Published ahead of print on October 11, 2007, doi:10.1164/rccm.200706-823OC
© 2008 American Thoracic Society doi: 10.1164/rccm.200706-823OC
Mortality Displacement in the Association of Ozone with MortalityAn Analysis of 48 Cities in the United States1 Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts Correspondence and requests for reprints should be addressed to Antonella Zanobetti, Ph.D., Department of Environmental Health, Exposure Epidemiology and Risk Program, Harvard School of Public Health, 401 Park Drive, Landmark Center, Suite 415, P.O. Box 15698, Boston, MA 02215. E-mail: azanobet{at}hsph.harvard.edu
Rationale: Although the association between mortality and particles is well established, fewer studies have been reported with ozone. The harvesting hypothesis posits that the deaths associated with an exposure are occurring in people who are dying already, and the effect of exposure is merely to move the death from one day to an earlier day, and has no other effects that would influence deaths. Objectives: The aim of this study was to analyze the effect of ozone on mortality, and the extent to which this is due to short-term mortality displacement in 48 U.S. cities between 1989 and 2000. Methods: Time series of mortality and ozone were investigated with a generalized linear model during the June–August months, controlling for season, day of the week, and apparent temperature. We examined an unconstrained and a smooth distributed lag with 21 lags of ozone, and effect modification for city-specific characteristics. Measurements and Main Results: We found a 0.3% (95% confidence interval, 0.2–0.4) increase in total mortality for a 10-ppb increase in 8-hour ozone at lag 0 during summer months. The effect increased to 0.5% (95% confidence interval, 0.05–0.96) when looking at the unconstrained distributed lag. The shape of the distributed lag indicates that all the effect is in the first week. Conclusions: We did not find mortality displacement due to ozone; rather, the effect size estimate when looking at 21 days of ozone was larger than when using a single day's ozone concentration. Therefore, these results indicate that risk assessments using the single day of ozone exposure are likely to underestimate, rather than overestimate, the public health impact.
Key Words: ozone • mortality displacement • distributed lag
Airborne particles have been associated with daily mortality in cities around the world. Fewer studies have reported associations with ozone. One early, large multicity study was conducted in Europe in 1997 (1). More recently, three meta-analyses commissioned by the U.S. Environmental Protection Agency (EPA) were published together which reviewed the state of the ozone mortality literature (2–4). In addition, one of the articles also included a large multicity analysis of U.S. data (2). In general, these studies reported associations with ozone, but the associations seemed to be primarily restricted to the summer period. Another multicity study looking at the potential for temperature confounding in these associations did not find evidence of confounding but again found an association only in the summer (5). The EPA is currently evaluating how to use these studies in health risk assessments. One key issue assessing the public health implications of these associations is the question of whether the deaths associated with increased ozone levels are merely being brought forward by a few days or weeks, or whether a more substantial advancement in mortality is involved. The so-called harvesting hypothesis is that the deaths associated with an exposure, in this case ozone, are occurring in people who are, for the most part, dying already, and the effect of exposure is merely to move the death from one day to a day only slightly earlier. If that is true, then exposure on that earlier day (which produced the mortality advancement) should be negatively associated with mortality on the later day, when fewer people will die than otherwise, because some of the deaths have been shifted forward. This suggests a straightforward way to examine the hypothesis (6). However, it is also worth making explicit an implicit assumption of the hypothesis, which refers to a frailty model (7). As stated, it assumes that ozone increases the rate of dying of a frail group of people who are critically ill and would die soon anyway, but has no other effects that would influence deaths. However, the pool of people who are at high risk of dying is not static. People are constantly joining the pool. For example, after a myocardial infarction, people are at elevated risk of dying. But this example also illustrates another point. The excess mortality risk after a myocardial infarction or pneumonia admission is high, but declines quickly. That is, death is not the only way to leave the high-risk pool. People can recover. Therefore, ozone could influence the rate of death by increasing the death rate out of the risk pool, or by increasing the recruitment rate into the risk pool, or by delaying the recovery rate out of the risk pool. If it only influenced the first path, one would expect to see fewer people die after an ozone episode depleted the risk pool. But if it increased recruitment, or delayed recovery, one might see an increase in the size of the risk pool (e.g., by triggering a myocardial infarction), resulting in somewhat elevated numbers of deaths in the period after the episode, rather than fewer. If all three processes occur, the net impact would depend on the relative size of each effect, and the temporal pattern of increased deaths after exposure might be complex. Several investigators have analyzed the mortality displacement hypothesis for particles using multiple methodologies, including distributed lag models, frequency domain regression, and moving average approaches (6, 8–12). These approaches have produced generally consistent results. No previous study has analyzed the mortality displacement due to ozone. The aim of the current study was to analyze the effect of ozone on mortality, and the extent to which there was short-term mortality displacement, or enhanced delayed mortality suggestive of ozone effects on the other two transition rates for the risk pool, in a large, multicity study using an unconstrained distributed lag model. This consists of putting in a regression model predicting daily deaths the ozone concentrations on the day of death, the previous day, and so forth, up to a lag of 20 days. We chose this approach to mortality because it seemed most clearly connected to the insight above that short-term mortality displacement implies a negative association between ozone and mortality at some lag. Although studies of mortality displacement for particles have examined effects for up to a month or more after the exposure, we limited our analysis to examining effects for up to 3 weeks after exposure, because the ozone effect is limited to warm months. If we enter all 21 terms in the model, the estimated coefficients are too noisy to provide any information about the shape of the effect versus lag; therefore, we also examined the shape of the distributed lag by using a smooth distributed lag (9).
