Time of the year has long been associated with morbidity and mortality.1–5 Seasonal extremes of heat and cold are associated with increased mortality and are linked to the well-described clinical entities of hyperthermia and hypothermia. Extreme high temperatures have been associated with deaths among the elderly and also young children,2 which are both large susceptible populations.6–8 In many temperate countries, death rates during the winter season are 10% to 25% higher than those in the summer.9 The seasonal differences in morbidity and mortality reflect factors beyond the weather, including, most importantly, seasonal patterns of respiratory infections. Consequently, assessments of the effect of weather on human health have usually included control for season and sometimes for influenza epidemics so as to examine shorter-term effects of weather while excluding seasonal confounding.10
Even excluding extremes, both higher and lower temperatures are associated with increased mortality in a pattern that depends on latitude.11,12 The effects of humidity on mortality have received little investigation.
To date, time-series studies on seasonal variation in health and disease status have largely used mortality as the outcome measure. However, effects on morbidity would also be anticipated given the present understanding of underlying mechanisms of temperature effects. Cold temperatures could lead to thrombosis through hemoconcentration,13 and physical activity during cold weather can precipitate angina pectoris and myocardial infarction. Human bodies adapt to heat by increasing cardiac output to increase skin surface blood circulation, which facilitates heat loss. Volume depletion or dehydration can limit this cardiovascular process,7 as can some widely used medications. Prolonged heat exposure is associated with heat cramps, heat syncope, heat exhaustion, heatstroke,14 and even acute renal failure.15 Blood viscosity and cholesterol have been found to increase with high temperatures.16
Given the potential severity of these consequences of heat exposure, it is not surprising that hospital admission rates have been found to be elevated during heat waves. For example, during the 1980 heat wave in Kansas City, a 5% increase in hospital admissions was observed,17 and during the 1995 Chicago heat wave, sharp increases in hospital visits for cardiovascular diseases and a wide range of underlying medical conditions were documented.18
In contrast to mortality, little research has addressed short-term weather variations and hospital admissions or other indicators of medical morbidity. Consequently, we have examined the effect of temperature on hospitalization in the elderly using the database of the Centers for Medicare and Medicaid Services, formerly the Health Care Financing Administration. We have applied methods that control for seasonal affects and that appropriately and flexibly model the relation of temperature and humidity with hospitalization.
Our approach, using contemporary time-series methods, flexibly controls for seasonal and other longer-term confounding and does not make rigid, a priori assumptions about the lag relationship between temperature and outcome. We thus improved on the methods of earlier studies of mortality, which generally used models with relatively fixed and ad hoc assumptions.12,19,20 For example, one strategy has been to examine temperature on the same day as the event and the mean of temperatures on several prior days.19 In some cases, the averaging has extended back for several weeks.20
We have used a regression spline model21 to flexibly control for season, with distributed lags22 to allow the effect of temperature or humidity to vary over time. This allows us to assess the occurrence of several phenomena without making strong assumptions as to the lag-risk relationship for temperature.
The term “harvesting” refers to an effect of an environmental agent on a susceptible pool with brief temporal advancement of an event (eg, mortality). If an exposure such as a heat wave depletes the susceptible pool, then we would anticipate a reduction of events for subsequent days until the pool size is reestablished. A “harvesting” effect has been documented in previous studies on heat mortality.23,24 Longer rather than shorter lags could be relevant if one effect of temperature extremes is to worsen established illnesses. These possibilities can be explored using a distributed lag model. This article takes such an approach in an analysis of hospitalization data for 12 U.S. cities.
Daily counts of urgent hospital admissions for cardiovascular disease in general (International Classification of Diseases, 9th revision [ICD-9] codes 390–429) and myocardial infarction (ICD-9 code 410) specifically in persons age 65 years and older were extracted from Medicare files obtained from the Health Care Financing Administration for the years 1986 through 1994. These records capture all hospital admissions for persons in this age group. These extractions were done for the metropolitan counties containing Atlanta, GA, Birmingham, AL, Canton, OH, Chicago, IL, Colorado Springs, CO, Detroit, MI, Houston, TX, Minneapolis–St. Paul, MN, New Haven, CT, Pittsburgh, PA, Seattle, WA, and Spokane, WA. Daily weather data were obtained from the nearest airport station (EarthInfo CD NCDC Surface Airways; EarthInfo Inc., Boulder, CO). As the major outcome variable, all cardiovascular admissions were examined together; in addition, admissions for acute myocardial infarction were considered separately, because numbers were sufficient and the diagnostic accuracy of that event is high.
