Introduction

Epidemiologic studies have consistently reported associations between exposure to air pollution, including particulate matter (PM), and adverse health effects (Pope and Dockery 2006; HEI 2010; Brook et al. 2010). There is evidence that fine particulate matter (PM2.5; PM with an aerodynamic diameter 2.5 μm and smaller) generated by combustion may be especially damaging to human health (Pope and Dockery 2006; Schlesinger et al. 2006). Estimates of the annual global mortality attributable to outdoor air pollution range from 0.8 million to over 4 million, with the majority of attributable deaths occurring in Asia (Cohen et al. 2005; Anenberg et al. 2010). However, due in part to challenges in population exposure assessment, relatively few epidemiologic studies have been conducted outside of North America and Europe (HEI 2004).

Mongolia’s population has undergone rapid urbanization since the mid-1990s, and this shift has had a major impact on the capital city, Ulaanbaatar, which is now home to 1.11 million of the nation’s 2.74 million inhabitants (National Statistical Office 2010). Population growth has led to major increases in the city’s air pollution emissions (Asian Development Bank 2006; Guttikunda 2007). Much of the population growth has been in the city’s low-income ger (traditional Mongolian dwelling) areas where coal and wood are burned for heat (World Bank 2004). Half of Ulaanbaatar’s population lives in a ger (Asian Development Bank 2006), and the city’s 160,000 gers each burn an average of 5 t of coal and 3 m3 of wood per year (Guttikunda 2007). Mobile sources also contribute to air pollution in Ulaanbaatar. From 1995 to 2005, the number of vehicles in Ulaanbaatar increased from 30,000 to 75,000 (Asian Development Bank 2006), and Mongolia is 1 of only 17 countries where leaded gasoline is still legally available (HEI 2010). The city’s other major air pollution sources include three coal-fueled combined heat and power plants, approximately 400 heat-only boilers, and wind-blown dust (World Bank 2004; Davy et al. 2011). A recent source apportionment study found that the majority of PM2.5 in Ulaanbaatar is produced by coal combustion (Davy et al. 2011).

Ulaanbaatar is located in a valley with mountains to the north and south (Asian Development Bank 2006; Davy et al. 2011). The topography, extensive pollution emissions, and frequent temperature inversions combine to cause very high pollution concentrations in the winter months. Given the limited applications of current exposure assessment techniques in developing countries and the limited data on Ulaanbaatar’s air quality in the literature (Davy et al. 2011), our objectives were to (1) characterize air pollution concentrations and temporal patterns; (2) assess the feasibility of using two current exposure assessment techniques, land use regression (LUR) modeling and mobile monitoring, in a developing country; (3) characterize spatial patterns in pollutant indicators of specific sources to identify “hot spots” and create exposure assessment tools; and (4) estimate the mortality attributable to outdoor air pollution in Ulaanbaatar.

Methods

Fixed-site monitoring

The government air pollution monitoring network in Ulaanbaatar has improved considerably in recent years. PM2.5, the criteria pollutant most relevant to human health, is now routinely monitored using tapered element oscillating microbalances (TEOMs) at four of the nine government monitoring sites in the city (the locations of the government monitoring sites are shown in Fig. 1). Air pollution data from June 1, 2009 to May 31, 2010 were obtained from the Ulaanbaatar City Environmental Monitoring Agency’s four PM monitoring sites (Fig. 1). Our main focus for analysis was site #1 because (1) it had the most complete data record for the period of interest (Table 1) and (2) it is centrally located and may, therefore, be the single site most representative of overall population exposure in Ulaanbaatar (Fig. 1). The PM10, PM2.5, and SO2 data from this site were used to characterize diurnal and seasonal patterns in pollution concentrations and to estimate the annual average PM2.5 concentration for use in the attributable mortality calculation described below. SO2 concentrations are reported in units of micrograms per cubic meter, but are also reported in parts per billion (assuming 0°C) for comparison with Ogawa passive sampler measurements. PM2.5 data from the city’s other three monitoring sites were only 50–77% complete (Table 1). To approximate the long-term PM2.5 concentrations at these sites, we replaced missing observations with the site- and season-specific median concentrations.

