Essays on Household Electri cation in Developing Countries · 2018. 10. 10. · Doctor of...

Essays on Household Electrification in Developing Countries

by

Manuel Fernando Barron

A dissertation submitted in partial satisfaction of the

requirements for the degree of

Doctor of Philosophy

in

Agricultural and Resource Economics

in the

Graduate Division

of the

University of California, Berkeley

Committee in charge:

Professor David Zilberman, ChairAssociate Professor Maximilian Auffhammer

Professor Edward Miguel

Spring 2014

1

Abstract

Essays on Household Electrification in Developing Countries

by

Manuel Fernando Barron

Doctor of Philosophy in Agricultural and Resource Economics

University of California, Berkeley

Professor David Zilberman, Chair

This dissertation studies issues on household electrification in developing coun-tries, an issue of utmost importance in the fields of development and energy eco-nomics. Chapter 1 provides the first empirical evidence that household electrifica-tion leads to direct and substantial welfare improvements via reductions in indoor airpollution. In the setting of a recent electrification program in northern El Salvador,we exploit a unique dataset on minute-by-minute fine particulate matter (PM2.5)concentration within the framework of a clean experimental design. Two years afterbaseline, overnight PM2.5 concentration was on average 63% lower among householdsthat were randomly encouraged to connect compared to those that were not. Thischange is driven by reductions in kerosene use and, as suggested by our conceptualframework, the effects are larger among heavier kerosene users. As a result, the in-cidence of acute respiratory infections among children under 6 fell by 35-45% amongencouraged households. Estimates of exposure measures suggest large health gainsfor all household members, but these gains are unequally distributed by gender. Inaddition, we show that when the electrification rate among the non-encouraged groupcaught up with that of the encouraged group, the effects in the former group weresimilar to those in the latter.

Chapter 2 studies the effects of household electrification on time use, and exploresthe consequences of this resource reallocation. In particular we study the effectson investment in education and job choice. We find that electrification increasesinvestment in education, by inducing school-age children to study more at home,which is reflected in improvements in a math skills index. Adult males reduce timeallocated to leisure and independent farm work, while increasing time in other laboractivities. This time reallocation is reflected in higher income. Electricity also leadsto increased ownership of electronic appliances like TV sets, that improve leisure

2

and access to information; as well as refrigerators, which improve nutrition throughimproved food storage and food safety, reinforcing the health gains from improvedindoor air quality.

Chapter 3 studies the role that a cost-sharing scheme like the vouchers used in ourexperiment play in adoption of electric connections and the potential consequenceson infrastructure financing. We draft a model showing the conditions under whichsharing part of the connection costs would increase the electric utility’s revenues,by accelerating the connection rate, thus increasing the flow of revenues earlier thanotherwise. In our the case at hand, offering a low discount resulted more profitablethan not sharing any connection costs, and also more profitable than offering highdiscounts.

i

To my mother, wife, and kids (whenever they decide to come!)

ii

Acknowledgments

I want to thank my dissertation committee for their constant support, thoughtfuladvice, and infinite patience: David Zilberman, Max Auffhammer, and Ted Miguel.Along the years I benefited from the thriving intelectual environment at Berkeley,but the following individuals deserve special mention: Pia Basurto, Kyle Emerick,Meredith Fowlie, Francois Gerard, Kelsey Jack, Nick Lam, Jeremy Magruder, ShaunMcRae, Eduardo Nakasone, Ajay Pillasiretti, and Elisabeth Sadoulet. Kirk Smith,and Catherine Wolfram. Previous versions of the papers in this dissertation havebeen presented at various seminars in UC Berkeley, IFPRI, PacDev (SFSU), IPWSD(Columbia), and NEUDC (Harvard).

My beautiful wife pushed me, most of the times gently, whenever I started losingsteam. Special thanks to my co-author, Maximo Torero, for the trust he placedin me, especially in the early stages of the project. We are especially indebtedto Victor Agreda and the IFPRI and DIGESTYC field staff for their efforts andinsightful feedback. MCC, DIGESTYC, and IFPRI provided valuable support inthe implementation of this study. I gratefully acknowledge funding from the UCBerkeley Graduate Division.

I owe my academic career to my mother, Mercedes Ayllón, who taught me froma young age the importance of education.

1

Chapter 1

Household Electrification andIndoor Air Pollution1

1.1 Introduction

In 2009, 1.3 billion people lacked access to electricity at home (International EnergyAgency, 2011). At night, households with no access to electricity make do mostlywith candles or kerosene lamps to satisfy their illumination needs. These sourcesof light provide poor illumination and, more importantly, emit high amounts ofpollutants harmful for human health. In fact, indoor air pollution (IAP) is thethird leading risk factor for global disease burden, after high blood pressure andsmoking (Lim et al., 2013).2 Given the stylized fact that lighting is one of thefirst uses of electricity in newly electrified areas (see e.g. Bernard, 2012; Barnes,2007; Independent Evaluation Group, 2008), electrification is expected to decreaseIAP levels by replacing traditional source of lighting, like kerosene, candles, andwoodsticks. These reductions and their potential health effects are often argued tobe one of the main benefits of electrification, but there is no solid empirical evidenceto date.

Our paper contributes to filling this gap by providing the first experimental esti-mates of the relationship between household electrification and indoor air pollution.To answer this key question we collected, within the frame of a clean experimentaldesign, a uniquely rich dataset that pairs minute-by-minute fine particulate matter(PM2.5) concentration with detailed data on household members’ time allocation.

1This chapter is joint work by Manuel Barron and Maximo Torero2Ambient pollution also has important negative health effects, as shown by Chay and Greenstone

(2003) and more recently by, e.g., Chen et al. (2013) and Hanna and Oliva (2011). This paper willconcentrate on indoor air pollution only.

CHAPTER 1. ELECTRIFICATION AND INDOOR AIR POLLUTION 2

This unique combination of data and experimental framework allows an accurateestimation of the lower bound of the changes in PM2.5 concentration driven by elec-trification and, moreover, to assess the magnitude of the health effects that thesechanges imply.

Our research question is situated within the broader area of the effects of elec-trification, an active area of research in which the debate is far from settled. Themassive amounts of resources allocated to rural electrification3 have usually beenjustified on the assumed benefits of electrification on health, education, and income,but most of the empirical evidence on which these claims are based is weak (Bernard,2012; Independent Evaluation Group, 2008), and the more recent literature showsmixed results. Some of the recent evidence suggests that electricity may save timein household chores, thus increasing female labor supply (Dinkelman, 2011; Gro-gan and Sadanand, 2012) or that it leads to improvements in educational outcomes,consumption, and income (e.g. Khandker, Barnes, and Samad; van de Walle et al.,2013), while others find no such relationships (e.g. Bernard and Torero, 2013; Bensch,Kluve, and Peters, 2011).

A major challenge in this literature is the identification of causal effects. Sincethe electric grid cannot be expanded randomly, recent studies use time variationand instrumental variables (IV), mainly geographic variables, to deal with the en-dogeneity of connection. Studies have used land gradient (Dinkelman, 2011; Groganand Sadanand, 2012), distance to hydroelectric dams (Grogan and Sadanand, 2012),distance to the electricity line (Samad, Khandker, and Barnes, 2009), and distanceto power generating plants and baseline electrification rate in the locality (van deWalle et al., 2013).

Our estimates provide the first experimental evidence that household electrifica-tion leads to substantial, immediate and sustained improvements on welfare throughan improvement of indoor air quality. Overnight PM2.5 concentration was 63 per-cent lower among voucher recipients than among non-recipients. Due to imperfectcompliance with voucher assignment, this is a lower bound of the true effect of elec-trification on PM2.5 concentration. Similarly, the probability probability of observingconcentrations above 67 µg/m3 between 5pm and 7am fell by at least 33 percent-age points. Although there are no clear benchmarks in the environmental healthliterature, we consider this a useful reference point because over a 24-hour period, aconcentration of 67 µg/m3 implies the same PM2.5 exposure as smoking one cigarette

3For instance, the World Bank recommends investing $10 billion per year between 2010 and2020 to production and distribution of electricity in rural areas in Africa (World Bank, 2009). Giventhat the institution aims to provide 250 million people across Africa with modern sources of energyby 2030 (World Bank, 2007), understanding the effects of electrification is of urgent importance.


(1200 µg/m3).4

The reductions in overnight PM2.5 concentration result in large and significantfalls in acute respiratory infections (ARI) among children under 6. Depending onthe exact specification voucher recipients report 37 to 44% lower incidence of ARIin the four weeks preceding the survey than non-recipients. To assess further healthimplications of the observed reductions in PM2.5 concentration among the populationover 6, in section 6.1 we use data from the time allocation module to estimate thechange in daily exposure to PM2.5. The resulting reductions in exposure to PM2.5 arelarge but unequally distributed among household members. Adult males benefit themost, with 59% lower exposure. Since adult females are still exposed to high PM2.5concentrations while cooking, they benefit the least, with reductions in exposureof 33%. The figures for children are 46% (males) and 39% (females). The dose-response function recently developed by Pope III et al. (2011) based on first andsecond-hand tobacco smoking associates the figures we find with large reductions inthe relative risk of lung cancer, 25% for adult females, 33% adult males, with therespective figures for children falling 25% (females) and from 30% (males). Althoughthe composition of PM2.5 generated by kerosene combustion is not the same as theone generated by cigarette smoking, the current scientific evidence cannot reject thattheir health effects are similar.