Study Population We obtained individual mortality data for 48 cities in the United States from the National Center for Health Statistics for 1989 to 2000. The mortality files provided information on the exact date of death, and the underlying cause of death. Our primary outcome was all-cause daily mortality; we also examined any cardiovascular, stroke, and any respiratory death.
Environmental Data We obtained local meteorological data (temperature and dew-point temperature) from the U.S. Surface Airways and Airways Solar Radiation hourly data (13), and we computed apparent temperature.
Statistical Methods The time series of daily counts of mortality and ozone were investigated with a generalized linear model, with a quasi Poisson link function to account for overdispersion. Such models generally control for season and other longer term patterns using natural splines with 4–8 degrees of freedom (df) per year; and, as we were examining quarter-years, we used 2-df splines to control for longer term trends in each summer. We used day-of-the-week indicator variables, and an unconstrained distributed lag for apparent temperature to take into account the effect of temperature today and the previous 7 days. We examined the dependence of daily deaths on 8-hour ozone concentrations on the day of death to use as a baseline analysis. We then refit models using ozone on the day of death, and up to the previous 20 days using an unconstrained distributed lag model. If the pollution-related deaths are only being advanced by a few days to a few weeks, we would see this "harvesting" effect expressed by a negative association between air pollution and deaths several days to weeks subsequently. The effect of air pollution, net of any such short-term rebound up to 20 days, is the sum of the positive and negative effect estimates for all 21 days. If this second estimate were smaller than the estimate using only lag 0 ozone, this would support the harvesting hypothesis. Several methods have been used to study the shape of the distributed lag. As previously described (9), we used a penalized quasi likelihood to estimated the coefficient of the smooth distributed lag. These models were fit using R statistical analysis software (R Project for Statistical Computing, http://www.r-project.org). In a second stage of the analysis, the city-specific results were combined using the meta-regression technique of Berkey and coworkers (14). We examined effect modification by city characteristics by entering them as predictor variables in the meta-regression. The results are expressed as percentage increase in deaths for a 10 ppb of 8-hour ozone concentrations. Additional detail on methods is provided in the online supplement.
The cities analyzed in this study are listed in the Tables E1 and E2 in the online supplement. The 48 cities chosen were distributed geographically across the United States; the smaller cities in terms of population were as follows: Terra Haute, Indiana; Boulder, Colorado; Provo/Orem, Utah; and Youngstown, Ohio. The biggest cities were Houston, Texas; Chicago, Illinois; New York, New York; and Los Angeles, California. Overall, we examined 1,614,124 deaths. Among the 48 cities, we have a wide range of ozone exposures and accompanying temperatures. For example, the 75th percentile of 8-hour ozone ranged from 19.8 ppb in Honolulu to 75.9 ppb in Los Angeles, whereas the 75th percentile for apparent temperature ranged from 20.6°C in Spokane to 34.6°C in Houston. City-specific tables (Tables E1 and E2) and box plots (Figures E1, E2, and E3) presenting the distributions of the environmental variables are reported in the online supplement. Table 1 presents the results of the association between mortality and ozone for the same day and the sum of the unconstrained distributed lag. We found a 0.3% (95% confidence interval [CI], 0.2–0.4) increase in total mortality for a 10-ppb increase in ozone at lag 0 during summer months. The effect increased to 0.5% (95% CI, 0.05–0.96) when looking at the unconstrained distributed lag.
