The Demand for Off-grid Solar Power: Evidence from Rural

India’s Lopsided Retail Power Market ∗

Robin Burgess†
Michael Greenstone‡
Nicholas Ryan§
Anant Sudarshan¶
March 20th, 2019
Abstract
Over a billion people, nearly all in South Asia and Sub-Saharan Africa, do not have electricity
in their homes. Solar panels have come down sharply in price, making off-grid, distributed
solar systems a plausible alternative to grid electrification. This paper reports the results of
a randomized-control trial that offered off-grid solar systems at varying prices in one hundred
sparsely electrified villages in Bihar, one of India’s poorest states. We find that demand for
solar is near zero at unsubsidized (2013) prices but rises with experimental cuts in price and
improvements in quality over time. The high price elasticity of demand for solar is due to
a surprisingly competitive retail market where off-grid solar competes with diesel generators
and the state-run grid. The model estimates show that solar is a stop-gap: relatively well-off
households strongly prefer the grid to off-grid electricity, and the willingness-to-pay of the
average household for solar is twice as high when the grid is absent. We use counterfactuals
to understand how grid policy, namely supply rationing and subsidized pricing, shapes this
conclusion. We find that massive subsidies to the grid limit solar market share. Removing
these subsidies to level the competitive playing field would devastate grid market share among
the rural poor and reduce electricity access. On supply, increases in rationed supply, from the
present mean of eleven hours per day, are not valued enough by households to cover their cost.
Our analysis can therefore rationalize why the state, to promote access for the rural poor,
serves them with a seemingly dysfunctional combination of subsidies and rationing.

I Introduction

In 1879, the year Thomas Edison invented a workable light bulb, the world’s population was

around 1.2 billion (The Maddison Project, 2013). Today, a higher number of people, nearly all in

South Asia and Sub-Saharan Africa, do not have electricity in their homes. A surge in initiatives by

governments and development organizations aims to provide these people with access to electricity.1

These initiatives are based on the idea that people value power and so widening access will benefit

households who are now unelectrified.

The advent of off-grid solar systems has opened a new way to electrification. The traditional

mode of electrification, large-scale grid expansion with infill connections of households over time,

has high fixed costs shared over a large population, generating large economies of scale. The
∗ We thank Manoj Sinha and Col. Baljit Singh of Husk Power Systems for their partnership in implementing the

project. We thank Rashi Sabherwal, Shruti Bhimsaria, Aditya Petwal and Rakesh Pandey for excellent research
assistance. We thank the Shakti Sustainable Energy Foundation, the LGT Venture Philanthropy Foundation, US
AID and the International Growth Center (IGC) for financial support. All views and errors are solely ours.
† London School of Economics, r.burgess@lse.ac.uk
‡ University of Chicago, mgreenst@uchicago.edu
§ Corresponding author. Yale University, nicholas.ryan@yale.edu.
¶ University of Chicago, Energy Policy Institute at Chicago – India, anants@uchicago.edu
1 USAID, for example, launched Power Africa in 2013 and DFID launched Energy Africa in 2015.

1
extension of the grid and widespread electrification in rural areas has typically lagged urban elec-

trification by many decades (International Energy Agency, 2017). Off-grid, decentralized solar also

has high fixed costs, but fixed costs which can be paid by even a single, isolated household. The

ready nature of solar technology has therefore spurred hope of a faster, greener path to universal

electrification.2 Former UN Secretary General Ban Ki-moon proclaimed “Developing countries can

leapfrog conventional options in favor of cleaner energy solutions, just as they leapfrogged land-line

based phone technologies in favor of mobile networks.”3

Will off-grid solar change how the poor gain electricity? How does the value of this innovation

depend on competition from other electricity sources? Off-grid solar systems are sharply limited

in the load they can serve and, lacking economies of scale, have very high costs per unit of energy

supplied. It is therefore not clear to what extent households can use solar as a substitute for the

grid, or only as a stop gap, while waiting until the grid arrives.

