Effects of household and participant characteristics on personal exposure and kitchen concentration of fine particulate matter and black carbon in rural Honduras

Abstract

Traditional cooking with solid fuels (biomass, animal dung, charcoals, coal) creates household air pollution that leads to millions of premature deaths and disability worldwide each year. Exposure to household air pollution is highest in low- and middle-income countries. Using data from a stepped-wedge randomized controlled trial of a cookstove intervention among 230 households in Honduras, we analyzed the impact of household and personal variables on repeated 24-hour measurements of fine particulate matter (PM_2.5) and black carbon (BC) exposure. Six measurements were collected approximately six-months apart over the course of the three-year study. Multivariable mixed models explained 37% of variation in personal PM_2.5 exposure and 49% of variation in kitchen PM_2.5 concentrations. Additionally, multivariable models explained 37% and 47% of variation in personal and kitchen BC concentrations, respectively. Stove type, season, presence of electricity, primary stove location, kitchen enclosure type, stove use time, and presence of kerosene for lighting were all associated with differences in geometric mean exposures. Stove type explained the most variability of the included variables. In future studies of household air pollution, tracking the cooking behaviors and daily activities of participants, including outdoor exposures, may explain exposure variation beyond the household and personal variables considered here.

1. Introduction

Biomass and coal remain primary cooking fuels for over 3 billion people worldwide, especially in low- and middle-income countries; the resulting household air pollution contributes to over 4 million premature deaths annually (Murray et al., 2020; WHO, 2021). Previous studies have associated household air pollution with health complications such as increased lower and upper respiratory infections, chronic obstructive pulmonary disease, childhood pneumonia and asthma, high blood pressure, and death among others (Murray et al., 2020; Quansah et al., 2017). Rural households and communities often fall behind urban areas in access to and adoption of clean cooking/energy opportunities (Pollard et al., 2018). The adverse impacts of household air pollution in rural settings may also be compounded by socio-cultural and economic challenges such as less diverse employment options, fewer opportunities to connect to national power grids or take advantage of private suppliers, and both the normalization of and freely available firewood and other biomass for collection and use.

Earlier studies have indicated that specific household and participant characteristics influence fine particulate matter (PM_2.5) and/or black carbon (BC) concentrations. PM_2.5 is associated with significant morbidity and mortality around the world, including household exposures (Murray et al., 2020). Previous studies have shown that BC is harmful to human health (Baumgartner et al., 2014; Janssen et al., 2011; United States Environmental Protection Agency, 2011) and there is sufficient evidence from cohort studies to substantiate associations between daily changes in BC and short-term health impacts such as cardiopulmonary hospitalization and cardiovascular mortality (WHO, 2012).

Many previous studies have used a cross-sectional analysis and did not record repeated measures. Based on a cross-sectional analysis of a biomass cookstove exposure study in Honduras, Clark et al. (2010) reported that stove quality and kitchen ventilation affected personal exposure and kitchen concentrations of PM_2.5 and carbon monoxide. In a cross-sectional study in Puno, Peru, Fandiño-Del-Rio et al. (2020) highlighted the number of rooms in participant households, roof material (e.g., metal sheets), and the existence of a chimney or recessed area for stove placement as variables leading to increased PM_2.5 exposure.

Though not as frequently studied when compared to PM_2.5 exposure, BC, formed through the incomplete combustion of carbonaceous fuels, is a component of increased interest in household air pollution, specifically in low- and middle-income countries. Exposure to BC in indoor settings is primarily due to biomass burning (Shupler et al., 2020) In a cross-sectional analysis of personal and kitchen BC concentrations from eight countries, Wang et al. (2021) found that personal (e.g., age, gender, hours spent cooking) and household (e.g., presence of electricity, stove types, and fuels used) characteristics together with country indicators explained 46% of household BC variation and 33% of personal BC variation. When household and personal PM_2.5 concentrations were included in models, the amount of variation explained increased to 60% and 54%, respectively. Additionally, in rural Northern India, Sharma and Jain (2020) found that personal BC exposure was highest for those using traditional cookstoves in enclosed when compared to semi-open and open style kitchens; while exposure was lower for all kitchen enclosure types with the improved cookstove in comparison to the traditional cookstove.

Few studies have used repeated measures on the same household, which can better capture variation in cooking patterns over time. One exception is a repeated measures study conducted in Beijing and Shanxi, China by Lee et al. (2021), who found that 24% and 17% of variation in 24-hour PM_2.5 and BC exposures, respectively, was explained by fuel source, smoking status, and outdoor air pollution concentrations. The most impactful variable was outdoor PM_2.5, which explained 16% of within-participant variation in personal PM_2.5.

Despite previous work to associate physical characteristics of the house and kitchen such as ventilation and building materials, characteristics of the household such as family size or social-economic status, and/or personal characteristics to household air pollution, exposure studies often vary in their approaches to measuring (e.g., study design, instruments deployed, and frequency and length of exposure data collection) and modeling (e.g., multivariable, predictive, and machine learning) exposure variables (Benka-Coker et al., 2021; Clark et al., 2010; Fandiño-Del-Rio et al., 2020; Northcross et al., 2015; Pollard et al., 2018). Our study was developed to explore the different household and personal characteristics affecting personal and kitchen measurements in a biomass stove environment using longitudinal data from rural Honduran communities. Compared to the work of Lee et al. (2021), who collected two to four 24-hour repeated measures per individual over a one-year period, our study included six repeated 24-hour measurements per participant over three years. The main objective of our study was to determine what variables explain variability in repeated measures of kitchen and personal 24-hour PM_2.5 and BC concentrations in rural Honduras. We hypothesized that the following would influence measured concentrations: presence of kerosene for lighting, presence of electricity, primary stove type used, relative size of household assets, demographic characteristics, season, and kitchen characteristics. This paper fits into the broader literature that seeks to identify variables that influence personal exposure and kitchen concentrations of PM_2.5 and BC among users of biomass cookstoves. Further, study findings highlight the public health importance to understanding intersections between household air pollution exposures and the personal or household characteristics that amplify or mediate such exposures. Our results provide key insights to improve the health efficacy of future cookstove and clean household energy interventions, transitions, and policies. To holistically understand and work at addressing the public health concerns related to household air pollution, we provide findings that exemplify the need for evaluation of multiple air pollutants, i.e., PM_2.5 and BC, and participant specific characteristics.

