UBC Faculty Research and Publications

Air Pollution and Daily Mortality in a City with Low Levels of Pollution Vedal, Sverre; Brauer, Michael; White, Richard; Petkau, John Jan 31, 2003

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


52383-Vedal_S_et_al_Air_pollution_daily_mortality.pdf [ 1.43MB ]
JSON: 52383-1.0220733.json
JSON-LD: 52383-1.0220733-ld.json
RDF/XML (Pretty): 52383-1.0220733-rdf.xml
RDF/JSON: 52383-1.0220733-rdf.json
Turtle: 52383-1.0220733-turtle.txt
N-Triples: 52383-1.0220733-rdf-ntriples.txt
Original Record: 52383-1.0220733-source.json
Full Text

Full Text

Environmental Health Perspectives • VOLUME 111 | NUMBER 1 | January 2003 45Air Pollution and Daily Mortality in a City with Low Levels of Pollution Sverre Vedal,1 Michael Brauer,2,3 Richard White,4 and John Petkau41Department of Medicine, National Jewish Medical and Research Center, Denver, Colorado, USA; 2Department of Medicine, 3School of Occupational and Environmental Hygiene, and 4Department of Statistics, University of British Columbia, Vancouver, British Columbia, CanadaAn association between short-term changes inambient pollutant concentrations, especiallyparticle concentrations, and daily mortality hasbeen observed in many settings (1,2). Thesesettings have included a range of mean particleconcentrations and of particle emission sourcesand different types and concentrations of otherambient pollutants. When concentration–response relationships between particle con-centration and daily mortality have been evalu-ated, most studies have observed that therelationship is reasonably linear, even down tothe lowest concentrations observed in anygiven study (3–5). One implication of suchobservations is that no clear threshold concen-tration can be identified below which noeffects of particle pollution are observed.Conclusions about the linearity of theconcentration–response relationship extendingdown to the lowest observed concentrationsshould be tempered by the relative scarcity ofobservations at the lowest, as well as the high-est, particle concentrations in most studies.This is reflected in the larger confidence inter-vals around the fitted relationship typicallyobserved at both low and high concentrationextremes (6). Further, error in measuring indi-vidual particle exposures due to the use ofonly a few pollution monitors to reflect indi-vidual exposures would be expected to blurany threshold concentration, if one exists.This occurs because some individuals willhave true exposures above their “thresholds”when measured exposures are below thesethresholds, and some will have true exposuresbelow their “thresholds” when measuredexposures are above these thresholds (7). Theobserved linearity of the concentration–response relationships may therefore partly bea result of measurement error.Vancouver, British Columbia, Canada, is alarge urban area that has low levels of air pollu-tion relative to other large urban areas (2,8).For example, the mean daily PM10 (particulatematter ≤ 10 µm in diameter) concentration inVancouver during 1994–1996 was lower thanthat reported for any of the 90 cities studied inthe National Morbidity and Mortality AirPollution Study (NMMAPS) (2). A large num-ber of observations are therefore available atvery low concentrations of ambient particlesand other pollutants, which should enhance theability to assess effects at the low end of theconcentration–response relationship. The pres-ence of an association at these low concentra-tions would argue in favor of the linearity of theconcentration–response relationship extendingdown to very low concentrations. Conversely,the absence of an association would argue that athreshold concentration is present.MethodsMortality data. Vancouver is a metropolitanarea with a population of approximately 1.8million located on the southwest coast ofBritish Columbia. Daily mortality data forthe lower mainland of British Columbia,including Vancouver, for the 3-year periodJanuary 1994–December 1996 were obtainedfrom the Centre for Health Services andPolicy Research at the University of BritishColumbia, with approval from the BritishColumbia Vital Statistics Agency, Ministry ofHealth Services, and included date of death,underlying cause of death according to theInternational Classification of Diseases, NinthRevision (ICD-9), local health area of resi-dence and of death, and date of birth. Totaldeaths were defined as all deaths except thosewith an underlying cause of trauma or suicide(ICD-9 codes 800–999). Respiratory deathswere defined as all deaths coded with ICD-9codes of 460–519, and cardiovascular deathswere defined as deaths with ICD-9 codes of390–459.Pollution and meteorology data. Meanhourly pollutant concentration data [inhal-able particulate matter (PM10), ozone, nitro-gen dioxide, sulfur dioxide, and carbonmonoxide] were obtained from the GreaterVancouver Regional District for the same 3-year period for which mortality data wereavailable. The pollutant monitoring networkspanned the entire region extending fromVancouver proper eastward to Chilliwack, 90km east of Vancouver. The network for con-tinuous pollutant monitoring included 10monitoring sites for PM10 [using tapered ele-ment oscillating microbalance (TEOM)monitors], 19 O3 monitoring sites, 12 SO2sites, 