We examined daily mortality rates for 28 major U.S. cities over 29 years between 1964 and 1998. Raw mortality totals were culled from the National Center for Health Statistics (1998) archives. These data include documentation on each death recorded in the United States as compiled from death certificates and include the date, place, and cause of death, and demographic factors such as the age, race, and sex of the decedent. Because few deaths are directly attributable to heat stroke and there has been no consistent standard in reporting heat-related mortality over our period of record (Ellis 1972; Ellis et al. 1980; Henschel et al. 1969), we use "all causes" mortality to include both heat stroke and any possible comorbid factors (Davis et al. 2002, 2003; Gover 1938; Kalkstein and Davis 1989; Kilbourne 1997; Kunst et al. 1993; Schuman 1972; Schuman et al. 1964). Over time, changes in the age structure of a city's population can significantly influence the daily mortality rate, thereby potentially biasing temporal comparisons. Furthermore, different cities have inherently different population demographics. To account for these variations both within and between cities over time, we standardized each day's mortality count relative to a hypothetical standard city with a population of 1 million people. The demographics of this standard city were based on the age distribution of the entire U.S. populace in the year 2000. We age-adjusted our data using the direct standardization method (Anderson and Rosenberg 1998). County-level population data were acquired from U.S. Census sources for 1960, 1970, 1980, 1990, and 2000 for 10 age classes (U.S. Department of Commerce 1973, 1982, 1992, 2001), and the population of intervening years was estimated via linear interpolation.
To examine temporal changes in heat-related mortality, we divided the time series into three "decades" of approximately equal length [1964-1966 and 1973-1979 (1960s-1970s), 1980-1989, 1990-1998]. Data from 1967-1972 were not used for this study because the date of death was not systematically reported, thereby requiring the exclusion of those years.
In our analysis we used large metropolitan areas with enough daily deaths to generate robust statistical samples. We used the 1990 definitions of the counties that comprised each metropolitan statistical area (MSA) and U.S. Census data to define the population of each city. For example, according to the 1990 MSA definition, Chicago, Illinois, comprised Cook, Du Page, and McHenry Counties. Urbanization has resulted in the addition of counties to some official MSA definitions over time, so rural counties not officially designated within an MSA in the 1960s, for example, were nevertheless included in our analysis to maintain temporal sampling consistency.
Weather varies significantly on a daily basis throughout most of the United States; therefore, proper analysis of weather-mortality relationships requires the use of daily mortality records linked to a representative weather observation site. Daily weather records were retrieved for the same 28 cities from a proximate U.S. National Weather Service observation station within each metropolitan area (Table 1). Because our analysis required hourly observations, only first-order observation stations could be used, which limited our station choice to only a single station within most of the MSAs.
Using energy balance principles, Steadman (1979, 1984) demonstrated that humans respond physiologically not only to temperature but to a combination of temperature and humidity, among other factors. Biometeorologists have therefore developed a variety of indices of atmospheric conditions in an effort to determine how humans react to environmental stressors (e.g., Gagge et al. 1986; Greenberg et al. 1983; Höppe 1993, 1999; Horikoshi et al. 1997; Jendritzky and Nübler 1981; Jendritzky et al. 2000). One such measure, the apparent temperature (AT) (Steadman 1979, 1984), combines air temperature and humidity into a single variable. This index of the relative "sultriness" of the air serves as the basis for the heat index, the most commonly used summer discomfort measure in the United States, which serves as one of the bases of the heat advisories issued by the U.S. National Weather Service.
In an effort to determine the weather variables most closely linked to high mortality, we plotted daily mortality rates versus several weather variables, including morning and afternoon air temperature, dew point temperature (a measure of the amount of moisture in the air), and AT for six representative MSAs in different climatic regions. In general, the strongest relationships were found with afternoon AT, although the results were similar using morning dew point temperature, in agreement with previous research (Kalkstein and Davis 1989; Smoyer et al. 2000). Therefore, we chose afternoon AT as the independent variable in this analysis. Hourly weather data were obtained for each city [National Climatic Data Center (NCDC) 1993; National Environmental Satellite, Data, and Information Service 2000], and data were extracted for 1600 hr local standard time (LST), approximating the time of daily maximum AT.
There is often a lag between the mortality response and a given weather event (Bull and Morton 1978; Gorjanc et al. 1999; Kalkstein and Davis 1989; Rogot and Padgett 1976). After exploring several possible lags (from 0 to 3 days), we used a 1-day lag throughout this study because this consistently provides the strongest relationship between weather and mortality.