Fig. 1
figure 1

Map of the study area including Ogawa monitoring locations and government-run PM2.5 monitoring sites

Full size image
Table 1 Percent of possible 30-min measurements collected from June 1, 2009 to May 31, 2010 at four monitoring sites operated by the Ulaanbaatar City Environmental Monitoring Agency
Full size table

Land use regression modeling

LUR models have become a very common exposure assessment tool in wealthy countries. These models are developed based on relatively spatially dense monitoring of one or more pollutants and the road configuration, population density, land use, elevation, and other geographic characteristics surrounding the measurement sites (Hoek et al. 2008). Based on the empirical relationship between concentrations and predictors at the measurement locations, it is possible to predict concentrations at unmeasured locations (Hoek et al. 2008). To assess the wintertime spatial patterns in traffic-related air pollution and pollution produced by coal burning, we measured 14-day average concentrations of NO2 and SO2, respectively, using two-sided passive Ogawa samplers attached to power poles, telephone poles, etc., at approximately 3 m above ground level. The samplers were deployed at 38 locations in Ulaanbaatar on February 24 and 25, 2010 and retrieved in the same order on March 10 and 11, 2010 (Fig. 1). Three of the 38 sites were colocated with government monitoring sites, and the remaining locations were selected based on local knowledge to cover the study area and capture a wide range of NO2 and SO2 concentrations (Fig. 1). After retrieval, the Ogawa samplers were shipped to Vancouver, Canada and analyzed by ion chromatography at the University of British Columbia School of Environmental Health laboratory. Based on four field blanks, we determined the limits of detection (LOD, calculated as three times the standard deviation of field blanks) for NO2 and SO2 to be 0.8 and 2.5 ppb, respectively.

Given the relative lack of available geographic information systems (GIS) data in Ulaanbaatar, we derived data from several sources. A digital elevation model (DEM) produced by Environmental Systems Research Institute (Redlands, CA, USA) and provided with ArcGIS 9.3 was used to calculate elevations. Land use data, which are commonly used as predictors in LUR models (Hoek et al. 2008), were not available for Ulaanbaatar. Therefore, to obtain information on land cover, we used Landsat Enhanced Thematic Mapper Plus (ETM+) satellite images (http://www.landcover.org), which have been used in previous LUR modeling efforts (Su et al. 2008a, 2009). The ETM+ images for Ulaanbaatar (path 131, row 27) were captured on August 13, 2006 and orthorectified by the United States Geological Survey. Using a tasseled cap transformation (Crist and Cicone 1984), ETM+ bands 1–5 and 7 were simplified into three dimensions: brightness (a measure of soil reflectance), greenness (a measure of the presence and density of green vegetation), and wetness (Su et al. 2009; Crist and Cicone 1984). The locations of the city’s ger areas were determined based on the road network, features in the DEM and ETM+ data, observations made during Ogawa sampler deployment, and local knowledge (Fig. 1). Data on roads were obtained from Open Street Map (http://www.openstreetmap.org/), and minor modifications were made based on local knowledge and the location of features in the ETM+ images. Roads were divided into two categories: Peace Avenue (the city’s busiest road and main east–west connector) and major roads. These GIS data layers (Fig. 1) were used to derive 46 potential predictors of NO2 and SO2 concentrations (Table 2).

