Xi’an, with an area of 9,983 km2
and a resident population > 8.1 million in 2005, is the capital of Shanxi Province, China. Xi’an is the largest city in northwestern China, and it experiences some of the worst air pollution among China’s cities (Cao et al. 2005
). Our study area was limited to the urban area of Xi’an, an area of 1,166 km2
with a resident population of > 2.7 million.
Mortality data. We obtained numbers of deaths among urban residents in Xi’an for each day for 1 January 2004 through 31 December 2008 from the Shanxi Provincial Center for Disease Control and Prevention (SPCDCP). In Xi’an, all deaths, regardless of whether they occur in a hospital or at home, must be reported to appropriate authorities before cremation of the remains. Hospital or community doctors must indicate the cause of death on a death certificate card that is sent to the SPCDCP. SPCDCP staff then classify the cause of death according to the International Classification of Diseases, 10th Revision [ICD-10; World Health Organization (WHO) 1992] as due to total nonaccidental causes (ICD-10 codes A00–R99), cardiovascular diseases (I00–I99), respiratory diseases (J00–J98), or injury (S00–T98). The Chinese government has mandated detailed quality assurance (QA) and quality control (QC) programs for the SPCDCP death registry.
Pollutant and meteorological data. For this study, we measured daily concentrations of PM2.5, organic carbon (OC), elemental carbon (EC), and 10 water-soluble ions [i.e., sodium ion (Na+), ammonium (NH4+), potassium ion (K+), magnesium ion (Mg2+) calcium ion (Ca2+), flouride (F–), choride (Cl–), nitrite (NO2–), sulfate (SO42–) and nitrate (NO3–)] for 1 January 2004 through 31 December 2008 (1,827 days). We also measured concentrations of 15 elements [i.e., sulfur (S), chlorine (Cl), potassium (K), calcium (Ca), titanium (Ti), chromium (Cr), manganese (Mn), iron (Fe), nickel (Ni), zinc (Zn), arsenic (As), boron (Br), molybdenum (Mo), cadmium (Cd), and lead (Pb)] for 1 January 2006 through 31 December 2008 (1,096 days).
monitoring site was located on the rooftop of the Chinese Academy of Sciences’ Institute of Earth Environment building in an urban-scale zone of representation (Chow et al. 2002
). The site was surrounded by a residential area where there were no major industrial activities nor local fugitive dust sources [see Supplemental Material, (http://dx.doi.org/10.1289/ehp.1103671
samples were obtained 10 m above the ground. Our previous studies suggest that the measured PM2.5
concentrations at this monitoring station are representative of the general status of PM2.5
pollution in Xi’an (Cao et al. 2005
Figure 1 Estimated percent increases [mean (95% CI)] in total, cardiovascular, and respiratory mortality per IQR increase in pollutant concentrations on the current day (lag 0) or the previous 1–3 days (lags 1, 2, and 3), adjusted for temporal trend, day (more ...)