2. Methods

2.1. Study Overview

We conducted a stepped-wedge randomized controlled trial of a cookstove intervention near La Esperanza, Honduras between August 2015 through May 2018 (Young et al., 2019). Study design and methods (Young et al., 2019) and analysis of the effect of the assigned intervention on PM_2.5 (Benka-Coker et al., 2021) and BC (Young et al., 2022) related to this research have been published elsewhere. Participants included 230 women that were self-identified as the primary cooks in their homes; women were between 24-59 years old at time of enrollment, not pregnant, non-smoking nor exposed to secondhand smoke, and used only traditional cookstoves with biomass fuel to cook (Young et al., 2019). The trial included six repeated measures of household air pollution and survey data taken approximately every six months over three years. Women were randomly assigned to one of two study arms by blindly drawing numbers from a bag during a community meeting at Visit 2. The stepped-wedge design allowed all participants to receive the cookstove intervention–a Justa stove with an engineered combustion-chamber and chimney. Arm 1 (n=115) received the stove intervention after the second visit and arm 2 (n=115) received the intervention after the fourth visit (Young et al., 2019). Participants’ traditional stoves were destroyed when the Justa stove was installed. Due to the engineering and improved ventilation, the Justa stove is considered an improved biomass stove and will be referred to as Justa throughout the paper.

This study was approved by the CSU Institutional Review Board (#12-3870H). Prior to enrollment, and at each study visit, all participants were reminded that the study was voluntary and personal data would be kept confidential; verbal consent was then obtained by study staff from all participants. The ClinicalTrials.gov identifier is NCT02658383.

2.2. Exposure Assessment

At each study visit, 24-hour gravimetric measurements of personal and kitchen PM_2.5 concentrations were collected. Kitchen air pollution monitors were placed slightly above the woman’s breathing zone when standing (76-127 cm away from stove) and away from the main plume of smoke, windows, and doorways (Young et al., 2019); this setup was standardized across all participant homes. Personal measurements were made using a monitor worn on a cloth necklace by participants as they went about their daily activities (Young et al., 2019). During visits 1-4, measurements were made using a 2.5 μm size-selective cyclone inlet (Triplex, BGI, Inc., NJ, USA), AirChek XR5000 Pump (SKC Inc., Eighty-Four, PA, USA), and Pallflex fiberfilm filters (Fiberfilm, Pall Corporation, NY, USA) (Benka-Coker et al., 2021; Pillarisetti et al., 2019). During visits 5-6 we switched to the Teflo filters (Teflo filters, VWR, Radnor, PA, USA) (Benka-Coker et al., 2021) for both personal and kitchen measurements and the Ultrasonic Personal Air Sampler (UPAS, Access Sensor Technologies, Fort Collins, CO, USA) for personal exposure monitoring due to product availability, improvements in technology and wearability, and based on measurement correlation of the instruments between paired samples (Pillarisetti et al., 2019). Overall, the two data collection methods were highly correlated (Spearman correlation of 0.91, 95% CI: 0.84, 0.95) (Pillarisetti et al., 2019). Using the same 37mm Pallfex and Teflo filters, BC concentrations were analyzed from the samples based on the change in transmission of 880 nm light through the filters before and after sampling (Transmissometer model OT-21, Magee Scientific, Berkeley, CA, USA) (Young et al., 2019). Due to instrument variability in the light scatter calculation, we obtained negative values for some low BC concentrations that led to the substitution of BC concentrations below the fifth percentile of personal and kitchen measurements with the value of the fifth percentile of the personal measurements (5th percentile, 0.089 μg/m³; n = 60 (5%) personal, and n = 24 (2%) kitchen samples substituted) (Young et al., 2022). Full calibration and limit of detection details for PM2.5 can be found in Benka-Coker et al. (2021) and for BC in Young et al. (2022). Pearson correlations between log transformed PM_2.5 and BC were calculated.

2.3. House, Household, and Personal Characteristics

At each visit, we conducted in-person questionnaires in Spanish to elicit information about characteristics of the house, household, and participant. Questions included participant demographics, income, and socio-economic status; cooking behaviors, preferences, and opinions; characteristics about the house (location, elevation, electricity, lighting), kitchen (ventilation, height, length, and width), and household exposures such as trash burning and smoke from neighbors; and participant health history and medications, physical activity level, and diet. Stove use monitors (SUMs; Thermochron iButtons, Embedded Data Systems, Lawrenceburg, KY, USA) were deployed to collect information on 24-hour stove use via temperature measurements every five minutes.

2.4. Statistical Analysis

2.4.1. Variable selection.

The initial nineteen variables of interest were selected based on literature review (Figure 1). We conducted an exploratory data analysis and chose variables that had at least 10 μg/m³ difference in geometric mean concentration between variable categories as those that may meaningfully contribute to variability in personal and kitchen PM_2.5 concentrations and to include in the multivariable model. This variable selection was conducted using PM_2.5 concentrations; the same variables were included in models for BC.

Figure 1.

Nineteen variables of interest selected for exploratory data analysis and 11 for mixed models.