Counts of daily admissions were modeled separately for each city using Poisson regression, with regression splines to control for covariates. This approach16 allows variables expected to have a nonlinear effect on the outcome to be modeled using piecewise polynomials. For example, a separate cubic polynomial can be used for each 3-month interval to control for seasonal variation in hospital admissions. The number of intervals chosen determines the overall number of degrees of freedom of fit. This type of model fit produces a flexible, but smooth, curve. Hospital admissions show a seasonal pattern, rising and falling each year, but the height and location of the winter peak can vary from year to year. By using a regression spline, rather than a trigonometric function, the model captures this pattern of cycling with year-to-year variation.25 We also controlled barometric by using a regression spline. This modeling approach has been widely used in studies of daily deaths and hospital admissions26–29 We used indicator variables to control for day-of-the-week effects.
Seasonal patterns of hospitalization can vary between cities, and consequently a separate number of degrees of freedom was chosen in each city to both eliminate seasonal patterns in the admissions and reduce the residuals of the regression to “white noise”30 (ie, to remove serial correlation). The number of degrees of freedom for barometric pressure was chosen separately in each location by minimizing Akaike's Information Criterion,31 a measure of fit.
For daily mean temperature and relative humidity, we fit models examining the effects of exposures on each day up to 3 weeks before the event, which should be sufficient to capture any delayed effects.
Distributed Lag Models
Distributed lag models have been extensively used in the social sciences,32 and Pope and Schwartz33 recently described their use in epidemiology. We have previously applied this methodology to estimate the distributed lag between all-cause mortality and air pollution in 10 U.S. locations,34 and between temperature and humidity and daily deaths.6 The distributed lag model flexibly allows for effects of temperature and humidity on hospital admissions not merely on the same day, but on subsequent days. Conversely, the effect of temperature (or humidity) on admissions depends on today's temperature, the 1-day lag effects of yesterday's temperature, and possibly on still previous days. Therefore, suppressing covariates for the moment, the unconstrained Poisson distributed lag model assumes:
where Xt-q is the temperature q days before the admission. In this study, we examined the effect of temperature in the 12 cities on admissions with latencies (lags) ranging from zero until 20 days before the admission.
An additional modeling issue is what type of terms to use to model the effect of weather on health. Linear terms for temperature, on the same or lagged days, or lagged moving averages, have been applied in many studies.35–37 Nonlinear terms (smooth functions) also have been used.38,39 Kalkstein and coworkers40 proposed grouping meteorologic factors using factor analysis to create categories of meteorologic variables. In general, the effects of temperature and humidity on human health are adverse at both extremes. This pattern suggests that any analysis should incorporate potential nonlinearity at each lag examined. To do so, we used both a linear and a quadratic term for temperature at each lag. Equation 1 can be recast as:
where the ωi are parameters to be estimated, to reflect the relative impact of temperature at each lag.
Because there is substantial correlation between temperatures on days close together, this regression model will have a high degree of collinearity. This will result in unstable estimates of the individual ωi, and hence poor estimates of the shape of the distribution of the effect over lag.
To gain more efficiency and more insight into the shape of the distributed effect of the temperature over time, it is useful to constrain the ωi. If this is done flexibly, substantial gains in reducing the noise of the unconstrained distributed lag model can be obtained with minimal bias.37 The most common approach is to constrain the shape of the variation of the ωi with lag number to fit some polynomial function. We adopted a fourth-degree polynomial for both the linear and quadratic temperature terms in our models, because that should be flexible enough to encompass any plausible pattern of delayed effect over time. The model uses 10 degrees of freedom to fit the surface describing the effect of temperature on hospital admissions for heart disease by day of lag between exposure and admission. Note that the use of, for example, the mean of the temperature 1 to 20 days before the admission as an exposure variable represents the constraint that β1 = β2 = .. = βq, which is a much stronger constraint. Linear and quadratic terms for relative humidity up to 20 days before the admission were also included in the model, subject to similar constraints.