This paper studies the value of electrification for the poor. The analysis uses a randomized

experiment in Bihar, India on the pricing and availability of off-grid solar power to estimate the

benefits of solar micro-grids for households. We then broaden our definition of electrification to

encompass household choices between grid electricity, solar micro-grids, people’s own private solar

systems and diesel power. This wider scope allows us to address policy-relevant questions about

how the regulation and pricing of different electricity sources affect household access and surplus.

Our setting is especially apt to describe household choices between competing sources of elec-

tricity. The experiment allows sharp estimates of household willingness-to-pay for electricity. We

also ran our experiment during a period of transformative change in the electricity sector of Bihar.

In fewer than four full years, the share of households with electricity from diesel generators fell

from 17% to 3%, the share with their own solar systems leapt from 5% to 21%, and the share on

the grid skyrocketed from 5% to 41%. Households in our data are poor and use electricity mainly

for lighting and charging their mobile phones. They are therefore willing to substitute rapidly

between sources of power that can serve these needs.

The analysis uses a randomized experiment on the pricing and availability of off-grid solar

power. The experiment was conducted in rural Bihar, India from 2013 through 2017 (Figure 1).

The implementing partner was Husk Power Systems (HPS), a private company that installs and

maintains solar micro-grids for a monthly charge.4 The experiment has three treatment arms: (i)

a pure control arm (34 villages), in which HPS did not offer the system, (ii) a normal price arm

(33 villages), where the system was offered at the prevailing price (initially INR 200 a month, later

cut to INR 160 per month) (iii) a subsidized price arm (INR 100 a month) (33 villages).
2 See, for example, “Africa Unplugged: Small-scale Solar Power is Surging Ahead”, The Economist, October

29th, 2016.
3 As written in “Powering Sustainable Energy for All,” The New York Times, January 11th, 2012.
4 A micro-grid serves up to five households. Each household on the micro-grid gets 25-40 watts of power for

5-7 hours each day. This small quantity of power is used to power a package, supplied with the system, of a
high-efficiency light bulb and an electrical outlet, typically used for mobile phone charging.

2
 We collect data to cover not only off-grid solar systems but the whole rural electricity landscape.

The primary source of data for the analysis is a three-round household panel survey covering

roughly three thousand households in the hundred sample villages. The surveys cover household

characteristics, electricity access and billing, energy services and measures of health and education.

To better understand the supply side of the market, we supplement our household surveys with

three other data sources. First, we surveyed diesel generator operators about their cost structure

and customer base. Second, we collected administrative data from HPS on consumer payments.

Third, we collected administrative data from the state utilities, which supply all grid electricity,

on grid extension, supply and customer payments.

Our analysis and findings are presented in two parts. First, we conduct a reduced-form impact

analysis on the demand for off-grid solar and its benefits for households. Second, we broaden the

scope of the analysis by using the experiment to estimate a demand system wherein households

choose over all competing sources of electricity. We then use the estimated demand system to

study how policies for different electricity sources interact in the marketplace

There are two main results of the impact analysis. First, household demand for micro-grid

solar electricity is highly price elastic. At the prevailing (2013) prices of INR 200 to INR 160 per

month the demand for HPS solar micro-grids was near zero. At the subsidized price of INR 100,

8% of households took up micro-grids. Our supply-side data suggest that competition from other

sources of electricity may put a ceiling on demand for solar. The main competition for solar at

baseline was diesel generators, which most often charged a price of Rs. 100 themselves, and the

grid was expanding rapidly in this period. Though only a fraction of households adopted solar

connections, we find that households in treatment villages are significantly more likely to own light

bulbs (i.e., to have any electricity connection), and that they use more electricity, purchase more

mobile phones and spend less money charging them. As we would expect, these effects are muted

in the normal price treatment arm where adoption was low.

The second main result of the impact analysis is that we do not find any indirect welfare benefits

of solar micro-grids. Our surveys measured welfare outcomes that may be indirectly affected by

electricity use in the areas of health and education. For health, electricity may displace kerosene

use and therefore reduce indoor air pollution and respiratory infections (Barron and Torero, 2017).

We find no significant impact of off-grid solar on self-reported respiratory problems. For education,

electric light may increase study time at home and therefore children’s test scores. We find no

significant impact of off-grid solar on children’s reading and math test scores. The point estimates

for test scores are positive and economically large, however: the experiment is not well-powered

here due to low take-up.