Table 2 Variables screened in developing LUR models for NO2 and SO2
Full size table

LUR models for NO2 and SO2 were developed using methods that have previously been applied to several North American cities (Henderson et al. 2007; Poplawski et al. 2008; Allen et al. 2011). We first calculated correlations between each potential predictor and the pollutant, then ranked the predictors in each subcategory (Table 2) by the absolute value correlation. We then removed any variables in a subcategory that were correlated (r > 0.6) with that category’s highest ranking variable. All remaining variables were entered into a stepwise multiple linear regression model, and the models were rerun as necessary to include only variables contributing at least 1% to the model R 2 and coefficients consistent with a priori assumptions (e.g., positive coefficients for road variables in the NO2 model). Model performance was evaluated based on the model-based R 2 and the R 2 and root mean square error from a “leave one out” approach in which the model was repeatedly calibrated based on all but one measurement then used to predict the excluded measurement. Residuals from both models were evaluated for normality and spatial autocorrelation (Moran’s I statistic), and variance inflation factors (VIF) were calculated for all models.

To account for any bias in the Ogawa measurements, we intended to adjust the Ogawa concentrations based on colocated NO2 and SO2 measurements at three government monitoring sites (Fig. 1). However, missing government data during the 2 weeks of Ogawa monitoring did not allow for such an adjustment. As a result, our LUR models provide an assessment of relative concentrations across the city, but the absolute concentrations have not been independently verified.

Mobile monitoring

We used a mobile monitoring approach to assess spatial patterns in PM2.5 concentrations resulting primarily from home heating (Larson et al. 2007; Su et al. 2008b; Lightowlers et al. 2008; Robinson et al. 2007). The details of the method are presented elsewhere (Larson et al. 2007). Briefly, on three evenings (between approximately 2000–2300), we drove predetermined routes that were selected based on local knowledge to capture different land uses (including areas with high and low ger density) and a wide range of PM2.5 concentrations. We drove the same route in opposite directions on February 24 and 25, 2010 and a different route on the evening of February 26, 2010. A nephelometer (Radiance Research M903, Seattle, WA, USA), blower, and air preheater were placed in the back seat of the vehicle and the nephelometer’s inlet was extended out the window. The nephelometer recorded the particle light scattering coefficient (b sp) at 15-s averages, while a global positioning system receiver (Garmin 60CSx, Olathe, KS, USA) connected to a magnetic antenna on the roof of the vehicle recorded the vehicle’s location at 5-s intervals. Additional evenings of monitoring would be needed to definitively characterize spatial PM2.5 concentration patterns; our goal was to assess the feasibility of the mobile monitoring technique in this setting.

We followed the approach of Larson et al. (2007) for removing the influence of temporal variation on the mobile measurements to allow for comparisons between measurements made at different times and during different evenings. Specifically, we used central-site TEOM data (from the City Environmental Monitoring Agency’s site #1) to adjust for within- and between-evening temporal trends. The temporally adjusted light scattering data were then spatially smoothed by calculating the average value in a 100-m radius around each measurement location. The temporally adjusted data from February 24 and 25 were then averaged. Finally, we converted the light scattering data into approximate PM2.5 concentrations using a previously published b sp–PM2.5 mass relationship from a study in Seattle, Washington (Liu et al. 2002) where wood burning is a major source of PM2 .5:

$$ {\text{P}}{{\text{M}}_{{2.5}}}\left( {\mu {\text{g}}/{{\text{m}}^3}} \right) = \left[ {\left( {{b_{\text{sp}}} \times {{10}^{{ - 5}}}{{\text{m}}^{{ - 1}}}} \right) + 0.39} \right]/0.27 $$
(1)

Because this approach provides only a semiquantitative estimate of PM2.5, results were grouped into concentration tertiles for mapping.