First, we computed summary statistics and conducted visual analyses of the distribution of measurements (personal and kitchen) across different levels of each variable using cumulative distribution function (CDF) and smoothed density plots. The variables measured only at baseline (i.e., Visit 1) included: participant age and years of schooling, home ownership, number of rooms in the house, and income sources. Variables measured at baseline and during each of the subsequent five visits included: number of people in the household, presence of kerosene for lighting, presence of electricity for lighting, cooking, and other uses, if and where trash was burned, exposure to smoke from neighbors, primary stove used (not assumed based on stove assignment), season, and if the household kitchen was enclosed or open and inside or outside of the house (Figure 1). Enclosure type was used as an estimation for kitchen ventilation where ‘enclosed’ meant the kitchen had four walls, the walls were sealed without gaps, and no open eaves were present; ‘semi-enclosed’ meant big gaps between the roof and eaves, door frames without doors, and/or windows that were frequently open, and ‘open’ meant few or no walls. For analysis, open and semi-enclosed were combined into one category. We investigated the possible effect of typical cooking time (hours per week) using self-reported survey information. We also measured cooking time during the sampling period via SUMs (Young et al., 2019), where cooking time was defined as proportion of time at temperatures above 28 degrees Celsius (°C). CDF plots for personal exposure and kitchen concentrations were also created for the SUM variable stratified by stove type(s) used (i.e., Justa, Justa+Traditional, Traditional).

We estimated socio-economic status via a weighted household-level asset index at baseline (see Supplemental Material Section A for description of asset index calculation). We explored the impact of seasonality and study phase on exposure. We also evaluated the number of livestock groups a household owned, and cows and horses individually, at each visit, but did not include these variables in the model due to considerable overlap with the asset index. Overall, the 11 variables that showed possible relationships with 24-hour PM_2.5 personal exposure or kitchen concentrations were included in the models (Figure 1). All data processing and analyses were conducted in R v.4.0.2 (R Core Team, 2021).

2.4.2. Linear Mixed Model.

We fit four linear mixed models separately for log-transformed personal and kitchen measurements of PM_2.5 and BC. The variables included were: whether the participating household had electricity (yes/no), the presence of kerosene for lighting (yes/no), primary (and secondary) stove type used, household weighted asset index (high/low), education level (no school; 6 years; greater than 6 years), self-reported cooking time (hours per week), stove use (categorized as 0%-60%, 60%-80%, or 80%-100% to capture different cooking trends without overinterpreting small differences in use percentages), income source (agriculture or multiple sources), location (4-level) and enclosure type (2-level) for the kitchen, and an indicator of season (rainy/dry). All models included a random effect for participant to account for repeated measurements. Because of the stepped-wedge design, visit number and date were highly correlated with stove type and were thus not included in the models. We used an α=0.05 level to determine statistical significance. We also looked at interactions between stove type and all other variables. Due to minimal difference in model results we did not include interaction terms in our final models.

We used the models to quantify the impact of each variable on geometric mean concentrations and on the exposure variation explained. Exponentiated model coefficients estimate the relationship between each variable and the geometric mean PM_2.5 and BC concentration. Overall explanatory ability of the model is quantified as both marginal and conditional R² (Nakagawa et al., 2017). Marginal R² quantifies the proportion of variation in the outcome explained by the fixed effects alone, whereas conditional R² quantifies the proportion of variation explained by the fixed effects and random effects together. To quantify the variability explained by each variable, we calculated the difference in marginal R² with and without the variables in the fully adjusted model using the r2glmm package in R (Jaeger, 2017). We additionally fit a univariate linear mixed model for each variable that included only that variable and the random intercept, to estimate the marginal percent variation explained by that variable in the absence of other variables (see Section G of the Supplemental Material). Together, the univariate and multivariable models provide information on the variability explained by each variable alone and in combination with the other variables. For comparison with Lee at al. (2021) we also calculated the “R²-within” and ‘R²-between” as described in that paper (Supplemental Material Section G).

3. Results

3.1. Characteristics of Study Population

Of the 230 women recruited for the study, two were removed from analysis because of missing information on both education level and stove-use monitoring data. Therefore, at baseline the study population included 228 women between the ages of 23 and 59, the majority (60%) were under 40 years old. Additional baseline characteristics of the cohort are provided in Table 1. Most families owned their homes (78%), and most families had four or more family members living in their home (86%). Most participants reported cooking in enclosed kitchens (76%); kitchens were mostly in a separate room attached to the house (60%) or in a separate building (28%). Related to energy use, 75% of participants reported the presence of electricity, and 11% stated that they regularly use kerosene for lighting. Lastly, after all installations had been completed (Visit 5), 39% of women were using the Justa stove (with or without a secondary improved stove) as their primary stove and 59% were using the Justa stove and had a secondary traditional stove.

Table 1.

Summary of baseline participant, household, and house characteristics for the 228 Honduran women included in the study

Participant Characteristics at Baseline a	Response Category	n	Percent
Age	< 40 years old	137	60%
	≥ 40 years old	91	40%
Education	No school	14	6%
	Compulsory schooling (6 years)	205	90%
	Greater than 6 years	9	4%
Income source	Agriculture	148	65%
	Multiple sources	80	35%
Who owns participants’ house	Immediate family	178	78%
	Neighbors	20	9%
	Extended family	8	3%
	Other or unknown	22	10%
Number of rooms in house	≥ 3 rooms	97	42%
	4-5 rooms	116	51%
	> 5 rooms	15	7%
Number of people in household	3 people or less	32	14%
	4-6 people	111	48%
	7-10 people	74	33%
	More than 10 people	11	5%
Asset Index b	Low (0-4 items)	118	51%
	High (5-9 items)	110	48%
Livestock c	None	101	47%
	Low (1-2 types)	106	44%
	High (3 or more types)	21	9%
Cooking time	0-14 hours/week	45	20%
	14-28 hours/week	136	60%
	28-42 hours/week	41	18%
	42-128 hours/week	6	2%
Stove Use Monitoring – % of stove use time >28 °C	0-60%	296	22%
	60-80%	607	44%
	80-100%	184	13%
	Missing Values	280	20%
Primary kitchen enclosure	Enclosed	173	76%
	Open (including semi-enclosed)	55	24%
Primary kitchen location	Inside main living area	22	10%
	Outside under veranda/porch	5	2%
	Separate building	65	28%
	Separate room attached to house	136	60%
Smoke from neighbors	No	175	77%
	Yes	53	23%
Burn trash	No	9	4%
	Yes	219	96%
Where is trash burned	Outside far from house	136	78%
	Outside near house	36	20%
	Other	3	2%
Presence of electricity	Has electricity	64	28%
	No electricity	164	72%
Presence of kerosene	Use kerosene for lighting	25	11%
	Kerosene not used for lighting	203	89%
Stove type used d	Justa	76	39%
	Justa + traditional	115	59%
	Traditional	5	2%

