Study locations and context
Ethiopia is situated in the Northeastern part of Africa and occupies an area of 1.1 million square kilometers ranging from 4620 m above sea level to 148 m below sea level [40]. The country possesses three major topographic-induced climatic zones, the hot lowlands (Kolla) located 1500 below, the temperate (Wayna Dega) which range 1500–2400, and the cool temperate highlands (Dega) located above 2400 m above sea level [40, 41].
The average annual temperature is approximately 15–20 and 25–30 °C for highlands and lowlands, respectively [41]. This trial was conducted in a low-income rural community of the Mecha Health and Demographic Surveillance System (MHDSS) site in Northwest Ethiopia. MHDSS site is a field research center established in 2013 by Bahir Dar University to carry out and support postgraduate level studies in the region. It is located 525 km away from the capital city of Ethiopia, Addis Ababa, towards Northwest and 40 km far away from the capital city of Amhara Regional State, Bahir Dar. According to the official population profile report of MHDSS, the study area comprises 132 clusters/Gots with a total of 65,086 populations within 20631 houses at the end of 2016.
Our earlier research work, carried out in the current study area, have also shown that all households use biomass fuel as the primary household energy source and the households in the included clusters are relatively homogenous in terms of basic socio-demographic and cooking-related characteristics [13, 19], which made them ideal populations for rationale comparison of the study groups in the current trial study. Furthermore, the presence of extra indoor (95.8%) and outdoor (38.1%) burning events such as coffee ceremony, burning incense, local alcohol/areqi making, burning rubbish, and charcoal production were common observable facts in the study area [13]. About 63% of the households use a separate kitchen, and most (89.1%) houses are owned privately [19].
Study design
As part of the wider stove trial project in Northwest Ethiopia (ClinicalTrials.gov Identifier: NCT03612362), a community-level cluster randomized controlled trial study with two arms of equal size was used to assess the effect of biomass-fuelled ICS intervention on the concentration of HAP compared with the continuation of an open burning TCS method. Cluster is a small village, termed as Got in Amharic (both national and local language), is the lowest administrative unit in Ethiopia, and is used as the smallest unit of enumeration areas by the Ethiopian national census authority. All eligible households in the selected clusters were enrolled as control or intervention for baseline and repeated follow-up visits approximately every 3 months for 1 year after receiving the intervention. The concentration of HAP at the individual household level was measured before the installation of Mirt ICS and, again in the same households, 4 times after the intervention households received the ICS intervention. The households with the TCS method were served as a control arm.
Eligibility criteria
All clusters/Gots and households under the MHDSS site were eligible for participation in the cookstove trial, and all households who were exclusive users of TCS for injera baking were eligible for participation in the trial. Only households who did not have any enclosed cooking quarter (kitchen) arrangement were excluded.
Sample size determination
To estimate the effect of ICS intervention on HAP over the follow-up period, the sample size was calculated based on previous publication [42] by considering a detectable difference of 30% in HAP concentration reduction by the ICS intervention to be worth pursuing, a two-sided alpha of 0.05, a power of 80%, and the common coefficient of variation (CoV) value of 0.7 for HAP measurement outcome in biomass fuel using households [42]. Accordingly, the estimated sample size (n) was found to be 171 households in each arm assuming individual randomization.
However, since this trial randomized the intervention over clusters instead of individual houses, the standard formulae for estimation of sample size might lead to an underpowered study which may be inconclusive. Thus, the calculated sample size assuming the individual randomization was inflated by a design effect (DE) value to reach the required level of statistical power under cluster randomization using the formula [43]:
$$ \mathrm{DE}=1+\kern0.5em \left[\left({\mathrm{CoV}}^2+1\right)\mathrm{m}'-1\right]\ \mathrm{ICC}. $$
Considering a CoV value of 0.2 for cluster size and an average number of eligible houses of 55 from the updated data of MHDSS and an intra-cluster correlation coefficient (ICC) value of 0.05 for HAP to cope with the unknown ICC was suggested by previous reviews of ICC values [44, 45]. Accordingly, the DE value became about 4, and the required sample size for the cluster randomized controlled trial (nc) became about 684 households per arm using the formula:
$$ {\mathrm{n}}_{\mathrm{c}}=\mathrm{n}\kern0.5em \left[1+\left(\left({\mathrm{CoV}}^2+1\right)\ \mathrm{m}'-1\right)\ \mathrm{ICC}\right]. $$
Then, with an additional 30% to account for any unpredictable events in the field due to equipment-related problems as well as for any lost to follow-up (LTF) events such as unexpected change of cooking behavior of the participant houses [42, 46], the required sample size becomes about 978 houses per arm. Finally, the number of clusters (K) required in each arm for unequal cluster sizes was determined using the formula [47]:
$$ \mathrm{K}=\mathrm{n}\kern0.5em \left[1+\left(\left({\mathrm{CoV}}^2+1\right)\mathrm{m}'-1\right)\ \left.\mathrm{ICC}\right]/\mathrm{m}'\right] $$
, which became about 18 clusters per arm, and this caused to increase the sample size to 990 houses per arm.
