A new approach to evaluate the impact of climate on human mortality.

The objective of this study is to introduce a new procedure to determine the impact of climate on human mortality with the use of a synoptic climatological approach. The holistic nature of synoptic categories allows for the simultaneous evaluation of numerous weather elements as they realistically appear within air masses. In addition, this approach allows for a better distinction between pollution-induced mortality and weather-induced mortality. A synoptic categorization was performed for St. Louis, Missouri, and each category was evaluated in terms of its mean daily mortality. Of 10 summer categories found in St. Louis, one possessed the highest mean mortality by far, and 8 of the top 10 mortality days in St. Louis occurred when this category was present. Further analysis determined that long, consecutive day periods of this hot, oppressive category are associated with a continuing rise in mortality. It was determined that the procedure described here has the potential to be used in a weather/mortality watch-warning system. Finally, it appears that day-to-day mortality fluctuations are much more sensitive to weather than to pollution concentrations, as the oppressive category associated with the greatest mortality possessed levels of six major pollutants that were not noteworthy.


Introduction
The impact ofvariable climate on human health and well-being has been the subject ofnumerous studies, with the majority ofthe work being performed by medical specialists and a small minority by climatologists. With the threat ofan anthropogenic-induced global warming, the interest in climate/human health studies has increased, and at least three comprehensive reports summarizing most of this research have appeared in recent years (1)(2)(3).
The majority of the climate/human health evaluations have concentrated on mortality, and virtually all of the studies correlate a number of climate variables with daily or weekly mortality statistics. For example, medical researchers have noted that mortality attributed to weather seems to vary considerably with age, sex, and race, although there is disagreement in defming the most susceptible population group (4)(5)(6). Some researchers have found that many causes other than heat stroke increase during extreme weather (7Z8). Evaluations have been performed for winter weather impacts as well, and mortality increases have been noted during extreme cold waves (6)(7)(8)(9)(10)(11). More recent studies have concentrated on interregional impacts of weather on human health, and the potential adverse effects ofglobal climate change (12,13). In addition, the sophistication of climatic modeling within these has increased, and "threshold temperatures," which represent the temperature beyond which mortality significantly increases, have been identified for numerous cities (14).
All of these studies suffer from at least two shortcomings. First, the use ofindividual weather elements to evaluate the impact of climate on human health may provide misleading results. Climate affects human health as a holistic unit composed si-*Center for Climatic Research, Department of Geography, University of Delaware, Newark, DE 19716. multaneously of numerous climatic elements that react in concert. Thus, the isolation ofindividual elements can represent an unrealistic situation that does not duplicate climatic reality. Second, the impact of pollution concentration is often ignored in a climate/mortality evaluation. It is possible that the mechanisms responsible for increasing mortality are truly pollution-oriented, but these are imbedded (and ignored) within the climate evaluation. It is not feasible to evaluate individual pollutants (e.g., SO2 or NO2 concentrations) along with individual climate elements (e.g., temperature or dew point) as independent variables within a health study because the climate and pollution variables are somewhat collinear.
The objective ofthis study is to introduce a new procedure to determine the impact ofclimate on human mortality that does not suffer from the problems outlined above. This involves the use of a synoptic climatological approach, which combines weather elements into groups or categories that are representative oftrue meteorological situations at a moment in time (15). Synoptic climatological approaches characterize similarities in general circulation patterns and active meteorological elements within a holistic framework, facilitating analyses ofclimate's impact on a variety of environmental variables such as human mortality. Thus, the impact ofthe totality ofweather on the environmental variable can be determined. Additionally, synoptic categories possess differential abilities to retain high concentrations of pollutants. Thus, certain synoptic categories are usually associated with low pollutant concentrations, while others are almost always highly polluted. Kalkstein and Corrigan (16) determined that the 9 most polluted days in Wilmington, Delaware, during the winters of 1974-1978 were associated with two synoptic categories that occurred only 15 % of the time during the period. It is clear that this approach can distinguish between offensive weather situations and offensive pollution situations.

