statistical significance (Sheridan & Dolney, 2003). For example, if
the Boston MG had 250 days classified as very hot, and the mean
mortality anomaly on those days was þ1.5 deaths, then each iteration would first randomly select 250 mortality anomalies in the
Boston MG. The average mortality anomaly from the random
sample would then be compared to the observed mortality
anomaly on the days classified as very hot (in this case þ1.5 deaths).
If 4 percent of the mortality anomaly samples had means greater
than 1.5 deaths, then the derived p value would be 0.04.
Socio-demographic determinants of heat-related mortality
Socio-demographic data were obtained from the 2000 U.S.
Census (U.S. Census, 2010a), which was conducted approximately
midway through the study period. The following sociodemographic variables were calculated for each MG: (1) mean
household income, defined as the average income per household;
(2) percent AfricaneAmerican, defined as the percent of population
who classified themselves as AfricaneAmerican; (3) percent elderly,
defined as the percent of population above the age of 65; (4) percent
infant, defined as the percent of population below the age of 1; (5)
percent single person households, defined as the percent of occupied
households with only one person; and (6) percent without high
school diploma, defined as the percent of population above 25
without a high school diploma. These variables were chosen
because previous studies have linked the underlying phenomena
with heat-related mortality (see Section 1). Correlation analysis and
linear regression were performed to determine the relative
importance of these socio-demographic variables in explaining the
variance in the mortality anomaly between the MGs.
50
D. Hattis et al. / Applied Geography 33 (2012) 45e52
Urbanization and heat-related mortality
Table 1
Heat-related mortality summary statistics.
There are several ways to operationalize urbanization. One
option is to use the U.S. Census Bureau’s urban/non-urban classifications. Individuals are classified by the U.S. Census as living in an
urban area if their residence lies within an urban cluster (a densely
settled area with a population between 2500 and 50,000 people) or
an urbanized area (a densely settled area with a population greater
than 50,000 people) (U.S. Census, 2010b). One downside of this
method is that it is a binary scale, which may be problematic because
it does not capture any variation between areas classified as urban or
non-urban. This may be problematic for our purposes because the
urban heat island effect may be more severe in a place like Boston
than in Pittsfield, a relatively small city located in western Massachusetts that is also classified as urban by the U.S. Census.
An alternative to this binary approach is to use a continuous
variable, like population density, as a proxy for urbanization.
However, using population density may not always represent the
urbanization of an area. Consider an area that has within it a small
densely populated city, as well as a large mountain range, where
few people live. Deriving the population density of this entire area
would underestimate the urbanization of the population because it
is “averaged out”, even though the vast majority of this area’s
inhabitants live in the densely populated city.
To resolve this issue, a weighted average procedure was used to
create a modified population density value for each MG. This was
done in three steps: (1) The population density of each census block
group within each MG was obtained; (2) these population densities
were multiplied by the percent of population they represented of
the entire MG; and (3) these values were summed. The equation for
this procedure is shown below:
Modified Population DensityMG ¼
n X
PopBG
BG
PopMG
PopBG
AreaBG
(4)
Where BG is a given block group within an MG, n is the total
number of block groups within the given MG, Pop is population, and
Area is the land area in square miles.
Correlation analysis and linear regression were performed to
determine whether the degree of urbanization (measured by
modified population density) was associated with increased heatrelated mortality.
Results
Total heat-related mortality in massachusetts
A summary of the characteristics of heat-related mortality in
Massachusetts is presented in Table 1. As expected, the mean
mortality anomaly increased with the temperature threshold: the
mean mortality anomaly statewide was 5.11 deaths on hot days,
Statewide mean mortality anomaly (deaths)
Percent Increase in Deaths
Sample size (number of days in study period
meeting each classification)
Number of MGs with significantly elevated
mortality anomalies (p < 0.05)
Hot
Days
Very Hot
Days
Extremely
Hot
Days
5.11
3.7%
436
6.26
4.6%
291
7.26
5.3%
146
15
13
9
6.26 deaths on very hot days, and 7.26 deaths on extremely hot
days. As the temperature classification became more stringent, the
sample size of days meeting each category decreased, and as
a consequence, a lower number MGs exhibited significant increases
in their mortality anomalies.
The spatial distribution of heat-related mortality
Figs. 2e4 show the spatial distribution of the average mortality
anomaly per million population on hot, very hot days, and
extremely hot days. The MGs are labeled by the largest municipality
they contain in terms of population, according to the 2000 U.S.
