Impacts of Improved Cookstove Interventions on Personal Exposure to Carbon Monoxide and Particulate Matter in Zambia

Abstract

Eighty‐four percent of sub‐Saharan African households rely on polluting fuels (e.g., wood, charcoal) for cooking, leading to high levels of household air pollution (HAP). While switching to modern fuels/stoves could decrease HAP levels, they are not always available or affordable. Improved biomass cookstoves could provide an intermediate step supporting transitions from traditional biomass to clean burning fuels/stoves. We conducted two stove intervention trials in Lusaka, Zambia using targeted marketing/incentives to motivate participants to use improved biomass stoves, either the Mimi Moto (pellet) or the EcoZoom (charcoal). Before the intervention, 65% of participants exclusively used charcoal, while 27% relied on electricity to some extent for cooking. We measured 24‐hr personal exposure to CO (n = 747) and PM_2.5 (n = 90) of primary cooks. We implemented several statistical approaches to estimate the effects of interventions on exposure: household‐specific endline minus baseline exposure, ranksum testing, difference‐in‐differences analyses, and cross‐sectional analyses. We found that switching from traditional charcoal stoves to either intervention stove was not associated with significantly reduced exposures. However, cooks using electric stoves independent of the intervention did have significantly lower CO exposures than those using traditional charcoal, with greater electric stove use corresponding to greater exposure reductions. Variability in exposure was dominated by seasonal, regional, and neighborhood differences rather than household stove/fuel choices. A focus on HAP exposure from cooking in urban settings is unlikely to yield expected exposure reductions. Policy makers should consider pollution reduction policies/interventions that target ambient air quality in tandem with HAP‐mitigating strategies to address air pollution health burden.

Plain Language Summary

In 2019, almost 7 million deaths worldwide were linked to air pollution. In sub‐Saharan Africa, 84% of households use polluting fuels like wood and charcoal for cooking, which leads to high levels of air pollution. While cleaner energy options such as electricity could reduce pollution, they are often not accessible or affordable. Improved biomass cookstoves might offer an intermediate solution, helping people transition from traditional fuels to cleaner options. We conducted two studies in Lusaka, Zambia to encourage traditional stove users to switch to improved biomass stoves, either the Mimi Moto (using wood pellets) or EcoZoom (using charcoal). We measured exposure to carbon monoxide and particulate matter among household cooks. We found that switching to improved stoves did not significantly reduce exposures compared to traditional charcoal use. However, cooks using electric stoves did have lower exposures, with greater exposure reductions as more meals were cooked with electricity. Factors such as season and neighborhood also had substantial impacts on exposure. This suggests that focusing only on reducing pollution from cooking in urban areas may not be enough to improve health. Policymakers should consider broader strategies that also target outdoor air quality to effectively address the health risks of air pollution.

Key Points

• In two interventions, the use of improved biomass stoves was not associated with reduced personal exposures

• Household cooks using existing electric stoves did have significantly lower exposures than traditional charcoal users

• Changes in exposure were dominated by seasonal, regional, and neighborhood differences rather than household stove/fuel choices

1. Introduction

In 2019, almost 7 million deaths globally were attributed to exposure to ambient particulate matter and household air pollution (HAP) primarily from cooking and heating (Fisher et al., 2021). In Africa, air pollution contributed to 1.1 million premature deaths, 63% of which were attributed to HAP (Fisher et al., 2021; Gouda et al., 2019; Murray et al., 2018). Eighty‐four percent of sub‐Saharan African households rely on polluting fuels (e.g., wood, charcoal, crop residues, dung) as their primary cooking fuel (Stoner et al., 2021) and inefficient stoves, which combined typically have poor combustion efficiencies leading to high levels of HAP exposure. HAP combustion emits harmful pollutants, such as carbon monoxide (CO) and fine (PM_2.5) and coarse (PM₁₀) particulate matter. Chronic exposure to these pollutants leads to respiratory illnesses, impaired neurocognitive ability, ischemic heart disease, stroke, obstructive pulmonary disease, lung cancer, and premature death (Kampa & Castanas, 2008; Lai et al., 2024; McCord et al., 2024; Raub & Benignus, 2002; Rosenthal, 2015; US Department of Health and Human Services, 2012; WHO, 2014; Wilbur et al., 2012). Addressing HAP is expected to decrease the health burden associated with HAP exposure. Additionally, mitigating HAP is likely to improve other well‐being metrics targeted by the Sustainable Development Goals including (3) good health and well‐being, (5) gender equality, (7) affordable and clean energy, (13) climate action, and (15) life on land (e.g., forest health) (Rosenthal et al., 2018; UN, 2015).

While switching to clean cooking fuels and stoves, such as electricity or liquefied petroleum gas (LPG), is expected to decrease HAP levels, these fuels are not always an option for resource‐limited households due to affordability, accessibility, and reliability (Puzzolo et al., 2016). Improved biomass cookstoves could provide an additional rung in the “energy ladder” (van der Kroon et al., 2013), an intermediate step supporting transitions from traditional biomass to clean burning fuels and stoves (Phillip et al., 2023). Improved biomass stoves burn solid fuel but include features that decrease emissions and/or exposures, such as increasing thermal efficiency by maximizing heat to cooking pot via insulation (e.g., clay liners) or using a chimney for ventilation. In field and laboratory tests, improved charcoal stoves have reported higher thermal and modified combustion efficiencies and lower emissions compared to traditional, uninsulated charcoal stoves (Coffey et al., 2017; Eilenberg et al., 2018; Jetter et al., 2012). Forced‐draft, gasifier stoves are also an attractive alternative as they have reported lower emissions measurements (Bilsback et al., 2018; Champion et al., 2021; Champion & Grieshop, 2019; Garland et al., 2017; Jetter et al., 2012; Parsons et al., 2022; Sandro et al., 2019) and offer measurable exposure reductions (Balakrishnan et al., 2015; Baumgartner et al., 2019; Jagger et al., 2017; Pope et al., 2021; Sambandam et al., 2015).

Portfolios of energy sources, fuels, and cooking technologies are more diverse in urban centers than rural areas of low‐income countries as wider fuel availability, electricity access, market access, and economic development provide households with numerous household cooking options (Bailis et al., 2020; Stoner et al., 2021; Trotter, 2016; Wiedinmyer et al., 2017). In 2020, primary fuel use in urban sub‐Saharan Africa was diverse, consisting of 30% charcoal, 28% biomass, 20% LPG, and 13% electricity (Stoner et al., 2021). Stove stacking, or the use of multiple fuels/stoves within a single household, often a combination of clean and polluting fuels/stoves, to meet cooking needs, is common (Ochieng et al., 2020; Shankar et al., 2020). However, charcoal is expected to remain the dominant primary fuel in urban sub‐Saharan Africa until at least 2030 (Stoner et al., 2021) due to slow market development, underdeveloped distribution channels, and the challenge of motivating households to change cooking behaviors.

Personal exposure measurements are often used as a proxy for health impacts of air pollution. HAP from cooking is thought to be a key contributor to personal exposure. Studies of personal exposure exploring the impact of fuel use changes have found that in urban settings, changes in exposure are hard to quantify due to higher background influences from other sources such as diesel vehicles and burning solid waste (McFarlane, Isevulambire, et al., 2021, McFarlane, Raheja, et al., 2021; Naidja et al., 2018; Petkova et al., 2013; Shikwambana & Tsoeleng, 2020). Fuel transitions and personal exposure in urban settings are understudied, at least in part due to these complications (Johnson et al., 2022; Ochieng et al., 2020; Weston et al., 2016). However, Africa's urban population is projected to triple by 2050, with the population becoming majority urban by 2033 (Murray et al., 2018; UN, 2019). Understanding how fuel and stove choices influence health in urban settings is imperative to mitigating overall air pollution exposures (Güneralp et al., 2017). Evaluation of cookstove interventions can provide direct insight into how exposure is impacted by transitioning from traditional to more efficient stoves by directly comparing exposures of the same or similar individuals using different fuels/stoves over time.

