The exposure of infants and children to carbon monoxide ... - Nature

Journal of Exposure Science and Environmental Epidemiology (2012) 22, 173–181 r 2012 Nature America, Inc. All rights reserved 1559-0631/12

www.nature.com/jes

The exposure of infants and children to carbon monoxide from biomass fuels in The Gambia: a measurement and modeling study KATHIE L. DIONISIOa,b, STEPHEN R.C. HOWIEc, FRANCESCA DOMINICId, KIMBERLY M. FORNACEe, JOHN D. SPENGLERb, SIMON DONKORc, OSARETIN CHIMAHc, CLAIRE OLUWALANAc, READON C. IDEHc, BERNARD EBRUKEc, RICHARD A. ADEGBOLAc,f AND MAJID EZZATIg,h a

Department of Global Health and Population, Harvard School of Public Health, Boston, Massachusetts, USA Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts, USA c Child Survival Theme, Medical Research Council, The Gambia Unit, Fajara, The Gambia d Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts, USA e Veterinary Epidemiology and Public Health Group, Royal Veterinary College, London, UK f Bill & Melinda Gates Foundation, Seattle, Washington, USA g Department of Epidemiology and Biostatistics, MRC-HPA Center for Environment and Health, Imperial College London, London, UK h Department of Epidemiology and Biostatistics, School of Public Health, Imperial College of London, London, UK b

Smoke from biomass fuels is a risk factor for pneumonia, the leading cause of child death worldwide. Although particulate matter (PM) is the metric of choice for studying the health effects of biomass smoke, measuring children’s PM exposure is difficult. Carbon monoxide (CO), which is easier to measure, can be used as a proxy for PM exposure. We measured the exposure of children r5 years of age in The Gambia to CO using small, passive, color stain diffusion tubes. We conducted multiple CO measurements on a subset of children to measure day-to-day exposure variability. Usual CO exposure was modeled using a mixed effects model, which also included individual and household level exposure predictors. Mean measured CO exposure for 1181 children (n ¼ 2263 measurements) was 1.04±1.46 p.p.m., indicating that the Gambian children in this study on average have a relatively low CO exposure. However, 25% of children had exposures of 1.3 p.p.m. or higher. CO exposure was higher during the rainy months (1.33±1.62 p.p.m.). Burning insect coils, using charcoal, and measurement done in the rainy season were associated with higher exposure. A parsimonious model with fuel, season, and other PM sources as covariates explained 39% of between-child variation in exposure and helped remove within-child variability. Journal of Exposure Science and Environmental Epidemiology (2012) 22, 173–181; doi:10.1038/jes.2011.47; published online 14 December 2011

Keywords: indoor air pollution, biomass fuels, global health, Africa, carbon monoxide, exposure assessment.

Introduction Nearly one-half of the world’s population and 80% of the population of sub-Saharan Africa use biomass fuels and coal for cooking (Smith et al., 2004). When burned in suboptimal conditions, these fuels emit high concentrations of healthdamaging pollutants, including particulate matter (PM) and carbon monoxide (CO). Women and young children may spend hours close to cooking fires, and hence have high exposures (Smith et al., 2004). Exposure to biomass and coal smoke leads to a range of adverse health outcomes including child pneumonia, the leading cause of child death worldwide (Dherani et al., 2008; Black et al., 2010).

1. Address all correspondence to: Professor Majid Ezzati, Department of Epidemiology and Biostatistics, MRC-HPA Centre for Environment and Health, Imperial College London, Norfolk Place, London W2 1PG, UK. Tel.: þ 44 (0)20 7594 0767. Fax: þ 44 (0)20 7594 3456. E-mail: [email protected] Received 13 April 2011; accepted 12 August 2011; published online 14 December 2011

Data on personal exposure to pollutants in biomass smoke are necessary for analysis of association with disease outcomes, but has been measured or estimated in few studies (Ezzati et al., 2000; Naeher et al., 2001; Balakrishnan et al., 2002; Bruce et al., 2004; Dasgupta et al., 2006; Mestl et al., 2007; Dionisio et al., 2008; McCracken et al., 2009). In particular, a metric of usual exposure is desired because exposure may vary from day to day and health effects may depend on cumulative exposure. Measuring personal exposure also allows for assessing its seasonality and its association with factors such as fuel, cooking and childcare habits, and housing structure. Such data are needed to develop models to predict exposure in large population-based studies and to plan interventions. PM exposure is generally considered the best measure of the hazardous effects of biomass smoke. Measuring children’s personal PM exposure is particularly difficult because current PM monitors are bulky and heavy, making it difficult for a small child to carry a monitor for many hours. Children’s exposure to CO can be directly measured with small, passive, relatively inexpensive color stain tubes, and

Dionisio et al.

can be used as an indicator of PM (Naeher et al., 2001; Dionisio et al., 2008; McCracken et al., 2009; Northcross et al., 2010). We report on a study that measured and modeled the exposure of children to CO, with emphasis on estimating ‘‘usual exposure’’ over a period of a few weeks, which may be relevant for effects on childhood pneumonia. We also examined the association of exposure with season (rainy or dry), and with household (e.g. fuel and location of cooking; insect coil and incense burning), childcare (e.g. frequency of carriage on mother’s back), and demographic (e.g. child’s age) factors.

