Credit, attention, and externalities in the adoption of energyefficient technologies by low-income households

Susanna B. Berkouwer§ Joshua T. Dean†

June 3, 2021

Latest version available here.

Abstract

Research in high-income countries has identified carbon taxes and nudges to increase attentionas effective policies to increase investment in energy efficient technologies. We show that thesepolicies will likely be less effective in low-income settings, which are expected to contributealmost all growth in global energy demand in the next several decades. In a randomized fieldexperiment with 1,000 households in Nairobi, Kenya, we study an energy efficient replacement fortheir primary energy-consuming durable: a charcoal cookstove. Using an incentive-compatibleBecker-DeGroot-Marschak mechanism we estimate a 40% reduction in charcoal spending, inline with engineering estimates, which corresponds to a 300% annual rate of return. Despitelarge private fuel savings of $236 over the stove’s two-year life span—almost as high as the $292in carbon emissions reductions—households are only willing to pay $12. Drawing attention topotential energy savings does not increase demand: households already internalize the privatefinancial benefits, perhaps because energy is a larger portion of household budgets. A threemonth loan more than doubles WTP; credit constraints prevent adoption of privately optimaltechnologies. This suggests carbon taxes and nudges to increase the salience of future energysavings are unlikely to significantly increase the adoption of energy efficient appliances in low-income settings, and may instead induce regressive increases in energy costs.JEL: C9, D9, O1, Q4, Q5

§The Wharton School, University of Pennsylvania; 322 Vance Hall, 3733 Spruce Street, Philadelphia PA 19104;sberkou@wharton.upenn.edu. †Booth School of Business, University of Chicago; joshua.dean@chicagobooth.edu. Forgenerous financial support we thank the Weiss Family Program Fund, the International Growth Centre, the Universityof Chicago Booth School of Business, Analytics at Wharton, Berkeley X-Lab, the Art Rosenfeld Fund for EnergyEfficiency, and the Kleinman Center for Energy Policy. We thank Michael Anderson, Max Auffhammer, AbhijitBanerjee, Severin Borenstein, Fiona Burlig, Lucas Davis, Esther Duflo, Pascaline Dupas, Armin Falk, Meredith Fowlie,Matt Lowe, Holger Gerhardt, Michael Greenstone, Marco Gonzalez-Navarro, Jonathan Gruber, Leander Heldring,Kelsey Jack, Seema Jayachandran, Supreet Kaur, David Levine, Jeremy Magruder, Edward Miguel, Matthew Pecenco,Chris Roth, Elisabeth Sadoulet, Jim Sallee, Frank Schilbach, Frederik Schwerter, Daniel Stein, Dmitry Taubinsky,Arthur van Benthem, Florian Zimmerman, Catherine Wolfram, and seminar participants for helpful comments. Wethank the Busara Center for Behavioral Economics, in particular Esther Owelle, Suleiman Amanela, and AnthonyOdhiambo, for superbly implementing field activities. We thank Boston Nyer and everyone at Burn Manufacturingfor their support and Kamen Velichkov for excellent research assistance. This study had IRB approval in Kenya(KEMRI/RES/7/3/1) and the US (Berkeley CPHS 2017-11-2534). To prevent increased Covid-19 transmission, allsurveys conducted by this research team in 2020 were conducted via phone and SMS. A Pre-Analysis Plan filed withthe AEA RCT Registry (ID: 2484) is included in the On-line Appendix.

1

1 Introduction

To meet the Paris Agreement targets cost-effectively, the IEA calls for the world’s huge potentialfor energy efficiency to drive half of targeted emissions reductions (2018). Existing research showsthat the uptake of energy efficiency investments remains inefficiently low, and that two of the mosteffective policies to increase uptake in higher income countries are carbon taxes that internalizenegative externalities and nudges that increase the salience of energy savings (Gerarden, Newell, andStavins 2017). But these policies may not be as effective in lower income settings, for two reasons.First, because energy is a larger portion of household budgets, attention to savings may alreadybe high. Second, credit market frictions, which are more prevalent in low-income settings, mayprevent the adoption of even privately beneficial technologies, limiting the impact of internalizingexternalities. Identifying the optimal policies to promote the adoption of energy efficiency in lowerincome contexts is important because almost all growth in global energy demand in the next severaldecades is expected to come from these countries.

We take up this question in a randomized experiment with 1,000 households in Nairobi, Kenya.We examine how households decide whether to adopt a more efficient version of their primaryenergy consuming durable, a charcoal stove. The efficient stoves, sold under the name Jikokoa, areessentially identical to traditional stoves, but improved insulation reduces the charcoal needed toobtain the same cooking temperature. We use the random price in a Becker, Degroot, and Marschak(1964) (BDM) mechanism as an instrument for adoption to estimate the private and social impactsof stove adoption, and use the willingness-to-pay (WTP) elicited by the BDM mechanism to studythe impact of attention and credit constraints on WTP for the energy efficient stove.

We find that the Jikokoa reduces household charcoal consumption by 40 percent. This yieldsprivate savings of USD 236 and reduces carbon emissions by 6.9 tons of CO2e over the two-yearlifetime of the stove, valued at USD 292 when using a social cost of carbon of USD 42 (U.S. EPA2016). Yet, control households are only willing to pay USD 12 to adopt the stove. A multi-prongedintervention drawing attention to potential energy savings does not increase demand. Householdsdouble their willingness to pay when offered a three month loan, closing the gap between savings andWTP during the loan period and highlighting the importance of credit constraints. Together, thesefindings suggest that without additional interventions, carbon taxes and nudges to make potentialsavings salient are unlikely to significantly increase the adoption of energy efficient appliances inlower income settings, and may instead induce regressive increases in energy costs. Instead, com-bining a smaller tax on the emitting good with a subsidy on the energy efficient technology wouldgenerate larger welfare gains.

We begin by estimating the private and social benefits of adoption. Using the random price inthe BDM mechanism as an instrument for adoption, we estimate a 40 percent reduction in charcoalspending, measured both by a recurring SMS survey of self-reported charcoal expenditures and theweight of charcoal ash generated by the household. This reduction yields savings of USD 118 peryear—about one month of income for the average respondent, and persists at least one year afteradoption. Households also spend one less hour cooking each day due to reduced startup time, and

2

report a 0.5 standard deviation reduction in health symptoms.1

The reduction in charcoal consumption also reduces annual emissions by 3.5 tons of CO2e perhousehold, which results in USD 146 worth of environmental benefits per year. Unlike research onenergy efficient technologies evaluated in high income contexts—which finds that high predictedpresent net benefits of energy efficiency investments are often not realized—we find private benefitsthat substantially exceed the cost of adoption when we quantify the energy efficiency gap.2 In spiteof these large private benefits, households are on average only willing to pay USD 12 for the stove.

We first assess whether this is the result of inattention to fuel savings. In high-income contexts,inattention has been found to be an important determinant of adoption.3 The effectiveness ofattention nudges as a policy lever may be higher or lower in low-income contexts. Energy is amuch larger portion of the household budget (20 percent in our sample), so households may attendto these savings more carefully and make optimal trade-offs.4 Other work suggests the cognitivestress of being poor can impair households’ decision-making capabilities.5 To measure the impactof inattention, we randomize subjects to a multi-faceted attention treatment designed to bring asmuch attention as possible to the savings. Subjects in this treatment arm receive SMSs in themonth leading up to the BDM every three days asking about their charcoal spending. Immediatelybefore stating their WTP, they complete an accounting exercise estimating their yearly savings andwhat they could spend it on. They then receive five minutes to contemplate the savings as thesurveyor enters the savings into the tablet, and the tablet then reminds them of their answers to theaccounting exercise during the WTP elicitation. In contrast, participants in the attention controlgroup receive only standard Jikokoa marketing materials. In spite of this intensive intervention, wefind no impact on WTP, suggesting individuals are already attentive to the savings potential.

If the private benefits are sufficiently large that the social benefits are secondary, and householdsattend to the private financial benefits, why is WTP only USD 12? We consider whether this is dueto credit constraints. This has been documented with the adoption of other technologies (Banerjee,Karlan, and Zinman 2015; Mel, McKenzie, and Woodruff 2008; Pitt and Khandker 1998), but lessso in the context of energy efficiency adoption (Stern and Stiglitz 2021). We offer a random subsetof subjects a three month loan at an interest rate of 1.16% per month to finance adoption.6 Toensure subjects are not simply short of cash-on-hand on the day of the experiment, all subjectsare notified of the purchasing opportunity one month in advance, and those in the credit controlare given an additional 12 hours after the second survey to make the payment for the stove. The

1We discuss additional attributes, such as food taste and durability, in Section 2.1. To the extent that theseprovide additional benefits, our estimate of under-adoption is an underestimate.

2McKinsey & Company (2009) first famously projected widespread availability of energy efficiency investments withnet positive returns. Gillingham and Palmer (2014) provide an overview and explanation of why this opportunity maybe small. For examples, see Fowlie, Greenstone, and Wolfram (2018), Christensen et al. (2019), Burlig et al. (2020),Davis, Martinez, and Taboada (2020), and Allcott and Greenstone (2012).

3Allcott and Taubinsky (2015), Allcott and Wozny (2014), Gillingham, Houde, and van Benthem (2019), Jessoeand Rapson (2014), and De Groote and Verboven (2019).

4Shah, Shafir, and Mullainathan (2015), Fehr, Fink, and Jack (2019), Goldin and Homonoff (2013), Dupas (2009).5Haushofer and Fehr (2014), Schilbach, Schofield, and Mullainathan (2016), Kremer, Rao, and Schilbach (2019),

Duflo, Kremer, and Robinson (2011), Kremer et al. (2013), and Liu (2013).6This was M-Shwari’s monthly interest rate (excluding a 7.5 percent service fee) during the study design period.

3

loan doubles WTP, from USD 12 to USD 25, which equals the average amount of savings they willaccrue during the three month period over which the loan relaxes credit constraints.

We finally consider whether this large response to credit is purely rational inter-temporal sub-stitution. Credit moves costs to the future and typically disperses one large payment across manyperiods, which may make a purchase more attractive for an agent who fails to attend to futurecosts or over attends to concentrated costs (Gabaix and Laibson 2017; Koszegi and Szeidl 2013;Dertwinkel-Kalt et al. 2019). To study inattention, we sub-randomize among the attention treat-ment group to either attend to gross savings or net savings. Those in the net savings group arereminded of their future loan payments during the WTP elicitation. We find that this reminder re-duces the effect of credit by USD 4, an economically and statistically significant difference: shiftingcosts to the future where they are not attended to appears to increase the impact of credit. To studythe importance of concentration bias, we sub-randomize the credit treatment into either weekly ormonthly deadlines. Those with monthly deadlines are given the option to pay early and to switch toweekly payments as a commitment device. This flexibility should make this arm weakly dominantexcept for individuals who are averse to the larger size of the payments. We find no difference acrossthese two arms, suggesting no role for concentration bias in our sample.

Taken together, these results highlight the importance of adapting policies to local contexts.In the past decade, South Africa, Chile, Colombia, and Mexico have all enacted a carbon tax(World Bank 2020), but a carbon tax may not be the optimal policy in low-income settings wherehouseholds face credit constraints. Given that WTP is already low relative to the private benefits,it is unlikely a carbon tax will increase adoption among these households. In fact, a carbon tax maybe regressive by increasing energy costs among the credit constrained, who tend to be the poorest.Given the lack of response to an intense intervention designed to increase attention, nudges to makesalient potential savings may also be ineffective. Instead, our results suggest policy makers wantingto increase the adoption of energy efficient technologies in these contexts should focus on increasingaffordability through subsidies or targeted financing, particularly when the private benefits are largeand inefficiencies due to imperfect targeting are minimal.

This paper proceeds as follows. Section 2 provides background on energy use in Kenya and theenergy efficient stove that we study. Section 3 presents a model of household technology adoption,and Section 4 presents the experimental design and methodology. Section 5 presents the private andsocial impacts of adoption, and Section 6 discusses the drivers of the adoption decision. Section 7discusses the implications of these results for optimal policy. Section 8 concludes.

2 Background: Household energy use in Kenya

Traditional charcoal cookstoves produce indoor air pollution that contributes to millions of deathseach year, and contribute to growing deforestation and climate change (WHO 2017; Pattanayaket al. 2019; Bailis et al. 2015), but more than 4 billion people still do not have access to moderncooking methods (World Bank 2020). By 2030, half of Africa’s population is expected to be living

4

in cities, where more than 80 percent of households rely on charcoal for daily cooking and heatingneeds (FAO 2017). We provide more detail on global and local charcoal consumption in OnlineAppendix Section A.

The share of household income spent on energy costs, also known as the energy burden, tendsto be largest among the poor. Energy spending comprises 3.5 percent of household income for themedian American household, but exceeds 7 percent for the poorest Americans (Drehobl and Ross2016). The share in low-income countries is often even higher: the energy burden for the medianhousehold in our study sample is 20 percent of household income.

Household adoption of energy efficient appliances has the potential to reduce these expendituresmeaningfully—but adoption remains low. The IEA (2018) estimates that cost-effective energyefficiency opportunities available today to households globally have the potential to save USD 201billion in avoided fuel expenditures and USD 365 billion in transport costs per year by 2040.

For the participants in our study, the household’s primary energy-using durable is a traditionalcharcoal stove, a ‘jiko’, which they use for daily cooking. According to Kenya’s Ministry of Energy(2019), around 42 percent of Kenyan households use a charcoal jiko at home, with the primaryalternatives being woodstoves (in rural areas) and liquefied petroleum gas (LPG) and kerosenestoves (in urban areas).

2.1 The energy efficient Jikokoa cookstove

We study the Jikokoa, a charcoal stove produced by Burn Manufacturing (‘Burn’). Burn’s firstmanufacturing facility in Nairobi began producing stoves in 2014. By 2021, they had sold more thana million energy efficient cookstoves, and were selling more cookstoves annually in East Africa thanany other company. Adoption so far has been primarily among Kenyans with higher socioeconomicstatus: we discuss how our sample differs from existing owners in Online Appendix Section B.

The Jikokoa was available for USD 40 in stores and supermarkets across Nairobi at the start ofour study. In theory, this could be a concern for eliciting truthful WTP above USD 40; however,less than one percent of respondents in the control group had a WTP of USD 40 or higher. Figure 1displays a traditional charcoal jiko as well as the energy efficient Jikokoa stove we study.

More than 98 percent of respondents had heard of the stove, primarily via television (66 percent),via a friend, neighbor, or family member (30 percent), on the radio (20 percent), or in a billboard,newspaper, or bus advertisement (10 percent). All respondents received a pamphlet, presented inFigure A1, containing the information about the Jikokoa that is widely advertised on billboardsand television and is accessible to literate and illiterate respondents.

The primary difference between the Jikokoa and the traditional jiko is that the main charcoalchamber of the energy efficient stove is constructed using improved insulation materials. The com-bustion chamber is made of a metal alloy that better retains heat, and a layer of ceramic woolinsulates the combustion chamber to cut heat loss. Parts are made to strict specifications, andcomponents fit tightly to minimize air leakage. These features were designed and tested by labo-ratories in Nairobi and Berkeley, which estimated that they double the charcoal-to-heat conversion

5

rate. Only half the charcoal should therefore be required to reach and maintain the same cookingtemperatures as the traditional jiko.

Adoption of the energy efficient stove does not require any behavioral adaptation or learning.The steps required for cooking are identical, and most adopters continue cooking the same typesand quantities of food as before.7 Both stoves use the same type of charcoal, so users can continueto purchase charcoal from their preferred charcoal vendors. When asked an open-ended questionabout the Jikokoa’s best features, 87 percent of respondents state financial savings, while only 52percent state reduced smoke and 22 percent state time savings. Figure A2 shows that the savingsare almost twice as salient to respondents as the next best attribute.