Zambia is an ideal study location to explore the transition from traditional to improved biomass stoves. Zambia has a large urban population (40%) providing insights on cooking fuel transitions for growing urban populations throughout sub‐Saharan Africa (Zambia Statistics Agency, 2022). Zambia has seen an upward trend in charcoal as the primary cooking fuel used over the past two decades, with ∼70% of the urban population using charcoal in 2018 (Jürisoo et al., 2019; Mulenga et al., 2019; Stoner et al., 2021; Zambia Statistics Agency et al., 2019). Further, Zambia has abundant forest resources, covering 54% of total land area, making it challenging to displace charcoal production, distribution, and sales (Phiri et al., 2019; Stoner et al., 2021). Lusaka, the capital city and a major commercial center in Zambia, is a useful case study of how complex urban air pollution sources influence exposure. Although urban Zambia has a 71% electrification rate, only 18% of the urban population primarily cooks with electricity (Zambia Statistics Agency et al., 2019). While this is a small fraction of the electrified population, it is almost 1.5 times the average percentage (13%) of urban households in sub‐Saharan Africa that use electricity primarily for cooking (Stoner et al., 2021). As of 2023, current policies in Zambia are pushing for greater LPG uptake due to electricity supply and reliability issues (USAID, 2023a). At present, only 0.3% of urban households rely on LPG primarily for cooking (Zambia Statistics Agency et al., 2019).

We address several gaps in the literature. First, we know of few studies that have attempted to quantify the impact of switching from traditional charcoal to an improved stove on exposure. Instead, most studies have focused on wood users in rural settings (Jack et al., 2021; Liao et al., 2021; Ochieng et al., 2017; Phillip et al., 2023; Piedrahita et al., 2019; Pilishvili et al., 2016; Pope et al., 2021; Quinn et al., 2017; Saleh et al., 2022; Shupler et al., 2020). Exposure measurements from studies of cooks primarily using charcoal varied. Wylie et al. (2017) measured CO and PM_2.5 exposures of cooks primarily using charcoal in urban Tanzania and found low levels of CO exposure (2 ppm) and moderate levels of PM_2.5 exposure (40 μg m⁻³). Wiedinmyer et al. (2017) also reported low CO exposures, averaging 1.3 ppm, in the urban, charcoal‐using cohort of their study in Ghana. Ellegård and Egnéus (1993) measured personal exposure of primary cooks in Lusaka, Zambia and found charcoal users had average CO and PM₁₀ exposures of 13 ppm and 380 μg m⁻³, respectively, during mealtimes. Delapena et al. (2018) found in urban Accra, Ghana, charcoal exclusive users had an average (standard deviation) PM_2.5 exposure of 30 (±24) μg m⁻³. However, no other studies to our knowledge have measured air pollution exposures of charcoal‐using cooks in urban sub‐Saharan Africa. Third, relatively few studies have directly quantified personal exposure (Chillrud et al., 2021; Delapena et al., 2018; Dutta et al., 2021; Shupler et al., 2020), but instead rely on non‐specific indicators of health outcomes related to interventions, such as blood pressure (Alexander et al., 2017; Jagger et al., 2019), hospital records (Coker et al., 2020), or indoor concentrations (Pope et al., 2021; Ubuoh & Nwajiobi, 2018). In Zambia, few studies have measured personal exposure related to cooking fuel/stove use, leading to reliance on proxies such as reported primary fuel (Bickton et al., 2020; Kadir et al., 2010; Patel et al., 2015), indoor concentrations (Mulenga et al., 2018), or models (Li et al., 2023) to estimate personal exposure to pollutants and impacts on health.

Our study quantifies exposures in Lusaka using data from a two‐phase cookstove intervention study focused on two private sector‐led efforts to encourage switching from traditional charcoal stoves to one of two improved biomass stoves: a forced‐draft micro‐gasifier stove (Mimi Moto) and an improved charcoal stove (EcoZoom). Our specific objectives are to (a) quantify personal CO and PM_2.5 exposures of urban cooks, (b) assess the impacts of the cookstove interventions on exposure, (c) examine seasonal differences and the influence of ambient air quality on exposure, and (d) determine significant predictors for personal exposure in an urban setting.

2. Methods

2.1. Study Design

Our study was a quasi‐experimental program evaluation conducted in Lusaka, Zambia in collaboration with two social enterprises, SupaMoto and VITALITE. Each enterprise promoted a different alternative biomass stove, with SupaMoto promoting the Mimi Moto and VITALITE promoting the EcoZoom. Baseline stoves, locally referred to as mbaulas, are traditional charcoal stoves made entirely of metal with no insulation in the combustion chamber.

The Mimi Moto is a biomass pellet‐fed, forced‐draft semi‐gasifier stove designed in the Netherlands. The Mimi Moto employs a two‐stage combustion process enabling better mixing of the combustible gases and air, leading to a uniform flame front and more complete combustion (Anderson et al., 2007). The Mimi Moto is designed to operate with densified biomass pellets, typically made of agricultural or forestry waste, such as sawdust. SupaMoto produces pellets at a factory in Ndola, Zambia. The Mimi Moto stove has been promoted in 25 countries throughout Asia and Africa (Mimi Moto, 2021) and is one of three biomass cookstoves to achieve best performance for emissions: Tier 5 (best) for CO emissions and Tier 4 (second‐best) for PM_2.5 emissions based on the 2018 international standard performance tiers (Champion et al., 2021; ISO, 2018; Parsons et al., 2022). Field emissions measurements of the Mimi Moto were similar to those of gas stoves (Champion & Grieshop, 2019).

The EcoZoom Mbaula Fresh is a charcoal stove that includes a metal lining and ceramic insert to minimize heat loss, thus decreasing the amount of charcoal needed. It has a manual vent that can be opened or closed to control air flow and temperature when cooking. It has been promoted in seven African countries as an alternative to traditional charcoal stoves. VITALITE estimates a 60%–70% reduction in charcoal consumption (VITALITE, 2022). While the EcoZoom has not been officially rated based on international standards, lab results from the University of Nairobi reported 45% thermal efficiency (Kithinji, 2015), compared to 15% for traditional charcoal stoves similar to those used in this region (Jetter et al., 2012).

The interventions aimed to motivate charcoal stove users to switch to cleaner biomass stoves, and in the case of SupaMoto, to transition from charcoal to pellets as their primary cooking fuel. We worked with each social enterprise to select two neighborhoods in Lusaka, one in which they had previously marketed their stove (intervention cookstove, ICS, users) and one in which they planned to market their stoves after our baseline data collection in 2019 (prospective users). SupaMoto marketed the Mimi Moto/biomass pellets in Matero and planned to market it in Kalingalinga after our baseline data collection; VITALITE marketed the EcoZoom in Kamanga with plans to market in Ng'ombe. After baseline data collection, the social enterprises provided incentives for households in prospective user neighborhoods to purchase the intervention stove. SupaMoto offered Kalingalinga households free solar lamps with consistent payments on their Mimi Moto stoves, and VITALITE offered Ng'ombe households a discount voucher for the EcoZoom stove. Both social enterprises directly marketed to customers through community and household level contacts. After incentives were offered, we returned to collect endline data to assess how effective the incentives were at encouraging adoption and stove use, and in turn, how stove use influenced exposure to CO and PM_2.5.

We implemented two sampling strategies to select households to participate in the study. Households in “ICS user” neighborhoods (Matero, Kamanga) were randomly selected to participate from a list of stove owners provided by each social enterprise. Households in “prospective user” neighborhoods (Kalingalinga, Ng'ombe) were selected using geographic sampling by mapping all residential dwellings in each neighborhood, assigning numbers to each household, and randomly selecting household numbers. For endline data collection, we returned to households where we collected baseline data. We added additional households at endline in Matero and Kalingalinga due to lost contact with baseline households and to bolster an initially small sample of Mimi Moto users from baseline household selection. New households were selected from SupaMoto's customer list. Figure S1 in Supporting Information S1 shows a map of study neighborhoods (Google Earth, 2023).

Baseline data collection was completed in July–August of 2019 in 1,377 households. Marketing took place between September 2019 and March 2020. Endline data collection was planned for 12 months after baseline but was delayed due to the COVID‐19 pandemic. We collected endline data in October–November of 2021 in 1,043 households. The delay resulted in baseline and endline data being collected in different seasons. Baseline data were collected during the cool, dry season and endline data during the start of the warm, wet season. Figure 1a shows a timeline for the data collection.

Figure 1.