Subjects and methods Study Area and Study Population Our study took place in The Gambia, in the greater Banjul area, and the Basse Health and Demographic Surveillance System area of the Upper River Division. The Gambia has a population of 1.4 million (2003 national census), about onehalf of which live in the greater Banjul area, including urban Banjul and less densely populated peri-urban settlements. The Basse region is located nearly 400 km east of Banjul and is predominantly rural. Study Design The study was approved by The Gambia Government– MRC Joint Ethics Committee (SCC/EC 1062) and was assessed as exempt by the Harvard School of Public Health Office of Human Research Administration. Overview Ideally, children’s usual exposure would be estimated using repeated measurements to account for daily variability. However, multiple measurements are costly and logistically difficult. We conducted multiple measurements of CO exposure in a subset of our study participants, with the remainder of study participants having one measurement. We used all measurements of CO exposure on the children together with data on predictors of CO exposure from a questionnaire in a statistical model to estimate usual exposure for all participants. We used the estimated usual exposure to examine associations with fuel, cooking location, and burning of insect coils and incense. Study Participants Study participants were from an epidemiological study of pneumonia in young children. Children between 2 and 59 months of age with severe pneumonia or with non-severe pneumonia were recruited from health facilities after informed consent was obtained from their parents. Individually matched healthy controls were recruited from each case’s neighborhood in a 1:1 ratio. The detailed methods for case finding, diagnosis criteria, and epidemiological analyses will be reported in a subsequent manuscript. The exposure analysis was blinded to the case vs 174

Exposure to CO from biomass fuels

control status of participants. At the time of this analysis, 1303 children had been enrolled and were eligible for exposure measurement.

Personal Exposure Measurement Measurements took place between July 2007 and January 2011, and were scheduled such that each case and their individually matched control were measured in the same week to the extent possible. A total of 1303 children were eligible for exposure measurement. Of these, a small number of children were excluded from the analysis for the following reasons: 28 (2%) because cooking fuel or location of cooking changed between recruitment and measurement and 37 (3%) because the child had migrated, the child was withdrawn from the study, or the child died before measurements were done. Exposure measurement was done after the control was recruited, at least 1 week after the case was discharged from the clinic or hospital, and after the caregiver had confirmed that the child’s activities were back to normal. This protocol was adopted to ensure that conditions affecting exposure were as similar as possible to conditions immediately before the disease episode. Before each measurement, the child’s caregiver was asked whether the child’s activity on that day was affected by any current or recent illness, and whether household activities during the coming week were expected to be different from usual, for example, due to celebration or travel. Measurement was postponed until the following week if the caregiver answered ‘‘yes’’ to either question (58 occasions). If a measurement was not successful, it was reattempted the following week. Fieldworkers also verified that children’s activities were the same as usual every 24 h during the measurement period. We conducted repeated measurements in a random subset of children to quantify within-child exposure variability. Repeated measurements were done over a period of 2–3 months to reflect the usual exposure that may be relevant for pneumonia. Specifically, randomly selected pairs of cases and controls were scheduled for a second CO measurement 2 weeks after the first measurement; third, fourth, and fifth measurements were spaced 3 weeks apart, from the date of the second measurement. We conducted a total of 2707 field measurements on 1236 children. Each measurement period was 72 h, with CO levels checked and recorded at 24 h intervals for quality control. We used a 72 h measurement period to account for day-to-day variation in exposure, without reaching the upper limit of detection (LOD) of the tubes based on the results of our pilot study (Dionisio et al., 2008). A total of 428 measurements were excluded for the following reasons: the child traveled after the measurement session began or could not be found at the end of the session; the CO tube was wet, discolored, broken, or lost; the child was sick and admitted to the clinic during the session; or the child was withdrawn from the study. For another 174 measurements, we used the reading at Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)


48 h because one of the above problems occurred between 48 and 72 h. This resulted in 2263 measurements (84%) on 1181 children (96%) included in our statistical analysis.

CO Exposure Measurement We measured integrated CO exposure over 72 h using a passive monitor, with CO color change levels checked and recorded at 24 h intervals. At each reading, the tube’s color change was measured to the nearest millimeter using a metric ruler. We used Drager CO 50/a-D Diffusion Tubes (Drager Safety AG & Co. KGaA, Luebeck, Germany), with a detection range of 50–600 p.p.m.-h. Tubes were placed inside a rubber sleeve sealed with end caps to prevent breakage, protect the subjects, and protect the tube from direct sunlight. The tube’s open end was placed over an opening in the lower cap, so that it was exposed to circulating air (Figure 1a). The rubber sleeve was placed inside a fabric cover and pinned to the back of the child’s shirt, between the child’s shoulders (Figure 1b). Although it would have been desirable to pin the CO tube on the front of the shoulder to be closest to the child’s breathing zone, this placement was deemed infeasible because it drew attention from the subject and other children, who would then play with the CO tube. Caregivers were asked to remove the tube while the child was bathing, and to hang the tube next to the child using a lanyard supplied by the fieldworkers while they were sleeping; they were instructed that the child should wear the tube at all other times. Color change of the diffusion tubes was measured to the millimeter with metric rulers. We fitted a third-order polynomial to the millimeter measurements corresponding to preprinted p.p.m.-h markings on each of the three batches of CO tubes used in the study. We used the fitted relationship to estimate cumulative (p.p.m.-h) exposures from each of the