Other differences may also affect adoption. If subjects perceive these attributes as additionalbenefits, these would bias us towards underestimating under-adoption. This appears to be thecase. In the pilot, we asked 153 subjects why they had not adopted the Jikokoa as an open endedquestion. Answers overwhelmingly had to do with affordability. Zero subjects said their food wouldtaste worse, zero subjects said that it was not healthy, zero said they worried it would break, foursaid they were not sure how to use it, and zero mentioned the stove’s appearance. This open-endedquestion was excluded from the main study for the sake of time, but we did ask about taste anddurability directly. All but three of the 1,018 subjects believe the food will taste the same or better.The median respondent in our sample (correctly) believes that the Jikokoa has an expected lifespanof three years and typically needs to buy a new traditional jiko each year. We therefore defineunder-adoption of the stove conservatively as purely the financial gap between costs and benefits.

2.2 Credit in Nairobi

Loans are common in this context (86 percent of respondents borrowed at least once last year—seeOnline Appendix Section A for more detail), but most respondents face significant credit constraints.More than one-third of respondents had sought out a loan in the past year and been refused, andmore than 50 percent of respondents said they would borrow more if the cost of borrowing was lower.People who had not taken a loan in the past year did not do so largely because they were worriedabout their ability to pay back the loan: for example, M-Shwari charges 7.5 percent for a loan andrequires repayment within one month.8 The company tracks past M-PESA usage and borrowingbehavior to place quantity constraints on individual borrowing. In practice, this means that almosta quarter of our sample would not be able to take out a loan today, even if they wanted to. Themedian amount available for short-term borrowing was less than USD 20; less than a quarter ofthe sample was able to borrow the full cost of the stove if they wanted to. In addition, the loansmentioned above are generally used for emergency situations—a respondent may wish to keep theircredit available for emergencies and not use it to fund technology adoption, as this would leave themvulnerable to unanticipated shocks.

7Respondents report improvements in food quality, but this is primarily enabled by savings from the stove.8Loans that are not repaid within one month are automatically re-registered as a new loan, with an additional 7.5

percent interest rate. If a borrower does not repay a loan within 120 days, they are reported to a local credit bureau(Bharadwaj, Jack, and Suri 2019).

6

3 Conceptual Framework

To fix ideas, consider an agent deciding whether to adopt a more energy efficient version of anappliance that costs price p and reduces future energy costs by ψt for T periods which the agentdiscounts using function D(t). Suppose the agent has access to a loan, l, in exchange for futurepayments, rt, and that the agent makes these inter-temporal substitution choices optimally. Theagent’s maximum willingness to pay is then price p∗ such that:

u(c0)− u(c0 − p∗ + l)︸ ︷︷ ︸Cost of adoption today

=T∑t=1

D(t)[u(ct + ψt − rt)− u(ct)︸ ︷︷ ︸Future benefits of adoption

] (1)

3.1 Carbon taxes to internalize negative externalities

Using this framework, we consider the common polices advocated in higher income contexts—acarbon tax and a nudge to increase the salience of savings—and how they may operate in lowerincome contexts.

The positive externalities from carbon emissions reductions generate social benefits from adop-tion that the agent does not consider. In this case, the agent’s willingness to pay will be lower thanis socially efficient. A carbon tax increases the private savings by internalizing the externalities. Itis easy to see in our framework how increased private benefits increases an unconstrained agent’swillingness to pay. The size of this increase depends on the relative size of the private and socialbenefits.

Previous research in the context of other technologies has shown that credit constraints canprevent households from being willing to pay even the privately optimal price. Consider an agentwho is constrained to borrow at most lcc with payments rcct and that this constraint binds. Theirmaximum willingness to pay, pcc, is defined by:

u(c0)− u(c0 − pcc + lcc) =T∑t=1

D(t)[u(ct + ψt − rcct )− u(ct)] (2)

In this setting, with concave utility, increasing the private benefits with a carbon tax will lead toa smaller increase in adoption than the unconstrained case. Households with limited ability tosubstitute the savings into the present will instead have to sacrifice more consumption in order toadopt. This will be particularly painful for low income households located on the steep portion ofthe utility function. If purchasing the efficient version of the appliance would put the agent belowa subsistence threshold, the carbon tax will have no impact on adoption.

3.2 Nudges to increase attention

Agents may also not attend to all of the future savings, instead perceiving the savings as θψt whereθ < 1. This decreases their willingness to pay for obvious reasons. In this case, a nudge may beable to increase attention and subsequently increase willingness to pay.

7

How well this will work in lower income settings depends both on how attentive agents are tothe savings and whether agents are credit constrained. If agents are more attentive to the potentialsavings at baseline because energy is a larger fraction of their budget, the nudge will be less effective.Conversely, if they are less attentive due to the many demands poverty places on their attention,nudges may be more effective.

Nudges will be less effective with credit constraints for the same reason these constraints impedethe effectiveness of a carbon tax. Increasing perceived benefits will increase adoption less whenagents cannot bring these benefits into the present period to avoid costly utility losses.

4 Experimental Design

We enrolled 1,018 respondents who live in the Dandora, Kayole, Mathare, and Mukuru informalsettlement areas around Nairobi, Kenya (see Figure A3 for respondent locations). These areasare among the lowest-income areas of Nairobi, and have not been targeted by sales teams of thecookstove company. Field officers walked around these areas and enrolled a convenience sampleby visiting the homes of potential respondents, until the required number of respondents had beenenrolled. To qualify for study participation, respondents had to use a traditional charcoal jiko astheir primary cooking technology and spend at least USD 3 per week on charcoal, although this isnot usually binding: most households that use a charcoal stove as their primary cooking technologyusually buy at least USD 0.50 per day, or USD 3.50 per week, worth of charcoal.

Table 1 presents summary statistics of socioeconomic variables. The median household in oursample earns a daily income of USD 5, and spends USD 0.70 (14 percent of income) on charcoal perday. 60 percent of study participants purchase charcoal at least once per day. Households buy anew jiko around once per year, for between USD 2–5. Within each household, field officers enrolledthe primary stove user. 95 percent of respondents in our sample are women, largely reflectingKenyan societal norms and expectations around household tasks. For more details on samplerepresentativeness please see Online Appendix Section B.

The following sections describe the timeline, randomized treatments, and measurement method-ologies, which are in line with the pre-analysis plan. The sample size for each treatment exceedsthat specified in the pre-analysis plan, which was determined using statistical power calculations.

4.1 Experimental timeline

The survey design centers around three in-person visits 25-30 days apart referred to as visit 1,visit 2, and visit 3, or the baseline, midline, and endline visits, respectively. One year later along-term endline survey is administered over the phone.9,10 Participants complete three additional

9The long-term endline was intended to be in-person, but this was changed to phone due to COVID-19 restrictions.10Due to logistical constraints, visits deviated moderately from this. 88 percent of visit 2 surveys were conducted

between 23-33 days of that respondent’s visit 1, and 90 percent of visit 3 surveys were conducted between 23-33 daysof that respondent’s visit 2. 90 percent of long-term endline surveys were conducted between 12.2 and 13.1 monthsof visit 2.

8

activities: 1) A recurring SMS survey asked about a respondent’s charcoal expenditures every threedays, implemented during the main study and after the long-term endline; 2) Collection of ash in abucket to measure charcoal usage; and 3) Loan payments by respondents who purchased the stoveusing credit. Figure 2 presents the timeline for these components.

During visit 1, the field officer completes an enrollment survey and gives the respondent apamphlet about the stove (Figure A1). To relax short-term liquidity constraints, the field officertells the respondent that a colleague will return one month later, and that if they would like tobuy the stove then, they should have sufficient cash on hand. While the respondent’s price is notdisclosed during visit 1, respondents are told that they “may receive a small discount.”

After visit 1, each respondent is randomly assigned a subsidized price for the stove, and israndomly assigned into one of three credit treatment groups and one of three attention groups. Re-spondents in the treated attention groups then start receiving SMSes about their charcoal spending(described in Section 4.4). To prevent contact imbalance, respondents in the attention control groupreceive placebo SMSes before switching to charcoal SMSes after visit 2.

During visit 2, the field officer implements the credit and/or attention treatments for this re-spondent (described in Section 4.2), and the BDM mechanism (described in Section 4.3). If therespondent wins the stove, they receive the stove during visit 2. Respondents in the credit controlgroup must pay their price during visit 2. Participants who won the stove and who are in one of thecredit treatment groups begin making their loan payments after visit 2. All participants are alsoasked to collect the charcoal ash generated by stove usage using the bucket provided.

During visit 3, the field officer implements the endline survey and weighs the ash collectionbucket. One year later, a field officer conducts a similar endline survey over the phone.

4.2 Credit and attention: Experimental treatment arms

We implement a 3-by-3 experimental design, cross-randomizing two credit treatments with twoattention treatments (Figure 3). Treatment is stratified by baseline charcoal spending.

Loan payments include an interest rate of r = 1.16 percent per month, which was the borrowinginterest rate of M-Shwari (exclusive of fees) at the time of our study. Respondents who do not maketheir payments are asked to return the stove.11 Regardless of credit treatment assignment, everyrespondent who purchased the stove received it during visit 2.

Credit control group (C0): Payment is due at visit 2.

Weekly deadlines (C1): Payment is due at 12 weekly deadlines, starting oneweek from visit 2. Payments may be made more frequently or earlier, as long asthe cumulative minimum is met by each weekly deadline.

11All respondents received SMSes reminding them of their upcoming payment deadlines in advance. If a respondentmissed a deadline, they were initially sent three reminders over a six day period. If they had not paid within oneweek after their missed deadline, a field officer would visit them and reclaim the stove. More detail on repayment isprovided in Section 6.3.

9

Monthly deadlines (C2): Payment is due at three 4-weekly deadlines, starting4 weeks from visit 2. Payments may be made more frequently or earlier, as longas the cumulative minimum is met by each monthly deadline. For example, theycould pay in weekly instalments. Respondents were told they would have theoption to commit to weekly deadlines (C1) after the WTP elicitation.

In addition to facilitating inter-temporal substitution, credit may make a product more attractiveby both moving the costs to the future and making them more diffuse. The difference in the coststreams of C1 and C2 allows us to test whether agents over attend to concentrated costs (Koszegiand Szeidl 2013, Dertwinkel-Kalt et al. 2019). Importantly, because those in C2 are free to chooseto either commit to weekly payments, make payments more frequently than monthly at their owndiscretion, or make monthly payments, they have strictly greater flexibility than those in C1. Thisshould make this option more attractive except for the fact that the payments are presented asmonthly totals. An individual exhibiting concentration bias will perceive these larger paymentnumbers as relatively less attractive and have a lower WTP.

While the conceptual framework does not consider the cost of credit or the cost of default, thecredit intervention implicitly addresses three channels through which individuals may face creditconstraints. It addresses quantity constraints by not constraining the size of the loan. By chargingan interest rate that excludes the fees charged by mobile lenders, it reduces the cost of borrowing.Finally, agents face limited penalties under default. If someone defaults on their loan, a field officercollects the stove, and the participant faces no additional repercussions.12

To test for inattention, we cross-randomize the three credit treatment arms with three treatmentarms designed to increase attention:

Attention control group (A0): Participants are informed that the stove man-ufacturer says that the stove reduces charcoal consumption by 50 percent. Theyare informed of the Ksh equivalent of these savings, based on the respondent’sstated average weekly charcoal spending. They are also given a calculator andallowed to perform calculations regarding their expected savings.

Attention to energy savings (A1): Participants receive everything that A0receives. In the month between visits 1 and 2, respondents are then asked abouttheir charcoal spending every three days via SMS.13 During visit 2, the respondentis asked to complete the attention sheet displayed in Figure A4 immediately prior

12Around 13 percent of respondents who adopted the stove using credit had paid less than 10 percent of therequired amount by the end of the payment period: for 56 respondents, we do not have any record of them payingany amount. In six cases, the field officer retrieved the stove. In all other cases, field officers were unable to trackdown the respondent. The research team does not have authority to implement any penalties or legal repercussions,other than repeatedly encouraging repayment.

13Respondents in the attention control group received placebo SMSes between visits 1 and 2. The timing andincentives were identical, but respondents were asked about their matatu (bus) travel instead of their charcoalexpenditures. Starting at visit 2, these respondents received the regular charcoal SMS survey.

10

to the BDM elicitation,14 writing down the amount of money they think theywould save each week for the next year if they owned an energy efficient stove.This can be expected to be around 50 percent of their expected spending eachweek. Since savings are proportional to spending, the respondent might expectlarger savings during weeks when they expect to cook more, for example duringreligious holidays, or when a temporary migrant returns home. The respondentthen calculates the expected savings for each of the twelve months, writes downhow they would use these savings for each month, and calculates the total amountsaved after a year. Respondents are then given a waiting period15 of five min-utes to think about these savings while the enumerator enters the numbers intoa tablet. The savings are subsequently shown on the tablet during the BDM elic-itation.16

Attention to energy savings minus costs (A2): Participants receive every-thing that A1 receives. In addition, during the BDM elicitation, alongside thesavings they are informed of the cost of adoption during each period. The cost perperiod is calculated and presented in line with the respondent’s credit treatmentassignment (C0, C1, C2). The net benefit (defined as cost minus savings) for eachperiod is also calculated and presented to the respondent.

The difference between A0 and {A1, A2} provides a test of whether respondents are inattentiveto savings and would thus potentially benefit from nudges to increase attention. The differencebetween A1 and A2 allows us to test whether credit may make the product more attractive bymoving costs into the future where agents may attend to them less (Gabaix and Laibson 2017).

One may be concerned that the attention treatment addresses math ability or provides awarenessof the technology rather than addressing attention alone. To alleviate this concern, respondentscomplete a short math test consisting of eight questions taken from Kenya’s Certificate of PrimaryEducation (KCPE) and Secondary Education (KCSE) standardized exams. This allows us to ruleout heterogeneity in the attention treatment by math ability. Only 10 people had not heard of theJikokoa stove at baseline, so we are unable to test whether the attention treatment works moreeffectively for respondents that were not aware of the technology at baseline.

One may also be concerned that participating in the experiment, even in the A0 condition, isitself an attention treatment, and that we may thus find no impact of attention because participationbrings respondents to full attention. While this is possible, the attention drawn to the product inthe A0 condition mirrors that done by Burn’s typical marketing efforts (handing out fliers, bus-side

1447 percent of respondents filled in the sheet entirely independently. 31 percent of respondents wrote in most ofthe sheet independently, but required some guidance by the field officer. 22 percent of respondents were illiterate andthe field officer completed their sheet on their behalf.

15Recent work has shown that a waiting period, defined as a delay between information about a prospective choiceand the choice itself, can lead to more forward-looking choices. For example, Brownback, Imas, and Kuhn (2019)find that a waiting period causes a 28 percent increase in healthy food purchases.

16Figure A5 provides examples of what is shown on the screen for three hypothetical respondents.

11

and billboard media, demonstrations by distributors, reaching out to ROSCAs, etc.). This suggeststhat any inattention of this nature is unlikely to persist in the market. Treatments A1 and A2 aimto examine whether attentional failures not addressed by traditional marketing strategies may leadto inefficiently low demand. To the extent that this remains a concern, the control WTP providesan upper-bound on the effect of participating in the experiment on WTP.

4.3 Becker, DeGroot, Marschak (BDM) mechanism

We implement the Becker, DeGroot, Marschak (BDM) mechanism defined in Becker, Degroot, andMarschak (1964), building on the implementations developed in Berry, Fischer, and Guiteras (2020)and Dean (2020). Between visits 1 and 2, each respondent is randomly allocated a hidden pricethat is printed and sealed inside an envelope with the respondent’s name on it. The prices rangefrom USD 0.01 to USD 29.99, but neither the respondent nor the field officer know the price insidethe envelope or the distribution of prices.17 To eliminate the need for contingent reasoning andto provide concreteness, rather than asking for the respondent’s maximum WTP in an open-endedfashion, the field officer conducts a binary search over the range of USD 0 – USD 40, asking “If theprice of the Jikokoa is 2,000 Ksh [USD 20.00] would you want to buy it?”, then proceeding to thenext price based on the respondent’s answer. After arriving at a final WTP the envelope is openedand the respondent buys the stove for the price in the envelope if and only if it is less than theirWTP. For more details see Online Appendix Section B.