Estimation of mortality attributable to air pollution

Mortality data by age and cause, as indicated by the International Classification of Diseases, 10th Revision (ICD-10), for the year 2009 were obtained from the Statistical Department at the Mongolian Government Implementing Agency/Department of Health. We used the World Health Organization’s (WHO) approach for environmental burden of disease calculations (Cohen et al. 2005; Ostro 2004) to estimate the deaths from lung cancer (ICD-10 code C34) and cardiopulmonary causes (ICD-10 codes I10–I70 and J00–J99) that are attributable to long-term exposure to outdoor air pollution in Ulaanbaatar. First, we estimated for both causes of death, the air pollution-attributable fraction, assuming mortality effects up to the existing long-term PM2.5 concentration (X = 75 μg/m3, the annual average concentration for June 1, 2009 to May 31, 2010 at Ulaanbaatar City Monitoring Agency site #1) relative to a counterfactual concentration (X o), assuming a log-linear concentration–mortality relationship (Ostro 2004; Pope et al. 2009) derived from the American Cancer Society (ACS) cohort study of air pollution and mortality (Pope et al. 2002). X o was initially set at 7.5 μg/m3, the lowest PM2.5 concentration observed in the ACS study. The number of deaths attributable to air pollution was then determined based on the attributable fractions and the number of deaths among those 30 years or older from lung cancer (114) or cardiopulmonary causes (2,007) in Ulaanbaatar in 2009. As a sensitivity analysis, we also calculated attributable mortality under alternative scenarios, including a linear concentration–mortality relationship (Ostro 2004; Kunzli et al. 2000), a counterfactual concentration (X o) of 3 μg/m3, and maximum truncation concentrations (X) of 96 μg/m3 (the annual average concentration in the more polluted traditional housing areas approximated from measurements at Ulaanbaatar City Monitoring Agency site #2) and 50 μg/m3 (i.e., we assumed no additional attributable mortality above 50 μg/m3 in consideration of the fact that the ACS study was conducted in the US where PM2.5 concentrations are relatively low).

Results

Government fixed-site data

The annual average concentrations of PM10, PM2.5, and SO2 measured at the City Monitoring Agency’s site #1 were 165.1, 75.1, and 50.5 μg/m3 (17.7 ppb), respectively. Concentrations were highest in winter (Fig. 2). For example, the mean (±SD) 24-h PM2.5 concentration in summer (June–August) was 22.8 ± 9.0 μg/m3, while in winter (December–February), the mean concentration was 147.8 ± 61.2 μg/m3. The 24-h PM2.5/PM10 ratios were also highly variable between seasons (Fig. 2), with a mean ratio of 0.26 ± 0.11 in summer and 0.78 ± 0.12 in winter.

Fig. 2
figure 2

Monthly distributions of 24-h average a temperature, b PM10, c PM2.5, d PM2.5/PM10 ratio, and e SO2 measured at the Ulaanbaatar City Environmental Monitoring Agency’s site #1 from June 1, 2009 to May 31, 2010

Full size image

In addition to seasonal variation, pollution levels also varied diurnally with two concentration peaks per day. In both summer and winter, the morning PM2.5 concentration peak occurred between approximately 0800 and 1000 (Fig. 3). In the summer, the maximum evening levels occurred between approximately 2000 and 2300, while in the winter, the highest evening concentrations were from approximately 2200 to 0200.

Fig. 3
figure 3

Diurnal patterns in PM2.5 concentrations from a June to August and b December to February at the Ulaanbaatar City Environmental Monitoring Agency’s site #1. PM2.5 concentrations are expressed as the ratio of hourly concentration to average concentration over the 3-month period (June–August average, 23 μg/m3; December–February average, 148 μg/m3)

Full size image

After replacing missing observations with site- and season-specific median values, the annual average PM2.5 concentrations at monitoring sites #2, #3, and #4 were 96, 67, and 57 μg/m3, respectively, with wintertime averages of 248, 172, and 153 μg/m3, respectively.

Land use regression models

Ogawa samplers were retrieved from 37 of the 38 sampling locations. All NO2 concentrations were above the LOD; two SO2 measurements below the LOD (2.5 ppb) were assumed to have a concentration of 1.25 ppb (LOD/2). We colocated our Ogawa monitors with government monitors at three locations, but adjustment of the Ogawa measurements, which are known to underestimate NO2 concentrations at cold temperatures (Hagenbjork-Gustafsson et al. 2010), was not possible due to large gaps in the government data during the 2 weeks of Ogawa monitoring. NO2 and SO2 were normally distributed with mean (±SD) concentrations of 10.7 ± 5.8 and 17.0 ± 11.8 ppb, respectively, and the two pollutants were moderately correlated (r = 0.50; p < 0.01). There was significant spatial autocorrelation in the measured concentrations of both NO2 (Moran’s I = 0.42, p < 0.01) and SO2 (Moran’s I = 0.50, p < 0.01).