The impact evaluation results suggest that solar micro-grids are valuable to some households,

but far from transformative. There are several objections that may be raised, under the wide

banner of external validity, about using these results to make claims about the value of electricity

3
for the poor in general. First, the poor might value electricity but choose sources other than

our experimental partner’s micro-grids, for example, grid electricity or their own personal solar

systems. Second, the experiment may have been run at a time when solar was just coming into

its own, and future innovation in the product may indeed be transformative. Third, the electricity

playing field in rural India is just not level, and, since solar demand will depend on competing

sources of supply, our experiment does not measure the true value of solar.

This third factor, an uneven playing field in the electricity sector, is critically important in

our setting. We use administrative data on grid power supply, customer payments and costs to

characterize a dysfunctional grid. The grid is absent in many villages. When it is present, the grid

supplies on average 11 hours of power per day and recovers in revenue only 42% of the variable cost

of supply to the rural customers in our sample. One may argue, then, that the only reason solar

gains any market share at all is that grid supply is so heavily rationed. Conversely, the government

heavily subsidizes grid electricity and has a high tolerance for theft, so perhaps off-grid solar cannot

gain market share because it is competing against a money-burning state utility.

Is solar’s market share small because the government is willing to lose money on grid supply,

or because households do not value the product? These different explanations for our results

have very different implications for policy. We therefore use the experiment to estimate a demand

system to allow us to pick apart what forces drive the reduced-form findings. We estimate a

nested logit model of electricity source demand in which households’ indirect utility depends on

their own characteristics and the characteristics of the sources available in their village. The grid,

diesel generators, solar micro-grids, own (private) solar systems and not having electricity make

up the possible choice set. The model combines several features to allow for a rich pattern of

household demand: (i) household panel data to micro-found demand heterogeneity, (ii) variation

in the unobserved quality of electricity sources at the technology-by-village-by-time level and (iii)

experimental variation in price. The combination of variation in unobserved product quality and

an experiment that varies price means we can transparently identify the structural parameters that

drive our counterfactual simulations.

The demand model estimates echo the reduced-form finding that households are highly price

elastic. Raising the price of the grid by two cups of tea per month reduces the household electri-

fication rate by 3 percentage points. The elasticity of grid market share with respect to price is

-0.71 and the elasticity of own solar market share with respect to price is -2.90. A second, novel

finding of the demand model estimates is that richer households have much stronger preferences for

the grid over other electricity sources. The mean willingness to pay for the unobserved attributes

of the grid, across all households, is much greater than that for the three off-grid sources, and

willingness to pay for the grid rises further with household wealth and income. For example, if a

poor household living in house with a thatch roof had a solid roof instead, which is a reliable wealth

proxy, the probability of that household choosing grid electricity would leap up by 11 percentage

4
points. The grid is the only electricity source that reliably supports higher-load appliances like

televisions and fans, in our data, so the preference of the rich for the grid is likely driven by demand

for these services.

The demand system allows us to predict household demand for counterfactual choice sets and

product characteristics. We consider counterfactuals that vary along several different dimensions:

the quality of off-grid solar, the availability of the grid and grid policy towards rationing and pricing.

We use the counterfactuals to understand how off-grid and on-grid sources of power interact.

There are two main findings from the counterfactual scenarios. The first finding is that solar

power is valuable in large part because the grid is incomplete and dysfunctional. We estimate that

household surplus from all electricity sources is 68% higher with solar in the choice set, relative to

a world without off-grid solar. If we project the quality of solar forward to 2022, then solar would

surpass the grid in market share and increase surplus by 91%, relative to a world without off-grid

solar. The total value of solar power to households in this scenario is Rs. 2171 per year (USD

33), or 3% of median annual household income. These values are contingent on the status quo,

incomplete grid. If we value solar when the grid is present in all villages, as compared to the grid

being absent from all villages, we find that the mean value of solar is twice as large when the grid

is absent. This finding and the preference of richer households for the grid both support the view

that solar power is a stop gap until the arrival of the grid rather than a permanent substitute. As

the grid is extended and households get richer, they will demand the grid over off-grid solar.

