- Email: [email protected]
Spatial and temporal patterns of particulate matter sources and pollution in four communities in Accra, Ghana
Science of the Total Environment 435–436 (2012) 107–114
Contents lists available at SciVerse ScienceDirect
Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv
Spatial and temporal patterns of particulate matter sources and pollution in four communities in Accra, Ghana Michael S. Rooney a, 1, Raphael E. Arku b, 1, Kathie L. Dionisio b, c, 1, Christopher Paciorek d, e, Ari B. Friedman f, Heather Carmichael g, Zheng Zhou b, c, Allison F. Hughes h, Jose Vallarino b, Samuel Agyei-Mensah i, John D. Spengler b, Majid Ezzati j, k,⁎ a
Harvard–MIT Division of Health Sciences and Technology, Cambridge, USA Department of Environmental Health, Harvard School of Public Health, Boston, USA c Department of Global Health and Population, Harvard School of Public Health, Boston, USA d Department of Biostatistics, Harvard School of Public Health, Boston, USA e Department of Statistics, University of California, Berkeley, USA f University of Pennsylvania, Philadelphia, USA g Harvard Medical School, Boston, USA h Department of Physics, University of Ghana, Legon, Ghana i Department of Geography and Resource Development, University of Ghana, Legon, Ghana j MRC–HPA Centre for Environment and Health, Imperial College London, London, UK k Department of Epidemiology and Biostatistics, School of Public Health, Imperial College London, London, UK b
H I G H L I G H T S ► ► ► ►
This is the first study of spatial patterns of PM levels and sources in multiple neighborhoods in a developing country city. We developed a novel spatio-temporal model to estimate pollution-source association, which can be used to predict PM levels. Wood stoves, fish smoking, trash burning, road capacity and surface were associated to higher PM2.5. Community SES was inversely associated with PM pollution.
a r t i c l e
i n f o
Article history: Received 3 November 2011 Received in revised form 18 April 2012 Accepted 22 June 2012 Available online 28 July 2012 Keywords: Air pollution Particulate matter Biomass Africa Geographic information system Spatial analysis
a b s t r a c t Sources of air pollution in developing country cities include transportation and industrial pollution, biomass fuel use, and re-suspended dust from unpaved roads. We examined the spatial patterns of particulate matter (PM) and its sources in four neighborhoods of varying socioeconomic status (SES) in Accra. PM data were from 1 week of morning and afternoon mobile and stationary air pollution measurements in each of the study neighborhoods. PM2.5 and PM10 were measured continuously, with matched GPS coordinates. Data on biomass fuel use were from the Ghana 2000 population and housing census and from a census of wood and charcoal stoves along the mobile monitoring paths. We analyzed the associations of PM with sources using a mixed-effects regression model accounting for temporal and spatial autocorrelation. After adjusting for other factors, the density of wood stoves, fish smoking, and trash burning along the mobile monitoring path as well as road capacity and surface were associated with higher PM2.5. Road capacity and road surface variables were also associated with PM10, but the association with biomass sources was weak or absent. While wood stoves and fish smoking were significant sources of air pollution, addressing them would require financial and physical access to alternative fuels for low-income households and communities. © 2012 Elsevier B.V. All rights reserved.
1. Introduction
⁎ Corresponding author at: MRC–HPA Centre for Environment and Health, Department of Epidemiology and Biostatistics, Imperial College London, Norfolk Place, London W2 1PG, UK. Tel.: +44 20 7594 0767; fax: +44 20 7594 3456. E-mail address: [email protected] (M. Ezzati). 1 These authors contributed equally to the research and manuscript. 0048-9697/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.scitotenv.2012.06.077
The spatial patterns of air pollution in cities in high-income countries are primarily related to distance from major traffic routes and industrial sources, and in some settings residential heating using biomass (Charron and Harrison, 2005; Hoek et al., 2001, 2002; Holmes et al., 2005; Levy et al., 2000, 2001; Su et al., 2008, 2007; Weijers et al., 2004). Sources of air pollution in developing country cities include transportation and
108
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
industrial pollution, biomass and coal fuel use for household and commercial purposes, and re-suspended dust from unpaved roads, with varying importance based on neighborhood location and socioeconomic status (SES) (Dionisio et al., 2010a). A few studies have examined spatial variability and sources of air pollution in developing country cities (Arku et al., 2008; Chowdhury, 2004; Dionisio et al., 2010a, 2010b; Engelbrecht et al., 2001; Etyemezian et al., 2005; Jackson, 2005; Padhi and Padhy, 2008; Saksena et al., 2003; van Vliet and Kinney, 2007; Zheng et al., 2005), but few have systematically examined variation in air pollution and its sources within neighborhoods, have analyzed the associations with sources, or have been in neighborhoods that span a range of SES. In previous work, we documented the within-neighborhood variation of particulate matter (PM) pollution and examined the effects of nearby sources on short-term local PM in four neighborhoods in Accra, Ghana (Dionisio et al., 2010b). This work could not be used to examine the association of the spatial patterns of PM with source variables that are collected only once, and hence predict pollution elsewhere in the city or even in other parts of the study neighborhood,
Nima (NM) Path length: 7.7 km
because it used real-time data on source status collected at the same time as PM measurements, e.g., whether there was heavy traffic at the time of measurement vs. data on road capacity. In this paper, we use new data on the distributions of sources in the same neighborhoods to estimate their associations with PM pollution over the full mobile monitoring path. There are three main contributions in this paper: First, to our knowledge this is one of the first studies of the spatial patterns of air pollution levels and of their sources in neighborhoods of varying SES in a developing country city. In particular, we present the first detailed data on the density of biomass stoves along roads in urban neighborhoods with different SES. Second, the unique data are used to estimate the pollution-source associations. In addition to helping understand the predictors of air pollution spatial patterns, these associations can be used, as a so-called “land-use regression” model, to predict PM pollution along roads and alleys in different neighborhoods. Third, we extend the methods for analysis of continuous mobile air pollution data to account for both spatial and temporal correlation of the measurements.
East Legon (EL) Path length: 9.4 km
Accra Fixed site Alley Local secondary road
% household using biomass
Connecting secondary road
0 - 40% 40 - 60%
Primary road
60 - 70% 70 - 80%
Divided multi-lane highway
> 80%
Asylum Down (AD) Path length: 8.9 km
Jamestown/Ushertown (JT) Path length: 8.5 km
Fig. 1. Map of Accra Metropolitan Area, study areas, and mobile monitoring path, delimited by census enumeration areas (EAs). EAs have nearly the same population; hence the area of an EA is inversely related to population density. Local secondary roads are smaller roads whose traffic is primarily for the purpose of reaching a local destination; connecting secondary roads are smaller roads used for passing through the neighborhood.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
2. Materials and methods 2.1. Study location Our study took place in four Accra neighborhoods: Jamestown/ Ushertown (JT), Asylum Down (AD), Nima (NM), and East Legon (EL) (Fig. 1). Detailed information on study neighborhoods are provided in previous work (Dionisio et al., 2010a, 2010b; Zhou et al., 2011). 2.2. Study design and data sources 2.2.1. PM data We conducted consecutive days of mobile monitoring in each neighborhood (7 days in most neighborhoods). The Accra air pollution study took place between 2006 and 2008; mobile measurements were done in April 2007 (before the main rainy season) in AD and in July–August 2007 (after the main rainy season) in the other neighborhoods. There were no unusual meteorological factors during the measurement periods. There were two monitoring tours on each day (one in the morning starting between 6:30 and 7:00, one in the afternoon starting at approximately 12:00). If it rained for more than approximately 30 min or if there was an equipment failure, the monitoring session was cancelled and repeated later so that the number of measurement days was as close as possible to 7. In each monitoring tour, we walked slowly along the full length of a pre-determined path, recording data at 1-minute intervals with a continuous real-time PM monitor and a GPS unit. The path for each neighborhood was designed to traverse different areas including main highways/roads, local roads, residential alleys and foot paths, and markets. The paths ranged 8–9 km in length, and the tours lasted 4.5–5.5 h in duration in different neighborhoods (Fig. 1). Further details on the measurement schedule are provided in a previous work (Dionisio et al., 2010b). We measured 48-hour integrated PM2.5 and PM10 concentrations at 2–3 roof-top fixed sites in each neighborhood using gravimetric methods. One fixed site in each neighborhood was located along a main road with the others located in residential areas. We also measured PM2.5 and PM10 continuously at as many fixed sites as possible. Further details on measurements at fixed sites are provided elsewhere (Dionisio et al., 2010a). Continuous PM was measured using DustTrak Model 8520 and SidePak AM510 monitors (TSI Inc., Shoreview, MN), and corrected against gravimetric PM as described in detail in a previous work (Dionisio et al., 2010a, 2010b). Using the timestamps of DustTraks or SidePaks, error-corrected continuous PM2.5 and PM10 data from fixed sites and mobile monitors were compiled into a single dataset with each record representing a unique date, minute, and location. Geographic coordinates for each mobile monitor data point were matched to PM data using the GPS date/timestamp. Because GPS units measure true location with error, GPS coordinates were “snapped” to the nearest point on the monitoring path; the location of the path was ascertained using a Trimble GeoXT GPS unit (Trimble Navigation Limited, Sunnyvale, CA) with a nominal error of less than 1 m. When a point was at or near the intersection of two roads that were both on the path, the snapping retained the temporal ordering of data points. 2.2.2. Source data along and near the mobile monitoring path We conducted a geo-coded census of wood and charcoal stoves along each of the four mobile monitoring paths. Traversing each path, we used a GPS unit to record the location of all stoves and stove clusters as well as other large PM sources (e.g. bakeries, fish smoking, trash burning). We counted all stoves visible from the path and those that could be identified by visible smoke, and recorded the following information: number of stoves in a cluster, stove type(s) (charcoal or firewood), stove size (small or large), and on/off status. We asked a set of pre-specified questions from the attendants about the days of a week and times of a day (divided into 6:00–10:00, 10:00–14:00, 14:00–17:00, 17:00–19:00 and 19:00–22:00) when the stove was used; the questions were asked of
109