(a) Timeline of data collection. Sampling structure and size for each stove company and their respective neighborhoods during (b) baseline and (c) endline data collection. At baseline, SupaMoto had existing intervention cookstove (ICS) users using the Mimi Moto in Matero with plans to market to prospective users in Kalingalinga. VITALITE had ICS users using the EcoZoom in Kamanga with plans to market to prospective users in Ng'ombe. Sample sizes are defined as follows: total N is the total number of household surveys collected in each neighborhood, CO N is the number of CO exposure measurements collected, and PM_2.5 N is the number of PM_2.5 exposure measurements collected. At endline, Matero and Kalingalinga sample sizes are separated by panel households (those sampled at baseline) and new households that were added during endline.

Data collection consisted of a structured household questionnaire that included questions about household demographics, cooking practices, household facilities, and economic decision making. The main decision maker (household head) and person most knowledgeable about cooking practices in the household (primary cook) were interviewed. A subset of households (44%) was convenience sampled for CO exposure monitoring. Households were selected based on the primary cooks' consent to wear the CO monitors for 24 hr. Among primary cooks that agreed to CO monitoring, a subset was selected for monitoring of PM_2.5 exposure, selected opportunistically based on availability of PM_2.5 monitors. Comparisons of the overall sample and the exposure monitoring sample are discussed in Section 3.1. Endline exposure data collection focused on households that had baseline CO (and PM_2.5) monitoring. Figures 1b and 1c show sample sizes for total, CO, and PM_2.5 household samples in each neighborhood for both phases of data collection. Exposure monitoring households were asked additional questions about cooking practices and activities of the primary cook during the 24 hr they were wearing the monitors.

2.2. Air Pollution Measurements

Primary cooks were outfitted with exposure monitors and instructed to wear them for 24 hr, unless sleeping or bathing, when they were instructed to keep the monitors nearby. If participants were only undergoing CO exposure monitoring, they were given a lanyard to wear the CO monitors around their necks or to clip to their outermost clothing layer to keep the monitor in their breathing zone. If monitored for both CO and PM_2.5, monitors were put in a bag strapped across cooks' chests to secure monitors in their breathing zone comfortably. Details of all air pollution instruments are in Table S1 in Supporting Information S1. To understand the implications of exposure concentration levels, we compared with the World Health Organization (WHO) global air quality guidelines, developed to offer guidance on limiting exposure to pollutants with large health risks (WHO, 2021). The Interim‐I (IT‐I) concentration targets for CO (24‐hr) and PM_2.5 (annual) are 6 ppm and 35 μg m⁻³, respectively.

We deployed Lascar EasyLog CO Data Loggers to measure CO exposure. Lascar loggers were calibrated before and after deployment for both baseline and endline field campaigns. The average of before‐ and after‐deployment calibration factors for each logger was used to correct the respective raw CO measurements. Additional details of CO calibrations are in Text S1 and Figure S2 in Supporting Information S1.

At baseline, we deployed RTI MicroPEMs to collect personal PM_2.5 exposure. MicroPEMs include a laser light scattering nephelometric sensor to measure real‐time PM_2.5 mass and a 25 mm Teflon filter for gravimetric correction of real‐time data (Chartier et al., 2017). MicroPEMs also include a temperature sensor. MicroPEMs were programmed to automatically turn off 24 hr after deployment. Any deployments that ran less than 20 hr (n = 8), had filters that were dropped or torn (n = 4), had real‐time data that did not have a corresponding filter correction (n = 3), or had real‐time data that included negatives (n = 6) were excluded from analysis.

At endline, due to COVID‐19‐imposed logistical constraints on preparation and training time, we deployed Atmotube Pro PM loggers instead of MicroPEMs to monitor PM_2.5 exposure. Atmotubes are operationally simpler than MicroPEMs because they only include a light scattering sensor (Sensirion SPS30) with a maximum PM_2.5 reading of 1,000 μg m⁻³ and include no filter measurement. Atmotubes also include a temperature sensor. While data from the Sensirion SPS30 compared favorably with that from reference grade monitors, with mean absolute errors (MAE) below 5.6 μg m⁻³ (Demanega et al., 2021; Motlagh et al., 2021; Roberts et al., 2022; SCAQMD, 2020), light scattering sensors typically require correction to provide consistent data quality. The most widely accepted correction practice is to collocate low‐cost sensors and regulatory monitors to create a sensor and setting‐specific correction, but there are limited regulatory monitors in southern Africa (Barkjohn et al., 2021; Jiao et al., 2016; Johnson et al., 2018; Zusman et al., 2020). To approximate this in Lusaka, we conducted a stationary collocation of Atmotubes and MicroPEMs to develop a deployment‐specific correction factor for each Atmotube. More details about the collocations are in Text S2 and Figures S3–S4 in Supporting Information S1.

Before endline data collection began, we installed two PurpleAir PA‐II monitors in Lusaka, one close to city center and one more central to the study neighborhoods, to measure ambient PM_2.5 during the campaign. PurpleAir PA‐II monitors use two nephelometric sensors (Plantower PMS‐5003) to measure real‐time PM_2.5 and transmit data to an online server via Wi‐Fi. While we did not install ambient monitors during baseline data collection, data from two PurpleAir PA‐II monitors installed in Kabwe, Zambia (no longer available as of December 2020), approximately 140 km north of Lusaka, were used as an indication of regional PM_2.5 levels for comparison with baseline exposure. Data for only the first week of baseline data collection was recoverable from the Kabwe PurpleAir monitors. Figure S1 in Supporting Information S1 shows the locations of PurpleAir monitors relative to study neighborhoods (Google Earth, 2023).

Similarly to the Atmotube Pros, PurpleAir monitors include no gravimetric filter measurement, and thus need to be evaluated and/or corrected. We were unable to perform collocations with the Lusaka PurpleAir monitors and filter measurements. Instead, to examine a representative range of “corrected” sensor results and constrain uncertainty, data were corrected using five different literature corrections, two derived using data from collocations in sub‐Saharan African cities (McFarlane, Isevulambire, et al., 2021, McFarlane, Raheja, et al., 2021) and three in the USA (Barkjohn et al., 2021; Holder et al., 2020; Magi et al., 2020). Any ambient concentrations reported here are the uncorrected PM_2.5 from the PurpleAir monitors along with the range of “corrected” concentrations generated by applying the different literature corrections. Details on corrections applied to the PurpleAir PM_2.5 data are in Text S3 and Figure S5 in Supporting Information S1.

2.3. Data Analysis and Statistical Modeling

Twenty‐four‐hour average exposure for CO during both phases and PM_2.5 at endline was calculated by averaging all real‐time data for the first 24 hr after cooks received the monitors. Baseline PM_2.5 24‐hr average exposure was determined from gravimetric filter measurements. Diurnal trends were developed from 1‐hr averaging of real‐time data. Any CO or PM_2.5 deployments that did not last at least 20 hr were excluded from analysis (n = 125, 13%). To test for differences between groups, we used Wilcoxon ranksum tests for all hypothesis testing at a 5% significance level; we chose this test because it is nonparametric, and our exposure sub‐sample sizes were relatively small and non‐normal. To test significance in comparisons of multiple groups (e.g., stove group), we used analysis of variance (ANOVA).

We applied four different approaches to analyze the impact of the two intervention stoves on cooks' CO and PM_2.5 exposures. First, we subtracted endline exposure from baseline exposure for households that had both temporal measurements for either CO or PM_2.5 to compare household‐specific differences in exposure. Next, we applied Wilcoxon ranksum testing to assess whether exposure was significantly different between cooks who did versus did not use the intervention stoves. Third, we conducted difference‐in‐differences (DiD) analyses in multiple permutations as a more robust estimate of intervention impact and to account for time‐varying factors and unobservable characteristics (Dimick & Ryan, 2014). Lastly, we implemented cross‐sectional generalized least squares (GLS) models across all household data (independent of intervention status) to quantify variables with significant associations with exposure.

DiD analysis is a common approach to estimating the impact of an intervention on observed outcomes. This approach allows us to compare changes in outcome (exposure) over time between control and intervention populations controlling for unobservable time varying factors (Dimick & Ryan, 2014). Each intervention stove was assessed independently. To minimize neighborhood‐level influences, we isolated DiD analyses to only the SupaMoto prospective user neighborhood, Kalingalinga, and the VITALITE prospective user neighborhood, Ng'ombe. We used a treatment‐on‐the‐treated (TOT) approach to determine which households were in control or treatment groups. Treatment versus control status was determined by stove use at endline: households that never used the intervention stoves were in the control group and households that started using the intervention stoves at endline were considered treated. We applied three estimation strategies using forward selection (i.e., adding new predictor variables each time in addition to previous variables) to include predictors we expected to influence exposure. Predictor variables used, their descriptions, and for which of the three DiD models they were included are described in Table 1. The simplest form for naïve DiD analysis (DiD Model 1) only included time, treatment group, and the interaction between time and treatment group, followed by the addition of the electric stove use variable (DiD Model 2), and the addition of household and other characteristics (DiD Model 3).