samples were collected using two battery-powered mini-volume samplers (MiniVol™ TAS; Airmetrics, Eugene, OR, USA) operating at a flow rate of 5 L/min (Cao et al. 2003
). We used a relatively low flow rate due to high PM loading in Xi’an. PM2.5
samples were collected on 47-mm Whatman quartz microfiber filters that were pre-heated at 900°C for 3 hr before sampling. The quartz-fiber filters were analyzed gravimetrically for mass concentrations. We analyzed a 0.5-cm2
punch from each sample for OC and EC using a Desert Research Institute (DRI) model 2001 thermal/optical carbon analyzer (Atmoslytic Inc., Calabasas, CA, USA) for eight carbon fractions following the IMPROVE (Interagency Monitoring of Protected Visual Environments) thermal/optical reflectance (TOR) protocol (Chow et al. 2004
). Levels of the five water-soluble cations (Na+
) and five water-soluble anions (F–
) were determined in aqueous extracts of the sample filters using an ion chromatograph (Dionex 600; Dionex, Thermo Fisher Scientific, Inc., Cambridge, England, UK). Cation concentrations were determined using a CS12A column (Dionex), and anions were separated by an AS11-HC column (Dionex). The elemental concentrations of these samples were then determined by energy dispersive X-ray fluorescence (ED-XRF) spectrometry using the PANalytical Epsilon 5 XRF analyzer (PANalytical B.V., Almelo, the Netherlands). Detailed descriptions of the sample pretreatment, specific methods, detection limits, and QA/QC have been discussed previously (Cao et al. 2003
; Shen et al. 2009a
To adjust for the effect of gaseous pollutants and weather on mortality, we obtained daily concentrations of sulfur dioxide (SO2) and nitrogen dioxide (NO2) from the Xi’an Environmental Monitoring Center, and daily mean temperature and humidity from the Xi’an Meteorological Bureau. The SO2 and NO2 concentrations were averaged from the available monitoring results across seven stations in our study area. According to the rules of the Chinese government, we assumed the monitoring data from these stations generally reflected the background urban air pollution of Xi’an rather than pollution from local sources.
Statistical methods. Due to different time periods for measuring PM2.5 constituents, we constructed two data sets to analyze the data: The first involved daily measurement of PM2.5, OC, EC, and ions for 1 January 2004 through 31 December 2008 and the second included daily concentrations of PM2.5 and constituent elements for 1 January 2006 through 31 December 2008.
Daily counts of deaths and air pollution levels were linked by date and analyzed with time–series analyses (Bell et al. 2004
). Because daily counts of deaths approximate a Poisson distribution and the relationship between mortality and explanatory variables is mostly nonlinear, we used overdispersed generalized linear Poisson models (quasi-likelihood) with natural spline (ns
) smoothers to analyze mortality, PM2.5
constituents, and covariate data.
In the basic model, we incorporated smoothed spline functions of time, accommodating both nonlinear and nonmonotonic relations between mortality and time and thus providing a flexible model to control for long-term and seasonal trends (Hastie and Tibshirani 1990
). Day of the week (DOW
) was included as a dummy variable (a variable that takes on the values 1 and 0; also called an indicator variable) in the basic models. Partial autocorrelation function (PACF
) was used to guide the selection of degrees of freedom (df) for the time trend until the absolute values of the sum of PACF
of the residuals for lag days of up to 30 reached a minimal value (Peng et al. 2006
; Touloumi et al. 2004
). We used residual plots and PACF
plots to examine residuals of the basic model for discernable patterns and autocorrelation.
After establishing the basic model, we introduced the PM2.5
constituents and covariates (including temperature, humidity, and SO2
concentrations) in the model. Based on previous literature (Dominici et al. 2006
), we used smoothed spline functions with 3 df (for the whole period of the study) to control for temperature and relative humidity. To examine the temporal relationship of PM2.5
constituents with mortality, we fitted the models with different lag structures from 0 lag days to 3 lag days because our previous work on PM2.5
and daily mortality in China showed little evidence of a significant association with a lag beyond 3 days (Kan et al. 2007
; Ma et al. 2011
). A lag of 0 days (lag 0) corresponds to the current-day PM2.5
, and a lag of 1 day (lag 1) refers to the previous-day PM2.5
. We used the smoothing spline, with 3 df for PM2.5
, to graphically describe its relationships with mortality. We compared the linear and spline models by computing the difference between the deviances of the fitted two models (Dominici et al. 2002
; Samoli et al. 2005
). We estimated associations of PM2.5
constituents with mortality before and after adjustment for PM2.5
mass. Finally, to examine the robustness of our choice on the optimal values of df for time trend, we performed a sensitivity analysis to test the impact of df selection on the regression results.
All analyses were conducted in R version 2.10.1 (http://www.R-project.org
) using the MGCV package. The results are presented as the percent change in daily mortality per interquartile range (IQR) increase of pollutant concentrations unless specified otherwise. Statistical significance was defined as p