Prior to the BDM each respondent completes two practice exercises, one for a bottle of lotion andone for a bar of soap (Figure A7). Each respondent is allocated a random price for the lotion and arandom price for the soap. Respondents are randomly assigned whether they would be offered thelotion using take-it-or-leave-it (‘TIOLI’) and the bar of soap using BDM, or vice versa. These twopractice take-up decisions serve two purposes. First, participants learn how the BDM mechanismworks relative to the TIOLI they are used to in stores and experience the binding outcome ofthe bidding process. Second, Figure A8 displays significant overlap in the demand curves elicitedthrough the TIOLI and BDM mechanisms for both goods, suggesting respondents understand theBDM mechanism and that the elicited WTP values reflect realistic decisions.18

The BDM mechanism serves two purposes. First, since the price is fixed ex ante, the mechanismis incentive compatible: it is in the respondent’s best interest to state their true WTP. Second,because the hidden prices are randomly assigned, adoption is random conditional on WTP. Ran-domized stove assignment allows us to estimate the causal impact of adoption on charcoal spending.

4.4 Measuring charcoal use

We use three measures of charcoal use. The primary outcome is a recurring SMS survey. Everythree days the respondent receives an SMS asking how much money they spent on charcoal in the

17Figure A6 displays the distribution of BDM prices across participants.18While one might be concerned these practice purchase opportunities would reduce liquidity, the average price

paid for these goods was USD 0.68—substantially less than the cost of the stove.

12

past three days.19 The SMS survey was administered for three months after visit 1 (visit 2 for theattention control group), and again for one month after the one year endline survey. Second, weask respondents to recall their recent charcoal expenditures during the visit 3 endline survey andthe one-year endline survey. Finally, to generate ground-truth comparisons of these self-reportedmeasures and address concerns about experimenter demand, respondents collect the ash generatedby stove usage between visits 2 and 3. Normally, when a respondent is done cooking a meal, theydispose of the charcoal ash in the trash. Instead, during visit 2, each respondent is given an empty20 liter bucket and asked to dispose of the used ash in the bucket rather than in the trash. Duringvisit 3, field officers weigh the bucket using a hand-held weighing scale.

4.5 Heterogeneity measures: time-inconsistency and risk

We measure time-inconsistency through an effort task allocation exercise closely following Augen-blick, Niederle, and Sprenger (2015). Figure A9 displays the effort task. Unfortunately, subjectsappear to have not fully understood the interest rates as the number of tasks allocated to visit 2 arerelatively constant across interest rates with several violations of monotonicity. Thus we treat theseheterogeneity results as suggestive and focus on the simplest choice where the interest rate is oneand classify respondents who choose to postpone additional tasks during visit 2 relative to their de-cisions in visit 1 as exhibiting time-inconsistency using a binary indicator. Using this methodology,57 percent of respondents are classified as exhibiting time-inconsistency.20

To measure risk preferences, at the end of visit 1, each respondent is offered USD 4 for their timethat day, and told they may invest any amount x ∈ [0, 4], which pays 3x or zero with 50 percentprobability each.21 Respondents who choose to invest x < 2 (68 percent) are classified as exhibitingrisk aversion.

5 Private and social impacts of adoption

In this section, we estimate the welfare effects of adoption. Since WTP cannot be interpreted asthe total welfare gain when agents are credit constrained (we discuss this more in Section 7.4), wecompute private and social benefits for the most plausible channels.

Out of the 955 respondents who completed visit 2, 570 (60 percent) adopted the stove, paying anaverage price of USD 15.22 The benefit from two years of ownership, discounting weekly at δ = 0.9

19To increase response rates, respondents receive USD 0.20 for every SMS they correctly respond to and a bonusof USD 2 for every 10 SMSes that they correctly respond to. An SMS is correct if it identifies charcoal spending inthe past 3 days. Incorrect messages include those that refer to quantities (e.g. ‘1 KG’ or ‘2 tins’); irrelevant messages(e.g. comments about the weather); amounts below USD 0.10 (these are assumed to be typographical errors); or anySMS beyond 1 per day (only the last correct SMS of the day qualifies).

20As is common in these allocations, some subjects also increase their allocation of tasks at visit two. This is alsotechnically time inconsistent behavior, but for the sake of simplifying exposition, we do not refer to it as such.

21This follows Gneezy and Potters (1997) and Charness, Gneezy, and Imas (2013). Respondents blindly pick oneof two pieces of paper from a small bag. If the paper says ‘Win’, they receive 3x what they invested. If the papersays ‘Lose’, they lose what they invested. Respondents always retain the uninvested amount (4− x).

22Respondents who did not were still able to buy it at retail prices. In the year after visit 2, about 4 percent did

13

annualized, totals USD 696. This exceeds the retail cost of USD 40. These discounted benefitsconsist of USD 213 in financial savings (Section 5.1), USD 263 in avoided greenhouse gas emissions(Section 5.2), and USD 221 in time savings (Section 5.3). We discuss health and other non-financialbenefits in Section 5.4.

These impacts persist at least one year after adoption (Section 5.5) and we cannot rule outthat they are homogeneous across agents (Section 5.6). Sections 5.7 and 5.8 rule out any welfare-reducing attributes and energy rebound effects. Finally, section 5.9 documents low and balancedattrition. Table A1 presents balance checks for the random treatments. None of the joint F-testsare significant.

5.1 Charcoal spending and usage

Figure 4 presents charcoal spending before and after the main visit, for adopters and non-adoptersof the energy efficient stove, as elicited using the SMS survey. Weekly spending decreases sharplyimmediately after adoption.

To estimate the causal effect of stove adoption on charcoal spending, we use the randomlyassigned BDM price as an instrument for adoption. Table 2 presents the results. In the first stagein Column (2), the BDM price strongly predicts adoption. Column (3) shows that adoption reducescharcoal spending by USD 2.27 per week.23,24 The 50 log point reduction corresponds to a 40 percentdecrease in charcoal consumption.25

Column (5) shows that adoption causes a 40 percent reduction in total ash generated betweenvisits 2 and 3. This estimate matches the independent estimate from the SMS data. Converting‘weekly charcoal spending’ (in Ksh) to ‘kilograms charcoal purchased’ (in KG) using local charcoalmarket prices, and comparing KG of charcoal purchased with KG of ash generated from charcoalusage, identifies a charcoal-to-ash conversion ratio of 1.6 percent. This matches the physico-chemicalproperties of charcoal (FAO 1987).

It is worth putting these savings into perspective. USD 2.27 per week corresponds to USD 118per year—one month of income for the average respondent. Net Present Value26 (NPV) after twoyears of stove ownership is USD 173 for the average respondent. Given that most respondents areon the steep part of the utility function, the marginal utility from these savings is likely large. Whenasked how they spent their charcoal savings, 53 percent of respondents report buying more food,23 percent report paying school fees, and 15 percent report buying household items such as soap orclothes.

so. After the conclusion of visit 3, Burn released a new model of the stove that was smaller and only cost USD 30.As a result, most respondents who bought a stove in the year after visit 2 paid less than USD 40.

23To accommodate values of 0, we use an Inverse Hyperbolic Sine (IHS) transformation to estimate the impact inproportional terms (Burbidge, Magee, and Robb 1988). IHS is defined as: sinh−1(x) = log(x+ (x2 + 1)1/2).

24One might be concerned that some of this reduction in spending is due to the burden of the stove purchase onthe household budget. This is weighed against by the fact that the effect remains stable more than a year later whentotal savings have outstripped cost of the stove and the similarity of the effects across treatment arms (Table A2).

25Figure A10 displays the instrumental variables results using SMS data over time graphically.26NPVi =

[∑Tt=1 δ

tψit]− PE , where ψit are savings. We use δ = 0.9 annualized and PE = 40 USD.

14

Our empirical estimates of charcoal savings align closely with ex-ante engineering predictions.Burn previously estimated that the Jikokoa uses 45 percent less charcoal than a traditional Kenyanstove (Ashden 2015). Our point estimate is a 40 percent reduction with a 95 percent confidenceinterval of 30-48 percent. We therefore cannot rule out that the engineering estimates accuratelypredict realized savings. This departs from extensive empirical work evaluating energy efficiencyinvestments in high income contexts—including Fowlie, Greenstone, and Wolfram (2018), Burlig etal. (2020), and Gillingham and Palmer (2014)—which find realized savings lacking when comparedto engineering estimates.27 This may be due to the inelastic demand for cooking, the homogeneityof the technology, and the simplicity of its implementation. Our estimate is also significantly largerthan those in many papers studying the adoption of improved cookstoves.28

Relative to a retail price of USD 40, these savings constitute an average internal rate of return(IRR) of 24.8 percent per month, or 294 percent per year.29 Table 3 shows that the IRR is larger thanmost relevant alternative investments that are likely available to households, including investmentsin business, agriculture, and education, according to recent literature. Research in the U.S. haseven found negative IRR for household investments in energy efficient technologies.

5.2 Environmental externalities

The median household spends USD 255 on charcoal per year. Using local market prices, thisimplies each household consumes 849 KG of charcoal per year. The stove’s 40 percent reductionthen corresponds to 6.9 metric tons of CO2e in avoided emissions over two years of usage.30 Usingthe EPA (2016) estimate for the 2020 social cost of CO2 of USD 42, adoption of a stove generatesUSD 292 in CO2e emission reductions over two years of usage. Focusing on only the environmentalbenefits, investing in a Jikokoa reduces greenhouse gas emissions at a cost of USD 5.76 per ton ofCO2e.

Anthoff, Hepburn, and Tol (2009) argue that the social cost of carbon will vary substantiallydepending on the income of the country to which the estimates are normalized, and the locallyrelevant SCC should be used to determine each country’s optimal abatement. Given low incomes inSub-Saharan Africa, the SCC would be lower than the U.S. EPA SCC. Given the lack of acceptedregional SCC’s, we refrain from quantifying this, noting only that this will increase the importanceof private benefits relative to social benefits.

27Christensen et al. (2019) analyse what may drive this wedge.28See Pattanayak et al. 2019, Hanna, Duflo, and Greenstone 2016, Levine et al. 2018, Mobarak et al. 2012, Burwen

and Levine 2012, Beltramo and Levine 2013, and Chowdhury et al. 2019 for examples.29The IRR corresponds to the discount rate where the Net Present Value of an investment, from time 0 to infinity

(we assume two years of use), equals 0. Specifically, IRR equals δ such that[∑T

t=1 δtψt]− PE = 0. Conversion

between monthly and annualized IRR conservatively (and in line with the literature) assumes no reinvestment.30The production of charcoal emits 7.2–9.0 KG of carbon dioxide equivalent, CO2e, per KG of charcoal (FAO,

2017). The combustion emits a further 2.2–2.6 KG (Bhattacharya, Albina, and Salam 2002). We use the mid-pointsin our calculations.

15

5.3 Time savings

Table 4 presents results from an instrumental variables regression, using the randomly assignedprice as an instrument for adoption, to estimate the causal impact of adoption on non-financialoutcomes. Column (2) estimates a one hour daily reduction in cooking time.31 Figure A11 displaysthe distribution of daily cooking time by treatment. To value these time savings, we use medianearnings of USD 3 per day and assume daily earnings scale linearly with hours worked starting atan 8-hour work day. The time savings correspond to USD 0.34 per day, or 107 percent of medianfinancial savings.

5.4 Health and other non-financial attributes

Columns (4) through (8) of Table 4 show that adoption of the Jikokoa causes a 0.5 standard deviationimprovement in health in both the 1-month endline survey and the 1-year endline survey.32 However,we exclude health benefits from our welfare totals because there is substantial uncertainty in medicalcosts, value of statistical life, and disability-adjusted life years.

Most respondents report that space heating generated by stove usage helps keep them warmduring winter months.33 Despite these heating benefits, few respondents ever burn charcoal purelywith the goal of heating their living space. Heating is thus a side benefit from cooking rather thana goal in and of itself. We therefore exclude this when enumerating benefits.

Two-thirds of respondents who adopted the stove said they did not change which foods theycook, and 71 percent said they cook the same quantity of food as before. 61 percent said that foodcooked with the energy efficient stove tasted slightly or a lot better, and fewer than 1 percent saidthe food tasted worse.

To study network effects, we evaluate whether households in geographic or social proximity torespondents in our sample have the Jikokoa. Column (3) of Table 4 documents no increased adoptionby neighbors, friends, or family of the respondent in the month after adoption. This confirms thatthe binding constraint is not information or perceptions of stove quality.

5.5 Long-term impacts

Previous studies of efficient cookstoves found declines in usage and benefits over time due to tech-nology breakdown, poor maintenance, or user tastes (Pillarisetti et al. 2014; Duflo, Kremer, andRobinson 2011). To test this, we re-survey respondents and conduct a charcoal SMS survey 12–14months after the main experiment.34

31Likely driven by a reduction in the time spent lighting charcoal, which is time-consuming for traditional jiko’s.32The health index consists of self-reported health and respiratory symptoms for the primary cookstove user and

any children (if applicable). The index is standardized for the control group to have a mean of 0 and a standarddeviation of 1. A higher value indicates more respiratory symptoms, and thus, poorer health. These results areself-reported and may therefore be biased due to experimenter demand or Hawthorne effects. In-person surveys tomeasure physiological health outcomes were cancelled due to COVID-19 safety considerations.

33The endline survey was conducted in June–July, which are Nairobi’s coldest months—temperatures can drop tothe single digits (◦C)—so heating benefits were likely large relative to the annual average.

34Of the 955 participants who completed visit 2, 875 (92 percent) completed the one-year endline survey.

16

Out of the 521 stove adopters re-surveyed one year later, 510 (98 percent) still had the Jikokoa.Reasons for loss include theft, fire, non-payment repossession, and giving the stove away voluntarily.27 percent of adopters still had a working traditional jiko at home, 46 percent said their jiko hadbroken and they simply did not replace it, and 23 percent had given their old jiko away as a gift.Only one person said they had sold their jiko, which is not surprising given the lack of a secondaryjiko market, likely due to its fragility and low cost.

Columns (6) and (7) of Table 2 show that twelve months after adoption, the stove continuesto cause a reduction in charcoal spending of USD 2.42 per week, corresponding to a 42 percentreduction (54 log points) relative to the control group. Savings appear constant over the long term.This improvement on previous cookstove technologies may be attributable to the Jikokoa’s ease ofuse and similarity to traditional jiko’s. It is also more durable, and adopters can access free repairservices provided in Nairobi if needed.35

Table A3 shows suggestive evidence of a positive impact on long-term financial assets, in particu-lar for respondents who at baseline had positive savings in a SACCO group, mobile money account,or formal banking account. Among those with positive savings at baseline, we find a 56 log pointincrease in savings, driven by larger SACCO payouts.

5.6 Heterogeneity

The stove appears to benefit almost all adopters. Figure A12 presents unconditional local quantiletreatment effects estimated using the method proposed by Frölich and Melly (2013). The effect isrelatively homogenous across the charcoal spending distribution. Figure A13 presents the sortedgroup average treatment effects estimated using the method in Chernozhukov et al. (2018), whichfinds the most heterogeneity possible using baseline observables.36 While the confidence intervalsare wide, the point estimates suggest the entire sample benefits from adopting the stove.

5.7 Ruling out welfare-reducing attributes

If stove adoption causes unpredictable negative impacts, increased adoption may reduce welfare.We rule this out for several reasons. First, Figure A14 shows that the relationship between WTPas measured during the BDM mechanism and stated WTP elicited during the endline survey isnearly identical for respondents who adopted the stove and those who did not. It is thereforeunlikely that there is substantial learning of any welfare-reducing hidden attributes post adoption.Second, during the endline survey 99 percent of stove adopters say they recommend the stoveto friends and family members. Fewer than 1 percent had ever considered selling it, suggesting

35Respondents can call the Jikokoa service number to locate their nearest repair shop. Damaged from customermisuses does not qualify for repair. This service is therefore not expected to induce moral hazard.