^aAt baseline, study Visit 1, participants were all women and the primary household cook, between 23-59 years old, not pregnant, non-smoking nor exposed to secondhand smoke, and used only traditional cookstoves with biomass fuel to cook

^b9 items including radio, cell phone, bicycle, tv, sewing machine, refrigerator, car, motorbike, and computer

^c5 livestock types including pig, cow, duck, turkey, and horse

^dData from visit 5, after destruction of traditional stove and installation of Justa stove for all participants

3.2. Fine Particulate Matter Exploratory Data Analysis

Table 2 presents summary statistics for personal and kitchen measurements of PM_2.5. There was a difference in personal exposure and kitchen concentration levels by primary stove type used in the household. Households that reported using only the Justa stove had geometric mean (geometric standard deviation [GSD]) personal and kitchen measurements of 44.8 (2.4) μg/m³ and 62.4 (2.7) μg/m³, respectively. These values were similar to the geometric mean (GSD) exposures of households that used both Justa and traditional stoves (personal: 47.6 [2.5] μg/m³, kitchen: 57.3 [2.8] μg/m³) and lower when compared with households that only used the traditional stoves (personal: 87.3 [2.4] μg/m³, kitchen: 184 [3.5] μg/m³). These differences are also evident from the CDF plots in Figure 2.

Table 2.

Summary statistics for personal and kitchen time weighted, 24-hour PM_2.5 measurements (μ/m³) by reported stove used and overall. These statistics include multiple measurements from each household.

	n	Arithmetic mean (SD)	Geometric mean (GSD]	Minimum	Median (25th, 75th)	Maximum
Personal
Justa	186	72.7 (110)	44.8 (2.4)	5.5	42.4 (26.2, 70.7)	802
Justa+Trad	327	92.8 (271)	47.6 (2.5)	5.5	45.2 (28.4, 75.4)	3652
Traditional	549	143 (296)	87.3 (2.4)	3.9	80.8 (50.5, 138.7)	5509
Overall	1062	115 (266)	64.5 (2.6)	3.9	60.8 (35.9, 109.4)	5509
Kitchen
Justa	181	114 (214)	62.4 (2.7)	5.1	53.5 (33, 113)	1924
Justa+Trad	325	108 (228)	57.3 (2.8)	5.0	51.3 (28.1, 98.7)	3357
Traditional	555	425 (711)	184 (3.6)	4.3	175 (70.7, 440)	5426
Overall	1061	274 (559)	107 (3.6)	4.3	90.3 (41.9, 243)	5426

Notes. SD = Standard deviation; GSD = Geometric standard deviation

Figure 2.

Cumulative distribution function plot for 24-hour personal exposure and kitchen concentration of PM_2.5 (μg/m³) by stove type and presence of electricity in the house.

We observed a difference in both personal exposure and kitchen PM_2.5 concentration with electricity use (Figure 2 and Supplemental Material Tables B.1 and B.2). Personal geometric mean (GSD) concentrations varied from 51.6 (2.2) μg/m³ for households with electricity to 70.0 (2.7) μg/m³ among households without electricity, while kitchen geometric mean concentrations varied from 82.1 (3.4) μg/m³ to 118 (3.7) μg/m³, respectively. Personal geometric mean (GSD) concentrations were 81.6 (2.8) μg/m³, 56.2 (2.4) μg/m³, and 68.6 (2.6) μg/m³, for the stove use categories of 0-60%, 60-80%, and 80-100%, respectively, and kitchen concentrations showed a similar pattern. The distribution of personal exposures and kitchen concentrations was similar by SUM time above 28 degrees when either the traditional or Justa plus traditional stove were used. However, participants using only the Justa stoves tended to have higher kitchen concentrations with higher use (Supplemental Material Figures F.1 and F.2). Differences in geometric mean concentration by other variables were generally smaller than those observed by stove type, presence of electricity, and stove use (Supplemental Material Tables B.1 and B.2).

3.3. Black Carbon Exploratory Data Analysis

BC measurements were moderately correlated with PM_2.5 measurements: r=0.63 for personal measurements and r=0.80 for kitchen measurements. Correlations were higher for when using traditional stove only (r=0.69 and r=0.82 for personal and kitchen, respectively) than when using Justa stoves with or without a secondary stove (r=0.51 and r=0.70).

Table 3 includes summary statistics for personal exposure and kitchen concentrations of BC. The trends in other variables for BC were largely similar to those for PM_2.5 for personal and kitchen measurements, with the notable exception of the presence of kerosene. Houses that reported the presence of kerosene for lighting showed higher geometric mean (GSD) kitchen concentrations of BC (25.7 [7.7] μg/m³) compared to those without kerosene present for lighting (18.0 [7.4] μg/m³) (Supplemental Material Table C.2). This is also evidenced in the CDF plots (Figure 3). We saw similar patterns in stove type where both personal and kitchen BC concentrations were lower for participants using only Justa stoves or Justa plus secondary traditional stoves; also, for electricity where those that didn’t report having electricity had higher personal and kitchen measurements (Figure 3). The CDF stratified by measured stove use and stove type shows that more stove use was correlated with slightly higher values for both personal exposure and kitchen concentrations for households using Justa only (Supplemental Material Figures F.3 and F.4).