19 NO2 sites, and 16 CO sites. Meanhourly meteorologic data on temperature (16sites), relative humidity (7 sites), barometricpressure (7 sites), and rainfall (8 sites) wereobtained from Environment Canada and theGreater Vancouver Regional District.Analysis. We filled in missing hourly datavalues for each pollutant and each meteoro-logic variable (except rainfall) separately usingthe EM algorithm, after preprocessing toremove systematic patterns in the data.Specifically, temporal patterns were firstremoved from the observed data at each ofthe available sites using sine and cosine func-tions with periods of 1, 1/2, 1/3, 1/4, 1/6,1/12 and 1/24 of a year, as well as indicatorsfor the hour of the day as predictors. WeAddress correspondence to S. Vedal, Division ofEnvironmental and Occupational Health Sciences,National Jewish Medical and Research Center, 1400Jackson Street, Denver, CO 80206 USA. Telephone:(303) 398-1520. Fax: (303) 398-1452. E-mail:vedals@njc.orgWe thank the British Columbia Vital StatisticsAgency, Ministry of Health Services, for providingthe vital statistics data for Vancouver.This work was supported by a grant from theBritish Columbia Lung Association. Received 5 October 2001; accepted 7 October 2002.ArticlesThe concentration–response relationship between daily ambient inhalable particle (particulatematter ≤ 10 µm; PM10) concentrations and daily mortality typically shows no evidence of athreshold concentration below which no relationship is observed. However, the power to assess arelationship at very low concentrations of PM10 has been limited in studies to date. The concen-trations of PM10 and other air pollutants in Vancouver, British Columbia, Canada, from January1994 through December 1996 were very low: the 50th and 90th percentiles of daily average PM10concentrations were 13 and 23 µg/m3, respectively, and 27 and 39 ppb, respectively, for 1-hr max-imum ozone. Analyses of 3 years of daily pollution (PM10, ozone, sulfur dioxide, nitrogen dioxide,and carbon monoxide) concentrations and mortality counts showed that the dominant associa-tions were between ozone and total mortality and respiratory and cardiovascular mortality in thesummer, and between nitrogen dioxide and total mortality in the winter, although some associa-tion with PM10 may also have been present. We conclude that increases in low concentrations ofair pollution are associated with increased daily mortality. These findings may support the notionthat no threshold pollutant concentrations are present, but they also raise concern that theseeffects may not be effects of the measured pollutants themselves, but rather of some other factor(s)present in the air pollution–meteorology mix. Key words: air pollution, mortality, nitrogen dioxide,doi:10.1289/ehp.5276 available via http://dx.doi.org/ozone, particulate matter. Environ Health Perspect 111:45–51 (2003). [Online 14 November 2002]removed any autoregressive (AR) structure inthe residuals from this initial fit by fitting anAR model; the AR order was chosen to ade-quately minimize residual autocorrelation,with fourth-order AR models being adequatefor all variables. We then applied the EMalgorithm to these residuals from all sitessimultaneously to take advantage of any spa-tial autocorrelation that might be present inthe variables. This yields an imputed residualfor each missing hourly value, which, whenadded to the predictions from the temporaland AR fits, provides the filled-in value. Theavailable rainfall data were converted to anindicator (present or absent) for each hour ateach site. The estimated probability of rainfallwas imputed for missing hourly values of theindicator using a logistic regression separatelyat each site, incorporating temporal patternsas described above for the other variables.The estimation procedure for missing datawas carried out on the log scale for the pollu-tion variables because their distributions wereskewed. To overcome the minor complicationof some sites reporting zero readings for somevariables, we added a small positive constant(one-half the smallest nonzero hourly mea-surement of that variable in the entire dataset) to all the observations before the datawere log transformed. After exponentiating totransform back to the original scale, we sub-tracted this same constant from each of thefilled-in hourly values.Because the mortality data were availableon a daily basis, all analyses were carried outwith daily data. We obtained overall daily val-ues of the pollutants (except O3) and themeteorologic variables by first averaging thefilled-in hourly values across the day at eachsite and then averaging these site-specificdaily average values across the available sites.For O3, the maximum hourly value over theday was first determined at each site, andthese site-specific maxima were then averagedacross the available sites. For rainfall, theoverall daily value is the proportion of hoursfor which rain was recorded.Because of the marked seasonality of someof the pollutants, we stratified analyses by sea-son. Only two “seasons” were chosen giventhe relatively moderate marine climate inVancouver: summer ranging from Maythrough September, and winter from Octoberthrough April. We used both Poisson regres-sion and generalized additive models