Figure 1. (a) CO tube and tube holder (with rubber sleeve and fabric cover). (b) Child wearing CO tube. Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)

Dionisio et al.

field measurements, which were in turn converted to exposure concentrations through division by the duration of measurement (e.g., 72 h). The lower LOD for the tubes was 1 mm of measured color change over 48 or 72 h, equivalent to 0.10–0.13 p.p.m. over 72 h and 0.15–0.20 p.p.m. over 48 h, depending on the batch-specific calibration curve. All field workers were trained and tested on measurement of color change. Specifically, on 70 CO tubes, multiple fieldworkers independently measured the length of color change. Mean standard deviation of multiple measurements on tubes was 0.99 mm, equivalent to 0.10–0.20 p.p.m.

Mixed Effects Model for Usual CO Exposure We applied a mixed effects model to the data on personal CO exposure. The mixed effects model uses data from children with more than one measurement to partition the total variance into within-child and between-child components (Laird and Ware, 1982). The model includes a random intercept for each child, and fixed effects for covariates that may be associated with CO exposure (see Table 3 for complete list). By including a random intercept we estimate the standard errors of the fixed effects accounting for the variability in the repeated CO measurements. The dependent variable was log transformed because the CO measurements had a log-normal distribution (Figure 2a). The complete model (denoted ‘‘complete model’’ hereafter) is represented by: lnðyij Þ ¼ b0 þ b1 ðfuelÞi þ b2 ðseasonÞij þ b3 ðW Þi þ b4 ðX Þi þ b5 ðZÞij þ b6 ðstudy siteÞi þ ai þ eij where yij is the jth measured CO exposure for child i; b0 is the overall intercept; ai is the random intercept for child i; eij is random error; fuel is type of fuel used most for cooking in the home (purchased firewood, collected firewood, charcoal, other); season is the season in which the measurement was done (rainy, dry); study site is the region where the measurement was done (Banjul, Basse); and W, X, and Z are vectors of other PM sources in the home, household-level covariates, and individual-level covariates, respectively (see Table 3 for more detail). ai and eij are modeled as independent and normally distributed, both having a mean of 0 and variances 2 2 (sb2) and swithin-child (sw2 ), respectively. The of sbetween-child intraclass correlation coefficient, r ¼ sb2/sb2 þ sw2 , measures the proportion of the total variance explained by betweenchild differences. Covariate data were from a household questionnaire administered to each child’s mother or primary caregiver by trained MRC fieldworkers (for the remainder of this paper, we use ‘‘caregiver’’ to indicate ‘‘mother or other primary caregiver’’). We imputed missing covariate values (see Table 1 for detailed information on missingness) using multiple imputation with Amelia II software version 1.2–17 (King et al., 2001; Honaker and King, 2010; Honaker et al., 175


Dionisio et al.

Table 1. Selected characteristics of 1181 subjects included in the statistical analysis.a Type of fuel used most for cooking Collected firewood Purchased firewood Charcoal Other

531 478 147 25

Study site Banjul Basse

688 (58.3%) 493 (41.7%)

Other PM sources Frequency of incense burning Every day Sometimes or never Unknown

646 (54.7%) 534 (45.2%) 1 (0.1%)

Insect coil burning Yes No

327 (27.7%) 854 (72.3%)

Trash burning in the family compound Yes No

209 (17.7%) 972 (82.3%)

Household-level covariates Duration of stove use per day (h) Unknown Location of cooking in rainy season Inside (inside main house or inside separate cookhouse) Outside (in open air or outside under roof/overhang) Location of cooking in dry season Inside (inside main house or inside separate cookhouse) Outside (in open air or outside under roof/overhang) Unknown

Figure 2. (a) Distribution of measured child CO exposure, excluding those below the LOD (12.2% of measurements). (b) Distribution of predicted annual child CO exposure from the parsimonious model.

