Airborne particles have been associated with increased hospital admissions for respiratory illness 1,2 and for cardiovascular illness 3–5 in a wide range of locations, as well as in situations with low exposure to sulfur dioxide and ozone. Similar associations have been reported for daily deaths. 6 These associations have led to a general acceptance that these associations are causal. 7,8 It has been suggested, however, that these events may not be displaced by a long period of time. 9 Rather, they are only being brought forward by a few days to a few weeks in persons who are likely to be hospitalized or to die in a short period of time. This pattern is sometimes termed harvesting. This paper discusses a framework for thinking about such a process and an analytical approach to test the hypothesis that most of the particulate-related events are only being advanced by a short period of time. This approach is then applied to data on respirable particles (particulate matter less than 10 μm in aerodynamic diameter; PM10) and daily deaths and hospital admissions for heart and lung disease in Chicago.
Suppose that the population can be divided into two pools, the low-risk and the high-risk pool. On each day, persons in the high-risk pool can transition out of it to hospital or death, can recover and transition to the low-risk pool, or can remain in the pool. In addition, every day, some people are recruited into the pool. This scheme is illustrated in Figure 1 for mortality, in which the transitions are labeled T1, T2, and T3. Under steady-state conditions, there is an equilibrium between the number of deaths each day (T3) and the daily net recruitment into the high-risk pool (T1–T2). Net recruitment is the difference between the number of persons entering the high-risk pool (T1) and those who revert to the relatively healthy pool (T2).
Any short-term increase in the mortality rate in the high-risk pool (T3) that is not accompanied by a counterbalancing increase in the net recruitment rate (T1–T2) would deplete the high-risk pool to a level below its equilibrium value. This depletion could occur if a transient environmental factor, such as a heat wave or pollution episode, increased mortality in the high-risk pool but had little adverse effect on healthy subjects.
On subsequent days during such an episode, or in response to another stressor, fewer people would die, because the mortality rate produced by the stressor would be applied to a smaller at-risk pool. This lower mortality would continue until the high-risk pool was replenished through net recruitment. As a result, the daily death count would be diminished on subsequent days. Figure 2 shows an example of a possible time course of daily deaths over time after an episode, with the initial increase followed by the subsequent reduction. If all of the increase is due to short-term harvesting, then the areas (number of deaths) above the baseline (the long-term mean) should equal the area (number of deaths) below the baseline.
This conceptualization makes clear an important point that has been little discussed previously. If the stressor increases net recruitment into the risk pool by, for example, increasing the rate or severity of serious respiratory or cardiovascular illness, the harvesting effect would be diminished and indeed the size of the risk pool could even increase. Unfortunately, without additional information, the individual transition probabilities, T1, T2, and T3, are unidentifiable. Rather, we can only estimate the net effect of air pollution as a result of its effect on all three. Nevertheless, the conceptualization is important in understanding that the effect of air pollution on the size of the risk pool may be complex and may vary over time.
If particulate air pollution is only advancing the date of adverse events by a few days, it is a much less serious public health concern. That would occur if the effects of particulate air pollution are restricted to a subset of the susceptible pool in which the probability of transition to the healthy status is essentially zero and airborne particles had little impact on the probability of entering the risk pool. This hypothesis is readily testable, and further, it is possible to estimate the number of events that are either avoided or displaced by more substantial periods.
Recently, two papers 10,11 have proposed conceptually similar approaches to assessing the effect of air pollution net of short-term harvesting, which differ in detail and implementation. Both are based on the observation that the harvesting hypothesis assumes that the rebound effect of lower-than-expected mortality shown in Figure 2 occurs relatively soon after the initial increase. If not, then mortality or morbidity is being displaced by a more substantial amount. The two approaches localize the harvesting period in time 11 or frequency 10 and examine the association between daily death and air pollution outside of those time scales or frequency ranges. Each has examined mortality in one city and reported that at longer time scales the effect sizes for particulate air pollution increase rather than decrease. To date, no one has examined harvesting in the association between air pollution and hospital admissions. I here apply the time-domain approach to a different city and to both hospital admissions and daily deaths.