Table 3.

Summary statistics for kitchen and personal time weighted, 24-hour BC measurements (μ/m³) by reported stove use and overall. These statistics include multiple measurements from each household.

	n	Arithmetic mean (SD)	Geometric mean (GSD)	Minimum	Median (25th, 75th)	Maximum
Personal
Justa	186	33.1 (188)	4.2 (6.2)	0.1	3.7 (1.9, 11.9)	2364
Justa+Trad	327	19.3 (78.)	3.3 (6.4)	0.1	3.4 (1.5, 10.4)	1026
Traditional	549	33.4 (72.6)	11.6 (4.5)	0.1	11.8 (5.2, 29.1)	874
Overall	1062	29.0 (104)	6.6 (5.9)	0.1	7.3 (2.5, 19.9)	2364
Kitchen
Justa	181	54.3 (175)	9.4 (5.9)	0.1	8.6 (2.6, 26.4)	1414
Justa+Trad	325	53.2 (178)	7.2 (6.8)	0.1	5.9 (2.3, 25.6)	1506
Traditional	555	134 (205)	40.9 (6.1)	0.1	43.2 (12.1, 160)	1100
Overall	1061	95.0 (195.9)	18.5 (7.4)	0.1	17.4 (4.4, 84.9)	1506

Notes. SD = Standard deviation; GSD = Geometric standard deviation

Figure 3.

Cumulative distribution function plot for 24-hour personal exposure and kitchen (μg/m³) concentrations of BC by stove type, presence of electricity, and presence of kerosene for lighting.

3.4. PM_2.5 Personal Exposure Multivariable Modeling Results

We saw significant differences in personal 24-hour PM_2.5 exposure based on stove type, electricity, season, and percent of time cooking on the primary stove at temperatures above 28 °C (Table 4). Houses using only a traditional stove had an estimated 107% (95% CI: 79%, 139%) higher geometric mean personal PM_2.5 exposure compared to houses using only the Justa stoves, when adjusting for the other variables in the model. Houses without electricity had a 26.8% (95% CI: 3%, 44%) higher geometric mean personal PM_2.5 exposure compared to houses with electricity. Stove use monitoring, recorded as percentage of time above 28 °C, showed that those using their stove 60-80% of the time had reduced exposure by 14.6% (95% CI: (−24%, −4%) and 80-100% of the time had 18.9% (95% CI: (0%, 41%) higher exposure compared to those with less than 60% of use time. Lastly, geometric mean exposures were found to be 14.2% lower (95% CI: −23%, −6%) in the rainy season compared to the dry season.

Table 4:

Percent change by variable in personal and kitchen 24-hour geometric mean PM_2.5 (μ/m³) concentrations. Resulting values from two mixed models, one for personal exposure and one for kitchen.

	Personal,		Kitchen
Variable	Percent Change in Geometric Mean (95% CI)	p	Percent Change in Geometric Mean (95% CI)	p
Electricity		.02		0.03
Presence of electricity	(Reference)		(Reference)
No electricity	26.8% (3%, 44%)		37.1% (3%, 83%)
Kerosene (for lighting)		0.64		0.99
Presence of kerosene	(Reference)		(Reference)
No kerosene	−4.45% (−20%, 15%)		0% (−21%, 27%)
Asset Index		0.28		0.97
High	(Reference)		(Reference)
Low	10.0% (−8%, 31%)		−0.4 (−22%, 28%)
Stove Type		<.001		<.001
Justa	(Reference)		(Reference)
Justa + Traditional	13.3% (−3%, 33%)		−2.6% (−21%, 19%)
Traditional	107% (79%, 139%)		212% (160%, 275%)
Income Source(s)		0.90		0.27
Multiple sources	(Reference)		(Reference)
Income only through agriculture	−0.9% (−12%, 11%)		8.4% (−6%, 25%)
Kitchen Enclosure		0.88		0.04
Fully enclosed	(Reference)		(Reference)
Open	−1.2% (−17%, 17%)		−21.3% (−37%, 1%)
Kitchen Location		0.38		<.01
Inside main living area	(Reference)		(Reference)
Outside under veranda/porch	−4.6% (−45%, 66%)		183% (34%, 496%)
Separate building	−19.8% (−32%, 5%)		−16.6% (−38%, 20%)
Separate room attached to house	−18.6% (−27%, 6%)		−11.4% (−37%, 26%)
Cooking Time		0.41		0.38
0-14 hours/wk	(Reference)		(Reference)
15-28 hours/wk	15.1% (3%, 37%)		−17.2% (−33%, 3%)
29-42 hours/wk	16.4% (−4%, 42%)		−16.0% (−44%, 8%)
43-128 hours/wk	8.1% (−31%, 69%)		−7.9% (−49%, 68%)
Stove Use Monitoring- % of stove use time > 28 °C		<.001		<.01
0-60%	(Reference)		(Reference)
60-80%	−14.6% (−24%, −4%)		−3.4% (−17%, 13%)
80-100%	18.9% (0%, 41%)		30.8% (5%, 63%)
Season		<.01		0.03
Dry	(Reference)		(Reference)
Rainy	−14.2% (−23%, −6%)		−12.2 (−22%, 0%)
Education		0.44		0.24
No school	(Reference)		(Reference)
Compulsory schooling	−17.5 (−39%, 12%)		−27.7% (−53%, 11%)
Greater than 6 years	−15.4 (−53%, 36%)		−9.7(−54%, 76%)

Based on conditional R², the overall model, including the random effects, explained 36.7% of variability in personal exposure. Based on marginal R², predictor variables alone explained 16.7% of the variation in personal exposure (Table 5), with the participant random effects accounting for an additional 20 percentage points of variation. The multivariable model showed that stove type accounted for 11.9% of variability when adjusting for other variables, indicating that it has an effect not attributable to the other variables in the model. In the same model, electricity (2%), stove use (2.5%), and kitchen location (1.4%) explained about 5.5% of the exposure variation.

