ArticlePlus
Click on the links below to access all the ArticlePlus for this article.
Please note that ArticlePlus files may launch a viewer application outside of your web browser.
Extremes of temperature are well known to be associated with short-term increases in daily deaths.1–8 Greater relative increases have been reported for the elderly,3,9 and, in some studies, for persons in lower social conditions.4,10 Social isolation also has been reported to increase risk.11,12 Less is known about the effects of chronic disease or other medical conditions on the risk of death associated with temperature extremes.
To study which factors modify a person's risk due to a time-varying exposure, it is possible to restrict analysis to persons who died. The case-only approach has been widely used in genetic epidemiology, and Armstrong13 has recently pointed out that it can be extended to study the acute effects of weather.
The underlying idea behind this approach is that if a condition increases the risk of dying on an unusually hot (or cold) day, a greater proportion of the people who died during those periods would be expected to have the condition, compared with people who died during milder weather conditions. Hence, for example, if a condition is a risk modifier for deaths during unusually hot days, then unusually hot days should be a predictor of the occurrence of that condition on death certificates using logistic regression. A more formal proof of the approach is provided by Armstrong.13
This approach has some important advantages. Because only cases are examined, variables that are typically associated with the likelihood of death are generally not potential confounders. To be a confounder, variables must also be associated with extreme temperature, which does not apply to such traditional risk factors as smoking history, blood pressure, and body mass index. Age, sex, social status, and prevalence of chronic conditions also would not be expected to fluctuate over time in association with weather. Although they are not potential confounders, these variables may be examined as potential effect modifiers.
Time-varying predictors of mortality risk, such as season, likewise should not confound the association with chronic conditions, because those conditions will not fluctuate over time. This is an advantage of the case-only approach, because it can reduce or eliminate the need for complex modeling of the change in risk of death over time as the result of factors other than temperature.
A limitation of the approach is that validity depends on the assumption that the modifier and the exposure are not associated in the base population that gave rise to the deaths. Because the presence of chronic conditions varies among individuals but does not fluctuate from day to day with temperature, this assumption (which is a major issue in studies of gene–environment interactions) is clearly met in studies of weather.
However, as Armstrong13 points out, if there are other time-varying predictors of death that also interact with the chronic conditions being examined, the simplification of the covariate model is not complete. This situation is most likely to occur for season, where persons with certain conditions may show enhanced seasonal increases in risk of death. In addition, contrasting extreme temperatures to all other temperatures may miss a more continuous dependence on the other temperatures. I examined whether age, sex, race, and medical conditions modified the response to extreme temperatures in a population-based study of persons 65 years of age and older. I also checked for evidence to support differential seasonal patterns and response to more common temperatures among these subgroups of the elderly and assessed the sensitivity of the initial results to control for these 2 interactions.
METHODS
In the United States, hospital admissions of all elderly residents are covered by Medicare, a national health insurance plan. Because the elderly are more sensitive to extreme weather conditions,10 and Medicare allows record linkage to help identify preexisting conditions, I started by obtaining Medicare records of all hospitalizations for heart or lung disease in Wayne County, Michigan (which contains Detroit and its surrounding communities), in persons 65 years of age and older between 1984 and 1998. These records capture all hospital admissions for persons in this age group. In the 1990 census, Wayne County had a population of 2,111,687, of whom 263,900 were 65 years of age or older. Unique identifiers for each beneficiary were used to identify multiple admissions per subject.
For each admission, the primary discharge diagnosis, as well as the secondary contributing diagnoses, were examined for the presence of the following conditions: myocardial infarction (MI; International Classification of Diseases [ICD]–9 code: 410); diabetes (ICD–9: 250); chronic obstructive pulmonary disease (COPD; ICD–9:490–496); congestive heart failure (CHF; ICD–9: 428); and pneumonia (ICD–9: 480–486). Asthma (ICD–9: 493) was included in COPD for this analysis because the distinction is difficult in this age group. Individuals for whom the condition was noted on any admissions were classified for the mortality follow-up analysis as having the condition. The analysis was limited to subjects who were discharged after at least 1 hospital admission and died subsequently with a validated date of death. This information also is contained in the Medicare files obtained from the Center for Medicare and Medicaid Services. The last known address was used to identify subjects resident in Wayne County. A total of 160,062 eligible deaths occurred in the follow-up period (up to December 31, 1999).