An indirect approach to examining harvesting, which is also of interest for other reasons, is to look separately at death occurring inside and outside of hospitals. If the deaths caused by particulate air pollution are primarily in desperately ill subjects who will die within a few days in any event, those subjects are much more likely to die in the hospital than persons whose condition was not as critical and who might have well recovered. Hence, looking at the relative impact of air pollution on deaths by location of death is an indirect test of how much of the air pollution-related deaths are only being brought forward by a few days.
Subjects and Methods
I computed daily death counts for the years 1988–1993 for Cook County, IL, using the Detail Mortality Tapes of the National Center for Health Statistics. In addition to all-cause mortality, I computed separate counts of deaths outside of the hospital and deaths of hospital patients. Similarly, I compiled daily counts of Medicare hospital admissions of persons aged 65 and older for heart disease [International Classification of Diseases, 9th revision (ICD-9) codes 390–429], pneumonia (ICD-9 480—487), and chronic obstructive pulmonary disease (COPD; ICD-9 490—496, except 493) using MEDPAR files obtained from the Health Care Financing Administration. Weather data were obtained from the weather station at O’Hare Airport and air pollution data from the Environmental Protection Agency’s Aerometric Information Retrieval System (AIRS) database.
The computation of PM10 exposure in Chicago was complicated by the differing monitoring schedules for different monitors. Some monitors operate on a daily basis, with the others operating every sixth day. If the monitors were simply averaged, the daily mean would jump on days when new monitors were included, merely because their annual averages differs from the monitoring stations that operate on a daily basis.
The variance of PM10 measurements also can differ from monitoring location to monitoring location. Day-to-day changes in the set of monitors included in the daily average would also result in changes in the day-to-day variation in the exposure measure that do not represent true changes in exposure but only changes in the sampling of monitors. To remove these influences, I used the following algorithm. The annual mean was computed for each monitor for each year and subtracted from the daily values of that monitor. I then standardized these daily deviances from each monitor’s annual average by dividing by the standard deviation for that monitor. The daily standardized deviations for each monitor on each day were averaged, producing a daily averaged standardized deviation. I multiplied this by the standard deviation of the deviations of all of the monitor readings for the entire year and added back in the annual average of all of the monitors. The mean of the PM10 on the concurrent and previous day was used as the exposure index.
Long-term trends and seasonal variations represent the patterns that have traditionally been filtered out of the data to avoid confounding by omitted factors (such as smoking or respiratory epidemics), which may have long-term trends or seasonal patterns that coincide with those of air pollution, but whose short-term fluctuations are unlikely to be correlated with air pollution. The remaining pattern has variations on the scale of a few months to a few days.
If the effect of air pollution is only to move a death (or hospital admission) forward by a few days, that would contribute to the regression coefficient in a standard analysis. On the other hand, if we constructed 30-day moving averages of air pollution and deaths, movements of deaths by a few days would not be expected to contribute much to the association between these two moving averages, because moving the death by a few days within a 30-day window would not change the 30-day moving average very much. This analysis takes advantage of this fact. For example, Figure 3 shows a simple case of harvesting, in which there is an effect followed by an immediate rebound. The circles represent the raw data and the solid line, is a 15-day smooth through the data. The harvesting has clearly been netted out. My approach takes advantage of this effect to estimate results net of any such short-term rebounds. Unfortunately, it can only estimate the net effect and does not estimate T1, T2, and T3 directly, as there is insufficient information to identify those three subcomponents.