The LUR model predictors for NO2 were satellite-based greenness, ger areas, major roads, and distance to city center (Table 3). The model-based R 2 was 0.74 and the cross-validation R 2 was 0.66 (Table 3). Of the 46 potential predictor variables, average greenness in a 1,000-m buffer had the strongest bivariate relationship (R 2 = 0.47) with NO2. The VIF for predictors in the NO2 LUR model were ≤1.41 and there was no significant spatial autocorrelation in the model residuals (Moran’s I = 0.03).

Table 3 Wintertime LUR models for NO2 and SO2 in Ulaanbaatar
Full size table

The final SO2 LUR model included satellite-based greenness and ger areas as predictors (Table 3). The model explained 78% of the variability in SO2 concentrations, with a cross-validation R 2 of 0.75 (Table 3). Ger area in a 2,000-m buffer had the strongest bivariate relationship with SO2 (R 2 = 0.67). The average satellite-based brightness in a 2,000-m buffer was also highly correlated with SO2 (R 2 = 0.55), although this variable was also correlated with ger areas and, therefore, did not appear in the final LUR model. The SO2 model predictors had VIF = 1.1 and there was no significant spatial autocorrelation in the model residuals (Moran’s I = −0.14).

The two LUR models captured the different spatial patterns for these pollutants, with higher NO2 concentrations around the city center and near major roads and higher SO2 concentrations in the ger areas north of the city (Fig. 4). There was a strong correlation (r = 0.96) between the modeled SO2 concentrations and wintertime (December–February) PM2.5 concentrations measured at four government-run sites, although this correlation was driven primarily by one influential observation from government site #2, which had relatively high modeled SO2 and a wintertime average PM2.5 concentration of 248 μg/m3.

Fig. 4
figure 4

LUR model predictions of wintertime a NO2 and b SO2 in Ulaanbaatar

Full size image

Mobile monitoring

Mobile monitoring was conducted on cold, calm evenings that were typical for the time of year (Fig. 2a). The mean temperature and wind speed on February 24, 25, and 26 were −22.5°C and 1.4 m/s, −19.6°C and 0.5 m/s, and −14.3°C and 0.8 m/s, respectively. We observed a wide range of light scattering values across Ulaanbaatar; the interquartile ranges of temporally adjusted approximate PM2.5 concentrations (converted from b sp) measured during mobile monitoring on February 24/25, 2010 and February 26, 2010 were both 85 μg/m3 (25th–75th percentiles, 110–195 μg/m3 on February 24/25 and 85–170 μg/m3 on February 26) (Fig. 5). The spatial patterns captured by mobile monitoring were generally similar to SO2 patterns predicted by the LUR model (Fig. 5). The mobile monitoring routes passed within 250 m of 25 Ogawa monitoring sites, and at these sites, the correlation between the 2-week average SO2 concentration and the nearest temporally adjusted light scattering coefficient was 0.55 (p < 0.01). For NO2, the correlation was 0.15 (p = 0.47).