Table 1.

List of Predictors Used in Statistical Models and Their Descriptions

Predictor	Description	DiD model	GLS model
Time	Either baseline or endline	1, 2, 3
Treatment	Treatment group, either control or treatment	1, 2, 3
Time x treatment	The interaction between time and treatment group	1, 2, 3
Electric stove use	If participant had any amount of electric stove use	2, 3
Season	Either “cool” or “warm”	3	CO, PM2.5
Cooking location	Either “indoors” or “not indoors”	3	CO, PM2.5
Ventilation	If there was ventilation while cooking (i.e., opens windows)	3	CO, PM2.5
Expenditure quintile	Annual expenditure quintiles determined from reported annual household expenditures	3	CO
Stove group	Eight stove groups as listed in Table 2 with traditional charcoal as reference		CO, PM2.5
Neighborhood	Four neighborhoods with Matero as reference		CO, PM2.5
Total SI hours	Number of hours at elevated concentrations considered to be stove influenced		CO, PM2.5

Note. DiD model column lists for which DiD model the predictors were included, while GLS model column lists for which GLS model the predictors were included, either CO or PM_2.5.

We transformed the dependent variable to the natural log of the pollutant exposure to account for log‐normally distributed data (histograms and quantile‐quantile plots shown in Figures S6 and S7 in Supporting Information S1). If the interaction term was significant, the model suggests the intervention had significant impact on participants' exposures. We implemented 12 DiD models, one for each prospective user neighborhood for each pollutant for each model form. Text S4 and Figure S8 in Supporting Information S1 provide additional model details, a schematic of all model runs completed, and results of additional runs performed.

Because there were substantially more (∼12 times) control households than treatment, we used propensity score matching to ensure control and treatment groups were comparable at baseline. Matches were made from baseline characteristics we expected to influence adoption and use of the intervention stoves: household head's age, gender, and education, household member count, and sum of electronic assets. Using a 1:many approach, we estimated the 12 DiD models using the matched groups within each prospective user neighborhood (Chan & StatsNotebook Team, 2020; Stuart et al., 2011). We also implemented models with the full, “unmatched” sample (Table S2 in Supporting Information S1) and the panel subset only (Table S3 in Supporting Information S1) and found no significant differences in results between these models and those with the propensity score matched sample.

Additionally, we fitted cross‐sectional GLS models, considering all households at baseline and endline, to understand the variables significantly influencing exposure in addition to stove use. Due to heteroskedasticity of our exposure data (Figure S9 in Supporting Information S1), we opted to use GLS regression over ordinary least squares (OLS) regression so as not to break OLS assumptions. We developed two models, one for each pollutant, with the dependent variable as the natural log of the pollutant exposure concentration. The final model form for CO regressed log‐transformed 24‐hr average CO exposure (ppm) against stove group, season, neighborhood, cooking location, ventilation, total stove influenced (SI) hours, and expenditure quintile (see Table 1).

Total SI hours were estimated using a previously developed approach (Chen et al., 2016; Islam et al., 2022) to determine specific periods and the duration during the day that were spent at elevated concentrations. The resulting times/duration are labeled “SI,” but it must be clear that the CO exposures may come from other sources, so this is an inexact estimate of the impact of cooking on exposure. However, this predictor was included because it provides some proxy for the amount of time spent in proximity to a combustion source. Details of SI analysis are included in Text S5 in Supporting Information S1. Additionally, we implemented GLS models with pollutant exposures averaged during SI periods only instead of 24‐hr average exposures (Figures S11–S12 in Supporting Information S1); however, analysis results did not substantially change compared to 24‐hr average exposures, and thus we opted to use 24‐hr average exposures in our final model form as this is the norm in other similar intervention studies (Lai et al., 2024).

Because PM_2.5 exposure sample sizes were much smaller than CO samples, we used forward selection to sequentially add the same predictors used in the CO model while maximizing sample size and statistical power. Because significance did not change among predictors as we added each predictor, we opted to only include stove group, season, neighborhood, cooking location, ventilation, and total SI hours in the final PM_2.5 model (Table 1).

This study was reviewed and approved by the Institutional Review Board (IRB) at the University of North Carolina at Chapel Hill (19‐0061) and the Humanities and Social Science Research Ethics Committee at the University of Zambia (2019‐MAY‐012); study participants provided written consent.

3. Results

3.1. Characteristics of Exposure Monitoring Participants

Table 2 shows the characteristics of the exposure monitoring subset of participants at baseline by neighborhood. Generally, households had a high electrification rate (87%), most primary cooks cooked indoors (56%), and 30% of participants cooked with electricity to some extent (i.e., at minimum as stove used second most often). To ensure our subset of exposure monitoring participants was representative of the total household sample, we used Wilcoxon ranksum tests to check for significant differences between the exposure subset and the total sample in eight different metrics collected from the survey in each neighborhood. No metric in any neighborhood was significantly different between the exposure subset and total sample, suggesting that the exposure subset was representative of the total household sample (Table S4 in Supporting Information S1).

Table 2.

Characteristics of Exposure Monitoring Participants at Baseline

	SupaMoto (Mimi Moto)		VITALITE (EcoZoom)		All exposure subset
Characteristics	Matero (ICS users)	Kalingalinga (prospective users)	Kamanga (ICS users)	Ng'ombe (prospective users)
Household size	5.4 (2.1)	5.1 (2.3)	5.5 (2.4)	5.5 (2.7)	5.3 (2.5)
Female primary cook	88%	90%	98%	92%	92%
Primary cook age	36 (15)	33 (13)	37 (13)	33 (13)*	34 (13)
Primary cook education level
Preschool	0.0%	0.0%	0.0%	0.0%	0.0%
Primary	14%	32%	39%	44%	38%
Secondary	86%	62%	54%	53%	57%
Post‐secondary/higher	0.0%	5.9%	6.8%	2.7%	4.4%
Owns home	59%	35%	53%	39%*	41%
Has electricity access	82%	91%	91%	81%	87%
Cooks indoors	71%	48%	73%	54%*	56%
Gap between wall and roof	50%	38%	31%	37%	36%
Ventilation during cooking	44%	39%	50%	46%	47%
Total SI hours	8.0 (5.2)	12 (5.6)*	14 (4.8)	12 (4.7)*	12 (5.2)
Stove group*
Traditional charcoal	5.9%	44%	1.1%	70%	46%
Improved charcoal	5.9%	8.8%	1.1%	0.9%	3.9%
Charcoal + electric	18%	30%	2.2%	22%	21%
Electric + charcoal	5.9%	14%	0.0%	4.5%	7.0%
Electric exclusive	0.0%	3.3%	0.0%	1.3%	1.8%
Mimi Moto any	65%	0.0%	0.0%	0.0%	2.1%
EcoZoom any	0.0%	0.0%	96%	0.0%	17%
Other	0.0%	0.0%	0.0%	1.3%	0.6%
Sample size	17	181	93	223	514

Note. Averages and standard deviations (in parentheses) are provided for continuous variables. Statistical comparisons are between intervention cookstove (ICS) users and prospective users within each social enterprise grouping (SupaMoto and VITALITE).

^∗ p < 0.05.

As understanding the influence of stove use on exposure measurements was the focus of our study, we defined stove groups carefully. Because of frequent self‐reported stove stacking in our sample, there were many combinations of primary (stove used most often) and secondary (stove used second most often) stove use. The stove use grouping we used balanced the complex stove group dynamics while minimizing the total number of stove groups. Details on how we formed stove groups are found in Text S6 and Figures S13–S14 in Supporting Information S1. The eight stove groups we finally considered were as follows: traditional charcoal, improved charcoal (excluding EcoZoom), charcoal + electric (primary traditional/improved charcoal, secondary electric), electric + charcoal (primary electric, secondary traditional/improved charcoal), electric exclusive, Mimi Moto use any (either primary or secondary use regardless of other stove), EcoZoom use any, and other (gas, wood, kerosene; <1% of total households).