The mechanism behind the PM2.5 reductions in our study setting is a substitutionaway from kerosene lighting. Electrification caused large reductions in kerosene ex-penditures, while changes in other traditional lighting sources like candles are smallin magnitude and not statistically significant. We find no evidence of changes incooking practices either. The reduction in kerosene use has important health im-plications, because although kerosene is usually considered a cleaner alternative tobiomass, emissions from kerosene-burning devices are considered extremely harmfulfor human health. Aside from PM2.5, kerosene emissions include carbon monoxide(CO), nitric oxides (NOX), and sulfur dioxide (SO2) (Schare and Smith, 1995; Fanand Zhang, 2001; Röllin et al., 2004). These pollutants can impair lung function andincrease infectious illness, asthma, and cancer risks Lam et al. (2012a). The reduc-tion in kerosene use also has important environmental consequences, since kerosenelighting (let alone other uses like cooking or heating) is responsible for 7 percent ofannual black carbon emissions globally (Lam et al., 2012b).

The reductions in PM2.5 concentration found in this study are not necessarily ob-vious ex-ante for four reasons. First, since households continue to use fuelwood for

4Concentration is the amount of PM2.5 per unit of space, usually measured in mg/m3 or µg/m3.

Exposure is the amount of PM2.5 that effectively enters the human body, usually measured in mgor µg.


cooking, and woodsmoke produces higher IAP concentration than lighting fuels, theresulting reductions in IAP may have been small and thus irrelevant for policy. Infact, given that there is still scarce evidence in the specialized literature, the evidenceof a strong, positive relationship between kerosene use and PM2.5 concentration ina sample in which 70% of households rely on fuelwood for cooking is a contributionto the environmental health literature.5 Second, it may be the case that only theheaviest kerosene users experience significant reductions, and thus the average reduc-tion may just be an illusion. We explore this issue and find even though the highestpolluters indeed get the highest gains, 80% of households experienced significant re-ductions in overnight PM2.5 concentration. Third, voucher recipients may have notadopted electricity at a high enough rate, thus the relationship between receiving avoucher and connecting to the grid may be too weak to reflect in IAP measurements.Fourth, households may be effective in dealing with IAP from lighting sources.6

The study closest to ours is Bernard and Torero (2013), in which the authorsimplemented a RED to study the effects of electrification in Ethiopia. We buildon their work with four main differences. First, our main outcome of interest isPM2.5, which is not measured in their study. Second, our design allows to controlfor externalities in adoption of electric connections, which due to characteristics ofthe Ethiopian electrification program is not possible in their study. Third, our studyallows to analyze dynamic effects of household electrification, since it includes abaseline and three follow-up follow-up survey.7,8

Our findings differ qualitatively from the typical findings in the literature onimproved cookstoves in two dimensions. First, the effects found in field studieson improved cookstoves are null or small, especially when compared to the effectsexpected from laboratory or controlled field studies.9,10 The second dimension is

5We explore this issue in closer detail in Barron and Torero (2013).6In fact, households in the study setting households usually place the kerosene lantern or ocote

sticks close to windows and doors. Ocote is a type of wood used for ligthing by the pooresthouseholds.

7In addition, a fourth follow-up survey is scheduled for November 2013 and a fifth one inNovember 2015

8An additional, although not major difference, is the study setting: El Salvadorian householdshave higher average income and more readily access to remittances than Ethiopian households, thusare potentially more responsive to the variations in cost of connection induced by our experiment.

9By “controlled field studies” we refer to field trials in which field officers or research assistantsconstantly visit households to make sure the improved cookstoves are being used the way they aresupposed to be used. Although this blocks the behavioral response of the household, the evidenceis informative as to the maximum potential benefits of new technologies under adequate use.

10The main driver of this divergence is household behavior: households do not adopt the im-proved stoves (WHO, 2006; Mobarak et al., 2012)11, or if they do adopt them, do not use themcorrectly (Wallmo, Jacobson et al., 1998). A large body of literature investigates the reasons (e.g.


that unlike the cookstove literature (e.g. Hanna, Duflo, and Greenstone, 2012), thereductions in PM2.5 observed in our sample are steady over time. The time resilienceof the effects we find strengthens the link between household electrification andhuman health discussed in the preceding paragraphs.12

The remainder of this document is organized as follows. Section 2 presents thestudy setting and discusses the data. Section 3 describes the conceptual frameworkthat guides our study. Section 4 presents the econometric approach. We discuss themain results in section 5. In section 6 we combine the findings on PM2.5 concentra-tion with time-use data to infer PM2.5 exposure and health implications. Section 7presents a note on the profitability of replicating our by a private agent, and section8 concludes.

1.2 The Study Setting

In this section we describe the study setting and give detailed information on theexperimental design. After a brief introduction on fine particulate matter and adescription of the channels through which it affects human health, we describe theprotocol and equipment used to measure it. Next we present descriptive statisticsof our baseline survey and show that groups resulting from the randomization havebalanced means across a wide set of observable characteristics.

The Electrification Program and Experimental Design

The study takes place during a recent grid extension and intensification program innorthern El Salvador, designed to be rolled-out in three phases according to con-struction costs and accessibility. In this program, the El Salvadorian governmentcovered all the installation costs up to the electric meter, and households had to payfor their internal wiring and a connection fee (for a safety certification). The feefor the safety certification is of around US$ 100. It is non-trivial for a household,amounting to roughly 20% of annual per capita income in our sample.

The experimental sample consists of 500 households located in subdistricts thatwere scheduled to be covered by the program during its first year. We generatedexperimental variation in the connection fee by offering discount vouchers to a ran-domly selected subsample. We randomly allocated 200 low-discount vouchers (20%

Masera, Dı́az, and Berrueta, 2005; Pokharel, 2003; Troncoso et al., 2007).12Although recent work suggests that exposure to IAP for short periods of time can have per-

manent negative effects on in-utero babies.


discount), 200 high-discount vouchers (50% discount), and left the remainder house-holds as control group (N=100). The exogenous variation in the connection feegenerated by the random voucher allocation deals with self-selection in connectionto the grid. Vouchers were valid for a discount towards the safety certification to bereimbursed after paying the full cost. Each voucher showed the name and address ofthe beneficiary, it was non-transferable, and it was valid for 9 months.13

The random voucher allocation also creates exogenous variation in the number ofvoucher recipients in a given neighborhood of household i (controlling for the numberof eligible neighbors), which generates variation in the number of new connectionsaround household i, so we can control for the role of spillovers on grid connection.The sign of the effect is theoretically ambiguous. On the one hand, observing theirneighbors connect to the grid may make households more prone to connect, througha combination of social learning and imitation effects.14 On the other, higher formalconnection rates in a neighborhood reduce the cost of getting an informal connection,so the number of vouchers around a household may increase the number of informalconnections. To estimate the role of spillovers on adoption, we use the number ofhousehold i’s neighbors that received a voucher in a given radius (0-100 meters, 100-200 meters, 200-300 meters), controlling by the number of eligible neighbors in thatradius. Eligible households are households with no electricity at baseline.

EHEIPCER, the household survey implemented for this study, is a fairly stan-dard survey that collected data on demographic characteristics, health, education,housing characteristics, energy use, income, consumption, among others. In par-ticular, it includes a detailed module on time allocation for up to four householdmembers: the male head, the female head, and up to two school-age children. Stricttraining sessions were conducted to ensure high quality in data collection, which wasconducted with handheld computers. Enumerators were trained and selected by theauthors with the assistance of DIGESTYC (the Salvadorian Bureau of Statistics)and IFPRI staff. The indoor air pollution data described below were collected bya subset of enumerators that underwent additional special training to this end.

The baseline household survey, designed using the 2007 Population Census asthe sampling framework, was collected in November and December 2009. It covered4,800 households all over northern El Salvador. Three follow-up surveys have been

13In a few cases there was a delay in the implementation of the program, so the expiration datewas extended for the households from those areas.

14Social learning would occur if households observed the private benefits of electrification (betterillumination, less smoke at night, better food availability, more enjoyable leisure time) from theirneighbors. Imitation effects (also known as “preferences interactions” in the literature) are similarto a “keeping-up with the Joneses” story: a household wants electricity because its neighbors haveit.


collected in the same months in 2010, 2011, and 2012 respectively. An additionalfollow-up survey is scheduled for November 2013, and a final round is scheduled tobe fielded in November 2015.