every third to fifth attendant in AD, the tenth attendant in JT and NM, and all consenting attendants in EL due to varying density of sources across neighborhoods. We conducted 6–8 rounds of source count in JT, NM, and EL and only one round in AD because the field team was smaller at the time of AD campaign. Questions were not repeated to attendants previously interviewed. We used a Trimble GeoXT GPS to measure the location of the mobile monitoring path and to mark its various segments based on (i) road capacity, categorized by alleys, local secondary roads, connecting secondary roads, primary roads, and multi-lane highways and (ii) road surface, categorized by paved, paved broken, packed dirt, and loose dirt. We used a geo-coded road map of Accra to measure the distance from each point along the monitoring path to the primary roads or multi-lane highways. 2.2.3. Community household fuel use and SES In addition to the source data along the monitoring path, we used a 10% sample of the Ghana 2000 population and housing census to calculate the following variables in each census enumeration area (EA) traversed by the mobile monitoring path: • density of households per hectare that used charcoal or wood for cooking; • density of households per hectare that burned their household trash as primary means of disposal; and • community SES index. There were 55 EAs along the path in JT, 22 in AD, 34 in NM, and 3 in EL. Following previous analyses of household data in developing countries (Gakidou et al., 2007; Wagstaff, 2000), we measured SES using an index based on type and tenure of dwelling, materials of outer walls, floor, and roof, toilet and bathing facilities, solid and liquid waste disposal methods, water source, and the number of persons per room and per bedroom; the SES index excluded household fuel type (Zhou et al., 2011). To calculate the SES index, we used principal component analysis (PCA), a statistical procedure that transforms a number of possibly correlated variables into a smaller number of uncorrelated variables, the principal components. We used the first principal component as the SES index, because it is a linear combination of the individual variables that explains the maximum variance (29%) across the households. The PCA was conducted using data from individual census records (each corresponding to one household), with household scores averaged to obtain EA SES. 2.2.4. Meteorological/weather variables Data on relative humidity (RH), wind speed and time since last rain were from a station near Accra's Kotoka International Airport (Fig. 1), maintained by the United States Department of Commerce National Oceanic and Atmospheric Administration (NOAA). We predicted RH and wind speed for hours with missing data using a simple linear interpolation when data were missing for less than 3 h. When more than 3 hours of data were missing, we used the average of RH for the same hour over 5 days before and 5 days after the missing value. We then fitted a cubic spline function to hourly RH and wind speed to obtain RH and wind speed values for each minute (Dionisio et al., 2010a). 2.3. Statistical analysis 2.3.1. Stove and other source density along the mobile monitoring path We used the stove count data to calculate a “wood stove density” and a “charcoal stove density” associated with each observation in the mobile monitoring PM dataset. Using the interview data, we assigned a probability of 0, 0.5, or 1.0 to the stove being on for each day of the week and time interval, with 0.5 indicating self-reported variable status at that time/day. We assigned probabilities to stoves without an interview using multiple imputation (R package AMELIA; available at http://gking.harvard.edu/amelia/) (Honaker, 2010) by creating 300 imputed datasets and using their average for each stove. Stove
110
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
density for each PM record depended on its location along the path and the time of measurement and was calculated in relation to the location of all stoves and the time that they were on as described below. Each stove or stove cluster whose probability of being on was non-zero contributed to the density, with its effect declining with distance. The decline in the effect of each stove was modeled using a multivariate normal density, with a standard deviation of 10 m. Studies in rural areas have shown that PM pollution from biomass stoves drops sharply within about this distance (Ezzati et al., 2000). Stove ∑i N ðdi j0;102 IÞci P t P d
i i density was calculated as , where N(·) is the multivars iate normal density; di is the longitudinal and latitudinal displacement between the mobile monitor and stove or stove cluster i in meters; ci is the number of stoves at stove cluster i (large stoves being counted doubly); P(ti) and P(di) are, respectively, the probabilities that source i would be on during the hour of the day and day of the week the measurement was taken; s is the number of times we counted sources along the path in each neighborhood. This metric is equivalent to the number of stoves that are in the proximity of the measurement point because it accounts for distances to various stove clusters. We calculated similar metrics for other sources along the walking path. Source densities were scaled to stoves per hectare.
2.3.2. Association of PM with sources We used regression analysis to quantify the association of sources and meteorological factors with PM along the mobile monitoring path. The regression model was specified by yi = Xiβ + bi + εi, where yi is the dependent variable vector (natural logarithm of PM concentrations) for neighborhood-day i, β is a vector of fixed-effect coefficients for the covariates plus the intercept term, Xi is the matrix of covariates associated with all PM records, bi is the neighborhood-day random intercept (Laird and Ware, 1982), and εi is the vector of within-neighborhoodday errors. We used the following model covariates: household wood and charcoal use densities in the EA where the measurement was carried out; stove density along the mobile monitoring path, calculated as above; road traffic capacity at the measurement location; distance to the closest major road; wind speed; road surface type; relative humidity; and time since the last precipitation. In addition to these covariates, we conducted the analysis both with and without adjusting for the ambient neighborhood PM, calculated as the average of concurrent PM measured at the fixed monitoring stations in each neighborhood (smoothed to remove perturbations lasting less than 30 min). PM concentrations were log-transformed to ensure that model residuals were normally distributed. For consistency with this transformation, we also log-transformed the density of sources along the walking path and the EA fuel use density. Residual diagnostics suggest the relationships between these variables and the outcome were reasonably linear when they were log-transformed but that, for a subset of them, the relationship was nonlinear without the log transformation. The neighborhood-day random intercept accounts for unobserved factors that may have affected PM for each neighborhood-day, such as local meteorological factors, day-to-day variations in economic activity, or even imperfect instrument calibration. In addition, because PM observations were taken at 1 min intervals during which the monitor moves for a short distance only, there may be both spatial and temporal correlation between successive observations within a single neighborhood-day. To account for this, we modeled the within neighborhood-day error (εi) as the sum of empirically-estimated spatially and temporally correlated components as well as an independent component. By using the sum of spatially and temporally correlated components, the spatial and temporal components are modeled as independent; i.e., the temporal components of the within-neighborhood-day variance are similar over space, and vice versa. This produces a spatio-temporal covariance that is additive in the temporal and spatial components. The temporal and spatial autocorrelation were modeled as exponential, with the parameters estimated empirically with other model parameters.
The model was fitted by maximizing the restricted log likelihood using the function lme() in the R package nlme. 3. Results Gravimetric-corrected ground-level PM varied substantially within each neighborhood, reaching as high as 200 μg/m3 for PM2.5 and 400 μg/m 3 for PM10 (Fig. 2) (Dionisio et al., 2010a). In general, PM concentrations increased with road capacity and were highest along the divided multi-lane highway, with median PM2.5 and PM10 of 53.4 and 144.5 μg/m 3, respectively (Fig. 3). PM2.5 concentrations along alleys were higher than smaller local secondary roads with some traffic. The density of charcoal stoves used along the walking path in JT and NM was about 7 and 20 times those of AD and EL, respectively (Fig. 4). Woodstoves were virtually absent along the mobile monitoring path in EL and were most common in JT, about 5 times more common than in NM and 10 times that of AD. There was little variation in stove use among days of the week, except on Sundays, when fewer stoves were in use in AD and JT. Stoves were used most in the morning, with about two thirds of all surveyed stoves used before 10:00, about half between 10:00 and 17:00, declining to about 10% after 19:00. This is consistent with demand for and sale of street food. In multivariate analysis, PM2.5 and PM10 levels increased with road capacity (Table 1). After adjustment for neighborhood average, divided multi-lane highways had 56% (95% CI 36–78) higher PM2.5 and 74% (95% CI 53–98) higher PM10 than alleys. PM2.5 and PM10 were highest on roads with a loose dirt surface, by 14–22% for different size fractions and in models with and without adjustment; the next highest pollution was on roads with a broken paved surface. Across the different models, PM levels decreased with distance from a main road; at 50 m from a main road, PM2.5 dropped by approximately 13–15% and PM10 by 14–16%. Community SES was inversely associated with PM pollution, with PM2.5 and PM10, respectively, 34% and 20% higher in the lowest SES neighborhoods in our sample than in the highest SES one. Non-traffic combustion sources generally did not have a significant association with PM10 along the walking path at p = 0.05, with the quantity of woodstoves, trash burning, and fish smoking along the walking path approaching this threshold. The seemingly negative association between community fuel use and PM10 in the model without adjustment for the neighborhood average is unexplained. The pair-wise correlations between the density of sources along the walking path (from source count) and community fuel use (from the census) were generally small and ranged between −0.19 and 0.56. Removing the density of sources along the walking path removed the negative associations for community charcoal use, making the coefficient null. Density of nearby woodstoves along the walking path was significantly associated with higher PM2.5. An additional 25 nearby woodstoves per hectare (the 95th percentile of densities in our data after accounting for the fact that pollution drops sharply with distance) had an effect of about 9-10% higher PM2.5 than an area without woodstoves in the two models. After adjusting for community SES and other variables, community fuel use had no significant association with PM2.5. Among larger sources, fish smoking and trash burning had a noticeable and statistically significant effect on PM2.5 levels along the walking path, with 37% (95% CI 17–61) and 27% (95% CI 1.0–59) higher PM2.5 for increasing the number of nearby fish smoking or trash burning spots to 5 per hectare, respectively, in the model without adjustment for the neighborhood average. An increase in RH from 25% to 75% was associated with 215% higher PM2.5 and 126% higher PM10. In the model without adjustment for the neighborhood average, increased wind speed was inversely associated with PM; compared to stagnant conditions, PM2.5 and PM10 levels were expected to be 33% (95% CI 26–39) and 28% (95% CI 20–35) lower, respectively, with a 10 mph breeze. Rain also had a mitigating effect on PM10.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
800
600
400
200
0
1000
Distance (meters) 800
600
400
200
0
1600
1800
Distance (meters) 1400
1200
1000
800
600
400
200
0
1200
Distance (meters) 1000
800
600
400
200
0
Distance (meters)
111
PM2.5 (µg/m3)
PM2.5
N=3,602
N=3,942
N=3,875
N=3,791
< 10 10 − 20 20 − 30 30 − 40 40 − 50 50 − 60 60 − 70 70 − 80 80 − 90 90 − 100 100 − 150 150 − 200 > 200 PM10 (µg/m3)
PM10
N=3,599
N=4,209
N=3,240
N=3,806
Asylum Down (AD)
East Legon (EL)
Jamestown (JT)
Nima (NM)
< 20 20 − 40 40 − 60 60 − 80 80 − 100 100 − 120 120 − 140 140 − 160 160 − 180 180 − 200 200 − 300 300 − 400 > 400
Fig. 2. Concentrations of PM2.5 and PM10 along the walking path in the study neighborhoods. For each neighborhood and PM size fraction, data from all monitoring days/tours were combined in a moving average, with a 50 m averaging interval. Sample sizes show the total number of minute-by-minute measurements in the neighborhood. Reproduced with permission from Environmental Health Perspectives (Dionisio et al., 2010a).
