Health Trade-Offs of Boiling Drinking Water with Solid Fuels: A Modeling Study

Abstract

Background:

Billions of the world’s poorest households are faced with the lack of access to both safe drinking water and clean cooking. One solution to microbiologically contaminated water is boiling, often promoted without acknowledging the additional risks incurred from indoor air degradation from using solid fuels.

Objectives:

This modeling study explores the trade-off of increased air pollution from boiling drinking water under multiple contamination and fuel use scenarios typical of low-income settings.

Methods:

We calculated the total change in disability-adjusted life years (DALYs) from household air pollution (HAP) and diarrhea from fecal contamination of drinking water for scenarios of different source water quality, boiling effectiveness, and stove type. We used Uganda and Vietnam, two countries with a high prevalence of water boiling and solid fuel use, as case studies.

Results:

Boiling drinking water reduced the diarrhea disease burden by a mean of 1,100 DALYs and 367 DALYs per 10,000 people for those under and over 5 y of age in Uganda, respectively, for high-risk water quality and the most efficient (lab-level) boiling scenario, with smaller reductions for less-contaminated water and ineffective boiling. Similar results were found in Vietnam, though with fewer avoided DALYs in children under 5 y of age due to different demographics. In both countries, for households with high baseline HAP from existing solid fuel use, adding water boiling to cooking on a given stove was associated with a limited increase in HAP DALYs due to the log-linear exposure–response curves. Boiling, even at low effectiveness, was associated with net DALY reductions for medium- and high-risk water, even with unclean stoves/fuels. Use of clean stoves coupled with effective boiling significantly reduced total DALYs.

Discussion:

Boiling water generally resulted in net decreases in DALYs. Future efforts should empirically measure health outcomes from HAP vs. diarrhea associated with boiling drinking water using field studies with different boiling methods and stove types. https://doi.org/10.1289/EHP15059

Introduction

Globally, many households face the challenges of both poor drinking water quality and unclean cooking fuels. Two billion people lack safely managed drinking water.¹ In 2019, 1.53 million deaths were attributed to diarrheal diseases,² and in 2016, 60% of all diarrheal deaths in low-and middle-income countries (LMICs) were attributed to improper water, sanitation, and hygiene (WASH).³ In addition, in 2020, 2.8 billion people cooked using polluting fuels (solid fuels and kerosene),⁴ resulting in high exposures to fine particulate matter (PM) with an aerodynamic diameter $\le 2.5\mathrm{\mu }\mathrm{m}$ (${\mathrm{PM}}_{2.5}$).⁵ Household air pollution (HAP) is associated with various negative health outcomes, including stroke,⁶ ischemic heart disease (IHD),⁷ chronic obstructive pulmonary disease (COPD),⁸ lung cancer (LC),⁹ and acute lower respiratory infections (ALRI),¹⁰ among others.¹¹ Air pollution was associated with 6.7 million premature deaths annually in 2019,¹² with 2.3 million deaths attributed to HAP exposures.^13,14 Both diarrheal diseases and lower respiratory infections are leading causes of death for children under 5 y of age.¹⁵ Technologies providing safe water and clean household energy that are considered “low cost” are often unaffordable to many households,¹⁶ placing solutions out of reach.

A third of households in a study of 67 LMICs reported treating their drinking water at home.¹⁷ Boiling water is the most common household water treatment,¹⁸ with an estimated 1.2 billion users (70% of all household water treatment users).¹⁹ Based on household surveys, boiling for water treatment is most common in the Western Pacific region and least common in the Eastern Mediterranean and African regions. Boiling is widespread in many Asian nations, including Indonesia (90.6% of households practicing water treatment reported boiling), Mongolia (95.2%), Uzbekistan (98.5%), and Vietnam (91.0%).¹⁸ Though boiling on the African continent is comparatively less common, many countries in Africa have high rates of boiling,^18,20 including Lesotho, Rwanda, Uganda (more than 80% of households reported boiling), Burundi, and Namibia (more than 60% reported boiling).²⁰ Boiling for water treatment has been widely promoted for decades for low-income countries and emergency situations.²¹ A key limitation of boiling is the potential recontamination of stored boiled water by contact with hands and utensils, because boiling does not provide residual protection. Improper boiling methods can also result in poor water quality,^22–24 and families often mix boiled and nonboiled water.²⁵

Other concerns raised by boiling water are the potential for increased air pollution exposures from fuel combustion²⁶ and high fuel costs.²⁷ Solid fuel use is prevalent in low-income settings for both cooking and boiling water. Reducing the use of solid fuels reduces ${\mathrm{PM}}_{2.5}$ concentrations,²⁸ yet transitions to clean fuels, such as liquified petroleum gas (LPG), have proven challenging.²⁹ Despite the perception that clean fuels are better and more convenient,³⁰ the associated financial burden (e.g., the cost relative to wood, which is often cheap or free) and other barriers often prevent their widespread adoption.³¹ One proposed solution is “improved cooking solutions” (ICS), or “cooking solutions that improve, however minimally, the adverse health, environmental, or economic outcomes from traditional solid fuel technologies.”³² ICS include natural- or forced-draft biomass cookstoves with improved combustion efficiency. For example, pellet-fed gasifier stoves emit $>90\%$ less (${\mathrm{PM}}_{2.5}$) than conventional stoves.^33–35 However, despite their potential to provide health and other benefits relative to traditional stoves, behavioral, technical, and economic challenges have limited their adoption.^36,37 Electric stoves are an additional clean option but are not common in low-income countries because of their cost and the poor availability and reliability of electrical grid connections. In Uganda, for example, $<1\%$ of the population cooks with electricity³⁸ due to the expense of electricity and lack of availability and/or subsidies.³⁹

Those working to mitigate risks from HAP and unsafe drinking water face similar challenges in designing, implementing, and securing the sustained use of interventions such as clean fuel and household water treatment.³⁷ However, few studies have examined these linked risks together. A randomized controlled trial in Rwanda combining a “rocket-style” biomass-burning natural draft cookstove and water filter found that the intervention reduced the prevalence of reported child diarrhea by 29% and that the benefits of the program outweighed the financial costs.^40,41 In a related cost–benefit analysis, the averted disability-adjusted life years (DALYs) from using the water filter and improved cookstove were found to be 239 and 556 per 10,000 people per year, respectively.⁴⁰ However, the study observed no significant reduction in 48-h personal exposure to ${\mathrm{PM}}_{2.5}$,⁴¹ consistent with challenges faced by stove-replacement programs observed elsewhere.⁴² A study in China measured the reduction in thermotolerant coliforms (TTC) in water from boiling using different methods and modeled air pollution from boiling water. The modeled mean 24-h ${\mathrm{PM}}_{2.5}$ kitchen concentration from boiling water with biomass combustion was $79{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$,²⁶ substantially above the World Health Organization (WHO) Interim-1 target of $35{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$.⁴³ Boiling with electric kettles was associated with the largest reduction in TTC. However, the study did not measure health outcomes. To date, no study has specifically investigated the trade-offs associated with drinking water treatment by boiling using solid fuels and compared the health risks. The overarching goal of this study was to develop a modeling framework to quantify the net health impacts from boiling drinking water with solid fuels, accounting for a range of HAP-associated health outcomes and for diarrhea associated with water contaminated with fecal matter. We then applied this modeling framework using available literature values for inputs for two countries, Uganda and Vietnam, selected as case studies.

Methods

Framework Definition and Test Population

DALYs are commonly used to quantify health burdens because they account for morbidity with differential disease severity⁴⁴ and mortality. In our study, we used DALYs as the primary metric to compare multiple risks.⁴⁵ Quantitative microbial risk assessment (QMRA) models are commonly used to determine the diarrhea risk associated with consuming water from a particular water source.^46,47 For HAP, the population attributable fractions based on exposure–response curves for individual diseases are used to calculate the burden of disease.^48,49 We used the QMRA and exposure–response curves as opposed to direct epidemiological evidence to calculate the DALYs for a range of conditions, including different levels of water boiling effectiveness, source water qualities, and HAP levels. Although epidemiological estimates exist for associations between boiling drinking water and child diarrheal outcomes, these provide an average over a range of conditions in the study population and do not allow estimating effects under specific scenarios and boiling efficiencies. Though the DALY calculation methods for these two risk factors are very different, they represent the best available established methods to calculate the risks.