Finally, based on information provided in the Medicare files, I examined whether sex, nonwhite status, or age greater than 85 years were modifiers of the effect of temperature extremes. In this analysis, the outcome variables were the presence or absence of the medical conditions noted previously or whether the decedents were nonwhite, female, or age 85 years and older.
Weather Data
Daily weather data were obtained from the Detroit airport station (EarthInfo CD NCDC Surface Airways, EarthInfo Inc., Boulder, CO). To examine the effect of extreme temperatures, I created several exposure variables. Hot days were defined as those at or greater than the 99th percentile of the minimum daily temperature. Minimum temperature was chosen because it indicates scenarios in which there is little relief at night. Recent studies suggest this better represents the types of events producing heat–related deaths.14 A second heat indicator was taken to be days greater than the 99th percentile of the 3-day moving average of minimum temperature because some recent studies have suggested that multiple days with extreme temperature may be more important than a single day.15,16 Cold days were defined as days at or less than the first percentile of daily maximum temperature,17 and a second indicator for the 3-day moving average was also constructed. Analyses were based on the temperature on the day of death, or for the 3-day moving average, on the day of death and the 2 preceding days.
Analytic Methods
Because this approach is relatively recent, for convenience I here replicate part of the argument of Armstrong.13 Readers are referred to that article for further details. Assume that the daily counts of deaths (Yt) among subjects can be modeled as a Poisson regression:
where Hot is an indicator variable for whether that day is at or greater than the 99th percentile of temperature and other predictors denote seasonal terms and any other predictors of changes in risk over time. In recent years, these parts of the model have become quite complex.
Then, if Hot = 0 and Diabetes = 0, the overall number of cases expected over the course of the study is
If Hot = 1 and Diabetes = 0, it is
If Diabetes is 1, then the same sums are obtained, except they are multiplied by exp(β2) if Hot is zero, and by exp(β2 + β3) if Hot is 1. Hence a two-by-two table of death counts by hot temperature and diabetes can be constructed, yielding an odds ratio (OR) for the interaction of exp(β3). But this two-by-two analysis corresponds to a logistic regression among deaths, predicting the presence of diabetes in the deceased with Hot as the exposure. That is,
;)
Hence, a logistic regression model predicting the presence of diabetes as a function of high temperatures would yield the same results as the Poisson model, without the need for the other time-varying covariates, which can be complex. Therefore, although the case-only analysis is conducted using only the subjects who died, the inference is to interactions in the Poisson model, which applies to the entire population that gave rise to the deaths. Hence, if the case-only analysis shows that a diagnosis of diabetes is a modifier of dying on extreme weather days, this indicates that persons with diabetes have a different risk than members of the baseline population who did not have diabetes.
Using data on deaths, I fit a logistic regression model with the indicators for both extreme weather conditions (hot and cold) as predictors and the presence or absence of the hypothesized modifying condition as the dependent variable. The analysis was performed separately using the indicators for extreme temperature on the day of death and the indicators for the extreme temperatures for the 3-day average ending on the day of death.
As noted by Armstrong,13 this approach can give inappropriate results if the modifier of interest also is a modifier of one of the other time-varying covariates. In that case, the main effect of the time-varying covariate would cancel out, but the interaction would remain. In particular, it seems likely that some chronic conditions may modify not merely the sensitivity of persons to extreme temperatures but also sensitivity to seasonal patterns, which reflect, for example, time spent indoors and proximity to other persons. If this is the situation, then a seasonal component should be included in the logistic model to capture, for example, the greater amplitude of the seasonal increase that may be seen in persons with the presence of a particular condition. That is, seasonal effect modification should be included.
It seems reasonable to assume that, however complicated the baseline seasonal pattern of mortality is, the additional nontemperature-related modifiers of risk by predisposing condition might be captured by a sine and cosine term with a 365.24-day period. I therefore used this approach.