Randomization and masking
Clusters were randomly allocated to intervention and control arms at a 1:1 ratio by an independent epidemiologist using a computer-generated randomization schedule. Intervention status was revealed after all baseline measurements had been completed as well as all study households recruited and assigned to their respective arm to ensure the allocation sequence was concealed from those assigning the arms. Also, participating households and data collectors were blinded to intervention status during study enrollment and baseline data collection. All eligible households within the clusters were included in the study to minimize the risk of selection bias; however, because of the typical feature of cluster design and nature of the intervention under study, blinding of the households receiving ICS intervention was not possible.
The major rationale for adopting a cluster randomized trial design was to prevent contamination [48] or unintentional spill-over of intervention effects from one treatment group to another of the trial if individual household randomization was used, as the concentration of HAP would inevitably be affected by the air pollution status from neighbor households. The other rationales were to increase administrative effectiveness and minimize costs [49].
Sampling and recruitment of households
The cluster sampling method was used to select 36 clusters randomly (18 clusters per arm) among the total 132 clusters in the MHDSS site, and all eligible households were included within the selected cluster (complete enumeration). The list of clusters/Gots and households was established from the MHDSS record, the selected households were identified using the permanent MHDSS site house number, and the actual participant households were recruited at the household level by field workers during the baseline survey after ensuring whether the households met the eligibility criteria.
A screening questionnaire was used by field data collectors upon their first visit to each household to ensure that the household was appropriate and willing to participate. When the household met the eligibility criteria, the study was explained to the heads of the household, and they were asked whether the household would be willing to participate in the study and use ICS technology for at least 12 months. Then, when the heads of the household agreed to be involved in the study, the field staff administered a written consent form at that time, and the consent procedure was conducted in Amharic (both national and local language). To achieve adequate participant enrolment, we utilized local energy experts and health extension workers to oversee the overall efforts in recruiting eligible houses.
Intervention
Trial descriptions and implementation
In general, there are about six primary types of biomass-fuelled improved cookstoves [50]. These are as follows:
-
1.
Rocket (also known as side-fed) cookstoves: these are fuelled with wood sticks or biomass residues that are continuously fed through the side of the stove, typically resting on a grate so that ash and charcoal can settle below. Air enters by natural or forced-draft through the same opening as the fuel (examples: Grameen Greenway Smartstove, Envirofit G-3300)
-
2.
Gasifier cookstoves: these are batch-or-continuously fed using processed fuel (examples: Awamu Troika, Mimi Moto, and Philips ACE 1)
-
3.
Charcoal cookstoves: these are batch-operated and fuelled with charcoal or carbonized biomass (examples: Kenyan Ceramic Jiko, Envirofit CH-2200, and Burn Jikokoa)
-
4.
Forced-draft/fan cookstoves: these cookstoves have air that is forced into the stove using a fan or a blower to enhance turbulence and promote cleaner combustion (example: BioLite HomeStove)
-
5.
Batch-operated cookstoves: these stove types are operated on a single load of fuel at a time
-
6.
Continuously fed cookstoves: these stoves require fuel to be loaded throughout the cooking process
In the Ethiopian context, two different household cooking devices are required traditionally. One stove is for baking the staple food of Ethiopia called injera, which is a unique type of yeast-risen flatbread, consumed widely in Ethiopia [38, 39], and another for other cooking purposes used nearly on a daily basis [13]. Replacing the open burning TCS method with an ICS method, locally called Mirt (best) improved cookstove (Fig. 1), was the intervention for this study which is the well-known commercially distributed type of ICS in Ethiopia for injera baking (i.e., the staple food of Ethiopia) [38, 39].