Methodology
The National Center for Health Statistics (NCHS) provided us with a very detailed mortality database that contains records of every person who died in the United States from 1964 to the present [refer to NCHS (17) for a detailed description of these records]. We selected St. Louis, Missouri, for our test location, as previous research suggests that the area is particularly sensitive to heat-related deaths (14). The data contain cause ofdeath, place of death, age, date of death, and race, and these were extracted for the St. Louis Standard Metropolitan Statistical Area for 11 years: 1964-66, 1972-78, and 1980 (although the data are complete for the entire record, during intervening years the date of death was omitted). The number of deaths for each day was tabulated and divided into categories based on age (all ages and elderly, classified as greater than 65 years old) and race (white and nonwhite).
All mortality data were adjusted to account for changes in the age, racial distribution, and total population ofthe St. Louis area during the period of record. A direct standardization procedure was used, and a standard population was constructed for each of the evaluated mortality categories [refer to Mausner and Bahn (18) and Lilienfield and Lilienfield (19) for a discussion of this standardization procedure].
An automated air-mass-based synoptic climatological index was developed for St. Louis to categorize each day into its particular synoptic category. The synoptic procedure is designed to classify days that are considered to be meteorologically homogeneous. This is accomplished by defining each day in terms of seven readily available meteorological elements, which include air temperature, dew point temperature, visibility, total cloud cover, sea-level air pressure, wind speed, and wind direction. There elements are measured four times daily (0100, 0700, 1300, 1900 hr local standard time), and the developed 28 variables represent the basis for categorization.
The developed temporal synoptic index (TSI) uses principal components analysis (PCA), a factor analysis technique that rewrites the original data matrix into a new set of components that are linearly independent and ordered by the amount of variance they explain [for a detailed explanation of this procedure, refer to Kalkstein et al. (20)]. Component loadings are calculated, which express the correlation between the original 28 variables and the newly formed components. Each day is then expressed by its particular set of "component scores," which are weighted summed values whose magnitudes are dependent on the weather observation for each day and the principal component loading. Thus, days with similar meteorological conditions will tend to exhibit proximate component scores.
A clustering procedure was then used to group those days with similar component scores into meteorologically homogeneous groups. There are numerous clustering methodologies available, but the one deemed most efficient in the development ofa synoptic climatological index is the average linkage method (20). Once the groups have been determined, average values are calculated for the 28 meteorological variables for all days within each particular group. Weather map classification is also possible by reviewing maps for those days within a group and describing general similarities.
The mean daily mortality for each synoptic category, along with the standard deviation, was then determined to ascertain whether particular categories were distinctively high or low. Potential lag times were accounted for by evaluating the daily synoptic category on the day ofthe deaths, as well as 1, 2, and 3 days before the day ofthe deaths. Daily mortality was then sorted from highest to lowest during the period of record to determine whether certain synoptic categories were prevalent near the top or bottom.
Mean daily pollution concentrations were also examined for each synoptic category in an attempt to determine if the synoptic situations associated with the highest mortailty are also the most polluted. The EPA made available mean daily pollution concentrations for six raw pollutants at several automatic recording monitoring stations in the St. Louis area. The monitor with the most complete record was selected for this study, and mean daily values of the following pollutants were extracted: total suspended particulates (TSP), NO., NO2, total oxidants, 03, and S02. It was also noted whether the highest mortality days within each synoptic category were the most polluted.

Synoptic Categorization
Ten major summer synoptic categories were uncovered for St. Louis and were identified based on their mean meteorological characteristics (Table 1). Two hot, oppressive synoptic situations, categories 6 and 9, were identified, with the latter being more extreme. Several anticyclonic categories were identified, which generally possess the most stable atmospheric situations. Two categories (4 and 8) were frequently associated with convective precipitation patterns and represented unstable atmospheric conditions. Mean pollution concentrations varied considerably between categories ( Table 2). The three anticyclonic synoptic patterns (categories 1, 2, and 3) generally possessed the highest concentrations for most of the pollutants. Category 2 exhibited the highest mean TSP, ozone, and NO, concentrations, while category 1 was highest for NO2 and oxidants. The only exception was for SO2, where the highest concentrations were found within maritime tropical category 5. The two hottest synoptic situations, categories 6 and 9, were generally not distinctive, except for a moderately high TSP reading for category 6. The most oppressive synoptic situation, category 9, generally possessed mean pollutant readings near or below the overall norm. A cursory evaluation ofthe very hottest category 9 days indicated that they were no more polluted than any of the other days within the category.