The second main finding from the counterfactual scenarios is that our household demand system

can rationalize the apparently dysfunctional state of the grid. At the status quo level of grid

adoption in our model, we estimate that the state utility loses Rs. 497 (USD 7.5) per household

per year, through a combination of explicit subsidies and theft. Why is the government willing

to lose so much money? We posit that the answer is that feasible alternative policies devastate

household electricity access. If the state utility were to exit rural areas altogether, so that it did not

lose any money, the share of households without electricity, from any source, would rise from 57 to

64 percent, and household consumer surplus from all sources would fall by 46%. If the state were

to continue to serve rural areas, but forbade theft and raised prices to cover the cost of supply, the

share of households without electricity from any source would rise to 62 percent, and the market

share of the grid would fall from 24% to a meager 4%. In this scenario, household surplus falls by

36%. Therefore pricing grid electricity at cost has nearly the same effect on surplus, for these poor

households, as not building a grid at all.

We also use the model to rationalize why the government rations power to rural areas so

severely. If the government increased power supply during peak hours, total surplus would fall, as

we estimate that household willingness-to-pay is less than the cost of the additional energy that

would be supplied. Finally, we imagine that the government may offer a budget-neutral “grand

bargain” in which the state utility supplies more power to rural areas during the peak hours, but

5
raises prices to cover the additional cost of energy. We find that such a grand bargain would also

reduce household surplus and therefore total surplus. Households value supply during peak hours,

but not enough to be willing to pay its cost.

The highly elastic demand that we observed in the reduced-form impact analysis, when passed

through our demand model estimates, therefore implies that the state budget constraint binds

against the goal of universal electrification. Even with massive subsidies electrification is incom-

plete. If the grid did not lose lots of money, hardly any of the rural poor would be willing to

connect. The demand system also describes why it may be hard for a government with a budget

constraint to offer supply for twenty-four hours a day: the people do not demand it.

Our paper contributes to the literature on the demand for and effects of electrification. Much

of the literature on rural electrification has focused on the spread of grid electricity, and found that

grid electrification has large effects on labor supply, productivity and welfare.5

The closest precedents to this paper are a handful of experiments on the demand for electricity

connections. We know of two experiments on demand for off-grid solar.6 The impact analyses in

these papers are broadly consistent with our findings that demand for off-grid solar and its possible

indirect welfare benefits for households are limited. A recent experiment in Kenya finds that grid

electrification is prohibitively costly for rural Kenyan households, even at heavily subsidized prices

(Lee, Miguel and Wolfram, 2016). This finding agrees with our finding of highly elastic demand

for off-grid electricity among rural Bihari households.

Our paper takes a couple steps to unify this literature. First, we estimate how households

value both grid and off-grid electicity together, in a single demand system, where other work has

considered each source in isolation. This allows us to consider the impact of off-grid solar or

grid extension on electrification in a setting where substitution between these sources is of great

importance. We expect that the choice betwen grid and off-grid technologies, implicitly governed

by the policy environment, will be simiarly important in most areas at the margins of electrification

today. Second, we study the supply side to try to rationalize the seemingly dysfunctional rural

electricity sector. Our model shows that the playing field for off-grid solar is indeed lopsided, due

to massive subsidies to the grid, but that these subsidies can be justified by the government placing

a value on access for the rural poor.

Our study also contributes methodologically to the development literature by placing a greater

emphasis on the external validity of experimental results. Field experiments have lately gotten

longer to address the realism and durability of effects.7 The analysis of experiments more often
5 Rud (2012) shows that electrification leads to structural transformation. Dinkelman (2011) demonstrates that

electrification can lead to increased employment and female empowerment. Lipscomb, Mobarak and Barham (2013)
find large effects of electrification on the UN Human Development Index and average housing values. Barron and
Torero (2017) find that household electrification reduces indoor air pollution.
6 Aklin et al. (2017) find that offering off-grid solar power in Uttar Pradesh, India increased electrification rates

by 7 pp but had no effect on socio-economic outcomes like expenditures, business creation or time studying. Grimm
et al. (2016) find that household willingness-to-pay is high relative to incomes but that nearly no households are
willing to pay market prices.
7 De Mel, McKenzie and Woodruff (2013) study firm formalization with a 31-month followup, Dupas and Robin-