We adopted these two methods (Figure 1), creating two modules, and used literature-derived distributions of the relevant HAP, QMRA, and demographic parameters as inputs (Table S1; Table 1). The “water risk module” uses a QMRA model to calculate the DALYs from drinking water contaminated by fecal matter before and after treatment by boiling. The “air risk module” uses an indoor box model to quantify the ${\mathrm{PM}}_{2.5}$ concentrations for different stoves and a modified form of the Household Air Pollution Intervention Tool (HAPIT; version 3.1.1)^49,52 to quantify the DALYs associated with HAP under various scenarios. Both modules employ Monte Carlo simulations to capture the influence of variability and uncertainty in input parameters on model outputs.

Figure 1.

Overview of methods to calculate and compare drinking water DALYs and HAP DALYs. Air health risks are for COPD, IHD, LC, ALRI, and stroke. Note: ALRI, acute lower respiratory infection; COPD, chronic obstructive pulmonary disease; DALYs, disability-adjusted life years; HAP, household air pollution; IHD, ischemic heart disease; LC, lung cancer; ${\mathrm{PM}}_{2.5}$, particulate matter with an aerodynamic diameter $\le 2.5\mathrm{\mu }\mathrm{m}$.

Table 1.

Household demographics for Vietnam and Uganda.

	Vietnam	Uganda
Average household size	5a	5a
Life expectancy: women	79.2b	69.2b
Life expectancy: men	70b	62.3b
Adults per household	4.6c	4c
Children under 5 y per household	0.4c	1c
Total country population	96,362,928b	41,117,856b
Cooks per household	1c	1c
Adults and children 5 y of age and over 5 y of age (5-and-overs) (not including cooks)	3.6c	3c
Children under 5 y of age (under-5s)	0.4c	1c
Percent solid fuel use	49%c	95%c

^a United Nations.⁵⁰

^b Institute for Health Metrics and Evaluation.⁵¹

^c Pillarisetti, Mehta, and Smith.⁵²

Globally, more than half of all deaths from HAP from solid fuel are from ALRI in children under 5 y of age,⁵³ and most diarrheal DALYs occur in children under 5 y of age.⁵⁴ Therefore, we conducted analyses for two age groups: those under 5 y of age (“under-5s”) and all other age groups combined (“5-and-overs”).

We designed the model to be used for any setting. However, we selected Uganda and Vietnam as case study countries because they are in distinct regions, have different population demographics, and high prevalences of boiling among household water treatment users (82% in Uganda,²⁰ 91% in Vietnam¹⁸) and solid fuel use (96% in Uganda, 35% in Vietnam).⁴ To simplify the model for both countries, we assumed one cook for each household and a household size of five people per household, based on average household size and composition for LMICs^49,55 and default values used in the HAPIT model.⁵² The number of adults and children under 5 y of age varied by country, with 4.6 people over 5 y of age and 0.4 children under 5 y per household for Vietnam, and 4 people over 5 y and 1 child under 5 y for Uganda.⁵² In addition to the household demographics, we also varied the life expectancy, background disease data, deaths, and DALYs for each country and kept the rest of the model inputs the same. The models were run for a sample population of 10,000 people.

Health Benefits from Boiling Drinking Water

The risk of illness from contaminated drinking water was characterized using a QMRA, focusing on selected reference pathogens.⁵⁶ Pathogens used in QMRAs are typically selected based on global public health relevance; because they are transmitted via environmental, waterborne, and foodborne routes,⁵⁷ the dominant pathogens may also depend on geographic region.⁵⁸ In addition, reference pathogens were selected to span different waterborne pathogen classes (viruses, bacteria, protozoa).^59,60 For this study, we selected pathogens identified as leading causes of diarrhea in multisite studies of diarrhea etiology in low-income countries.^61–63 Based on these studies, the locations selected for our modeling exercise, and previous QMRA studies,^59,64,65 we selected a virus (rotavirus), a protozoan (Cryptosporidium), and a bacterium (Campylobacter) to quantify the risk from exposure to fecally contaminated water.

A QMRA for drinking water involves four steps^{47,59,65–68}: a) hazard identification, b) exposure assessment, c) dose–response quantification, and d) risk assessment. For the hazard identification, we used a uniform distribution of E. coli levels for each water quality category, using most probable number (MPN) ranges of: $0\text{\hspace{0.17em}MPN}/100\mathrm{mL}$ for “safe” water, $1–10\text{\hspace{0.17em}MPN}/100\mathrm{mL}$ for “low-risk” water, $11–100\text{\hspace{0.17em}MPN}/100\mathrm{mL}$ for “medium-risk” water, and $101–\mathrm{1,000}\text{\hspace{0.17em}MPN}/100\mathrm{mL}$ for “high-risk” water.⁶⁹ These untreated drinking water categories defined baseline water qualities for the model. We used fecal indicator bacteria to pathogen ratios for the three pathogens (see Table 2) from the literature to estimate the abundance of the selected pathogens in untreated water.

Table 2.

Ratio of E. coli to individual pathogens modeled and their dose–response relationships, illness probabilities, and percent susceptible, based on the literature.

Pathogen	Ratio of E. coli to pathogen	Dose–response parameters and model	Probability of illness given infection, percent susceptible
Campylobacter	1:0.6657,64	$\mathrm{\beta }$-Poisson $\mathrm{\alpha }$: $\text{mean}=1.51\times {10}^{-1}$, $\text{SD}:5.9\times {10}^{-2}$, ${\mathrm{N}}_{50}$: $\text{mean}=1.69\times {10}^{-3}$, $\text{SD}:2.78\times {10}^{-3}$, ${\mathrm{P}}_{\text{Di}}=0.3$59,64,66,70,71,72	30%, 100%59,64
Rotavirus	$1:5\times {10}^{-6}$ 64	$\mathrm{\beta }$-Poisson $\mathrm{\alpha }$: $\text{mean}=2.48\times {10}^{-1}$, $\text{SD}:1.46\times {10}^{-1}$, ${\mathrm{N}}_{50}$: $\text{mean}=8.16$, SD: 6.65, ${\mathrm{P}}_{\text{Di}}=0.51$59,66,70	50%, 13%59,64
Cryptosporidium	$1:{10}^{-6}$ 57,64	Exponential $r$: $\text{mean}=3.44\times {10}^{-1}$, $\text{SD}:1.46\times {10}^{-1}$, ${\mathrm{P}}_{\text{DI}}=0.7$59,71,73	70%, 100%59,64

Note: SD, standard deviation.

We then estimated boiling effectiveness for the selected pathogens. The modeled intervention was boiling at different levels of microbiological effectiveness to account for different field conditions and household practices. The boiling effectiveness was quantified using log reduction values (LRV), defined as the base-10 logarithm of the ratio of influent to effluent pathogen concentrations. The effectiveness of boiling drinking water was compiled from a literature review of field and lab studies; see Table S2 for the full set of values from the literature review and Table 3 for values used in model. This literature review included studies of boiling from nine different countries from three continents, as well as laboratory studies. We chose a range of boiling effectiveness for use in the model, ranging from ineffective to laboratory level effectiveness. Water boiling studies reported LRVs for different fecal indicator organisms, including E. coli, TTC, and fecal coliforms (FC). Because our modeled water quality was characterized using E. coli, we converted TTC and FC LRVs to E. coli LRVs⁷⁶ based on assumed ratios. Although E. coli to FC ratios can vary in different seasons,⁷⁷ the ratios were assumed to be constant throughout the year. LRV values in the literature (converted to E. coli) ranged from 6 for lab-level boiling to $-0.26$ for ineffective boiling. We multiplied E. coli LRVs by a pathogen-specific factor to quantify removal of the respective pathogens (Cryptosporidium, Campylobacter, and rotavirus) (Table 4).

Table 3.

Literature sources of log reduction values of E. coli and thermotolerant coliforms used in our study.