We used the I2 statistic to assess the proportion of total variation in the effect estimates that was due to the between-cities heterogeneity (15) (see the online supplement), and we found that 4 and 21% of the total variability when using the lag 0 of ozone and sum of the distributed lag, respectively, were attributable to between-cities differences. Figure 1 presents the city-specific estimates for total mortality, together with the overall effect. The plot has been sorted by the amplitude of the confidence interval to highlight the heterogeneity of the data. The estimates are reasonably homogeneous across cities.
The cities with the larger confidence intervals tend to be the smallest cities, which are Boulder, Provo, Terra Haute, Honolulu, and Colorado Springs; on the other end, we have the biggest cities such as Los Angeles, New York, and Houston. When looking at cause-specific mortality, we found at lag 0 a 0.5% (95% CI, 0.3–0.6) increase in cardiovascular disease (CVD) mortality, a 0.5% (95% CI, 0.3–0.8) increase in respiratory mortality, and a 0.4% (95% CI, 0.01–0.7) increase in stroke mortality. Using an unconstrained distributed lag, the effect was a 0.5% (95% CI, –0.01 to 1.0) increase in cardiovascular mortality, a 0.6% (95% CI, –0.4 to 1.6) increase in respiratory mortality, and a 2.2% (95% CI, 0.7 to 3.7) increase in stroke mortality. We found little evidence of heterogeneity (small I2 statistic) and did not find any city characteristics to be significant modifiers of the risk (Table 2), but we found a trend toward a larger effect in cities with a lower percentage of central air conditioning of 0.69% (95% CI, 0.03 to 1.35) in cities in the lowest quartile, compared with 0.37% (95% CI, –0.2 to 0.95) in cities in the highest quartile of air conditioning.
Figure 2 presents the shape of the smooth distributed lag across the 48 cities. Each triangle represents the percentage increase in total mortality for a 10-ppb increase in ozone for each of the 21 lags. We found an immediate effect that decays during the first week, with the highest effect at lag 0, declining to zero by Day 6. We did not observe a negative effect due to mortality displacement; rather, the values between lags 7 and 17 are almost zero; the last lags have very high confidence intervals and we cannot extrapolate what the curve could look like afterward. The sum of the coefficients in the smoothed distributed lag was the same (0.51% increase per 10 ppb ozone) as in the unconstrained model.
We conducted several sensitivity analyses. To test the sensitivity to our use of apparent temperature instead of temperature and relative humidity to control for weather, we used the latter two variables, and found essentially identical results (results not shown). To test whether the lagged effect of ozone exposure really disappeared after lag 3, as indicated by Figure 2, we performed two sensitivity analyses. First, we used the unconstrained distributed lag model, as before. However, instead of summing all the coefficients from lag 0 to lag 20, and meta-analyzing that sum across cities, we created two sums of coefficients in each city: from lag 0 to lag 3, and from lag 4 to lag 20. We meta-analyzed both sums across cities. For all-cause mortality, we found an effect size of 0.53% (95% CI, 0.28 to 0.77) for lag 0–3 and –0.02% (95% CI, –0.35 to 0.3) for lag 4–20, which confirms the finding of no effect after lag 3 shown in Figure 2. For cause-specific deaths, we found insignificant suggestions of some harvesting for CVD and respiratory deaths, and a significant indication of a prolonged response for stroke. These results together with the cause-specific mortality results are reported in Table 1. In addition, we fitted a cubic polynomial distributed lag model instead of the penalized spline model, and obtained essentially an identical curve (results not shown).