The model accounted for 63% of PM2.5 and 72% of PM10 variability when neighborhood average was included, and for 41% and 33% when neighborhood average was not included, indicating that its predictions
1000 N=3270
N=489
PM2.5 (µg/m3)
N=5840
N=3469
N=2136
100
are substantially improved in the presence of fixed site monitors. Note that part of the unexplained variance includes daily fluctuations (modeled as a random intercept term), which may or may not be of interest in understanding location-specific PM exposures. Model residuals from observations made at the same time and location were approximately 90% correlated, and it was estimated that about 70% of this effect was due to time and 20% due to space. The autocorrelative effects of time and space dropped to 5% of their maximal values at lags of 13 min and 100 m for PM2.5, and 14 min and 190 m for PM10 (unadjusted model). These estimates provide guidance on appropriate lags for collecting statistically independent samples in an urban environment. Autocorrelation seemed to occur in both time and space, with more temporal effect. Specifically, about 70% and 65% of variance of PM2.5 and PM10 model error terms were, respectively, explained by the temporal component and 20% by the spatial one.
10
4. Discussion
1000
N=489 N=3017
N=2113
PM10 (µg/m3)
N=5684
N=3384
100
10
Alley
Local Sec. Connect. Sec. Primary Div. Multi−Lane
Fig. 3. Distributions of PM2.5 and PM10 by road capacity. Sample sizes show the total number of minute-by-minute measurements along that road type.
In previous work, we used data from a mobile monitoring campaign to show that PM concentrations had significant spatial variability within and between neighborhoods in Accra (Dionisio et al., 2010b). The low-SES neighborhood of JT had the highest pollution, followed by segments of the mobile monitoring path along primary roads in other neighborhoods. Our previous work also demonstrated that burning woodstoves and congested or heavy traffic had the strongest effects on real-time local PM concentration at approximately pre-selected points along the walking path, but did not analyze the remaining parts of the paths (Dionisio et al., 2010b). In the current work, we investigated the associations between PM concentrations and sources along the complete walking path in these four neighborhoods, with source data collected once and hence reflecting “usual” vs. real-time conditions. Specifically, we used the population and housing census to obtain data on biomass use in small areas (census EAs) and conducted our own census of PM sources along the monitoring path. Except for road surface, which is time-invariant, comparison with the previous work is not possible due to the differences between
112 M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
Fig. 4. Locations of stoves along and near the walking path, and stove use by day of week and time of day.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
113
Table 1 Model coefficients for multivariate analysis of the association of PM with biomass sources, road capacity and surface, and meteorological covariates. Ln denotes natural logarithm. Dependent variable: Ln(PM2.5)
Constant Ln(neighborhood average) Ln(distance to nearest main road (m)) Community SES Density of biomass stoves and other sources along walking path (per hectare) Ln(charcoal stove density) Ln(wood stove density) Ln(bakery density) Ln(fish smoking density) Ln(flour mill density) Ln(trash burning density) Density of community sources from the census (per hectare) Ln(wood stove density) Ln(charcoal stove density) Ln(trash burning density) Road typea Alley Local secondary roads Connecting secondary roads Primary roads Multi-lane highways Road surfacea Paved Paved broken Packed dirt Loose dirt Meteorological factors Wind speed (mph) Relative humidity Ln(hours since rain)
Model 1
Model 2
Coefficient
p-Value
Coefficient
p-Value
1.194 (0.596, 1.793)
b0.001
−0.036 (−0.053, –0.019) −0.041 (−0.060, –0.022)
b0.001 b0.001
−0.042 (−0.567, 0.482) 0.568 (0.503, 0.633) −0.041 (−0.058, –0.024) −0.040 (−0.059, –0.021)
0.875 b0.001 b0.001 b0.001
−0.003 (−0.008, 0.003) 0.013(0.005, 0.020) −0.018 (−0.045, 0.008) 0.051 (0.026, 0.076) −0.004 (−0.031, 0.023) 0.038 (0.002, 0.074)
0.330 b0.001 0.175 b0.001 0.780 0.041
−0.002 (−0.007, 0.004) 0.012(0.004, 0.019) −0.023 (−0.049, 0.002) 0.040 (0.015, 0.064) −0.004 (−0.030, 0.022) 0.033 (−0.002, 0.068)
0.552 0.002 0.073 0.002 0.776 0.065
−0.003 (−0.009, 0.003) −0.015 (−0.033, 0.003) 0.002 (−0.003, 0.008)
0.358 0.107 0.405 0.003
−0.002 (−0.008, 0.004) −0.010 (−0.028, 0.007) 0.003 (−0.003, 0.009)
0.510 0.250 0.308 b0.001
0 0.032 0.089 0.128 0.390
0 0.033 (−0.008, 0.073) 0.088 (0.036, 0.140) 0.150 (0.085, 0.214) 0.444(0.309, 0.579)
(−0.009, 0.073) (0.036, 0.142) (0.062, 0.193) (0.246, 0.535) b0.001
0 0.054 (−0.006, 0.114) 0.022 (−0.017, 0.060) 0.199 (0.137, 0.260)
b0.001 0 0.047(−0.012, 0.105) 0.017 (−0.021, 0.055) 0.199(0.138, 0.260)
−0.040 (−0.050, −0.030) 3.591(3.155, 4.026) 0.079 (−0.028, 0.186)
b0.001 b0.001 0.147
−0.009 (−0.018, 0.001) 2.295(1.900, 2.691) 0.047(−0.042, 0.135)
0.067 b0.001 0.303
2.039 (1.433, 2.646)
b0.001
−0.038 (−0.055, −0.022) −0.027 (−0.046, −0.009)
b0.001 0.004
0.772 (0.327, 1.218) 0.550 (0.492, 0.608) −0.045 (−0.061, −0.029) −0.025 (−0.043, −0.007)
b0.001 b0.001 b0.001 0.006
Dependent variable: Ln(PM10) Constant Ln(neighborhood average) Ln(distance to nearest main road (m)) Community SES Density of biomass stoves and other sources along walking path (per hectare) Ln(charcoal stove density) Ln(wood stove density) Ln (bakery density) Ln(fish smoking density) Ln(flour mill density) Ln(trash burning density) Density of community sources from the census (per hectare) Ln(wood stove density) Ln(charcoal stove density) Ln(trash burning density) Road typea Alley Local secondary roads Connecting secondary roads Primary roads Multi-lane highways Road surfacea Paved Paved broken Packed dirt Loose dirt Meteorological factors Wind speed (mph) Relative humidity Ln(hours since rain)
−0.002 (−0.007, 0.004) −0.007 (0.000, 0.013) 0.004 (−0.023, 0.030) 0.021 (−0.007, 0.048) −0.011 (−0.038, 0.015) 0.026 (−0.012, 0.063)
0.545 0.057 0.797 0.138 0.403 0.179
0.001 (−0.006, 0.005) 0.005 (−0.002, 0.012) 0.003 (−0.024, 0.029) 0.012 (−0.015, 0.038) −0.008 (−0.033, 0.018) 0.022 (−0.015, 0.059)
0.840 0.154 0.845 0.398 0.565 0.249
−0.006 (−0.012, 0.000) −0.018 (−0.036, 0.001) 0.000 (−0.005, 0.006)
0.036 0.044 0.930 b0.001
−0.006 (−0.012, 0.000) −0.003 (−0.018, 0.013) 0.001 (−0.005, 0.006)
0.067 0.716 0.734 b0.001
0 0.028 0.101 0.135 0.513
0 0.024 0.098 0.158 0.551
(−0.012, 0.068) (0.050, 0.152) (0.071, 0.198) (0.374, 0.653)
(−0.016, 0.063) (0.047, 0.148) (0.096, 0.221) (0.422, 0.681) b0.001
0.026 0 0.103 (0.045, 0.162) 0.014 (−0.023, 0.050) 0.129 (0.070, 0.187) −0.032 (−0.043, −0.022) 3.259 (2.827, 3.692) 0.104 (−0.004, 0.211)
0 0.102 (0.044, 0.160) 0.008 (−0.029, 0.044) 0.129 (0.071, 0.188) b0.001 b0.001 0.059
−0.006 (−0.014, 0.003) 1.628 (1.251, 2.006) 0.098 (0.037, 0.160)
0.217 b0.001 0.002
Because dependent variable is Ln(PM), [exp(coefficient) − 1] is the percent increase/decrease in PM due to each variable. For small coefficients, those whose absolute value is up to about 0.25, this value is approximately equal to the coefficient itself. Model 2 is adjusted for neighborhood average (estimated as average of smoothed concentrations at all fixed sites) and Model 1 is without this variable. See Materials and methods section for details. a p-Values are given for the complete categorical variable using an F-test. The p-value for the categorical variable tests whether the inclusion of the variable in the model is significant.