Label used in this paper’s results	Country, water source, and reported treatment	E. coli a	TTC
Lab level	Australia, laboratory, heated for 1 min, ${75}^{\mathrm{o}}\mathrm{C}$	6.0074	—
Best field	Cambodia, surface water, household reported boiling	2.0075	—
Average field	Cambodia, well, household reported boiling	1.5075	—
Low field	Peru, urban various sources, household reported boiling	Converted to 0.35	0.4424
Ineffective field	Peru, rural various sources, household reported boiling	Converted to 0.0480	0.0624

Note: —, no data; LRV, log reduction value; TTC, thermotolerant coliforms.

^a All values were converted to E. coli for comparison in the model if not reported as “E. coli LRV.”

Table 4.

Effect of boiling on E. coli, Campylobacter, cryptosporidium, and rotavirus.

OrganismsSource	LRV52	Temperature	Contact time (seconds)
E. coli 74	6	75°C	60
Campylobacter 74	5	63°C	300
Cryptosporidium 74	8	75°C	60
Rotavirus78	8	95°C	60

Note: LRV, log reduction value.

For the exposure assessment, we used Equation 1 to calculate the daily exposure to pathogens (${\mathrm{E}}_{\mathrm{d}}$, in MPN),⁶⁸

$$E_d=C\times \frac{1}{R}\times I\times {10}^{-LRV}\times V,$$ (1)

where C is the concentration of pathogen in the source water (MPN per liter), R is the fractional mean analytical recovery (to account for success in microorganism counts) of the pathogen in the sample, I is the fraction of pathogens that are infectious,^59,68 LRV is the pathogen-specific log reduction by boiling water, and V (liters) is the daily volume of water consumed per person. We assumed I and R to be 1.

We calculated the probability of illness from the daily exposure using dose–response relationships from the literature.^{59,64,66,70,71} The dose–response parameters were assumed to be the same regardless of age (Table 2).

We calculated the probability of developing disease (diarrhea) per single exposure (${\mathrm{P}}_{\mathrm{D}}$) as the product of the probability of disease given infection (${\mathrm{P}}_{\text{DI}}$) and the probability of infection (${\mathrm{P}}_{\mathrm{i}}$) (Table 2). The probability of developing illness over period n, (${\mathrm{P}}_{\mathrm{D},\mathrm{n}}$), was calculated using the probability of disease [${\mathrm{P}}_{\mathrm{D},\mathrm{n}}=1-{\left(1-{\mathrm{P}}_{\mathrm{D}}\right)}^{\mathrm{n}}$], and the annual symptomatic cases was calculated by multiplying the yearly probability of developing illness (${\mathrm{P}}_{\mathrm{D},\text{yearly}}$) times the exposed population.^47,65,68

For the risk assessment, DALYs from diarrhea (years lived with disability) were calculated by taking the disability weight times the duration of disease times the probability of disease.^64,65,68 Disability weights to quantify disease severity were obtained from the literature (Table 5). When calculating years of life lost, the disability weight is 1 and the duration was the life expectancy minus the age at death.⁶⁴ To make the life expectancy represent the demographics of the country of interest, we adjusted this method so that the remaining life expectancy at the time of death was selected using stratified random sampling, based on the population age distribution in that country, separated into two categories, children $<5$ y of age and all other ages. For disease burden, severity and duration varied by pathogen (Cryptosporidium, Campylobacter, or rotavirus).⁶⁴

Table 5.

Burden of disease for each pathogen.

Pathogen	Outcomes	Severity64	Duration (days)64	Duration (y) based on lifetime (56 here)64	DALYs per case64	Likelihood of outcome64
Campylobacter	Gastroenteritis: population	0.067	5.1	0.014	0.0009	94%
Campylobacter	Gastroenteritis: doctor-diagnosed	0.39	8.4	0.023	0.009	6%
Campylobacter	Death from gastroenteritis	1	—	Life expectancy from population distribution	—	0.1%
rotavirus	Mild diarrhea	0.1	7	0.02	0.0019	86%
rotavirus	Severe diarrhea	0.23	7	0.02	0.0044	14%
rotavirus	Death from diarrhea	1	—	Life expectancy from population distribution	—	0.7%
Cryptosporidium	Watery diarrhea	0.067	3.4	0.009	0.0024	100%
Cryptosporidium	Death	1	—	Life expectancy from population distribution	—	0.4%

Note: —, no data; DALYs, disability-adjusted life years.

Health Impacts from HAP

In the air risk module, first an air pollution box model was used to estimate the 24-h ${\mathrm{PM}}_{2.5}$ concentrations in the household⁷⁹ and then DALYs were estimated using the framework of the HAPIT model.⁴⁹ We picked the HAPIT model because it is readily available and used in previous studies. Most of the disease burden for children under 5 y of age comes from ALRI, with a very small fraction of the disease burden from children under 5 y ($<1\%$) coming from COPD and stroke, which are accounted for in the model.³ All five diseases in our model (ALRI, COPD, IHD, stroke, and LC) affect the over-5 population.

For the air risk module, we considered two baseline scenarios: one with a household “already cooking” on a traditional woodstove with no water boiling, and one in which a household is “not cooking, and also not boiling water.” We considered the following cookstoves as replacements for this baseline: a) improved wood, b) charcoal, c) gasifier, d) LPG, and e) electric.^30,80,81 In the model, we assumed 100% adoption (e.g., LPG completely replaces traditional wood for all household energy needs). Electric stoves were included to serve as an ideal counterfactual (i.e., completely clean) cooking technology. We considered three categories of stove use per day: “cooking only” (no water boiling), “water boiling only” (no cooking), and “cooking and water boiling.”

Our model assumes that stove energy is only used for cooking and water heating and that all cooking takes place indoors. The required cooking energy (delivered to pot) for both the Ugandan and Vietnamese households was assumed to be log-normally distributed, with a mean [standard deviation (SD)] of 11 (5.5) megajoules (MJ) per day.^79,82 The daily energy for water heating ($E_{WH}$; Equation 2) was calculated as the energy needed to heat the water from ambient temperature ($T_a$, assumed to be 15°C) to boiling (100°C) and then boiled for one minute. The volume of water heated (V, liters) was assumed to be normally distributed with mean (SD) 3.12 (1.17) liters.⁸³ This was converted to mass using density ($\mathrm{\rho }$, $\mathrm{1,000}\mathrm{g}/\mathrm{L}$). $C_p$ is the heat capacity of water ($4.186\mathrm{J}/\mathrm{g}/\mathrm{K}$). Finally, the stove was assumed to boil the water for 1 min, so the product of 60 s (t) and the power of the stove (P, in watts) was added to get the total energy demand for heating water in a household.

$$E_{WH}=V\times \mathrm{\rho }\times C_p\times \left(100-T_a\right)+P\times t.$$ (2)

Laboratory studies of boiling effectiveness have been conducted for numerous different temperatures and durations of boiling. Generally, heating water at higher temperatures for longer periods of time results in greater LRVs in comparison with lower temperatures and shorter periods.⁷⁴ The recommended boiling time in the literature ranges from 1 to 25 min.²² Though pathogens are deactivated at temperatures $<100°\mathrm{C}$,²⁷ in field studies of water boiling, households typically heated water until they saw bubbles and often longer.²² In this study, we assumed that once water reaches boiling temperature (calculated as described above), it is boiled at 100°C for 1 min, based on US CDC^21,84 and WHO²⁷ recommendations, to reach the assumed LRV. The influence of elevation on boiling temperature was not considered in this analysis, but this would impact the temperature and thus the length of time needed for boiling for a given LRV.