My baseline analysis regressed the counts of deaths or admissions on PM10 controlling for time trends, day-of-week effects, and weather (temperature and relative humidity). The regression was done using a generalized additive model 12 to allow for potentially nonlinear effects, which were modeled using loess smooth functions of the variables (except PM10). The span for the smooth function of time was chosen to remove seasonal and long-term (for example, multiyear) trends and to minimize autocorrelation in the residuals. The methods have been described previously. 5 For example, Figure 4 shows the residuals from the regression of daily deaths, demonstrating that seasonality has been removed. The spans for the other terms were chosen to minimize Akaike’s information criterion. This approach has been described previously in more detail. 5 Because most studies have found a 2-day mean of PM10 more predictive than a single day’s exposure, the exposure variable was taken to be the mean of PM10 on the concurrent and previous day.
The baseline analysis provides an estimate of the effect of PM10 on deaths and hospital admissions that ignores the issue of harvesting or changes in the size of the susceptible pool.
I used STL, the seasonal and trend decomposition program introduced by Cleveland et al, 13 to decompose my mortality and hospital admission data into different time scales. STL iteratively fits loess 14 smooths to a time series using different smoothing windows. This method decomposes a single time series (for example, daily deaths) into a series of independent time series representing the daily fluctuations in that series that are due to patterns with different time scales. A loess smooth is a generalization of a moving average that uses weights that decrease to zero at the edges of the window (number of days that are averaged, also called the span) to reduce distortion. Further details about loess smoothing have been published earlier. 15 I chose the first window (120 days) to remove long-wavelength and seasonal patterns. The residuals from this filtering have no seasonality (for example, Figure 5, which shows the results for daily deaths). A similar filtering process was applied to the PM10 and weather data. These are my seasonally detrended series.
The next set of regressions examined the same outcomes net of short-term harvesting. This analysis was done by using a 15-day window smooth of the seasonally detrended deaths and hospital admissions as the outcomes, as well as similarly detrended predictor variables. These variables have the long-term and seasonal patterns removed but are not sensitive to short-term displacement of the deaths or hospital admissions. This approach has been described in more detail elsewhere. 11 Again, a generalized additive model was fit, modeling the detrended hospital and death data as smooth functions of the detrended weather variables and as a linear function of the detrended PM10. If an association remains with PM10 in this analysis, then the previous associations cannot be due entirely to very short-term harvesting. Moreover, the difference in the effect-size estimate between the two regressions provides an estimate of how much of the effect is harvesting vs how much represents mortality or morbidity events avoided or postponed by more substantial amounts of time.
To examine harvesting on different time scales, the detrended data were separately smoothed and the above analysis repeated, using smoothing windows of 30, 45, and 60 days instead of 15 days. These examine the association between PM10 and outcome net of progressively longer displacement of the events.
Table 1 shows the distribution of the data in this study. Table 2 shows the baseline estimates of the effect of PM10 on all-cause mortality and deaths inside and outside of the hospital. The effect-size estimate for deaths outside of the hospital is larger than for deaths inside the hospital. The effects on hospital admissions for heart disease, COPD, and pneumonia are also shown. PM10 is associated with increased risk of all of the outcomes. Figure 6 shows what happens to the effect-size estimates for deaths as we exclude short-term fluctuations on progressively longer time scales. All-cause mortality shows an increase in effect size at longer time scales. There is a considerable difference when the deaths are divided by the location of occurrence. The effect of PM10 on deaths inside of hospitals shows little trend as we move to longer time scales, whereas deaths outside of the hospital increase more steeply with increasing time scale.
The results for hospital admissions (Figure 7) also show no evidence that most of the effect is short-term harvesting. For heart disease and pneumonia admissions, the effect-size estimates are essentially stable at all time scales, whereas for COPD admissions they roughly triple at the longer time scales.
These mortality results confirm those previously reported in Boston and Philadelphia for mortality. When short-term rebound effects are averaged out, the effect size for particulate air pollution increases rather than decreases. Moreover, this finding is supported by the alternative approach of examining deaths inside and outside of the hospital separately. Deaths in the hospital, where short-term harvesting is most likely, are less sensitive to air pollution than are deaths outside of the hospital.