for countdata (9) to estimate the effects of the air pollu-tants on daily mortality. A systematic batteryof preliminary analyses were carried out sepa-rately on the total, respiratory, and cardiovas-cular mortality counts in summer and winterto identify appropriate forms to adjust fortemporal trends and meteorologic effects. Theprimary approach taken to removing the tem-poral trends was to fit a loess smooth. Initially,we explored a wide range of window widthsusing total mortality, with more detailedexploration subsequently of window widthsranging from 90 to 180 days for all mortalityoutcomes. Examination of the partial autocor-relation functions suggested a window widthof 120 days sufficed to remove almost all ofthe autocorrelation in the residuals and wastherefore used for all subsequent loess-basedtemporal adjustments. We then explored vari-ous combinations of loess smooths of theArticles • Vedal et al.46 VOLUME 111 | NUMBER 1 | January 2003 • Environmental Health PerspectivesFigure 1. The time series of daily total, respiratory, and cardiovascular deaths from January 1994 through December 1996 show an annual cyclical pattern that ismost pronounced for respiratory deaths [dotted lines separate summer (S) from winter (W)]. The corresponding residual time series plots after use of loesssmoothing of long-term temporal trends show that no meaningful cyclical patterns persist.W S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S W1994 1995 1996 1994 1995 19961994 1995 1996 1994 1995 19961994 1995 1996 1994 1995 1996Year of study Year of studyYear of study Year of studyYear of study Year of study605040302012108642025201510520–23210–1–2–33210–1–2–3Cardiovascular deathsRespiratory deathsTotal deathsResidualsResidualsResidualsTime series Time series after loess smoothingmeteorologic variables as additional adjust-ments. The impact of these adjustments wasmuch more modest than that for the temporaltrends in the winter, but comparable in thesummer. In each case, there was little tochoose among many of the different possibleadjustments. Results reported here are formodels in which the meteorologic adjustmentis a sum of separate loess smooths of tempera-ture, relative humidity, barometric pressure,and rainfall on the same day, each based onthe default window width of one-half.To allow for easy interpretation and sig-nificance testing, we then entered linear termsfor each pollutant in separate models for daylags ranging from no lag (same day) to 2 days.We examined the simultaneous effects onmortality of two pollutants by fitting modelswith linear terms for each of the two pollutantstogether.We performed sensitivity analyses toassess the influence on the results of using adifferent approach to accounting for temporaltrends. For these analyses, a linear trend andsine and cosine terms with periods of 1 year,and 6, 4, and 3 months were included in themodels instead of a loess smooth to capturethe temporal trends. Exploration of variousloess smooths of the meteorologic variablesindicated that the same sum of separate loesssmooths of temperature, relative humidity,barometric pressure, and rainfall on the sameday provided adequate adjustment for themeteorologic variables. We used S-Plus (Insightful Corp., Seattle,WA, USA) for all analyses. Model fits usingloess smooths were based on stringent conver-gence criteria (10–9 for both the local scoringand backfitting algorithms) to eliminaterecently identified difficulties with the use ofthe default convergence criteria in the S-Plusgeneralized additive models (GAM) module(10). Standard errors for the pollutant effectsin these model fits were determined by aMonte Carlo approach (based on 1,000 sim-ulated Poisson data sets for each model fit) toavoid reliance on the potentially inaccurateapproximation used to evaluate these stan-dard errors in the S-Plus implementation ofGAM (10,11).ResultsMortality, pollution, and meteorology. Totaldaily deaths showed an annual cyclical pat-tern with peaks occurring in the winter(Figure 1). Numbers of daily deaths rangedlargely from 30 to 40 per day. Respiratorydeaths showed an even more pronouncedcyclical pattern, with peaks again occurring inthe winter. Cardiovascular deaths also showedan annual cyclical pattern, but this was muchless pronounced than for respiratory deaths.After removal of long-term temporal trendsusing a loess smooth (with a window width of120 days) separately for summer and winter,residual plots showed that the annual patternsseemed to be adequately removed (Figure 1).Each of the monitored pollutants andmeteorologic variables had some data miss-ing. For PM10, 8.9% of the monitor-hourshad missing concentration data. For O3, SO2,NO2, and CO, 4.7%, 6.4%, 6.5%, and 5.0%of the monitor-hours, respectively, were miss-ing. For temperature, humidity, barometricpressure, and rainfall, 5.3%, 12.9%, 14.7%,and 2.7% of the hourly measurements,respectively, were missing. We estimated themissing hourly data using an EM algorithmas detailed in “Methods,” with the overalldaily values based on the complete set of pol-lution and meteorologic data, both measuredand estimated, used in the analyses.The overall daily air pollution concentra-tions and the meteorologic measures exhibitedvarious temporal patterns (Figure 2). Of