Table 5.

Percent of variance in 24-hour PM_2.5 personal exposure and kitchen concentrations explained by each variable in multivariable mixed models.

Variable	Personal Variance Explained (%)	Kitchen Variance Explained (%)
Stove type	11.9	20.4
Electricity	2.0	1.2
Measured stove use >28° C	2.5	1.2
Kitchen location	1.4	2.7
Season	0.6	0.2
Assets	1.0	0.3
Income	0.9	2.1
Education	0.4	0.5
Cooking time	0.1	1.0
Kitchen enclosure	0	0.2
Kerosene	0.1	0.1
Total Model
Predictors alone	16.7	24.1
Predictors and random effects	36.7	48.9

Note: Total model predictors alone represent the variance explained by fixed effects (marginal R²). Predictors and random effects are the variance explained by the entire model, including both fixed and random effects (conditional R2).

3.5. PM_2.5 Kitchen Concentration Multivariable Modeling Results

We identified statistically significant differences in 24-hour PM_2.5 kitchen concentrations for the following characteristics: presence of electricity, stove type, kitchen enclosure, primary stove location, season, and stove use (Table 4). Houses using only traditional stoves had a 212% (95% CI: 160%, 257%) higher geometric mean kitchen PM_2.5 concentration compared to houses using only improved Justa stoves. Houses without electricity had a 37.1% (95% CI: 3%, 83%) higher geometric mean kitchen PM_2.5 than those with electricity. Geometric mean concentrations were found to be lower by 12.2 (95% CI: −22%, 0%) in the rainy season compared to the dry season. Households with 60-80% of measured stove use time had similar kitchen concentrations to those with 0-60% of stove use (−3.4%, 95% CI: −17%, 13%). For households with 80-100% of stove use time, kitchen concentrations were higher by 30.8% (95% CI: 5%, 63%). Additionally, kitchens considered to be in open (or semi-enclosed) areas had 21.3% (95% CI: −37%, −1%) lower geometric mean kitchen PM_2.5 concentrations than those in enclosed spaces. Kitchens outside under a veranda or on a porch had 183% (95% CI: 34%, 496%) higher geometric mean kitchen PM_2.5 compared to houses with kitchens inside the main living area (Table 4).

Based on conditional R², the overall model, including random effects, explained 48.9% of variability in kitchen concentrations. Based on the marginal R², predictor variables alone explained 24.1% of the variation in kitchen concentrations (Table 5). We observed that stove type accounted for 20.4% of variability when adjusting for other variables, indicating that it had an effect not attributable to the other variables in the model. In the same multivariable model, kitchen location accounted for 2.7% of variability, income 2.1%, and electricity and stove use 1.2% each. Variability explained by the other variables was small (Table 5).

3.6. Black Carbon Personal Exposure and Kitchen Concentration Modeling Results

BC results were similar to those for PM_2.5. Briefly, when comparing percent change in geometric means for both personal and kitchen measurements we saw significant differences by stove type, kerosene, and season. For personal exposures, we additionally found differences with measured stove use and for kitchen concentrations with kitchen enclosure type and kitchen location (Table 6). Based on the marginal R², our predictors of personal BC exposure explained 18.9% of personal variance in the multivariable model with predictors alone, and based on the conditional R² 36.8% of variance when including random effects (Table 7). Our predictors for kitchen BC exposure explained 22.2% of variance in the multivariable model with predictors alone, and 47.0 % of variance when including random effects (Table 7).

Table 6:

Percent change by variable in personal and kitchen 24-hour average BC (μ/m³) concentrations. Resulting values from two multivariable models, including random effects, one for personal exposure and one for kitchen concentration.

	Personal		Kitchen
Variable	Percent Change in Geometric Mean (95% CI)	p	Percent Change in Geometric Mean (95% CI)	p
Electricity		0.08		0.07
Presence of electricity	(Reference)		(Reference)
No electricity	91.4% (31%, 180%)		51.0% (−54%, 138%)
Kerosene (for lighting)		0.04		0.05
Presence of kerosene	(Reference)		(Reference)
No kerosene	−30.6% (−51%, 2%)		−30.7% (−52%, 1%)
Asset Index		0.75		0.74
High	(Reference)		(Reference)
Low	5.5% (−24%, 46%)		6.9% (−28%, 59%)
Stove Type		<.001		<.001
Justa	(Reference)		(Reference)
Justa + Traditional	−21.5% (−42%, 6%)		−23.3% (−45%, 6%)
Traditional	193.3% (123%, 286%)		351.5% (337%, 504%)
Income Source(s)		0.32		0.15
Multiple sources	(Reference)		(Reference)
Income only through agriculture	11.2% (−10%, 38%)		18.2% (−6%, 49%)
Kitchen Enclosure		0.85		0.03
Fully enclosed	(Reference)		(Reference)
Open	2.9% (−25%, 42%)		−32.7% (−53%, 2%)
Kitchen Location		0.80		<.01
Inside main living area	(Reference)		(Reference)
Outside under veranda/porch	11.0% (−61%, 317%)		669.7% 143%, 2337%)
Separate building	−15.6% (−49%, 40%)		−7.4% (48%, 65%)
Separate room attached to house	−18.0% (−50%, 33%)		3.6% (−41%, 80%)
Cooking Time		0.80		0.47
0-14 hours/wk	(Reference)		(Reference)
14-28 hours/wk	2.9% (−26%, 43%)		−20.9% (−44%, 11%)
28-42 hours/wk	9.3% (−25%, 58%)		−10.3% (−39%, 33%)
42-128 hours/wk	−25.4% (−68%, 75%)		−27.2% (−73%, 88%)
Stove Use Monitoring-Time ± 28°C		<.01		0.21
0-60%	(Reference)		(Reference)
60-80%	−19.5% (−46%, 1%)		17.7% (−8%, 51%)
80-100%	20.4% (−13%, 67%)		35.8% (−4%, 93%)
Season		<.001		0.03
Dry	(Reference)		(Reference)
Rainy	61.5% (35%, 93%)		23.5% (3%, 49%)
Education		0.62		0.18
No school	(Reference)		(Reference)
Compulsory schooling	−17.1 (−52%, 45%)		−44.9 (−72%, 8%)
Greater than 6 years	7.6 (−55%, 158%)		−26.6 (−73%, 110%)