2010). Five separate imputations were done, with outputs averaged for continuous variables and with the mode used for categorical variables in the analysis. We used principal components analysis (PCA) to calculate a socioeconomic status (SES) index for each child (Wagstaff, 2000; Filmer and Pritchett, 2001). Variables included in the PCA were: whether or not mother earns any income herself, quantity of household assets (radios, iron beds, iron roofs, watches, bicycles, horse/donkey carts, televisions, and cars), and total number of animals owned by the family (cattle, sheep, goats, horses, and donkeys). The first principal component had relatively large coefficients for quantity of radios, iron beds, iron roofs, watches, bicycles, horse/donkey carts, cars, 176

(45.0%) (40.5%) (12.4%) (2.0%)

Number of tobacco smokers in the house Unknown Individual-level covariates Child’s sex Male Female

5.9±1.9 (n ¼ 1 172) n¼9

992 (84.0%) 189 (16.0%)

945 (80.0%) 235 (19.9%) 1 (0.1%) 0.5±0.9 (n ¼ 1 180) 1 (0.1%)

625 (52.9%) 556 (47.1%)

Frequency of carriage on the mother’s back during cooking Most of the time 77 (6.5%) Sometimes 592 (50.1%) Never 512 (43.4%) a Missingness for covariates included in the regression analysis but excluded from the table above ranged from 0.0 to 0.6%, with the exception of father’s occupation, for which 2% of data were missing.

and total number of animals owned by the family. Thus, the first principal component represents SES of a household based on the quantity of possessions and animals owned by the Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)


Dionisio et al.

(use of insect coils, burning of trash, and burning of incense; Table 2). The proportion of within- and between-child variance explained by each model compared with the nocovariate model is given by O2 ¼ 1s2/s20 (Xu, 2003), where s20 is the variance component (either within- or between-child variance) from the no-covariate model and s2 is the variance component from the comparison model.

household, and was used as the child’s SES index in subsequent analysis (see Supplementary Table 1 for more detail). The above model uses the observations from children with multiple measurements interchangeably. There may be a temporal relationship among the measurements, for example, because of temporally correlated determinants or time since illness. We used paired t-tests to examine whether measurement sequence affected exposure. Specifically, we tested for differences between the first and second measurements for all children with at least two CO measurements, between the first and third measurements and the second and third measurements for all children with at least three CO measurements, and so on. At the 95% confidence level, 1 of the 10 separate t-tests was statistically significant at p ¼ 0.05, providing evidence of a lack of systematic differences due to measurement sequence. To evaluate how covariates explain exposure variability, we compared models with an increasing number of covariates to a model that only included the random intercept for child (the ‘‘no-covariate’’ model), specified as below. lnðyij Þ ¼ b0 þ ai þ eij

Model Validation To assess the goodness of fit (or prediction accuracy) of the linear mixed effects model, we dropped a random subset of observations to create the appearance of missing data. The remaining data were used to estimate the model, which was in turn used to predict the excluded exposures. Data were dropped from each quartile of measured CO exposures (10% of observations) to assess model performance in all exposure ranges. We repeated this 100 times, with each run dropping a different random subset of data. For each model run, we calculated the following statistics for the known but dropped exposures: Mean square error (MSE) between predicted and measured exposure; Spearman’s correlation coefficient between predicted and measured exposure.

In addition to the no-covariate and complete models, we used models that included fuel, season, and other PM sources Table 2. Linear mixed effects model performance and validation. Model specification

Model fit

Model validation MSEd,e

Correlationf,e

0.394

0.935

0.003

0.188

0.345

0.866 (7.4%)

0.264

0.041

0.291

0.325

0.796 (14.9%)

0.370

0.219

0.040

0.393

0.291

0.753 (19.5%)

0.444

0.196

0.054

0.457

0.271

0.746 (20.2%)

0.464

O2a w

O2b b

Within-child variance (sw2 )

Between-child variance (sb2)

No-covariate model g ln(yij) ¼ b0+ai+eij

0.555

0.361

NA

NA

Fuel model g ln(yij) ¼ b0+b1(fuel)i+ai+eij

0.557

0.293

0.004

Fuel+season model g ln(yij) ¼ b0+b1(fuel)i+b2(season)ij+ai+eij

0.532

0.256

Parsimonious model g,h ln(yij) ¼ b0+b1(fuel)i+b2(season)ij+b3(W)i+ai+eij

0.533

0.525

Complete model g,h,i ln(yij) ¼ b0+b1(fuel)i+b2(season)ij+b3(W)i+b4(X)i+ b5(Z)ij+b6(study site)i+ai+eij

rc

Abbreviation: MSE, mean square error. O2w ¼ Within-child variance explained relative to no-covariate model. O2b ¼ Between-child variance explained relative to no-covariate model. c r ¼ sb2/(sb2+sw2 ). d MSE between the actual CO measurement and the predicted CO value. Number in parentheses is percent change in MSE of each model relative to the nocovariate model. e Statistics presented are the means of multiple model validation runs. f Spearman’s correlation coefficient comparing actual CO measurement and the predicted CO value. g yij ¼ CO exposure (p.p.m.) for child ‘‘i’’ with ‘‘j’’ measurements. h Wi ¼ Other PM sources (use of insect coils, burning of trash, burning of incense). i Xi ¼ Household-level covariates, Zij ¼ individual-level covariates; see Table 3 for details. The first portion of the table provides summary statistics for the model’s explanatory power and fit of data. The second portion summarizes the results of the model validation. a

b

Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)

177

Dionisio et al.