The Mirt is made of cement and volcanic ash and is an unvented ICS type designed by the Ethiopian Energy Studies Research to be used only for cooking injera in Ethiopia with a life span of at least 5 years [38, 39]. It is big in size and a fixed stove type which requires firewood to be loaded all over the cooking process [39]. It can save the quantity of fuel up to 31% compared to the TCS method [51] through efficient energy conversion in an enclosed combustion chamber which can also decrease PM2.5 emission up to 50% in μg/m3 [38, 39].
Households who were randomized to the intervention arm were identified using the permanent MHDSS house number for convenient appointment date, and the intervention was delivered at the beginning of the study period to all eligible households allocated in the intervention arm. However, since the firewood is often self-collected and inexpensive, fuel was procured by the recipients of this trial in both arms. Also, the fuel requirement was practically planned to be attained by every recipient household in the Ethiopian Mirt ICS implementation program.
All the trial stoves were manufactured by a local licensed firm and installed on-site by the installation teams. Demonstration in the use of the Mirt ICS was provided to each household during the time of installation, and the intervention was promoted regularly throughout the follow-up period by the local energy experts’ team of ICS monitors. Correspondingly, the control households were continued to use the usual open burning TCS method in an equal number of randomly allocated clusters.
Since the life span of Mirt ICS is about 5 years [38, 39], the length of both the intervention and the follow-up period was 1 year to safely account for seasonal factors that have a major effect on the magnitude of HAP as well as to maintain a balance between achieving sufficiently long follow-up period for HAP outcome measurement and short follow-up period to decrease attrition.
Trial adherence and compliance monitoring
Adherence of study households’ to the trial protocol was assessed through self-reporting and direct observation by trained field workers along with the local energy expert team. At each follow-up visit, the field workers observed and recorded the type and condition of the cookstove currently being used (i.e., no stove change or no observed breakage resulting in no use). Additionally, the primary cook was asked whether the cookstove intervention was in good working order (i.e., no reported breakage resulting in no use). Only trivial maintenance problems were detected regarding protocol adherence, and timely responses were carried out by the installation teams to avoid the possible detrimental effect of non-adherence through improving intervention protocol adherence. Besides, trial protocol compliance was checked by the local energy experts’ team of stove monitors’ through unannounced visual inspection visits in homes of both arms to enhance data validity.
Participant retention strategies
Once the eligible households were enrolled, a variety of strategies was used to avoid premature withdrawal of participants and the associated complexities in the analysis and interpretation of findings due to missing data. In this regard, active community engagement was established through the Ethiopian health extension program and local health development army team structure to promote participant retention and complete follow-up for the entire study period. Interest in the study was maintained through periodic communications about the intervention protocol adherence during the regular local health development army team meetings and throughout the home visits by field health workers as well as by the local energy experts.
Also, HAP measurement events were scheduled at a regular appointment of home visits to limit the participants’ burden related to follow-up visits, and at the start of the trial, control households were informed that they would receive the ICS intervention at the end of the study period to maintain justice and achieve a high level of post-recruitment participant retention.
Trial safety monitoring
The Mirt ICS intervention [38, 39], which was tested by this trial, was not involved in any drug or a medical procedure as well as not known to increase the risk of any adverse event. Nevertheless, an interim analysis was included in the protocol for safety and efficacy monitoring.
On the other hand, the Mirt ICS intervention [38, 39] was expected to reduce HAP-related health effects [39] and we expected participation in the intervention arm might reduce risks to the participant. Thus, any adverse events data deemed related to the trial intervention were collected and reported immediately during the routine household visits to take appropriate management as well as to inform the conduct of the ongoing and future studies. The collected data were also reviewed for safety by an independent Data Safety and Monitoring Board (DSMB) to determine whether there were grounds to stop the trial early for adverse events. Nevertheless, the Board found no grounds to stop the trial early due to adverse events.