thepollutants, O3, with higher concentrations insummer, and CO, with higher concentra-tions in winter, exhibited the most seasonalvariability. Air pollution concentrations wereuniformly low (Table 1). The pollution vari-ables, apart from O3, were highly correlated(Table 2). Correlations between pollutantswere generally similar for the summer andwinter seasons, with the exception of O3, forwhich correlations with other pollutants werenegative in winter but positive in summer.Single-pollutant models. Linear terms forthe pollution variables were added singly tothe regression models in which separate loesssmooths were used to remove the long-termArticles • Low levels of air pollution and mortalityEnvironmental Health Perspectives • VOLUME 111 | NUMBER 1 | January 2003 47W S W S WW1994 1995 1996Year of study20100–10Temperature (C)Humidity (%)Pressure (kPas)S W1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of study1994 1995 1996Year of studyW S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S WW S W S W S WRain (% hr/day)1008060401041031021011009998100806040200302010PM10 (µg/m3 )3530252015105604020Max O3 (ppb)NO2 (ppb)SO2 (ppb)1510501.51.00.5CO (ppm)Figure 2. The time series of the overall daily valuesof the five air pollutants and the four meteorologicmeasures [dotted lines separate summer (S) fromwinter (W)] show varying degrees of an annualcyclical pattern. kPas, kiloPascals. Concentrationsof the air pollutants were uniformly low. temporal patterns and adjust for the same daytemperature, relative humidity, barometricpressure, and rainfall. Estimated effects fortotal, respiratory, and cardiovascular deaths byseason for a standard deviation change in con-centration for each pollutant at lags of 0, 1,and 2 days are presented in Figure 3. In thesummer, we observed a statistically significanteffect on total deaths only for O3 at lag 0, anda nearly significant effect (p < 0.10) for SO2 atlag 0. Statistically significant effects on respira-tory deaths in the summer were observed forPM10, O3, and SO2, but effects of NO2 andCO were also nearly significant. Effects oncardiovascular deaths were seen only for O3,and then only for lag 0 (p < 0.10). In the win-ter, we observed significant effects on totaldeaths for PM10 (lag 2), NO2 (lag 2), and SO2(lag 1). No significant deleterious effects onrespiratory deaths were observed. Effects oncardiovascular deaths were observed for NO2(lags 1 and 2) and SO2 (lag 1).Two-pollutant models. Because pollutantconcentrations are often highly correlated,findings from single-pollutant models may bedifficult to interpret. We fit models includinglinear terms for all pollutant lag pairs (lags 0,1, 2) to attempt to identify independent pol-lutant effects. We focused attention on thosemodels in which both pollutant effects wereobserved in the single-pollutant models. Inthe summer, of the significant effects detectedin the single-pollutant models, only theeffects of O3 on total mortality at lag 0 andon respiratory mortality at lag 2 remained sta-tistically significant in all two-pollutant mod-els (Figure 4 shows the estimates of effectfrom the most relevant two-pollutant mod-els). In the winter, only the effect for NO2 ontotal mortality at lag 2 was largely unchangedafter the addition of other pollutants (Figure4). Other effects that were significant in thesingle-pollutant models were substantiallydiminished after addition of some of theother pollutants in the two-pollutant models:for example, the greatly diminished effect ofPM10 at lag 1 on respiratory mortality in thesummer after the addition of O3, and thegreatly diminished effect of SO2 at lag 1 ontotal mortality in the winter after the additionof NO2 (Figure 4).Sensitivity analyses. As detailed in the“Methods,” another approach taken to remov-ing the long-term temporal trends was to fit alinear trend and sine and cosine functionswith frequencies ranging from 1 year to 3months to the mortality series. Residuals fromthese fits showed that the annual pattern fortotal, respiratory, and cardiovascular deathswas removed (data not shown). Although, ingeneral, there was little difference in estimatesof pollution effect between the two approachesto removing the long-term temporal trends,there were exceptions. For example, Figure 5illustrates the smaller effect estimates obtainedfor total mortality in the summer with thesine–cosine function approach. If differenceswere observed, loess smooth models almostalways resulted in larger estimates of pollutioneffect. However, no qualitative differences inthe study findings resulted from the use of thetrigonometric function approach rather thanloess smoothing for the removal of long-termtemporal trends.DiscussionIn Vancouver, where ambient concentrationsof all major air pollutants are low relative toother large urban areas (2,8), the principalfinding of this 3-year study was that increasesin the concentrations of some of the gaseousair pollutants, particularly O3 in the summer,were associated with increases in daily mor-tality. For PM10, the only relatively robusteffect was for increased respiratory mortalityin the summer at lag 1, but even that wassensitive to inclusion of O3 (at lag 2) in thetwo-pollutant model.Estimated effects of ambient particulatematter have been less