Fine Particulate Matter (PM2.5)

The main outcome of interest in this paper is fine particulate matter (PM2.5), whichis particulate matter (PM) with a diameter of 2.5 microns or less (1 micron = 0.001millimeters). Particulate matter, also known as particle pollution, is a complexmixture of small particles and liquid droplets composed by potentially hundreds ofchemicals. Given the complexity of their composition, particles are mainly classifiedaccording to their size. Particles with a diameter between 10 and 2.5 microns arealso known as “coarse particulate matter” or “inhalable coarse particulate matter”(PM10), while particulates with a diameter under 2.5 microns are known as “fineparticulate matter” (PM2.5). As reference, a human hair has a diameter of between50 to 70 microns, 20 to 30 times larger than the cutoff point for fine particulatematter.

Particle size is inversely linked to its potential for causing health problems. BothPM10 and PM2.5 can pass through the throat and nose and enter the lungs, but beingsmaller, PM2.5 can get deeper into the lungs and also enter the bloodstream, thuscausing more damages to health than PM10. Both PM10 and PM2.5 have been shownto cause or aggravate heart and lung diseases. Further, there is evidence that theyweaken the immune system, making the body more vulnerable to disease in general.Thus, it affects cognitive ability and productivity.15

PM2.5 Measurement

A central part of the project consisted in collecting data on overnight PM2.5 con-centration. We obtained PM2.5 measurements in two subsamples of households, oneexperimental and one non-experimental. The experimental subsample includes PM2.5data on 141 randomly selected households from the 500 households that were con-sidered for voucher allocation. The reasons for not selecting the whole sample were

15Other than by size, particles that compose particulate matter are usually classified as primaryor secondary. Primary particles are emitted directly by a source, like kerosene lamps, cookstoves,unpaved roads, or construction sites. The outcome of interest in this paper falls in this category.Secondary particles, on the other hand, are formed in the atmosphere, as a result of sulfur dioxidesand nitrogen oxides emitted from power plants, industries and automobiles. Secondary particlesaccount for most of the particulate matter in developed countries, while the converse is arguablytrue in developing regions.


logistical and budgetary. Measurements for these households were collected withrounds 3 and 4 of the household survey. The non-experimental subsample is formedby 200 households of EHEIPCER households from neighboring subdistricts in thesame departments as the experimental sample (San Miguel and Chalatenango) thathad not connected to the grid by September 2010. Measurements in these householdscollected with rounds 2, 3, and 4 of the household survey. The non-experimentalsample is formed by households that had not connected to the grid by round 2.Descriptive statistics of both subsamples are reported in Appendix Table A1.

In each household we measured minute-by-minute PM2.5 concentration between1700 hours and 0700 hours the next morning in the main evening living area. Themain evening living area is defined as the room where household members spentmost of their time awake during the evenings. In the majority of cases this wasthe living room. Measurements were conducted with the University of California atBerkeley Particle and Temperature Sensor (UCB-PATS). The UCB-PATS is a small,portable and non-intrusive datalogging particle monitor for indoor environments. Ituses a photoelectric detector to measure PM2.5 concentrations down to 25 µg/m

3.The UCB-PATS records PM2.5 concentration, relative humidity and temperature ata 1 min time resolution. For details on the development and performance of theUCB-PATS see Litton et al. (2004), Edwards et al. (2006), and Chowdhury et al.(2007).

Experienced and meticulously trained enumerators visited the selected house-holds, explained the purpose of the study and obtained consent to place the UCB-PATS in the home. The protocol implemented to measure PM2.5 concentration issimilar on the protocol applied by Northcross et al. (2010) for cookstoves. It isa standard protocol in the cookstove literature but there is no standard protocolin place for measuring indoor air pollution emitted by kerosene lanterns.The mon-itor was placed in the room where most household members spent most of theirtime awake during the evenings. In practice, this was mostly the living room. In ahandful of cases it was the master bedroom. The monitor was placed on a wall 1m(horizontally) from the place where the lamp is usually located in the evenings, atleast 1.50m away from any working doors or windows, and at a height of 1.50m abovethe ground. In the baseline measurement enumerators took pictures of the picturesof the placement to ensure placing the monitor in the same place in the follow-upvisits. This reduces the risk of generating artificial variation in PM2.5 concentrationsby placing the monitor in different locations.

In follow-up measurements, the enumerators used pictures from previous roundsto place the monitors in the same place as the baseline measurement. The enu-merators filled a data sheet with exact details on the height, distance, set-up time,pick-up time, among other information. The monitors were placed in the homes be-


fore 1600 hrs. If the monitor was placed in a home between Monday and Thursday,it was picked up the next morning starting around 0800 hrs. If it was placed in ahome on a Friday, it was picked up the coming Monday starting around 0800 hrs.This procedure was followed to comply with labor regulations in the governmentsector. In a subsample of households the measurement took place between 5pm ona Friday and 7am the coming Monday. Following the standard practice in the envi-ronmental health literature, the resulting PM2.5 concentration for those householdswas averaged across the three days.

Descriptive Statistics and Balance

According to the 2007 National Census around 80% of the El Salvadorian populationhad access to electricity. Although this figure is high, there are strong correlationsbetween socioeconomic status, electrification, and use of traditional fuels for lightingor cooking. Figure 1 shows that the poorest municipalities are the ones with thelowest electrification rates and the highest use of traditional fuels for cooking andlighting.

To illustrate the relationship between kerosene use and indoor air quality in ourstudy setting, Figure 2 shows a non-parametric regression of overnight PM2.5 as afunction of monthly kerosene expenditure (more details on the variables and thesamples in the next section). There is a clear positive relationship between these twovariables, suggesting that reductions in kerosene use could generate important im-provements in indoor air quality. Kerosene provides an important source of variationin PM2.5 even with 70% of households using wood for cooking.

Table 1 shows descriptive statistics split by treatment arm. Column 1 showsthe means for the control group, column 2 shows the means for the households thatreceived a 20% discount, and column 4 shows the means for households that receiveda 50% discount. Columns 3 and 5 test for differences between each of the treatmentarms and the control group. Household heads are on average 50 years old, 69%of them are male and have 2.4 years of schooling on average. Literacy rates amonghousehold heads are low, with only 54% of them reporting being literate. The averageage in the households is 30.8 and households are composed by 4.5 members, with atotal dependency ratio of roughly 0.45. Annual income is around US$770 per head,roughly US$2.11 per person per day.

The main source of energy expenditure is kerosene (US$2.11 per month) mainlyused for lighting, and propane (US$2.09 per month), mainly used for cooking, fol-lowed by candles (US$0.46/month) and car battery recharging (US$0.08/month),used to power TV sets. Use of wood for cooking was reported by 70% of households.Thirty-eight percent of households had informal access to electricity at baseline. In-


formal connections consist on a series of extension cables connected to each other andplugged into a neighbor’s sockets. They are at most enough for two lightbulbs andsome times a television set. For our purposes, households with informal connectionswere treated as off-grid. This can attenuate the effects of electrification on indoor airquality if we think that households with informal connections rely less on kerosenefor lighting than those with no connection at all. However, since it is difficult forthe government or the electric utility to determine if a household has informal accessto the grid, we argue that the results from our strategy are more relevant for policypurposes.

Households were also balanced regarding their ex-ante perceptions towards energysources. The vast majority agreed that electricity illuminates better than kerosene(96%) and that woodsmoke generates respiratory problems (87%). Between 30 to40% of respondents said that kerosene is not an expensive source of lighting, and20-30% said it as the best way to illuminate their household.

1.3 Conceptual Framework

In this section we present the conceptual framework that guides our analysis. Themodel is based in the work by Zilberman and Liu (2011), who model the demand fordurables with indivisibilities. We apply their model for the case in which a householddecides whether to use traditional fuels or electricity for lighting. We make only aminor modification, which is that utility depends on air quality a, and that traditionalfuels generate lower air quality.16

Each household maximizes its utility function

U(l, a, z) (1.1)

, which depends on artificial illumination, l, indoor air quality a, and a compositegood z that will work as numeràire. The shape of the utility function is such that∂U(·)∂i

> 0, ∂2U(·)∂i2

< 0, i = l, a, z.Illumination can be produced using two alternative processes: traditional (kerosene

lamps) or modern (electricity). Each household uses only one process: either tradi-tional fuels or electricity. Let D be a dichotomous variable that takes the value ofone if the household uses electricity and 0 if it uses traditional fuels for lighting.

The traditional process does not require purchase of a durable, but the moderndoes. In exchange for the higher price, the modern process offers a lower cost per

16We could introduce a second modification, that the marginal utility of pollution depends uponthe comparison with the neighbors. This would help understand the role of peer effects, but we donot study this in this paper.


lumen than the traditional process, and produces no pollution. For simplicity, weassume a ∈ [0, a∗], with a∗ being the best ambient environment possible given thatthe household relies on biomass for cooking.17

The household’s budget constraint is given by:

[(1−D)p0 +Dp1]l + z = R−DI (1.2)

, where p0 is the price per lumen with the traditional technology, p1 is the pricewith electricity (p0 > p1), I, and R is household income.

