Evidence on clean cooking subsidies for exclusive use, exposure, and affordability from a randomized control trial in rural Tanzania

In the first study arm, participants will receive an unconditional cash transfer. Using data from their baseline survey (household size and estimated monthly expenditure), we will estimate the cost for the participant to exclusively use LPG. Then, we will calculate the necessary UCT to ensure that exclusive use would cost the participant ~5% of her monthly expenditure. We rely on the main participant (94% women)’s expenditure, rather than household as often considered, as our previous work in this field site revealed that the woman was entirely responsible for purchasing cooking fuel from her income/stipends.

The second study arm is the subsidy treatment group in which participants will receive vouchers to obtain discounted LPG. To determine the subsidy, we will calculate what the participant would have received as a UCT, that is the monthly amount needed to make sure exclusively LPG use would only cost 5% of expenditure. We will then discount each 6kg cylinder by a percentage so that the overall subsidy over the course of each month (if used) is equal to the UCT. That is, for each participant the UCT and voucher amounts would be of equal value on a monthly basis regardless of which group they are assigned. The values are not equal across respondents, given that different household sizes and expenditures will affect both treatment arms’ approaches. Subsidy amounts are participant specific.

The final study arm is the control group. They will not receive a UCT or a fuels subsidy; however, at the end of the study, they will receive an equivalent package of household items. We will pilot in Shirati to understand what package of goods would be culturally appropriate.

2026-12-01

2027-12-01

Our key outcome variables are total annual LPG refills, biomass use, exclusive LPG use, and 24-hr exposures to particulate matter 2.5 (PM2.5), carbon monoxide (CO), black carbon (BC), and nitrogen dioxide (NO2).

Total annual LPG refills will be constructed from continuous LPG purchase data (receipts recording the date and cylinder size of each exchange) collected from both respondents and LPG providers in Shirati.

Biomass use comprises two continuous variables: household consumption of firewood and of charcoal. At bi-weekly household surveys, we will ask participants to report their consumption in standard, purchasable units (firewood bundles; charcoal in black sacks, buckets, or large sacks). We will weigh typical examples of each unit and triangulate these weights with household-reported cost data to estimate the average kilogram weight corresponding to a given monetary amount of firewood or charcoal.

Exclusive LPG use will be constructed as a binary variable, drawing on LPG purchase data, participant responses, stove use monitor data, and enumerator observation.

PM2.5 exposure will be measured in a subset of participants, who will wear PM2.5 exposure monitors for 24 hours. Time-weighted exposure concentrations will be constructed from initial and final filter masses and sample volume.

CO exposure will be measured in a subset of participants, who will wear CO exposure monitors for 24 hours.

BC exposure will be derived ex post in the laboratory from the PM2.5 filters.

NO2 exposure will be measured in a subset of participants, who will wear NO2 exposure monitors for 24 hours; exposures will be derived using colorimetric methods.

Further details are provided in our pre-analysis plan.

In this proposed yearlong RCT, we will enroll households across Shirati, Tanzania, who had previously participated in an LPG RCT (n~511) (Gill-Wiehl et al. 2025), and had received two 6 kg LPG cylinder and burner kits. After obtaining consent to the study, we will conduct a baseline survey that collects information on socio-demographic characteristics and current cooking fuel use. Each respondent will be randomly assigned to one of three study arms.

Before treatment is revealed or delivered, we will ask a random subset of individuals to wear PM2.5, CO, and NO2 monitors for 24 hours, to obtain baseline exposure estimates. Those selected for exposure will also have stove use monitors on their main polluting stove for the study period. For treatment households, we return the next week to either collect the participant’s mobile money number where UCTs will be sent on the first of each month or provide the participant with a distinct ID voucher for them to receive discounted LPG from the main LPG retailers across Shirati. Voucher participants will only be allowed a certain number of subsidized LPG each month, depending on the respondent’s monthly expenditure.

After the baseline survey and exposure measurement, survey enumerators will return every other week to conduct follow-up surveys that collect information on cooking use and the respondents’ expenditure. At baseline, prior to randomization, and at four additional points throughout the study (to capture seasonality), we will ask a random subset to wear PM2.5, CO, and NO2 monitors for 24 hours to obtain exposure estimates. We estimate BC exposures through postsampling transmisometry. After a year, we will analyze the results, develop focus groups and semi-structured interview guides, and conduct qualitative work.

Not available

Randomization through R 4.5.2

Main cooks of households

No

511 main cooks

26 surveys from 511 main cooks ~ 13286 surveys Five rounds of exposure monitoring for 100 participants ~ 500 exposure estimates per pollutant (2,000 exposure estimates) LPG purchase data and stove use monitoring data is continuous throughout the study period

211 in control; 150 in the UCT arm, 150 in the voucher arm

Our primary outcome of interest is monthly LPG use (kgs/month). We test the null hypothesis that LPG use is unaffected by treatment on average. We conducted power calculations using a two sample t-test. Since treatment is assigned at the individual household level, observations can be treated as independent. We assume a two-sided test at a 5% significance level. We defined a minimum detectable effect (MDE) size as the difference between control and treatment uptake rates by arm. The only other published experiment on LPG subsidies indicates that providing subsidized LPG may increase exclusive LPG use from 5.8 kgs/month to 7.4 kgs/month (Table 2 in Jeuland et al. 2023). Thus, we ensure we have power for both interventions to detect at least a 1.6kg/month increase in LPG use. We calculated the necessary sample size to detect this MDE with 80% power (with an alpha of 0.05 and a beta of 0.2) to be 106 participants in each treatment arm and 149 (106*square root (# of treatment arms)). We overpower to account for attrition and the risk that the study in India’s results will not apply to our setting. We enroll 150 in each treatment arm and 211 in the control arm. Our primary pollutant of interest is PM2.5 as the most rigorously studied pollutant within health air quality research. We test the null hypothesis that personal exposure to PM2.5 (ug/m^3) is unaffected by treatment on average. Since treatment is assigned at the individual household level, observations can be treated as independent. We assume a two-sided test at a 5% significance level. We define a MDE as the difference in ug/m^3 throughout the post-intervention period by arm. Numerous studies have evaluated the impact of clean or improved stove interventions on PM2.5 exposure; however, for our power calculations, we rely on those from the Household Air Pollution Intervention Network (HAPIN) trial as the most recent, robust estimate, spanning four different rural field sites. HAPIN found that treatment was associated with a 66% reduction in PM2.5 exposure by the first and second follow-up visits (over roughly a six month period) (102.5 vs. 35.8 ug/m^3) (Johnson et al. 2022). Johnson et al. found that in their study arm standard deviations (SD) were 108 for the control and 55 for the treatment. We calculate a pooled standard deviation of 88. We calculated the necessary sample size to detect this MDE with 80% power (with an alpha of 0.05 and a beta of 0.2) to be 53 participants, but overpower to 100 (58 across the treatment arms and 42 in control).