36We use LASSO on half the sample to predict charcoal spending with and without the stove based on baselineobservables. In the other half of the sample, we predict the treatment effect for each individual as the differencebetween these two predicted spending levels. We use this predicted treatment effect to split the sample into thirdsand estimate the treatment effect for each group. Repeating this process over 1,000 random splits stratified by WTPand taking the medians gives the estimates in the figure.

17

limited hidden welfare-reducing stove attributes.37 Finally, by investing in the stove a householdmay forgo an alternate investment with higher returns. By comparing our findings with relatedliterature, Table 3 documents that alternative investments that this population may have access toare unlikely to generate higher returns than the energy efficient cookstove.

5.8 Ruling out a rebound effect

The rebound effect describes technological progress where improvements in efficiency are partly offsetby increased usage (Jevons 1866), at an extreme even increasing usage of the input, referred to asthe Jevons paradox. In energy efficiency, this is often due to either an income effect—individualsuse savings generated from the investment to use the appliance more—or a substitution effect—usage of the appliance is now relatively cheaper (Borenstein 2015; Gillingham, Rapson, and Wagner2015; Chan and Gillingham 2015). A rebound would complicate interpretation. Increased usagewould reduce net savings, but increase utility derived from the technology, which would need to befactored into welfare calculations. Rebound may also increase energy usage by increasing usage ofother energy-consuming durables due to an income effect. This is likely small in this context giventhat the cookstove is each household’s primary energy consuming durable.

Rebound is likely small in this context for three reasons. First, cooking is inelastic: a regressionof log of time spent cooking on log of income yields a coefficient that is not statistically significantfrom zero. In the endline survey, 71 percent of adopters report no change in the amount of foodthey cook. 23 percent state that the amount “increased slightly,” but likely do so without cookingfor longer or using more charcoal, instead simply adding more food into the pot they were alreadycooking with. Second, a rebound effect would generate a wedge between engineering estimates andrealized energy efficiency gains. To the contrary, our empirical estimates align closely with ex-anteengineering predictions. Finally, a rebound effect would cause an increase in the time spent cooking,whereas we find a decrease of 54 minutes per day.

5.9 Attrition

Selective attrition might bias results. We test for attrition and do not find meaningful variation.Of the 1,018 respondents who were enrolled during visit 1, 955 (94 percent) completed visit 2 and924 (91 percent) completed visit 3. Table A4 confirms that attrition for surveys and median SMSresponse is balanced across all socioeconomic and treatment groups except age and stove adoptionstatus. Figure 5 displays limited attrition in the SMS survey.

6 Drivers of adoption

Average WTP for the stove is USD 12. This is lower than average private discounted savings of USD213, and substantially lower than total private and social benefits of USD 696. Figure 6 displays

37To limit experimenter demand, field officers repeatedly reminded the respondent that they were part of a universityresearch team, working independently from the cookstove company, and that responses were anonymous.

18

the histogram of WTP for the pure control group, and the corresponding demand curve, definedQ(PE) = Pr(PE ≤WTPi).

We compare this against a hypothetical breakeven demand curve, defined Q(PE) = Pr(PE ≤p∗i ). A fully rational, unconstrained risk-neutral agent has a maximum WTP of p∗i =

∑Tt=1D(t)γsi.

We use γ = 40 percent as estimated in Section 5.1 and use baseline charcoal spending as a proxyfor counterfactual spending si. We conservatively limit T to the three-month period over which ourloan relaxes credit constraints. We assume exponential discounting D(t) = δt, with an annualizeddiscount factor of 0.9.38 The difference between the breakeven demand curve and the pure controldemand curve in Figure 6 indicates significant under-adoption. Any reduction in the wedge betweenthe two demand curves caused by a treatment addressing a particular constraint or bias can beinterpreted as its contribution to the under-adoption gap.

Sections 6.1 and 6.2 investigate how credit and attention affect under-adoption. Section 6.3discusses loan repayment patterns, and Section 6.4 discusses how risk aversion affects adoption.39

6.1 Credit doubles WTP, while attention to benefits has no impact

Column (1) of Table 5 shows that access to credit increases WTP by USD 12.58, or 104 percentrelative to the control group.40 Households are unable to take advantage of investments even whenthe private financial returns are 300 percent per year on average, and relatively homogeneous acrossagents. Credit alone appears to be sufficient to fully close the energy efficiency gap over the 3-monthperiod of the loan. Figure 7 presents these results graphically.

This rate of return is higher than local lending rates. This suggests that these constraints aredue to more than inefficiently high costs of borrowing. One possibility is that credit providers arechoosing to ration by quantity rather than price. A related possibility is that agents are sufficientlypresent biased that they fail to maintain liquid buffer stocks. This would both prevent them fromhaving cash-on-hand to finance stove adoption, and would lead to lines of credit being allocated tosmoothing shocks (Kremer, Rao, and Schilbach 2019).

The attention to benefits treatment has zero impact on WTP.41,42 Interestingly, Table A6 showsthat this lack of response is in spite of a 0.2 standard deviation increase in stated beliefs aboutsavings. Given the lack of impact of the attention treatment on WTP, in combination with theheavy handed treatment, the increase in stated beliefs may be the result of experimenter demand.This lack of impact on WTP may be because this is a high-stakes decision: when mistakes arecostly, an individual living in poverty may be relatively more attentive to adoption decisions (Shah,

38Figure A15 confirms that the results are robust for δ = 0.5 and δ = 1 due to the short time horizon. Even usingthe relatively more conservative discount parameters recently estimated by Balakrishnan, Haushofer, and Jakiela(2020) in a similar context using a money now/money later design (δ = 0.942 per week and β = 0.924), the lowcontrol WTP cannot be explained by time preferences alone.

39These analyses match those proposed in the Pre-Analysis Plan, available in the On-line Appendix.40Table A5 in the appendix provides a full breakdown of primary treatment effects.41Given that our attention to benefits treatment was designed in part to address concentration bias, it is reassuring

that we also do not find any evidence of concentration bias in costs. We discuss this further in Section 6.2 below.42The estimates rule out an effect larger than USD 1.70.

19

Shafir, and Mullainathan 2015; Fehr, Fink, and Jack 2019).Figure A16 shows that there is no relationship between WTP and stove benefits, whether ex-

pected or realized. Interestingly, even after credit constraints are relaxed, we do not find a relation-ship between expected savings and WTP among those offered credit. One explanation for this isthat subjects’ beliefs about average savings are well calibrated but they fail to appreciate hetero-geneity in individual circumstance. Some evidence for this comes from our subjective belief data.In the attention control group at the second visit, the average respondent’s median belief is thatthey would save USD 89 over the next year by purchasing the stove; however, the average standarddeviation of their beliefs was USD 15. This suggests that respondents are reasonably correct onaverage, but exhibit significant uncertainty about their beliefs.

6.2 The psychology of credit

Credit changes the structure of costs: it postpones costs to the future, and reduces the maximumcost incurred in any single period. In addition to relaxing credit constraints, credit may thereforework in part through psychological channels. We provide results on two potential psychologicalmechanisms: inattention to future costs and concentration bias.

6.2.1 Inattention to future costs

First, we consider whether the impact of credit is mediated by inattention to future costs. If this isthe case, the impact of credit should be smaller among agents who are reminded of future costs whilewe elicit WTP. To test this, we interact the credit treatment with the attention to costs treatment.Column (2) of Table 5 shows that attention to future loan payments reduces the impact of credit onWTP by USD 3.73. Relative to an impact of credit on WTP of USD 12.58 on agents in the controlgroup,43 inattention contributes around 30 percent of the total impact of credit. This mechanismis economically meaningful: the large impact of credit is in part driven by inattentiveness to costswhen these are incurred in the future, rather than through relaxing credit constraints alone.

Attention to costs among the credit control group has a moderately positive impact (USD 2.34)on WTP, significant at the 10 percent level. These individuals may observe that costs will beincurred in only a single period, whereas benefits will be accrued over many periods. This maymake the adoption decision look more attractive.

There is suggestive evidence that this is linked to time-inconsistency as elicited through an efforttask allocation exercise that builds on Augenblick, Niederle, and Sprenger (2015) (see Section 4.5for more detail). While we unfortunately have too many violations of monotonicity with respect tothe interest rate to be confident in these results, an interesting pattern emerges when focusing onthe easiest decisions where the subjects faced a one-to-one trade-off in effort tasks. Columns (3)

43One may be concerned that highlighting the future costs may have changed respondents’ beliefs about thelikelihood of the enforcement of repayment. We think this is unlikely as the repeated visits by the team had alreadysignaled we were serious about following up with the respondents, and for those in the other attention treatmentswe still explained the repayment system before eliciting WTP. We also find no evidence of this ex-post as repaymentrates across the attention arms are indistinguishable.

20

and (4) of Table 5 show that both the effect of credit and the mediating impact of being remindedof future costs are stronger for agents exhibiting time inconsistency. Column (5) presents the tripleinteraction. This suggests that agents may be revising down effort tasks at visit two because theydo not attend to future costs rather than due to their time preferences. Future research shouldinvestigate this question more completely.

6.2.2 Concentration Bias

We test for concentration bias (Koszegi and Szeidl 2013, Dertwinkel-Kalt et al. 2019) by comparingWTP under weekly and monthly loan deadlines. Figure 8 separately displays the demand curves forrespondents in the monthly and weekly credit groups. Respondents paying with weekly deadlinesare willing to pay on average USD 1.29 more for the stove than those paying with monthly deadlines.While this is consistent with theory, it is economically small and not statistically significant. Thissuggests concentration bias is not at play in a meaningful way, and respondents are largely able tocorrectly perceive the size of costs, regardless of how these are presented to them.

6.3 Repayment and default

Respondents are free to choose the frequency and amount of each payment, but are required to meetcumulative minimums by the relevant deadlines. Respondents who miss a deadline are remindedvia repeated SMSes in the following days. Most respondents pay within three days of their officialdeadlines. By the end of the payment period, the median respondent who adopted the stove andwas in the credit treatment group had paid 98 percent of their price. 13 percent had paid less than10 percent. Repayment rates are reasonably high, and most default is not immediate, suggestingstrategic default likely does not drive adoption. Figure A17 shows repayment rates.

Still, default rates are high enough to warrant concern among potential lenders, and this mayexplain why access to credit is generally limited and costly in this context. In its annual reports,BRAC (one of the world’s largest microlenders) often reports repayment rates of around 98 percentin Bangladesh. However, they define repayment as having paid off the loan within one year ofits initial disbursement, regardless of the frequency or size of missed payments during that one-year period—shorter-term repayment rates are closer to 90 percent. BRAC’s Liberia, Sierra Leone,and Uganda offices reports 86, 82, and 93 percent repayment rates, respectively. Kenya’s AkibaMashinani Trust reports repayment rates of 90 percent for livelihood loans and 76 percent forhousing loans. Our results suggest that repayment rates in this study are roughly in line withrepayment rates among existing lending agencies.

The loans offered in this experiment charged an interest rate of r = 1.16 percent per month,or 3.5 percent over the three months. In practice lenders could charge higher interest to recoverlosses from these high default rates. Across all respondents who adopted the stove in the credittreatment groups, on average 72 percent of the loan was repaid. Thus, a lender would need tocharge an interest rate of 39.2 percent (such that 0.72 = 1

1+r ), on top of the market lending rate of1.16 percent, for a total interest rate of 40.1 percent.

21

We recalibrate the demand cure at this breakeven interest rate. We assume WTP is the max-imum total cost an individual is willing to accept in order to adopt the good. Then, for eachindividual we calculate the maximum price they would be willing to pay for the stove assumingthey had to pay an additional 33 percent on top of the purchase price in interest. For example, anindividual willing to spend USD 40 in total to obtain the stove is only willing to pay USD 30 as apurchase price if they must also pay 33 percent in interest. Figure A18 displays this recalibrateddemand curve. Using the new interest-excluding price as WTP, the impact of credit on WTP is71 percent relative to the control group, rather than the original increase of 104 percent when onlyfactoring in the market lending rate.

In the data, however, repayment rates are not constant: respondents with a higher price havelower repayment rates. The median respondent in the high subsidy group had paid 100 percent,whereas the median respondent in the low subsidy group had paid 67 percent. When controllingfor price, WTP is not correlated with repayment rate, suggesting price itself is the mechanism. Thehigh breakeven interest rate of 40.1 percent would likely cause higher default rates, in turn requiringa higher interest rate. Extrapolating observed patterns in this way reveals that there is no interestrate at which a lender could breakeven for any price above USD 15.

6.4 Risk aversion

Under uncertainty, risk aversion can reduce technology adoption (Oliva et al. 2020). Column (1) ofTable 6 demonstrates that, even after controlling for socioeconomic characteristics like income andbaseline savings, the WTP of agents exhibiting risk aversion is on average USD 2 lower.

This raises the question of whether part of the response to our loan is risk averse agents usingthe loan as a form of insurance. Participants in our study do not face any financial or other penaltiesfor payment delays, other than having to return the stove if they cannot continue the payments.44

Our credit may therefore be attractive to respondents who are risk averse, in which case the impactof credit on WTP would be larger for risk averse agents. To the contrary, Column (2) of Table 6indicates that risk aversion does not affect the impact of credit. The low risk of credit may be lessimportant in this context because respondents in the credit control group also had the option toreturn the stove at no cost under Burn’s warranty program.

7 Policy implications

The International Energy Agency (2018) estimates that 44 percent of all global emissions reductionsby 2040 could come from energy efficiency gains. The energy efficient technology in this paper isprivately profitable, and households attend to these savings, but credit constraints prohibit adoptionfor most agents. These results have implications for the effectiveness of climate policies in low-incomecontexts, including carbon taxes, nudges, and subsidies for energy efficient technologies.

44Carney et al. (2018) call this a new-asset collateralized. Once agents adopt, their reference point changes andrepayment increases as the endowment effect applies to the new asset. Their predictions line up with what we observe.

22

7.1 Carbon taxes and technology subsidies

In a first-best setting, the efficient solution to negative externalities is a Pigovian tax on the emittinggood (Pigou 1920). Low-income country governments are increasingly implementing carbon taxesto reduce emissions of greenhouse gases and local environmental pollutants. For example, SouthAfrica, Chile, and Mexico have all enacted a carbon tax since 2014, each covering at least 40 percentof domestic greenhouse gas emissions (World Bank 2020).

An agent facing binding credit constraints cannot respond optimally to the incentives set bythe tax. A carbon tax is therefore unlikely to achieve optimal abatement. Worse, a carbon taxwould have important equity implications by increasing energy prices, in particular for those withthe tightest credit constraints, who are often the poorest and therefore also have the highest energyburden. Most likely, a carbon tax would be regressive.

Instead, combining a smaller tax on the emitting good with a subsidy on the energy efficienttechnology would generate larger welfare gains. The relative sizes of the tax and the subsidy willdepend on the size of the externality, the size of credit constraints, and the correlation betweenusage and credit constraints, as this affects targeting. By lowering the cost of the energy efficienttechnology at adoption, a subsidy targets credit constrained agents more effectively than a tax.Factoring in private savings and avoided environmental damages, a subsidy for the energy efficientcookstove would generate USD 19 of welfare gains for every USD 1 of government expenditure.

7.2 Nudges

Given the lack of response to an intense intervention designed to draw attention to potential energysavings, it is unlikely that adoption can be promoted through lighter-touch nudges designed to drawattention to savings. This does not mean that all nudges are unlikely to be helpful. More context-specific nudges, for example decision-aids to assist with optimizing the allocation of credit, may bebeneficial. Future work should examine these alternatives.