In addition, the effect of temperature may be modified by the predisposing condition on a more continuous basis. To account for this effect modification, I included linear and quadratic terms for apparent temperature. Apparent temperature is a physiologically derived index of thermal stress that incorporates temperature and humidity, analogous to the wind-chill factor. It is based on research indicating that the physiologic impact of humidity occurs by modifying the body's ability to shed heat in hot weather through a reduction in evaporative loss. This can be translated into an equivalent temperature at moderate humidity that has the same thermal stress. Apparent temperature is an empirically derived construct capturing that phenomena16 and was constructed using the 24-hour average temperature and dew point temperature. Although apparent temperature can be corrected for high wind speeds, such high winds did not occur during any of the periods in this analysis, and so the correction was not necessary. All analyses were conducted 3 times (1) using the basic case only analysis, (2) adding seasonal sine and cosine terms, and (3) adding both seasonal terms and a nonlinear continuous dependence on apparent temperature.
RESULTS
Table 1 shows baseline descriptive statistics for the subjects followed in this study. More than half of the deaths were among women. As expected, the prevalence of the preexisting medical conditions among these subjects who died is considerably higher than their prevalence in the general population of this age.
TABLE 1: Baseline Descriptive Statistics For Deaths (1984–1999) in Detroit of Medicare Beneficiaries With Previous Admissions for Heart or Lung Disease
Detroit has a continental climate with a broad range of temperature. Hence, as displayed in Table 2 the coldest days were quite cold (maximum temperature less than −8.9°C), and the hottest days quite hot (minimum temperature >22.8°C).
TABLE 2: Distribution of Temperature (°C) Variables in Detroit, 1985–1999
Table 3 shows the results of the basic analysis looking at very hot and very cold days, defined either for 1 day or for a 3-day moving average. Persons older than 84 years of age showed greater effects of extreme cold, but not of extreme heat, compared with the effects on younger Medicare participants. Women were similarly at greater risk of death during periods of extreme cold. Nonwhites showed the greatest evidence of effect modification, with ORs of 1.22 (95% confidence interval [CI] = 1.09–1.37) for hot days and 1.25 (1.12–1.40) for cold days. Among medical conditions, persons with diabetes showed increased susceptibility to very hot days (1.17; 1.04–1.32), whereas persons with COPD had greater susceptibility on very cold days (1.19; 1.07–1.33), and marginally greater susceptibility on very hot days. Persons who survived MIs, in contrast, were less susceptible than Medicare patients who were hospitalized for other conditions on cold days (0.83; 0.69–0.99). Note that these are relative odds and do not mean that the absolute risk for these subjects is not elevated, just that it is elevated less. Subjects who previously had been admitted for CHF or pneumonia showed little evidence of having risks different from the average.
TABLE 3: Estimated Effect Modification by Subject Characteristics of the Risk Attributable to Extreme Temperature
Table 4 (available with the electronic version of this article) shows the results of the sensitivity analysis of control for interactions with season and continuous temperature terms. An extra digit is reported to facilitate evaluation of how much the effect sizes changed. Overall, the results for extreme temperatures were very stable when these other interactions where added. The strongest evidence for a change was the increase in the estimated risk of extreme cold weather among persons with COPD from an OR of 1.19 when examined alone to 1.26 when additionally controlling for temperature and season. Those risk estimates are well within each other's CIs.
Although control for these other interactions did not affect the estimated differential susceptibilities to extreme temperatures, there was some evidence for an independent effect of the additional variables. In particular, seasonal interactions were seen for persons older than 84 years of age, persons with COPD, nonwhite subjects, and persons with CHF. The relative risks for these terms were modest, typically between 1.01 and 1.03 for the difference between winter peaks and the annual average (data not shown).
Armstrong13 speculated in his article introducing the case-only approach, that the need to control for additional factors, such as seasonal interactions, might have an important impact on the power of the study to detect susceptibility to weather. This does not appear to be the case in these data. Table 5 (available with the electronic version of this article) shows, for several selected models, the standard errors for the extreme temperature terms in the base model, after including interactions for season alone, and for both season and temperature. Although the standard errors increased in the more complex models, the effect on precision was quite modest.