6
uses structural models to understand how an intervention works and to go from reduced-form

treatment effects to policy counterfactuals (Duflo, Hanna and Rya, 2012; Bryan, Chowdhury and

Mobarak, 2014). Our experiment offered long-term subsidies on solar technology, on a standing

basis for everyone in treatment villages, and tracked take-up and welfare outcomes over three-

and-a-half years. The experimental variation therefore closely mimics real-world variation in the

price of a technology. We use this variation to recover the price elasticity of demand for electricity

sources in a discrete choice demand model.8 We use the model estimates to study counterfactuals

on grid policy that are well beyond the boundaries of the experiment itself.

The rest of the paper is organized as follows: Section II describes the experimental design and

the data, Section III presents the reduced form results from the randomized control trial, Section IV

details our model of demand for electricity connections, Section V presents the demand estimates

and counterfactuals and Section VI concludes.

II Experimental Design, Data and Context

This section describes the setting for the study, the experimental design and the data collection.

We then use the data to characterize the electricity sources competing within our study sample.

We conducted our experiment in Bihar, one of the poorest and most energy deprived states in

India. Table 1 compares the percapita GDP, access to electricity, and annual per-capita electricity

consumption of the United States, India, Sub-Saharan Africa, and Bihar. Bihar’s per capita GDP

is a quarter of that for India as a whole and its electricity consumption per capita only 15% of

the Indian average. The mean Bihari electricity consumption of 122 kWh per capita per year is

a quarter of that in Sub-Saharan Africa. At this average level of consumption, each person can

power two light bulbs totaling 60 watts for six hours per day through the year.

II.A. Experimental Design

We partnered with Husk Power Systems (HPS) to vary the availability and price of solar micro-

grids in a randomized-control trial. Husk Power was founded in 2007 to provide electricity in rural

areas using biomass gasifiers as generators to generate power from agricultural waste, such as rice

husks (hence the name of the company). These biomass plants could only serve a village if demand

was sufficiently great (e.g., 100 households) and were subject to fuel supply disruptions. HPS made

a strategic decision to add a solar micro-grid product to its portfolio for flexibility in reaching a

wider set of villages.

The HPS solar product is a solar micro-grid. The micro-grid consists of a 240 watt panel is
sona (2013) study household savings with a nearly 3-year follow-up
8 We are not aware of prior work that uses an experiment to estimate price sensitivity in a discrete choice demand

model. Kremer et al. (2011) is a close precedent that experimentally varies the quality of a good (a local water
source) and uses observable variation in walking distance to water sources as a proxy for price.

7
shared among 6 to 9 households. Each households gets its own 3.2 volt rechargeable battery and

its own meter. Meters have key pads to unlock access to the battery and therefore the system and

households must purchase codes each month to keep the system unlocked. Each household on the

micro-grid gets 25 to 40 watts of power for 5 to 7 hours each day. This quantity of power is very

small and is used to supply a package, supplied with the system, of a high-efficiency light bulb and

an electrical outlet, typically used for mobile phone charging.

Our experiment sample consisted of 100 villages distributed across three districts in Bihar, as

shown in Figures 1 and ??. These villages were chosen to fulfill three criteria. First, they were

not listed as electrified villages by the government. This implied that either grid connections were

entirely unavailable, or very few households were connected to the grid. Second, villages were

chosen to be reasonably close to existing HPS operating sites, so that micro-grid services could

be feasibly expanded to these areas. Third, sample villages had not already been offered HPS

micro-grids. The total population of households in all 100 villages was 48,979.

The experiment is a cluster-randomized control trial at the village level. The experiment had

three treatment arms that vary the availability and price of an HPS system: a control arm (34

villages), in which HPS did not offer micro-grids, a normal price arm (33 villages), in which HPS

offered micro-grids at the prevailing price of INR 200 per month, and a subsidized price arm (33

villages), in which HPS offered micro-grids at the reduced price of INR 100 per month. While the

prevailing price at the start of the experiment was INR 200, HPS later cut this price—within this

experimental arm only—down to INR 160, due to a lack of demand at the higher price. Therefore

the normal price arm consists of an initial price of INR 200 and a later price of INR 160.