Within the exposure subset, we did observe statistically significant differences in household characteristics across neighborhoods. We compared the SupaMoto/Mimi Moto neighborhoods (Matero and Kalingalinga) and the VITALITE/EcoZoom neighborhoods (Kamanga and Ng'ombe) separately. As expected, stove group composition was significantly different between Matero and Kalingalinga and between Kamanga and Ng'ombe. Additionally, total SI hours were also significantly different between Matero and Kalingalinga and between Kamanga and Ng'ombe, with Matero cooks experiencing the lowest SI hours (8 hr) and Kamanga cooks the most (14 hr) on average. While no other variables in Table 2 were statistically significantly different between Matero and Kalingalinga, there are other differences worth noting. Approximately 60% of Matero households owned their home, compared to 35% of Kalingalinga households. Seventy‐one percent of Matero households cooked indoors, compared to 48% of Kalingalinga households. Kamanga and Ng'ombe exposure subsets were significantly different for primary cook age, home ownership, and cooking location, with Kamanga cooks being on average older, with a greater proportion of home ownership, and cooking indoors more often. Primary cooks in ICS neighborhoods on average cooked indoors more (73%) than those in prospective neighborhoods (51%), possibly negating some potential exposure reductions from using alternative stoves.

3.2. Personal Exposure by Neighborhood

We examined 24‐hr average CO and PM_2.5 exposures by neighborhood during both data collection phases. Per our study design, at baseline, Matero homes would be using the Mimi Moto, Kamanga homes the EcoZoom, and Kalingalinga/Ng'ombe homes traditional charcoal. At endline, after marketing in prospective neighborhoods, Mimi Moto stoves would be used in Matero and Kalingalinga and EcoZoom stoves in Kamanga and Ng'ombe. In reality, at baseline only 65% of Matero homes in the exposure subset were using the Mimi Moto and 96% of Kamanga homes were using the EcoZoom. Due to the prevalence of electricity use, only 44% and 70% of homes in Kalingalinga and Ng'ombe, respectively, were exclusively using traditional charcoal at baseline (Table 2). By endline, Mimi Moto use decreased to 40% in Matero and EcoZoom use decreased to 54% in Kamanga. In prospective neighborhoods, 17% of homes in Kalingalinga used the Mimi Moto (n = 15) and 21% of homes in Ng'ombe used the EcoZoom (n = 22) at endline. Of the households that started using intervention stoves, only 17 had both baseline and endline CO measurements, and none had endline PM_2.5 measurements.

First, we present CO and PM_2.5 exposures by neighborhood for both baseline and endline phases in Figure 2. Box and whisker plots shown here and throughout denote the following: boxes represent the interquartile range, middle line in boxes represents the 50th percentile, whiskers represent 1.5 times the interquartile range, and diamonds represents the average. Sample sizes are noted for each group and data from any group with fewer than 10 data points is shown as points with the group average as a diamond.

Figure 2.

Box and whisker plots of 24‐hr average CO exposure at (a) baseline and (b) endline and 24‐hr average PM_2.5 exposure at (c) baseline and (d) endline by neighborhood. Gray dashed lines denote the World Health Organization IT‐I Guideline for the respective pollutant. “+” denote neighborhoods with prospective users at baseline, that is, users not using intervention cookstoves per study design.

At baseline, Matero cooks, mostly using the Mimi Moto as their primary stove, had 78% lower CO exposure on average (3.8 ppm) compared to cooks in the other neighborhoods (17 ppm). Eighty percent of cooks in Matero experienced CO below the IT‐I Guideline, while only 40%, 19%, and 20% of cooks in Kalingalinga, Kamanga, and Ng'ombe did, respectively. Kamanga cooks, predominantly using the EcoZoom, had the highest CO exposure, at 19 ppm. PM_2.5 exposures were similar across all neighborhoods, averaging 89 μg m⁻³ during baseline, or 2.5 times greater than the IT‐I Guideline.

Like baseline CO exposure trends, Matero cooks also had the lowest CO exposure on average (3.9 ppm) compared to other neighborhoods (8.6 ppm) at endline. However, endline CO exposure in Kalingalinga, Kamanga, and Ng'ombe were on average 26%, 59%, and 52% lower than respective baseline exposures. PM_2.5 exposure at endline was consistent between neighborhoods, as it was at baseline, but also approximately 55% lower than baseline PM_2.5 exposure on average.

3.3. Personal Exposure by Stove Group

Because we found that use, adoption, and disadoption of intervention stoves varied widely within neighborhoods and was not consistent with expectations (e.g., widespread adoption of intervention stoves in Kalingalinga and Ng'ombe), we now explore exposure outside of the neighborhood delineation, specifically focusing on stove use. Figure 3 shows box and whisker plots of CO and PM_2.5 exposure by stove group for both baseline and endline. A few unexpected patterns stand out and highlight the complexity of stove use and transitions in our study households.

Figure 3.

Box and whisker plots of 24‐hr average CO exposure at (a) baseline and (b) endline and 24‐hr average PM_2.5 exposure at (c) baseline and (d) endline by stove group. Diamonds represent the average of each group. Gray dashed lines denote the World Health Organization IT‐I Guideline for the respective pollutant. Stove group abbreviations are as follows: “MM any” is the Mimi Moto any group, “EcoZ any” is EcoZoom any, “Trad. charc” is traditional charcoal, “Impr. charc” is improved charcoal, “Charc + elec” is charcoal + electric, “Elec + charc” is electric + charcoal, and “Elec exc” is electric exclusive.

At baseline, exclusive electric users and Mimi Moto users had comparable CO exposures, 1.6 and 3.6 ppm, respectively. Generally, cooks using exclusive charcoal (EcoZoom, traditional charcoal, improved charcoal) had the highest average CO exposures, and as users displaced charcoal with electricity, their exposures decreased. EcoZoom (19 ppm) and traditional charcoal users (18 ppm) had the highest average CO exposures, followed by, in decreasing order, improved charcoal (14 ppm), charcoal + electric (14 ppm), and electric + charcoal (9 ppm). A similar trend for endline CO exposure holds, except for Mimi Moto users: improved charcoal (10 ppm), EcoZoom (8.9 ppm), and traditional charcoal (8.0 ppm) users have the highest CO, followed by Mimi Moto users (7.1 ppm), then charcoal + electric (6.8 ppm), electric + charcoal (1.6 ppm), and electric exclusive (1.1 ppm). Therefore, we observe that as electricity use displaced charcoal, CO exposure decreased. Surprisingly, CO exposure for the Mimi Moto user group increased between baseline and endline, while it decreased for all other stove groups. This increase was likely due to the addition of Mimi Moto users in Kalingalinga at endline, per the intervention study design. As further explored in Section 3.5.3, there were neighborhood‐level differences in exposures, with Matero associated with significantly lower exposures. There was a weaker trend in baseline PM_2.5 exposure as electricity use increased, with traditional charcoal, EcoZoom, and charcoal + electric users having the highest PM_2.5 exposure (average 93 μg m⁻³) and electric + charcoal users with the lowest (59 μg m⁻³). During endline, Mimi Moto users had the highest PM_2.5 exposure on average (52 μg m⁻³) and charcoal + electric users had the lowest (28 μg m⁻³).

3.4. Daily Trends in Personal Exposure

The divergent associations between stove use and 24‐hr average CO versus 24‐hr average PM_2.5 exposures led us to examine the daily (diurnal) variation in each pollutant to observe trends. Figure 4 shows diurnal trends of CO exposure, PM_2.5 exposure, ambient PM_2.5, personal temperature (from the PM_2.5 monitor's temperature sensor), and ambient temperature during baseline and endline phases. For exposure and temperature profiles, the dark line shows the diurnal median and the shaded band shows the interquartile range. For ambient PM_2.5 diurnal trends, the darker line denotes the median uncorrected PM_2.5 from the PurpleAir monitors and the shading denotes the range of median diurnal PM_2.5 trends estimated by applying various corrections of these sensors from the literature (Text S3 in Supporting Information S1).

Figure 4.

CO exposure concentrations at (a) baseline and (b) endline, PM_2.5 exposure and ambient concentrations at (c) baseline and (d) endline, and personal and ambient temperatures at (e) baseline and (f) endline throughout the day. Baseline ambient PM_2.5 data were from two PurpleAir monitors in Kabwe, Zambia (no longer available as of December 2020), approximately 140 km north of Lusaka. Endline ambient PM_2.5 data were from two PurpleAir monitors we installed in Lusaka, Zambia. Darker line denotes median, and shading denotes the interquartile range for all diurnal plots except ambient PM_2.5 in panels (c) and (d). For ambient PM_2.5 diurnal plots, the darker line denotes the median uncorrected PM_2.5 from the PurpleAir monitors and the shading denotes the median range of corrected PM_2.5 from literature corrections (Text S3 in Supporting Information S1). The yellow star in (a) denotes the “nighttime peak” period (18:00 to 23:00) during baseline, apparent both in (a) CO personal exposure and (e) personal temperature.