Finally, let pollution be a function of l and the technology used to produce it,such that

a = a∗ − (1−D)× f(l) (1.3)

Even in households have identical preferences, heterogeneity in income will gener-ate heterogeneity in the optimal demand for artificial illumination, which determinesambient quality a. Zilberman and Liu (2011) show how reductions in I increase thedemand for the durable for a case in which utility depends on l and z. In our analysisa acts as an additional incentive to switch to the modern technology.18 In a recentpaper, Greenstone and Jack (2013) argue that households in developing countriesact as if they had small valuation towards health, so the extra incentive to switchgenerated by a is likelly to be small. In the extreme case where it is zero, the op-timal demand for artificial illumination reverts to the solution in in Zilberman andLiu (2011).

Random voucher allocation described in section 1.2 reduced I for a random subsetof households, increasing their demand for electric connections. This increase, inturn, decreased kerosene use, thus reducing a. The change in a that arises fromvoucher allocation is what constitutes the treatment effect of interest in this paper.

Treatment effect is the change in a induced by the change in connections cost I.From (1.3), it is straightforward to show that

∂a

∂I= −∂D

∂If(l) < 0 (1.4)

Thus, empirically, the effect of a reduction in cost of connection will be a directfunction of pre-electrification level of overnight PM2.5 concentration and the prob-ability of taking-up the voucher. This means that the standard assumption of aconstant treatment effect will likely be violated given the heterogeneity in kerosene

17Cooking may affect indoor ambient quality even if it is conducted outside, since the fumes canfilter into the dwelling depending on wind conditions.

18By adding further assumptions to the shape of the utility function we can derive explicitdemands for l and z, and the resulting quantity of a, but this is not the main aim of this paper.


use observed in the sample. In the empirical approach, discussed in the next section,we will first implement the traditional estimators that assume constant treatmenteffect and explore heterogeneity in the traditional way (by interacting treatment withnumber of kerosene lamps owned by the household) and then apply a non-parametricprocedure that allows estimating the full distribution of treatment effects.

1.4 Empirical Approach

In this section we describe the econometric approach on which our empirical estimatesare based. First we describe the standard parametric approach used for our mainresults, both for the experimental and the non-experimental sample.

Standard Parametric Approach

First Stage: Voucher Allocation and Electrification - In this section wepresent the first-stage relationship between voucher allocation and household decisionto adopt electricity. Since the effect of the voucher is expected to change in time(being stronger at the beginning and wearing off when they expire) we estimate anadoption regression per round. For each t = 2, 3, 4, we estimate:

connectedit = β20V 20i + β50V 50i + γ100V0−100i + λ100N

0−100i + ω

′Xi0 + εit (1.5)

where connectedit indicates that household i had a formal connection at time t,V 20i indicates whether the household received a 20% discount voucher, and V 50iindicates a 50% discount. To control for spillovers we included V 0−100i , the numberof households that received a voucher within a 100 meter radius of household i, andN0−100i , the total number of eligible households around 100 meter radius of householdi, Xi0 are individual baseline characteristics.

In some specifications we estimate an equivalent regression stacking all the roundstogether, and interact each coefficient with a round fixed effect, which allows us toestimate a fixed effects panel data model.

connectedit =4∑

t=2

β20t V 20i+4∑

t=2

β50t V 50i+4∑

t=2

γ100t V0−100i +

4∑t=2

λ100t N0−100i +µi+νt+εit

(1.6), where the variables denote the same as in (2.3) but in addition µi and νt denote

individual and year fixed effects. Given that the vouchers are randomly allocated,


both approaches give the same point estimates. There are slight differences in thestandard errors. The main advantage of (1.6) over (2.3) is that it allows testinghypotheses on the coefficients by rounds.

Alternatively, in specifications in which the point estimate of the vouchers aresimilar for both values, we pool both types of voucher together. This does not changethe conclusions, but it makes the analysis of the reduced form more straightforwardlater on:

connectedit = β × voucheri + γ100V 0−100i + λ100N0−100i + ω′Xi0 + εit (1.7)

At each round households can be classified as follows: Always-Takers, Never-Takers, and Compliers. Always-takers are households that would have connectedeven without the discount. Next, Never-takers are households that would not connecteven if they received the discount. Finally, Compliers are households that wouldconnect if and only if they receive the discount.

Note that the status of a household as a complier, never-taker or always-taker, iscontingent on the period. In particular, a household may be a complier at round tand an always-taker at round s > t. For instance, take a household that without thevoucher would have decided to connect in period 3, but with the voucher decides toconnect in period 2. This household is a complier in period 2, but an always-takerin period 3. Despite its not possible to know which household is a complier, the sizeof the complier subpopulation is given by the estimates of the β coefficients in (2.3),(1.6) and (1.7).

Second Stage: Electrification and PM2.5 concentration - To study the effectsof electricity, we exploit the experimental variation in connection fee describe in thepreceding section to instrument for connection to the grid. At each round, the firststage is given by (2.3), and the second stage is given by:

yit = δcônnit + ψ′Xi0 + εit (1.8)

, where yit indicates the outcome of interest (measures of PM2.5), Xit includesbaseline covariates, and εit is a disturbance term. δ is the main coefficient of in-terest, as it gives the causal effect of connection on the outcome for the populationof compliers. The complier subpopulation may be small because either (i) there issmall take-up among the encouraged group, or (ii) there is large take-up among thenon-encouraged group. In our case, the small complier subpopulation (14% of thesample in round 2, 18% in round 3 and 14% in round 4) is due to large take-up inthe non-encouraged group (35% take-up by round2, 50% by round 3, 70% by round


4).

Reduced Form: Voucher Allocation and PM2.5 concentration - Due to asmall complier subpopulation, the IV the point estimates are noisy, which generateslarge standard errors. To avoid relying in noisy estimates, our main results will bebased in the reduced form estimates. Effectively, these estimates report the effect ofreceiving a voucher on the outcome of interest, and as such, these estimates are infor-mative and valuable from a policy perspective. In addition, note that given imperfectcompliance, these estimates represent a lower bound of the effects of electrification.

The reduced form is given by:

yit = βvoucheri + γV0−100i + λN

0−100i + ω

′Xi0 + εit (1.9)

We estimate two variants of (1.9). First, we will exploit the minute-by-minutenature of the data. To allow arbitrary structure in the covariance matrix within ahousehold, we cluster the standard errors at the household level. Second, we col-lapse the data at the household level and run the regressions at the (in the spiritof a “between” estimator) to show that the significance in the coefficients on thevoucher variables is not driven by the large number of observations, we also collapsethe data.19 Given that the first-stage analysis shows a positive and significant rela-tionship between voucher allocation and grid connection, we find it difficult to argueagainst electrification being the channel through which vouchers affect IAP.

Non-experimental Approach: Fixed Effects Estimation - In addition tothe first-stage and reduced-form analysis, we present non-experimental estimatesas further supportive evidence that the channel through which voucher allocationcausally affects PM2.5 concentration is in fact household electrification. For thispurpose we use a sample of households from neighboring regions used in Barron andTorero (2013). These are approximately 250 EHEIPCER households that had notconnected to the grid by round 2. The first PM2.5 measurement was conducted inround 2, and follow-up measurements were taken in rounds 3 and 4. Appendix 1shows some summary statistics describing the sample.

To test for a longitudinal relationship between electrification and PM2.5 we esti-mate the following regression:

yit = δ × connectedit + λt + µi + εit (1.10)19The difference between this procedure and the “between” estimator is that the latter requires

collapsing the minute-by-minute data from all the rounds to a single point for each household andestimating the effect based on the cross-sectional variation in the resulting sample.


, where yit is the outcome of interest for household i at time t, connectedit in-dicates whether the household had a connection to the grid, λt capture round fixedeffects, and µi capture household fixed effects. Causal identification of the parameterof interest δ in this setting requires assuming that connectedit is uncorrelated withthe disturbance term εit after controlling for time invariant characteristics. This cannever be tested, but one way to show supportive, albeit indirect, evidence on this as-sumption is showing that the outcome variable followed the same time trend betweenconnected and off-grid households before the former group got a connection.

To test for parallel pre-electrification trends, we use the fact that we observethe households in the non-experimental sample in four points in time, from 2009 to2013 (although there are no PM2.5 measurements in 2009). By 2010 no householdin this sample had a grid connection. We define three groups based on the timingof grid connection: T1, which is formed by households that connected by 2011; T2,formed by households that connected by 2012; and C, formed by households that didnot connect during the study period. Under the assumption of parallel pre-treatmenttrends in PM2.5 concentration, the trend in T2 should be parallel to that of C between2010 and 2011.

yit = θ2 × T2i + λt + µi + εit (1.11)

In the regressions that test parallel pre-electrification trends in PM2.5 we compareT2 to C from 2009 to 2010. We also compare it by hour of the day (from 5pm to 6pm,from 6pm to 7pm, up to 6am to 7am) and fail to reject the hypothesis of parallelpre-electrification trends (i.e., in every specification θ2 results non-significant).