DISCUSSION
Chronic Conditions
The primary goal of this analysis was to examine the susceptibility to extreme temperatures of persons with chronic illnesses, assessed by using record linkage in the Medicare system. Perhaps the most interesting finding is that persons with diabetes appear to be more susceptible to very hot days (OR = 1.17; 95% CI = 0.04–1.32) but not very cold days. This finding was robust with control for possible interactions with season and with continuous temperature. An estimated 15.6 million Americans have diabetes, and the prevalence among the elderly is approximately 18%.18 The prevalence of diabetes is rising in many countries. Hence, this susceptibility represents a potentially growing public health problem.
The reasons for the increased response to very hot days among individuals with diabetes are unclear. People attempt to adapt to heat by increasing cardiac output, which increases skin surface blood circulation, and heat loss. Blood viscosity and cholesterol increase with high temperatures.19 Persons with diabetes have impaired autonomic control and endothelial function and, thus, it is possible that extreme thermal stress in these subjects in addition to the consequent increased demands on the circulatory system, interact with those functional impairments to increase risk of fatal events.
The increased susceptibility of persons with COPD to dying on extremely cold days is also noteworthy, although perhaps more explicable. The lungs of persons with COPD are typically colonized by bacteria, and cold weather can easily exacerbate respiratory infections.20 In addition, cold can induce bronchospasm,21 as well as increase platelet and red cell counts, and blood viscosity.19 Because persons with COPD often have cardiovascular complications, these effects on blood components may play a role.
Sociodemographic Factors
This study found greater risks of dying on both hot and cold days in nonwhites. This result is consistent with many previous studies4,7,10 and probably represents associations with physical deprivation and other social factors that limit adaptability.11,12
Another interesting finding is that women have increased risk of mortality on very cold days. This sex differential was not seen in the Eurowinter study.22 However, that study examined the slope of increased mortality as temperature fell continuously to levels less than 18°C; the excess risk may only be apparent at extreme temperatures such as examined here. In a study of clothing use in winter across Europe, it was reported that women wore 0.14 clo less thermally protective clothing than men,23 which might explain such an effect.
Interaction with Season
Seasonal terms were predictors of whether a death occurred in individuals 84 years of age or older, who were not white, who had COPD, or who had heart failure. These results indicate that persons in those categories, at least in Detroit, show seasonal patterns of deaths that are different from the general population. In particular, they show greater amplitude in the summer-to-winter increase in the risk of death, which may reflect greater susceptibility to infection or risk factors (other than extreme cold) that vary from season to season. In light of these findings, it seems prudent to recommend that future studies of short-term environmental factors, such as weather or air pollution, examine whether seasonal interactions confound any interactions.
It was encouraging that in this study seasonal interactions did not appear to be a confounder. The effect estimates for very hot or very cold days changed little when a seasonal interaction term was included. Further, any modest pattern was more suggestive of a slight increase in the estimated effect sizes after control for these additional interactions. Although such findings need to be confirmed in more locations, including areas with different climates, they support the use of the case-only approach to examine susceptibility. The modest size of the changes in the effect estimates for the extreme temperature terms is likely related to the modest size of the seasonal interactions. The effect size estimates for season suggested 1–3% higher risks than for the overall Medicare population. Of course, this may differ in other cities.
Control for continuous temperature, even using a nonlinear relation, did not reduce the estimated impact of extremely hot or cold days, suggesting that such continuous terms do not adequately model the effects of extremes on susceptible subgroups. Hence, the examination of indicator variables or other approaches to capture the effects of extreme temperatures on mortality risk seems to be needed.
Qualifications
It should also be emphasized that these results are expressed as relative odds. For example, the finding of a protective association of cold days for persons who survived MIs does not mean that those persons do not have an increased risk of dying associated with extreme cold. Rather, it indicates that their risk on those days is lower than that of other persons aged 65 years and older, who have been admitted to hospital for a heart and lung condition, but who have not had an MI. The question of risk, among survivors of MI, as a result of extreme cold compared with more moderate temperatures cannot be addressed by the case-only approach, which constitutes an important limitation. A previous article by Braga and coworkers2 examined deaths in Wayne County and reported elevations on both hot and cold days.
Another limitation is that this study focused, by design, on extreme temperature days, and these results cannot be extrapolated moderately hot or cold days. Finally, and these results may not be generalizable to other geographic areas. They do, however, illustrate the utility of the approach.