consistently observed insome more recently reported time-series stud-ies, whereas gaseous pollutant effects have beenobserved more consistently (12–16). It is notknown whether these effects reflect those ofthe gaseous pollutants themselves, or whetherthe gaseous pollutants are acting as surrogatemarkers of pollutant sources that contain moretoxic compounds. In Vancouver, given the lowconcentrations of these pollutants, it seemsunlikely that the observed effects are due to themeasured pollutants themselves. The meanconcentrations of both PM10 and O3 (mean24-hr average was 13.7 ppb) during the 3-yearperiod of the study (1994–1996) were lowerthan those in any of the 90 cities studied inNMMAPS (2). Further, only one of the citiesin NMMAPS in which data on carbonmonoxide were available had a lower meanconcentration of carbon monoxide thanVancouver. Mean concentrations of NO2 andSO2 in Vancouver were lower than 72% and76% of the cities in NMMAPS, respectively.Because the mean concentrations reported inNMMAPS were trimmed means in which theupper and lower 10th percentiles wereexcluded, and because the distribution of con-centration data is typically skewed to the right,it would be expected that the trimmed meanswould be lower than those calculated from thecomplete set of data. If trimmed values hadbeen calculated for Vancouver, mean concen-trations relative to the cities included inNMMAPS would have been even lower.In Vancouver, the most notable effects onmortality were those associated with increasesin O3 concentrations. It is noteworthy that,of the pollutants, O3 was least strongly corre-lated with the other pollutants. Two-pollu-tant models with O3 also typically had thelowest correlations between pollutant effectestimates; effect estimate correlations may bemore relevant because they reflect pollutantcorrelations after long-term temporal trendsand effects of meteorology have beenremoved. The weaker correlations associatedwith O3 may have contributed to the abilityArticles • Vedal et al.48 VOLUME 111 | NUMBER 1 | January 2003 • Environmental Health PerspectivesTable 1. Distribution of daily mortality counts and overall daily pollutant and meteorologic measures.Min 10% 50% 90% Max Mean (SD)Total mortality 16 27 35 43 60 35.0 (6.6)Respiratory mortality 0 1 4 7 13 3.8 (2.2)Cardiovascular mortality 4 9 14 19 28 14.0 (4.0)PM10 (µg/m3) 4.1 7.8 13.1 22.8 37.2 14.4 (5.9)O3 [ppb (1-hr max)] 3.1 14.6 27.3 39.4 75.1 27.4 (10.2)NO2 (ppb) 4.3 11.8 16.1 22.9 33.9 16.9 (4.5)SO2 (ppb) 0.3 1.2 2.4 4.9 15.4 2.8 (1.7)CO (ppm) 0.3 0.4 0.5 0.9 1.9 0.6 (0.2)Temperature (°C) –8.8 2.7 10.1 17.8 24.2 10.2 (6.0)Relative humidity (%) 24.2 65.4 79.9 91.3 97.3 78.6 (11.1)Barometric pressure (kPas) 97.4 100.7 101.7 102.4 103.9 101.6 (0.7)Rainfall (% hours/day) 0 0 2.6 48.1 97.9 14.7 (20.9)Abbreviations: Max, maximum; Min, minimum; kPas, kiloPascals.Table 2. Pearson correlations among the overall daily pollutant and meteorologic measures by season.aPM10 O3 SO2 NO2 CO Temperature Humidity Pressure RainfallPM10 0.48 0.76 0.84 0.71 0.61 –0.35 0.01 –0.47O3 –0.32 0.44 0.45 0.12 0.41 –0.59 –0.02 –0.33SO2 0.78 –0.41 0.80 0.67 0.59 –0.36 0.13 –0.47NO2 0.73 –0.38 0.68 0.81 0.45 –0.24 –0.01 –0.31CO 0.76 –0.65 0.83 0.78 0.28 0.12 –0.01 –0.19Temperature –0.11 0.28 –0.13 –0.34 –0.29 –0.42 0.00 –0.41Humidity –0.38 –0.38 –0.21 –0.21 –0.05 0.21 –0.20 0.55Pressure 0.40 –0.25 0.35 0.24 0.28 –0.04 –0.21 –0.40Rainfall –0.55 –0.02 –0.40 –0.31 –0.31 0.13 0.56 –0.40aSummer (May–September) above diagonal; winter (October–April) below diagonal.to detect effects of O3 that were relativelyinsensitive to effects of other pollutants. Thedifficulty of disentangling the effects of theother pollutants is not surprising in view ofthe high correlations among them. Because ofconcerns regarding the interpretability of esti-mated effects in models containing severalstrongly correlated variables, we did notattempt to estimate effects from models thatincluded any more than two pollutant terms.Findings from a study of the acute mortal-ity effect of short-term increases in PM10 con-centrations in the 88 largest cities in theUnited States have recently been reported(2,10) (in only 88 of the 90 cities was an effectfor PM10 estimated). Because these cities wereselected only on the basis of population sizeand availability of pollution concentrationdata, and a standardized approach was taken tothe analysis, these data provide the most com-prehensive picture to date of the consistency ofacute PM10 effects on mortality. An overalleffect of a 0.21% increase in total mortality (ata lag of 1 day) for each 10 µg/m3 increase inPM10 concentration was estimated. However,a substantial degree of heterogeneity of effectacross cities was observed, with the estimatedeffect being zero or negative (that is, a decreasein mortality effect associated with an increasein PM10 concentration) in 32 of the 88 cities(36%) in the single-pollutant models. Thepresence of a positive effect in any given citywas not clearly related to the average PM10concentration in that city, although there wasa trend for stronger effects to be present incities with lower PM10 concentrations, sug-gesting that the inability to observe a PM10effect in the 33% of cities was not due to thePM10 concentrations being