Air pollution dispersion from biomass stoves to neighboring homes in Mirpur, Dhaka, Bangladesh

This investigation was conducted in a densely populated (35,200 people/km², 2011) low-income area of Mirpur, Dhaka, Bangladesh, where the type of cooking fuel used is heterogeneous [3]. Mirpur homes are typically arranged in compounds, with several one-room homes immediately adjacent to each other, a shared latrine, and a small central courtyard or walkway; compounds are often adjacent to other compounds with shared walls (Fig. 1).

We recruited clusters of index and neighbor homes. Index homes exclusively used biomass fuels (dung, crop residue/grass, rice husk, dried leaves, coal, charcoal, wood, and/or bamboo) and a traditional stove for cooking. Only one index home was recruited per compound. Once the index home was identified, we recruited three to four neighbor homes, defined as homes immediately surrounding the index home in any direction that primarily used one of the following cleaner fuels for cooking: natural gas, biogas, electricity, and/or kerosene. Index and neighbor homes located on an upper level (above ground level) of an apartment building were ineligible, because of potential variation in sources, levels, and movement of air pollutants on upper levels of a multi-story building compared to ground level. Eligible respondents intended to reside in their current home for the subsequent one week. If more than four neighbor homes were eligible, we prioritized recruiting those neighboring homes that had main entrances nearest to the main entrance of the index home. It is possible that other nearby homes that we did not recruit may have used biomass-burning stoves.

Once eligible index and neighbor homes were identified, we sought informed consent from the adult who usually cooks in each household. As many participants were unable to read, a consent form was read aloud to them, and they were asked to either sign or provide a thumbprint as a mark of consent. The first ten clusters of households (index and neighbor homes) visited that met all eligibility criteria and consented to participate were enrolled. The study protocol was approved by the institutional review boards at icddr,b and the University at Buffalo. No interventions were performed, therefore this is not registered as a trial.

After obtaining written informed consent, study staff administered a questionnaire to each participant to elicit information on demographics, stove type and location, and behaviors that may affect exposure to air pollution, including cigarette smoking, ventilation behavior (opening and closing windows and doors, turning fans on and off), cooking fuel use, and burning substances other than for cooking. The participant was then asked to show his/her home and cooking space. Study staff measured the size of the main sleeping space of the home using a measuring tape. Air quality in the main sleeping space serves as a proxy for individual exposure, as many homes consist of only one room: the sleeping space where family members spend substantial time. Staff recorded the location of the main cooking area, including distance from the home to the cooking area (if external to the home) and the distance from index to neighbor homes in steps (approximately 0.4 m per step). Study staff drew sketches of each cluster, including the index home and all neighboring homes of the index home, including participants and non-participants. Locations of primary stoves, walls, doors, and windows were indicated on the sketch.

After administering the baseline questionnaire and taking measurements of dimensions of the home, study staff installed two sets of PM_2.5 and CO monitors for each index and neighbor home: one as close as possible to the primary stove, and one inside the main sleeping space, above the participant’s bed. Additional monitors were placed at one outdoor location per cluster. The outdoor monitors were located between the main entrances of the index home and neighbor home(s), under a roof or eave and out of plain sight, in order to protect instruments from rain and theft. Due to the variety in structure of local homes, we were unable to standardize the locations of these outdoor monitors. PM_2.5 and CO monitors recorded measurements once per minute for at least 24 h. We used University of California, Berkeley Particle Monitors (University of California, Berkeley, Berkeley, CA, USA) to measure PM_2.5 concentrations and Lascar CO monitors (Lascar Electronics, Salisbury, UK) to measure CO concentrations. The lower limit of detection of the University of California, Berkeley Particle Monitors is 50 μg/m³, which is twice the WHO-recommended 24-h mean PM_2.5 concentration of ≤25 μg/m³ [2, 3, 10, 21‐23]. The lower limit of detection for Lascar CO monitors is 0 ppm, with a resolution of 0.5 ppm. The WHO recommended 24-h mean CO concentration is ≤6.11 ppm [24].

Study staff conducted detailed observations of household activities for approximately 8 h on the working day following the initial interview, concurrent with PM_2.5 and CO monitoring. Study staff recorded cooking activities and fuels used, ventilation activities, smoking, and burning substances other than for cooking (e.g. for fragrance, lighting, mosquito control, or garbage disposal). Study staff was stationed near each index home in order to prioritize recording activities occurring at the index home. However, they moved about the cluster as needed in order to record activities occurring at the participating neighbor homes. Residents were asked to continue normal daily activities as much as possible while monitoring was taking place.

We calculated geometric mean PM_2.5 and CO concentrations at each monitor location for the following times: between 5 and 6 am (baseline), during observed biomass cooking in the index home, the observation period during which no biomass cooking occurred in any home, and 24 h of monitoring time. Previous research has shown that the lowest PM_2.5 concentrations are observed in Mirpur from 5 to 6 am [4], so the 5–6 am hour served as a baseline for this study. We used data from observations to identify biomass cooking times in the index home. Although neighbor homes primarily used cleaner fuels, our observations indicated that some biomass fuel use occurred when cleaner fuels were unavailable. For our measure of biomass cooking time in an index home, we excluded any times during which an enrolled neighbor home was also observed to cook with biomass fuel.