Furthermore, since these are EHEIPCER households, we also have their full sur-vey data starting from 2009. This allows testing for parallel pre-electrification trendsin both T1 and T2 compared to C in energy expenditure by source (kerosene, candles,gas) and in use of wood for cooking between 2009 and 2010.

yit = θ1 × T1i + θ2 × T2i + λt + µi + εit (1.12)

Neither θ1 nor θ2 result significant in any specification, thus we fail to reject thehypothesis of parallel pre-electrification trends in every case.

Non-parametric Approach: Changes-in-Changes

As discussed earlier, treatment effect is likely to vary depending on each house-hold’s pre-electrification PM2.5 concentration, which would violate the assumptionof constant treatment effect. To examine possible violations to this assumption,we employ changes-in-changes (CIC), a non-linear, non-parametric generalization of


differences-in-differences (DID) developed by Athey and Imbens (2006) that, amongother features, does not impose the assumption of constant treatment effect. In thissection we summarize the main points of CIC and towards the end we draw a parallelwith this study.

The intuition behind this methodology is straightforward when compared to DID.The DID approach uses the change in means in the control group to estimate thecounterfactual change in means in the treatment group and derive an average treat-ment effect on the treated group. Under less restrictive assumptions than DID, CICuses the change in the full distribution of the outcome variable in the control groupto estimate the counterfactual change in the distribution in the treatment group,and derives not only the average treatment effect on the treated, but also the distri-bution of those treatment effects. This feature makes it especially appealing for ourpurposes. Unlike DID, where time and group dimensions are treated symmetrically,CIC allows the effects of time and treatment to differ systematically across groups.It also estimates the entire counterfactual distribution of effects of the treatment onthe treated and of the treatment on the control group.20

The model assumes that the outcome variable Y for an individual in the absenceof the intervention is given by

Y N = h(U, T ) (1.13)

, where h is a “production function” that depends on the period T , and an unob-servable variable U . Y N is the outcome of interest under no treatment. In this case,treatment assignment will be defined by voucher allocation. We can think of theunobservable U as being household’s ability and preferences toward health or a cleanenvironment.

The identification in CIC relies on the assumption that in each time period theproduction function is monotone in an unobservable U . The distribution of U may infact differ between treatment and control groups but, absent treatment, householdswith the same U must have the same value of Y N irrespective of their voucherassignment. Note that Y N is the counterfactual value of Y11 under no treatment.

Formally, CIC is based on the following assumptions21:

Assumption 1 The model: The outcome of an individual in the absence ofinterventions satisfies the relationship Y N = h(U, T )

20In addition, CIC allows for the possibility that the treated group took-up treatment becauseof greater benefits than the control group. We do not make use of this property, since voucherswere randomly allocated.

21Athey and Imbens (2006) shows that DID can be obtained by adding assumptions...


Assumption 2 Strict monotonicity: The production function h(u, t), whereh : U× {0, 1} → R, is strictly increasing in u for t ∈ {0, 1}

Assumption 3 Time invariance within groups U ⊥ T |G

Assumption 4 Support. U1 ⊆ U0.

Assumption 1 requires outcomes not to depend on the group indicator directlyand that all relevant unobservables can be captured in a single index U . In this study,groups are defined by voucher allocation. Given random group assignment, there isno reason to believe that outcomes would depend directly on the group indicator(other than because vouchers increase the probability of electrification, and that willchange the production function from hN to some other function hI). Assumption 2requires higher unobservables to correspond to strictly higher outcomes. Householdswith higher ability or higher valuation for health could have better indoor environ-ment, by using less kerosene or better ventilation rates, for instance. The assumptionabout the unobservables cannot be tested but it is consistent with basic intuition,and it allows for non-additive structures that arise in economic models. Assump-tion 3 requires the population of agents within each group not to change over time.In other words, it requires any differences between the groups to be stable, whichseems feasible given random voucher allocation. This allows trend in one group tobe used to estimate the counterfactual trend in the in the other. Note that underthis assumption any change in the variance of outcomes over the time period will beattributed to changes over time in the production function.22 Finally, assumption4 implies that Y10 ⊆ Y00 and YN11 ⊆ Y01, i.e. we can only estimate the distribu-tion function of Y N11 on the support Y01. The minute-by-minute dataset on PM2.5concentration gives adequate support in both groups and in both time periods.

We start by defining F−1Y (q). For q ∈ [0, 1] and for a random variable Y withcompact support Y,

F−1Y (q) = inf{y ∈ Y : FY (y) ≥ q} (1.14)

Athey and Imbens (2006) show that under the assumptions discussed above thedistribution of F−1Y is identified and that the counterfactual value density Y

N for thetreated group in the second period is given by:

FY N ,11(y) = FY,10(F−1Y,00 (FY,01 (y))

)(1.15)

22As noted by Athey and Imbens (2006), this contrasts with DID with full independence (whichrules out changes in variance), and DID with mean independence (which ignores these changes).


The mechanics of this calculation are straightforward. Start by taking a value ofthe outcome variable y0 in the treated group before treatment, corresponding to apercentile q0 in FY,10. We want to find the counterfactual distribution of Y11 for thetreated group, FN11 , which we do not observe. To estimate F

N11 , first find the quartile

q′0 corresponding to y0 in the CDF of Y00. Next, measure the change in the outcomevariable between periods 0 and 1 for the percentile q′0 and note the change in y0, ∆y0.Then calculate y1 = y0 + ∆y0 and attribute it to the original q0. The point (y1,q0)lies on the counterfactual CDF of Y11. The full counterfactual CDF F

N11 is obtained

by repeating this procedure for all values of y in the domain.The average treatment effect in CIC is then given by23,24:

τCIC ≡ E[Y11 − Y N11

](1.16)

= E [Y11]− E[FY,10

(F−1Y,00 (FY,01 (y))

)](1.17)

In section 1.5 we show that the connection rate among the non-encouraged group(i.e. households that did not receive discount vouchers) started catching up with thatof the encouraged group by round 4 of the survey. We exploit this fact to analyzewhether overnight PM2.5 concentration in the non-encouraged group also caught upwith the encouraged group.

As discussed above, vouchers gave incentives to adopt electricity in rounds 2 and3 instead of in later rounds. Thus, vouchers can be thought of as an incentive notto connect in round 4. Effectively, the electrification rate among voucher recipientsbarely changed between rounds 3 and 4. On the contrary, non recipients had afaster electrification rate between those rounds, almost catching up with voucherrecipients. Note that this change in electrification rates was also exogenously inducedby random voucher allocation. But in this case we should think of voucher recipientsas “control” group between rounds 3 and 4, and non-recipients as “treatment” groupin that period. Hence, we will use voucher recipients in rounds 3 and 4 to estimateFY,00 and FY,01 and non-recipients to estimate FY,10 and FY,11, respectively.

1.5 Results

This section presents the main results of the paper, starting with the experimentalestimates on the reduction in PM2.5 concentration among voucher recipients, which

23This contrasts with the treatment effect in DID: τDID = E [Y11]− (E [Y10]− (E[Y01]− E [Y00])24 Athey and Imbens (2006) draw a parallel with quantile estimation, which they name quantile

differences in differences (QDID). The authors show that CIC has more desirable properties thanQDID, so we will concentrate on CIC.


because of imperfect compliance are a lower bound on the effects due to electrifi-cation. Next, we show some non-experimental estimates to support the fact thatthe changes are in fact driven by electrification and are not an artifact of the data.Finally we present the non-parametric estimates that show that the non-encouragedgroup caught up with the encouraged group towards the end of the study (by whenelectrification rates were similar in both groups), thus increasing our confidence inthe internal validity of the estimated treatment effects.

PM2.5 Experimental Estimates

Adoption of Electric Connections - Table 2 reports the first stage results. Boththe low- and high-discount vouchers increase the probability of adoption of a formalconnection. Individual discount vouchers made households 11 to 19 percentage pointsmore likely to connect to the grid. The effect of low-discount and high discountvouchers is roughly similar. This suggests that either demand for connections isinelastic in this price range, or that vouchers are operate through, like increasingawareness about the electrification program, reducing credit constraints, or nudginghouseholds against procrastinating in the decision to connect.

There are important spillover effects in rounds 2 and 3. Holding the number ofeligible neighbors constant, an additional voucher allocated within 100m of householdi increased the probability of connection by 4 percentage points. Since we had noprior on the reach of the spillovers, we also explored larger radii. The coefficient onvouchers within 100-200m radius is not statistically significant, as shown in columns3 and 4. Adding the number of vouchers in the 200-300m radius to the specificationin column 4 has the same effect (not reported).