Once the energy demand was estimated, an air pollution box model⁷⁹ was used to calculate the indoor kitchen ${\mathrm{PM}}_{2.5}$ concentration (${\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$) over 24 h. The indoor air pollution model was based on a single-box model with variability captured using a Monte Carlo simulation and calculated room concentrations assuming a well-mixed room and a single emission source.⁷⁹ Model inputs included the daily cooking time, the emission rate, kitchen air exchange rate and volume, and concentration from proceeding time step.⁷⁹ The time for cooking and water heating was calculated using the energy demand divided by the assumed stove power (watts) and thermal efficiency (Table 6). We assumed that cooking and water heating each occurred once per day to produce all daily energy for cooking and water heating. The emission rates were calculated using the relevant emission factor (grams ${\mathrm{PM}}_{2.5}$ per kilogram fuel), stove power and thermal efficiency, and heating value of fuel (Table 6). We selected a range of stoves and fuels ranging from basic, “unimproved” models (e.g., traditional wood stove) to modern, clean option (e.g., gasifiers, LPG, and electric stoves). To simplify the model, we also assumed an ambient concentration⁵² to which the indoor concentration decays after the cooking event and a second event (water heating) increases it again. Because the ambient concentration could vary drastically depending on the location (such as urban or rural) and also varies between Uganda and Vietnam,⁹⁴ for simplicity we assumed a background concentration of $12{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$,⁹⁵ to represent a low ambient value with minimal health impacts. If a single stove did not provide sufficient power to heat the specified daily water and food, allowing time for household concentrations to return to ambient levels between stove uses, a second stove was used in the model simultaneously, and in this scenario, we doubled emission rates. Although emission factors vary between stove operation stages (e.g., start-up vs. steady operation),^96–99 we assumed a constant average emission factor for the entire burning process. This approach may underestimate 24-h emissions, because it does not reflect multiple starting events, which can produce high emissions.¹⁰⁰ For an example of the 24-h ${\mathrm{PM}}_{2.5}$ kitchen concentration for cooking and water heating, see Figure S1.

Table 6.

Emission factors, stove power, and thermal efficiency of stoves and fuel heating value.

Stove	Category	${\mathrm{PM}}_{2.5}$ EF $\text{Mean}\pm \text{SD}$, g/kga	Stove power (watts)	Thermal efficiency (% $\pm \text{SD}$)5	Fuel heating value (MJ/kg) mean, min, max, $\pm \text{SD}$b (Distribution)
Traditional three-stone fire	Traditional wood	$7.1\pm 1.3$ 85,86	1,95885,86	$14.8\pm 1.8$ 85,86	14.780, 13.500, 15.88387 (Triangular)
Jiko	Charcoal	0.4788	6,05688	$23.30\pm 0.29$ 88	35.989 (Point)
MimiMoto	Gasifier	0.07990	3,05690	37.990	$16.1\pm 0.18$90 (Normal)
Chitetezo Mbaula	Improved wood	3.691	1,116.791	2091	14.780, 13.500, 15.88387 (Triangular)
LPG	Clean	$0.053\pm 0.04$ 92	1,52992	$54.4\pm 1.5$ 92	45.78093 (Point)

Note: EF, emission factors; LPG, liquified petroleum gas; MJ, megajoules; PM, particulate matter; ${\mathrm{PM}}_{2.5}$, PM with an aerodynamic diameter $\le 2.5\mathrm{\mu }\mathrm{m}$; SP, stove power; SD, standard deviation; TE, thermal efficiency.

^a Distributions are log normal.

^b Dry basis heating value is given for biomass; lower heating value is used for LPG.

To calculate the health burden from HAP, we adapted the approach of HAPIT.^49,52 HAPIT (version 3.1.1)¹⁰¹ model inputs include 2010 background disease data, 2020 population data, 2010 solid fuel use data, and 2013 average household size.⁴⁹ We modified the inputs to use more recent 2019 background disease data^51,102 in addition to our modeled pre-and postintervention ${\mathrm{PM}}_{2.5}$ exposures. The preintervention exposure is from the traditional wood stove, and the postintervention exposure is from one of the five cleaner stoves (charcoal, improved wood, gasifier, LPG, or electric). The relative risks (RRs) were calculated for each exposure level, and this was used to calculate the population attributable fraction, the attributable burden, and the averted burden associated with an intervention (e.g., switching from a baseline to cleaner stove).^49,52 Personal exposure is estimated by multiplying the modeled 24-h ${\mathrm{PM}}_{2.5}$ kitchen concentration by an estimated ratio of personal exposures to kitchen concentrations,^53,103,104 with separate ratios applied for the cook, noncook, and children under 5 y in a household (Table 7).

Table 7.

Personal exposure to concentration ratios.

Person	Women	Women to children	Women to men
Ratio52	0.742	0.85	0.61

The burden of disease attributable to household ${\mathrm{PM}}_{2.5}$ pollution was calculated for LC, IHD, stroke, ALRI, and COPD using 2019 background disease data, deaths, and DALYs^51,102 for Uganda and Vietnam, respectively, and the disease-specific integrated exposure–response functions from Burnett et al.,¹⁰⁵ using data and code from the Institute for Health Metrics and Evaluation (IHME) website.¹⁰⁶ Uncertainty was taken into account using results from Burnett et al., in which the integrated exposure–response models were fit to 1,000 sets of source type-specific RR values, creating 1,000 sets of parameter estimates.¹⁰⁵ In our modeling, we randomly drew from these 1,000 sets of parameter estimates as part of the Monte Carlo simulation approach taken for the entire model. Although there is no clear safe level of ${\mathrm{PM}}_{2.5}$ exposure,¹⁰⁵ we used the distribution of $5.8–8.8{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ as a counterfactual “no effect” level for this model,¹⁰⁴ which is the same level used by the HAPIT model,^49,52 though we used a distribution instead of the point estimate HAPIT uses.^49,52

The RR and the existing fraction of each country’s population exposed to solid fuels (i.e., fraction exposed equals the percent solid fuel users for each country) were used to calculate the attributable fraction (AF) (Equation 3).⁵² In calculating the attributable fraction in the model, the fraction exposed is country-specific and fixed, but the RR varies with air pollution level.

$$AF=\frac{FractionExposed\times \left(RR-1\right)}{FractionExposed\times (RR-1)+1}.$$ (3)

The attributable fraction was multiplied by the DALYs or deaths from a given disease in the country or region of that specific population (given location and age group) to calculate the attributable burden associated with HAP.⁴⁹ To calculate the fraction of DALYs from children under 5 y of age, we used the fractions of children in that population and relevant under-5 DALYs for each disease.

Simulations, Analysis, Statistical Tests, and Sensitivity

When comparing HAP and drinking water DALYs, we defined “net DALYs” as the increase in HAP DALYs minus the decrease in drinking water DALYs resulting from boiling drinking water. Positive net DALYs means that the HAP DALYs increase is greater than the water DALY decrease, indicating a net increase in disease burden. Negative net DALYs means the water DALYs decrease is greater than the HAP increase, indicating a net health benefit.

We conducted Monte Carlo simulations to capture the influence of variability and uncertainty in the inputs. Each simulation draws from distributions for parameters of a stove and fuel type and water boiling effectiveness and was run 10,000 times in R (version 4.3.1; R Development Core Team). The mean, SD, and 95% confidence intervals of model output are reported. In addition, to understand the variation in the outputs, the coefficient of variation (COV) (SD over mean) was calculated. Environmental parameters are often log-normally distributed,¹⁰⁷ so we used the Shapiro-Wilk test to test the normality of our resulting distributions of drinking water and HAP DALYs before making statistical comparisons between different scenarios. If an output distribution was log-normal, it was log-transformed for hypothesis testing. We used the $t$-test to compare the risks across the different scenarios (i.e., different stove types or water boiling scenarios). $p$-Values $<0.05$ were considered statistically significant.

A sensitivity analysis was conducted for the HAP and drinking water QMRA models to identify the specific impact of individual input variables on model output. Each input parameter was individually evaluated by varying the assigned value between a minimum and maximum value determined based on an assessment of the variability or uncertainty of the parameter from our literature review (Tables S3 and S4). Input parameters were then ranked in order of their influence on output values by taking the ratio of the output values for input at its minimum and maximum values.¹⁰⁸

Results

DALYs from HAP

The simulations with the lowest to highest 24 h average ${\mathrm{PM}}_{2.5}$ kitchen concentrations (for cooking and water heating scenarios) were electric, LPG, gasifier, charcoal, improved wood, and traditional wood, with values ranging from $12{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ (for the electric stove) to $\mathrm{5,144}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for cooking and water heating (data provided in Excel Table S11). The average 24-h ${\mathrm{PM}}_{2.5}$ concentration for the “worst-case,” traditional wood stove scenario was lower for water heating alone ($\mathrm{1,790}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$) than for cooking alone ($\mathrm{3,367}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$), whereas the concentration associated with both activities together was essentially the same as the sum of the two activities considered separately.