real-time and usual data, e.g. current presence of a burning stove vs. usual stove density or current heavy traffic vs. road capacity. The limitations of our study include the relatively short measurement period which does not allow for assessing how spatial patterns
of PM pollution and association with sources vary over seasons; the absence of evening and night-time data; the lack of census data concurrent with our study; potential error in self-reported days and times of stove use; and the use of DustTraks whose measurements
114
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
are subject to error which may have been only partially addressed by our correction factors. Finally, while some of the predictors needed to apply our results to other neighborhoods are available from routine sources like the census, source data along the path requires additional data collection. Despite the apparent complexity, such data can be collected once at a cost that is substantially lower than many other emission inventories routinely used in air pollution studies such as those of residential woodstove use in high-income countries. Despite these limitations that are common to many field research studies due to logistical difficulties, this unique combination of data showed that PM concentrations along the mobile monitoring path had a stronger association with sources along the path – especially with the presence of wood stoves, fish smoking, and trash burning – than they did with census biomass variables. The associations with sources along the walking path were stronger for PM2.5 than for PM10. Our findings on the absence of an association with census variables are perhaps different from those in the Greater Vancouver area (Canada), where residential woodstove use was a predictor of PM pollution (Larson et al., 2007). The reasons for this difference need further investigation and may include the following: first, there were many biomass sources along the walking path in Accra (Fig. 4) used for cooking street food, which is rare in high-income settings. These nearby sources may have had a more important role in local concentrations than the density of household biomass use due to their proximity. Second, we had data on hours of stove use for sources along the path but not for those from the census; thus our path source density was temporally variable but those for EAs from the census were not. This may have increased the relative predictive power of the density of sources along the path, because the PM data were also time varying. We also found associations between PM pollution and road capacity, consistent with studies in high-income countries (Levy et al., 2001; Weijers et al., 2004) and those in developing countries that had found higher PM along major traffic routes than in non-traffic areas (Padhi and Padhy, 2008; van Vliet and Kinney, 2007). None of these studies had accounted for non-traffic combustion sources.
5. Conclusions We found that, after adjusting for other factors, the density of wood stoves, fish smoking, and trash burning as well as road capacity and surface were associated with PM2.5 in Accra neighborhoods. Trash burning occurs because there are few places for safe disposal of solid waste and because the waste piles are removed too infrequently by the city authorities. The wood stoves along the walking path are used for cooking and selling street food, often by members of poor households (Zhou et al., 2011); fish smoking is also a source of income as a small business, for its owners and workers. While both were sources of air pollution, addressing them would require financial and physical access to alternative fuels for low-income households and communities (Bailis et al., 2005; Barnes et al., 2005; Zhou et al., 2011). Acknowledgment Funding for this research was provided by the National Science Foundation (grant 0527536). Laboratory support was provided by the Harvard NIEHS Center for Environmental Health. We thank Nana Prempeh and Adam Abdul Fatah for field assistance, and the Legal Resources Center and the Department of Geography and Resource Development at the University of Ghana for valuable help with logistical arrangements.
References Arku RE, Vallarino J, Dionisio KL, Willis R, Choi H, Wilson JG, et al. Characterizing air pollution in two low-income neighborhoods in Accra, Ghana. Sci Total Environ 2008;402:217–31. Bailis R, Ezzati M, Kammen DM. Mortality and greenhouse gas impacts of biomass and petroleum energy futures in Africa. Science 2005;308:98-103. Barnes DF, Krutilla K, Hyde WF. The urban household energy transition : social and environmental impacts in the developing world. Washington, DC: RFF Press; 2005. Charron A, Harrison RM. Fine (PM2.5) and coarse (PM2.5–10) particulate matter on a heavily trafficked London highway: sources and processes. Environ Sci Technol 2005;39:7768–76. Chowdhury MZ. Characterization of fine particle air pollution in the Indian subcontinent. Atmospheric Chemistry: Georgia Institute of Technology; 2004. Dionisio KL, Arku RE, Hughes AF, Vallarino J, Carmichael H, Spengler JD, et al. Air pollution in Accra neighborhoods: spatial, socioeconomic, and temporal patterns. Environ Sci Technol 2010a;44:2270–6. Dionisio KL, Rooney MS, Arku RE, Friedman AB, Hughes AF, Vallarino J, et al. Within-neighborhood patterns and sources of particle pollution: mobile monitoring and geographic information system analysis in four communities in Accra, Ghana. Environ Health Perspect 2010b;118:607–13. Engelbrecht JP, Swanepoel L, Chow JC, Watson JG, Egami RT. PM2.5 and PM10 concentrations from the Qalabotjha low-smoke fuels macro-scale experiment in South Africa. Environ Monit Assess 2001;69:1-15. Etyemezian V, Tesfaye M, Yimer A, Chow JC, Mesfin D, Nega T, et al. Results from a pilot-scale air quality study in Addis Ababa, Ethiopia. Atmos Environ 2005;39: 7849–60. Ezzati M, Saleh H, Kammen DM. The contributions of emissions and spatial microenvironments to exposure to indoor air pollution from biomass combustion in Kenya. Environ Health Perspect 2000;108:833–9. Gakidou E, Oza S, Vidal Fuertes C, Li AY, Lee DK, Sousa A, et al. Improving child survival through environmental and nutritional interventions: the importance of targeting interventions toward the poor. JAMA 2007;298:1876–87. Hoek G, Brunekreef B, Fischer P, van Wijnen J. The association between air pollution and heart failure, arrhythmia, embolism, thrombosis, and other cardiovascular causes of death in a time series study. Epidemiology 2001;12:355–7. Hoek G, Brunekreef B, Goldbohm S, Fischer P, van den Brandt PA. Association between mortality and indicators of traffic-related air pollution in The Netherlands: a cohort study. Lancet 2002;360:1203–9. Holmes NS, Morawska L, Mengersen K, Jayaratne ER. Spatial distribution of submicrometre particles and CO in an urban microscale environment. Atmos Environ 2005;39:3977–88. Honaker J. What to do about missing values in time series cross-section data. Am J Pol Sci 2010;54:561–81. Jackson MM. Roadside concentration of gaseous and particulate matter pollutants and risk assessment in Dar-es-Salaam, Tanzania. Environ Monit Assess 2005;104: 385–407. Laird NM, Ware JH. Random-effects models for longitudinal data. Biometrics 1982;38: 963–74. Larson T, Su J, Baribeau AM, Buzzelli M, Setton E, Brauer M. A spatial model of urban winter woodsmoke concentrations. Environ Sci Technol 2007;41:2429–36. Levy JI, Houseman EA, Ryan L, Richardson D, Spengler JD. Particle concentrations in urban microenvironments. Environ Health Perspect 2000;108:1051–7. Levy JI, Houseman EA, Spengler JD, Loh P, Ryan L. Fine particulate matter and polycyclic aromatic hydrocarbon concentration patterns in Roxbury, Massachusetts: a community-based GIS analysis. Environ Health Perspect 2001;109:341–7. Padhi BK, Padhy PK. Assessment of intra-urban variability in outdoor air quality and its health risks. Inhal Toxicol 2008;20:973–9. Saksena S, Singh PB, Prasad RK, Prasad R, Malhotra P, Joshi V, et al. Exposure of infants to outdoor and indoor air pollution in low-income urban areas — a case study of Delhi. J Expo Anal Environ Epidemiol 2003;13:219. Su JG, Larson T, Baribeau AM, Brauer M, Rensing M, Buzzelli M. Spatial modeling for air pollution monitoring network design: example of residential woodsmoke. J Air Waste Manag Assoc 2007;57:893–900. Su JG, Brauer M, Ainslie B, Steyn D, Larson T, Buzzelli M. An innovative land use regression model incorporating meteorology for exposure analysis. Sci Total Environ 2008;390:520–9. van Vliet E, Kinney P. Impacts of roadway emissions on urban particulate matter concentrations in sub-Saharan Africa: new evidence from Nairobi, Kenya. Environ Res Lett 2007:2. Wagstaff A. Socioeconomic inequalities in child mortality: comparisons across nine developing countries. Bull World Health Organ 2000;78:19–29. Weijers EP, Khlystov AY, Kos GPA, Erisman JW. Variability of particulate matter concentrations along roads and motorways determined by a moving measurement unit. Atmos Environ 2004;38:2993–3002. Zheng M, Salmon LG, Schauer JJ, Zeng L, Kiang CS, Zhang Y, et al. Seasonal trends in PM2.5 source contributions in Beijing, China. Atmos Environ 2005;39:3967–76. Zhou Z, Dionisio KL, Arku RE, Quaye A, Hughes AF, Vallarino J, et al. Household and community poverty, biomass use, and air pollution in Accra, Ghana. Proc Natl Acad Sci U S A 2011;108:11028–33.