By fitting the same model in 12 different locations, and combining effect size estimates by lag over the cities, we can obtain an estimate of the distribution of the effect of temperature and humidity over time. To combine results across cities, we used inverse variance weighted averages, including a random variance component to incorporate heterogeneity.41 The analysis was done with S-PLUS software (Insightful, Inc., Seattle, WA).
Table 1 presents the descriptive characteristics of the cities in the study. Together, their total population was nearly 19 million, including 2.2 million persons age 65 years and older. During the study period, there were a total of 1.3 million cardiovascular admissions. Houston was both the warmest and the most humid city in the study. Minneapolis, on the other hand, was the coldest city and had the widest range of temperature. The smallest variation of temperature was observed in Seattle.
Table 2 presents the correlation coefficients of temperature with humidity and barometric pressure. Temperature and humidity were negatively correlated in most of the cities, except Birmingham and New Haven. Spokane and Seattle showed the largest negative correlations. Temperature and barometric pressure generally had small negative correlations, except in Colorado Springs.
The patterns of the dependence of hospital admissions on lag did not vary by city. Figure 1 illustrates this similarity, showing the covariate-adjusted (including humidity) estimated effects in 4 of the 12 cities, representing various regions and climates. Consequently, the estimates in the 12 cities were combined in a metaanalysis, shown in Figure 2.
The effect of hot temperatures appeared to be immediate (Fig. 2). The positive effect observed at lag 0 was followed by a period of lower-than-average admissions, returning to the baseline after a week. This pattern, referred to as the harvesting effect, was not seen for the protective effect of cold weather. In particular, the estimated increased risk of a 24-hour mean temperature of 80°F (compared with 0°F) on the day of admission was 1.15 (95% confidence interval [CI], 0.96–1.37). The coefficients for both the linear and quadratic temperature terms were significant at this lag; the pointwise confidence intervals are wider at this temperature because it is well above average, and confidence intervals for pointwise estimates grow as one diverges from the mean. In contrast, lag 6 on Figure 2 shows the estimated risk 6 days after a day with mean temperature of 80°F. This risk was 0.95 (0.89–1.00), showing the lower-than-expected admission count following the higher-than-expected count a few days earlier.
There was not a clear pattern across city for the effect of humidity on cardiovascular admissions, either in terms of lag structure or in terms of differences between high and low humidity. The overall estimate of the relative humidity effect on admissions is presented in Figure 3. There is no evidence of a consistent association of humidity with the number of daily admissions. With stratification of the cities by weather characteristics, effects were still not seen (data not shown).
Figure 4 presents the overall estimate of the effect of temperature on myocardial infarction admissions in the 12 cities. Overall, the analysis shows a similar pattern but with smaller effect sizes than for all cardiovascular admissions.
Mortality has been the outcome measure considered in most studies of weather or temperature and health, and the present study is one of the first to systematically examine temperature and morbidity. We analyzed delayed (up to 3 weeks) and potentially nonlinear effects of weather on hospitalization across geographically diverse cities in the United States. Controlling for season, hotter temperatures were associated with increased admissions for cardiovascular disease. For the hottest temperatures, some of these admissions appeared to be displaced from occurring on subsequent days, a pattern consistent with a harvesting effect. For less-extreme temperatures, there was no suggestion of harvesting.
These results are consistent with mortality studies that consistently show a rise in cardiovascular deaths during heat waves followed by lower-than-expected mortality, ie, a harvesting effect.5,23,42 We did not find evidence for harvesting with extreme cold temperature, also consistent with the mortality studies.5,43
In view of findings for mortality, we were surprised to find a linear relation of temperature with hospitalization, rather than the nonlinear relation as found for mortality. In contrast, for all-cause mortality5,37 and for cardiovascular and respiratory mortality,44,45 the temperature-related excess increases as temperatures deviate in either direction away from a temperature at which no increased risk is observed.
Several hypotheses can be offered for the difference between our findings for hospitalization and the studies of mortality. The mortality outcomes are more nonspecific, and analyses have included both heart disease and respiratory disease deaths; the former could increase more monotonically with temperature than do respiratory deaths, which are highest in the winter season. The combination of these 2 dominant causes of death might produce a U- or J-shaped curve.