Fig. 5
figure 5

Temporally adjusted light scattering tertiles obtained during mobile monitoring in Ulaanbaatar on a February 24 and 25, 2010 and b February 26, 2010. Approximate PM2.5 concentrations for each tertile are given in parentheses. For comparison, the mobile data are superimposed on the modeled SO2 surface shown in Fig. 4b

Full size image

Mortality attributable to long-term air pollution exposure in Ulaanbaatar

There were 6,426 total deaths in Ulaanbaatar in 2009, including 1,885 (29.3%) from cardiovascular disease (ICD-10 codes I10–I70), 269 (4.2%) from respiratory causes other than lung cancer (ICD-10 codes J00–J99), and 117 (1.8%) from lung cancer (ICD-10 code C34). Among those 30 years or older, we estimated that 40% (95% CI, 17–56%) of lung cancer deaths and 29% (12–43%) of cardiopulmonary deaths in Ulaanbaatar are attributable to outdoor air pollution. These attributable fractions correspond to 578 (232–857) cardiopulmonary deaths and 45 (19–64) lung cancer deaths annually, or 9.7% of the total mortality in Ulaanbaatar, attributable to air pollution (Table 4). Calculations using alternative assumptions resulted in estimates that generally deviated from the base scenario by <50%. For example, a counterfactual concentration of 3.0 μg/m3 (instead of 7.5 μg/m3) increased the estimated number of attributable deaths by 27% (to 792, or 12.3% of total mortality).

Table 4 Estimates of mortality attributable to long-term PM2.5 air pollution exposure in Ulaanbaatar under different model scenarios
Full size table

Discussion

The annual average concentrations of PM10 and PM2.5 in central Ulaanbaatar (165.1 and 75.1 μg/m3, respectively) are approximately seven to eight times the WHO air quality guidelines of 20 and 10 μg/m3, respectively (Krzyzanowski and Cohen 2008). Although there is no WHO guideline for annual SO2 concentrations, the annual average concentration of 50.5 μg/m3 (17.7 ppb) in Ulaanbaatar far exceeds even the 24-h guideline of 20 μg/m3 (7 ppb). Importantly, the concentrations measured in the city center are considerably lower than those measured in one of the city’s ger areas.

These PM concentrations place Ulaanbaatar among the most polluted cities in the world (HEI 2004). For example, Ulaanbaatar’s annual average PM10 concentration of 165 μg/m3 is comparable to late 1990s levels (approximated as half the concentration of total suspended particles; Cohen et al. 2005) in megacities such as Kolkata, Delhi, and Beijing and exceeds levels in cities such as Mexico City and Buenos Aires (Gurjar et al. 2008). Despite its extraordinarily high air pollution concentrations, Ulaanbaatar has received very little research attention (Davy et al. 2011).

The high pollution concentrations are driven by conditions during winter, when 24-h PM2.5 concentrations frequently exceed 150 μg/m3 (and approach 250 μg/m3 in traditional housing areas) and SO2 levels are frequently above 80 μg/m3. The high PM2.5/PM10 ratios (≥0.65) in winter are comparable to previous wintertime measurements in polluted urban areas such as 0.71 in Beijing (Zhang et al. 2010) and 0.69 in Shanghai (Zhang et al. 2006) and suggest a major contribution from combustion-derived particles (Davy et al. 2011). The lower summertime PM2.5/PM10 ratios (≤0.35) are similar to observations in arid locations impacted by wind-blown dust (Eliasson et al. 2009) and suggest a relatively large influence of crustal particles. The diurnal PM2.5 patterns varied by season, with an evening peak in the winter that occurs later in the day and lasts longer than the evening peak in summer. The wintertime diurnal pattern in Ulaanbaatar is similar to developed country communities impacted by emissions from residential wood combustion (Robinson et al. 2007; Krecl et al. 2008).

Rapidly developing cities often have different urban designs and air pollution sources than cities in high-income regions, and few studies have attempted to characterize spatial patterns of air pollution in developing cities (Padhi and Padhy 2008; Dionisio et al. 2010; Etyemezian et al. 2005). We developed LUR models for NO2, a marker of traffic emissions, and SO2, a marker of coal combustion, in Ulaanbaatar. The model-based R 2 of our NO2 model was 0.74, which is within the wide range (0.51–0.97) reported in previous studies (Hoek et al. 2008). Our SO2 model-based R 2 of 0.78 is higher than the values (0.66 and 0.69) reported in the two previously published LUR SO2 models (Wheeler et al. 2008; Atari et al. 2008).