Mortality Associations
An evaluation of mortality for the 10 synoptic categories uncovered some startling results (Table 3). After examining the potential lag-time relationships, it was determined that a 1-day lag existed between the causal synoptic mechanism and the mortality response (the between-category differential in mean daily mortality was greatest for a 1-day lag situation), with mean mortality values up to 52% higher than the other categories. The distinction between category 9 and the other synoptic situations was especially strong for elderly and nonwhite mortality. Interestingly, the second hottest synoptic situation, category 6, possessed mortality means that were almost identical to the other nondistinctive categories. This seems to imply that weatherinduced mortality rises dramatically only after a particular oppressive threshold is exceeded. It appears that only category 9 possesses conditions that exceed this threshold for St. Louis.
It is noteworthy that almost all the highest mortality days in St. Louis occurred when category 9 was present (Table 4). For total mortality, 8 ofthe 10 highest days occurred under the domination of category 9, even though this oppressive category occurred only 7% ofthe time during the period of study. This finding suggests that a vast majority of deaths that can be attributed to weather occur only when category 9 is present. For the 20 highest mortality days, category 9 was present 11 times. Although this is a bit less impressive, this still represents 8 times the expected occurrence of category 9 over a typical 20-day period. Category 6 only occurred once within the top 20 mortality days, further suggesting that the second hottest synoptic category is not oppressive enough to induce any extra deaths. Categories 3 (anticyclonic, marine influence) and 5 (maritime tropical, cloudy) are present among the top 20 days, but their frequencies are  not much different from their expected occurrence over a typical 20-day period.
Standard deviations in mortality within the 10 synoptic tyI indicate further that category 9 is distinctive ( Table 5). The stE dard deviations for all the synoptic situations are somewl similar with the very apparent exception of category 9, whi possesses values about three times higher than the other ca gories. This indicates that, although most ofthe highest mortal days are associated with category 9, numerous days within t synoptic type possess much lower mortality totals. An evaluati of all days within category 9 confirms this assumption, a numerous category 9 days exhibit mortality totals at sometimes below the means of the other synoptic types.
In an attempt to understand why daily mortality shows su considerable variation within category 9, a detailed with: category evaluation was performed for total and elderly morta ty. The objective was to determine which meteorological a pollution-oriented aspects of synoptic category 9 are respon ble for the highest mortality. To  with the temperature hovering near 95 F for many hours appears to have a greater influence on mortality than a day where the temperature exceeds 100T for a short time but falls rapidly to normore tolerable levels [refer to Kalkstein and Davis (14) for details on the computation]. A final nonmeteorological independent variable was included, "day in sequence," which noted how a particular category 9 day was positioned within a consecutive day my sequence. For example, if the day in question was the third of a category 9 consecutive day sequence, it was assigned a value of pes 3. All single-day category 9 days that were not part of a conan-secutive sequence were assigned a value of 1. hat The results explained much ofthe variance in mortality within ich category 9 ( Table 6). The day in sequence variable explained the te-largest proportion of the variance in both the total and elderly lty mortality models, and the direct relationship indicated that a his category 9 day near the end ofa consecutive day sequence is most ion nd ibble 5   bThis value represents an adjusted R2, which has been adjusted for degrees of freedom in the model.
