Funding: supported by the Harvard Environmental Protection Agency (EPA) PM Center, Grant R827353, and National Institute for Environmental Health Science (NIEHS) ES0002
Competing interests: none declared
Statistics from Altmetric.com
- AT, apparent temperature
- COPD, chronic obstructive pulmonary disease
- HF, heart failure
- ICD, International Classification of Diseases
- IHD, ischaemic heart disease
- MI, myocardial infarction
- NMMAPS, National Morbidity and Mortality Air Pollution Study
- NO2, nitrogen dioxide
- PM10, particulate matter of aero-diameter equal or less than 10 μm
Short term increases in concentrations of airborne particles, complex mixtures of solid and liquid droplets, have been associated with immediate increases in mortality events.1–4 The largest study of air pollution, the National Morbidity and Mortality Study (NMMAPS), examined this association in all the cities in the United States.5 The important evidence coming from this study confirmed previous findings on the positive association between particulate matter of aero-diameter equal or less then 10 μm (PM10) and daily deaths, and proved that such relation was little changed when controlling for any gaseous pollutant. They also confirmed that none of the gaseous pollutants had an independent effect on the daily mortality. However, because the NMMAPS included cities that only monitored PM10 one day in six, it could only examine lags one day at a time, thus it could not address whether the mortality association persisted across several days.
Most of the evidence for the positive effects of PM10 has been generally shown in association with all-cause mortality,6–8 or mortality from all respiratory and/or cardiovascular disease.9–12 Such a broad definition of cause of death can possibly underestimate the magnitude of the association that one would observe had a more specific classification of diseases been used. Only a few studies to date have examined the acute effects of particulate matter and cause specific daily mortality,13–17 and fewer were multi-city studies.
Differing results with the more specific definition of mortality causes may aid our understanding of the possible mechanisms that induce the PM10 related events. Particles are thought to be involved in complex inflammatory responses in animals and humans, which could lead to cardiovascular and/or pulmonary effects, and possibly to cardiovascular and/or pulmonary events.18–24 Experimental studies have shown that exposures to concentrated air particles induce pulmonary inflammatory response, such as alveolar cytokine release, responsible for bronchospasm, and reduction of pulmonary function.25–27 The mechanisms leading to cardiovascular effects have been suggested to be a result of increasing blood clotting factors such as increased serum levels of fibrinogen,25,26 increased inflammatory factors like C-reactive protein,22 or cardiotoxicity by mechanisms such as the increase of plaque formation and decrease of plaque stability,28 vasoconstriction and temporary coronary occlusion,29,30 and ischaemia.28 Increased reactive oxygen species in the lung and heart have been produced by exposure to concentrated air particles as well.31 Epidemiological studies have suggested similar mechanisms. Increased concentrations of particulate matter have been associated with plasma viscosity18 and fibrinogen32 which are suspected risk factors for cardiovascular disease. Results for C-reactive protein concentrations have been mixed.22,33,34
In the present study we examined the acute effects of particulate matter on all-cause and cause specific daily mortality for a list of diseases. For this, we obtained detailed mortality files for a 12 year period between 1989 and 2000. We used publicly available sources to obtain air pollution data and weather data for the same time period. A case-crossover design analysis was used to evaluate the acute effects of daily PM10 concentrations on mortality. Short term effects of particulate matter on daily mortality were examined using lag models. Several studies have examined factors that may explain heterogeneity in the effects of PM10 by location (city).8,35,36 We examined the potential for effect modification of the PM10—daily mortality estimates by city characteristics, specifically the prevalence of central air conditioning, population density, standardised mortality rates, the proportion of elderly in each city, daily minimum apparent temperature in summer (mean and variance), daily maximum apparent temperature in winter (mean and variance), and percentage of primary PM10 from traffic sources.