During baseline, CO exposure peaked at three distinct time periods of around 9:00, 13:00, and 21:00, approximately at mealtimes (Figure 4a). The median nighttime peak (denoted by the yellow star) was 14 times greater than the daytime peaks. In contrast to CO, PM_2.5 exposure only peaked twice a day, at ∼7:00 and ∼19:00 (Figure 4c). While there is uncertainty in the exact magnitude of ambient PM_2.5 due to the range of applied literature corrections, exposure PM_2.5 peak times aligned well with ambient PM_2.5 peak times. Endline exposure diurnal trends were like baseline: CO exposure (Figure 4b) peaked three times a day (∼9:00, ∼13:00, and ∼20:00) and PM_2.5 exposure (Figure 4d) peaked twice a day (∼7:00 and ∼20:00). However, endline exposure magnitudes of both pollutants were on average 62% lower than baseline exposures, as shown in Figure 2. Like baseline, endline PM_2.5 exposure peak times were concurrent with those for ambient PM_2.5. Both PM_2.5 exposure and ambient PM_2.5 were lower at endline than baseline; thus, overall trends in exposure concentrations were consistent with ambient concentrations, regardless of season. To assess the correlation between personal and ambient PM_2.5, we compared 24‐hr average PM_2.5 exposure with ambient PM_2.5 averaged for the same 24‐hr period (Figure S15 in Supporting Information S1). Generally, PM_2.5 exposure and 24‐hr ambient PM_2.5 were moderately correlated, with a pearson correlation coefficent (r) of 0.54, with endline measurements more strongly correlated (r = 0.62) than baseline (r = 0.13).

We compared ambient temperature (measured by ambient PM_2.5 monitors) to personal temperature (measured by PM_2.5 exposure monitors) to gain additional insights into exposure trends. At baseline, ambient temperature was consistent with expectations: increase at sunrise, peak during midday, and decline late afternoon (Figure 4e). Personal temperature trends were generally consistent with ambient before 18:00, but after 18:00 there was another small peak in personal temperature. Thus, cooks experienced temperatures warmer than ambient at night. This small peak in personal temperature (denoted by the yellow star in Figure 4a) aligns with the CO exposure nighttime peak. We reason that because baseline data collection occurred during the cool season, it is likely this large nighttime peak in CO exposure and personal temperature was due to households' use of stoves for heating at night. Seventy‐two percent of households reported heating their homes during baseline, with the majority (97%) using traditional charcoal to do so. This hypothesis is also supported by examining CO exposure stratified by cooler versus warmer nights. Nighttime average CO exposure concentration from days where the average ambient temperature from 18:00–23:00 was below 17°C (median during study) was 60% greater than when the ambient temperature was above 17°C, suggesting that higher exposure during cooler nights was a result of using stoves for space heating (Figure S16 in Supporting Information S1).

The ambient temperature profile during endline was again consistent with expectation while personal temperature strayed from the ambient temperature trend, yet differently than it did at baseline (Figure 4f). Endline personal temperatures were consistent with ambient temperatures in the morning and evening but notably below ambient temperatures in the afternoon. Endline data collection occurred during the warm season, and this temperature deviation suggests participants were spending the afternoon indoors or in the shade to avoid the heat. Additionally, the nighttime CO peak during endline was less extreme than at baseline and more consistent in magnitude with other mealtime peaks. Nighttime personal temperature was consistent with ambient temperature, suggesting participants were not using stoves for heat at night and thus not increasing CO nighttime exposure outside of cooking.

3.5. Impacts of Intervention Stoves

3.5.1. Endline‐Baseline Exposure Differences

We applied four different approaches to assess the impact of the intervention stoves on exposures. First, we subtracted baseline exposure from endline exposure for households that had both measurements for either CO or PM_2.5 to compare household‐specific differences in exposures. We grouped households measured at both baseline and endline into four categories based on their stove use status between measurement phases: kept using original (same as at baseline, e.g., traditional charcoal, electric), kept using intervention (Mimi Moto or EcoZoom), stopped using intervention, or started using intervention. Figure 5 shows endline minus baseline exposure measurements for CO (ppm) and PM_2.5 (μg m⁻³) by stove use status group.

Figure 5.

Differences between endline and baseline (a) CO and (b) PM_2.5 exposures for each stove use group. Negative values indicate endline concentrations lower than baseline concentrations. The “kept using original” group includes households that kept any stove from baseline to endline, excluding intervention stoves.

Differences in CO exposure concentrations were negative, consistent with neighborhood‐level comparisons shown in Figure 2, where endline CO exposures were generally lower than baseline. Those who stopped using intervention stoves had the least negative difference in CO exposure on average, suggesting that disadoption of the intervention stove increased stove‐associated exposures and at least partially offset reductions in CO exposure from seasonal differences. However, based on ANOVA, CO exposure differences between these stove use groups were not statistically significant. For PM_2.5 exposure differences, only 15 households had both baseline and endline measurements, limiting our ability to draw significant conclusions. Generally, PM_2.5 exposure differences were negative, and the seasonal differences were far greater than inter‐household or ‐group differences.

3.5.2. Difference‐in‐Differences Estimates

Next, we applied a DiD analysis approach. As a first step, we compared only endline exposures between control and treatment groups for both intervention stoves, where the intervention was defined as a household using the Mimi Moto or EcoZoom. We defined the control groups as households that did not use the respective intervention stoves, while the treatment groups were households that did use the stoves. We performed Wilcoxon ranksum tests for significant differences between both control‐treatment comparisons (Mimi Moto and EcoZoom), one for each pollutant during endline, with the average differences in outcomes shown in Table 3. No control‐treatment comparison showed difference at a 5% significance level. CO exposure was similar for intervention stove users and non‐users, differing by less than 2 ppm on average. PM_2.5 exposure was greater for Mimi Moto users compared to non‐users and lower for EcoZoom users compared to non‐users. However, this simple hypothesis testing does not account for other variables that likely influenced exposure.

Table 3.

Average Differences of CO and PM_2.5 Exposures for SupaMoto and VITALITE Treatment‐Control Comparisons at Endline

			Outcome average difference (treatment‐control)
	Treatment group	Control group	CO (ppm)	PM2.5 (μg m−3)
SupaMoto	Mimi Moto users	Non‐Mimi Moto users	−1.3	16
VITALITE	EcoZoom users	Non‐EcoZoom users	0.63	−5.1

Note. Significance was determined from Wilcoxon ranksum tests. *p < 0.05.

Table 4 shows DiD analysis p‐values for interaction terms for all prospective user neighborhood‐pollutant‐model combinations. DiD Model 1 included only the basic predictors in a DiD analysis: time, treatment group (i.e., using the intervention stove or not), and the interaction between time and treatment group. DiD Model 2 added electric stove use as a control variable and DiD Model 3 added four additional control variables: season, cooking location, ventilation, and expenditure quintile. No combinations had a significant interaction term, suggesting that either there was no effect of either intervention or that the effect size was too small for our sample size and data set to demonstrate. Therefore, as a final analysis step that leveraged all study data, we implemented full sample cross‐sectional regression models to explore associations with exposure outcomes.

Table 4.

P‐Values of the Interaction Term Coefficients for the Full Sample

		P‐value for interaction term
Outcome	N	DiD Model 1	DiD Model 2	DiD Model 3
Kalingalinga CO	231	0.14	0.16	0.17
Kalingalinga PM2.5 a	n/a	n/a	n/a	n/a
Ng'ombe CO	317	0.83	0.73	0.83
Ng'ombe PM2.5 b	n/a	n/a	n/a	n/a

Note. Outcomes are the four prospective user neighborhood‐pollutant combinations for DiD Models 1, 2, and 3. *p < 0.05.

^a No PM_2.5 measurements for treatment group in Kalingalinga at endline.

^b No PM_2.5 measurements for treatment group in Ng'ombe at baseline.