Census. The spatial distribution of heat-related mortality was
similar on hot, very hot, and extremely hot days. The greatest heatrelated mortality was generally experienced in southwest/southcentral Massachusetts (in the Southbridge and Springfield MGs),
southeast Massachusetts (in the Brockton and New Bedford MGs)
and in some municipalities in and around Boston (for example the
Boston and Framingham MGs). LISA tests found significant clustering of high mortality anomaly values in southwest/south-central
Massachusetts on hot and very hot days. A LISA test on extremely
hot days showed significant negative spatial autocorrelation
centering in the Framingham MG.
Explaining the variation in heat-related mortality across space
Correlation analysis and linear regression were performed to
determine the extent to which socio-demographic variables and
urbanization could explain the variation in heat-related mortality
between the MGs. The results reported here use the average
mortality anomaly on hot days per million population as the
dependent variable (refferred to hereafter as the mortality anomaly
on hot days), because this classification contained the largest
sample size of days, and is thus expected to show the most reliable
variations in heat-related mortality.
Correlation analysis
Table 2 shows a bivariate Pearson’s correlation matrix, which
contains the dependent variable (mortality anomaly on hot days),
Table 2
Bivariate Pearson’s correlation matrix for all variables used in the regression models.
Mortality Anomaly on Hot Days
Mean Household Income
% AfricaneAmerican
% Elderly
% Infant
% Single Person Households
% Without High School Diploma
Modified Population Density
a
b
Mortality Anomaly
on Hot Days
Mean Household
Income
1
0.213
0.337b
0.387a
0.305
0.273
0.178
0.066
1
0.307
0.179
0.216
0.376a
0.773a
0.189
Denotes Statistical Significance at p < 0.05.
Denotes Statistical Significance at p < 0.1.
% AfricaneAmerican
1
0.224
0.080
0.529a
0.317b
0.709a
%
Elderly
%
Infant
% Single Person
Households
% Without High
School Diploma
1
0.565a
0.355b
0.024
0.176
1
0.479a
0.037
0.071
1
0.243
0.747a
1
0.308
Modified
Population
Density
1
D. Hattis et al. / Applied Geography 33 (2012) 45e52
Table 3
Regression models using the mean mortality anomaly per million population on hot
days as the dependent variable. Standardized regression coefficients are reported,
along with their associated p values in parenthesis.
Model 1
Mean Household Income
% AfricaneAmerican
% Elderly
% Infant
% Single Person Households
% Without High School Diploma
Modified Population Density
R2
Adjusted R2
Moran’s I of Residuals
Number of Observations
a
0.339
(0.299)
0.783
(0.005a)
0.413
(0.124)
0.274
(0.235)
0.042
(0.924)
0.362
(0.241)
0.453
(0.290)
0.457
0.276
0.121
29
Model 2
Model 3
0.446
(0.011a)
0.487
(0.006a)
0.679
(0.005a)
0.481
(0.006a)
0.331
(0.147)
0.338
0.287
0.006
29
0.393
0.320
0.091
29
Denotes Statistical Significance at p < 0.05.
along with all of the socio-demographic variables described in
Section 2.6, and the modified population density variable described
in Section 2.7. Percent AfricaneAmerican and percent elderly were
both positively correlated with the mortality anomaly on hot days,
and were significant at the p < 0.1 and p < 0.05 levels respectively.
Additionally, many of the independent variables exhibited strong
correlations with one another. For example, mean household income
was strongly negatively correlated with percent without high school
diploma (r ¼ 0.773), and modified population density displayed
a strong positive correlation with percent single person households
(r ¼ 0.747).
Linear regression
Several diagnostic tests were used to assess whether our data
conformed with the assumptions of linear regression: A JarqueBera procedure was undertaken to test for non-normality (Jarque
& Bera, 1987), a BreuschePagan test was undertaken to test for
heteroskedasticity (Breusch & Pagan, 1979), and Moran’s I tests
were used to test for spatial autocorrelation in the regression
variables and residuals (Moran, 1950). None of these tests showed
statistically significant deviations from the regression assumptions,
so no adjustments were made.
Table 3 shows the results of three linear regression models.
Model 1 included all seven independent variables (the sociodemographic variables described in Section 2.6, and the modified
population density variable described in Section 2.7). Only percent
AfricaneAmerican was statistically significant (p ¼ 0.005). For
Model 2, a stepwise regression procedure was used with all the
independent variables from Model 1. Percent AfricaneAmerican
and percent elderly were found to be positively associated with the
mortality anomaly on hot days (p values of 0.011 and 0.006
respectively). Model 3 used the two variables included in the
stepwise procedure (percent AfricaneAmerican and percent
elderly), along with modified population density as independent
variables. Percent AfricaneAmerican and percent elderly were still
positively associated with the mortality anomaly on hot days (p
values of 0.005 and 0.006 respectively), but here the modified
population density coefficient was negative, although still statistically insignificant (p ¼ 0.147). As shown in Table 2, modified
population density showed a slight, statistically insignificant
bivariate correlation of 0.066 with the mortality anomaly on hot
days (p ¼ 0.734).