The study included a period of 12 years between 1 January 1989 and 31 December 2000. We selected 20 cities in the United States with sufficient mortality and daily air pollution data: Birmingham, Alabama; Boulder, Colorado; Canton, Ohio; Chicago, Illinois; Cincinnati, Ohio; Cleveland, Ohio; Colorado Springs, Colorado; Columbus, Ohio; Denver, Colorado; Detroit, Michigan; Honolulu, Hawaii; Minneapolis, Minnesota; Nashville, Tennessee; New Haven, Connecticut; Pittsburgh, Pennsylvania; Provo, Utah; Salt Lake City, Utah; Seattle, Washington; Terra Haute, Indiana; and Youngstown, Ohio.
Mortality files were obtained from the National Center for Health Statistics (NCHS) for the period between 1989 and 2000. The mortality files provided information on the exact date of death, and the underlying cause of death. For this study we selected all-cause daily mortality excluding any deaths from accidental causes (ICD code, 10th revision V01–Y98). Specific cause mortality case groups were derived from this previous group for the following underlying causes of death: heart disease (I01–I51), ischaemic heart disease (IHD) (I20–I25), myocardial infarction (MI) (I21, I22), dysrhythmias (I46–I49), heart failure (HF) (I50), stroke (I60–I69), respiratory disease (J00–J99), pneumonia (J12–J18), and chronic obstructive pulmonary disease (COPD) (J40–J44, J47).
Data for PM10 were obtained from the United States Environmental Protection Agency Aerometric Information Retrieval System (US EPA AIRS) for the same period of years.37 US EPA AIRS retrieves data from multiple monitors in each county for daily and hourly average measurements. The sampling method for hourly data uses monitors which are susceptible to a greater loss of semi-volatile particles from traffic, than the integrated 24 hour (daily) monitors. The hourly monitors heat the sample to 50°C to remove water from the mass, and in doing so it volatilises the semi-volatile PM components. Even more complicated is the fact that the amount of semi-volatile mass for a given sample in the hourly measurements will likely vary by location, season, and even time of day, thus requiring constant adjustments by a reference method. Therefore, the integrated method provided the highest quality data for our analysis.38,39
Air pollution data from multiple monitors in each city were used to estimate a common value for the city, using an averaging method based on an algorithm described previously by Schwartz.35 However, before applying this algorithm, we made sure that all monitors in one city represented general ambient exposures, and not rather a local source. For this, we calculated the correlation between city monitors. We obtained multiple correlation coefficients for each monitor (correlated with all other monitors in the city), from which we extracted the median values. Those monitors falling in the low 10th centile of the distribution of the median values, across all cities, were excluded from the analyses.
While we selected cities with daily PM10 monitoring, daily concentrations of PM10 were occasionally missing. Even small numbers of missing data can present a problem when examining multiple lags of exposure in a model. We used regression models to predict missing PM10 daily values using local meteorological data obtained from the United States Surface Airways and Airways Solar Radiation hourly data (NESDIS) (including extinction coefficient,40 a measure of light scattering by fine particles), and the PM10 concentration the day before and the day after. Prior to filling in the missing values, we made sure that we used only months missing less than 25% of the days, and PM10 daily values equal or less than 300 μg/m3.
Statistical methods of analysis
A case-crossover design was carried out to examine the association between daily mortality and PM10. The approach, first described by Maclure,41 is similar in concept to a case-control study, except that now cases and controls are the same subject, but in different times. A case is defined by the time of the event (day of the event in our case), and controls are chosen as the times (days) free of the event, selected to be close to the event day in time. Bateson and Schwartz showed that such matching can control even strong confounding by season and time trends.42,43 We chose these control days by using a time stratified approach.44 This involves choosing control days to be all other days (or a sample of all other days) of a fixed time strata, typically the same month of the same year as the event. Using monthly strata has been shown to produce unbiased estimates of effect sizes and coverage probabilities even in the presence of strong seasonal confounding.45 Control days have different exposures from that of the case, since exposures change from day to day. As in a matched case-control study, conditional logistic regression is applied to the matched pairs to compare different characteristics between the case day (day of the event) and its control days (days free of the event), including exposure characteristics. Because of its design, the case-crossover approach controls for factors that change slowly over time,42,43 or factors that are individual characteristics which may, in another case, be potential confounders.41 In our study, controls were selected within the month of the event, leaving two days between each control day to eliminate any serial correlation.