The seasonality of respiratory infections would result in an association with cold temperature if season were not controlled given the seasonality of temperature. However, within the winter period, respiratory disease is not strongly temperature-dependent and the timing of annual outbreaks is not strongly associated with monthly mean temperatures.46
Finally, extreme cold can be accompanied by inclement weather, effectively reducing access to a hospital or inclination to be admitted. A greater effect of such influences of weather on behavior would be anticipated for office visits than for hospital admissions for more severe illness.47 However, if the failure to find increased risk on cold days were related to persons postponing their admissions, one would expect to find a peak several days after a cold day, which is not seen in our results. Also, an increase in mortality from myocardial infarctions has also been associated with heavy snowfall,48–50 likely related to overexertion; mortality from pneumonia, however, involves neither of these behavioral attributes. To test these hypotheses, systematic joint examination of cause-specific mortality and hospital admissions in multiple locations will be needed.
We selected cities that experience a wide range of temperatures. The 95th percentiles of temperature ranged from 30°C in Houston to 20.6°C in Seattle, whereas the 5th percentile of temperature ranged from 7.2°C in Houston to −13.3°C in Minneapolis. The similar patterns for the temperature dependence by lag across this range suggest there is little modification of the effect of temperature by annual mean climate or geography across the United States. This apparent lack of modification differs from results for daily deaths in the United States.5,38 In contrast to the effects of temperature, consistent patterns were not seen for the relation of humidity to daily hospital admissions.
The distributed lag approach proved informative in characterizing the lag structure for temperature and hospitalization. The results shown in Figure 2 suggest that modeling out to lags of up to 21 days was more than sufficient, and that days beyond 7 to 10 days previously had little effect. This lag pattern is consistent with results reported by Braga and coworkers5 for mortality using 12 U.S. cities. They found the temperature effects did not persist beyond 2 weeks. A similar study in 11 eastern U.S. cities38 incorporated multiple lagged intervals and found that most mortality was explained by mean temperature on the same day as death. Little was gained in model fit by including the prior 3 days of temperature or dew point, and very little, if any, improvement in fit was found by including values preceding death by more than 3 days.
These analyses were partially motivated by the need to develop models that can be used to project temperature-related morbidity as a result of global climate change. Global warming is likely to bring not only warmer temperatures on average (ranging from 1.7 to 4.9°C by the year 2100),51,52 but greater frequency of extreme weather events, including hot days. Therefore, our use of mean daily temperatures, rather than maximum or minimum temperature, could limit the interpretation of these results for purposes of prediction. On the other hand, in their 11-city study, Curriero et al.38 included a variable, “daily spread in temperature,” to examine if within-day temperature variability contributed to predicting mortality and found no significant change in model fit.
A limitation of this study is the failure to control for confounding by air pollution. Unfortunately, the pollutant most convincingly associated with acute cardiovascular hospital admissions is airborne particles, which are measured only one day in 6 in many cities and only for certain years in other cities. In addition to the loss of power that would result from including a variable with so much missing data in our regressions, the missing data would also prevent us from examining harvesting or longer-term responses. Hence, pollution was not included in this analysis.
The present analyses focused on heart disease admissions in the elderly, a morbidity indicator that should be highly sensitive because of the susceptible population considered. Future research should address other morbidity outcome measures such as hospital admissions for respiratory disease (including influenza), diabetes, and other potentially temperature-related outcomes. Comparative studies should be conducted using other climate indicators such as apparent temperature or synoptic air masses to optimize the potential for predictability and effective early warning. Finally, personal exposure assessment studies, which measure the temperature proximate to a person during the course of a day, will improve the accuracy of predicting temperature's effect on human health.53 If people go outdoors less on cold days, for example, this would introduce increasing error into our exposure measure as temperature falls. Improving our understanding of such issues will improve our ability to accurately estimate the effects of cold weather.
1.Ellis FP. Mortality from heat illness and heat-aggravated illness in the United States. Environ Res. 1972;5:1–58.
2.Buechley RW, Van Bruggen J, Truppi LE. Heat island equals death island? Environ Res. 1972;5:85–92.
3.Doyle R. By the numbers: deaths from excessive cold and excessive heat. Sci Am. 1998;February:P26.
4.Patz JA, McGeehin MA, Bernard SM, et al. The potential health impacts of climate variability and change for the United States: executive summary of the report of the health sector of the US national assessment. Environ Health Perspect. 2000;108:367–376.