7.3 Expanding access to credit

Expanding access to credit might also increase adoption, but this depends on the source of thecredit constraints. On the one hand, if lenders are engaging in quantity rationing and individuals areoptimally allocating the credit they have access to, then expanding access to credit will likely increaseadoption given the high rate of return. On the other hand, if households are credit constrained dueto self-control problems, then expanding the quantity of credit may not increase adoption. In thisworld, households may choose to use the additional credit to smooth the shocks to consumption thatthey face due to their inability to maintain a buffer stock of savings (Kremer, Rao, and Schilbach2019).

A related question is why profit maximizing firms are not offering product-specific credit. As inthe experiment, these targeted expansions of credit would sidestep concerns about misallocation ofcredit. Informal conversations with decision makers in the sector suggest that of primary importance

23

are fears of over-extension if technology firms expand into the credit sector. The primary strengthof energy efficient technology companies is developing and marketing these technologies—extendinginto activities beyond this scope may jeopardize the quality of those products. Furthermore, a largegap exists between the formal and informal sectors. A manufacturing company interested in offeringcredit to its customers may be more likely to partner with an existing formal banking institution,but the population studied in this paper are almost entirely served by informal financial providers.

7.4 The policy interpretation of willingness-to-pay

Policy-makers and researchers often infer welfare gains of a product or intervention from beneficiaryWTP. In the context of environmental and health economics, WTP is often used to value environ-mental attributes or individual health outcomes. These valuations are often relied upon by policymakers in setting policy priorities.

However, market frictions may create a wedge between WTP and ability-to-pay (ATP). Banerjee(1997) discusses how credit constraints can increase the gap between WTP and ATP and exacerbatered tape in the context of government bureaucracy. Given how constrained the subjects in our ex-periment are as measured both by their response to a three month loan and our back-of-the-envelopecalculations of private benefits, revealed preference methods may substantially underestimate re-alized welfare gains. These results should caution policy-makers and researchers from relying onWTP measures for welfare estimates in this and similar settings.

8 Conclusion

In high-income contexts, carbon taxes and nudges to increase attention are effective policy toolsto increase adoption of energy efficient technologies. Yet, these tools may not be effective in lowerincome contexts where credit constraints can prevent the adoption of even privately beneficial tech-nologies, and the levels of attention to potential energy savings may be very different. We studythis question in the context of an energy efficient household technology in Nairobi, Kenya.

We estimate that the technology reduces household charcoal spending by 40 percent, saving theaverage household USD 118 per year. At a retail price of USD 40, this corresponds to an IRRof 300 percent per year. More than 60 percent of respondents report using the savings for criticalhousehold expenditures such as food items and child school fees. This means governments looking toreduce poverty by increasing household adoption of profitable technologies may find that providingfinancing can provide tangible opportunities for welfare gains for poor households. The stove alsogenerates significant carbon emissions reductions of USD 146 per year—but importantly, the privatebenefits alone justify adoption. Despite this, we identify significant under-adoption: participants inthe control group are only willing to pay USD 12 for the stove.

We then investigate whether drawing attention to potential energy savings increases adoption.We find that an intense intervention designed to increase attention to energy savings had no effecton WTP. This is in contrast to many papers in the energy literature and the development literature

24

that would predict significant behavioral biases, particularly for energy efficiency adoption decisionsamong low-income households. This may be due to the decision’s high financial consequences: themedian respondent saves one month of income per year. There is modest evidence in the literaturethat when stakes are higher, cognitive performance among the poor improves. It may also be thatenergy expenditures are easier to track when inputs and outputs are strongly correlated—charcoalusage is relatively easy to track when its sole usage is for charcoal cookstoves. This is analogousto gasoline usage to power a vehicle, and may explain why our findings align with some evidenceshowing households correctly evaluate costs against future gas prices when deciding whether topurchase a more energy efficient vehicle.

We next consider whether households are prevented from adopting due to credit constraints.Access to credit more than doubles WTP for the stove, suggesting households are significantlycredit constrained. Carbon taxes may not achieve optimal levels of emissions abatement whenpotential adopters are prevented from adopting through liquidity constraints. The first best policywould be to increase energy efficiency through credit, but this may be difficult to implement inpractice due to information asymmetry and adverse selection in credit markets, and the informalnature of many low-income economies. We also show that inattention to future periods can affectthe use of credit, suggesting simply expanding access to credit may not lead to the efficient outcome.

Instead, a subsidy on energy efficient technologies can effectively target credit constrained agentsby lowering the upfront cost of adoption. More broadly, we emphasize that more research is neededto better understand energy constraints in low-income contexts, and that environmental policiesthat have been studied primarily in high-income contexts must recognize how structural differencescan affect policy efficacy.

Low and middle-income countries are expected to propel future energy demand. Energy effi-ciency is often touted as a technology that can benefit households financially while also reducing car-bon emissions, yet adoption remains low. We illustrate that policy makers cannot rely on householdsto adopt privately cost-saving energy efficient technologies, and that a carbon tax may therefore notachieve efficient levels of adoption. Households in our study would like to adopt more energy efficientversions of their primary energy durable but are unable to do so due to credit constraints. If policymakers were able to address the underlying credit constraints, or provide financing or subsidies forhigh-return energy efficient technologies, this would allow low-income households to take advantageof technologies that are already available to them. This paper shows that these interventions havethe potential to improve environmental outcomes and generate significant benefits for the poor.

25

Figures

Figure 1: Traditional jiko (‘stove’) and energy efficient stove

On the left is the traditional jiko. On the right is the energy efficient stove. The two stoves use the same type ofcharcoal and the same process for cooking food, hence the energy efficient stove requires essentially no learning toadopt. After usage, the user disposes of the ash using the tray at the bottom. The central chamber of the energyefficient stove is constructed using insulating materials, creating a higher charcoal-to-heat conversion rate. Engineersex-ante predict that the energy efficient stove uses only half the charcoal to reach and maintain the same cookingtemperatures as the traditional jiko.

Figure 2: Timeline

Visit 1- Enrollment

Visit 2- Treatments- BDM elicitation- Stove transaction

Visit 3- Endline survey

Phone survey- Long-term

Endline survey

Loan payments (C1, C2)

Charcoal SMS survey

Charcoal SMS survey (A1, A2)

Commuting SMS survey (A0)

Charcoal ash collection in buckets

Charcoal SMS survey

t = 0t = - 30 t = 30 t = 60 t = 90 t = 365 t = 425

Timeline of the four main study components: 1) Three in-person visits, timed one month after each other, and a1-year endline phone survey; 2) A recurring SMS survey about charcoal spending (a control group received placeboSMSes about an unrelated topic (their commuting time) for the first month); 3) Ash collection in buckets for onemonth, to measure charcoal consumption; 4) Loan payments (for respondents who purchased the stove and used aloan to do so).

26

Figure 3: Experimental Treatment Arms

Credit Control

Credit Treatment

Weekly Deadlines

Monthly Deadlines

Attention Control

Attention Treatment

Energy Savings

Energy Savings – Costs

96 98 98

96 97 96

145 146 146

We enroll 1,018 respondents and randomly assign them to one of three credit treatments and one of three attentiontreatments. Respondents in the credit control group must pay for the stove during visit 2 and receive the stove thatday. Respondents in the credit treatment group still receive the stove during visit 2 but pay for it over 3 months.Respondents in the attention control group receive basic information about the stove. Respondents in the attentiontreatment group are prompted to report charcoal spending every three days in the month before WTP is elicited, toforecast 12 months of savings and spend time thinking about how they could use the savings. Respondents in thetreatment to costs group also think through costs associated with adoption. Treatment assignment is stratified bybaseline charcoal spending.

Figure 4: Energy efficient stoves reduce energy spending

0

1

2

3

4

5

6

7

8

9

10

Wee

kly

ch

arco

al s

pen

din

g (

US

D)

−21 −14 −7 0 7 14 21 28 35 42 49 56 63Days since main visit

Adopted stove

Did not adopt

95% CI

Weekly charcoal spending by adopters and non-adopters of the energy efficient stove before and after the main visit(visit 2). Charcoal spending is elicited through a recurring 3-day SMS survey. Adoption of the stove causes charcoalexpenditures to drop by USD 2.27 per week (40 percent relative to the control group). The causal estimates presentedin Figure A10 are similar.

27

Figure 5: Attrition of SMSes by adoption and treatment status

0.2

5.5

.75

1F

ract

ion

SM

Ses

res

po

nd

ed

0 20 40 60Days since Visit 2

High BDM price (> USD 15)Low BDM price (< USD 15)

0.2

5.5

.75

1F

ract

ion

SM

Ses

res

po

nd

ed

0 20 40 60Days since Visit 2

Did not adoptAdopted Jikokoa

0.2

5.5

.75

1F

ract

ion

SM

Ses

res

po

nd

ed

0 20 40 60Days since Visit 2

Attention controlAttention treatment

0.2

5.5

.75

1F

ract

ion

SM

Ses

res

po

nd

ed

0 20 40 60Days since Visit 2

Credit controlCredit treatment

We test for attrition in SMS responses by Jikokoa adoption and treatment status for all three treatments (credit,attention, and subsidy). Table A4 provides results from statistical tests that confirm balanced attrition across allthree treatment groups, adoption status, and most socioeconomic characteristics. We test for attrition in more detailin the recurring SMS survey. Of 955 respondents who completed visit 2, 838 (88 percent) responded correctly to atleast one SMS over the course of the study. Among these respondents, we received correct responses to 44 percent ofpost-adoption charcoal SMSes. The composition of respondents varies across SMS cycles: some responded to manySMSes while others responded to only a few. Figure A19 presents a histogram of SMSes each respondent correctlyresponded to during the first 48 days after visit 2. 446 (46 percent) responded to at least half of all SMSes. Duringthe SMS survey conducted one year after the main experiment, more than 75 percent of respondents who completedvisit 2 responded to at least one SMS, and the median respondent responded to six.

28

Figure 6: Under-adoption of the energy efficient technology

010

2030

4050

US

D

0 .2 .4 .6 .8 1

Q

Breakeven demand curve (3−month horizon)

Pure control demand curve

Pure control histogram of WTP

The histogram represents WTP elicited through the BDM mechanism for the control group, and the smooth linerepresents the corresponding demand curve. The dotted line represents the breakeven demand curve for all agents,if agents were willing to pay precisely their savings over a 3-month period. The gap between the two curves can beinterpreted as under-adoption of the energy efficient technology. The breakeven demand curve assumes annualizeddiscount rates δ = 0.9. Figure A15 presents robustness checks for annual discount factors δ = 0.5 and δ = 1.

29

Figure 7: Impacts of experimental treatments on WTP

Panel A

0

10

20

30

40

50

US

D

0 .2 .4 .6 .8 1

Q

No credit, no attention

Credit treatment, no attention

Panel B

0

10

20

30

40

50

US

D

0 .2 .4 .6 .8 1

Q

No credit, no attention

No credit, attention treatment

Graphs show the cumulative distribution of WTP for the control and treatment groups for both experimental treat-ments. Panel A presents results by credit treatment status among people in the attention control group only. PanelB presents results by attention treatment status among people in the credit control group only. Access to credit in-creases WTP by USD 13 (104 percent relative to control). Attention to benefits does not affect WTP. The breakevendemand curve assumes annualized discount rates δ = 0.9. Figure A15 presents robustness checks for annual discountfactors δ = 0.5 and δ = 1.

30

Figure 8: Test of concentration bias

0

10

20

30

40

50

US

D

0 .2 .4 .6 .8 1Q

No Credit

Credit (Weekly)

Credit (Monthly)

An agent exhibiting concentration bias would prefer payment plans where costs are dispersed across a larger numberof smaller payments rather than concentrated in a smaller number of larger payments. We test for this by randomlyassigning respondents to pay by either weekly or monthly deadlines. WTP is presented for the two credit treatmentsseparately. Credit increases WTP by 13 USD on average but the framing of deadlines (weekly or monthly) does notaffect adoption, suggesting concentration bias is not at play.

31

Tables

Table 1: Summary statistics

Mean SD 25th 50th 75th

Household size 4.73 2.08 3 4 6Age 37.24 11.83 29 35 44Female respondent 0.95 0.21 1 1 1Completed primary education 0.70 0.46 0 1 1Completed secondary education 0.26 0.44 0 0 1Respondent income (USD/week) 24.26 25.25 11 21 35Household income (USD/week) 47.79 39.25 21 35 60Energy spending (USD/week) 8.60 3.89 6 8 10Charcoal spending (USD/week) 5.65 2.93 4 5 7Savings (USD) 74.83 129.45 1 30 82Current cookstove price (USD) 3.46 1.93 3 3 4

Summary statistics of key socioeconomic characteristics for all 1,018 study participants. ‘Savings’ includes savings inbank account, mobile money account, or informal group savings.

Table 2: Causal impact of stove adoption on weekly charcoal spending

OLS First StageIV Estimate (1-month

endline)IV Estimate (1-year

endline)(1) (2) (3) (4) (5) (6) (7)USD Bought Stove USD IHS(USD) IHS(KG) USD IHS(USD)

BDM Price (USD) 0.003 -0.029∗∗∗

(0.013) (0.001)

WTP (USD) -0.003 0.024∗∗∗ -0.003 -0.001 0.003 0.008 0.004(0.010) (0.001) (0.011) (0.003) (0.003) (0.016) (0.003)

Bought Cookstove (=1) -1.939∗∗∗ -2.273∗∗∗ -0.493∗∗∗ -0.479∗∗∗ -2.423∗∗∗ -0.539∗∗∗

(0.294) (0.295) (0.071) (0.083) (0.427) (0.091)

Observations 7870 913 7870 7870 796 6982 6982Control Mean 5.717 4.964 2.154 1.545 5.303 2.212Socioeconomic controls Yes Yes Yes Yes Yes Yes YesData Source SMSes Midline SMSes SMSes Buckets SMSes SMSes

Results from an instrumental variables regression that uses the (randomly assigned) BDM price as an instrumentfor stove adoption to estimate the causal impact of adoption on weekly charcoal expenditures. Columns (1) and (2)present the OLS and first-stage estimates, respectively. Column (3) uses weekly charcoal expenditures in USD asthe outcome variable. Column (4) uses the inverse hyperbolic sine (IHS) conversion of the USD amount. A 0.498IHS reduction corresponds to a 40 percent reduction relative to the control group. Column (5) uses the IHS of theweight of the charcoal bucket one month after stove adoption as the outcome variable. Columns (6) and (7) conductthe same analyses as columns (3) and (4) respectively, but using data from the SMS survey conducted one yearafter the main visit. Socioeconomic controls include baseline savings, income, risk aversion, credit constrainedness,number of adults and children. In regressions using SMS data, errors are clustered by respondent. SE in parentheses.∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

32

Table 3: Empirical rate of return estimates from selected literature

Authors Year CountryAnnualized

IRR

Berkouwer and Dean 2019 Kenya 294%

Energy EfficiencyAllcott and Greenstone 2017 USA -4%Fowlie, Greenstone, Wolfram 2018 USA -10%–0%Davis, Martinez, Taboada 2018 Mexico less than -8%

FirmsBigsten, Isaksson, Soderbom, et al. 2000 Africa1 10–35%McKenzie and Woodruff 2006 Mexico 36–180%McKenzie and Woodruff 2008 Mexico 240–396%De Mel, McKenzie, Woodruff 2008 Sri Lanka 55–63%Kremer, Lee, Robinson 2013 Kenya 113%Fafchamps, McKenzie, Quinn, Woodruff 2014 Ghana 180%Banerjee and Duflo 2014 India 105%Blattman, Fiala, Martinez 2014 Uganda 30–50%Blattman, Green, Jamison, et al. 2016 Uganda 8–24%

AgricultureUdry and Anagol 2006 Ghana 30–50%Duflo, Kremer, Robinson 2008 Kenya 52–85%

EducationBigsten, Isaksson, Soderbom, et al. 2000 Africa1 1–5%Duflo 2001 Indonesia 8.8–12%

OtherBaird, Hicks, Kremer, Miguel2 2016 Kenya 32%Haushofer, Shapiro3 2016 Kenya 15%

1Cameroon, Ghana, Kenya, Zambia, Zimbabwe. 2Deworming. 3Unconditional cash transfers. Thistable shows annualized internal rate of return (IRR) estimates from recent literature. The IRRcorresponds to the interest rate where the Net Present Value of an investment, from time 0 toinfinity, equals 0. Conversion between monthly and annualized IRR conservatively (and in line withthe literature) assumes no reinvestment.