Measurements Below LOD In all, 12.2% of measured CO exposures were below the LOD of the Drager tubes. Although these measurements were not distinguishable from zero, the underlying (unobserved) exposure had a probability of being nonzero. We estimated the probability of a child having a CO measurement that was above or below the LOD based on all measurements. We used a logistic mixed effects model, including a random intercept for child and the same fixed effects as the complete mixed effects model. We also used the same logistic model to assess predictors of a child’s exposure being in the top 75th percentile of measurements (exposure 41.3 p.p.m.), that is, predictors of being high risk. The linear exposure model was fitted using measured data above the LOD, and used to predict usual exposure for all children, with the estimate conditional on a child having had a measurement above the LOD. This estimate was then multiplied by the outcome of the logistic model for having a measurement above the LOD to calculate each child’s expected exposure, accounting for the fact that each child had a probability of having an undetectable exposure. We estimated the uncertainties in the predicted CO exposures to account for the uncertainty associated with both of the above models. The uncertainty from the linear mixed effects model was estimated using the posterior predictive distributions of CO exposures (Gelman and Hill, 2007), represented by 500 draws from this distribution. We used these draws together with 500 draws from a Bernoulli distribution based on the probability of being above the LOD from the logistic model to calculate the overall uncertainty in the estimated CO exposure of individual children resulting from both models. All analyses were conducted in R version 2.10.0.

Results and discussion Study Population The mean and median ages of children included in analysis at the time of their first exposure measurement were 19.7 and 17.6 months, respectively (range: 2.4–61.2 months). Of the 1181 children included in the statistical analysis, 671 had 1 completed CO measurement (completed measurement defined as the 72 h measurement or the 48 h measurement if the former was not available), 216 had 2 completed measurements, and 294 had Z3 completed measurements, leading to a total sample size of 2263. A total of 1088 (48%) of the measurements were in the rainy months (July–October) and the remainder during the dry months. In all, 80.0% of study households cooked in an enclosed cookhouse, which may be attached to or separate from the main house, during the dry season, with the proportion rising to 84.0% in the rainy season; the remainder cooked outdoors (Table 1). More than 95% of households used biomass as 178


their main fuel, with 85.5% using either collected or purchased firewood and 12.4% using charcoal (Table 1). Fuel type changed little by season (Supplementary Table 2). Stoves were used for an average of 5.9 h per day (Table 1). Wood is burned in an open three-stone fire, and charcoal in a metal stove. More than one half of households burned incense every day, 27.7% of households burned insect coils, and 17.7% burned trash in the compound (Table 1).

Children’s CO Exposure The mean of all 2263 measured CO exposures was 1.04±1.46 p.p.m. (geometric mean ¼ 0.74 p.p.m.); it was 1.18±1.50 p.p.m. for those above the LOD. Mean of annual usual CO exposures, calculated as a time-weighted average of usual exposures in the rainy season and the dry season, from the above two models was 0.95±0.50 p.p.m. (geometric mean ¼ 0.85 p.p.m.). Mean measured CO exposure was 0.67±1.07 p.p.m. in Basse (n ¼ 881) and 1.27±1.62 p.p.m. in Banjul (n ¼ 1382; Figure 3a). Mean CO exposure in our study was lower than that of children in the highlands of Guatemala, where mean 48 h exposure was B2 p.p.m. (McCracken et al., 2009). Possible explanations for the higher exposures in Guatemala are the highland location, where temperatures are cooler, possibly leading to differences in duration and location of cooking and housing characteristics. The two studies also used different brands of CO tubes. Using the PM–CO relationship in a pilot study (Dionisio et al., 2008), CO exposure of 1 p.p.m. corresponds to a PM2.5 exposure of 184 mg/m3, with each additional p.p.m. of CO exposure adding 64 mg/m3 to PM2.5 exposure. The measured exposures had a distinct seasonal pattern, and were higher during the rainy months of July through October (when mean exposure was 1.33±1.62 p.p.m.; Figure 3b). The role of such exposure seasonality on that of pneumonia (Selwyn, 1990; Brewster and Greenwood, 1993) should be investigated. CO exposure may be higher during the rainy season due to longer stove use, more time spent near or inside the cookhouse, or differences in fuel, for example, wet vs dry fuel. Season of measurement was associated with higher exposure in the mixed effects model and was also an important predictor of having an exposure above the LOD. Measurements done in the rainy season had 10.2 (95% CI: 6.4, 16.2) times the odds of being above the LOD relative to those done in the dry season (Table 3). There was no apparent trend in CO exposure by age (data not shown). Cooking with charcoal or purchased firewood were both associated with higher child CO exposure (Table 3). The specific finding on the effect of purchased vs collected firewood requires further investigation, although we note that main type of firewood used differed by study site; in Banjul 61% of households used purchased firewood, whereas in Basse 85% of study households used collected firewood. Purchased firewood may be smoldered to conserve fuel, potentially creating more CO. Using insect coils and burning Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)


Dionisio et al.