Table 7.

Percent of variance in 24-hour BC personal exposures and kitchen concentrations explained by each variable in multivariable mixed models.

Variable	Personal Variance Explained (%)	Kitchen Variance Explained (%)
Stove type	9.7	14.4
Electricity	1.8	0.6
Income	0.1	0.1
Measured stove use >28° C	0.7	0.1
Season	1.8	0.3
Assets	0	0
Kitchen location	0	1.3
Kerosene	0.3	0.2
Education	0	0.4
Cooking time	0	0
Kitchen enclosure	0	0.5
Total Model
Predictors alone	18.9	22.2
Predictors and random effects	36.8	47.0

Note. Total model predictors alone represent the variance explained by fixed effects (marginal R²). Predictors and random effects are the variance explained by the entire model, including both fixed and random effects (conditional R²).

4. Discussion

Overall, our hypotheses that household stove type and season would be predictive of differences in both kitchen and personal 24-hour PM_2.5 and BC concentrations were supported. Although we selected out predictors by comparing difference in PM_2.5 values, we found that they explained a similar amount of variation in BC concentrations. Due to this, we used the same variables to explore variance in both personal and kitchen PM_2.5 and BC measurements. Electricity in the house and use of the Justa stove were associated with lower PM_2.5 and BC concentration and kerosene was associated with an increased BC kitchen concentration, as hypothesized. Our findings are consistent with the Pollard et al. (2018) study that found the presence of electricity affects PM_2.5 exposure. We found that stove use as reported by participants vs stove use monitor results showed varying levels of association with personal exposures and kitchen concentrations. These results confirm the importance of using stove use monitors to collect more accurate data, and highlights the need for improvements to understanding and eliciting stove use behavior information specifically with the intention of behavior change toward long-term clean household energy adoption. Higher self-reported typical cooking times resulted in higher personal exposures and lower kitchen exposures. While households with the longest time spent cooking as shown by the stove use monitor (>80% of time over 28°C) resulted in higher concentrations. However, households with 60-80% of measured stove use time showed similar or lower concentrations compared to household with lower measured use.

We found it interesting that the presence of electricity explained differences in geometric mean PM_2.5 exposures, but the presence of kerosene for lighting did not when comparing across both personal and kitchen results. This may be due to who is primarily responsible for varying household activities, such as the use of kerosene by the non-primary cooks or the types of activities related to electricity. Additional reasons may be related to low numbers of households regularly using kerosene in the study, the particular use of kerosene such as to start fires, or for kerosene use for smaller tasks that may not have had a large impact on exposure. Household activity logs may be useful to identify who is carrying out what activities, for how long, and when during the day. We observed that presence of electricity and kerosene were negatively correlated (Supplemental Material Table D.1). We also anticipated that electricity may be related to assets, and that those with higher assets would show similar results and association with lower exposure. However, the asset index was not correlated with electricity and did not result in significant differences in geometric means nor contributed to variance explained for personal exposure or kitchen concentrations for either PM_2.5 or BC. Although exploratory data analysis showed that participants with higher weighted assets had lower exposure, model results indicated that this accounted for an insignificant amount of variance in exposure once other variables were accounted for. For BC, results showed that both the presence of electricity and the presence of kerosene for lighting, affected kitchen concentrations. This finding provides further evidence that components of PM formed through combustion, such as BC, should be studied in combination with PM_2.5 to understand the relationship more accurately between kerosene, and other variables, and exposure to air pollution.

The stove enclosure type and location (i.e., kitchen characteristics) also impacted geometric mean kitchen concentrations. This corroborates other findings that building structure and ventilation impacts exposure (Clark et al., 2010; Fandiño-Del-Rio et al., 2020; Sharma & Jain, 2020). We saw lower kitchen BC and PM_2.5 concentrations in open kitchen types when compared to enclosed kitchens. Additionally, despite having their primary traditional stoves destroyed during Justa installation, we have self-reports of secondary, less permanent traditional stove use in conjunction with the Justa stove. However, secondary traditional stoves were most frequently located outdoors, and our results indicate that personal exposure and kitchen concentration varied minimally between solely Justa use vs Justa plus traditional stove use (Figure 2).

Higher PM_2.5 and BC concentrations by measured stove use was observed with the highest amount of stove use (>80%), but results were less consistent for moderate amounts of stove use (60-80%). Reasons for this may be reflected in the frequency and use of stoves. For example, if participants used the stove more frequently for shorter time periods, these results may have been influenced by frequency of starting the stove, the largest exposure occurring at “startup” when the stoves were ignited (Fedak et al., 2019). Lastly, the chosen temperature threshold may not fully capture all cooking events across all households.