Figure 2 shows the DALYs calculated from these ${\mathrm{PM}}_{2.5}$ values for the different stove and use scenarios for each country. Though the total DALYs per 10,000 people were similar between the two countries, the number of HAP DALYs associated with under-5s were higher in Uganda in comparison with Vietnam, with the under-5s DALYs in Uganda making up 22%–50% (depending on scenario) of total, vs. under-5s DALYs in Vietnam making up 1%–2.4% of total DALYs. DALYs became significantly lower as stove type shifted from traditional wood to LPG stoves for cooking and water heating, with 96.8%, 97.1%, 97.3%, and 99.9% changes in DALYs for Uganda 5-and-overs, Uganda under-5s, Vietnam 5-and-overs, and Vietnam under-5s, respectively, eliminating almost all the DALYs associated with HAP. For the electric stove, because it was assumed that the electric stove contributes no additional ${\mathrm{PM}}_{2.5}$, DALYs were 100% lower when the electric stove replaced other stoves. There were only small differences in DALYs when a traditional wood stove was replaced with an improved wood stove. For example, DALYs were 2.9% lower (for Uganda under-5s) and 11.0% lower (Vietnam 5-and-overs) for water heating and cooking when switching from traditional wood stove to improved stove, suggesting the limited potential for health impacts from replacing traditional stoves with another basic wood stove.

Figure 2.

Mean DALYs from indoor air pollution for different stove and use scenarios for (A) Uganda and (B) Vietnam. Error bars show $\pm 1\text{\hspace{0.17em}SD}$. Baseline: no stove use. DALYs are calculated for scenarios of cooking only, water boiling, and water boiling plus cooking. Simulation was run for 10,000 iterations of our Monte Carlo simulation, for a population of 10,000 people. Under-5s refers to children $<5$ y of age, and 5-and-overs refers to all other age groups combined. The bars show Cooking (label “A”), Water Heating (label “B,” and Water Heating & Cooking (label “C”), Under-5s (label “$<5$”), and 5-and-overs (“$\ge 5$”). The stoves presented are electric, LPG, Gasifier, Charcoal, Imp. Wood, and Trad. Wood. Data for figure provided in Supplemental Excel File tabs Excel Table S1 (Uganda) and Excel Table S2 (Vietnam). Note: DALYs, disability-adjusted life years; Imp. Wood, improved wood; LPG, liquified petroleum gas; SD, standard deviation; Trad. Wood, traditional wood.

In contrast to the additive nature of the ${\mathrm{PM}}_{2.5}$ concentration when aggregating water heating and cooking, DALYs were not additive due to the log-linear nature of the exposure–response curves.^109,110 As a result, the DALYs associated with “only cooking” and “water heating and cooking” were similar (e.g., mean of 377 and 387 per 10,000 people, respectively, for traditional wood stove in Uganda for all ages), suggesting that the additional exposure from boiling adds a minimal increment to the HAP impact. However, results in Figure 2 show that the relative increment (fractional increase in DALYs from adding boiling) was slightly larger for cleaner cooking options.

DALYs from Water Contamination

The total DALYs from drinking water for Uganda and Vietnam were similar, but a larger share of the DALYs came from children under 5 y in Uganda (23%–27% in comparison with 11%–14% in Vietnam, across different LRVs and source water values) because of the higher number of children per household in Uganda. DALYs from drinking water were a strong function of untreated water quality and boiling effectiveness (drinking water DALYs are provided in Excel Tables S12 and S13 for Uganda and Vietnam, respectively). For example, for 5-and-overs in Uganda and low-risk water, health impacts ranged from 51 DALYs per 10,000 people per year for lab-level boiling to 219 DALYs per 10,000 people per year from ineffective boiling, in comparison with 228 DALYs without boiling. For high-risk water for 5-and-overs, DALYs ranged from 59 per 10,000 people per year for lab-level boiling to 1,128 per 10,000 people per year for ineffective boiling, in comparison with 1,160 per 10,000 people per year without boiling. Using lab-level boiling in comparison with untreated water greatly reduced 5-and-over DALYs in both Uganda and Vietnam. For under-5s in Uganda, lab-level boiling in comparison with no boiling lowered DALYs by 94%, 89%, and 74% for high-, medium-, and low-risk water, respectively. The percent change in DALYs for under-5s in Vietnam were similar (94%, 78%, and 74% lower, respectively).

Comparison of Water and Air Pollution DALYs

In most of our scenarios, the benefits of clean water outweighed the impacts of HAP from boiling that water. Figure 3 reports the changes in DALYs for water and HAP exposures under various boiling and stove scenarios. The higher fraction of DALYs for children under 5 y of age in Uganda for both water and air reflected their higher proportion in the population. In addition, the relative contribution of the under-5s category was much greater for the HAP risk (driven by ALRI) than for water.

Figure 3.

Mean change in DALYs from boiling drinking water with different stove types for under-5s and 5-and-overs in Uganda (A) and Vietnam (B) for different stove types and use (assuming households either previously cooked and did not cook), and for different water source qualities and boiling effectiveness. The error bars show $\pm 1\text{\hspace{0.17em}SD}$ from Monte Carlo simulations. Simulation was run for 10,000 iterations of our Monte Carlo simulation, for a population of 10,000 people. The first set of bars show different water boiling categories. The number in parenthesis is the respective LRV, with ineffective Field LRV 0.048 (label “1”), Low field LRV 0.464 (label “2”), Average Field LRV 1.5 (label “3”), Best Field LRV 2 (label “4”), and Lab Study LRV 6 (label “5”). DALYs are separated into under-5s (label “$<5$”), and 5-and-overs (label “$>5$”) showing the different age groups. The second group of bars show the stove category, including LPG (label “A”); Gasifier (label “B”); Charcoal (label “C”); Wood (Improved) (label “D”); and Wood (Traditional) (label “E”). The keys with the corresponding letter labels are provided for the Water Boiling Categories, Stove Categories, and Age Categories. Data for figure provided in Excel Table S3 (Uganda) and Excel Table S4 (Vietnam). Note: DALYs, disability-adjusted life years; LPG, liquefied petroleum gas; LRV, log removal value; MPN, most probable number; ppl, people; SD, standard deviation.

All source water categories boiled with lab-level effectiveness ($\text{LRV}=6$) reduced the waterborne DALYs more than HAP DALYs increased for all stove types and stove use scenarios (water heating without cooking and water heating with cooking). For the scenario of a household that already cooked with this type of stove and then also started using the stove to boil water, we found that if the source water was boiled with at least the “low field” effectiveness (LRV of 0.5 or greater), the decrease in drinking water DALYs from boiling was greater than the increase in HAP DALYs (for all scenarios of water quality, water boiling, and stove types). In other words, we observed a net benefit for all water risk categories and all stove types for a household that used the same stove for cooking and for boiling water with at least low field effectiveness. If a household used the stove only for boiling water and did not cook, there was a net benefit if medium- or high-risk water was boiled with at least average field effectiveness for all stove types.

Log Removal Rates Needed for Health Benefits from Water Boiling

As another way to examine net benefits of boiling, the model was also used to determine “break-even” points, where increases in HAP DALYs were equal to the associated decrease in drinking water DALYs. Figure 4 shows the reduction in drinking water DALYs for 5-and-overs in Uganda plotted against LRVs for low-risk, medium-risk, and high-risk water qualities. The increase in HAP DALYs is also shown for two stoves, traditional and LPG, and for two different use scenarios, only water boiling and water boiling plus cooking. Where these two sets of curves cross can be considered the break-even point, or the minimum LRV required to provide a net reduction in DALYs for a given stove scenario. High-risk water had a significant decrease in DALYs even with low LRVs. In the case of 5-and-overs in Uganda, the intersection of the curves and HAP lines show that, in comparison with only cooking in a home, boiling drinking water with an LRV of 0.18 or greater resulted in a net reduction in DALYs for all stove types and source water types; this reduction was true even for the dirtiest stove (traditional wood stove) and low-risk water. In the case that a household was not cooking, but started boiling water, LRVs of 0.2, 0.3, and 1 were needed for high-, medium-, and low-risk water, respectively, to offset DALYs from HAP with a traditional wood stove. The LRV cutoff needed to achieve net health benefits increased as source water quality improved (for example, an LRV of 0.2 is needed for high-risk water I comparison with an LRV of 1 for low-risk water).