51
Linear regression residuals
As is apparent in Table 2, the three regression models had
Moran’s I values that were either negative (although none of these
values were statistically significant) or close to 0. Fig. 5 shows the
residuals for Model 2 (with all coefficients statistically significant).
We could not identify any missing explantory variables that could
explain the spatial pattern of the residuals.
Discussion and conclusions
This study is one of only several to our knowledge that has
analyzed the spatial distribution of heat-related mortality in relation
to both urbanization and relevant socio-demographic variables.
Mortality statewide was elevated on days that exceeded the
apparent temperature thresholds. The reported figures, however,
may be overstated. Previous studies indicate that high apparent
temperatures are associated with pollution episodes, which also
impact mortality (Kovats & Hajat, 2008). For example, a study of
Mexico City found that controlling for air pollution and respiratory
epidemics decreased the impact of apparent temperature on
mortality by 50% (O’Neil et al., 2005). As a result, failure to account
for pollution may produce results that overestimate the impact of
apparent temperature on mortality. However, it should be noted
that higher temperatures increase the concentrations of many air
pollutants, including Ozone and Volatile Organic Compounds
(Bernard, Samet, Grambsch, Ebi, & Romieu, 2001).
Previous investigations have found significantly elevated heatrelated mortality in urban areas relative to surrounding rural/
suburban areas (Jones et al., 1982; Tan et al., 2010). However,
neither of these studies reported results after controlling for relevant socio-economic variables. This study found a very slight
positive (but statistically insignificant) correlation between an
MG’s modified population density and its mortality anomaly on hot
days. However, after controlling for the two statistically significant
variables from the stepwise regression (percent AfricaneAmerican,
and percent elderly), a negative relationship was present between
modified population density and the mortality anomaly on hot days
(this also was statistically insignificant). These results suggest that,
at least in Massachusetts, an area’s demographics may be more
important to its heat-related mortality than its level of urbanization, at least as captured by the specific variables used in this study.
The comparative vulnerability of urban and rural areas appears
to vary greatly by location and heat event. Up until this point, little
research has been done on heat vulnerability in rural areas. It is
conceivable that the vulnerability of rural areas may be influenced
by drivers that have little importance in urban areas. For example,
heat vulnerability in rural areas may be influenced by the prevalence of outside laborers (Sheridan & Dolney, 2003). Further
research is needed to determine the factors that affect heatrelated mortality in rural areas, especially in light of expected
temperature increases.
References
Anselin, L. (1995). Local Indicators of Spatial AssociationsdLISA. Geographical
Analysis, 27(2), 93e115.
Baccini, M., Biggeri, A., Accetta, G., Kosatsky, T., Katsouyanni, K., Analitis, A., et al. (2008).
Heat effects on mortality in 15 European cities. Epidemiology, 19(5), 711e719.
Basu, R., & Samet, J. M. (2002). Relation between elevated ambient temperature and
mortality: a review of the epidemiologic evidence. Epidemiologic Reviews, 24(2),
190e202.
Basu, R., & Ostro, B. D. (2008). A multicounty analysis identifying the populations
vulnerable to mortality associated with high ambient temperature in California.
American Journal of Epidemiology, 168(6), 632e637.
Bernard, S. M., Samet, J. M., Grambsch, A., Ebi, K. L., & Romieu, I. (2001). The potential
impacts of climate variability and change on air pollution-related health effects in
the United States. Environmental Health Perspectives, 109, 199e209.
52
D. Hattis et al. / Applied Geography 33 (2012) 45e52
Breusch, T. S., & Pagan, A. R. (1979). A simple test for heteroscedasticity and random
coefficient variation. Econometrica, 47(5), 1287e1294.
Buck, A. L. (1981). New equations for computing vapor pressure and enhancement
factor. Journal of Applied Meteorology, 20(12), 1527e1532.
Centers for Disease Control and Prevention (CDC). (1993). Heat-related deathse
United States, 1993. Morbidity and Mortality Weekly Recap, 42, 558e560.
Centers for Disease Control and Prevention (CDC). (2001). Heat-related deathseLos
Angeles County, California, 1999e2000, and United States, 1979e1998. Journal
of the American Medical Association, 286(8), 911e912.
Changnon, S. A., Kunkel, K. E., & Reinke, B. C. (1996). Impacts and responses to the
1995 heat wave: a call to action. Bulletin of the American Meteorological Society,
77(7), 1497e1506.
Davis, R. E., Knappenberger, P. C., Michaels, P. J., & Novicoff, W. M. (2003). Changing
heat-related mortality in the United States. Environmental Health Perspectives,
111(14), 1712e1718.