In the most simplistic harvesting model, in which airborne particles can only deplete the pool of susceptibles, these results would be inexplicable. As illustrated in Figure 1, however, there are transitions (T1 and T2) into and out of the risk pool that do not involve death. If air pollution were to increase the net recruitment rate by more than the death rate, these results would follow. Hence, these results indirectly support such a conclusion and suggest that models that fail to incorporate T2, or the potential for air pollution to affect T1 and T2, will be inadequate. How plausible is it that such phenomena are happening?
Air pollution could increase the size of the risk pool or, viewed continuously, shift the susceptibility of the population to the right if it, for example, increased the intensity of infectious illness. Air pollution has been associated with increases in physician visits 16 and hospital admissions 17 for infectious respiratory illness and with increased school absences in children. 18
In animal studies, Zelikoff and colleagues 19 have reported that exposure of rats to modest concentrations of ambient particles 48 hours after infection with Streptococcus pneumoniae more than doubled both bacterial burden and the area of the lung with pneumonia. S. pneumoniae is the most common organism associated with human pneumonia. Exposure to particulate matter has also been shown to exacerbate influenza infections in animal models. 20 Such infections may matter for more than pneumonia deaths. An analysis in Philadelphia reported that cardiovascular deaths on high-air pollution days were much more likely to have respiratory contributing causes than cardiovascular deaths on low-air pollution days, 21 and COPD deaths are often due to an infection.
Recent epidemiologic studies have reported that airborne particles are associated with changes in heart rate variability that are known risk factors for arrhythmias and death. 22–24 PM2.5 exposure has also been associated with defibrillator discharges in subjects with implanted defibrillators. 25 Instillation of combustion particles into the lungs of rats has been shown to induce arrhythmia and death. 26 Dogs exposed to particles from ambient air have shown changes in electrocardiogram patterns as well. 27 Hence, there is a biological basis for believing that airborne particle exposure increases the risk of sudden death, which may then be triggered by other factors. That is, it may increase the risk pool.
Other recent studies support the hypothesis that particulate exposure has systemic effects on humans and animals. Exposure of animals to concentrated air particles has been shown to increase peripheral neutrophil counts, a cellular marker of inflammation. 28 Recently, Salvi and co-workers 29 exposed human subjects to 300 μg/m3 of diesel particles for 1 hour and reported increases in neutrophils, not just in the lung, but peripherally. Peters and co-workers 30 have reported that plasma coagulability increased with increased exposure to airborne particles. These reports all suggest that particle exposure can lead to systemic upregulation of processes that increase the risk of adverse events. When this change occurs in a population with a distribution of pre-existing other conditions and with constant stimulation by other stressors, it is reasonable to assume that it may temporarily increase the pool of people susceptible to death or more serious illness.
Air pollution may also reduce the transition out of the risk pool to the healthier population. For example, it has been reported to increase the duration of respiratory illnesses. 31
The analysis of deaths occurring inside and outside of hospitals is informative in two ways. First, the baseline results themselves speak to the harvesting issue. Persons who are lingering on the edge of death, who will die in a few days if they do not die today, are disproportionately in the hospital today. If most air pollution-related deaths were short–term harvesting, we would expect a greater impact on this population. In fact, airborne particles had a greater impact on deaths outside of the hospital in Chicago. This effect is similar to results previously reported in Philadelphia 20 and in London. 32 Moreover, the increase in effect size seen as we move to longer time scales was seen only for the out-of-hospital deaths. This result is consistent with particle exposure increasing the risk of sudden death (due to other triggers) in ambulatory subjects with apparently good health but underlying heart disease.
The results for hospital admissions also showed no evidence of net harvesting. For COPD admissions, there was evidence of a larger effect at longer time scales.