33

Non-monetary impacts

Table 4: Non-monetary outcomes: Drivers and impact of stove adoption

WTP(USD)

Minutescooking per day

Adoptionsin network

Health SymptomsIndex (1-month follow-up)

Health SymptomsIndex (1-year follow-up)

(1) (2) (3) (4) (5) (6) (7) (8)

Health beliefs (index) -0.064(0.630)

Savings beliefs (USD) 0.016∗∗(0.008)

Jikokoa (=1) -53.866∗∗∗ -0.176 -0.524∗∗∗ -0.563∗∗∗ -0.512∗∗∗ -0.557∗∗∗ -0.562∗∗∗(14.582) (0.166) (0.105) (0.111) (0.118) (0.094) (0.097)

Continued old stove use (=1) 0.171∗ 0.149∗ 0.050(0.088) (0.089) (0.097)

Charcoal usage (KG/month) 0.038∗∗(0.017)

Observations 913 913 913 913 913 913 855 855Control Mean 11.879 192.093 0.318 -0.000 -0.000 -0.000 0.000 0.000Socioeconomic controls Yes Yes Yes Yes Yes Yes Yes Yes

Column (1) tests whether baseline beliefs about financial and health benefits affect WTP. Columns (2) through (6) present causal estimates of theimpact of stove adoption on various outcomes measured one month after adoption, using the randomly assigned price as an instrument for adoption.Adoptions in network indicates whether any of the respondent’s friends, family, or neighbors purchased the Jikokoa in the past 1 month. The healthsymptoms index measures self-reported respiratory outcomes for adults and children normalized such that the control group has a mean of 0 anda standard deviation of 1. Columns (7) and (8) report results from the same IV regression but using a health symptoms index measured one yearafter adoption. Socioeconomic controls include baseline savings, income, risk aversion, credit constrainedness, number of adults and children. SE inparentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

34

Table 5: Inattention to Future Costs Mediates Impact of Credit

WTP(USD)

(1) (2) (3) (4) (5)

Credit 12.58∗∗∗ 12.54∗∗∗ 10.44∗∗∗ 14.66∗∗∗ 10.20∗∗∗

(0.69) (1.27) (1.90) (1.73) (1.51)

Attention to benefits 0.40 -1.11 -0.98 -0.60 -0.74(0.86) (1.48) (2.30) (1.96) (1.48)

Attention to costs -0.12 2.34∗ 2.85 1.82 2.24∗

(0.79) (1.36) (2.24) (1.69) (1.35)

Attention to benefits X Credit 2.30 1.77 1.97 1.94(1.81) (2.87) (2.37) (1.81)

Attention to costs X Credit -3.73∗∗ -2.27 -4.67∗∗ -1.86(1.67) (2.71) (2.10) (1.91)

Time inconsistent -2.52∗∗

(1.13)

Time inconsistent X Credit 4.48∗∗∗

(1.55)

Attention to costs X Time inconsistent X Credit -3.07∗

(1.63)

Observations 955 955 411 544 955Control Mean 12.12 12.12 13.14 10.98 13.14Sample Full Full TI=0 TI=1 Full

This table shows the causal impact of credit (pooling the two credit treatment arms) and attention treatments on willingness-to-pay (WTP) elicited during theBecker-DeGroot-Marschak (BDM) mechanism. For the ‘attention to benefits’ treatment, the indicator variable ‘Attention to benefits’ is set to 1 and the indicatorvariable ‘Attention to costs’ is set to 0. For the ‘attention to benefits minus costs’ treatment, both indicator variables are set to 1. Agents are defined as exhibitingtime inconsistency (TI=1) if they choose to postpone effort tasks during visit 2. Socioeconomic controls include baseline savings, income, risk aversion, creditconstrainedness, number of adults and children. SE in parentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

35

Table 6: Risk averse agents have lower adoption, but respond similarly to credit

WTP(USD)

(1) (2)

Credit 12.58∗∗∗ 12.61∗∗∗

(0.69) (1.22)

Risk aversion -2.02∗∗∗ -2.00∗

(0.71) (1.21)

Risk aversion X Credit -0.04(1.47)

Belief about stove durability (years) 0.41∗∗ 0.41∗∗

(0.16) (0.16)

Observations 955 955Control Mean 12.12 12.12Sample All All

This table shows results from a regression estimating how risk aversion and beliefs about stove durability affectWTP. Risk aversion affects WTP directly but does not meaningfully affect the impact of credit. Socioeconomiccontrols include baseline savings, income, risk aversion, credit constrainedness, number of adults and children. SE inparentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

References

African Population and Health Research Center. 2014. “Population and Health Dynamics in Nairobi’sInformal Settlements: Report of the Nairobi Cross-sectional Slums Survey”.

Allcott, Hunt, and Michael Greenstone. 2012. “Is There an Energy Efficiency Gap?” Journal ofEconomic Perspectives 26 (1): 3–28.

— . 2017. “Measuring the welfare effects of residential energy efficiency programs”. NBER WorkingPaper #23386.

Allcott, Hunt, and Dmitry Taubinsky. 2015. “Evaluating Behaviorally Motivated Policy: Experi-mental Evidence from the Lightbulb Market”. American Economic Review 105 (8): 2501–2538.

Allcott, Hunt, and Nathan Wozny. 2014. “Gasoline Prices, Fuel Economy, and the Energy Paradox”.The Review of Economics and Statistics 96 (5): 779–795.

Anthoff, David, Cameron Hepburn, and Richard S.J. Tol. 2009. “Equity weighting and the marginaldamage costs of climate change”. Ecological Economics 68 (3): 836–849.

Ashden Clean Energy for Women and Girls Award, supported by UK aid. 2015. Winner case studysummary: BURN Manufacturing, Kenya.

Augenblick, Ned, Muriel Niederle, and Charles Sprenger. 2015. “Working over Time: Dynamic In-consistency in Real Effort Tasks”. The Quarterly Journal of Economics 130 (3): 1067–1115.

Bailis, Robert, Rudi Drigo, Adrian Ghilardi, and Omar Masera. 2015. “The carbon footprint oftraditional woodfuels”. Nature Climate Change 5:266–272.

Baird, Sarah, Joan Hamory Hicks, Michael Kremer, and Edward Miguel. 2016. “Worms at Work:Long-run Impacts of a Child Health Investment”. Quarterly Journal of Economics 131 (4): 1637–1680.

36

Balakrishnan, Uttara, Johannes Haushofer, and Pamela Jakiela. 2020. “How soon is now? Evidenceof present bias from convex time budget experiments”. Experimental Economics 23 (2): 294–321.

Banerjee, Abhijit V. 1997. “A Theory of Misgovernance”. The Quarterly Journal of Economics 112(4): 1289–1332.

Banerjee, Abhijit V., and Esther Duflo. 2005. “Growth Theory through the Lens of DevelopmentEconomics”. Chap. 7 in Handbook of Economic Growth, ed. by Philippe Aghion and StevenDurlauf, 1:473–552.

Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman. 2015. “Six Randomized Evaluations ofMicrocredit: Introduction and Further Steps”. American Economic Journal: Applied Economics7 (1): 1–21.

Becker, Gordon M., Morris H. Degroot, and Jacob Marschak. 1964. “Measuring utility by a single-response sequential method”. Systems Research and Behavioral Science 9 (3): 226–232.

Beltramo, Theresa, and David I. Levine. 2013. “The effect of solar ovens on fuel use, emissions andhealth: results from a randomised controlled trial”. Journal of Development Effectiveness 5 (2):178–207.

Berry, James, Greg Fischer, and Raymond P. Guiteras. 2020. “Eliciting and Utilizing Willingnessto Pay: Evidence from Field Trials in Northern Ghana”. Journal of Political Economy 128 (4):1436–1473.

Bharadwaj, Prashant, William Jack, and Tavneet Suri. 2019. “Fintech and Household Resilience toShocks: Evidence from Digital Loans in Kenya”. NBER Working Paper #25604.

Bhattacharya, S.C., D.O. Albina, and P. Abdul Salam. 2002. “Emission factors of wood and charcoal-fired cookstoves”. Biomass and Bioenergy, no. 23: 453–469.

Bigsten, Arne, Anders Isaksson, Mans Soderbom, Paul Collier, Albert Zeufack, Stefan Dercon,Marcel Fafchamps, Jan Willem Gunning, Francis Teal, Simon Appleton, Bernard Gauthier,Abena Oduro, Remco Oostendorp, and Cathy Pattillo. 2000. “Rates of return on physical andhuman capital in Africa’s manufacturing sector”. Economic Development and Cultural Change48 (4): 801–827.

Blattman, Christopher, Nathan Fiala, and Sebastian Martinez. 2014. “Generating skilled self-employmentin developing countries: Experimental evidence from Uganda”. Quarterly Journal of Economics129 (2): 697–752.

Blattman, Christopher, Eric P. Green, Julian Jamison, M. Christian Lehmann, and Jeannie Annan.2016. “The Returns to Microenterprise Support among the Ultrapoor: A Field Experiment inPostwar Uganda”. American economic journal: Applied economics 8 (2): 35–64.

Borenstein, Severin. 2015. “A Microeconomic Framework for Evaluating Energy Efficiency ReboundAnd Some Implications”. Energy Journal 36 (1): 1–22.

Brownback, Andy, Alex Imas, and Michael Kuhn. 2019. “Behavioral Food Subsidies”.Burbidge, John B., Lonnie Magee, and A. Leslie Robb. 1988. “Alternative Transformations to Handle

Extreme Values of the Dependent Variable”. Journal of the American Statistical Association 83(401): 123–127.

Burlig, Fiona, Christopher Knittel, David Rapson, Mar Reguant, and Catherine Wolfram. 2020.“Machine learning from schools about energy efficiency”. Journal of the Association of Environ-mental and Resource Economists 7 (6): 1181–1217.

Burwen, Jason, and David I. Levine. 2012. “A Rapid Assessment Randomized-Controlled Trial ofImproved Cookstoves in Rural Ghana”. Energy of Sustainable Development 16 (3): 328–338.

Carney, Kevin, Xinyue Lin, Michael Kremer, and Gautam Rao. 2018. “The endowment effect andcollateralized loans”. Working paper.

37

Chan, Nathan W., and Kenneth Gillingham. 2015. “The Microeconomic Theory of the ReboundEffect and Its Welfare Implications”. Journal of the Association of Environmental and ResourceEconomists 2 (1): 133–159.

Charness, Gary, Uri Gneezy, and Alex Imas. 2013. “Experimental methods: Eliciting risk prefer-ences”. Journal of Economic Behavior and Organization 87:43–51.

Chernozhukov, Victor, Mert Demirer, Esther Duflo, and Ivan Fernandez-Val. 2018. “Generic MachineLearning Inference on Heterogenous Treatment Effects in Randomized Experiments”. NBERWorking Paper 24678: 1–44.

Chowdhury, Sourangsu, Sagnik Dey, Sarath Guttikunda, Ajay Pillarisetti, Kirk R. Smith, and LarryDi Girolamo. 2019. “Indian annual ambient air quality standard is achievable by completelymitigating emissions from household sources”. Proceedings of the National Academy of Sciencesof the United States of America 116 (22): 10711–10716.

Christensen, Peter, Paul Francisco, Erica Myers, and Mateus Souza. 2019. “Decomposing the Wedge:Mechanisms driving the gap between projected and realized returns in energy efficiency pro-grams”. Working Paper.

Cubitt, Robin, and Daniel Read. 2007. “Can intertemporal choice experiments elicit time preferencesfor consumption?” Experimental Economics 10 (4): 369–389.

Davis, Lucas W., Sebastian Martinez, and Bibiana Taboada. 2020. “How Effective is Energy-efficient Housing? Evidence From a Field Trial in Mexico”. Journal of Development Economics143:102390.

De Groote, Olivier, and Frank Verboven. 2019. “Subsidies and Time Discounting in New TechnologyAdoption: Evidence from Solar Photovoltaic Systems”. American Economic Review 109 (6):2137–2172.

Dean, Joshua T. 2020. “Noise, Cognitive Function, and Worker Productivity”. Mimeo.Dean, Mark, and Anja Sautmann. 2019. “Credit Constraints and the Measurement of Time Prefer-

ences”. Forthcoming. The Review of Economics and Statistics.Delavande, Adeline. 2014. “Probabilistic Expectations in Developing Countries.” Annual Review of

Economics 6 (1): 1–20.Delavande, Adeline, Xavier Giné, and David McKenzie. 2011a. “Eliciting Probabilistic Expectations

with Visual Aids in Developing Countries: How Sensitive Are Answers to Variation in ElicitationDesign?” Journal of Applied Econometrics 26:479–97.

— . 2011b. “Measuring Subjective Expectations in Developing Countries: A Critical Review andNew Evidence”. Journal of Development Economics 94 (2).

Dertwinkel-Kalt, Markus, Holger Gerhardt, Gerhard Riener, Frederik Schwerter, and Louis Strang.2019. “Concentration bias in intertemporal choice”. Under revision.

Drehobl, Ariel, and Lauren Ross. 2016. “Lifting the High Energy Burden in America’s Largest Cities:How Energy Efficiency Can Improve Low Income and Underserved Communities”. Report byEnergy Efficiency For All and the American Council for an Energy-Efficient Economy.

Duflo, Esther. 2001. “Schooling and Labor Market Consequences of School Construction in Indone-sia: Evidence from an Unusual Policy Experiment”. American Economic Review 91 (4): 795–813.

Duflo, Esther, Michael Kremer, and Jonathan Robinson. 2008. “How High Are Rates of Returnto Fertilizer? Evidence from Field Experiments in Kenya”. American Economic Review 98 (2):482–488.

— . 2011. “Nudging Farmers to Use Fertilizer: Theory and Experimental Evidence from Kenya”.American Economic Review 101 (6): 2350–2390.

Dupas, Pascaline. 2009. “What Matters (and What Does Not) in Households’ Decision to Invest inMalaria Prevention?” American Economic Review 99 (2): 224–230.

38

Fafchamps, Marcel, David McKenzie, Simon Quinn, and Christopher Woodruff. 2014. “Microenter-prise growth and the flypaper effect: Evidence from a randomized experiment in Ghana”. Journalof Development Economics 106:211–226.

Fehr, Dietmar, Gunther Fink, and B. Kelsey Jack. 2019. “Poverty, Seasonal Scarcity and ExchangeAsymmetries”. NBER Working Paper #26357.

Food and Agriculture Organization of the United Nations. 1987. “Simple technologies for charcoalmaking”. Forestry Paper 41.

— . 2017. “The Charcoal Transition”.Fowlie, Meredith, Michael Greenstone, and Catherine Wolfram. 2018. “Do Energy Efficiency Invest-

ments Deliver? Evidence from the Weatherization Assistance Program”. Quarterly Journal ofEconomics 133 (3): 1597–1644.

Frölich, Markus, and Blaise Melly. 2013. “Unconditional Quantile Treatment Effects Under Endo-geneity”. Journal of Business and Economic Statistics 31 (3): 346–357.

Gabaix, Xavier, and David Laibson. 2017. “Myopia and Discounting”. NBERWorking Paper #23254.Gerarden, Todd D., Richard G. Newell, and Robert N. Stavins. 2017. “Assessing the Energy-

Efficiency Gap”. Journal of Economic Literature 55, no. 4 (): 1486–1525.Gillingham, Kenneth, Sebastien Houde, and Arthur van Benthem. 2019. “Consumer Myopia in

Vehicle Purchases: Evidence from a Natural Experiment”. Forthcoming, American EconomicJournal: Economic Policy.