below a certainthreshold concentration below which effectswere not present.Based on our single-pollutant models, aneffect of PM10 was present for respiratorymortality in the summer and for total mortal-ity in the winter. However, these effects weresensitive to the addition of other pollutants inthe two-pollutant models. The effects ofPM10 were present even though the averagedaily PM10 concentration in Vancouver from1994 to 1996 (14.4 µg/m3) was lower thanthat of any of the 88 American cities during1987–1994, where average daily PM10 con-centrations ranged from a low of 15.3 µg/m3to a high of 53.2 µg/m3 (2). In a recent analy-sis of the concentration–response relationshipin the 20 largest of these U.S. cities (4), theinvestigators argued that no evidence for athreshold for PM10 could be found, at leastabove a concentration of 15 µg/m3. Similarfindings were reported for an analysis of all88 cities (5). The effects of PM10 in thisstudy are consistent with the contention thatthere is no threshold concentration for PM10,nor for some of the other pollutants.Because the absence of a concentration–response threshold for most air pollutantsseems biologically implausible, considerationshould be given to possible reasons that time-series studies seem to be unable to detect suchthresholds. First, a “blurring” of a thresholdmight be expected if the use of ambient pol-lutant concentrations measured with a fewArticles • Low levels of air pollution and mortalityEnvironmental Health Perspectives • VOLUME 111 | NUMBER 1 | January 2003 49Figure 3. Estimated increases (and 95% confidence intervals) in daily mortality (total, respiratory, and cardiovascular) corresponding to a standard deviationincrease in each pollutant from single-pollutant models for three lag periods for summer (May–September) and winter (October–April) periods.LagTotal mortality (summer)86420–2–4–6Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2PM10 O3 SO2 NO2 COLag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2Total mortality (winter)420–2–4Lag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2Lag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2Respiratory mortality (summer) Respiratory mortality (winter)Cardiovascular mortality (summer) Cardiovascular mortality (winter)Lag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2Lag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 23020100–101050–5–10Percent change/SD change1050–5–106420–2–4–6PM10 O3 SO2 NO2 COPM10 O3 SO2 NO2 CO PM10 O3 SO2 NO2 COPM10 O3 SO2 NO2 CO PM10 O3 SO2 NO2 COPercent change/SD changePercent change/SD changePercent change/SD changePercent change/SD changePercent change/SD changepollution monitors results in error in themeasurement of exposure. Such measurementerror must be present to some extent. Recentfindings of a simulation study based on actualambient and personal monitoring data sug-gest that measurement error can have a sub-stantial effect on the ability to detect athreshold (7). Others have not found thatmeasurement error causes any difficulty inidentifying a threshold within a meta-analysiscontext using a collection of cities (3). Thedetection of effects in Vancouver, where con-centrations of all of these pollutants are low,suggests that measurement error may not besolely responsible for the inability to detectthreshold concentrations in settings withhigher pollutant concentrations.Second, ambient concentrations of air pol-lutants may be acting as surrogate measures ofexposure to other agents or to specific pollu-tion sources that are in fact responsible for theobserved effects. For example, there is noknown mechanism whereby exposure toambient O3 might produce adverse cardiaceffects, although some have been suggested(17), yet O3 was associated with cardiac mor-tality in Vancouver. One could postulate thatambient O3 concentrations in this setting maybe reflecting other pollutants in the photo-chemical smog mix that might potentiallyhave adverse cardiac effects, although theidentity of these agents is not known. Noapparent threshold O3 concentration mighttherefore be observed if O3 were acting assuch a surrogate measure. Similarly, if PM10concentrations or concentrations of other pol-lutants are acting as surrogate measures ofanother unmeasured toxic pollutant or pollu-tants, or of specific sources, absence of anapparent threshold for these pollutants couldalso be observed. It has recently been suggested(18) that in some settings concentrations ofgaseous pollutants may be better measures ofexposure to particle pollution than the particlemass concentrations, in which case the appar-ent gaseous effects merely reflect unmeasuredeffects of particles. For example, others havealso reported an inverse association betweenO3 concentrations and mortality in the winter(19) (see Figure 3). It has been proposed thatthis effect is an example of negative confound-ing due to the negative correlation betweenwintertime O3 and fine PM (20). This seemsunlikely in our data given the wintercorrelation between O3 and PM10 of –0.32(Table 2), although the negative correlationwith PM2.5 could have been stronger. Therewas little change in the estimate of effect ofO3 on respiratory mortality in the winter inthe two-pollutant models (data not shown).We had