Table 3. Estimates of the coefficients and 95% confidence intervals for the linear mixed effects model and the logistic mixed effects model. Linear mixed effects modela

Dependent variable Constant

ln(yij)b Coefficient (95% CI) 0.63 (1.11, 0.14)

Type of fuel used most for cooking Collected firewood 0.00 Purchased firewood 0.20 (0.07, 0.33) Charcoal 0.37 (0.19, 0.54) Other 0.22 (0.09, 0.53) Measurement season Dry Rainy Study site Banjul Basse Other PM sources Frequency of incense burning Sometimes or never Every day Insect coil burning No Yes

0.00 0.48 (0.40, 0.56)

0.00 0.24 (0.39, 0.08)

Figure 3. (a) Measured child CO exposure by site. (b) Measured child CO exposure by month. Eight measured CO exposures greater than 10 p.p.m. are not shown because they reduced the resolution of the vertical axis. Total sample size for both plots is 2263 measurements.

incense increased the odds of an exposure greater than the LOD by 2.1 (95% CI: 1.2, 3.6) and 1.5 (1.0, 2.1), respectively; both factors were also associated with higher child CO exposure (Table 3). In the dry season, cooking outside had a nonsignificant association with lower exposure and having above LOD exposure. Surprisingly, a significant yet opposite association was seen in the rainy season (OR ¼ 2.8; 1.1, 7.3; Table 3). The frequency of carriage on the mother’s back during cooking was not associated with higher exposure or having exposure above the LOD. Previous studies (Armstrong and Campbell, 1991; de Francisco et al., 1993; O’Dempsey et al., 1996) had found an association between this variable and Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)

1.00 1.13 (0.68, 1.89) 3.17 (1.18, 8.52) 1.29 (0.28, 5.90)

1.00 10.16 (6.38, 16.19)

1.00 0.66 (0.36, 1.20)

1.00 1.46 (1.02, 2.10)

0.00 0.34 (0.24, 0.45)

1.00 2.08 (1.22, 3.56)

0.01 (0.02, 0.04)

Location of cooking in rainy season Inside (inside main house 0.00 or inside separate cookhouse) Outside (in open 0.29 (0.08, 0.49) air or outside under roof/overhang) Location of cooking in dry season Inside (inside main house 0.00 or inside separate cookhouse) Outside (in open air 0.12 (0.31, 0.07) or outside under roof/overhang) Number of tobacco 0.04 (0.01, 0.09) smokers in the house Household socioeconomic 0.01 (0.02, 0.05) status index Individual-level covariates Child’s sex Male Female

logit(pij)c OR (95% CI) 1.51 (0.19, 12.19)

0.00 0.10 (0.01, 0.19)

Trash burning in the family compound No 0.00 Yes 0.22 (0.34, 0.10) Household-level covariates Duration of stove use per day (hours)

Logistic mixed effects model for exposure above LODa

0.00 0.03 (0.06, 0.12)

1.00 0.83 (0.49, 1.40)

0.94 (0.83, 1.08)

1.00

2.82 (1.09, 7.31)

1.00

0.76 (0.35, 1.65)

1.06 (0.83, 1.35) 1.21 (1.04, 1.41)

1.00 0.97 (0.67, 1.40)

179


Dionisio et al.

Table 3. Continued Linear mixed effects modela

Age of child at time of measurement (days)

Logistic mixed effects model for exposure above LODa

0.00 (o0.01, o0.01)1.00 (1.00, 1.00)

Frequency of carriage on the mother’s back during cooking Most of the time 0.00 1.00 Sometimes 0.03 (0.15, 0.22) 1.13 (0.52, 2.42) Never 0.02 (0.21, 0.18) 1.02 (0.46, 2.28) Primary caregiver of the child Mother 0.00 Other relative 0.39 (0.62, 0.16)

1.00 1.36 (0.42, 4.48)

Amount of time child spends outside during daylight hours Most or all 0.00 1.00 Some, a little, or none 0.04 (0.09, 0.17) 0.83 (0.48, 1.42) a Adjusted for household-level covariates (number of people living in the family compound, number of people eating from the same pot as the child, number of people living in the child’s house, number of wives the father has, number of mother’s children who are alive (excluding study subject)); individual-level covariates (time spent near fire (calculated as described in Supplementary Text), mother’s age (years), mother’s ethnicity, father’s ethnicity, mother’s occupation, and father’s occupation). b yij ¼ CO exposure (p.p.m.) for child ‘‘i’’ with ‘‘j’’ measurements. c pij ¼ Probability of CO exposure for the ‘‘jth’’ observation for child ‘‘i’’ being greater than the LOD.