Figure 4.

The absolute value of net change in DALYs from boiling in water (reduction in Drinking Water DALYs and increase in HAP DALYs) vs. boiling LRV for homes in Uganda, considering 5-and-overs (population 5 y of age or older) only. The increases in DALYs, in comparison with baselines of both cooking and not cooking, are shown for LPG and traditional stoves. Water risk levels of low, medium, and high E. coli levels are shown. Inset shows a zoomed-in version of the large panel with same axes labels to show trade-offs at low LRVs. This plot presents the average value of the Monte Carlo simulation for each LRV. Water DALYs-High risk is labeled with “A,” Water DALYs-Medium Risk is labeled with “B,” Water DALYs-Low Risk is labeled with “C,” Traditional Stove Only Water Heating is labeled with “1,” Traditional Stove Cooking plus Water Heating is labeled with “2,” and LPG Only Water Heating is labeled with “3.” Data for figure are provided in Excel Table S5. Note: DALYs, disability-adjusted life years; HAP, household air pollution; LPG, liquified petroleum gas; LRV, log removal value.

Sensitivity Analysis

Estimated reductions in DALYs from water boiling were most sensitive to the following parameters (listed from greatest to lowest sensitivity): source water E. coli levels, the assumed ratios of Cryptosporidium, rotavirus, and Campylobacter to E. coli, age at death, dose–response parameters used in the QMRA, the LRVs, and the water volume ingested. For HAP DALYs, the input parameters with greatest influence (ranked from greatest to least) were stove emission factor, household air exchange rate, fuel heating value, and room volume. For the sensitivity of the QMRA dose–response parameters, we varied the parameters for each of the three pathogens and the level of risk per ingested dose. Of the three pathogens considered in the QMRA, rotavirus had the highest risk per dose of all the pathogens, and as a result, rotavirus exposure had the largest influence on QMRA DALYs (see Figures S2, S3, S4, and S5).

Discussion

In this study, we found that in most scenarios (even at low LRVs and with high-emitting cookstoves), boiling medium- and high-risk drinking water resulted in a net decrease in total DALYs. It was estimated that 1.1 billion people drink water that is of at least moderate risk,¹¹¹ so boiling would likely benefit these 1.1 billion people. In addition, though 89% of the world’s population uses an improved drinking source, such as household connections, public standpipes, boreholes, protected wells, springs, and rainwater collection,¹¹² and the odds of contamination are lower for so-called improved sources,¹¹³ 10% of improved sources are still considered high risk.¹¹¹ Thus, boiling could benefit even those using improved sources. However, in situations in which a high-emitting stove is used to boil water, and cooking is not already taking place in a home or cooking is done on a clean stove, boiling could have a net negative impact, showing the importance of considering risks from both water and air jointly.

Adoption of a new technology or practice is a major challenge for both water and stove interventions and will impact the generalizability of our model results. For example, in an intervention study in Rwanda, use of water filters and improved stoves was measured by self-report and spot-check observations, and though most household used the water filter, the majority continued to use traditional stoves.¹¹⁴ In our model, switching from traditional wood and improved stoves to LPG or electric stoves resulted in 99% reductions of ${\mathrm{PM}}_{2.5}$ and 96% reductions in the HAP DALYs, showing the huge potential for benefits from clean stoves, including for boiling. However, there are still many barriers to adopting LPG, including cost¹¹⁵ and social and cultural perceptions.^116,117 In addition, households often combine different stove and fuel types, known as stove stacking,^117–119 limiting the health benefits of a stove intervention. Traditional stoves continue to be used because of their additional benefits, including heating the living space, lighting the home, heating water for bathing and washing, drying, smoking food, getting rid of waste, keeping insects and animals away, and social gatherings.¹²⁰ This concomitant usage can be especially relevant for water boiling, because previous studies have observed that households continued the use of fuelwood to heat water after LPG and electricity are available,^120,121 possibly because of the cost¹²² of such an energy intensive task. Despite these challenges, the use of polluting stoves such as biomass has continued to decrease, from 53% using polluting stoves in 1990 to 36% in 2020.⁴ However, there have been increases in charcoal use in many areas,⁴ which our model suggests only moderately reduces ${\mathrm{PM}}_{2.5}$ relative to traditional wood use.

Our model assumed that all cooking and boiling occurs indoors. However, in some areas, when using certain fuels, boiling outdoors is common,¹²³ which would reduce HAP exposure from boiling water. In addition, our model assumed that cooking and water boiling happened at two distinct times, whereas the same fire could be used for both cooking and water boiling, offsetting some emissions. However, emission rates are based on fuel consumption,¹²⁴ which in either case is required to boil water and will result in PM emissions.

Boiling effectiveness, which varies widely, is key in determining whether boiling has a net health benefit. High adherence to a water treatment method has also been shown to be critical to realizing health benefits.¹¹⁸ None of the field studies of boiling used for this study observed “lab-quality” LRVs, and many field studies reported very low, even negative, LRVs for boiling.²³ Measured LRVs comparing posttreatment to untreated water (from the source) are influenced by source water quality before it is boiled, post-boiling storage, dipping, and other sources of recontamination.^22–24 Thus various factors can affect the end water quality despite treatment, limiting the potential benefits of boiling and resulting in a lower effective LRV. Our results showed that boiling is most beneficial in terms of a net reduction in scenarios with high-risk water; however, homes with high-risk drinking water may have the highest risk of recontamination due to poor sanitation practices.¹²⁵ Therefore, it is important to maintain improvements in water quality post boiling if water is not immediately consumed.^22,23,126

Benefits from increasing LRVs from different water treatment methods strongly depended on compliance of use¹²⁷ and full adoption of the intervention. It has been found that a few days of untreated water consumption after drinking treated water can completely negate the annual health benefits of drinking treated water.¹²⁸ The literature review conducted for this study on boiling LRVs suggested that societies and cultures with higher prevalence of boiling (e.g., Vietnam) have the highest LRVs.²² In countries like Zambia, where boiling is not prevalent²⁰ and boiling is promoted as an intervention, the LRVs tend to be much lower.²³ This finding suggests that introducing boiling as a water treatment method in a community that does not have a history of boiling presents additional challenges and likely reduces the chances of effective boiling and a net decrease in DALYs.²³

One notable alternative to boiling is chlorination, which is cheap and effective, provides residual protection,^129,130 and does not impact air quality. Chlorination has been shown to reduce the risk of child diarrhea and reduce risk of stored water contamination by E. coli.¹³¹ Residual protection could result in higher real-world LRVs. Unlike boiling, chlorination can be applied in several different ways, such as at the source and before drinking.¹³² However, there are several limitations related to chlorination, including challenges with the supply chain¹³³ and managing its use (including frequency and amount of added chlorine).¹³⁴ Both boiling and chlorination have trade-offs; however, our study suggests that the minimal HAP impacts associated with boiling are likely not the central factor in deciding whether boiling or chlorination is the best option for a household.

The cost of stoves and fuels were not considered in this study. However, cost is a major factor in whether a household adopts drinking water treatment¹³⁵ or improved stoves.^136,137 A cost–benefit analysis comparing boiling with various stove/fuel options could expand on our findings and add another dimension to the trade-offs we explore.