no data on PM2.5 during the 3 yearsof the study to allow us to specifically addressthis hypothesis. Additional monitoring inmultiple settings is needed to determinewhether this is likely.Third, because meteorology is a strongdeterminant of air pollutant concentrations inan urban setting, with variation in pollutantemissions being relatively minor, ambient pol-lutant concentrations might conceivably alsobe acting as surrogate measures of meteoro-logic factors. If some of the observed effects ofthe pollutants are due to meteorology ratherthan to the pollutants themselves, again, athreshold for the pollutants might not beobservable. The addition of daily meteorologicvariables to the regression models in this studywas aimed at controlling for effects of meteo-rology in the estimation of air pollutanteffects. Alternative approaches to controllingfor the effects of meteorology have not signifi-cantly affected estimates of pollution effects inother studies (21), suggesting that significant,uncontrolled confounding by meteorology isunlikely. However, when effects of pollutantsare detected at low concentrations at which noadverse effects would be expected (that is, nothreshold concentrations are detected), andthe primary determinant of pollutant concen-trations is meteorology, then a case can bemade that the pollutant concentrations areArticles | Vedal et al.50 VOLUME 111 | NUMBER 1 | January 2003 • Environmental Health PerspectivesFigure 4. Estimated increases (and 95% confidence intervals) in daily mortality (total, respiratory, and car-diovascular) for a standard deviation increase in each pollutant from selected two-pollutant models. Corr,correlation between the pollutant effect estimates.Corr–0.3220100–10Percent change/SD changeO3lag 2Total (summer)86420–2Corr–0.67Cardiovascular (winter)Respiratory (summer)Corr–0.58Percent change/SD changeCorr–0.13Corr–0.22PM10lag 1PM10lag 1SO2lag 1O3lag 2SO2lag 1O3lag 0SO2lag 06420–2–4Corr–0.76Corr–0.47SO2lag 1PM10lag 2NO2lag 2PM10lag 2NO2lag 2SO2lag 1SO2lag 1NO2lag 16420–2 Corr–0.69Total (winter)Figure 5. Estimated increases (and 95% confidence intervals) in daily total mortality for a standard deviation increase in pollutant concentration estimated fromthe single-pollutant models for the summer using (A) loess smoothing or (B) a trigonometric function approach to removing long-term temporal trends. Estimatesof effect using the trigonometric functions resulted in smaller estimates of pollution effect in this case.Lag50–5–10Percent change/SD changeLag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2PM10 O3 SO2 NO2 COLag Lag Lag Lag Lag0 1 2 0 1 2 0 1 2 0 1 2 0 1 2CO50–5–10A BPercent change/SD changePM10 O3 SO2 NO2serving as better measures of the meteorologicfactors influencing mortality than the meteo-rologic measures themselves (in this case dailytemperature, humidity, barometric pressure,and rainfall). Exposure measurement error ofthe relevant meteorologic factors as measuredby pollutant concentrations would also pre-sumably be less than that of the ambient pol-lutants. Given the multitude of adverse healtheffects attributed to changes in meteorology(22,23), confounding by meteorology stillseems plausible.Confidence in the study findings can beenhanced by demonstrating that the methodof data analysis did not substantially influencethe findings. The findings did not exhibitmuch sensitivity to the approach taken toremoving the long-term temporal trends fromthe data. Findings may have been influenced,however, by the decision to use a seasonallystratified analysis. The decision to stratify theanalysis by season was prompted by the obvi-ous annual cycles in much of the time seriesdata (Figures 1 and 2). Rather than attempt-ing to incorporate this seasonal complexity ina single model, an approach that may notsucceed in adequately removing all of the sea-sonal correlations between the time-varyingmeasures (24), it seems justified to stratify byseason. A potential disadvantage of stratifica-tion is the loss of statistical power and theassociated instability of the estimates of effect,but this concern seems relatively unimportantin the current study.For loess smoothing of the long-termtemporal trend, our Monte Carlo estimates ofthe standard errors of the pollutant effectswere typically about 10% larger than thosereported by the S-Plus GAM module, thoughfor a few models they were as much as 30%larger. A number of additional pollutanteffects would have been judged to be statisti-cally significant based on the smaller standarderrors reported by the S-Plus GAM module.The latter should not be used for model fitsinvolving loess smoothing (10,11). For thepresent context of overdispersed Poissonregression models, the only currently availablealternative for evaluating accurate standarderrors is a Monte Carlo approach. This iscomputationally intensive and time-consum-ing. This fundamental limitation of theS-Plus GAM module may encourage the useof parametric approaches to smoothing (e.g.,natural splines) for which explicit evaluationof exact standard errors of pollutant effects isstraightforward even for this context.In conclusion, assuming that the findingsfrom Vancouver are generalizable to othercities with low pollutant concentrations,increases in air pollutant