5.Basu R, Samet J. The relationship between elevated ambient temperature and mortality: a review of the epidemiologic evidence. Epidemiol Rev. 2002;24:190–202.
6.Centers for Disease Control and Prevention. Heat-related deaths—United States. MMWR Morb Mortal Wkly Rep. 1993;42:558–560.
7.Greenberg JH, Bromberg J, Reed CM, et al. The epidemiology of heat-related deaths, Texas—1950, 1970–1979. Am J Public Health. 1983;73:805–807.
8.Fish PD, Bennett GCJ, Millard PH. Heatwave morbidity and mortality in old age. Age Aging. 1985;14:243–245.
9.McMichael AJ, Githeko A, Akhtar R, et al. Human health. In: McCarthy J, Canziani O, Leary N, et al., eds. Climate Change 2001: Impacts, Adaptation, and Vulnerability. New York: Cambridge University Press; 2001.
10.Basu R, Samet JM. An exposure assessment study of ambient heat exposure in an elderly population in Baltimore, Maryland. Environ Health Perspect. 2002;110:1219–1224.
11.Braga ALF, Zanobetti A, Schwartz J. The time course of weather-related deaths. Epidemiology. 2001;12:662–667.
12.Keatinge WR, Donaldson C, Cordioli E, et al. Heat related mortality in warm and cold regions of Europe: observational study. BMJ. 2000;321:670–673.
13.The Eurowinter Group.Cold exposure and winter mortality from ischaemic heart disease, cerebrovascular disease, respiratory disease, and all causes in warm and cold regions of Europe. Lancet. 1997;10:1341–1346.
14.McGeehin MA, Mirabelli M. The potential impacts of climate variability and change on temperature-related morbidity and mortality in the United States. Environ Health Perspect. 2001;109(suppl 2):185–189.
15.Semenza JC. Acute renal failure during heat waves. Am J Prev Med. 1999;17:97.
16.Keating WR, Coleshaw SR, Easton JC, et al. Increased platelet and red cell counts, blood viscosity, and plasma cholesterol levels during heat stress, and mortality from coronary and cerebral thrombosis. Am J Med. 1986;81:795–800.
17.Jones ST, Liang AP, Kilbourne EM, et al. Morbidity and mortality associated with the July 1980 heat wave in St. Louis and Kansas City, MO. JAMA. 1982;247:3327–3331.
18.Semenza JC, McCullough J, Flanders DW, et al. Excess hospital admissions during the 1995 heat wave in Chicago. Am J Prev Med. 1999;16:269–277.
19.Samet JM, Dominici F, Curriero FC, et al. Fine particulate air pollution and mortality in 20 US cities, 1987–1994. N Engl J Med. 2000;343:1742–1749.
20.Keatinge WR, Donaldson GC, Bucher K, et al. Cold exposure and winter mortality from ischaemic heart disease, cerebrovascular disease, respiratory diseases, and all causes in warm and cold regions of Europe. Lancet. 1997;349:1341–1346.
21.Hastie T, Tibshirani R. Generalized Additive Models. London: Chapman and Hall; 1990.
22.Schwartz J. The distributed lag between air pollution and daily deaths. Epidemiology. 2000;11:320–326.
23.Kunst AE, Looman CW, Mackenbach JP. Outdoor air temperature and mortality in The Netherlands: a time-series analysis. Am J Epidemiol. 1993;137:331–341.
24.Rooney C, et al. Excess mortality in England and Wales, and in Greater London, during the 1995 heatwave. J Epidemiol Community Health. 1998;52:482–486.
25.Schwartz J. Generalized additive models in epidemiology. In: International Biometric Society, Invited Papers. 17th International Biometric Conference; August 8–12, 1994; Hamilton, Ontario, Canada. Washington, DC: International Biometric Society; 1994:55–80.
26.Schwartz J. Assessing confounding, effect modification, and thresholds in the association between ambient particles and daily deaths. Environ Health Perspect. 2000;108:563–568.
27.Hajat S, Haines A, Atkinson RW, et al. Association between air pollution and daily consultations with general practitioners for allergic rhinitis in London, United Kingdom. Am J Epidemiol. 2001;153:704–714.