The vast majority of existing LUR models were developed in high-income countries (Hoek et al. 2008), and very few LUR models have been developed for Asian cities (Kashima et al. 2009; Chen et al. 2010). One challenge to LUR modeling in developing settings is the lack of data on spatial predictors, but satellite-based ETM+ data, which have global coverage and are freely available, show promise for overcoming this limitation. In our analysis, average greenness in a 1,000-m buffer explained 47% of the variance in NO2, while brightness in a 2,000-m buffer explained 55% of the SO2 variance. ETM+ may even be useful for LUR models in developed countries, as demonstrated by the inclusion of satellite-based soil brightness in recent LUR models for Los Angeles (Su et al. 2009).

A few limitations of our LUR models should be considered. First, the models are based on 37 observations, which is fewer than the 40–80 recommended in a recent LUR review (Hoek et al. 2008). Additional observations may have resulted in a model with more variables that captured additional complexity in the spatial patterns, particularly for NO2 due to its high spatial variability (Fig. 4b). Second, although ETM+ data from 2006 were the most recent available, they may have missed important recent land cover changes in this rapidly growing city. Nevertheless, the high correlations between the ETM+ ground cover classifications and both NO2 (greenness predicted 47% of the variance) and SO2 (brightness predicted 55% of the variance) indicate the value of these data for predicting spatial pollution gradients. Third, we were unable to calibrate our Ogawa measurements with government monitors, so our LUR surfaces can be used for assessing spatial patterns but the absolute concentrations may be inaccurate. Finally, LUR models are generally developed based on multiple sampling campaigns to assess long-term conditions (Hoek et al. 2008), but our models are based on a single monitoring session. While additional monitoring would more definitively characterize wintertime spatial patterns, we have demonstrated the feasibility of developing LUR models in a rapidly developing Asian city.

Spatially dense fixed-location monitoring is an expensive and logistically challenging way to capture within-city spatial variations in PM2.5, and mobile monitoring represents a promising alternative, particularly in developing countries with limited resources for environmental monitoring. For example, Dionisio et al. (2010) measured spatial PM2.5 patterns in Accra, Ghana by walking 7.7–9.4 km paths while recording PM2.5 and latitude/longitude at 1-min resolution. They identified nearby wood and charcoal stoves, congested and heavy traffic, loose dirt road surface, and trash burning as important PM2.5 sources. In Ulaanbaatar, we piloted a vehicle-based mobile nephelometer monitoring technique that was originally developed for capturing spatial patterns of wood smoke PM2.5 in North American cities (Larson et al. 2007; Su et al. 2008b). Although 10–20 evenings of monitoring may be needed to definitively identify spatial patterns (Larson et al. 2007; Su et al. 2008b), we captured a wide range of PM2.5 concentrations during three evenings of monitoring. In spite of the differences in technique, pollutant, and averaging time, the spatial patterns in PM2.5 identified by mobile monitoring and SO2 patterns identified by Ogawa measurements and an LUR model were generally similar (Fig. 5). The identification of PM2.5 and SO2 “hot spots” in the city’s ger areas is consistent with major emissions of these pollutants in these areas and with source apportionment results, suggesting that coal is the dominant source of PM2.5 in Ulaanbaatar (Davy et al. 2011). Moreover, the similarities between SO2 and PM2.5 spatial patterns (Fig. 5), agreement between modeled SO2 and measured PM2.5 at four government monitoring sites, and similarities between Ogawa SO2 measurements and mobile light scattering coefficient measurements suggest that our SO2 LUR model may provide a tool for PM2.5 exposure assessment in Ulaanbaatar, although this needs to be verified with additional PM2.5 monitoring.