This analysis is consistent with those of the two previous studies 10,11 showing that the effect of particulate air pollution is not primarily the shifting of adverse events by a few days or weeks. This consistency is reassuring, because although both approaches (filtering in the time and frequency domain to remove short-term rebounds) are similar, they are quantitatively different. In particular, when a time series is subjected to a Fourier transformation, information about the shape of the 365-day seasonal pattern (unless it is perfectly sinusoidal) is shifted into other, higher, frequencies. Hence, the control for seasonality is not the same in the two approaches. Similarly, I used smoothing (see Figure 3) to remove short-term rebounds from the mortality data, whereas Zeger and co-workers 10 eliminated all sinusoidal components whose wavelength was short. These approaches do not do the identical job in removing short-term rebound patterns. Hence, these differences provide some indication that the major conclusions, that the effect sizes are not reduced, are robust. Indeed, all three studies suggest that the effect-size estimates from the existing time-series literature are smaller than the effect-size estimates that are net of short-term fluctuations in the size of the risk pool.
Although these estimates indicate that the deaths and hospital admissions are on average being brought forward by a nontrivial amount of time, they cannot tell us precisely what that time is. In particular, both this methodology and that of Zeger and co-workers 10 cannot examine mortality displacement past 3–4 months because of the necessity to control for season. This limitation is a major drawback of the time-series approach. What the two methods do show is that out to 3 or 4 months, there is no diminution of the effect.
1. Schwartz J. Air pollution and hospital admissions for the elderly in Birmingham, Alabama. Am J Epidemiol 1994; 139: 589–598.
2. Sunyer J, Saez M, Murillo C, Castelsaque J, Martínez F, Anto JM. Air pollution and emergency room admissions for chronic obstructive pulmonary disease: a 5-year study. Am J Epidemiol 1993; 137: 701–05.
3. Schwartz J, Morris R. Air pollution and hospital admissions for cardiovascular disease in Detroit, Michigan. Am J Epidemiol 1995; 142: 23–35.
4. Burnett RT, Dales R, Krewski D, Vincent R, Dann T, Brook JF. Associations between ambient particulate sulfate and admissions to Ontario hospitals for cardiac and respiratory diseases. Am J Epidemiol 1995; 142: 15–22.
5. Schwartz J. Air pollution and hospital admissions for heart disease in eight U.S. counties. Epidemiology 1999; 10: 17–22.
6. Pope CA, Dockery DW, Schwartz J. Review of epidemiologic evidence of health effects of particulate air pollution. Inhal Toxicol 1995; 7: 1–18.
7. Committee on the Medical Effects of Air Pollutants, Department of Health. Health Effects of Non-Biological Particles. London: Her Majesty’s Stationery Office, 1995.
8. U.S. Environmental Protection Agency. Air Quality Criteria for Particulate Matter. EPA/600/AP-95/001b. Research Triangle Park, NC: U.S. Environmental Protection Agency, 1995.
9. McMichael AJ, Anderson HR, Brunekreef B, Cohen AJ. Inappropriate use of daily mortality analyses to estimate longer-term mortality effects of air pollution. Int J Epidemiol 1998; 27: 450–453.
10. Zeger SL, Dominici F, Samet J. Harvesting-resistant estimates of air pollution effects on mortality. Epidemiology 1999; 10: 171–175.
11. Schwartz J. Harvesting and long-term exposure effects in the relationship between air pollution and mortality. Am J Epidemiol 2000; 151: 440–8.
12. Hastie T, Tibshirani R. Generalized additive models. London: Chapman and Hall, 1990.
13. Cleveland RB, Cleveland WS, McRae JE, Terpening I. Seasonal-trend decomposition procedure based on LOESS. J Offic Stat 1990; 6: 3–73.
14. Cleveland WS, Devlin SJ. Robust locally-weighted regression and smoothing scatterplots. J Am Stat Assoc 1988; 74: 829–836.
15. Schwartz J. Short term fluctuations in air pollution and hospital admissions of the elderly for respiratory disease. Thorax 1995; 50: 531–538.