Contents lists available at SciVerse ScienceDirect
Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv
Spatial and temporal patterns of particulate matter sources and pollution in four communities in Accra, Ghana Michael S. Rooney a, 1, Raphael E. Arku b, 1, Kathie L. Dionisio b, c, 1, Christopher Paciorek d, e, Ari B. Friedman f, Heather Carmichael g, Zheng Zhou b, c, Allison F. Hughes h, Jose Vallarino b, Samuel Agyei-Mensah i, John D. Spengler b, Majid Ezzati j, k,⁎ a
Harvard–MIT Division of Health Sciences and Technology, Cambridge, USA Department of Environmental Health, Harvard School of Public Health, Boston, USA c Department of Global Health and Population, Harvard School of Public Health, Boston, USA d Department of Biostatistics, Harvard School of Public Health, Boston, USA e Department of Statistics, University of California, Berkeley, USA f University of Pennsylvania, Philadelphia, USA g Harvard Medical School, Boston, USA h Department of Physics, University of Ghana, Legon, Ghana i Department of Geography and Resource Development, University of Ghana, Legon, Ghana j MRC–HPA Centre for Environment and Health, Imperial College London, London, UK k Department of Epidemiology and Biostatistics, School of Public Health, Imperial College London, London, UK b
H I G H L I G H T S ► ► ► ►
This is the first study of spatial patterns of PM levels and sources in multiple neighborhoods in a developing country city. We developed a novel spatio-temporal model to estimate pollution-source association, which can be used to predict PM levels. Wood stoves, fish smoking, trash burning, road capacity and surface were associated to higher PM2.5. Community SES was inversely associated with PM pollution.
a r t i c l e
i n f o
Article history: Received 3 November 2011 Received in revised form 18 April 2012 Accepted 22 June 2012 Available online 28 July 2012 Keywords: Air pollution Particulate matter Biomass Africa Geographic information system Spatial analysis
a b s t r a c t Sources of air pollution in developing country cities include transportation and industrial pollution, biomass fuel use, and re-suspended dust from unpaved roads. We examined the spatial patterns of particulate matter (PM) and its sources in four neighborhoods of varying socioeconomic status (SES) in Accra. PM data were from 1 week of morning and afternoon mobile and stationary air pollution measurements in each of the study neighborhoods. PM2.5 and PM10 were measured continuously, with matched GPS coordinates. Data on biomass fuel use were from the Ghana 2000 population and housing census and from a census of wood and charcoal stoves along the mobile monitoring paths. We analyzed the associations of PM with sources using a mixed-effects regression model accounting for temporal and spatial autocorrelation. After adjusting for other factors, the density of wood stoves, fish smoking, and trash burning along the mobile monitoring path as well as road capacity and surface were associated with higher PM2.5. Road capacity and road surface variables were also associated with PM10, but the association with biomass sources was weak or absent. While wood stoves and fish smoking were significant sources of air pollution, addressing them would require financial and physical access to alternative fuels for low-income households and communities. © 2012 Elsevier B.V. All rights reserved.
1. Introduction
⁎ Corresponding author at: MRC–HPA Centre for Environment and Health, Department of Epidemiology and Biostatistics, Imperial College London, Norfolk Place, London W2 1PG, UK. Tel.: +44 20 7594 0767; fax: +44 20 7594 3456. E-mail address: [email protected] (M. Ezzati). 1 These authors contributed equally to the research and manuscript. 0048-9697/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.scitotenv.2012.06.077
The spatial patterns of air pollution in cities in high-income countries are primarily related to distance from major traffic routes and industrial sources, and in some settings residential heating using biomass (Charron and Harrison, 2005; Hoek et al., 2001, 2002; Holmes et al., 2005; Levy et al., 2000, 2001; Su et al., 2008, 2007; Weijers et al., 2004). Sources of air pollution in developing country cities include transportation and
108
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
industrial pollution, biomass and coal fuel use for household and commercial purposes, and re-suspended dust from unpaved roads, with varying importance based on neighborhood location and socioeconomic status (SES) (Dionisio et al., 2010a). A few studies have examined spatial variability and sources of air pollution in developing country cities (Arku et al., 2008; Chowdhury, 2004; Dionisio et al., 2010a, 2010b; Engelbrecht et al., 2001; Etyemezian et al., 2005; Jackson, 2005; Padhi and Padhy, 2008; Saksena et al., 2003; van Vliet and Kinney, 2007; Zheng et al., 2005), but few have systematically examined variation in air pollution and its sources within neighborhoods, have analyzed the associations with sources, or have been in neighborhoods that span a range of SES. In previous work, we documented the within-neighborhood variation of particulate matter (PM) pollution and examined the effects of nearby sources on short-term local PM in four neighborhoods in Accra, Ghana (Dionisio et al., 2010b). This work could not be used to examine the association of the spatial patterns of PM with source variables that are collected only once, and hence predict pollution elsewhere in the city or even in other parts of the study neighborhood,
Nima (NM) Path length: 7.7 km
because it used real-time data on source status collected at the same time as PM measurements, e.g., whether there was heavy traffic at the time of measurement vs. data on road capacity. In this paper, we use new data on the distributions of sources in the same neighborhoods to estimate their associations with PM pollution over the full mobile monitoring path. There are three main contributions in this paper: First, to our knowledge this is one of the first studies of the spatial patterns of air pollution levels and of their sources in neighborhoods of varying SES in a developing country city. In particular, we present the first detailed data on the density of biomass stoves along roads in urban neighborhoods with different SES. Second, the unique data are used to estimate the pollution-source associations. In addition to helping understand the predictors of air pollution spatial patterns, these associations can be used, as a so-called “land-use regression” model, to predict PM pollution along roads and alleys in different neighborhoods. Third, we extend the methods for analysis of continuous mobile air pollution data to account for both spatial and temporal correlation of the measurements.
East Legon (EL) Path length: 9.4 km
Accra Fixed site Alley Local secondary road
% household using biomass
Connecting secondary road
0 - 40% 40 - 60%
Primary road
60 - 70% 70 - 80%
Divided multi-lane highway
> 80%
Asylum Down (AD) Path length: 8.9 km
Jamestown/Ushertown (JT) Path length: 8.5 km
Fig. 1. Map of Accra Metropolitan Area, study areas, and mobile monitoring path, delimited by census enumeration areas (EAs). EAs have nearly the same population; hence the area of an EA is inversely related to population density. Local secondary roads are smaller roads whose traffic is primarily for the purpose of reaching a local destination; connecting secondary roads are smaller roads used for passing through the neighborhood.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
2. Materials and methods 2.1. Study location Our study took place in four Accra neighborhoods: Jamestown/ Ushertown (JT), Asylum Down (AD), Nima (NM), and East Legon (EL) (Fig. 1). Detailed information on study neighborhoods are provided in previous work (Dionisio et al., 2010a, 2010b; Zhou et al., 2011). 2.2. Study design and data sources 2.2.1. PM data We conducted consecutive days of mobile monitoring in each neighborhood (7 days in most neighborhoods). The Accra air pollution study took place between 2006 and 2008; mobile measurements were done in April 2007 (before the main rainy season) in AD and in July–August 2007 (after the main rainy season) in the other neighborhoods. There were no unusual meteorological factors during the measurement periods. There were two monitoring tours on each day (one in the morning starting between 6:30 and 7:00, one in the afternoon starting at approximately 12:00). If it rained for more than approximately 30 min or if there was an equipment failure, the monitoring session was cancelled and repeated later so that the number of measurement days was as close as possible to 7. In each monitoring tour, we walked slowly along the full length of a pre-determined path, recording data at 1-minute intervals with a continuous real-time PM monitor and a GPS unit. The path for each neighborhood was designed to traverse different areas including main highways/roads, local roads, residential alleys and foot paths, and markets. The paths ranged 8–9 km in length, and the tours lasted 4.5–5.5 h in duration in different neighborhoods (Fig. 1). Further details on the measurement schedule are provided in a previous work (Dionisio et al., 2010b). We measured 48-hour integrated PM2.5 and PM10 concentrations at 2–3 roof-top fixed sites in each neighborhood using gravimetric methods. One fixed site in each neighborhood was located along a main road with the others located in residential areas. We also measured PM2.5 and PM10 continuously at as many fixed sites as possible. Further details on measurements at fixed sites are provided elsewhere (Dionisio et al., 2010a). Continuous PM was measured using DustTrak Model 8520 and SidePak AM510 monitors (TSI Inc., Shoreview, MN), and corrected against gravimetric PM as described in detail in a previous work (Dionisio et al., 2010a, 2010b). Using the timestamps of DustTraks or SidePaks, error-corrected continuous PM2.5 and PM10 data from fixed sites and mobile monitors were compiled into a single dataset with each record representing a unique date, minute, and location. Geographic coordinates for each mobile monitor data point were matched to PM data using the GPS date/timestamp. Because GPS units measure true location with error, GPS coordinates were “snapped” to the nearest point on the monitoring path; the location of the path was ascertained using a Trimble GeoXT GPS unit (Trimble Navigation Limited, Sunnyvale, CA) with a nominal error of less than 1 m. When a point was at or near the intersection of two roads that were both on the path, the snapping retained the temporal ordering of data points. 2.2.2. Source data along and near the mobile monitoring path We conducted a geo-coded census of wood and charcoal stoves along each of the four mobile monitoring paths. Traversing each path, we used a GPS unit to record the location of all stoves and stove clusters as well as other large PM sources (e.g. bakeries, fish smoking, trash burning). We counted all stoves visible from the path and those that could be identified by visible smoke, and recorded the following information: number of stoves in a cluster, stove type(s) (charcoal or firewood), stove size (small or large), and on/off status. We asked a set of pre-specified questions from the attendants about the days of a week and times of a day (divided into 6:00–10:00, 10:00–14:00, 14:00–17:00, 17:00–19:00 and 19:00–22:00) when the stove was used; the questions were asked of
109