28.Zanobetti A, Schwartz J. Race, gender, and social status as modifiers of the effects of PM10 on mortality. J Occup Environ Med. 2000;42:469–474.
29.Katsouyanni K, Touloumi G, Samoli E, et al. Confounding and effect modification in the short-term effects of ambient particles on total mortality: Results from 29 European cities within the APHEA2 project. Epidemiology. 2001;12:521–531.
30.Schwartz J. Air pollution and hospital admissions for heart disease in eight US counties. Epidemiology. 1999;10:17–22.
31.Akaike H. Information theory and an extension of the maximum likelihood principal. In: Petrov BN, Csaki F, eds. 2nd International Symposium on Information Theory; Akademiai Kiado, Budapest; 1973.
32.Judge G, Griffiths WE, Hill RC, et al. The Theory and Practice of Econometrics. New York: John Wiley and Sons; 1980.
33.Pope CA III, Schwartz J. Time series for the analysis of pulmonary health data. Am J Respir Crit Care Med. 1996;154:S229–S233.
34.Schwartz J. The distributed lag between air pollution and daily deaths. Epidemiology. 2000;11:320–326.
35.Schwartz J, Dockery DW. Increased mortality in Philadelphia associated with daily air pollution concentrations. Am Rev Respir Dis. 1992;145:600–604.
36.Saldiva PHN, Pope CA III, Schwartz J, et al. Air pollution and mortality in elderly people: a time-series study in São Paulo, Brazil. Arch Environ Health. 1995;50:159–163.
37.Kelsall JE, Samet JM, Zeger SL, et al. Air pollution and mortality in Philadelphia, 1974–1988. Am J Epidemiol. 1997;146:750–762.
38.Schwartz J. Air pollution and daily mortality: a review and meta analysis. Environ Res. 1994;64:36–52.
39.Schwartz J, Dockery DW, Neas LM. Is daily mortality associated specifically with fine particles? J Air Waste Manage Assoc. 1996;46:927–939.
40.Kalkstein LS, Tan G, Skindlov J. An evaluation of objective clustering procedures for use in synoptic climatological classification. J Climate Appl Meterol. 1987;26:717–730.
41.Berkey CS, Hoaglin DC, Mosteller F, et al. A random-effects regression model for meta-analysis. Stat Med. 1995;14:395–411.
42.Kalkstein LS, Greene JS. An evaluation of climate/mortality relationship in large US cities and the possible impacts of a climate change. Environ Health Perspect. 1997;105:84–93.
43.Huynen MTE, Martens P, Schram D, et al. The Impact of heat waves and cold spells on mortality rates in the Dutch population. Environ Health Perspect. 2001;109:463–470.
44.Curriero FC, Heiner KS, Samet JM, et al. Temperature and mortality in eleven cities of the eastern United States. Am J Epidemiol. 2002;155:80–87.
45.Braga ALF, Zanobetti A, Schwartz J. The effect of weather on respiratory and cardiovascular deaths in 12 US cities. Environ Health Perspect. In press.
46.Langford IH, Bentham G. The potential effects of climate change on winter mortality in England and Wales. Int J Biometeorol. 1995;38:141–147.
47.Wichmann HE, Mueller W, et al. Health effects during a smog episode in West Germany in 1985. Environ Health Perspect. 1989;79:89–99.
48.Spitalnic SJ, Jagminas L, Cox J. An association between snowfall and ED presentation of cardiac arrest. Am J Emerg Med. 1996;14:572–573.
49.Gorjanc ML, Flanders WD, VanDerslice J, et al. Effects of temperature and snowfall on mortality in Pennsylvania. Am J Epidemiol. 1999;149:1152–1160.
50.Glass RT, Zack MMJ. Increase in deaths from ischaemic heart disease after blizzards. Lancet. 1979;1:485–487.
51.IPCC, Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change. In: Houghton J, et al., eds. Climate Change 2001: The Scientific Basis. New York: Cambridge University Press; 2001:881.
52.Wigley TM, Raper SC. Interpretation of high projections for global-mean warming. Science. 2001;293:451–454.
53.Basu R, Samet JM. An exposure assessment study of ambient heat exposure in an elderly population in Baltimore, Maryland. Environ Health Perspect. 2002;110:1219–1224.
© 2004 Lippincott Williams & Wilkins, Inc.