Household air pollution outcome assessment
Fine particulate matter with a diameter less than 2.5 μm in diameter (PM2.5) is a key pollutant associated with both health and climatic impacts [52], and the latest WHO indoor air quality guideline uses PM2.5 concentration as a key household air pollutant [20, 53]. For this reason, the status of HAP was determined by measuring the concentration of indoor PM2.5 using a low-cost, light-scattering particulate matter monitoring device called Dylos DC1700 air quality monitor [54, 55].
The performance of the Dylos DC1700 monitor has been previously evaluated for different scenarios in both indoor and outdoor environments [54, 56,57,58], and it was also found to perform well at both rural and urban locations for measuring PM2.5 concentration [54, 55, 59,60,61]. The monitor has been also utilized as a reference instrument to calibrate low-cost PM2.5 sensors as reported in previous studies [62], and its performance did not seem to have been impacted by aerosol composition [63], relative humidity [64], and temperature [65]. Moreover, this monitoring device is cost-effective, portable, both electric power cable and battery operated, quieter, and easier to use, and does not need laboratory facilities [54, 56, 57].
The Dylos light-scattering monitors were calibrated under actual conditions of deployment by conducting co-located PM2.5 concentration in μg/m3 measurements in randomly selected sub-sample of homes (one household per cluster) using a light-scattering monitor which allowed PM2.5 mass concentration measurements in μg/m3 following adjustment by a gravimetric method to obtain a local calibration factor as a reference to convert and correct the light-scattering (photometric) measurements.
Data collection
At baseline, continuous indoor PM2.5 concentration monitoring was performed among 2031 households for one cooking hour using digital Dylos DC1700 monitors by trained environmental health officers after undergoing a 2-day HAP monitoring training. The training included how to operate the Dylos monitor, how to monitor indoor PM2.5 concentration using Dylos sampling protocol, how to record the acquired indoor PM2.5 concentration data, and how to apply standard operating procedures (SOPs). Practical exercises were carried out, and all monitors were tested as part of the training. Two senior environmental health professionals were assigned to supervise the entire HAP monitoring activity, and the overall coordination was handled by the investigators of the research project.
To measure indoor PM2.5 concentration, the Dylos monitors were placed in the main cooking quarter (kitchen) at least 1 m away from the edge of the stove, at a height of 1.5 m above the floor, 1.5 m away from doors, windows, and other openings horizontally [36], and at a safe location to minimize the risk of interrupting normal household activities or being disturbed. Similarly, co-located PM2.5 concentration measurements were also conducted for the one cooking hour in a randomly selected sub-sample of 36 households using an adjusted light-scattering monitor which allowed PM2.5 mass concentration measurements in μg/m3 of sampled air.
The HAP monitoring team members have recorded the sampling date, household ID, starting and completing time of monitoring, and the indoor PM2.5 concentration. The PM2.5 concentration data were also downloaded to a PC file by supervisors using the Dylos Logger software at the end of each monitoring day and sent to the principal investigator together with other daily records for cross-checking. Independent variables data were also collected through direct observations and face-to-face interviews using a structured questionnaire on main cooking area characteristics such as the location of the main cooking quarter, cookstove type, and frequency of the cooking event.
Subsequently, after the baseline survey and implementation of the intervention, a series of micro-environmental indoor PM2.5 concentration measurements were carried out by the trained environmental health officers using similar measuring devices and protocol for 1 year every 3-month interval in both arms. If houses were unavailable during the scheduled visit, repeated visits were made on any day of the same week. The duration of the follow-up period was determined to be 1 year to account for seasonal factors that might have a major effect on the magnitude of micro-environmental HAP.
Data quality assurance
The pragmatic approach that we followed ensures the generalizability of the study findings to the wider population by maintaining the validity and reliability of findings. In this regard, a variety of measures were also taken to ensure data quality. To begin with the outcome variable, indoor PM2.5 concentrations were assessed in the same manner in both arms by similarly trained environmental health officers using calibrated monitors under actual conditions of deployment. Equal numbers of intervention and control households were visited every morning and afternoon in each HAP monitoring day.