every third to fifth attendant in AD, the tenth attendant in JT and NM, and all consenting attendants in EL due to varying density of sources across neighborhoods. We conducted 6–8 rounds of source count in JT, NM, and EL and only one round in AD because the field team was smaller at the time of AD campaign. Questions were not repeated to attendants previously interviewed. We used a Trimble GeoXT GPS to measure the location of the mobile monitoring path and to mark its various segments based on (i) road capacity, categorized by alleys, local secondary roads, connecting secondary roads, primary roads, and multi-lane highways and (ii) road surface, categorized by paved, paved broken, packed dirt, and loose dirt. We used a geo-coded road map of Accra to measure the distance from each point along the monitoring path to the primary roads or multi-lane highways. 2.2.3. Community household fuel use and SES In addition to the source data along the monitoring path, we used a 10% sample of the Ghana 2000 population and housing census to calculate the following variables in each census enumeration area (EA) traversed by the mobile monitoring path: • density of households per hectare that used charcoal or wood for cooking; • density of households per hectare that burned their household trash as primary means of disposal; and • community SES index. There were 55 EAs along the path in JT, 22 in AD, 34 in NM, and 3 in EL. Following previous analyses of household data in developing countries (Gakidou et al., 2007; Wagstaff, 2000), we measured SES using an index based on type and tenure of dwelling, materials of outer walls, floor, and roof, toilet and bathing facilities, solid and liquid waste disposal methods, water source, and the number of persons per room and per bedroom; the SES index excluded household fuel type (Zhou et al., 2011). To calculate the SES index, we used principal component analysis (PCA), a statistical procedure that transforms a number of possibly correlated variables into a smaller number of uncorrelated variables, the principal components. We used the first principal component as the SES index, because it is a linear combination of the individual variables that explains the maximum variance (29%) across the households. The PCA was conducted using data from individual census records (each corresponding to one household), with household scores averaged to obtain EA SES. 2.2.4. Meteorological/weather variables Data on relative humidity (RH), wind speed and time since last rain were from a station near Accra's Kotoka International Airport (Fig. 1), maintained by the United States Department of Commerce National Oceanic and Atmospheric Administration (NOAA). We predicted RH and wind speed for hours with missing data using a simple linear interpolation when data were missing for less than 3 h. When more than 3 hours of data were missing, we used the average of RH for the same hour over 5 days before and 5 days after the missing value. We then fitted a cubic spline function to hourly RH and wind speed to obtain RH and wind speed values for each minute (Dionisio et al., 2010a). 2.3. Statistical analysis 2.3.1. Stove and other source density along the mobile monitoring path We used the stove count data to calculate a “wood stove density” and a “charcoal stove density” associated with each observation in the mobile monitoring PM dataset. Using the interview data, we assigned a probability of 0, 0.5, or 1.0 to the stove being on for each day of the week and time interval, with 0.5 indicating self-reported variable status at that time/day. We assigned probabilities to stoves without an interview using multiple imputation (R package AMELIA; available at http://gking.harvard.edu/amelia/) (Honaker, 2010) by creating 300 imputed datasets and using their average for each stove. Stove
110
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
density for each PM record depended on its location along the path and the time of measurement and was calculated in relation to the location of all stoves and the time that they were on as described below. Each stove or stove cluster whose probability of being on was non-zero contributed to the density, with its effect declining with distance. The decline in the effect of each stove was modeled using a multivariate normal density, with a standard deviation of 10 m. Studies in rural areas have shown that PM pollution from biomass stoves drops sharply within about this distance (Ezzati et al., 2000). Stove ∑i N ðdi j0;102 IÞci P t P d
i i density was calculated as , where N(·) is the multivars iate normal density; di is the longitudinal and latitudinal displacement between the mobile monitor and stove or stove cluster i in meters; ci is the number of stoves at stove cluster i (large stoves being counted doubly); P(ti) and P(di) are, respectively, the probabilities that source i would be on during the hour of the day and day of the week the measurement was taken; s is the number of times we counted sources along the path in each neighborhood. This metric is equivalent to the number of stoves that are in the proximity of the measurement point because it accounts for distances to various stove clusters. We calculated similar metrics for other sources along the walking path. Source densities were scaled to stoves per hectare.
2.3.2. Association of PM with sources We used regression analysis to quantify the association of sources and meteorological factors with PM along the mobile monitoring path. The regression model was specified by yi = Xiβ + bi + εi, where yi is the dependent variable vector (natural logarithm of PM concentrations) for neighborhood-day i, β is a vector of fixed-effect coefficients for the covariates plus the intercept term, Xi is the matrix of covariates associated with all PM records, bi is the neighborhood-day random intercept (Laird and Ware, 1982), and εi is the vector of within-neighborhoodday errors. We used the following model covariates: household wood and charcoal use densities in the EA where the measurement was carried out; stove density along the mobile monitoring path, calculated as above; road traffic capacity at the measurement location; distance to the closest major road; wind speed; road surface type; relative humidity; and time since the last precipitation. In addition to these covariates, we conducted the analysis both with and without adjusting for the ambient neighborhood PM, calculated as the average of concurrent PM measured at the fixed monitoring stations in each neighborhood (smoothed to remove perturbations lasting less than 30 min). PM concentrations were log-transformed to ensure that model residuals were normally distributed. For consistency with this transformation, we also log-transformed the density of sources along the walking path and the EA fuel use density. Residual diagnostics suggest the relationships between these variables and the outcome were reasonably linear when they were log-transformed but that, for a subset of them, the relationship was nonlinear without the log transformation. The neighborhood-day random intercept accounts for unobserved factors that may have affected PM for each neighborhood-day, such as local meteorological factors, day-to-day variations in economic activity, or even imperfect instrument calibration. In addition, because PM observations were taken at 1 min intervals during which the monitor moves for a short distance only, there may be both spatial and temporal correlation between successive observations within a single neighborhood-day. To account for this, we modeled the within neighborhood-day error (εi) as the sum of empirically-estimated spatially and temporally correlated components as well as an independent component. By using the sum of spatially and temporally correlated components, the spatial and temporal components are modeled as independent; i.e., the temporal components of the within-neighborhood-day variance are similar over space, and vice versa. This produces a spatio-temporal covariance that is additive in the temporal and spatial components. The temporal and spatial autocorrelation were modeled as exponential, with the parameters estimated empirically with other model parameters.
The model was fitted by maximizing the restricted log likelihood using the function lme() in the R package nlme. 3. Results Gravimetric-corrected ground-level PM varied substantially within each neighborhood, reaching as high as 200 μg/m3 for PM2.5 and 400 μg/m 3 for PM10 (Fig. 2) (Dionisio et al., 2010a). In general, PM concentrations increased with road capacity and were highest along the divided multi-lane highway, with median PM2.5 and PM10 of 53.4 and 144.5 μg/m 3, respectively (Fig. 3). PM2.5 concentrations along alleys were higher than smaller local secondary roads with some traffic. The density of charcoal stoves used along the walking path in JT and NM was about 7 and 20 times those of AD and EL, respectively (Fig. 4). Woodstoves were virtually absent along the mobile monitoring path in EL and were most common in JT, about 5 times more common than in NM and 10 times that of AD. There was little variation in stove use among days of the week, except on Sundays, when fewer stoves were in use in AD and JT. Stoves were used most in the morning, with about two thirds of all surveyed stoves used before 10:00, about half between 10:00 and 17:00, declining to about 10% after 19:00. This is consistent with demand for and sale of street food. In multivariate analysis, PM2.5 and PM10 levels increased with road capacity (Table 1). After adjustment for neighborhood average, divided multi-lane highways had 56% (95% CI 36–78) higher PM2.5 and 74% (95% CI 53–98) higher PM10 than alleys. PM2.5 and PM10 were highest on roads with a loose dirt surface, by 14–22% for different size fractions and in models with and without adjustment; the next highest pollution was on roads with a broken paved surface. Across the different models, PM levels decreased with distance from a main road; at 50 m from a main road, PM2.5 dropped by approximately 13–15% and PM10 by 14–16%. Community SES was inversely associated with PM pollution, with PM2.5 and PM10, respectively, 34% and 20% higher in the lowest SES neighborhoods in our sample than in the highest SES one. Non-traffic combustion sources generally did not have a significant association with PM10 along the walking path at p = 0.05, with the quantity of woodstoves, trash burning, and fish smoking along the walking path approaching this threshold. The seemingly negative association between community fuel use and PM10 in the model without adjustment for the neighborhood average is unexplained. The pair-wise correlations between the density of sources along the walking path (from source count) and community fuel use (from the census) were generally small and ranged between −0.19 and 0.56. Removing the density of sources along the walking path removed the negative associations for community charcoal use, making the coefficient null. Density of nearby woodstoves along the walking path was significantly associated with higher PM2.5. An additional 25 nearby woodstoves per hectare (the 95th percentile of densities in our data after accounting for the fact that pollution drops sharply with distance) had an effect of about 9-10% higher PM2.5 than an area without woodstoves in the two models. After adjusting for community SES and other variables, community fuel use had no significant association with PM2.5. Among larger sources, fish smoking and trash burning had a noticeable and statistically significant effect on PM2.5 levels along the walking path, with 37% (95% CI 17–61) and 27% (95% CI 1.0–59) higher PM2.5 for increasing the number of nearby fish smoking or trash burning spots to 5 per hectare, respectively, in the model without adjustment for the neighborhood average. An increase in RH from 25% to 75% was associated with 215% higher PM2.5 and 126% higher PM10. In the model without adjustment for the neighborhood average, increased wind speed was inversely associated with PM; compared to stagnant conditions, PM2.5 and PM10 levels were expected to be 33% (95% CI 26–39) and 28% (95% CI 20–35) lower, respectively, with a 10 mph breeze. Rain also had a mitigating effect on PM10.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
800
600
400
200
0
1000
Distance (meters) 800
600
400
200
0
1600
1800
Distance (meters) 1400
1200
1000
800
600
400
200
0
1200
Distance (meters) 1000
800
600
400
200
0
Distance (meters)
111
PM2.5 (µg/m3)
PM2.5
N=3,602
N=3,942
N=3,875
N=3,791
< 10 10 − 20 20 − 30 30 − 40 40 − 50 50 − 60 60 − 70 70 − 80 80 − 90 90 − 100 100 − 150 150 − 200 > 200 PM10 (µg/m3)
PM10
N=3,599
N=4,209
N=3,240
N=3,806
Asylum Down (AD)
East Legon (EL)
Jamestown (JT)
Nima (NM)
< 20 20 − 40 40 − 60 60 − 80 80 − 100 100 − 120 120 − 140 140 − 160 160 − 180 180 − 200 200 − 300 300 − 400 > 400
Fig. 2. Concentrations of PM2.5 and PM10 along the walking path in the study neighborhoods. For each neighborhood and PM size fraction, data from all monitoring days/tours were combined in a moving average, with a 50 m averaging interval. Sample sizes show the total number of minute-by-minute measurements in the neighborhood. Reproduced with permission from Environmental Health Perspectives (Dionisio et al., 2010a).