We used ANOVA to describe whether geometric mean PM_2.5 and CO concentrations varied by monitor location at the pre-specified times. Our independent variable was location of the monitor, categorized as follows: near the index stove, in the index home’s main sleeping space, monitors (near the stove and in the main sleeping space) of neighbor homes that share a wall with the index home, outdoor monitors, and monitors (near the stove and in the main sleeping space) of neighbor homes that do not share a wall with the index home. Although PM_2.5 and CO are expected to dissipate differently, we still expected that the order of dissipation would be the same, that is, highest concentrations for both pollutants would be highest in cooking space of biomass homes, then living space, then neighbor homes with a shared wall, then outside, then neighbor homes with no shared wall. Therefore, the ranking of environments is the same for both pollutants. We added a constant of 0.001 for all CO values in order to calculate the geometric mean, which requires nonzero values; the lowest possible value of the geometric mean CO concentration was therefore 0.001 ppm. We substituted a value equal to half the limit of detection for all PM_2.5 readings at or below the limit of detection, a common method of analyzing data from instruments with high limits of detection [25]. Therefore, the lowest possible value of the geometric mean PM_2.5 concentration was 25 μg/m³. We have provided a supplemental table comparing results using this method to results substituting a value equal to the limit of detection for all PM_2.5 readings at or below the limit of detection (Additional file 5).

Using the PM_2.5 and CO measurements at the index stove monitor as the referent category, we used linear regression to describe the relationship between the location of a monitor (index home, neighbor home—shared wall, outdoor, neighbor home—no shared wall) and geometric mean PM_2.5 and CO concentrations during index biomass cooking and 24 h of monitoring. We examined the following as potential confounders: building material(s) of the home, having a smoker living in the home, distance to the index home (in steps), area of the home, having a secondary biomass stove, ambient temperature, and relative humidity. Covariates that changed the bivariate measure of association by 10% were retained in the final models. We examined ventilation of the home (presence of at least one window in the home) and location of the index stove (indoor or outdoor) as potential effect modifiers by stratifying results by these factors.

We also performed spatial variogram analysis to describe the relationship between geometric mean PM_2.5 and CO concentrations in index and neighbor homes, limited to index stove cooking times, and Euclidian distance from a PM_2.5 or CO source (the biomass-burning stove). Variograms are used to describe variability in PM_2.5 or CO concentrations as a function of the distance between the source of pollution (biomass stove) and the monitor.

We recruited a total of 44 homes for this study (10 index homes and 34 neighboring homes). Index homes and neighboring homes shared many similar characteristics, including smoking in the home, low education levels, and high proportion of homes with only one room (Table 1). Neighbor homes were more likely to have a window compared to index homes. Most cooking areas were located inside the main housing structure, but some were located in a separate room from the sleeping space. Separation of the sleeping space from the stove was more common in index compared to neighbor homes; four of ten index home stoves were located outside. The most common type of biomass fuel used in index homes was wood (n = 9), followed by small sticks of wood or bamboo (n = 1). Supplemental fuels used in index stoves included: paper (n = 6), kerosene (n = 5), wood (n = 5), plastic (n = 3), and small sticks of wood or bamboo (n = 2). Although neighbor homes almost exclusively used electricity as their primary cooking fuel, 21 of 34 had a secondary biomass stove, likely used when electricity was unavailable [3].

We observed cooking events, window opening events, door opening events, and fan use events over approximately eight hours of structured observation for each cluster (Table 2). Cooking patterns varied between index and neighbor homes. Of 84 observed cooking events, 17 (20%) were biomass cooking events, 13 (76%) of which occurred in nine of ten index homes (we did not observe cooking in one index home). We observed 65 electricity cooking events (77% of cooking events), all of which occurred in neighbor homes. The median duration of cooking events with electricity (65 min) was considerably shorter than for those using biomass fuel (137 min), however, the total amount of time spent cooking was similar among the two fuel groups (median of 189 and 185 min for electric and biomass stoves). Homes that cooked with electricity tended to cook several times during the observation period, whereas most of those using biomass fuel only cooked once. We observed no smoking or burning of substances other than for cooking during structured observations in either index or neighbor homes.

Participants from index and neighbor homes had similar ventilation behaviors. Twenty-three participants had a single window. Of these, twenty-two were kept open during the entire structured observation with the remaining window kept open for 292 min (nearly five hours). Likewise, doors were open for the vast majority (482 min, approximately six hours) of the structured observation period. Fan use events were the most frequently observed events (median 111 min among index homes, 115 min among neighbor homes); many participants turned their fans on and off throughout the day. Participants had their fans on for a median of over six hours during the eight-hours observation periods in both index and neighbor homes. All participants who had a window kept it open during cooking, and almost all participants had their doors open (n = 42, 95%) and fans on (n = 38, 86%) during cooking.