The effects of the individual discounts and voucher density decrease by round4. This is to be expected, since by this time roughly 80% of treatment group hasconnected to the grid, closer to the equilibrium connection rate. We exploit the factthat non-encouraged households started catching up with the encouraged group byround 4 of the survey to test whether these groups caught up also in outcomes. Thisis a further test of randomization that is rarely possible to do in empirical papers.

IV Results - Table 3 reports the results of the IV estimation for rounds 3 and4. The dependent variable is PM2.5 concentration (in logs) between 5pm and 7am,followed by sub-periods: from 5 to 10pm, when members are likely to need artificialillumination and are awake; from 10pm to 4am, when most of household members aresleeping; and finally from 4am to 7am, when household members are awake againand may still need artificial illumination. For comparison purposes, we also haveincluded the non-experimental estimates that will be discussed in the next section.


The IV point estimates are negative and large. They are also consistent with thenon-experimental estimates but their standard errors make them too large to be in-formative. Hence, our analysis will focus on the reduced form.

Reduced Form - Figure 3 shows the Kernel density of PM2.5 concentration byvoucher allocation for the experimental sample in rounds 3 and 4 (the two points intime for which we have experimental PM2.5 data). Panel (a) shows the densities forround 3, while Panel (b) shows the density for round 4. The difference in distributionsin Panel (a) is evident: the distribution of voucher recipients has more mass towardsthe lower PM2.5 levels, while non recipients have relatively more mass on highervalues. In turn, Panel (b) shows that the differences fade out by round 4. Thisis consistent with the first stage point estimates of the vouchers being large andsignificant at round 3 and smaller at round 4. For this reason, we will first analyzethe round 3 data for the experimental estimates. later incorporate the data fromround 4 using CIC models.

Figure 4 plots local polynomial estimates of minute-by-minute PM2.5 concentra-tion by voucher allocation at round 3. Voucher recipients show lower and less variablemean values than non-recipients. PM2.5 concentration among non-recipients is higherup to around 11pm, decreases slightly between 11pm and 5am, and increases againstarting at 5am.

Table 4 shows PM2.5 concentration by treatment arm and presents the mainresults of this paper. The dependent variable in columns (1) and (2) is the meanPM2.5 concentration (in logs) between 5pm and 7am. In column (1) the observationis a household-minute. To allow for arbitrary structure of the covariance matrixwithin a household, we cluster the standard errors at the household level. Column(2) shows that the results are practically unaltered if we average PM2.5 concentrationsover households and run the regressions at the household level. Average PM2.5 was63% lower among voucher recipients compared to non-recipients.25

Columns (3)-(8) show the probability of observing PM2.5 concentrations abovethe threshold pointed in the respective column. Although there are no standardbenchmarks in the environmental health literature, we consider a concentration of67 µg/m3 a useful reference point because over a 24-hour period, this concentrationimplies the same PM2.5 exposure as smoking one cigarette (1200 µg/m

3). Voucherrecipients were 34 percentage points less likely to experience concentrations abovethis benchmark than non-recipients. This effects are huge if we consider that the

25In this sample we removed 17 households that reported average overnight PM2.5 concentrationsabove 4 mg/day. Given the small overall number of households we take the conservative approach ofexcluding them. When they are included in the sample, the estimated reductions are much larger,but still within the confidence intervals in Table 4 (of the order of 90%).


mean for the control group was 36 percentage points. These effects are also presentusing more “natural” reference points, like 50 and 75 µg/m3.

Treatment Effect Heterogeneity - The conceptual framework in section 3raised the possibility of heterogeneity in treatment effects, which we now turn toexamine. First, we follow the stander practice of interacting treatment with thesuspected source of heterogeneity, and next we estimate a CIC model to explore thedistribution of treatment effects.

Table 5 shows the results of these regressions. Effectively treatment effect isdirectly related to the number of kerosene lamps owned by the household at baseline.The effect of electrification on overnight PM2.5 concentration was not significantamong households that reporting owning no kerosene lamps. The same is observedfor the probability of experiencing concentrations above 67 µg/m3. Households thatreported owning one lamp experienced reductions of 53% (= e−0.763 − 1) in averagePM2.5, while households with 2 lamps experience reductions of 68%. Similarly, theprobability of observing concentrations above 67µg/m3 fell by 47 percentage pointsamong households owning one lamp and by 75 percentage points among householdsowning two kerosene lamps. The sample size change across these regressions becauseof missing data on the number of kerosene lamps for some households.

The CIC estimator for the average treatment on the treated is -0.70. Givenrandom group assignment, this is also the average treatment effect. The CIC esti-mator is consistent with the -0.63 found among voucher recipients by round 3. Thisstrengthens the internal validity of our findings to the extent eliminating the dif-ferences in electrification rates led to eliminating the differences in overnight PM2.5concentration.

Figure 5 analyzes the variation in magnitude of treatment effects along the distri-bution of overnight PM2.5 concentration. The percentage reductions in the outcomevariable are of considerable size all throughout the distribution. However, thereseems to be variability in the treatment effect along the distribution. The reductionis significant starting roughly from the 20th percentile, and the size of the effect startsincreasing (becoming more negative) starting at the 60th percentile. This is consis-tent with the intuition behind our study setting: treatment effects are significantabove a certain threshold of indoor air pollution, and higher polluters experiencelarger reductions.

Non-Experimental Estimates

In this section we show the results of non-experimental estimates to provide addi-tional support to the experimental estimates from the previous section. We do not


claim these estimates as causal, but we do show that there exists a solid relationshipbetween electrification, kerosene consumption and overnight PM2.5 concentration ina longitudinal setting. We show four main things. First, households that connectto the grid also reduce kerosene use and indoor air pollution. Second, these changeare not observed before households connect to the grid. Third these reductions aresimilar between groups irrespective of the timing of electrification. Fourth, there isno reversion to pre-electrification PM2.5 levels.

The non-experimental subsample shows a negative correlation between electrifi-cation and PM2.5. In 2010, the (geometric) mean PM2.5 concentration was 142 [95%Confidence Interval (CI) 122-165] µg/m3 (N=201). In 2011, PM2.5 concentration was185 (149-230) µg/m3 among non-connected households (N=119) and 128 (86-190)µg/m3 among connected households (N=46). In 2012, PM2.5 concentration was 108(86-136) µg/m3 among non-connected households (N=78) and 97 (75-125) µg/m3

among connected households (N=90).26

In this subsection we will use electrification status at a given point in time todefine groups, and we will use statistical tools to compare outcomes between thosegroups. None of the households in this sample had a connection by round 2 of theEHEIPCER survey (2010). T1 is the group of households that connected betweenrounds 2 and 3 of the survey (i.e., between 2010 and 2011). T2 is the group ofhouseohlds that connected between rounds 3 and 4 (between 2011 and 2012). T3 isthe group of households that remain unconnected during the study period.

Figure 5 shows the kernel density of PM2.5 concentration between 1700 and 0700hours by group. Panel (c) shows the kernel density estimates for T3 households.The density in 2010 and 2011 clearly overlap with each other. The two-sampleKolmogorov-Smirnov test for equality of distributions (KS) generates a p-value of.176; thus the null hypothesis of equality of distributions cannot be rejected at con-ventional confidence levels. The density in 2012 shows some differences with respectto the density in 2010; with an associated KS p-value< .001 in this case, thus reject-ing the null hypothesis of equality of distributions. Panel (b) shows the same for T2households. The densities overlap between 2010 and 2011 (KS p-value=.380), butthe PM2.5 density estimate shifted to the left 2012 (KS p-value< .001). For T1, thedensities corresponding to the 2011 and 2012 measurements fall to the left of the2010 distribution (KS p-value=.006 and < .001, respectively).

26Kerosene expenditure also shows a negative correlation with electrification in this subsample.In 2010, mean kerosene expenditure was 5.07 (4.57-5.61) US$/month in 2010. In 2011, the figureswere 5.54 (4.85-6.32) US$/month for non-connected households and 3.87 (.15-99.42, due to only2 non-zero observations) US$/month. In 2012, mean kerosene expenditure was 5.65 (4.62-6.92)US$/month among non-connected households and 2.52 (.93-6.81) US$/month among connectedhouseholds.


To allow for direct comparison with respect to the each group’s baseline values,the variables in Figure 6 have been standardized by subtracting the baseline meanand dividing by the baseline standard deviation of their respective group. Panel (a)shows the change in average monthly expenditure on kerosene in 2011 compared tothe 2010 levels by treatment arm, with 95% confidence intervals. There is no changein mean between the 2009 and the 2010 measurements for any of the groups. T3 doesnot show any change in mean kerosene expenditures in any of the surveys comparedto 2010. T1 shows a large reduction between 2010 and 2011, which is maintainedby 2012. T2 shows no change by 2011 (when the group is still off-grid), but a largereduction by 2012 (when the group connects to the grid).