The model accounted for 63% of PM2.5 and 72% of PM10 variability when neighborhood average was included, and for 41% and 33% when neighborhood average was not included, indicating that its predictions
1000 N=3270
N=489
PM2.5 (µg/m3)
N=5840
N=3469
N=2136
100
are substantially improved in the presence of fixed site monitors. Note that part of the unexplained variance includes daily fluctuations (modeled as a random intercept term), which may or may not be of interest in understanding location-specific PM exposures. Model residuals from observations made at the same time and location were approximately 90% correlated, and it was estimated that about 70% of this effect was due to time and 20% due to space. The autocorrelative effects of time and space dropped to 5% of their maximal values at lags of 13 min and 100 m for PM2.5, and 14 min and 190 m for PM10 (unadjusted model). These estimates provide guidance on appropriate lags for collecting statistically independent samples in an urban environment. Autocorrelation seemed to occur in both time and space, with more temporal effect. Specifically, about 70% and 65% of variance of PM2.5 and PM10 model error terms were, respectively, explained by the temporal component and 20% by the spatial one.
10
4. Discussion
1000
N=489 N=3017
N=2113
PM10 (µg/m3)
N=5684
N=3384
100
10
Alley
Local Sec. Connect. Sec. Primary Div. Multi−Lane
Fig. 3. Distributions of PM2.5 and PM10 by road capacity. Sample sizes show the total number of minute-by-minute measurements along that road type.
In previous work, we used data from a mobile monitoring campaign to show that PM concentrations had significant spatial variability within and between neighborhoods in Accra (Dionisio et al., 2010b). The low-SES neighborhood of JT had the highest pollution, followed by segments of the mobile monitoring path along primary roads in other neighborhoods. Our previous work also demonstrated that burning woodstoves and congested or heavy traffic had the strongest effects on real-time local PM concentration at approximately pre-selected points along the walking path, but did not analyze the remaining parts of the paths (Dionisio et al., 2010b). In the current work, we investigated the associations between PM concentrations and sources along the complete walking path in these four neighborhoods, with source data collected once and hence reflecting “usual” vs. real-time conditions. Specifically, we used the population and housing census to obtain data on biomass use in small areas (census EAs) and conducted our own census of PM sources along the monitoring path. Except for road surface, which is time-invariant, comparison with the previous work is not possible due to the differences between
112 M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
Fig. 4. Locations of stoves along and near the walking path, and stove use by day of week and time of day.
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
113
Table 1 Model coefficients for multivariate analysis of the association of PM with biomass sources, road capacity and surface, and meteorological covariates. Ln denotes natural logarithm. Dependent variable: Ln(PM2.5)
Constant Ln(neighborhood average) Ln(distance to nearest main road (m)) Community SES Density of biomass stoves and other sources along walking path (per hectare) Ln(charcoal stove density) Ln(wood stove density) Ln(bakery density) Ln(fish smoking density) Ln(flour mill density) Ln(trash burning density) Density of community sources from the census (per hectare) Ln(wood stove density) Ln(charcoal stove density) Ln(trash burning density) Road typea Alley Local secondary roads Connecting secondary roads Primary roads Multi-lane highways Road surfacea Paved Paved broken Packed dirt Loose dirt Meteorological factors Wind speed (mph) Relative humidity Ln(hours since rain)
Model 1
Model 2
Coefficient
p-Value
Coefficient
p-Value
1.194 (0.596, 1.793)
b0.001
−0.036 (−0.053, –0.019) −0.041 (−0.060, –0.022)
b0.001 b0.001
−0.042 (−0.567, 0.482) 0.568 (0.503, 0.633) −0.041 (−0.058, –0.024) −0.040 (−0.059, –0.021)
0.875 b0.001 b0.001 b0.001
−0.003 (−0.008, 0.003) 0.013(0.005, 0.020) −0.018 (−0.045, 0.008) 0.051 (0.026, 0.076) −0.004 (−0.031, 0.023) 0.038 (0.002, 0.074)
0.330 b0.001 0.175 b0.001 0.780 0.041
−0.002 (−0.007, 0.004) 0.012(0.004, 0.019) −0.023 (−0.049, 0.002) 0.040 (0.015, 0.064) −0.004 (−0.030, 0.022) 0.033 (−0.002, 0.068)
0.552 0.002 0.073 0.002 0.776 0.065
−0.003 (−0.009, 0.003) −0.015 (−0.033, 0.003) 0.002 (−0.003, 0.008)
0.358 0.107 0.405 0.003
−0.002 (−0.008, 0.004) −0.010 (−0.028, 0.007) 0.003 (−0.003, 0.009)
0.510 0.250 0.308 b0.001
0 0.032 0.089 0.128 0.390
0 0.033 (−0.008, 0.073) 0.088 (0.036, 0.140) 0.150 (0.085, 0.214) 0.444(0.309, 0.579)
(−0.009, 0.073) (0.036, 0.142) (0.062, 0.193) (0.246, 0.535) b0.001
0 0.054 (−0.006, 0.114) 0.022 (−0.017, 0.060) 0.199 (0.137, 0.260)
b0.001 0 0.047(−0.012, 0.105) 0.017 (−0.021, 0.055) 0.199(0.138, 0.260)
−0.040 (−0.050, −0.030) 3.591(3.155, 4.026) 0.079 (−0.028, 0.186)
b0.001 b0.001 0.147
−0.009 (−0.018, 0.001) 2.295(1.900, 2.691) 0.047(−0.042, 0.135)
0.067 b0.001 0.303
2.039 (1.433, 2.646)
b0.001
−0.038 (−0.055, −0.022) −0.027 (−0.046, −0.009)
b0.001 0.004
0.772 (0.327, 1.218) 0.550 (0.492, 0.608) −0.045 (−0.061, −0.029) −0.025 (−0.043, −0.007)
b0.001 b0.001 b0.001 0.006
Dependent variable: Ln(PM10) Constant Ln(neighborhood average) Ln(distance to nearest main road (m)) Community SES Density of biomass stoves and other sources along walking path (per hectare) Ln(charcoal stove density) Ln(wood stove density) Ln (bakery density) Ln(fish smoking density) Ln(flour mill density) Ln(trash burning density) Density of community sources from the census (per hectare) Ln(wood stove density) Ln(charcoal stove density) Ln(trash burning density) Road typea Alley Local secondary roads Connecting secondary roads Primary roads Multi-lane highways Road surfacea Paved Paved broken Packed dirt Loose dirt Meteorological factors Wind speed (mph) Relative humidity Ln(hours since rain)
−0.002 (−0.007, 0.004) −0.007 (0.000, 0.013) 0.004 (−0.023, 0.030) 0.021 (−0.007, 0.048) −0.011 (−0.038, 0.015) 0.026 (−0.012, 0.063)
0.545 0.057 0.797 0.138 0.403 0.179
0.001 (−0.006, 0.005) 0.005 (−0.002, 0.012) 0.003 (−0.024, 0.029) 0.012 (−0.015, 0.038) −0.008 (−0.033, 0.018) 0.022 (−0.015, 0.059)
0.840 0.154 0.845 0.398 0.565 0.249
−0.006 (−0.012, 0.000) −0.018 (−0.036, 0.001) 0.000 (−0.005, 0.006)
0.036 0.044 0.930 b0.001
−0.006 (−0.012, 0.000) −0.003 (−0.018, 0.013) 0.001 (−0.005, 0.006)
0.067 0.716 0.734 b0.001
0 0.028 0.101 0.135 0.513
0 0.024 0.098 0.158 0.551
(−0.012, 0.068) (0.050, 0.152) (0.071, 0.198) (0.374, 0.653)
(−0.016, 0.063) (0.047, 0.148) (0.096, 0.221) (0.422, 0.681) b0.001
0.026 0 0.103 (0.045, 0.162) 0.014 (−0.023, 0.050) 0.129 (0.070, 0.187) −0.032 (−0.043, −0.022) 3.259 (2.827, 3.692) 0.104 (−0.004, 0.211)
0 0.102 (0.044, 0.160) 0.008 (−0.029, 0.044) 0.129 (0.071, 0.188) b0.001 b0.001 0.059
−0.006 (−0.014, 0.003) 1.628 (1.251, 2.006) 0.098 (0.037, 0.160)
0.217 b0.001 0.002
Because dependent variable is Ln(PM), [exp(coefficient) − 1] is the percent increase/decrease in PM due to each variable. For small coefficients, those whose absolute value is up to about 0.25, this value is approximately equal to the coefficient itself. Model 2 is adjusted for neighborhood average (estimated as average of smoothed concentrations at all fixed sites) and Model 1 is without this variable. See Materials and methods section for details. a p-Values are given for the complete categorical variable using an F-test. The p-value for the categorical variable tests whether the inclusion of the variable in the model is significant.