We monitored PM_2.5 and CO for a total of 95 locations, at least 24 h each, from 44 households. Due to equipment error, five PM_2.5 recordings had missing data for 12 or more hours. Additionally, one index stove produced extremely high concentrations of PM_2.5 and CO (over 10 times the mean of other homes during biomass cooking), so we removed the monitors from that home from our analyses. Data analysis was performed using the data from the remaining 88 PM_2.5 recordings and 93 CO recordings. In one cluster, we did not observe biomass cooking in the index home during our observation period. For this cluster, we estimated index cooking time based on elevated CO concentrations in the index home at a common cooking time during the day prior to observation, as biomass cooking was the only obvious source of sustained elevated CO concentrations for all homes in the study.

At baseline (5-6 am), geometric mean concentrations at all monitor locations were at the lowest possible values for both PM_2.5 (25 μg/m³) and CO (0.001 ppm) (Table 3). During index stove biomass cooking events, increased geometric mean PM_2.5 concentrations were recorded at all monitors in all homes. Geometric mean concentrations of PM_2.5 during index stove biomass cooking events were highest at the index stove (mean 326.3, SD 211.6 μg/m³), followed by the sleeping space in the index home (mean 322.7, SD 408.1 μg/m³), neighbor home that shares a wall with the index home (mean 278.4, SD 423.7 μg/m³), outdoor (mean 154.2, SD 80.6 μg/m³), then neighbor home that does not share a wall with the index home (mean 83.1, SD 49.2 μg/m³) (ANOVA p-value = 0.005)—Additional file 3 shows PM_2.5 concentrations for one cluster during biomass cooking. Concentrations of CO during index stove biomass cooking events were also highest at the index stove (mean 12.3, SD 13.2 ppm), followed by index home (mean 11.2, SD 18.8 ppm), neighbor home that shares a wall with the index home (mean 3.6, SD 7.3 ppm), outdoor (mean 0.7, SD 1.3 ppm), then neighbor home that does not share a wall with the index home (mean 0.2, SD 0.6 ppm) (ANOVA p-value< 0.0001)—Additional file 4 shows CO concentrations for one cluster during biomass cooking. During observation periods with no biomass cooking, we observed no differences in PM_2.5 concentrations at different locations (p = 0.6), although CO concentrations were slightly lower at outdoor locations and in neighbor homes that did not share a wall with the index home (p = 0.1). During 24 h of monitoring, average geometric mean PM_2.5 and CO concentrations were also highest at the index stove, then in the index home and outdoor, then in neighbor homes (p = 0.1 for PM_2.5, p = 0.007 for CO). There were no differences in 24-h PM_2.5 and CO concentrations between neighbor homes with and without a shared wall. Additional file 5 compares PM_2.5 concentrations 1) substituting readings at or below the limit of detection with half of the limit of detection (25 μg/m³). as in the main analysis with 2) substituting readings at or below with limit of detection with the limit of detection (50 μg/m³). Baseline and 24-h means were about 25 μg/m³ higher than in the main analysis, as expected, but the difference between the two methods was not as great during index stove cooking and non-cooking times. Associations of PM_2.5 concentration by stove location were similar using both methods.

Multivariable linear regression models to examine the association between location of the monitor and PM_2.5 and CO concentrations included two covariates: distance to the index home in steps (index homes had a value of 0 steps) and the presence of a secondary biomass stove in the home. We did not account for smoking in the home as no smoking was observed to take place during biomass cooking times. In exploratory analyses, PM_2.5 and CO concentrations were not substantially different during electricity cooking periods compared to periods of no cooking, and kerosene cooking events were few, so we did not account for electricity or kerosene cooking in our models.

During index stove biomass cooking events, we observed a trend of lower geometric mean PM_2.5concentrations with monitor locations further from the index stove overall (p for trend 0.03) (Fig. 2 and Additional file 1). In stratified models, we did observe a trend of lower geometric mean PM_2.5 in clusters with no windows in the index home (p for trend 0.03), but this trend was clearer in clusters with at least one window in the index home (p for trend = 0.006), particularly for neighbor homes that share a wall with the index stove. We observed a trend similar to the overall data in clusters with indoor index stoves (p for trend 0.04), but not in clusters with outdoor index stoves (p for trend 0.8). We did not observe any associations between location of the monitor and 24-h geometric mean PM_2.5 concentrations (Additional file 2). Consistent with spatial dependence in PM_2.5 from biomass burning, variogram analysis of PM_2.5 concentrations showed a steadily increasing semivariance in PM_2.5 up to distances of around five meters from the biomass stove, suggesting that spatial dependence in PM_2.5 was limited at further distances (Fig. 3).

Reduced CO concentrations during index stove cooking were associated with location of monitor relative to index stove (p for trend 0.006) overall, in homes with at least one window (p for trend 0.005), and in clusters with indoor index stoves (p for trend 0.03) (Fig. 4 and Additional file 1). Geometric mean CO concentrations were very low, 0.03 ppm or less, over 24 h. However, we did observe a trend of lower CO concentrations at monitoring locations further from the index stove overall (p for trend 0.1) and in homes with at least one window (p for trend 0.1), though not significant (Additional file 2). Semivariance appeared to increase at distances beyond than that observed with PM_2.5, consistent with the lighter CO particles travelling farther than PM_2.5 (Fig. 5).