Panel (b) shows the association with fine particulate matter. PM2.5 concentra-tion did not change significantly in the control group between baseline and follow-up.Households in T1 show significant reduction in average PM2.5 concentration at thefirst and second follow-ups. Households in T2 show no reduction in PM2.5 concentra-tion by the first follow-up, but a significant reduction in mean PM2.5 concentrationby the second follow-up. This reduction is not statistically different from the averagereduction experienced by households in T1.

This graph shows three striking facts. First, Both PM2.5 and kerosene expendi-ture change when electrification status changes. Conversely, none changes if electri-fication status doesn’t change (except for an increase in PM2.5 among the Controlgroup in 2011). Second, the average changes among households in T2 are similarto those experienced by households in T1. Third, the new (lower) levels of keroseneconsumption and PM2.5 observed for T1 in 2011 are maintained in 2012.

The results are presented in Table 3, briefly discussed above. Connection to thegrid is associated with 53% reduction in PM2.5 concentration between 5pm and 7am.This estimate is strongly significant. The point estimate is slightly larger in theevenings (57%), when PM2.5 concentrations are 300 µg/m

3. The estimate for latenight is also high, at 51%, but PM2.5 concentration is 140 µg/m

3, around half of theevening period. The reduction in PM2.5 in the early mornings (4am-7am) is 42%, stilllarge but slightly lower than the other subperiods. Average PM2.5 concentration inthis subperiod is the highest, at 350 µg/m3, which could indicate cooking breakfast.As we discussed earlier, cooking fuel did not change with electrification.

Using the data for T2 and T3 to test for differential pre-treatment trends inPM2.5 and the 2010 EHEIPCER wave to test for differential pre-treatment trends inexpenditure in kerosene or candles, use of wood, candles, we cannot reject the nullhypothesis of parallel pre-treatment trends in any of the tests we performed.


Mechanisms

This subsection analyzes changes in traditional fuel use induced by electrificationand suggests kerosene as the main channel through which electrification affectedovernight PM2.5 concentration. Table 6 reports the effects of electrification on energyuse. Our findings conform with the stylized fact that newly electrified householdsuse electricity first for illumination. Electrification reduces the probability of usingkerosene by 60 percentage points by round 2 (significant at the 95% of confidence).The reduction by round 3 is smaller, 22 percentage points (not significant), and byround 4 is larger again, at 43 percentage points (significant at the 95%). Monthlyexpenditure in kerosene decreased accordingly, by $3.57 at round 2, $2.07 at round3 and $1.94 at round 4. The use of kerosene in kWh per month also decreasedwith electrification, by between 50-80% of the baseline value by rounds 2, 3, and 4,respectively.

Other substitutes to electricity, like candles (for illumination) and car battery (tooperate TV sets) show no significant changes, although in some cases the standarderrors are too large to reject sizable effects. These sources are less important in thehousehold’s energy budget than kerosene, so detecting an effect would require largersample sizes. As mentioned in the preceding section, similar patterns arise in thenon-experimental sample.

Next, we examine whether changes in cooking practices could have also generatedthe observed changes in overnight PM2.5 concentration. This is unlikely since the useof wood for cooking was around 85% and did not vary during the study period. Aswe saw in Table 1, 60% of the population agreed that cooking with electricity wasvery expensive. The bottom panels of Table 6 show that electrification did notlead to statistically significant changes neither in the use of wood for cooking nor inthe probability of cooking outdoors. The coefficient on electrification in the use ofwood regression for round 2 is larger than desirable, but the coefficients in rounds3 and 4 are close to zero. Similarly, the coefficient on electrification correspondingthe regression on cooking outdoors in round 4 is non-signficant but large, whilethe coefficients for rounds 2 and 3 are close to zero. Although we cannot rejectmoderate changes in two of the six regressions on cooking practices, it seems mostof the variation in overnight PM2.5 concentration is due to the large and statisticallysignificant changes in kerosene consumption.

Discussion on Effect Size

Taken at face value, these effects may seem too large. Given the first-stage andreduced-form coefficients, the implied IV estimator of the effect of electrification on


overnight PM2.5 concentration is given by:

βIV =βRFβFS

≈ −0.600.20

= −3.0 (1.18)

, which implies a reduction in overnight PM2.5 concentration of 95%27. However,

in this section we argue that an effect this size is not out of the question and presentsome evidence in support of that claim. Note that the outcome of interest is PM2.5concentration between 5pm and 7am in the room of main use at night, in most casesthe living room. In this room, and during this period, kerosene lamps arguablyaccount for the largest share of PM2.5 emissions. Note that Table 6 shows largedecreases in the intensive and extensive margin of kerosene use, which should reflecton large drops in overnight PM2.5 concentration.

28

Next, we show that the drops implied by the model are consistent with theraw percentage changes in PM2.5 concentration among households in our sample.We have successful PM2.5 measurements for 85 households in rounds 3 and 4. Wecalculated the change in PM2.5 concentration for these households. Figure 8 shows thehistogram of these changes.29 Forty households show reductions of 50% or higher,14 show reductions of up to 50%, and 30 households show increases in overnightPM2.5 concentration. A few households show large percentage increases, but theselarge increases correspond to households with specially low levels of overnight PM2.5concentration at round 3. In levels, these increases are rather small. Conditional onshowing an increase in PM2.5, the average increase in levels was 0.051 mg/m

3, (95%CI 0.004 - 0.099 mg/m3).

27In this case it is not possible to rely on the approximation yT−yCyC ≈∂lny

∂connected = β1, sinceβ1 = −3 is of considerable magnitude. The exact percentage change is given by:

ln yT − ln yC = ln(yTyC

)= β1

yTyC

= eβ1

yT − yCyC

= e−3 − 1 = −0.95

28Some emissions from cooking may filter during the day and linger in the living room, and, aswe have seen, cooking practices did not change with electrification. However, there is no reason tobelieve that PM2.5 concentration in the living room would depend more on filtrations from biomassduring the day than from direct use of kerosene lamps in that room during the evenings.

29Three households show percentage increases higher than 2.5. They are not included in thefigure.


1.6 Implications for Health Outcomes

This section analyzes the health implications of the observed reductions in PM2.5 con-centration. First, we show that acute respiratory infections (ARIs) among childrenunder the age of 6 were lower among voucher recipients. Next, we combine PM2.5concentration with time allocation data to construct measure of exposure to PM2.5for four typical household members (adult male, adult female, male child, femalechild), which allows gauging further health implications of the observed changes inPM2.5 concentration.

Acute Respiratory Infections Among Children

Lower respiratory infections cause 2.8 million deaths globally in 2010 (Lozano et al.,2013) and thus they constitute a major public health concern. In this subsection weshow that the reductions in overnight PM2.5 concentration generated by householdelectrification had sizable effects on respiratory infections among children under sixyears old. The experimental sample includes 380 children in this age-range. Despitethis relatively small sample size, there are large and statistically significant (at the90% of confidence) reductions in the incidence of acute respiratory infections amongchildren.

The dependent variable indicates whether the child had an episode of acute res-piratory infection in the four weeks prior to the survey (self reported). When theexplanatory variables are voucher, round, and their interactions, we find that vouch-ers led to a reduction of 16 percentage points at round 3. Including subdisdrict fixedeffects increases the magnitude to 19 percentage points. Adding baseline charac-teristics and the number of vouchers within 100m results in reductions of 18 and17 percentage points. In all cases the reductions are significant at the 90%. Thesefigures are large, and even more impressive when compared to the baseline sincethey represent reductions of between 37 to 44 percent of the mean incidence amongnon-recipients at round 3.

The coefficient on voucher at round 4 is positive across all specifications, implyingincreases in ARI incidence of 8 to 11 percentage points among voucher recipients,and it is significant at the 90% in the first specification. Moreover, in no specificationthe null hypothesis that the coefficients on voucher in both rounds have the samemagnitude (but opposite signs) can be rejected. This may initially seem a puzzlebut it merely reflects that the electrification rate among non-recipients caught upwith that of voucher recipients by round 4. It is worth noting that it is not thecase that ARIs bounced back up to their original levels, since ARI incidence reducedfrom 44% to 10% between rounds 3 and 4. So, consistent with the analysis of PM2.5


concentration, this shows that the effects of electrification are similar irrespective ofwhether a household received a voucher.

PM2.5 Exposure and Health Risks

In this subsection we construct measures of PM2.5 exposure based on PM2.5 concen-tration and time allocation data from the household survey. The resulting exposuremeasures are lower for the encouraged group (voucher recipients), but the gains areunequally distributed among household members. Adult males experience reductionsin exposure to PM2.5 nearly twice as large as adult females, mostly due to inequalityin time spent cooking, where PM2.5 concentration is highest, and outdoors, wherePM−2.5 concentration lowest. It is important to keep in mind that given imper-fect compliance with voucher assignment, the reduced form coefficients (the effect ofvoucher on overnight PM2.5 concentration) constitute lower bounds of the effects ofelectrification.