pneumonia, leading to the hypothesis of this childcare behavior being associated with higher exposure. Possible reasons for the differences between our result and previous studies include residual confounding in the previous study, CO being an inappropriate exposure metric for studying health effects, and that being carried on the mother’s back exposes a child to high intensity peaks that may independently be relevant for health effects (Ezzati and Kammen 2001). The role of childcare behavior in exposure, on average and for peak exposure, and its effects on disease risk requires further investigation. The most salient predictors of exposure being above the 75th percentile were similar to the predictors for being greater than the LOD (Supplementary Table 3). Season was associated with exposure being in the highest quartile, with an OR of 4.2 (95% CI: 3.1, 5.7) for measurements in the rainy season compared with the dry season. Insect coil burning was also a significant predictor of an exposure in the highest quartile. Children living in homes using purchased firewood or charcoal as the main fuel for cooking have a 2.0 (95% CI: 1.2, 3.2) or 3.8 (2.1, 7.1) times increased odds, respectively, of having a CO exposure above the 75th percentile relative to homes where collected firewood is used as the main cooking fuel (Supplementary Table 3). In addition to the above predictors, trash burning has an unexplained significant effect of lowering the odds of a high CO exposure (OR ¼ 0.5; 0.3, 0.7; Supplementary Table 3). Reasons for these results can be described by the same explanations given above. 180

Model Performance and Validation Adding covariates to the model led to explaining a larger proportion of the between-child variance as compared with the no-covariate model (ranging from 18.8% for the fuel model to 45.7% for the complete model; Table 2). Correspondingly, with an increasing number of covariates, a smaller proportion of the total variance was explained by differences in between-child characteristics, seen in the decreasing intraclass correlation (r; Table 2). Of the additional 45.7% of between-child variance that is explained by the complete model as compared with the no-covariate model, 39.3% is due to the addition of season and other PM sources as covariates in the ‘‘parsimonious’’ model. Adding location of cooking and study site provided little additional explanatory power compared with the parsimonious model. The MSE between the predicted and measured CO exposures decreases as covariates are successively added to the model, indicating an increase in predictive validity. Specifically, the MSE decreased by 14.9% for the fuel þ season model, 19.5% for the parsimonious model, and 20.2% for the complete model relative to that of the nocovariate model (Table 2). There was also an increasing trend in the Spearman’s correlation coefficient, as the models increased in complexity from the fuel model to the complete model, ranging from 0.264 to 0.464 (Table 2). We note that both the MSE and the Spearman’s correlation coefficient statistics underestimate the model performance, because of the error in measured exposures as an indicator of usual exposure. Limitations Similar to all field measurement studies, our work is affected by some limitations. The lack of equipment for measuring children’s exposure in developing countries led us to use CO diffusion tubes, which have functional limitations and also require the measurement of color change by field workers. Although we conducted thorough training and testing of field workers, our data may be affected by a small amount of inter-reader variability. CO concentrations measured by Drager tubes depend on temperature, for example, concentrations should be multiplied by 1.1 when used at 30 1C. As we did not have access to reliable daily temperature data at each study site and did not adjust concentrations for temperature, our results should be considered slightly conservative. Data on regression model covariates were from questionnaires, and hence may be subject to uncertainty despite extensive training of field workers. This would lead to prediction outcomes with greater uncertainty than if covariates were measured without error. Lastly, although the authors collected 41 CO exposure measurement on 57% of the study subjects, it would have been desirable to collect 3–4 exposure measurements on every child in the study to better explain the within-child variability, which accounts for a large proportion of overall variation in exposure measurements. Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)


Conclusions Our results indicate that on average children younger than 5 years of age in The Gambia who participated in our study have a relatively low CO exposure. Nonetheless, burning of insect coils in the home, using charcoal, and current season being rainy are associated with high exposure levels in this setting. We also found that the day-to-day variability in exposures related to cooking are large, making each single measurement a poor indicator of usual exposure (Ezzati et al., 2000; McCracken et al., 2009). Future studies should use a combination of measurements and statistical methods to reduce the error in estimated usual exposure. Further, the parsimonious model performed nearly as well as the complete model in predicting child CO exposure, indicating that a limited number of measurements, together with data on season, fuel, and other PM sources, would provide most of the information needed for estimating usual exposure.

Conflict of interest The authors declare no conflict of interest.

Acknowledgements This work was supported by a grant from the National Institute of Environmental Health Sciences (1R21ES01785501). We thank the households who participated in the study for their help and hospitality, our field workers and field supervisors for valuable assistance in data collection, the Biomedical Engineering Department at the MRC for technical assistance throughout the study, Jose Vallarino for information on methods and instruments for personal exposure measurement, and Mariel Finucane for advice on presentation of statistical results. We also thank Grant Mackenzie for operational support, Kim Mulholland, Philip Hill, Brian Greenwood, and Peter Smith for advice on case– control study design, and Nigel Bruce and Kirk Smith for advice on exposure measurement.

References Armstrong J.R., and Campbell H. Indoor air pollution exposure and lower respiratory infections in young Gambian children. Int J Epidemiol 1991: 20(2): 424–429. Balakrishnan K., Sankar S., Parikh J., Padmavathi R., Srividya K., and Venugopal V., et al. Daily average exposures to respirable particulate matter from combustion of biomass fuels in rural households of southern India. Environ Health Perspect 2002: 110: 1069–1075. Black R.E., Cousens S., Johnson H.L., Lawn J.E., Rudan I., and Bassani D.G., et al. Global, regional, and national causes of child mortality in 2008: a systematic analysis. Lancet 2010: 375(9730): 1969–1987.