The statistical modelling involved two hierarchical stages. Effects of PM10 on daily mortality were first estimated through city specific models, in the first stage analyses. Risk of mortality was evaluated in association with PM10 exposure concentrations on the same day of the event (lag 0), one (lag 1), and two days (lag 2) prior the event. The models included one lag at a time (single lag models), or all three lags. The latter is described as a “distributed lag model” and is a way to estimate unbiased time trends of the exposure-response relation, by controlling for confounding in exposure effects in one lag by exposures in other lags.46 All models included “day of week” as indicator variables, and quadratic spline functions for apparent temperature (AT) on the same day and the day before death, to control for effect of weather,47,48 in the form: ATij + ATij2 + (ATij − city specific mean ATij)+2, where i represents city, and j represents lag of AT. Given the humidity, AT is defined as the perceived body temperature, and was calculated as: AT = −2.653 + (0.994*Ta) + (0.0153*Td2), where Ta is air temperature and Td is dew point temperature.49,50
Another way to control for effect of whether, is to match on AT.51 We carried out additional analyses, where control days were matched on AT (on the same day, rounded in °C), month, and year as the event day. The models included additionally indicator variables for “day-of-week”, and a quadratic spline function for AT one day before the event.
Estimates obtained from the distributed lag models were used to estimate the cumulative effects in each city, computed by summing the coefficients across all three days, that of the event, and two days before.46 The overall variance was computed as:
is the city-specific lag-specific slope, i represents city, and j = 0, 1, 2 defines lag.
In the second stage analyses, maximum likelihood meta-regression models estimated summary effects for mortality causes and PM10 associations across the 20 cities, allowing for heterogeneity in city specific response using a random effect model.52
We examined the potential for effect modification of the cumulative PM10-mortality estimates by city characteristics including the prevalence of central air conditioning, population density, standardised mortality rates, the proportion of elderly in each city (greater than 65, 75, or 85 years old), daily minimum apparent temperature (mean and variance) in summer (June, July, August), daily maximum apparent temperature (mean and variance) in winter (December, January, February), and percentage of primary PM10 from traffic sources. These factors were examined in the second stage random meta-regressions.8,35,36
The estimates of PM10-mortality association reported in this paper refer to percentage increase in mortality per any 10 μg/m3 increment of PM10 concentration. SAS statistical software package was used for all data management, and statistical analyses of the first stage. S-plus was used to implement the Berkey method for random meta-regression analyses.52
The 20 selected cities showed variability in the city-specific characteristics (table 1). In several cities, the higher population density corresponded well with higher percentage of primary PM10 from traffic sources, suggesting highly urbanised areas. Some cities had relatively high percentages of central air conditioning including Birmingham, Cincinnati, Columbus, Nashville, and Terra Haute, which was also consistent with higher summer AT. Overall mean concentration of PM10 varied by city and across time within city. Birmingham, Cleveland, Provo, and Salt Lake City had higher and more variable concentrations of daily PM10.
Average annual mortality events per 100 000 were higher in Birmingham, Minneapolis, Nashville, Pittsburgh, and Terra Haute. Mortality from any respiratory disease was about 10% of all-cause mortality (table 2), and of these, 40– 50% were deaths from chronic obstructive pulmonary disease (COPD), and about 35% were deaths from pneumonia. One third of all deaths were from heart disease, the majority of which were ischaemic heart disease (IHD) (70%), and in small proportions heart failure (HF) (6%), and dysrhythmias (6%). Almost half of ischaemic heart disease mortality cases were diagnosed as acute myocardial infarction (MI). Mortality from stroke was about 7% of all-cause mortality (table 2). In total, we studied 1 896 306 deaths.