concentrations, evenwhen concentrations are low, are associatedwith adverse effects on daily mortality.Although this observation may support theargument that there are no threshold concen-trations of air pollution below which adverseeffects cannot be detected, it also raises con-cern that the associations are not reflecting theeffects of the measured pollutants, but rathersome factor or combination of factors, such as,for example, unmeasured air pollutants oruncontrolled features of meteorology that arecorrelated with the measured pollutants.REFERENCES AND NOTES1. Bascom R, Bromberg PA, Costa DA, Devlin R, DockeryDW, Frampton MW, Lambert W, Samet JM, Speizer FE,Utell M. Health effects of outdoor air pollution. Part 1. AmJ Respir Crit Care Med153:5–50 (1996).2. Samet JM, Zeger SL, Dominici F, Curriero F, Coursac I,Dockery DW, Schwartz J, Zanobetti A. Morbidity and mor-tality from air pollution in the United States. The NationalMorbidity, Mortality and Air Pollution Study (NMMAPS).Health Effects Institute Research Report No. 94, part II.Cambridge, MA:Health Effects Institute, 2000. 3. Schwartz J, Zanobetti A. Using meta-smoothing to esti-mate dose-response trends across multiple studies, withapplication to air pollution and daily death. Epidemiology11:666–672 (2000).4. Daniels MJ, Dominici F, Samet JM, Zeger SL. Estimatingparticulate matter-mortality dose-response curves andthreshold levels: an analysis of daily time-series for the 20largest US cities. Am J Epidemiol 152:397–406 (2000).5. Dominici F, Daniels M, Zeger SL, Samet JM. Air pollutionand mortality: estimating regional and national dose-response relationships. J Am Stat Assoc 97:100–111 (2002).6. Samet JM, Zeger SL, Berhane K. Particulate air pollution anddaily mortality: replication and validation of selected studies.The Phase I Report of the Particle Epidemiology EvaluationProject. Cambridge, MA:Health Effects Institute, 1995. 7. Brauer M, Brumm J, Vedal S, Petkau AJ. Exposure misclassification and threshold concentrations in timeseries analyses of air pollution health effects. Risk Anal22:1183–1193 (2002).8. Brook J, Dann T, Burnett R. The relationship among TSP,PM10, PM2.5, and inorganic constituents of atmosphericparticulate matter at multiple Canadian locations. J AirWaste Manage Assoc 47:2–19 (1997).9. Hastie TJ, Tibshirani RJ. Generalized Additive Models.London:Chapman and Hall, 1990.10. Dominici F, McDermott A, Zeger SL, Samet JM. On the useof generalized additive models in time-series studies of airpollution and health. Am J Epidemiol 156:193–203 (2002).11. Ramsay T, Burnett R, Krewski D. The effect of concurvityin generalized additive models linking mortality and ambi-ent air pollution. Epidemiology (in press).12. Mar TF, Norris GA, Koenig JQ, Larson TV. Associationsbetween air pollution and mortality in Phoenix, 1995–1997.Environ Health Perspect 108:347–353 (2000).13. Burnett RT, Cakmak S, Raizenne ME, Stieb D, Vincent R,Krewski D, Brook JR, Philips O, Ozkaynak H. The associa-tion between ambient carbon monoxide levels and dailymortality in Toronto, Canada. J Air Waste Manage Assoc48:689–700 (1998).14. Burnett RT, Cakmak S, Brook JR. The effect of the urbanambient air pollution mix on daily mortality rates in 11Canadian cities. Can J Public Health 89:152–156 (1998).15. Moolgavkaar SH. Air pollution and mortality in three U.S.counties. Environ Health Perspect 108:777–784 (2000).16. Touloumi G, Katsouyanni K, Zmirou D, Schwartz J, Spix C,Ponce deLeon A, Tobias A, Quennel P, Rabczenko D,Bacharova L, et al. Short-term effects of ambient oxidantexposure on mortality: combined analysis within theAPHEA project. Am J Epidemiol 146:177–185 (1997).17. Gong H, Wong R, Sarma RJ, Linn WS, Sullivan ED,Shamoo DA, Anderson KR, Prasad SB. Cardiovasculareffects of ozone exposure in human volunteers. Am JRespir Crit Care Med 158:538–546 (1998).18. Sarnat JA, Koutrakis P, Suh H. Assessing the relationshipbetween personal particulate and gaseous exposures ofsenior citizens living in Baltimore, MD. J Air WasteManage Assoc 50:1184–1198 (2000).19. Samet JM, Dominici F, Curriero FC, Coursac I, Zeger SL.Fine particulate air pollution and mortality in 20 U.S. cities.N Engl J Med 343:1742–1749 (2000).20. Sarnat JA, Schwartz J, Suh HH. Fine particulate air pollu-tion and mortality in 20 U.S. cities [Letter]. N Engl J Med344:1253–1254 (2000).21. Pope CA, Kalkstein LS. Synoptic weather modeling andestimates of the exposure-response relationship betweendaily mortality and particulate air pollution. Environ HealthPerspect 104:414–420 (1996).22. Rogot E, Padgett SJ. Associations of coronary and strokemortality with temperature and snowfall in selected areas ofthe United States, 1962–1966. Am J Epidemiol 103:565–575(1976).23. Kunst AE, Looman CWN, Mackenbach JP. Outdoor airtemperature and mortality in the Netherlands: a time-series analysis. Am J Epidemiol 137:331–341 (1993).24. Chock DP, Winkler SL, Chen C. A study of the associationbetween daily mortality and ambient air pollutant concen-trations in Pittsburgh, Pennsylvania. J Air Waste ManageAssoc 50:1481–1500 (2000).Articles • Low levels of air pollution and mortalityEnvironmental Health Perspectives • VOLUME 111 | NUMBER 1 | January 2003 51


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items