Strengths and Limitations

Our model is designed to give insights into possible trade-offs when addressing multiple environmental health risks. Combining two very different models meant we had integrated data from a variety of sources and years. We tried to incorporate the default or previously used values for the HAPIT and QMRA models when possible, but in some instances used different or newer data, such as the background disease data. Because the focus of this study was to develop a model framework applicable to many countries and scenarios with a focus on exploring health trade-offs rather than detailed contextual differences, this study used many of the same values (e.g., emission factors, fuel efficiency, water E. coli levels) for Uganda and Vietnam. Some inputs were country-specific, such as demographic parameters. Use of context-specific data on parameters such as stove type, source water quality, and boiling effectiveness, would enable a more accurate evaluation of risks. Some such data are available for our case study countries, such as stove emission factors^138,139 and source drinking water quality^140,141 specific to Vietnam and Uganda, water boiling effectiveness in Vietnam,²² and studies of water boiling habits in Uganda.¹⁴² Nonetheless, it is important to note that the focus of the modeling study was to compare risk trade-offs under different scenarios rather than to provide accurate risk estimates for the specific countries we modeled.

An additional limitation is that the QMRA model did not cover all possible waterborne disease outcomes and their sequelae. Specifically, malnutrition was not included in our analysis, which was shown to significantly increase the health burden from diarrhea.¹⁴³ Given that our models indicated a net reduction in DALYs from boiling under most scenarios, including additional outcomes associated with consumption of microbiologically contaminated drinking water would have further strengthened our conclusions.

In addition, some related past studies could be points of comparison, including studies of household air pollution levels in Vietnam¹⁴⁴ and Uganda.¹⁴⁵ Our study could be refined by using available data specific to Uganda and Vietnam. However, data for the same settings that cover the multiple dimensions in our analysis (e.g., boiling and air quality data for individual or nationally representative settings in a given country) are not currently available. Therefore, targeted field study of water boiling, pathogen levels in pre- and posttreated water, and HAP concentrations could be beneficial to evaluate boiling in a real-world setting.

Various assumptions used in our HAP modeling could be addressed with more specific input data or targeted measurements. For example, some of the stoves used in our model had poor efficiency, low stove power, and low fuel heating values, and thus they are challenged to meet a household’s water boiling needs. For these stoves, we assumed that the stove is used up to 24 h for the household’s cooking needs (as opposed to using multiple stoves for a shorter period). This assumption simplifies the model and avoids assumptions about stove use timing but is not realistic. The emission factors used for the stoves are averages, even though emission factors change throughout the combustion process.^96,97,99 However, limited data are available on phase-specific emission factors, which necessitated the use of averages. Emission rates were assumed to be the same for cooking and water heating for this study but are likely different, because it has been shown that varying cooking styles (e.g., frying vs. boiling) are associated with different emission rates.¹⁴⁶ However, lacking cooking- and boiling-specific emissions for Uganda and Vietnam, we used average values as a reasonable assumption. Another important simplification is our assumption that a single stove was used for all household uses, ignoring stove stacking.^117,147

The box model used to calculate 24-h ${\mathrm{PM}}_{2.5}$ has been found to overestimate concentrations,^79,148,149 and so modeled exposures are likely biased high. Depending on where the actual ${\mathrm{PM}}_{2.5}$ exposure falls on the exposure–response curve, this characteristic of the box model means we could be over- or underestimating the net change in DALYs due to the added activity for boiling. For example, if the ${\mathrm{PM}}_{2.5}$ with and without boiling falls above where the curves “plateau,” the difference in DALYs we estimated could be smaller than in reality. Our modeled values can be compared with past published field measurements. We estimated mean 24-h ${\mathrm{PM}}_{2.5}$ concentrations ranging from $\mathrm{1,797}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ to $\mathrm{5,212}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^{-3}$, $721{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ to $\mathrm{2,068}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$, and $14{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ to $16{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for traditional wood, improved wood, and LPG stove use cases, respectively. In a cross-sectional study in India in four states, the measured mean 24-h ${\mathrm{PM}}_{2.5}$ concentrations were $590{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ in the kitchen for a traditional wood stove and $179{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ in the kitchen for LPG.¹⁵⁰ In another study in India of an improved stoves intervention, 24-h ${\mathrm{PM}}_{2.5}$ for LPG stoves ranged from 70 to $103{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ and was around $500{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for households using a traditional three-stone fire.¹⁵¹ Mean personal exposure concentrations measured in the HAPIN trial, where high LPG uptake was observed in intervention households, were $71.5{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ and $24.1{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ before and after the intervention, respectively.¹⁵² A study in Nepal found 24-h ${\mathrm{PM}}_{2.5}$ in kitchens to be $80{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for electric stoves and $630{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for traditional mud wood stoves.¹⁵³ These findings suggest that we both overestimate the impacts of traditional cooking and the benefits of “realistic” clean stove adoption in many settings. For example, our model calculated substantially lower values for LPG-using households in comparison with most field studies, suggesting we may have overestimated the benefits of LPG use. Our modeled values for traditional wood stoves were higher than those in the field studies. Our estimated 24-h ${\mathrm{PM}}_{2.5}$ concentrations for improved biomass stoves were similar to those observed for traditional wood stoves in several field studies, suggesting a bias in model parameters used as defaults. The bias is potentially at least partly due to stove stacking in the field studies, because many study households used multiple stoves.¹⁵³

The observation that modeled ${\mathrm{PM}}_{2.5}$ values for traditional stoves were much higher than typically observed in field studies suggests that assumptions in the HAP model (as used in global assessments) lead to estimates substantially higher than reality. However, in our application, such a high bias in model results is likely conservative (overestimates HAP impacts when boiling with traditional biomass) and further emphasizes our finding that water boiling typically has a net benefit, even if using a very high-emitting stove. The relative differences we found between exposures associated with different stoves was fairly representative of real-world observations. For example, a 48% reduction in cooking-area ${\mathrm{PM}}_{2.5}$ was measured during an improved wood stove intervention study in Rwanda,¹¹⁴ relative to our average modeled reductions of 62% for ${\mathrm{PM}}_{2.5}$ concentrations for a similar scenario ($\mathrm{3,490}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ and $\mathrm{1,334}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for traditional and improved stoves, respectively). However, as acknowledged above, the nonlinearity of the ${\mathrm{PM}}_{2.5}$ exposure–response curves means that, depending on the pre- and postboiling combination, the modeling could over- or underestimate the DALYs associated with the addition of boiling. Future work can explore modeled DALY estimations further using a recalibrated model or representative field observations.

Our results can also be compared to previous indoor air quality modeling studies. The study presenting the household air pollution box model used here estimated kitchen concentrations during cooking, with 24-h averages ranging from $15{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for LPG to $\mathrm{1,975}{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for traditional stoves.⁷⁹ The estimated ${\mathrm{PM}}_{2.5}$ concentrations for wood stoves from our study using the model are higher due to different emission factor, stove power, and thermal efficiency inputs for the traditional stoves. A study in China used this air pollution box model in a more sophisticated way with substantially more household data to refine assumptions. This study used inputs from past field studies in China, assigned a ventilation index based on observations of cooking-area ventilation (presumably via a chimney or other active ventilation) and information on cooking and living-area location to scale the model air exchange rate. This study reported average modeled kitchen ${\mathrm{PM}}_{2.5}$ concentrations of $79{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for households using biomass to boil water, with the three homes with poor ventilation having an average of $148{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$.²⁶ Our estimated ${\mathrm{PM}}_{2.5}$ from water heating was significantly higher, likely due to the inclusion of a ventilation parameter and stove emission factors specific to China,^26,154 so a direct comparison is likely not warranted. However, this study provides an example of how HAP modeling can be refined for future studies of specific locations.

Our choice of integrated exposure response (IER) is also a limitation. We used the Burnett et al. IER functions¹⁰⁵ because that is what the HAPIT model uses.⁴⁹ This choice is a limitation because the updated meta-regression-Bayesian, regularized, trimmed (MR-BRT) curves were used in more recent assessments.¹³ Although these more recent exposure–response curves are available, we elected to use the IER curves for consistency with the HAPIT tool and the many existing studies using them to allow more direct comparison with our findings.