Our analysis also estimated the marginal percent variation explained by each variable in absence of other variables, including a random intercept. In linear mixed models, unlike linear (fixed effect) regression, estimates of variance explained cannot be easily partitioned among a set of correlated predictor variables. Therefore, we use the multivariable models with and without the variable to assess the variability explained by each variable in combination with the other variables. These results showed that the selected variables (and random effect) account for a large, but not majority, portion of variation in PM_2.5 personal exposure (37%) and kitchen concentration (49%), and BC personal exposure (37%) and kitchen concentration (47%). This is similar to the results of Lee et al. (2021), who found variables and random effects accounted for 29% of personal PM_2.5 variation and 24% of personal BC variation. This is less than the variability explained in a previous cross-sectional study in rural Honduras, but that study used a subjective measurement of stove quality (Clark et al., 2010). Our findings also corroborate a recent study by Wang et al. (2021) that found characteristics of participants and their households explained household (46-60%) and personal (33-54%) BC variation. Nishihama et al. (2021) presented a different approach that focused on out-of-sample predictive accuracy alone, and not on understanding relationships and variance decomposition (like ours does). Although a different measure, their cross-validated R² (0.42) is of similar magnitude to the in-sample amount of variance explained in our context of 37% of personal exposure.

Across studies, it is important to note that the magnitude of variance explained is also impacted by the amount of variability present in the sample of measurements. Studies that include a variety of cooking technologies may be able to explain a larger proportion of variability due to differences in stove type. We suggest that further research is needed to evaluate additional household and personal characteristics (of participants) around daily activities both indoors and outdoors that may vary by social, spatial, and temporal contexts. Contextual information from different geographies, i.e., communities, regions, and countries may provide valuable evidence on the range of variables that affect household air pollution, including personal exposures.

Because the individual random-effect term accounts for between-person variation, the unexplained variability in our results is primarily due to within-person differences between the repeated measures (Supplemental Material Section G). This may be due to major time-varying variables that were not measured, or simply day-to-day variability in cooking that is reflective of standard practice by study participants. Outdoor air quality was not measured in our study and may be a contributing variable to both personal exposure and kitchen concentrations of BC (Wang et al., 2021) and PM_2.5 (Lee et al., 2021). Other studies have also reported a large amount of within-household variation in PM_2.5 measurements from biomass stoves (Keller et al., 2020; Lee et al., 2021). The large amount of within-household variation that cannot be explained by predictor variables supports the use of mixed models (as opposed to household averages) for assessing long-term average exposures (Keller & Clark, 2022). The information about variables impacting concentrations gained from the multivariable models we have presented could be used to inform models for predicting long-term averages.

Previous work indicates that personal PM_2.5 exposure decreased over time (Benka-Coker et al., 2021). The stepped-wedge design meant households were allocated the Justa stove after visits 2 and 4, and exposure decreased more significantly in subsequent visits. Further, concentrations were more likely to be close to the personal PM_2.5 24-hour IT-3 WHO guideline of 37.5 μg/m³ (WHO, 2021) after the Justa stove was introduced. There was a larger range of values for 24-hour PM_2.5 kitchen concentrations. Geometric mean kitchen concentrations were higher than personal exposures and showed a slight decrease over time. Kitchen concentrations rarely dipped below the WHO interim or AQG targets (WHO, 2021).

Collecting accurate, repeated measurements toward understanding long-term household air pollution data in low- and middle-income countries is difficult due to the often-remote settings, varied instrumentation and data collection timeframes, and the extensive/expensive resources required (Johnson et al., 2021; Nishihama et al., 2021). However, those most affected by the negative health outcomes of household air pollution remain in rural, remote areas in lower income countries. Therefore, efforts to understand and effectively implement intervention programs that influence behavior change around energy adoption and clean cooking toward decreases in household air pollution are critical, but not without challenges. A recent study suggested that predicting exposures from house, household, or personal characteristics may help to improve household air pollution interventions and circumvent resource intensive and challenging exposure data collection (Johnson et al., 2020). This goal will require more work to identify the characteristics responsible for variation in exposure across time, space, and demographics. The advancement and increased availability of exposure monitoring devices that are cheaper and less invasive continue to greatly advance the study of household air pollution.

5. Conclusions

Our study adds to the limited body of literature on repeated measures of PM_2.5 and BC exposures to understand longer term effects of household air pollution. We present additional evidence that stove type, season, and kitchen and household characteristics affect personal exposure and kitchen concentration in homes using biomass fuel sources. However, the study indicates substantial unexplained variance, especially in relation to the role of within-person variability over time. In addition to exposure measurements, future interventions need to carefully explore how the daily indoor and outdoor activities and characteristics of participants may increase or mitigate their exposure to poor air quality. If variables can be accurately predicted and tested through empirical data collection, then home characteristics can be adapted accordingly as can program/study interventions toward adoption of cleaner cooking practices and improved human health outcomes.

Our study results also provide insights for future cookstove and household clean energy intervention efficacy and policy. For example, cookstove interventions may lead to higher overall public health improvements by focusing on whole house energy transitions such as clean cookstoves and kerosene for lighting via solar electricity or promoting enclosed kitchen construction. Further, results can lead to improved targeting of policy subsidies such as access for all communities to electricity and affordable clean cooking options especially during certain seasons that communities have increased exposure, such as the dry season in Honduras. Another example is subsidization to minimize the barriers for private-business to reach last-mile communities or for small and medium-size enterprises to expand and diversify. Our study provides important implications for health research and practice through understanding the roles of personal and household characteristics on exposure to poor indoor air quality in rural Honduras.

Abbreviations

BC

Black carbon

CI

Confidence interval

CDF

Cumulative distribution function

CSU

Colorado State University

GSD

Geometric standard deviation

PM_2.5

Fine particles <2.5 μm in aerodynamic diameter

SUM

Stove use monitor (electronic temperature logger)

UPAS

Ultrasonic Personal Air Sampler

WHO

World Health Organization