Gillingham, Kenneth, and Karen Palmer. 2014. “Bridging the Energy Efficiency Gap: Policy Insightsfrom Economic Theory and Empirical Evidence”. Review of Environmental Economics and Policy8 (1): 18–38.

Gillingham, Kenneth, David Rapson, and Gernot Wagner. 2015. “The Rebound Effect and EnergyEfficiency Policy”. Review of Environmental Economics and Policy 10 (1): 68–88.

Gneezy, Uri, and Jan Potters. 1997. “An experiment on risk taking and evaluation periods”. TheQuarterly Journal of Economics 112 (2): 631–645.

Goldin, Jacob, and Tatiana Homonoff. 2013. “Smoke Gets in Your Eyes: Cigarette Tax Salience andRegressivity.” American Economic Journal: Economic Policy 1 (5): 302–236.

Hanna, Rema, Esther Duflo, and Michael Greenstone. 2016. “Up in Smoke: The Influence of House-hold Behavior on the Long-Run Impact of Improved Cooking Stoves”. American Economic Jour-nal: Economic Policy 8 (1): 80–114.

Hanna, Rema, and Paulina Oliva. 2015. “Moving Up the Energy Ladder: The Effect of an Increasein Economic Well-Being on the Fuel Consumption Choices of the Poor in India”. AmericanEconomic Review Papers and Proceedings 105 (5): 242–246.

Haushofer, Johannes, and Ernst Fehr. 2014. “On the psychology of poverty”. Science 344 (6186):862–867.

Iiyama, Miyuki, Audrey Chenevoy, Erick Otieno, Teddy Kinyanjui, Geoffrey Ndegwa, Jan Vanden-abeele, Mary Njenga, and Oliver Johnson. 2014. “Achieving sustainable charcoal in Kenya: Har-nessing the opportunities for cross-sectoral integration”. World Agroforestry Center and Stock-holm Environment Institute.

International Energy Agency. 2018. “Market Report Series: Energy Efficiency 2018”.Jessoe, Katrina, and David Rapson. 2014. “Knowledge Is (Less) Power: Experimental Evidence from

Residential Energy Use”. American Economic Review 104 (4): 1417–1438.Jevons, William Stanley. 1866. “The Coal Question”. Macmillan and Co.Kenya National Bureau of Statistics. 2018. “Basic Report: Kenya Integrated Household Budget

Survey”.Koszegi, Botond, and Adam Szeidl. 2013. “A Model of Focusing in Economic Choice”. The Quarterly

Journal of Economics 128 (1): 53–104.

39

Kremer, Michael, Jean N. Lee, Jonathan M. Robinson, and Olga Rostapshova. 2013. “The Returnto Capital for Small Retailers in Kenya: Evidence from Inventories”. Mimeo, Harvard.

Kremer, Michael, Gautam Rao, and Frank Schilbach. 2019. “Behavioral development economics”.Handbook of Behavioral Economics 2:345–458.

Levine, David, Theresa Beltramo, Garrick Blalock, Carolyn Cotterman, and Andrew M. Simons.2018. “What Impedes Efficient Adoption of Products? Evidence from Randomized Sales Offersfor Fuel-Efficient Cookstoves in Uganda”. Journal of the European Economic Association 16 (6):1850–1880.

Liu, Elaine M. 2013. “Time to change what to sow: Risk preferences and technology adoptiondecisions of cotton farmers in China”. Review of Economics and Statistics 95 (4): 1386–1403.

McKenzie, David, and Christopher Woodruff. 2006. “Do Entry Costs Provide an Empirical Basis forPoverty Traps? Evidence from Mexican Microenterprises”. Economic Development and CulturalChange 55 (1): 3–42.

— . 2008. “Experimental Evidence on Returns to Capital and Access to Finance in Mexico”. WorldBank Economic Review 22 (3): 457–482.

McKinsey & Company. 2009. Pathways to a low carbon economy: Version 2 of the Global GreenhouseGas Abatement Curve.

Mel, Suresh de, David McKenzie, and Christopher Woodruff. 2008. “Returns to Capital in Microen-terprises: Evidence from a Field Experiment”. The Quarterly Journal of Economics 123 (4):1329–1372.

Mobarak, A. M., P. Dwivedi, R. Bailis, L. Hildemann, and G. Miller. 2012. “The Low Demandfor New Cookstove Technologies”. Proceedings of the National Academy of Sciences 109 (27):10815–10820.

O’Donoghue, Ted, and Matthew Rabin. 2015. “Present Bias: Lessons Learned and To Be Learned”.The American Economic Review 105 (5): 273–279.

Oliva, Paulina, B. Kelsey Jack, Samuel Bell, Elizabeth Mettetal, and Christopher Severen. 2020.“Technology Adoption under Uncertainty: Take-Up and Subsequent Investment in Zambia”. TheReview of Economics and Statistics 102 (3): 617–632.

Pattanayak, S. K., M. Jeuland, J. J. Lewis, F. Usmani, N. Brooks, V. Bhojvaid, A. Kar, L. Lip-inski, L. Morrison, O. Patange, N. Ramanathan, I. H. Rehman, R. Thadani, M. Vora, andV. Ramanathan. 2019. “Experimental evidence on promotion of electric and improved biomasscookstoves”. Proceedings of the National Academy of Science of the United States of America116 (27): 13282–13287.

Pigou, Arthur Cecil. 1920. The economics of welfare. London: Macmillan.Pillarisetti, Ajay, Mayur Vaswani, Darby Jack, Kalpana Balakrishnan, Michael N. Bates, Narendra

K. Arora, and Kirk R. Smith. 2014. “Patterns of Stove Usage after Introduction of an Ad-vanced Cookstove: The Long-Term Application of Household Sensors”. Environmental Scienceand Technology 48 (24): 14525–14533.

Pitt, Mark M., and Shahidur R. Khandker. 1998. “The impact of Group-Based Credit Programs onPoor Households in Bangladesh: Does the Gender of Participants Matter?” Journal of PoliticalEconomy 106 (5): 958–996.

Republic of Kenya Ministry of Energy. 2019. “Kenya Cooking Sector Study: Assessment of theSupply and Demand of Cooking Solutions at the Household level”.

Schilbach, Frank, Heather Schofield, and Sendhil Mullainathan. 2016. “The Psychological Lives ofthe Poor”. American Economic Review: Papers & Proceedings 106 (5): 435–440.

Shah, Anuj K., Eldar Shafir, and Sendhil Mullainathan. 2015. “Scarcity frames value”. PsychologicalScience 26 (4): 402–412.

40

Stern, Nicholas, and Joseph E. Stiglitz. 2021. “The Social Cost of Carbon, Risk, Distribution, MarketFailures: An Alternative Approach”. NBER Working Paper 28472.

U.S. Environmental Protection Agency. 2016. “Technical Update of the Social Cost of Carbon forRegulatory Impact Analysis Under Executive Order 12866”.

Udry, Christopher, and Santosh Anagol. 2006. “The Return to Capital in Ghana”. American Eco-nomic Review 96 (2): 388–393.

World Bank Group. 2020a. “State and trends of carbon pricing 2020”.— . 2020b. “The State of Access to Modern Energy Cooking Services”.World Health Organization. 2017. “The Global Impact of Respiratory Disease”.

41

Appendix Figures

Figure A1: Informational pamphlet

To reduce information asymmetries prior to the start of surveying, all participants received this leaflet containinginformation about the Jikokoa stove at baseline. The graphic with charcoal tins indicating that the Jikokoa uses only50 percent of a regular stove was designed to be understandable by literate and illiterate respondents.

Figure A2: Most salient benefits of Jikokoa stove

A. Important stove qualities B. Best attributes of Jikokoa

0.5

11.

52

Av

erag

e p

refe

ren

ce r

atio

Save moneyLess smokeEase of usePriceTasteLooks

0.2

5.5

.75

1F

ract

ion

of

resp

on

den

ts

Save money

Less smoke

Save time

Taste

Durable

I’ll be modern

Others will like it

The most salient positive attribute about the Jikokoa stove is that it saves money. This attribute is almosttwice as preferred as health benefits from reductions in smoke emissions, which itself is more than twiceas large as any other attribute.

42

Figure A3: Respondent locations

The 1,018 respondents enrolled in our study reside in one of four low-income neighborhoods in the eastern partof Nairobi: Dandora, Kayole, Mathare, and Mukuru. Respondents are randomly allocated to credit and attentiontreatment arms prior to the start of visit 2.

43

Figure A4: Attention Sheet

This figure displays the first nine weeks of the attention to benefits sheet as completed by a respondent. They arefirst asked to write down how much they expect to save each week, which may vary due to variations in for examplesocial and work engagements. They then calculate and write down the total expected savings for each month, andwhat they would do with these savings. Finally, respondents calculate their total annual savings by adding all 12monthly amounts, and write this at the top of the sheet. “kununua chakula” = “buy food” . “Kununulia watoto textbooks” = “buy the children textbooks”. Respondents in attention treatment groups A1 and A2 complete this sheet forall 52 weeks. 47 percent of respondents filled in the sheet entirely on their own. 31 percent of respondents filled inthe sheet themselves, but required guidance by the field officer. The remaining 22 of respondents were illiterate andthe field officer filled in the numbers on their behalf, while discussing the answers with the respondent. Responsesare statistically indistinguishable across these three groups. KES 100 ≈ USD 1 at the time of surveying.

44

Figure A5: Attention Treatments

Panel A: Attention control (A0)

Panel B: Attention to energy savings (A1)

Panel C: Attention to energy savings minus costs (A2)

The on-screen information during the BDM decision process for a hypothetical respondent in the weekly credittreatment group (C1). The payment amounts vary over 12 cycles of Yes/No questions depending on the respondent’sanswers. The benefits stay constant. The conversion from total cost to weekly payment amounts incorporatesinterest. Ksh (KES) 100 ≈ USD 1 at the time of surveying. While ‘pay -10’ might appear to cancel out, this waswell understood in context when the field officer talked through the funding stream.

45

Figure A6: BDM Hidden price distribution

0

50

100

150

200

250

Nu

mb

er o

f R

esp

on

den

ts

0 5 10 15 20 25 30BDM Hidden Price (USD)

The distribution of prices Pi used in the BDM elicitation mechanism. 6 percent of participants areallocated a price drawn from U [3.50, 4.50], 39 percent of participants are allocated a price drawn fromU [10, 12], and 44 percent of participants are allocated a price drawn from U [25, 27]. The remaining pricesare drawn from a uniform distribution over the entire interval U [0.01, 29.99]. Respondents buy the stoveif and only if WTPi ≥ Pi. For more detail see Online Appendix Section B.

Figure A7: Becker-DeGroot-Marschak (BDM) and Take-it-or-leave-it (TIOLI) methods practiceround items

Two practice rounds help respondents understand the BDM mechanism and allow us to compare BDM responseswith a TIOLI auction. Prior to the start of the BDM each respondent completes two practice exercises, one for abottle of lotion and one for a bar of soap. These particular brands of lotion and soap are commonly used by ourrespondents and widely sold for a retail price of $1.19 and $1.48, respectively. Each respondent is allocated a randomprice PL ∼ N(0.74, 0.35) for the lotion, truncated at USD [0.01, 1.10], and a random price PS ∼ N(0.89, 0.42) for thesoap, truncated at USD [0.01, 1.30], reflecting their respective retail prices (On the first three days of implementation,the practice prices for the lotion and soap were lower, averaging around USD 0.47 and USD 0.51 respectively. Becauseof higher than expected demand for both products, we increased prices to the higher amounts starting on the fourthday.) 50 percent of respondents were first asked to respond to a take-it-or-leave-it (‘TIOLI’) offer for purchasingthe lotion, and were then asked to complete a practice BDM exercise with the soap. The remaining 50 percent firstresponded to a TIOLI offer for the soap and then completed a BDM exercise with the lotion. The full script for theTIOLI and BDM practice rounds can be found in the On-line Appendix.

46

Figure A8: Demand curves elicited through Becker-DeGroot-Marschak (BDM) and Take-it-or-leave-it (TIOLI) methods

0

.5

1

1.5Pr

ice

(USD

)

0 .2 .4 .6 .8 1Q

BDM (Lotion)BDM (Soap)TIOLI (Lotion)TIOLI (Soap)

We test whether TIOLI and BDM elicit the same demand curves by cross-randomizing these with two goods. TheBDM demand curve is defined as Pr(WTP ≥ Pi). The TIOLI demand curve takes average adoption rates acrossintervals of 50 observations. The overlap of the two curves suggests that the BDM mechanism elicits WTP responsesthat are in line with respondents’ real behavior during a TIOLI decision. A statistical test cannot reject that averagetake-up within most price bins is equal for both elicitation methods.

Figure A9: Effort Task

A. Blank B. Complete

Example of one blank and one completed effort task. Respondents mark the number of times each symbol appears.Respondents are asked to use tick marks in order to prevent more educated or literate participants from gaining anadvantage. Most respondents took between one to two minutes to complete one task. The answers had to be within10 percent of the correct number of ticks to be marked as ‘complete’. Each respondent is required to complete atleast three tasks during each visit.

47

Figure A10: Causal IV estimate of adoption on spending over time

−1

−.75

−.5

−.25

0

.25

.5

.75

1

IV C

oef

fici

ent

on

wee

kly

ch

arco

al s

pen

din

g (

Lo

g)

−15 0 15 30 45 60Days since main visit

Instrumental Variables (IV) estimates for the causal effect of stove adoption on average weekly charcoal spendingover time. Estimates prior to adoption have larger standard errors because only respondents in the attentiontreatment groups participated in the SMS survey about charcoal expenditures prior to visit 2, which reduces thesample size.

Figure A11: Time savings from stove adoption

0 60 120 180 240 300 360

Daily time cooking (minutes)

Control(8% adoption)

Subsidy and credit treatment(93% adoption)

Time spent cooking per day, self-reported during the endline survey. We plot usage by randomly assigned treatmentsince stove adoption is correlated with WTP, but compliance is high. The instrumental variables regression suggeststhat adoption of the stove causes households to reduce their daily time spent cooking by 56 minutes. This is theopposite of the rebound theory in energy efficiency, which states that household usage of an appliance goes up afteradoption of a more energy efficient version, since the marginal cost of usage goes down. The reduction is likely dueto the ease of lighting the stove, rather than a change in cooking time itself.

48

Figure A12: Unconditional Local Quantile Treatment Effects on Charcoal Spending

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

0 25 50 75 100Quantile

Unc

ondi

tion

al L

ocal

Qua

ntile

Tre

atm

ent

Effe

ct

Unconditional local quantile treatment effect estimates on inverse hyperbolic sin charcoal spending reported via SMSand point-wise 95% confidence intervals using the method proposed by Frölich and Melly (2013). Confidence intervalswere computed using a block bootstrap clustered at the respondent level. The results show that the effect is relativelyhomogeneous across the distribution of charcoal spending.

Figure A13: Sorted Group Average Treatment Effects (GATES) on Charcoal Spending

-1.00

-0.80

-0.60

-0.40

-0.20

0.00

0.20

1st 2nd 3rdTercile of Treatment Effect

Est

imat

ed T

reat

men

t E

ffect

Sorted group average treatment effects on charcoal spending estimated using the method proposed by Chernozhukovet al. (2018). The sample was randomly split into auxiliary and main samples stratifying by quintile of WTP. Usingthe auxiliary sample, we use LASSO to form predictions of endline charcoal spending based on observed baselinemeasures of energy consumption, household socioeconomic status, demographics, health and energy prices separatelyfor stove adopters and non-adopters. Using these models, we form predictions for the main sample and difference thepredictions to create a predicted treatment effect. We then group the observations into terciles of predicted treatmenteffect and estimate the treatment effect of adopting a stove for each group using inverse propensity weighting basedon quintile of WTP. We then repeat this procedure 1000 times and take medians to form final estimates. The resultssuggest relative homogeneity in the treatment effect. 49

Figure A14: Change in WTP from midline to endline

0

10

20

30

40

50

Rev

ised

WT

P (

US

D)

0 10 20 30 40 50Original BDM WTP (USD)

Did not buy stove Bought stove

Respondent WTP as measured during the BDM mechanism and as stated during the endline survey. The 45 degreeline is displayed in blue: respondents on this line stated the same WTP during the BDM and in the endline survey.Conditional on BDM WTP, stove adoption is random. Endline WTP is similar for stove adopters and non-adopters.This rules out substantial learning or hidden attributes.