16. Schwartz J, Spix C, Wichmann HE, Malin E. Air pollution and acute respiratory illness in five German communities. Environ Res 1991; 56: 1–14.
17. Pope CA III. Respiratory disease associated with community air pollution and a steel mill, Utah Valley. Am J Public Health 1989; 79: 623–628.
18. Ransom MR, Pope CA III. Elementary school absences and PM10
pollution in Utah Valley. Environ Res 1992; 58: 204–219.
19. Zelikoff JT, Nadziejko C, Fang T, Gordon T, Premdass C, Cohen MD. Short term low-dose inhalation of ambient particulate matter exacerbates ongoing pneumonococcal infections in streptococcus pneumoniae-infected rats. In: Phalen KF, Bell YM, eds. Proceedings of the Third Colloquium on Particulate Air Pollution and Human Health. Air Pollution Health Effects Laboratory, University of California, Irvine, CA, vol 8, pp 94–101.
20. Clarke RW, Hemenway DR, Frank R, Kleeberger SR, Lonohre MV, Jakab GJ. Particle associated sulfate exposure enhances murine influenza mortality. Am J Respir Crit Care Med 1997; 155: 245.
21. Schwartz J. What are people dying of on high air pollution days? Environ Res 1994; 64: 26–35.
22. Liao D, Creason J, Shy C, Williams R, Watts R, Zweidinger R. Daily variation of particulate air pollution and poor cardiac autonomic control in the elderly. Environ Health Perspect 1999; 107: 521–525.
23. Gold DR, Litonjua A, Schwartz J, Verrier M, Milstein R, Larson A, Lovett E, Verrier R. Cardiovascular vulnerability to particulate pollution. Am J Respir Crit Care Med 1998; 157: A261.
24. Pope CA III, Verrier RL, Lovett EG, Larson AC, Raizenne ME, Kanner RE, Schwartz J, Villegas GM, Dockery DW. Heart rate variability associated with particulate air pollution. Am Heart J 1999; 138 (5 Pt 1):890–899.
25. Peters A, Liu E, Verrier RL, Schwartz J, Gold DR, Mittleman M, Baliff J, Oh A, Allen G, Monahan K, Dockery DW. Air pollution and incidences of cardiac arrhythmia. Epidemiology 2000; 11: 11–17.
26. Watkinson WP, Campen MJ, Costa DL. Cardiac arrhythmia induction after exposure to residual oil fly ash particles in a rodent model of pulmonary hypertension. Toxicol Sci 1998; 41: 209–216.
27. Godleski JJ, Lovett EG, Sioutas C, Killingsworth CR, Krishnamurthi GG, Hatch V, Wolfsom M, Ferguson St, Koutrakis P, Verrier RL. Impact of inhaled concentrated ambient air particles on canine electrocardiographic patterns. Health Effects Institute Annual Meeting, Health Effects Institute, Cambridge, MA, 1997;15.
28. Gardner SY, Costa DL. Particle-induced elevations in white blood cell count and plasma fibrinogen levels in rats. Am J Respir Crit Care Med 1998; 157: A152.
29. Salvi S, Blomberg A, Rudell B, Kelly F, Sandstrom T, Holgate ST, Frew A. Acute inflammatory responses in the airways and peripheral blood after short-term exposure to diesel exhaust in healthy human volunteers. Am J Respir Crit Care Med 1999; 159: 702–709.
30. Peters A, Doring A, Wichmann HE, Koenig W. Increased plasma viscosity during an air pollution episode: a link to mortality? Lancet 1997; 349: 1582–1587.
31. Schwartz J. Air pollution and the duration of acute respiratory symptoms. Arch Environ Health 1992; 42: 116–122.
32. Her Majesty’s Public Health Service. Mortality and Morbidity during the London Fog of December 1952. Report No. 95 on Public Health and Medical Subjects. London: Her Majesty’s Stationery Office, 1954.
© 2001 Lippincott Williams & Wilkins, Inc.