In the United States and other countries, mortality is higher in winter than in summer (Donaldson and Keatinge 1997; Eurowinter Group 1997; Langford and Bentham 1995; Laschewski and Jendritzky 2002; Lerchl 1998). This inherent seasonality could bias an analysis of heat-related mortality. For example, an early- or late-season heat wave (in April or October) could be linked with anomalously high mortality counts relative to mid-summer simply because death rates are generally higher in April and October. To remove this inherent seasonal cycle in mortality and produce a stable baseline for comparisons, we converted the data from daily mortality totals into daily mortality anomalies by subtracting from each day's mortality count the median mortality for the month in which the deaths occurred. We used the monthly median rather than the mean because the daily frequency distribution is often nonnormal, especially in months with several high mortality days. Through this technique, we enhanced the possibility of identifying relationships between daily ATs and daily mortality. Furthermore, by generating monthly mortality anomalies, we effectively standardized our dependent variable by removing the long-term trend of declining death rates, thus facilitating comparisons of heat-related mortality between decades.
Plots of daily mortality versus temperature indicate that death rates increase near the extremes of the temperature distribution in most cities with temperate climates (Alderson 1985; Bull 1973; Bull and Morton 1978; Curwen 1991; Kalkstein and Davis 1989; Khaw 1995; Kunst et al. 1993; McKee 1990; Rogot and Padgett 1976; Wyndham and Fellingham 1978). This observation led Kalkstein and Davis (1989) to propose the concept of a "threshold temperature," or the air temperature beyond which mortality increases above the baseline level, for either warm-season or cold-season mortality. Examination of mortality on the subset of days with ATs beyond this threshold enhanced our ability to link mortality to daily weather parameters. Our emphasis in this analysis was on warm-season mortality, so we only calculated thresholds at the high end of the AT distribution.
Specifically, we define a "threshold AT" as the AT at and above which mortality rates are significantly higher than the baseline rate (which is zero for deseasoned data). We aggregated the daily mortality anomalies into overlapping 2°C AT class intervals. When the mean class mortality anomaly exceeded zero based on a one-sample, one-tailed t-test ( 0.05) and remained elevated for all higher ATs, the mean AT within the lowest class with significantly above-normal mortality was defined as the threshold AT. For example, Figure 1A shows the relationship between daily all-causes mortality and 1600 hr LST AT for Chicago during the 1960s-1970s decade. Although in general there is no relationship between these variables, there is an apparent increase in mortality at high ATs. This is made more evident by computing the mean mortality within overlapping 2°C AT interval widths (Figure 1B). Because mortality increases significantly for ATs at and above 30°C, this value is calculated as the threshold AT for Chicago for the 1960s-1970s decade.
For each city, the death rates for all days above the threshold AT were summed by decade and then averaged to generate an MSA-specific decadal mean annual value. This excess mortality above the baseline approximates "heat-related" mortality rates for weather events in which the threshold AT is equaled or exceeded. Annual excess deaths were compared across decades based on bootstrapped estimates of the standard deviation for each city and decade. Bootstrapping is a nonparametric statistical procedure by which robust parameter estimates can be obtained from relatively small data samples (Efron and Gong 1983). Frequency distributions of a parameter are generated by randomly selecting observations from a sample, with replacement, multiple times. In our case, each standard deviation was estimated using 10,000 replicates, and determinations of statistically significant differences across decades were based on confidence bands defined by two standard deviations from the mean (Wilks 1995).
In previous work (Davis et al. 2002, 2003), we used a constant threshold AT for each city, defined in the 1960s and 1970s "decade" as a baseline, to determine if weather-mortality relationships had changed over time. But in this research, we employed a threshold AT that varies by decade. The resulting estimate of excess mortality should thus represent the average annual number of heat-related deaths per MSA within each decade. Ideally, one might hope to allow the threshold AT to vary from year to year; however, sample size and statistical robustness considerations, arising from the lack of warm and humid days in some years and locations, make an annual threshold calculation difficult and necessitate aggregation, which we chose to use at the decadal scale.
Temporal variations in excess deaths related to heat can arise from a number of factors, one of which is a changing climate that could influence exposure rates. To examine background climate change and related heat stress, annual trends in summer (June, July, and August) 1600 hr LST ATs were calculated for each city using least-squares linear regression. Statistical significance is based on a 0.05 level.