Figure A15: Robustness of under-adoption gap for varying annual δ

0

20

40

60

80

US

D

0 .2 .4 .6 .8 1Q

No Credit

Credit

Breakeven demand curve (d=0.5)

Breakeven demand curve (d=1)

Breakeven demand curve (BHJ)

Demand curves assuming annual discount factors of δ = 0.5 and δ = 1 and summing over 12 weeks—the term of theloan. The BHJ curve uses the discount parameters estimated by Balakrishnan, Haushofer, and Jakiela (2020) in asimilar context with a money now or money later design (δ = 0.942 per week and β = 0.924). Even using BHJ’srelatively more conservative estimates, present bias and exponential discounting alone cannot explain the low WTPamong the control group.

50

Figure A16: Treatment effect by WTP

Panel A Panel B Panel B

0

5

10

15

20

25

30

WT

P (

US

D)

3 6 9 12Baseline Charcoal Spending

(USD per Week)

Did not adopt

Adopted stove

95% CI

0123456789

10

En

dli

ne

char

coal

sp

end

ing

(US

D p

er W

eek

)

0 10 20 30 40 50WTP (USD)

Did not adopt

Adopted stove

95% CI

0

3

6

9

En

dli

ne

char

coal

sp

end

ing

(US

D p

er W

eek

)

3 6 9 12Baseline Charcoal Spending

(USD per Week)

Did not adopt

Adopted stove

95% CI

In Panel A, expected returns does not predict WTP. In Panel B, WTP does not predict realized returns. Panel Cconfirms that adoption affects the relationship between baseline and endline charcoal spending.

Figure A17: Loan repayment patterns

Panel A Panel B

0

.1

.2

.3

.4

.5

.6

Fra

ctio

n o

f re

spo

nd

ents

0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1Fraction paid at end

0

.2

.4

.6

.8

1

Av

erag

e fr

acti

on

of

req

uir

ed p

aid

0 1 2 3 4 5 6 7 8 9 10 11 12Weeks since Midline Visit

Weekly Deadlines

Monthly Deadlines

Panel A displays the histogram of fraction repaid made by those in the credit treatments. The median respondentrepaid 98% of the loan, and the mean repayment rate was 72%. Panel B displays the average fraction of the requiredamount that respondents had paid for every day of the 3-month payment period. These data exclude the 6 respondentswho returned the stove because of an inability to pay.

51

Figure A18: Repayment with breakeven interest rate

0

10

20

30

40

50

US

D

0 .2 .4 .6 .8 1Q

No credit

Credit treatment

Credit treatment, breakeven interest

This figures displays the demand curves in Panel A of Figure 7 as well as a hypothetical credit treatment demandcurve under a breakeven interest rate. The breakeven interest rate is the rate at which a lender would breakeven,conditional on default rates observed in the data. Interest-excluding demand is calculated by subtracting borrowingcosts from WTP.

Figure A19: Number of SMSes replied to, by respondent

0

50

100

150

Nu

mb

er o

f re

spo

nd

ents

0 5 10 15 20Number of SMSes replied to (out of 16)

The number of SMSes each respondent correctly responded to during the first 48 days (16 3-day SMS cycles) aftervisit 2. Out of 955 respondents who completed visit 2, 124 (13 percent) did not respond to any SMSes in thisperiod. 446 respondents (47 percent) responded to at least half of all SMSes. Table A4 confirms that socio-economiccharacteristics (with the exception of age) do not predict SMS non-response.

52

Figure A20: Eliciting beliefs about future savings

Respondents are first asked to estimate the maximum amount they may save in the best possible scenario. They arethen asked to allocate 20 beans on the interval [0,max] according to how likely the believe each outcome is. Aboveis a hypothetical response that a respondent might give. We discuss this in more detail in Table A6.

53

Appendix Tables

Table A1: Balance test for attention, credit, and subsidy treatments

SampleMean

AttentionTreatment

CreditTreatment

SubsidyTreatment N

(1) (2) (3) (4)Sex (female=1) 0.95 0.02 -0.01 -0.01 1011

[0.21] (0.01) (0.01) (0.01)Respondent age 37.22 -0.25 0.20 0.07 1011

[11.82] (0.82) (0.79) (0.74)Number of household residents 4.74 -0.04 0.13 0.07 1011

[2.09] (0.15) (0.14) (0.13)Number of child residents 2.58 -0.03 0.11 0.09 1011

[1.72] (0.12) (0.11) (0.11)Savings in bank, mobile, ROSCA (USD) 74.85 2.46 6.08 -12.13 1011

[129.51] (9.00) (8.64) (8.14)Household income (USD/week) 47.24 -2.32 -3.88∗ -1.69 1005

[34.65] (2.41) (2.32) (2.19)Total energy consumption (USD/week) 8.55 0.07 0.12 -0.16 1011

[3.59] (0.25) (0.24) (0.23)Charcoal consumption (USD/week) 5.59 0.04 0.04 -0.09 1011

[2.60] (0.18) (0.17) (0.16)Price of old jiko (USD) 3.39 0.07 -0.02 0.07 1006

[1.33] (0.09) (0.09) (0.08)Risky investment amount (0-4 USD) 1.20 0.04 -0.10 -0.11∗ 1011

[1.00] (0.07) (0.07) (0.06)Joint F-Test 0.78 0.48 0.14

Each row and each treatment column represents an individual regression of the row variable on an indicator forreceiving the treatment in the column. For legibility, in this table the sub-treatments for attention and credit arepooled. For the subsidy treatment balance check, we define treatment as BDM price Pi ≤ USD 15 (50 percentof respondents). Treatment assignment was stratified on baseline charcoal spending. The three treatments appearto be balanced on observable demographic and socioeconomic characteristics. SD in brackets. SE in parentheses.∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

54

Table A2: Causal impact of stove adoption on weekly charcoal spending

OLSIV Estimate

(1-month endline)IV Estimate

(1-year endline)

(1) (2) (3) (4) (5) (6)USD USD IHS(USD) IHS(KG) USD IHS(USD)

Control -2.890∗∗∗ -2.857∗∗∗ -0.636∗∗∗ -0.593∗∗∗ -2.996∗∗∗ -0.772∗∗∗

(0.524) (0.704) (0.188) (0.195) (0.931) (0.223)

Credit Only (pooled) -2.017∗∗∗ -2.588∗∗∗ -0.581∗∗∗ -0.456∗∗ -2.085∗∗ -0.439∗∗

(0.521) (0.682) (0.171) (0.181) (0.940) (0.221)

Attention Only (pooled) -1.563∗∗∗ -1.896∗∗∗ -0.409∗∗∗ -0.528∗∗∗ -2.324∗∗∗ -0.582∗∗∗

(0.420) (0.515) (0.105) (0.150) (0.848) (0.146)

Credit and Attention (pooled) -2.139∗∗∗ -2.248∗∗∗ -0.472∗∗∗ -0.414∗∗∗ -2.562∗∗∗ -0.493∗∗∗

(0.395) (0.534) (0.128) (0.154) (0.746) (0.166)

Observations 7870 7870 7870 796 6982 6982Control Mean 5.717 4.964 2.154 1.545 5.303 2.212Socioeconomic controls Yes Yes Yes Yes Yes YesData Source SMSes SMSes SMSes Buckets SMSes SMSes

Results from an instrumental variables regression that uses the (randomly assigned) BDM price as an instrumentfor stove adoption to estimate the causal impact of adoption on weekly charcoal expenditures. Column (1) presentsthe OLS results. Column (2) uses weekly charcoal expenditures in USD as the outcome variable. Column (3) usesthe inverse hyperbolic sine (IHS) conversion of the USD amount. Column (4) uses the IHS of the weight of thecharcoal bucket one month after stove adoption as the outcome variable. Columns (5) and (6) conduct the sameanalyses as columns (4) and (5) respectively, but using data from the SMS survey conducted one year after themain visit. Socioeconomic controls include baseline savings, income, risk aversion, credit constrainedness, numberof adults and children. In regressions using SMS data, errors are clustered by respondent. SE in parentheses.∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

55

Table A3: Causal impact of stove adoption on long-term financial assets

Mobile Savings Formal Savings SACCO Total Savings(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)USD >0 (=1) Log(USD) USD >0 (=1) Log(USD) USD >0 (=1) Log(USD) USD >0 (=1) Log(USD)

WTP (USD) 0.042 -0.000 0.003 0.091 0.002∗ 0.015 0.008 0.001 0.002 0.148 0.001 -0.002(0.070) (0.002) (0.013) (0.085) (0.001) (0.035) (0.326) (0.002) (0.007) (0.345) (0.002) (0.010)

Bought Cookstove (=1) 1.570 0.121∗ 0.356 -1.013 -0.020 -0.065 14.581 0.108 0.177 15.430 0.114∗ 0.564∗(2.049) (0.069) (0.424) (2.465) (0.035) (0.933) (9.586) (0.069) (0.203) (10.130) (0.061) (0.291)

Observations 822 822 372 838 838 59 829 829 436 792 792 580Control Mean 5.283 0.430 1.365 3.724 0.077 2.873 34.720 0.507 3.649 41.249 0.710 3.014Socioeconomic controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Results from an instrumental variables regression that uses the (randomly assigned) BDM price as an instrument for stove adoption to estimate the causal impactof adoption on savings. Columns (1) - (3) consider savings in mobile banking accounts. Columns (4) - (6) look at amounts in formal banking savings accounts.Columns (7) - (9) consider money that is available for withdrawal for SACCO, merry-go-round, and ROSCA participants. Columns (10) - (12) show results forsavings available in mobile and formal accounts, SACCO, merry-go-rounds, and ROSCA. In Column (12) a log increase of 0.445 corresponds to an increase of 56percent over the control mean. Socioeconomic controls include baseline savings, income, risk aversion, credit constrainedness, number of adults and children, andmean and standard deviation of health beliefs. Data for the outcome variables is from the one-year endline survey. SE in parentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

56

Table A4: Socio-economic characteristics do not predict attrition

Attrited(Visit 2)

Attrited(Visit 3)

Attrited(SMSes-1)

Attrited(SMSes-2)

BDM Treatment (Price <= 15 USD) 0.015 -0.003 0.005 0.025(0.014) (0.018) (0.031) (0.031)

Credit Treatment 0.012 0.004 -0.042 0.001(0.015) (0.019) (0.033) (0.033)

Attention Treatment 0.030∗ 0.034∗ 0.002 0.017(0.016) (0.019) (0.035) (0.035)

Sex (female=1) -0.010 -0.001 -0.017 0.043(0.035) (0.042) (0.075) (0.075)

Respondent age -0.002∗∗ -0.002∗∗∗ -0.003∗∗ -0.004∗∗∗(0.001) (0.001) (0.001) (0.001)

Number of household residents -0.004 -0.007∗ -0.009 -0.008(0.003) (0.004) (0.008) (0.008)

Number of child residents -0.006 -0.009∗ -0.002 -0.007(0.004) (0.005) (0.009) (0.009)

Savings in bank, mobile, ROSCA (USD) -0.000 -0.000 -0.000 -0.000(0.000) (0.000) (0.000) (0.000)

Household income (USD/week) -0.000 -0.000 -0.000 0.001(0.000) (0.000) (0.000) (0.000)

Total energy consumption (USD/week) -0.000 -0.003 0.002 -0.002(0.002) (0.002) (0.004) (0.004)

Charcoal consumption (USD/week) -0.002 -0.004 0.001 -0.002(0.003) (0.003) (0.006) (0.006)

Price of old jiko (USD) 0.002 -0.004 -0.005 0.017(0.005) (0.007) (0.012) (0.012)

Risky investment amount (0-4 USD) 0.002 0.002 -0.001 -0.020(0.007) (0.009) (0.016) (0.016)

Jikokoa (=1) -0.033∗∗∗ -0.062∗ -0.042(0.012) (0.033) (0.033)

Joint F-Test p-Value 0.77 0.98 0.82 0.57Sample Mean 0.06 0.09 0.46 0.48

Each coefficient in each of the Attrited columns represents a separate regression testing whether the outcome variablepredicts attrition. Attrited (SMS) equals one for respondents who responded than fewer of the median number ofSMSes. SE in parentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

57

Table A5: Impact of experimental treatments on WTP

WTP(USD)

(1) (2) (3) (4) (5) (6) (7) (8)

Credit (pooled) 12.50∗∗∗ 12.47∗∗∗(17.99) (9.76)

Credit (C1 only) 13.13∗∗∗ 13.13∗∗∗ 13.32∗∗∗(16.42) (16.40) (14.17)

Credit (C2 only) 11.85∗∗∗ 11.86∗∗∗ 11.55∗∗∗ -1.287(14.73) (14.73) (12.19) (-1.57)

Attention (pooled) 0.251 0.341(0.30) (0.27)

Attention (A1 only) 0.357 0.496 1.254(0.35) (0.57) (1.15)

Attention (A2 only) 0.180 0.288 -0.127 -0.141(0.20) (0.36) (-0.16) (-0.14)

Attention (pooled) X Credit (pooled) 0.0496(0.03)

Observations 943 943 943 943 943 943 663 628Control Mean 12.33 20.45 12.12 12.33 20.45 24.58 20.75 20.75Sample Full Full Full Full Full Full A1 & A2 C1 & C2

Impact of pooled treatments on WTP. Socioeconomic controls include baseline savings, income, risk aversion, creditconstrainedness, number of adults and children. SE in parentheses. ∗ ≤ 0.10,∗∗≤ .05,∗∗∗≤ .01.

58

Table A6: Attention Treatment on Beliefs

To elicit respondents’ beliefs about the savings generated by the stove over a one-year period ofownership we use methods developed in Delavande (2014), Delavande, Giné, and McKenzie (2011b),Delavande, Giné, and McKenzie (2011a), and Mel, McKenzie, and Woodruff (2008). Figure A20displays the beliefs sheet we employed. So as not to anchor respondents to a specific point on thesheet, prior to commencing the beliefs exercise respondents were first asked what they believedwould be the maximum they could possibly save in a year. Respondents then receive the beliefssheet and are given the following instructions:

“There are 20 beans. I would like you to take these beans and place them on this sheetaccording to how much you think you will save over the next year, if you receive aJikokoa. If you put some beans in one of the groups, it means you think you mightsave that amount of money. For example:• If you put all your beans in one group, it means you are sure that you will saveexactly that amount of money this year if you had the Jikokoa.• If you do not put any beans in one of the groups, it means you are confident thatyou will not save this amount of money.• As you add beans to a group, it means that you think the likelihood that you willsave that amount increases.• If you put 1 bean in every group, it means you think all of the outcomes areequally likely.”

EndlineBeliefs

(1) (2)

Baseline beliefs 0.09∗∗∗ 0.09∗∗∗

(0.03) (0.03)

Attention (pooled) 0.12∗

(0.07)

Attention (A1 only) 0.21∗∗∗

(0.08)

Attention (A2 only) 0.05(0.07)

Observations 943 943Control Mean 0.00 0.00

Beliefs in each period are standardized for the control group to have mean 0 and standard deviation 1. Controls includerandomized credit treatment and price assignments and socioeconomic controls, which include baseline savings,income, risk aversion, credit constrainedness, number of adults and children.

59