3.5.3. Cross‐Sectional Regression Models

Figure 6 shows parameters related to stove use, season, neighborhood, cooking practices, and annual expenditures from CO and PM_2.5 cross‐sectional models. Coefficients and confidence intervals for each parameter were back‐transformed from the log‐transformed dependent variable. Categorical parameters are listed versus their reference, where a value of one represents 100% of reference value. A value below one represents a decrease in exposure compared to reference, while a value greater than one represents an increase relative to reference. Confidence intervals overlapping one signify we have no confidence the parameter was a significant predictor of exposure difference relative to the reference.

Figure 6.

Dot and whisker plots showing back‐transformed regression coefficients and confidence intervals for models of exposure to (a) CO and (b) PM_2.5. The y‐axis categories represent each predictor variable and its reference, with predictor categories denoted by background color. For binary variables, references are listed second. “Trad charcoal” represents traditional charcoal users. A value of one means predictor variables are 100% of reference. A confidence interval overlapping with one indicates no confidence (at the 95% level) that the predictor variable is significantly different than the reference. Predictors shown in red are statistically significantly different than reference.

The influence of electric stove use on exposure was notable. Primary cooks using electric stoves exclusively or primarily with charcoal stoves secondarily (electric + charcoal) had on average 78% and 44% lower CO exposures, respectively, compared to traditional charcoal users. Cooking with electric stoves secondarily with primary charcoal (charcoal + electric) was associated with 7.2% lower CO exposures, although not significantly. Thus, primary electric stove use, even if supplemented with charcoal, led to significantly lower CO exposures compared to traditional charcoal. Also, there was a clear decreasing trend in exposure as cooks replaced charcoal with electric use, with the smallest decrease for charcoal + electric users and the greatest decrease for electric exclusive users. Intervention stoves, on the other hand, were not associated with significantly lower exposures. Primary cooks using Mimi Moto and EcoZoom stoves to any extent had higher exposures on average than traditional charcoal users, 22% and 3.8%, respectively, although neither difference was statistically significant. Use of other stoves (wood, gas; <1% of sample) was associated with a 34% reduction in mean CO exposure, but this difference was not significant.

In addition to stove use, other contextual variables were associated with changes in CO exposure. Season was a significant predictor, with CO exposures in the warm season (endline) on average 37% lower than those in the cool season (baseline). Primary cooks living in Matero had lower CO exposures than cooks in all other neighborhoods, with Kalingalinga, Kamanga, and Ng'ombe cooks exposed to 1.4‐, 1.6‐, and 1.8‐times higher CO concentrations on average. However, only living in Kamanga and Ng'ombe was associated with statistically significantly higher CO exposure on average. Primary cooks who reported not cooking indoors (i.e., outdoors, on veranda) had significantly lower CO exposures on average (23%) than those who did cook indoors. Lastly, primary cooks who spent more time cooking or exposed to “SI” concentration levels had significantly higher CO exposures on average; for every additional SI hour, CO exposure increased 15% on average.

Our PM_2.5 model found no significant association between any stove group and mean exposure. A few stove groups had non‐significantly higher mean PM_2.5 exposures compared to traditional charcoal, specifically improved charcoal, Mimi Moto any, and EcoZoom any. We expect this is due to the small PM_2.5 exposure monitoring sample sizes; no stove group other than traditional charcoal and charcoal + electric had sample sizes >12. Season was the only significant predictor in the PM_2.5 model, with measurements during the warm season associated with 64% lower exposure than the cold season on average. Cooking location, ventilation during cooking, and total SI hours were not significantly associated with differences in PM_2.5 exposure concentrations.

To more directly assess the effect of ambient PM_2.5 on personal PM_2.5 exposure and the interpretation of our observations, we re‐ran the PM_2.5 GLS model described in Section 2.3 with 24‐hr average ambient PM_2.5 in place of season (Figure S17 in Supporting Information S1) as a predictor. Ambient PM_2.5 was significantly associated with personal PM_2.5, with a 1 μg m⁻³ increase in ambient PM_2.5 associated with a 1.7% (1.0%–2.3%) increase in personal exposure on average. Generally, other predictors remained not significant, consistent with the base‐case PM_2.5 model (Figure 6b), except for charcoal + electric versus traditional charcoal and Kamanga versus Matero. However, these associations are only marginally significant and are likely due to a decreased sample size when using ambient PM_2.5 as a predictor (ambient data were not available for all exposure measurements); for example, this model form only had four Kamanga households. It is notable that the neighborhood variables had relatively large changes with this change in predictor, suggesting that the interaction of sampling day and season may be at least partially driving the observed inter‐neighborhood variation in exposure (e.g., in CO, where we do not have ambient observations). Cross‐sectional modeling for both pollutants indicated that context had a greater influence on exposure than use of one of the intervention stoves. For CO exposure, the use of electricity, neighborhood, season, cooking location, and SI hours were all significant predictors. For PM_2.5 exposure, only season was significant, although our ability to discern differences is limited by much smaller sample sizes than for CO exposure.

Because we observed a decreasing trend in CO exposure as cooks moved from secondary electric use, to primary, to exclusive use, we wanted to better assess how replacing traditional charcoal with electric stoves decreased exposure. Using 3‐day cooking logs, where we recorded each stove participants used for every meal the past three days, we developed stove use count variables, which summed the total number of meals cooked on a particular stove over the 3‐day period. We re‐ran the CO GLS model described in Section 2.3 with the predictors shown in Figure 6a but replaced the categorical stove group variable with the stove use count variables. Results are shown in Figure S18 in Supporting Information S1; generally, we found the same trends as the base‐case CO regression model: warm season, cooking not indoors, spending less time at “SI” concentrations, and living in Matero were associated with significantly lower exposures. Number of meals cooked with an electric stove was the only significant stove use variable and suggested that replacing one meal cooked with traditional charcoal with an electric stove was associated with a 7.9% reduction in CO exposure, on average.

4. Discussion

We conducted a quasi‐experimental program evaluation in Lusaka, Zambia to assess the impacts on personal exposure of switching from traditional charcoal stoves to one of two alternative biomass stoves, the Mimi Moto or the EcoZoom, and long‐term use of alternative stoves. However, use and adoption rates of intervention stoves were generally low, with only 5.1% and 17% of households in our exposure monitoring subset using the Mimi Moto or EcoZoom, respectively, and only 12% of prospective households (n = 30) adopting an intervention stove between baseline and endline. Additionally, approximately 43% of households reported stove stacking, complicating the relationship between stove use and exposure. One hypothesis for why Mimi Moto use was so low in our study was that switching from traditional charcoal to the Mimi Moto requires both a stove and fuel (pellets) switch. Because the EcoZoom uses charcoal, which is the primary (75% of population) biomass fuel used in urban Zambia (Zambia Statistics Agency et al., 2019), switching from a traditional charcoal stove to the EcoZoom could be less burdensome (e.g., same price of fuel, same fuel purchasing location, same fuel handling practices) on households than switching to both a different stove and fuel.

Electricity access played an important role in our analysis. On average, 87% of households in the exposure subset had access to electricity. Because of this high electrification rate, we found considerable electric stove stacking, particularly in prospective user neighborhoods. At baseline, over 30% of households were making some use of an electric stove, with 1.8%, 7.4%, and 25% using electric stoves exclusively, primarily, or secondarily, respectively.

We implemented four analytical approaches to evaluate the effect our cookstove interventions had on primary cooks' personal exposure. No method suggested that switching to either the Mimi Moto or EcoZoom stove resulted in statistically significantly lower exposures. However, we observed that the more electric stoves were used in place of traditional charcoal, the greater the CO exposure reductions. This pattern has been observed in other studies replacing traditional biomass with clean stoves (Chillrud et al., 2021; Islam et al., 2022), suggesting switching to low‐emitting clean stoves, such as electricity or gas, could be a more beneficial focus for future cookstove intervention studies. In urban Zambia, encouraging exclusive electric stove use for cooking could be a focus, as electric stoves are the most popular clean stove alternative as of 2019, and electrification is already prevalent (Zambia Statistics Agency et al., 2019). However, concerns with electricity supply, reliability, and seasonal load shedding remain (USAID, 2023a). While dramatic expansion of fossil‐powered electricity could lead to increases in air pollutant and greenhouse gas emissions, recent work has found that such an increase would be overwhelmed by reductions in air pollutants and short‐lived climate forcing pollutants due to shifts away from solid fuel use and charcoal production (Floess et al., 2023). Alternatively, as of 2023, actions are underway to increase LPG uptake in Zambia, with a target of 40% of cooking with LPG by 2030 (USAID, 2023a), by removing customs duty on gas cylinders (USAID, 2023b). When fully adopted, LPG has the potential to dramatically reduce air pollution exposures (Johnson et al., 2022), and while it is not considered as clean as electricity, LPG is far cleaner than cooking with biomass fuels (Pillarisetti et al., 2022). However, since <1% of urban households in Zambia primarily used LPG in 2019 (Zambia Statistics Agency et al., 2019), substantial changes will be needed to enable this rapid transition in LPG adoption and use.