Ellis, F. P. (1972). Mortality from heat illness and heat-aggravated illness in the
United States. Environmental Research, 5(1), 1e58.
Gaffen, D. J., & Ross, R. J. (1998). Increased summertime heat stress in the US. Nature,
396, 529e530.
Isaaks, E. H., & Srivastava, R. M. (1989). Applied geostatistics. New York: Oxford
University Press.
Jarque, C. M., & Bera, A. K. (1987). A test for normality of observations and
regression residuals. International Statistical Review, 55, 163e172.
Jones, T. S., Liang, A. P., Kilbourne, E. M., Griffin, M. R., Patriarca, P. A.,
Wassilak, S. G. F., et al. (1982). Morbidity and mortality associated with the July
1980 heat wave in St. Louis and Kansas City, Mo. Journal of the American Medical
Association, 247(24), 3327e3331.
Kalkstein, L. S., & Davis, R. E. (1989). Weather and human mortality: an evaluation of
demographic and interregional responses in the United States. Annals of the
Association of American Geographers, 79(1), 44e64.
Kovats, R. S., & Hajat, S. (2008). Heat stress and public health: a critical review.
Annual Review of Public Health, 29(1), 41e55.
Massachusetts Department of Public Health. (2010). Registry of vital records and
statistics. http://www.mass.gov/dph/rvrs Accessed 08.02.10.
Medina-Ramon, M., Zanobetti, A., Cavanagh, D. P., & Schwartz, J. (2006). Extreme
temperatures and mortality: assessing effect modification by personal characteristics and specific cause of death in a multi-city case-only analysis. Environmental Health Perspectives, 114, 1331e1336.
Moran, P. A. P. (1950). Notes on continuous stochastic phenomena. Biometrica, 37,
17e23.
McGeehin, M. A., & Mirabelli, M. (2001). The potential impacts of climate variability
and change on temperature-related morbidity and mortality in the United
States. Environmental Health Perspectives, 109, 185e189.
Naughton, M. P., Henderson, A., Mirabelli, M. C., Kaiser, R., Wilhelm, J. L.,
Kieszak, S. M., et al. (2002). Heat-related mortality during a 1999 heat wave in
Chicago. American Journal of Preventive Medicine, 22(4), 221e227.
National Climate Data Center (NCDC). (2010). 3505/CDO: Integrated surface
data, hourly (DSI0e3505). http://www.ncdc.noaa.gov/oa/mpp/digitalfiles.
html Accessed 04.03.10.
O’Neill, M., Hajat, S., Zanobetti, A., Ramirez-Aguilar, M., & Schwartz, J. (2005). Impact
of control for air pollution and respiratory epidemics on the estimated associations of temperature and daily mortality. International Journal of Biometeorology, 50(2), 121e129.
Poumadére, M., Mays, C., Le Mer, S., & Blong, R. (2005). The 2003 heat wave in
France: dangerous climate change here and now. Risk Analysis, 25(6),
1483e1494.
Reid, C. E., O’Neill, M. S., Gronlund, C. J., Brines, S. J., Brown, D. G., Diez-Roux, A. V.,
et al. (2009). Mapping community determinants of heat vulnerability. Environmental Health Perspectives, 117(11), 1730e1736.
Semenza, J. C., Rubin, C. H., Falter, K. H., Selanikio, J. D., Flanders, W. D., Howe, H. L.,
et al. (1996). Heat-related deaths during the July 1995 heat wave in Chicago.
New England Journal of Medicine, 335(2), 84e90.
Sheridan, S., & Dolney, T. J. (2003). Heat, mortality, and level of urbanization:
measuring vulnerability across Ohio, USA. Climate Research, 24(3),
255e265.
Smoyer, K. E., Rainham, D. G. C., & Hewko, J. N. (2000). Heat-stress-related mortality
in five cities in Southern Ontario: 1980e1996. International Journal of Biometeorology, 44(4), 190e197.
Steadman, R. G. (1984). A universal scale of apparent temperature. Journal of Climate
and Applied Meterology, 23, 1674e1687.
Tan, J., Zheng, Y., Tang, X., Guo, C., Li, L., Song, G., et al. (2010). The urban heat island
and its impact on heat waves and human health in Shanghai. International
Journal of Biometeorology, 54(1), 75e84.
Tol, R. S. J. (2002). Estimates of the damage costs of climate change. Part 1:
benchmark estimates. Environmental and Resource Economics, 21(1),
47e73.
U.S. Census. (2010a). Your gateway to census 2000. http://www.census.gov/main/
www/cen2000.html Accessed 13.08.10.
U.S. Census. (2010b). Glossary. http://factfinder.census.gov/home/en/epss/glossary_
u.html Accessed 27.12.10.