Our analysis of the association of ozone with daily deaths produced results at lag 0 that were very similar to results from other studies. For example, the meta-analyses of Levy and colleagues (3) and Ito and coworkers (4) reported effect size estimates of 0.41 and 0.39%, respectively, for a 10-ppb increase in maximum hourly ozone, compared with our results, with an estimate of 0.31% for a 10-ppb increase in the 8-hour ozone. Therefore, there does not appear to be anything about our selection of cities, control for more lags of temperature, or restriction of analyses to 3 months that greatly disturbed the association from that previously reported. When we examined models that looked at 21 days of ozone simultaneously in an unconstrained distributed lag model, we found effect size estimates, for all causes and cause-specific mortality, that were somewhat larger than the effect size estimates using a single day's ozone concentration. If the ozone-related mortality represented short-term mortality displacement, we would have expected a decrease in effect size. Therefore, these results indicate that risk assessments using the single day of ozone exposure are likely to underestimate, rather than overestimate, the public health impact. Given the consistency of the findings of association of summertime ozone and daily deaths (1, 3, 4), the lack of confounding by particles (16), the lack of confounding by temperature (5), and the lack of confounding by more lags of temperature in this analysis, our failure to find evidence of short-term mortality displacement indicates that it would be reasonable to include mortality reduction as one of the health improvements associated with ozone reduction. The results for cause-specific mortality were intriguing, with insignificant, but suggestive, indications of some harvesting at lags 4–20 for heart and lung disease deaths, and a significant persistent association with stroke mortality at the longer lags. This may be due to sampling variability, with the all-cause mortality results being the most reliable. However, persons with different disease states may respond differently to ozone, with harvesting seen for some conditions, and persistent increases in risk for others. These findings should be replicated in other locations, and the mechanisms should be investigated. The shape of the distributed lag as shown by the plot of the smooth distributed lag, as well as the sensitivity analyses, indicates that for all-cause mortality, all the effect is in the first week. This suggests that the relevant ozone effects occur quickly, and do not persist for an extended time. This is consistent with a previous report examining hospital admissions for pneumonia and chronic obstructive pulmonary disease, which found the ozone effect on exacerbation of lung illness was seen within 2 days (17). Ozone is known to produce respiratory inflammation, and such inflammation may inhibit recovery from infection, or produce systemic responses. For example, a recent panel study of students living on a campus with an air monitor found that ozone was associated with increased levels of C-reactive protein, fibrinogen, 8-hydroxy-2'-deoxyguanosine, and plasminogen activator inhibitor 1, and decreased heart rate variability (18). A previous article concluded that particle and ozone exposure may decrease vagal tone, resulting in reduced heart rate variability (19); similarly, another study showed that exposures to PM2.5 (particulate matter with aerodynamic diameter less than 2.5 µm) and O3 are associated with decreased heart rate variability (20). A recent review of toxicologic studies concluded that when rats are exposed to typical concentrations of ozone, they demonstrate decreases in heart rate and other related indices, such as metabolism, minute ventilation, blood pressure, and cardiac output. The investigators concluded that, although there is only limited experimental evidence that addresses the underlying mechanisms of these responses, there is some indication that they may be related to stimulation of pulmonary irritant receptors and that they may be at least partially mediated via the parasympathetic nervous system (21). The finding of increased markers of systemic inflammation, thrombosis, oxidative stress, and impaired autonomic function support a plausible association with cardiovascular mortality. These associations have been reported with ozone exposures shortly before the measurement, which is consistent with the effects of ozone on mortality being predominantly within a few days of exposure. We did not see the overall effect (summed over the 21 d) of ozone modified by any of the city characteristics we examined. This could be explained by the fact that the estimates were reasonably homogeneous across cities (I2 < 0.21). Therefore, there was little variation across cities that could be explained by city-level covariates. However, this finding is interesting by itself, because similar multicity studies of particulate air pollution have reported larger heterogeneity (16, 22), and reported associations with some city characteristics. More extensive examination of individual-level factors that modify risk is clearly needed. There are limitations to this analysis. Foremost is that time series analyses, by their nature, cannot provide much information on how long the ozone-associated deaths are advanced, given that the displacement clearly averages more than 20 days. However, it is worth noting that persons with myocardial infarctions, pneumonia, and many of the other conditions contributing to the ozone–mortality association generally have elevated short-term risks of mortality that decline rapidly if they survive for 3 weeks. The ozone-attributable deaths are therefore likely to mostly be advanced by a substantial period. There is evidence of adaptation to ozone, with, for example, reduced effects of ozone on lung function (23–26). We have been unable to address this in our study, but it is an important question to determine whether this holds for the mortality associations as well. We did not control for PM10 because we believe there is evidence demonstrating that it is not a confounder (4, 5, 27), and the use of it as a covariate would preclude an analysis of distributed lags of ozone, because in most cities PM10 is only measured 1 day in 6. Although PM10 is not a confounder of the ozone–mortality associations, sulfate particles are more tightly correlated with ozone, as they are formed by similar photochemical processes. We were unable to control for this potential confounder, because the EPA does not routinely measure it. The same holds for other photochemical oxidants. However, ozone–mortality associations have been reported in both the eastern United States, where sulfate levels are high, and the West Coast, where they are low, suggesting that these observations are not likely to be entirely due to sulfate confounding. We believe the association of ozone with daily deaths in the summer is well established, and, as it does not represent merely short-term mortality displacement, is an issue of public health concern.
This article has an online supplement, which is accessible from this issue's table of contents at www.atsjournals.org Originally Published in Press as DOI: 10.1164/rccm.200706-823OC on October 11, 2007 Conflict of Interest Statement: Neither author has a financial relationship with a commercial entity that has an interest in the subject of this manuscript. Received in original form June 5, 2007; accepted in final form October 11, 2007
This article has been cited by other articles:
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
TOP
ABSTRACT
AT A GLANCE COMMENTARY