We start by defining exposure to a pollutant as the amount (in mg) of pollutantthat effectively makes its way into a person’s respiratory system. (As discussedearlier, PM2.5 enters the deep lung and the bloodstream.) For a particular activityconducted for a given amount of time, exposure is estimated as the product of theconcentration in the environment where this activity took place multiplied by theinhalation rate while performing said activity. Daily exposure can be estimated byadding over all activities performed during the day, as in the following equation:

Exposure =J∑

j=1

timej × concentrationj × inhalationj (1.19)

where timej is time spent in activity j (hours), concentrationj is PM2.5 concen-tration in the room while performing activity j (mg/m3), and inhalationj is theestimated inhalation rate while performing activity j (m3/hour).

The detailed information on time allocation of the household members gives ourstudy an important advantage over the typical study in this literature.30 It allowsto estimate fairly accurately the time individuals spent indoors and, moreover, thetype of activity, they were performing. Knowing the type of activity and the timeallocated to said activity allows inputting an average inhalation rate adequate forthat particular activity during that time, instead of just inputting the average intakeper day. For a given pollutant concentration, this leads to better estimates of theamount of pollutants that effectively makes its way into a persons respiratory system.

30This is even more so when compared to outdoor pollution studies, in which time spent outdoorsis an unobservable variable


The time allocation data was collected for up to four household members: thehousehold head, his or her spouse, and up to two school-age children. This allowsestimating PM2.5 exposure for four “synthetic individuals”: adult female, adult male,female child, and male child. Table 8 presents average time allocation in four type ofactivities for each of our synthetic individuals. The male and female heads report 8.9hours of sleep per day, while the children report 9.5-10 hours of sleep per day. Timeat home during the evening (awake) is similar for all members (slightly lower forthe children, who sleep 0.5-1.0 hours more than the household heads). The starkestdifferences are observed in time spent in the kitchen and time spent outside the home.While female head reports 2.5 hours per day in the kitchen, the male head reportsspending an average of just five minutes. On the other hand, the female head reportsspending an average of 2.7 hours outside the home, while the male head reports 7.8hours (consistent with the length of a work day).

The differences in time allocation that arise from this analysis already suggestthat adult females are more exposed to PM2.5 (as well as other pollutants) since theyspend considerably more time in the kitchen than any other household member. Onthe other hand, males spend almost one third of the time outside the home. Themain activity is farming and walking to and from the farm, where it can be safelyassumed that exposure to PM2.5 is negligible.

Next, we make explicit the assumptions about the PM2.5 concentration in theenvironments where these activities were likely conducted. As shown earlier, averagePM2.5 concentration in the living room during the evenings is 0.40 mg/m

3. We takethis as representative of any room in the household, except the kitchen, between1700 hours in the evening and 0700 the next morning. Based on the subsample ofhouseholds for which we have 3-day measurements, we estimate the average PM2.5concentration in the living room during daytime (from 0700 to 1700 hrs) to be 0.26mg/m3. We take this as representative of the rooms in the household during daytime,again with the exception of the kitchen. Since we did not collect data on PM2.5 con-centration in the kitchen, we use 0.90 mg/m3, which corresponds to average PM2.5in the kitchen in Guatemalan households (Northcross et al., 2010). This figure seemsan adequate assumption in our context since it corresponds to a neighboring regionwhere households also rely of fuelwood for cooking. This makes our exposure esti-mates adequate for households that rely on wood for cooking, and even conservativegiven that its not uncommon to find cases where the average concentration is above2.0 mg/m3. We assume household members will not be exposed to PM2.5 wheneverthey are not home. This assumption seems not to be too restrictive for the pop-ulation in our study setting, since most of the time outside the home is spent inoutdoors activities like farming, and very little time is spent conducting activitiesoutside the home that suggest exposure to PM2.5 (e.g. visiting friends at night).


The third and final component in (1.19) is the inhalation rate. Since inhalationrate depends on age, we estimated it for the sample averages: 43 for the female head,47 for the male head, 11 for the female child, and 13 for the male child. Air inhalationrates per activity are based on the EPA Exposures Handbook (EPA, 2011). Mostactivities conducted at home can be classified as “light activity tasks” by the EPA.Light activities include cooking, washing dishes, ironing, watching TV, doing deskwork, writing and typing, and walking at a speed of up to 2.5 mph (2.9 km/h). Theaverage inhalation rate for these activities is 0.78 m3/hour, while the average airinhalation rate while sleeping is 0.30 m3/hour, again simliar for the four syntheticindividuals. The inhalation rate for activities conducted outside the home will varygreatly, depending on the intensity of these activities. For instance, walking to workcould be classified as light or medium intensity, depending on the speed at which theperson is walking. Farming, on the other hand, could be classified as medium to highintensity, but if lunch breaks would be light activity. However, the assumption madeearlier about PM2.5 concentration being zero outside the home makes the inhalationrates of these activities irrelevant for total exposure.

With these three components we estimated exposure rates for the four syntheticindividuals. Estimated exposure measures are highest for the female head, at 5.68mg/day, and lowest for the male head, at 3.20 mg/day. The exposure measuresfor children lie in between, with females 4.23 mg of PM2.5 per day males to 3.72mg/day. Taken plainly as units of PM2.5, these concentrations are equivalent to 8.0cigarettes a month for the male head, 14.2 for the female head, 10.6 for the femalechild, and 9.3 for the male child31. The scientific evidence is yet inconclusive as towhether generated by PM2.5 cigarette is worse or not than that generated by kerosenecombustion.

The changes in exposure are large for all members (all of them above 30%),but these gains are unequally distributed across household members. The male headbenefits the most, with a reduction in exposure of 59%, while the female head benefitsthe least, with a reduction of 33%. As pointed above, these differences owe to femalesspending more time than males in the kitchen, where pollutant concentration ishighest, while males spend more time than females outside the home, where pollutantconcentration is lowest.

To date, there are no dose-response functions linking exposure to PM2.5 fromkerosene combustion to health outcomes. However, Pope III et al. (2011) presentsan estimate of a dose-response function linking PM2.5 from first- and second-handtobacco smoking to lung cancer and cardiovascular diseases. In Panel (b) of Table 8we present the relative risks that would be associated to the exposure levels found in

31One cigarette is estimated to have 12mg of PM2.5.


Panel (a) if the health effects of PM2.5 from kerosene combustion were similar to thosefrom tobacco smoking. It is worth noting that the dose-response function estimatedby Pope III et al. (2011) is non-linear and it has support in exposures of from 0.18 to0.90 mg/day and then above 18 mg/day (but not between 0.90 and 18mg/day, whileour estimates range from 3.2 to 5.7) so we need to rely on the linear interpolation ofthe values up to 18mg/day. This caveat does not seem a strong weakness since thelinearization is highly accurate in this neighborhood, with R-squared values of 0.99(lung cancer), 0.96 (ischemic heart disease), 0.86 (cardiovascular disease), and 0.80(cardiopulmonary disease).

The changes in exposure are associated with a decrease in relative risk of lungcancer (compared to a person with no exposure to PM2.5) from 4.0 to 3.1 for thefemale head, nearly a 25% reduction. The relative risk for the male head falls by33%, while the reduction for female child is 25% and for the male child is almost 30%.The estimated reductions in the relative risk of ischemic heart disease, cardiovasculardisease, and cardiopulmonary disease are between 3 to 4%. Consistent with theresults in lung cancer, these changes are higher for adult males, but the differencesat these levels of exposure are relatively small.

1.7 Conclusions

This paper provides the first experimental evidence that electricity leads to importantimprovements in welfare, through substantial, immediate and sustained improve-ments in indoor air quality. To this end, we pair a clean experimental design withan exceptionally rich dataset with minute-by-minute data on fine particulate matter(PM2.5) concentration and time allocation for a sample of households in Northern ElSalvador during a recent electrification program. In this program, households had topay a $100 fee to connect to the grid (20% of the sample annual per capita income).By randomly allocating discount vouchers towards this fee we generated exogenousvariation in the cost of connection to the grid.

Given imperfect compliance with voucher assignment, our experiment produceda lower bound for reductions in average PM2.5 concentration, of the order of 60%. Inaddition, in our experimental sample, voucher recipients are 38 percentage points lesslikely to report PM2.5 concentrations above 67 µg/m

3 in a given minute than non-recipients (over a 24-hour period, this concentration implies the same PM2.5 exposureas one cigarette). The most salient mechanism behind the improvements in IAP isa substitution away from kerosene lighting, while there are no discernible changesin the use of other traditional lighting sources or in cooking practices. Given thatthe mechanism is clear in this context, our results suggest that other clean artificial

Essays on Household Electri cation in Developing Countries · 2018. 10. 10. · Doctor of...

Documents

Transcript of Essays on Household Electri cation in Developing Countries · 2018. 10. 10. · Doctor of...