These analyses were partially motivated by the need to better understand the potential health impacts 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),24,25 but greater frequency of extreme weather events, including hot days. Understanding who is susceptible to those extreme events will be important in minimizing their public health impact.
REFERENCES
1. Rooney C, McMichael AJ, Kovats RS, et al. Excess mortality in England and Wales, and in Greater London, during the 1995 heatwave.
J Epidemiol Community Health. 1998;52:482–486.
2. Braga AL, Zanobetti A, Schwartz J. The time course of weather-related deaths.
Epidemiology. 2001;12:662–667.
3. Diaz J, Garcia R, Velazquez de Castro F, et al. Effects of extremely hot days on people older than 65 years in Seville (Spain) from 1986 to 1997.
Int J Biometeorol. 2002;46:145–149.
4. Curriero FC, Heiner KS, Samet JM, et al. Temperature and mortality in 11 cities of the eastern United States.
Am J Epidemiol. 2002;155:80–87.
5. Ellis FP. Mortality from heat illness and heat-aggravated illness in the United States.
Environ Res. 1972;5:1–58.
6. Doyle R. By the numbers: deaths from excessive cold and excessive heat.
Sci Am. 1998;February:26.
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. Huynen M, 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.
9. Fish PD, Bennett GCJ, Millard PH. Heatwave morbidity and mortality in old age.
Age Aging. 1985;14:243–245.
10. O'Neill MS, Zanobetti A, Schwartz J. Modifiers of the temperature and mortality association in seven US cities.
Am J Epidemiol. 2003;157:1074–1082.
11. Semenza J, Rubin C, Falter K, et al. Heat-related deaths during the July 1995 heat wave in Chicago.
N Engl J Med. 1996;335:84–90.
12. Klinenberg E. Denaturalizing disaster: a social autopsy of the 1995 Chicago heat wave.
Theory Society. 1999;28:239–295.
13. Armstrong B. Fixed factors that modify the effects of time-varying factors: applying the case-only approach.
Epidemiology. 2003;14:467–472.
14. Kalkstein LS, Jamason PF, Greene JS, et al. The Philadelphia Hot Weather-Health Watch/Warning System: Development and Application, Summer 1995.
Bull Am Meteorol Soc. 1996;77:1519–1528.
15. Saez M, Sunyer J, Castellsague J, et al. Relationship between weather temperature and mortality: a time series analysis approach in Barcelona.
Int J Epidemiol. 1995;24:576–82.
16. Kalkstein L, Valamont K. An evaluation of summer discomfort in the United States using a relative climatological index.
Bull Am Meteorol Soc. 1986;67:842–848.
17. Kalkstein LS, Davis RE. Weather and Human Mortality: an evaluation of demographic and inter-regional responses in the United States.
Ann. Assoc. Am. Geographers. 1989;79:44–64.
18. NIH.
Diabetes Statistics. Vol. 2003. Bethesda, MD: National Institutes of Health; 1999.
19. Keatinge W, Coleshaw S, Easton J, et al. Increased platelet and red cell counts, blood viscosity, and plasma cholesterol levels during heart stress, and mortality from coronary and cerebral thrombosis.
Am J Med. 1986;81:795–800.
20. Tyrell D, Barrow I, Arthur J. Local hyperthermia benefits natural and experimental common colds.
Br Med J. 1989;298:1280–1283.
21. Millqvist E, Bengtsson U, Bake B. Occurrence of breathing problems induced by cold climate in asthmatics-a questionnaire survey.
Eur Respir J. 1987;71.
22. Group TE. 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.
23. Donaldson G, Rintamaki H, Nayha S. Outdoor clothing: its relationship to geography, climate, behavior, and cold-related mortality in Europe.
Int J Biometeorol. 2001;45:45–51.
24. Wigley TM, Raper SC. Interpretation of high projections for global-mean warming.
Science. 2001;293:451–454.
25. IWG. Contribution of Working Group I to the Third Assessment Report of the Intergovernmental Panel on Climate Change.
Climate Change 2001: The Scientific Basis. Cambridge: Cambridge University Press; 2001.