CThis is the day-in-sequence variable. Refer to text for explanation.
important. Thus, a long consecutive day sequence of category 9 days imposes the greatest amount ofstress as expressed by increasing mortality through the sequence. Afternoon wind speed was the second-ranking independent variable in both models, and low wind speeds seem to exacerbate stressful conditions. This finding runs counter to some studies that indicate that higher wind speeds at very warm temperatures increase stress on the body (21,22). Not surprisingly, minimum temperature was a statistically significant variable and was directly related to mortality within category 9. However, maximum temperature was not significant in either the total or elderly models. Visibility was quite important in the total mortality model, and fell just short of statistical significance at the 0.05 level. Interestingly, visibility was directly related, indicating that visibility-reducing high pollution concentrations are not important contributors to increased mortality. Rather, this relationship might imply that the heat load imposed by a transparent atmosphere that increases insolation might be important by further increasing stress.

Discussion and Conclusion
One of the major findings from this synoptic evaluation is evidence suggesting that fluctuations in daily mortality appear to be much more sensitive to stressful weather thanto high levels of pollution. Those anticyclonic synoptic categories with the highest mean pollution concentrations were not associated with unusually high numbers of deaths. Conversely, category 9, which was associated with the highest daily mortality totals, did not possess high mean pollution concentrations for the six pollutants examined. Further, those days within category 9 that were among the top 20 mortality days through the period of study exhibited pollution concentrations somewhat near (and sometimes below) the mean.
These findings run counter to research involving mortality and air pollution in London (23) and morbidity and air pollution in Los Angeles (24). The London study detected a significant relationship between high mortality and British smoke (particulates), while the Los Angeles study noted that chest discomfort, eye irritation, and headache probabilities were associated with high NO2 and SO2 concentrations.
It is quite possible that pollution levels in London are often higher than those in St. Louis, as stringent U.S. clean air laws might keep pollutant levels below health-damaging thresholds.
To determine ifthis was the case, a cursory evaluation similar to the St. Louis analysis was performed for 9 additional U.S. cities (Table 7), and synoptic indices were developed for each of them. Preliminary results indicated that 7 of the 10  a To have a moderate or high weather/mortality signal, a city must possess a synoptic category that has an unusually high mean daily mortality. To have a moderate or high pollution/mortality signal, a city must possess a synoptic category with an elevated pollution level that has an unusually high mean daily mortality. In this preliminary analysis, this was determined subjectively. Another interesting result is the absence of maximum temperature as a significant variable within the category 9 regression models, a surprising finding considering the importance of minimum temperature within both models. This supports a contention made in a previous nonsynoptic climate/mortality study that suggests that oppressive nighttime conditions offering little relief after a very hot day might be more stressful than the maximum temperature itself (14).
The results ofthis evaluation strongly suggest that the synoptic climatological approach is more robust than traditional climatological procedures in the evaluation ofweather/mortality relationships. It was noteworthy that only 1 of 10 synoptic categories which are present in St. Louis during summer is oppressive enough to raise daily mortality beyond baseline levels. This supports the notion that mortality does not increase linearly as weather becomes increasingly stressful. Rather, threshold conditions exist beyond which mortality significantly increases in abrupt fashion.
It is possible that the procedure described here has the potential to be used in a weather/mortality watch-warning system, designed to warn of impending meteorological conditions that could raise mortality significantly among high-risk groups such as the elderly. For example, if a category 9 synoptic situation is approaching, a within-category evaluation can be performed based on the regression analyses presented in Table 6. According to Box and Wetz (25), a regression model may be considered to be a worthwhile predictor ifthe actual F-ratio is larger than some multiple ofthe critical F The magnitude ofthis multiple is dependent on the degrees of freedom. Using the criteria set forth by Box and Wetz (25), the two models presented in Table 6 are acceptable predictors for summer mortality. At this point in the study, no model validation has yet been performed, and the results ofthis paper should be considered for analytical content rather than predictive capabilites. However, the next stage in this research will involve a detailed evaluation ofthe nine additional cities to add further credence to this methodology. In addition, model validation on a data set not used in model development is neccessary to determine if, in fact, this procedure has predictive value in a weather/mortality watch-warning system.