Based on 2009 mortality statistics, we conservatively estimated that 623 deaths in Ulaanbaatar were attributable to air pollution. This represents 9.7% of the 6,426 total deaths in the city and, notably, 4.0% of the 15,522 annual deaths for the entire country. Calculations using alternative assumptions produced estimates that generally deviated from the base scenario by <50%. The exceptions were estimates based on a linear concentration–response relationship and no truncation, which were up to 94% higher than our base scenario estimate. These scenarios may be unrealistic (Ostro 2004), given evidence suggesting that PM2.5 mortality effects are nonlinear across the wide range of concentrations considered here (Pope et al. 2009).

Our estimate of attributable mortality probably underestimates the true public health burden of air pollution in Ulaanbaatar for several reasons (Kunzli et al. 2000, 2008). First, we did not consider the effects of indoor air pollution or outdoor pollutants other than PM2.5 (Anenberg et al. 2010; HEI 2004; Rylance et al. 2010). In addition, due primarily to data limitations, we did not consider nonmortality endpoints that have been linked to air pollution such as cardiovascular disease, impaired lung development, incident asthma, asthma exacerbations, bronchitis, hospitalizations, and school absences (Pope and Dockery 2006; Brook et al. 2010; Allen et al. 2009; Gauderman et al. 2004; Clark et al. 2010; Perez et al. 2009). We also considered mortality impacts only among those 30 years or older, thus excluding, for example, attributable infant mortality (Kaiser et al. 2004). Finally, we made the a priori decision to use PM2.5 data from a centrally located government monitoring site to estimate outdoor concentrations for the attributable mortality calculation. Our SO2 LUR model and mobile PM2.5 monitoring (Figs. 4b and 5) suggest that this site is located in a relatively unpolluted area of Ulaanbaatar. As a result, both outdoor concentrations and attributable mortality may be underestimated, especially given that half the city’s population lives in a ger (Asian Development Bank 2006), and the higher attributable mortality estimates (10.6–13.1% of total mortality) based on concentrations at site #2 may be more appropriate.

The strengths and weaknesses of quantitative impact assessment methods have been discussed extensively (Brunekreef et al. 2007; Perez and Kunzli 2009; Sahsuvaroglu and Jerrett 2007; O'Connell and Hurley 2009). Estimates of attributable mortality are often misinterpreted as “avoidable” deaths, but it is more appropriate to interpret these as estimates of “postponable” deaths (Brunekreef et al. 2007). Some have suggested that changes in life expectancy (calculated from life tables) are a better and more interpretable indicator of the mortality impacts of long-term air pollution exposure (Brunekreef et al. 2007; Perez and Kunzli 2009; Boldo et al. 2006). If age-specific population and death statistics can be obtained, changes in life expectancy can be calculated (e.g., WHO AirQ 2.2.3 or http://www.iom-world.org/research/iomlifet.php). Unfortunately, the mortality data provided by the Statistical Department at the Mongolian Government Implementing Agency/Department of Health were aggregated for those older than 65 years. Therefore, we restricted our impact assessment to the attributable mortality calculations. Despite its limitations, attributable mortality is a commonly used metric for impact assessment. For example, it also allows for a comparison of the 623 deaths attributable to air pollution annually in Ulaanbaatar with the mortality attributable to other risk factors such as suicide (199 deaths in 2009), transportation accidents (185), and homicide (179).

Conclusions

Due in part to challenges and limitations in population exposure assessment, few epidemiologic studies of air pollution and health have been conducted in developing countries. We successfully applied current, cost-effective exposure assessment techniques in Ulaanbaatar, Mongolia, one of the most polluted cities in the world, which suggests that these techniques are feasible in other rapidly developing cities. Based on satellite-based land cover and other predictors, we developed LUR models that identified strong spatial concentration gradients consistent with a major contribution from home heating in Ulaanbaatar’s low-income traditional housing (ger) areas. Temporal patterns, mobile PM2.5 monitoring, and PM2.5/PM10 ratios supported this finding. Air pollution represents a major threat to public health in Ulaanbaatar, and reductions in home heating emissions should be the primary focus of future air pollution control efforts.