Single lag model results showed, in general, increased risk of mortality with concentrations of PM10 the day before and two days before of the event (lags 1 and 2) (table 3). Because of the potential bias due to confounding from exposures in other days, the results from the unconstrained distributed lag models presented more valid estimates of the distribution of the effects of PM10 over time (table 3). All-cause mortality increased with increasing concentrations of PM10 the day before and two days before the event, with effects essentially zero on the same day. When we examined broad categories of mortality causes, we saw different patterns. Respiratory mortality effects were distributed over all three lags. In contrast, cardiac deaths were associated with exposure the day before and two days before the death. The more specific causes had more mixed patterns, but the standard errors of the effect estimates for these rarer outcomes were large, limiting interpretation of the distributed lag pattern. Interesting to note was the strong positive association for mortality from heart failure with concentrations two days prior the event.
The cumulative increase for all-cause mortality due to PM10 concentrations across the three days was 0.45%, which was about a quarter more than what was seen with the strongest single day lag. All respiratory mortality increased 0.87% across the three days. The largest impact of PM10 concentrations in these three days was on the deaths from pneumonia (1.24%). Deaths from any heart disease showed a 0.50% increase. The cumulative increase for IHD over the three days was somewhat greater than for all heart disease mortality. Mortality from stroke showed a 0.46% relative increase, although imprecisely estimated.
Matching on AT resulted in less power, thus it was difficult to interpret the results for cause specific mortality. However, the effects on all-cause mortality did show a similar pattern with the results of models that had included AT as a covariate. Overall, the estimate for all-cause mortality was 0.52% (95% CI 0.31% to 0.72%).
We found no great impact of city-level characteristics on the overall estimates of PM10—cause specific mortality, although power was weak for cause specific deaths. However, we observed differences in the PM10 effects for all-cause mortality for several of these characteristics (table 4). Percentage increase in daily mortality was higher for denser populated areas (90% greater comparing locations with 3000 persons/mile2 to locations with 1000 persons/mile2), and higher for those cities with a higher fraction of primary PM10 from traffic (threefold comparing the estimate for the 75th centile with that of the 25th centile of the distribution). Higher variability in summer apparent temperature was associated with a greater impact of PM10 (about sixfold for the comparison of the interquartile range, 25th to 75th centile), as were lower winter temperatures (44% increase comparing the 75th and 25th centiles of the distribution). These important findings are also illustrated in fig 1. There was an indication that a greater proportion of the elderly in the city also modified the effect of PM10 on daily mortality (proportion of those older than 65 or 75 years, with the difference between the first and third quartiles being 56% and 51% respectively) (table 4).
In a large, multi-city study, we examined the effects of PM10 on daily mortality across three days: on the day of the event (death), and one and two days before, using a case-crossover design. We observed increased all-cause mortality for increasing concentrations of PM10 on the day before and two days before the event.
Some previous studies have been criticised for reporting the PM10 effects on mortality for the exposure lag showing the strongest association. It has been argued that this overestimates the effect of exposure. However, we found that considering the multi-day effects increased our estimates of the short term effects of PM10 on all-cause mortality by about one quarter, compared to even the strongest lag. This finding is important for risk assessment, suggesting that effect estimates from studies that used a single day rather than a distributed lag exposure, will be lower. The effect of PM10 on mortality from respiratory causes persisted in all three days. This was also true for pneumonia deaths, whereas deaths from COPD were predominantly associated with PM10 concentrations the two days before. Deaths from all heart disease showed association primarily with concentrations of PM10 the two days before (lags 1 and 2). However, mortality from myocardial infarction was only associated with concentrations on the same day. We found a larger, but imprecise association with heart failure deaths. Stroke daily deaths were mainly associated with the day before concentrations (lag 1), although imprecisely estimated. Because of insufficient power, we could not make stronger conclusions about most of the specific causes of death. However, it is important to note the doubling of the effect for respiratory mortality when compared to all-cause mortality. The effect was as large in an analysis of all-cause mortality-PM10 association, where we used an interaction term for all respiratory deaths (0.72%; 95% CI 0.28% to 1.16%).