Our results can also be compared with direct epidemiological studies. A systematic review of cooking with gas or electricity instead of solid fuels found that cooking with gas instead of polluting fuels lowered the risk of ALRI or pneumonia by 46%, whereas cooking with electricity instead of gas further decreased the ALRI risk by 26%.¹⁵⁵ We estimated a much higher reduction in DALYs in under-5s (over 95%) switching from the polluting wood stove to the gas stove or electric stoves or all three of our scenarios in Uganda (cooking only: 98% and 99%; cooking plus water boiling: 97% and 99%; water boiling only: 98% and 99%). The ranges for adults were similar. Other modeling studies have estimated significant reductions in indoor air pollutants and correspondingly extensive health benefits.^156,157 For example in Mozambique, a modeling study estimated that a rural natural draft stove reaching just 10% of households would avert 200 premature deaths and 14,000 DALYs.¹⁵⁷ A modeling study of Cameroon included a scenario in which 58% of the households adopted LPG, which averted 28,000 DALYs.¹⁵⁶ Unfortunately, field studies studying health impacts from gas stoves and other HAP stove interventions have frequently observed limited improvements in health.¹⁵⁸ For example, one study saw no improvement in severe childhood pneumonia with gas stove use, and the air pollution exposures remained above WHO health-based target values.¹⁵⁹ Central limitations of our study are our assumption of full adoption (100% use) of the “intervention” stove (improved biomass, gasifier, gas, and electric), thus no stove stacking, and not including exposures from other sources and high ambient air pollution levels. An additional limitation is that a significant portion of ambient ${\mathrm{PM}}_{2.5}$ exposure can be attributed to emissions from household solid fuel uses,^160,161 which is not accounted for in our model; therefore, our estimated health impacts from burning solid fuels might be underestimated. Further modeling efforts could attempt to incorporate these factors into risk estimates.

We compared our drinking water estimates with results from two epidemiological studies. Cohen et al. found risk ratios of 0.61 [95% confidence interval (CI): 0.46, 0.80] for boiling in an electric kettle, and 0.40 (95% CI: 0.08, 2.14) for boiling in a pot,¹⁶² corresponding to a 39% and 60% reduction in reported diarrhea, respectively. The source water in this study varied, with approximately half of those boiling water using an improved water source, though improved water sources may still be contaminated. In another study, Psutka et al. found (nonstatistically significant) relative risks of 2.38 (95% CI: 0.75, 7.54) and 1.26 (95% CI: 0.68, 2.31) for a household not boiling water and a child having diarrhea sometime in the past 7 d and past month, respectively. These correspond to relative risks of 0.42 (95% CI: 0.13, 1.33) and 0.79 (95% CI: 0.43, 1.47), or a 58% and 21% reduction in the risk of diarrhea, for households reporting boiling water.²³ In the second study, nearly all of the source water was contaminated.¹⁶³ In our study, for under-5s in Uganda for moderately effective boiling, we estimated mean reductions of 56% to 72% in DALYs depending on source water quality; for low-effectiveness boiling, we estimated reductions of 29% to 35%. Therefore, comparisons of DALY reductions from our study and reductions in diarrhea risk in these epidemiological studies suggest that our estimates based on moderate- and low-effectiveness boiling scenarios (the likely conditions for population-based epidemiological studies) are relatively consistent with empirically measured health benefits from boiling.

Environmental factors outside the home may also have an important influence on HAP not addressed in our model. For example, although ambient air pollution levels in LMICs vary widely across countries and between urban and rural areas,^109,160 for simplicity we assumed a fixed background ambient ${\mathrm{PM}}_{2.5}$ concentration; further analysis could vary this. If actual ambient air pollution levels are higher, ${\mathrm{PM}}_{2.5}$ exposure levels will be shifted higher on the exposure–response curves, resulting in a smaller additional risk from the increment of ${\mathrm{PM}}_{2.5}$ from boiling water. We also considered only cooking and water heating and did not consider space heating, even though many countries in the world where boiling is common are located in cold regions.^18,20 In addition, it is assumed that boiling is only used for drinking water, not for making tea or bathing. However, if a household is burning fuel for space heating or other reasons and boiling occurs simultaneously, no additional HAP DALYs would be incurred to boil water.

Several simplifying assumptions made in the water module also influence our results. For example, though we separated by population over and under the age of 5 y, we assumed a constant probability of death for all ages. However, this approach likely leads to an underestimation of the years of life lost due to death from diarrhea because children have the greatest risk of death and disease from diarrheal diseases.¹⁶⁴ Children do not have fully developed immune systems, so they are more susceptible to diarrheal illnesses, which is not accounted for in the dose–response curves used in the modeling. Therefore, if child-specific dose–response curves were available and used, the benefits of boiling would likely be greater.

We estimated pathogen levels based on fecal indicator bacteria (i.e., E. coli) levels, which are imperfectly correlated with the selected pathogens.¹⁶⁵ Further, pathogens present in drinking water vary widely by country and across seasons.¹⁶⁶ We expect that the results would vary widely based on the pathogens selected, as has been found in previous studies.¹⁶⁷ Further, the risks associated with exposure to rotavirus would be dependent on the vaccination status of the population.¹⁶⁸ However, even in a population fully vaccinated against rotavirus, the DALYs associated with rotavirus in our model are intended to represent illness from a variety of enteric viruses. Further, rotavirus prevalence as high as 40% has been documented in untreated groundwater used for drinking.¹⁶⁹ Our results were highly sensitive to parameters related to pathogen risk. Thus, pathogen-specific water quality data would be helpful to characterize risks associated with pre- and postboiling water. Limited data exist, but new methods (with their own associated limitations) are expanding the potential for pathogen-specific data, such as the use of TaqMan array cards for environmental samples.¹⁷⁰ In addition, water quality in LMICs can be highly temporally variable, with spikes in contamination, and intermittent exposure to contaminated drinking water can negate the benefits of improved water access.

DALYs are used for comparative risk assessment, including environmental risks such as HAP and water contamination. DALYs allow comparison across risk categories and are the common metric used in the Global Burden of Disease estimates.¹⁷¹ Estimates for DALYs from air pollution and water are calculated using very different approaches, which we adopted for this study. In addition, though DALYs account for the years of life lost and years lived with disability, they do not perfectly capture the acute vs. chronic nature of disease, which is an important difference between HAP and water risks. Using these different methods to compare different risks and outcomes is an inherent limitation in the comparative risk assessment approach. However, we took this approach because it is the most widely used to compare diverse risk factors. Future work can examine other methods to calculate and compare the risks from drinking water and HAP, including field studies to provide context-specific findings to compare against the broad insights from this scoping analysis.

Conclusions

To our knowledge, this is the first study to compare water and air pollution DALYs for scenarios of household cooking and water boiling to determine whether boiling water has a net benefit regardless of stove and fuel type. Our model identified certain scenarios that, when risks for water and air are considered separately, could suggest DALY reductions but that due to combined effects ultimately result in net positive DALYs. For example, if a household that does not cook their food is encouraged to boil water on a wood stove, they would experience increased air pollution exposure that may outweigh the benefits of ineffectual water boiling. However, this scenario is likely an exception, and our modeling suggests that boiling even at low log removal rates has net health benefits for medium- and high-risk source water, even if using a relatively high-emitting stove. In contrast, when treating low-risk source water, boiling effectiveness and cookstove type influence whether net health benefits result from the practice. We recommend further field investigations to jointly assess health effects associated with contaminated drinking water and HAP and the development and evaluation of interventions to mitigate both exposures. Because of uncertainties and assumptions in the model, we recommend that country- or context-specific inputs be used and ideally studies of boiling and household air pollution with empirical measurements of water and air quality be conducted to improve understanding of the health risks and trade-offs and evaluate our model. Specifically, more information on household-specific parameters, including pathogen contamination in water, stove types, cooking and boiling energy requirements, and parameters linking emissions to HAP and exposure could improve the model. Future efforts should empirically measure air pollution concentrations (to estimate health impacts) and diarrhea during field studies of different boiling methods and stove types to both provide better input parameters and improve our ability to model risk trade-offs.

Conclusions and opinions are those of the individual authors and do not necessarily reflect the policies or views of EHP Publishing or the National Institute of Environmental Health Sciences.