In addition to electric stove use, other variables influenced exposure in this urban setting. The ambient versus exposure temperature diurnal trends suggest that in the cool season, households used stoves for space heating, likely leading to higher nighttime CO exposures. PM_2.5 exposure diurnal trends showed consistency in peak times and magnitudes with ambient PM_2.5 concentrations during both seasons, suggesting personal exposure was dominated by local and regional sources other than in‐home cooking emissions (Dionisio et al., 2010; Johnson et al., 2024). Generally, exposures were lower during our endline versus baseline observations. This could be misinterpreted as indicating intervention effectiveness. However, closer analysis suggests that seasonal differences between background conditions during baseline and endline campaigns dominated the influence of intra‐household variability. Excluding CO exposure in Matero, average exposures during the cold season (baseline) were higher than IT‐I Guidelines (CO: 6 ppm; PM_2.5: 35 μg m⁻³) for both CO (15 ppm) and PM_2.5 (94 μg m⁻³), while warm season (endline) exposure averages (CO: 8 ppm; PM_2.5: 39 μg m⁻³) were similar to IT‐I Guidelines, reinforcing that measurements during multiple seasons are likely necessary to understand exposure implications of cookstove use and transitions. Lastly, differences between neighborhoods themselves were significant in determining exposure. CO exposure concentrations were the lowest in Matero. Matero was the ICS user neighborhood for SupaMoto, and thus this could be misinterpreted as Mimi Moto users having lower exposure than other stove user groups. However, even CO exposure for traditional charcoal users was lower in Matero (5.1 ppm) compared to the other neighborhoods (16 ppm), suggesting that lower CO exposure in Matero was attributable to neighborhood characteristics and not stove use. Some characteristics that could be associated with reduced exposures in Matero compared to other neighborhoods were higher primary cook education level, greater percentage of home ownership, and certain home characteristics, like easier/greater kitchen ventilation (Table 2). Additionally, these neighborhood characteristics may be associated with less traffic, a greater proportion of paved roads, and lower prevalence of burning activity (i.e., trash burning), which can substantially contribute to exposure (Dionisio et al., 2010).

The incorporation of data from other sensors during our study provided meaningful insight into the influences of seasonal, regional, and personal factors on exposure. Ambient monitoring proved crucial in interpreting the lack of relationship between PM_2.5 exposure and stove use. The use of temperature sensors to infer behavior during cool season nights and warm season afternoons also expanded our interpretation of personal CO exposure measurements. Our observations of using stoves for space heating are an important consideration when promoting alternative cookstoves. Stoves are used for more than just cooking, and for an intervention to be successful, the alternative stove will need to replace all uses of the traditional, such as space heating or provision of other energy services.

There are several limitations of our study. First, there was generally a low fraction of exposure monitoring that was unevenly distributed across neighborhoods, pollutants, and stove groups. Due to equipment constraints, we only collected CO and PM_2.5 measurements from 31% and 3.7% of our full sample, respectively. Substantially more exposure measurements were collected in prospective user neighborhoods (Kalingalinga, Ng'ombe) compared to ICS neighborhoods (Matero, Kamanga). Thus, only 28% of the full sample and 22% of the exposure monitoring subset used either intervention stove. Additionally, due to the COVID‐19 pandemic, endline measurements were collected after over two years and during a different season (warm) than baseline (cool). Therefore, we find a notable decrease in household participation from baseline to endline and that many baseline‐endline exposure differences were associated with seasonal differences rather than stove use. Ambient PM_2.5 data during baseline was only recoverable for the first week of data collection, limiting our ability to compare ambient PM_2.5 and personal PM_2.5 exposure during the cool season. Because we used real‐time CO concentrations to estimate SI hours (Text S5 in Supporting Information S1), it is not surprising that total SI hours were significantly associated with increased CO exposure. A model without SI hours as a predictor finds no difference in the significance of the interventions and a slightly stronger effect of neighborhood (versus the reference, Matero). This suggests that our SI hour metric may be influenced by neighborhood‐level CO sources/concentrations and is, as noted, an imperfect indicator of actual stove exposure time. Lastly, as the COVID‐19 pandemic occurred during our study, there were likely external influences on stove adoption and use that were not captured in our study, such as changes in exposure sources other than cooking, financial decision making, fuel cost and availability, and household energy demand and cooking practices.

Generally, quantifying the impacts of improved biomass stove use on exposure in an urban setting was challenging. Our urban setting had greater influences from other pollutant sources, diverse fuel availability and prevalent stove stacking, and significant neighborhood‐level differences compared to rural areas. Thus, attributing decreases in exposure to stove use other than exclusive electric stove use, including the intervention stoves, was challenging.

5. Conclusions

This study fills a gap in the literature by measuring exposures of urban sub‐Saharan African cooks who primarily use traditional charcoal stoves. We found that electric stove use was the only stove use‐related factor that significantly reduced exposures. We expected primary cooks who used intervention stoves, particularly the Mimi Moto, to have significantly lower exposures, as would be expected due to greatly reduced emissions measured in the field (Champion & Grieshop, 2019). However, small sample sizes, low adoption and use rates, stove stacking, seasonal variability, and the influence of urban and regional air pollution sources made it difficult to conclude if these cookstoves can reduce exposures. Our results illustrate that cooking is an important driver of exposure, particularly for CO. However, we only found evidence of this when contrasting with the “extreme clean” option, electricity, and other factors such as seasonality and neighborhood had an as‐big or bigger effect. Therefore, a narrow focus on HAP in such settings is unlikely to yield the needed and expected exposure reductions, but may still have positive effects, such as reducing climate impacts (Floess et al., 2023). A greater focus on improving ambient air quality is needed to address exposure in urban sub‐Saharan African settings.

Inclusion in Global Research Statement

This study was conducted as part of the Energy Poverty PIRE in Southern Africa (EPPSA), a 5‐year grant funded by the National Science Foundation's Partnerships in International Research and Education (PIRE) program in collaboration with Copperbelt University in Kitwe, Zambia and the Centre for Energy, Environment, and Engineering Zambia (CEEEZ) in Lusaka, Zambia. The objectives of EPPSA are to assess the impacts of energy poverty interventions, examine multiple dimensions of energy poverty, determine the most effective scale for interventions to enhance environmental outcomes and human well‐being, and encourage interdisciplinary collaborative research and training in low resource settings. Additionally, this project was funded by US Agency for International Development Partnerships for Enhanced Engagement in Research (PEER) grant cycle 7 between Stockholm Environmental Institute and CEEEZ. We thank our collaborators at Copperbelt University and CEEEZ for recruiting and training enumerators, organizing and implementing data collection, and providing invaluable local and regional perspectives on energy poverty solutions. We thank our enumerators and supervisors for collecting survey information and distributing exposure monitors during both campaign phases. We are grateful to SupaMoto, Zambia, and VITALITE, Zambia who implemented the cookstove interventions discussed in this study. We thank our study participants for responding to additional surveys and committing to wearing exposure monitors for 24 hr.

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Supporting information

Supporting Information S1

Parsons, S. , Hayes, W. , Kabwe, G. , Yamba, F. , Serenje, N. , Bailis, R. , et al. (2025). Impacts of improved cookstove interventions on personal exposure to carbon monoxide and particulate matter in Zambia. GeoHealth, 9, e2024GH001178. 10.1029/2024GH001178

Data Availability Statement

All data reported in this study, including personal exposure measurements, household questionnaire responses, and ambient PM_2.5 data, are preserved and available on Dryad (Parsons et al., 2025). Version 6.0.3 of Jupyter Notebook (under Python version 3.7.6) used for data processing, analysis, and visualization is preserved on Zenodo (Parsons et al., 2024) and published on GitHub at https://github.com/stephanieparsons14/Impacts‐of‐improved‐cookstove‐interventions‐on‐personal‐exposure‐to‐CO‐and‐PM‐in‐Zambia.