The findings of the study add to the previous literature in several ways. It confirms in a large multi-city US study with extended follow up (to 2000) what has previously been reported in Europe.8 The effects of PM10 on daily deaths persisted over at least two days, and incorporating that persistence resulted in higher effect estimates. The other largest study in the United States that had looked at the PM10-mortality association, due to the monitoring limitation could not examine the question of persistent effects in the course of more than one day.5,9 The results of the present study corroborated previous results reported by Braga and colleagues13 and Schwartz,46 which had used more limited data.
The persistent PM10 effects for respiratory mortality, and in particular pneumonia deaths supported epidemiological findings for an exacerbation of inflammation in immunocompromised subjects.53 Zelikoff and colleagues suggested that PM10 may act as an immunosuppressive agent which can disrupt normal pulmonary antimicrobial defence mechanisms, and illustrated this in animal models.54 They showed that even single exposures to ambient particulate matter are able to compromise normal recovery mechanisms.
The findings for myocardial infarction were consistent with the suggested mechanisms for exposure related triggering of coronary events within 24 hours.55–57 We confirmed the reports of Hoek and colleagues58 and Goldberg and colleagues59 that heart failure appears a particularly sensitive outcome. However, those previous studies could only investigate single locations. We also provided suggestive evidence that mortality from stroke was associated with PM10 in the United States. This has been previously only reported in Asian populations.14–17
We have also confirmed the findings reported in a more limited study by Schwartz,51 that choosing control days matched to the same temperature as the event days, did not change the PM10 effects. This indicates that the observed PM10-mortality associations were unlikely to be due to temperature confounding.
We considered the hypothesis that heterogeneity in the city PM10 effects could be due to certain city sociodemographic, physical, and environmental characteristics. The study of Janssen et al had reported differing of the PM10 effects on cardiovascular hospital admissions due to differences in the proportion of central air conditioning in each city.36 They also observed important difference by source of primary PM10, with traffic sources showing a greater effect, suggesting higher toxicity. Another study by Katsouyanni and colleagues looked at several groups of such city characteristics.8 They reported a higher effect of PM10 on mortality in cities with a lower ratio of PM10 to nitrogen dioxide (NO2) in the pollution mix, higher mean daily temperature, lower mean relative humidity, lower standard mortality rates, and higher proportion of susceptibles (age 65 years or older).
We found densely populated cities to have a greater effect of PM10 on daily mortality. We hypothesised that there were two reasons why denser cities had more daily deaths due to air pollution exposure. First, there may be a greater toxicity of the source particles, or, second, the population in denser cities may be more susceptible. Denser populated cities, in general, tend to be more urban, and thus a larger percentage of the primary PM10 is from traffic sources. The finding we obtained for percentage of primary PM10 from traffic corroborated this hypothesis. These results corresponded well with results of Laden and colleagues,60 who found greater toxicity from traffic particles in a source apportionment mortality study. This is also broadly consistent with the results of Katsouyanni and colleagues.8 The lower ratio of PM10 to NO2, which modified the particle effect in that study, could be interpreted as indicating a city with a larger proportion of particles from traffic, which generates most of the NO2 in urban areas. It may also explain the lower effect size estimates reported from Eastern Europe,4,8 where coal combustion accounts for a much larger proportion of the particles than in Western Europe. The difference in the toxicity of particles from different sources could be a result of the differences in specific components of particles.26
The second hypothesis, that such a difference in effects could be due to a more susceptible population in denser cities, was only partly supported when examining the age distribution in our study, as well as when examining the potential for modifying the PM10 effects of age standardised mortality rates, and the proportion of those greater than 65, 75, or 85 years old, in each city. Katsouyanni and colleagues had seen the greater effect of PM10 in cities with lower age standardised mortality rates,8 and with higher proportion of older ages (greater than 65 years). No modifying effect of age standardised mortality rates was seen in our study. However, the study findings suggested that a greater impact of PM10 concentrations was present in cities with a larger proportion of those older that 65 or 75 years.