real-time and usual data, e.g. current presence of a burning stove vs. usual stove density or current heavy traffic vs. road capacity. The limitations of our study include the relatively short measurement period which does not allow for assessing how spatial patterns
of PM pollution and association with sources vary over seasons; the absence of evening and night-time data; the lack of census data concurrent with our study; potential error in self-reported days and times of stove use; and the use of DustTraks whose measurements
114
M.S. Rooney et al. / Science of the Total Environment 435–436 (2012) 107–114
are subject to error which may have been only partially addressed by our correction factors. Finally, while some of the predictors needed to apply our results to other neighborhoods are available from routine sources like the census, source data along the path requires additional data collection. Despite the apparent complexity, such data can be collected once at a cost that is substantially lower than many other emission inventories routinely used in air pollution studies such as those of residential woodstove use in high-income countries. Despite these limitations that are common to many field research studies due to logistical difficulties, this unique combination of data showed that PM concentrations along the mobile monitoring path had a stronger association with sources along the path – especially with the presence of wood stoves, fish smoking, and trash burning – than they did with census biomass variables. The associations with sources along the walking path were stronger for PM2.5 than for PM10. Our findings on the absence of an association with census variables are perhaps different from those in the Greater Vancouver area (Canada), where residential woodstove use was a predictor of PM pollution (Larson et al., 2007). The reasons for this difference need further investigation and may include the following: first, there were many biomass sources along the walking path in Accra (Fig. 4) used for cooking street food, which is rare in high-income settings. These nearby sources may have had a more important role in local concentrations than the density of household biomass use due to their proximity. Second, we had data on hours of stove use for sources along the path but not for those from the census; thus our path source density was temporally variable but those for EAs from the census were not. This may have increased the relative predictive power of the density of sources along the path, because the PM data were also time varying. We also found associations between PM pollution and road capacity, consistent with studies in high-income countries (Levy et al., 2001; Weijers et al., 2004) and those in developing countries that had found higher PM along major traffic routes than in non-traffic areas (Padhi and Padhy, 2008; van Vliet and Kinney, 2007). None of these studies had accounted for non-traffic combustion sources.
5. Conclusions We found that, after adjusting for other factors, the density of wood stoves, fish smoking, and trash burning as well as road capacity and surface were associated with PM2.5 in Accra neighborhoods. Trash burning occurs because there are few places for safe disposal of solid waste and because the waste piles are removed too infrequently by the city authorities. The wood stoves along the walking path are used for cooking and selling street food, often by members of poor households (Zhou et al., 2011); fish smoking is also a source of income as a small business, for its owners and workers. While both were sources of air pollution, addressing them would require financial and physical access to alternative fuels for low-income households and communities (Bailis et al., 2005; Barnes et al., 2005; Zhou et al., 2011). Acknowledgment Funding for this research was provided by the National Science Foundation (grant 0527536). Laboratory support was provided by the Harvard NIEHS Center for Environmental Health. We thank Nana Prempeh and Adam Abdul Fatah for field assistance, and the Legal Resources Center and the Department of Geography and Resource Development at the University of Ghana for valuable help with logistical arrangements.
References Arku RE, Vallarino J, Dionisio KL, Willis R, Choi H, Wilson JG, et al. Characterizing air pollution in two low-income neighborhoods in Accra, Ghana. Sci Total Environ 2008;402:217–31. Bailis R, Ezzati M, Kammen DM. Mortality and greenhouse gas impacts of biomass and petroleum energy futures in Africa. Science 2005;308:98-103. Barnes DF, Krutilla K, Hyde WF. The urban household energy transition : social and environmental impacts in the developing world. Washington, DC: RFF Press; 2005. Charron A, Harrison RM. Fine (PM2.5) and coarse (PM2.5–10) particulate matter on a heavily trafficked London highway: sources and processes. Environ Sci Technol 2005;39:7768–76. Chowdhury MZ. Characterization of fine particle air pollution in the Indian subcontinent. Atmospheric Chemistry: Georgia Institute of Technology; 2004. Dionisio KL, Arku RE, Hughes AF, Vallarino J, Carmichael H, Spengler JD, et al. Air pollution in Accra neighborhoods: spatial, socioeconomic, and temporal patterns. Environ Sci Technol 2010a;44:2270–6. Dionisio KL, Rooney MS, Arku RE, Friedman AB, Hughes AF, Vallarino J, et al. Within-neighborhood patterns and sources of particle pollution: mobile monitoring and geographic information system analysis in four communities in Accra, Ghana. Environ Health Perspect 2010b;118:607–13. Engelbrecht JP, Swanepoel L, Chow JC, Watson JG, Egami RT. PM2.5 and PM10 concentrations from the Qalabotjha low-smoke fuels macro-scale experiment in South Africa. Environ Monit Assess 2001;69:1-15. Etyemezian V, Tesfaye M, Yimer A, Chow JC, Mesfin D, Nega T, et al. Results from a pilot-scale air quality study in Addis Ababa, Ethiopia. Atmos Environ 2005;39: 7849–60. Ezzati M, Saleh H, Kammen DM. The contributions of emissions and spatial microenvironments to exposure to indoor air pollution from biomass combustion in Kenya. Environ Health Perspect 2000;108:833–9. Gakidou E, Oza S, Vidal Fuertes C, Li AY, Lee DK, Sousa A, et al. Improving child survival through environmental and nutritional interventions: the importance of targeting interventions toward the poor. JAMA 2007;298:1876–87. Hoek G, Brunekreef B, Fischer P, van Wijnen J. The association between air pollution and heart failure, arrhythmia, embolism, thrombosis, and other cardiovascular causes of death in a time series study. Epidemiology 2001;12:355–7. Hoek G, Brunekreef B, Goldbohm S, Fischer P, van den Brandt PA. Association between mortality and indicators of traffic-related air pollution in The Netherlands: a cohort study. Lancet 2002;360:1203–9. Holmes NS, Morawska L, Mengersen K, Jayaratne ER. Spatial distribution of submicrometre particles and CO in an urban microscale environment. Atmos Environ 2005;39:3977–88. Honaker J. What to do about missing values in time series cross-section data. Am J Pol Sci 2010;54:561–81. Jackson MM. Roadside concentration of gaseous and particulate matter pollutants and risk assessment in Dar-es-Salaam, Tanzania. Environ Monit Assess 2005;104: 385–407. Laird NM, Ware JH. Random-effects models for longitudinal data. Biometrics 1982;38: 963–74. Larson T, Su J, Baribeau AM, Buzzelli M, Setton E, Brauer M. A spatial model of urban winter woodsmoke concentrations. Environ Sci Technol 2007;41:2429–36. Levy JI, Houseman EA, Ryan L, Richardson D, Spengler JD. Particle concentrations in urban microenvironments. Environ Health Perspect 2000;108:1051–7. Levy JI, Houseman EA, Spengler JD, Loh P, Ryan L. Fine particulate matter and polycyclic aromatic hydrocarbon concentration patterns in Roxbury, Massachusetts: a community-based GIS analysis. Environ Health Perspect 2001;109:341–7. Padhi BK, Padhy PK. Assessment of intra-urban variability in outdoor air quality and its health risks. Inhal Toxicol 2008;20:973–9. Saksena S, Singh PB, Prasad RK, Prasad R, Malhotra P, Joshi V, et al. Exposure of infants to outdoor and indoor air pollution in low-income urban areas — a case study of Delhi. J Expo Anal Environ Epidemiol 2003;13:219. Su JG, Larson T, Baribeau AM, Brauer M, Rensing M, Buzzelli M. Spatial modeling for air pollution monitoring network design: example of residential woodsmoke. J Air Waste Manag Assoc 2007;57:893–900. Su JG, Brauer M, Ainslie B, Steyn D, Larson T, Buzzelli M. An innovative land use regression model incorporating meteorology for exposure analysis. Sci Total Environ 2008;390:520–9. van Vliet E, Kinney P. Impacts of roadway emissions on urban particulate matter concentrations in sub-Saharan Africa: new evidence from Nairobi, Kenya. Environ Res Lett 2007:2. Wagstaff A. Socioeconomic inequalities in child mortality: comparisons across nine developing countries. Bull World Health Organ 2000;78:19–29. Weijers EP, Khlystov AY, Kos GPA, Erisman JW. Variability of particulate matter concentrations along roads and motorways determined by a moving measurement unit. Atmos Environ 2004;38:2993–3002. Zheng M, Salmon LG, Schauer JJ, Zeng L, Kiang CS, Zhang Y, et al. Seasonal trends in PM2.5 source contributions in Beijing, China. Atmos Environ 2005;39:3967–76. Zhou Z, Dionisio KL, Arku RE, Quaye A, Hughes AF, Vallarino J, et al. Household and community poverty, biomass use, and air pollution in Accra, Ghana. Proc Natl Acad Sci U S A 2011;108:11028–33.