The HAP monitoring team was in regular contact with the investigators of the research project with scheduled meetings, and additional communications were done as needed for feedback and quality control. To minimize the risk of bias, a specific monitoring protocol containing a detailed description of the standard operating procedures (SOPs) was used to reduce the level of error associated with PM2.5 concentration monitoring and other data collection through assuring consistency in measurements. Clusters were randomly allocated to intervention and control arms, all eligible households within the clusters were included in the study, allocation sequence was concealed from those assigning participant households to arms, and primary outcome assessors were blind to the intervention at baseline. A single licensed firm manufactured all the trial ICSs, and the same installation teams administered the intervention in both arms.
In addition, all initially randomized participants were analyzed in the arm they were assigned to (i.e., intention-to-treat analysis principle). Furthermore, the methodological soundness such as the large sample size that took ICC value into account, longitudinal study design, and baseline data collection on the primary outcome and related risk factors to be adjusted through Generalized Estimation Equation (GEE) modeling can help us to achieve an effective balance of potential confounders between both arms. Finally, this manuscript was reported following both the guidelines of Consolidated Standards of Reporting Trials (CONSORT) 2010 statement extension to cluster randomized trials and Template for Intervention Description and Replication (TIDieR) checklist [66] to address the essential study design components and intervention aspects of this trial report.
Statistical analysis methods
Baseline cooking characteristics and all Dylos indoor PM2.5 concentration data in 0.01 ft3 were entered into the Statistical Package for Social Sciences (SPSS) for analysis. The Dylos DC1700 monitor measures fine particles in two size ranges. These are the small range channel, which measures 0.5-μm particulates or greater, and the large range channel, which measures particulates of 2.5 μm or greater [54, 55]. Thus, the Dylos HAP concentration data was easily calculated by subtracting the large channel value from the small channel to find the Dylos indoor PM2.5 concentration data in 0.01 ft3.
Then, linear regression analysis was performed between the Dylos HAP concentration data in 0.01 ft3 and the co-located monitor which allowed PM2.5 concentration measurements in μg/m3 of sampled air [60, 63]. The analysis result showed a strong linear relationship with a conversion factor of PM2.5 concentration in μg/m3 = [(6.22) (Dylos PM2.5 concentrations in 0.01 ft3) (10−2)]. This provides a conversion factor of 6.22, and the resulting conversion factor was used to convert the Dylos indoor PM2.5 concentration data in 0.01 ft3 to equivalent PM2.5 concentration data in μg/m3 to make the Dylos results comparable with other HAP monitors and to follow the WHO standard of measuring HAP in units of mass per volume in μg/m3.
To quantify the magnitude of clustering for HAP outcome at baseline, cluster-level ICC value was also calculated using a multilevel mixed-effects model (i.e., mixed-effects linear regression estimation method in STATA), which directly estimates between and within-cluster variances to calculate ICC for continuous variables. Using this method, the cluster-level ICC value for indoor PM2.5 concentration in μg/m3 was found to be 0.0346, which indicates that only 3.46% of the total variability in indoor PM2.5 concentration is explained by the between cluster-level variation, showing the fact that group-level characteristics are not required to explain the outcome variable. Therefore, we considered the individual households as the unit of analysis and interpretation [67] in determining the effect of ICS intervention on the longitudinal indoor PM2.5 concentration in μg/m3 compared with the continuation of the open burning TCS method.
The effect of ICS intervention on the repeated response of indoor PM2.5 concentration between the two arms was estimated using linear regression with GEE modeling approach among the intention-to-treat (ITT) households. The GEE analysis method is the ideal method for longitudinal data analysis due to its computational simplicity and robustness to misspecification of the repeated measures’ correlation structure. Although the GEE method is understood to be robust against a wrong choice of working correlation structure (WCS), the best WCS of the outcome variable was chosen through a critical examination of the observed correlations between subsequent measurements to get a more precise estimation of the intervention effect [68]. Using this method, an exchangeable correlation matrix was found to be most appropriate to fit the observed data. Quasi-likelihood under the independence model criterion (QIC) technique was also employed to uphold the goodness of model fitness by choosing a model with a smaller QIC value.
As a final point, our GEE analysis model has simultaneously included a continuous outcome variable of repeatedly measured indoor PM2.5 concentration μg/m3 with a binary indicator of treatment allocation (i.e., control versus intervention) as well as other indicator variables such baseline indoor PM2.5 concentration, location of cooking quarter, secondary cookstove type used for other cooking purposes, and frequency of injera baking events measured at baseline.