A new observation of ours was a higher effect of PM10 in cities with greater summer temperature variability. High variability in temperature may be related to exposure differences, as more days with open windows may occur in the summer time in cities with such variability, compared to cities that are consistently hot in the summer, where continual air conditioning use is more common. Higher ventilation is associated with a steeper slope of the association between personal exposure to particles and ambient concentrations.61 The finding of a higher effect of PM10 in Mediterranean cities in the multicentre European study of air pollution (APHEA II) may reflect the same exposure phenomena,8 as air conditioning is rare in those cities, and high ventilation rates are the norm in the summer. Alternatively, this modification may result from some interaction between temperature exposure and exposure to air pollution.62 However, Samet and coworkers did not observe any interaction between temperature and air pollution in Philadelphia.63 Cooler winter temperatures are associated with increased mortality risk;48,64,65 they also appeared to modify the effects of air pollution in our study.
We note that some city characteristics may vary over a 12 year period, and we were limited on this by data availability. However, population density, and apparent temperature may not, on average, vary by much in a 12 year period.
Most of the specific groups of diseases examined in this study are strongly influenced by other risk factors, such as individual social habits (smoking and drinking), and individual characteristics. In our study, we controlled for this by design, using a case-crossover sampling. In such study each case serves as its control on the event-free days. This way we matched perfectly on potential confounders that change slowly over time.41–43
Particulate matter has been associated with short term increases in morbidity and mortality.
The study found increased all-cause, respiratory, and heart disease mortality in association with particulate matter of aero-diameter ⩽10 μm (PM10), with effects being persistent over several days. Matching control days on temperature did not change this association.
Several city characteristics modified the effects of PM10 on daily mortality, notably the percentage of particles from traffic.
This study did not address other air pollutants, in multiple pollutant scenarios. We made this choice for several reasons. First, the NMMAPS study, using more cities, and hence with more power, showed that gaseous pollutants were not confounders of the effects of PM10.5,9 The PM10 effects were not significantly changed by the presence of other co-pollutants in the models, in two or multiple pollutant models, nor were any of the other pollutants significant predictors of daily deaths, even in single pollutant models. Second, a very recent study by Schwartz controlled for confounding by co-pollutants in a case-crossover analysis by choosing control days that were matched on each gaseous pollutant in turn, and again, found no significant change in the estimates of PM10-mortality association, suggesting a causal relation between PM10 and daily mortality.66 Finally, two studies by Sarnat and colleagues have shown that ambient concentrations of gaseous pollutants may be surrogates for exposure to particulate matter, and indeed have stronger associations with exposure to particles than with exposure to the respective gases.61,67 Under such circumstances, controlling for an ambient gaseous pollutant, risks putting two surrogates for the same thing in a regression model, which is inappropriate.
In summary, the study showed important associations between all-cause and cause specific mortality and PM10, in a multi-city study in the United States including the most recent 12 year period of available data. The findings of our study corroborated those of earlier studies reported in Europe and the United States,4–6,8,9,13 including findings of immediate effects for MI, and effects spread over two to three days for most other causes. The observed persistent effect of PM10 does support the findings of several other studies.13,46 It also confirmed previous findings that certain characteristics which are particular to a location, either geographically or demographically, can affect the mortality-air pollution association.8,35,36 In particular, we found evidence that traffic particles are more toxic. On the other hand, we did not confirm the European findings of effect modification by age standardised mortality rates. In any event, such differing effects are important to consider in policy matters of regulating air pollution standards.
Reducing levels of particulate matter may contribute to reducing mortality associated with it.
Attention must be paid to traffic related air pollution, as suggestive of higher toxicity.
Funding: supported by the Harvard Environmental Protection Agency (EPA) PM Center, Grant R827353, and National Institute for Environmental Health Science (NIEHS) ES0002
Competing interests: none declared
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.