Dionisio et al.

Brewster D., and Greenwood B. Seasonal variation of paediatric diseases in The Gambia, West Africa. Ann Trop Paediatr 1993: 13(2): 133–146. Bruce N., McCracken J., Albalak R., Schei M.A., Smith K.R., and Lopez V., et al. Impact of improved stoves, house construction and child location on levels of indoor air pollution exposure in young Guatemalan children. J Expo Anal Environ Epidemiol 2004: 14(Suppl 1): S26–S33. Dasgupta S., Huq M., Khaliquzzaman M., Pandey K., and Wheeler D. Who suffers from indoor air pollution? Health Policy Plan 2006: 21(6): 444–458. de Francisco A., Morris J., Hall A.J., Armstrong Schellenberg J.R.M., and Greenwood B.M. Risk factors for mortality from acute lower respiratory tract infections in young Gambian children. Int J Epidemiol 1993: 22: 1174–1182. Dherani M., Pope D., Mascarenhas M., Smith K.R., Weber M., and Bruce N. Indoor air pollution from unprocessed solid fuel use and pneumonia risk in children aged under five years: a systematic review and meta-analysis. Bull World Health Organ 2008: 86(5): 390–398C. Dionisio K.L., Howie S., Fornace K.M., Chimah O., Adegbola R.A., and Ezzati M. Measuring the exposure of infants and children to indoor air pollution from biomass fuels in the Gambia. Indoor Air 2008: 18(4): 317–327. Ezzati M., and Kammen D.M. Indoor air pollution from biomass combustion and acute respiratory infections in Kenya. Lancet 2001: 358(9282): 619–624. Ezzati M., Saleh H., and Kammen D.M. The contributions of emissions and spatial microenvironments to exposure to indoor air pollution from biomass combustion in Kenya. Environ Health Perspect 2000: 108(9): 833–839. Filmer D., and Pritchett L.H. Estimating wealth effects without expenditure dataor tears: an application to educational enrollments in states of India. Demography 2001: 38(1): 115–132. Gelman A., and Hill J. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press, New York, 2007. Honaker J., King G., and Blackwell M. Amelia II: A Program for Missing Data. http://gking.harvard.edu/amelia. Accessed 2010. Honaker J., and King G. What to do about missing values in time series crosssection data. Am J Pol Sci 2010: 54(2): 561–581. King G., Honaker J., Joseph A., and Scheve K. Analyzing incomplete political science data: an alternative algorithm for multiple imputation. Am Pol Sci Rev 2001: 95: 49–69. Laird N.M., and Ware J.H. Random-effects models for longitudinal data. Biometrics 1982: 38(4): 963–974. McCracken J.P., Schwartz J., Bruce N., Mittleman M., Ryan L.M., and Smith K.R. Combining individual- and group-level exposure information: child carbon monoxide in the Guatemala Woodstove Randomized Control Trial. Epidemiology 2009: 20(1): 127–136. Mestl H.E.S., Aunan K., Seip H.M., Wang S., Zhao Y., and Zhang D. Urban and rural exposure to indoor air pollution from domestic biomass and coal burning across China. Sci Total Environ 2007: 377: 12–26. Naeher L.P., Smith K.R., Leaderer B.P., Neufield L., and Mage D.T. Carbon monoxide as a tracer for assessing exposures to particulate matter in wood and gas cookstove households of highland Guatemala. Environ Sci Technol 2001: 35: 575–581. Northcross A., Chowdhury Z., McCracken J., Canuz E., and Smith K.R. Estimating personal PM2.5 exposures using CO measurements in Guatemalan households cooking with wood fuel. J Environ Monit 2010: 12: 873–878. O’Dempsey T.J., McArdle T.F., Morris J., Lloyd-Evans N., Baldeh I., and Laurence B.E., et al. A study of risk factors for pneumococcal disease among children in a rural area of West Africa. Int J Epidemiol 1996: 25(4): 885–893. Selwyn B.J. The epidemiology of acute respiratory tract infection in young children: comparison of findings from several developing countries. Rev Infect Dis 1990: 12(Suppl 8): S870–S888. Smith K.R., Mehta S., and Maeusezahl-Feuz M. Indoor air pollution from household solid fuel use. Comparative Quantification of Health Risks: Global and Regional Burden of Disease Attributable to Selected Major Risk Factors. World Health Organization, Geneva, 2004, 1435–1493. Wagstaff A. Socioeconomic inequalities in child mortality: comparisons across nine developing countries. Bull World Health Organ 2000: 78: 19–29. Xu R. Measuring explained variation in linear mixed effects models. Stat Med 2003: 22(22): 3527–3541.

Supplementary Information accompanies the paper on the Journal of Exposure Science and Environmental Epidemiology website (http://www.nature.com/jes)

Journal of Exposure Science and Environmental Epidemiology (2012) 22(2)

181