The offers of these connections and prices were available to all households within a village,

regardless of whether they had previously expressed interest in HPS’s product or whether they

were surveyed as part of our experimental data collection. Sales of solar micro-grid connections

began in January 2014.

II.B. Data and Covariate Balance

Our primary source of data is a household panel survey with three waves. We also collected

administrative data on payments by customers to HPS, survey data on supply costs for diesel

generator operators, and administrative data on power supply and customer payments from the

electricity distribution company.

The household survey sample was drawn to represent households who stated an interest in a

solar connection. In August 2013, a customer identification survey was conducted in all sample

villages, to elicit the stated willingness of households to pay for a solar micro-grid connection. A

random sample of 30 households per village was drawn from those who reported they would be

interested in paying for a solar connection at a price of Rs. 100 per month. In practice, this

8
identification of potential customers was barely restrictive. Households were not required to put

down a deposit and were not held to their statement of interest when the product was eventually

offered. Over 90 percent of households with no electricity or diesel-based electricity said they would

be interested in using micro-grids. Over 70 percent of households with a grid electricity connection

or home solar panels, expressed an interest in the solar connections.9 We therefore expect that our

survey sample represents all or nearly all the potential solar users in the population.

The household survey panel consists of two thick rounds that we call baseline and endline

1 and one thin round that we call endline 2. The baseline household survey was conducted in

November and December 2013 and covered demographics, income, assets and appliance ownership,

electrification status and select measures of education and health. The survey covered all energy

sources used by households, payments associated with these sources, and other characteristics such

as capacity or hours of supply. The endline 1 survey was conducted in May through July, 2016 and

used nearly the same survey instrument. The endline 2 survey was conducted mainly as a check-up

on demand for solar a year after the first endline. This wave did not include characteristics of other

sources or detailed household covariates of the first two waves.

We collected administrative data from HPS on monthly payments from customers from January

2014 through April 2016, which we match with our household survey data to estimate demand for

HPS solar over time between our survey waves.

We surveyed diesel generator operators to get monthly information on cost structure, hours of

operation, connection plans, and number of customers, over a two-year period from January 2014

to 2016.

We also construct a time-series of electricity consumption and payments of households con-

nected to the electricity grid. We collect grid consumer IDs in our third household survey, as use

these to find sample households in administrative data. The time series includes monthly units

billed, which is based on units consumed for metered households, and average units consumed for

unmetered households. We observe the value of each bill and payment.

Table 2 shows the balance of covariates in our baseline survey across treatment arms for de-

mographic variables (Panel A), wealth or demand proxy variables (Panel B) and energy access

(Panel C). The first three columns show the mean values of each variable in the control, normal

price and subsidized price arms, with standard deviations in brackets. The next two columns show

the differnces between the normal price and control arms and subsidized price and control arms,

respectively, with the standard error of the difference. The final column shows the F -stat and

p-value from a test of the null that the treatment dummies are jointly zero at baseline.

The joint test rejects the null of equality of treatment arms at the ten percent level for three

out of twelve variables at baseline, somewhat higher than would be expected by chance. For
9 As Section ?? makes clear, the share of households who expressed an interest in using solar micro-grids was

vastly higher than actual take-up rates.

9
example, households in subsidy villages are more likely to have solid or pukka houses than control

households and to have solid roofs (these variables are likely to be highly correlated). The overall

rate of electrification does not differ by treatment arms (p-value 0.54, variable “Any elec source

(=1)”, but households in the subsidy treatment arm are more likely to have electricity from the

grid and somewhat less likely to have it from a diesel generator. We address this slight imbalance

by including baseline household covariates as controls in both our reduced-form and structural

estimates.

Considering the control group, the baseline covariates describe a population poorer and more

rural than Bihar as a whole, as would be expected given our sample selection criteria. Self-reported

household incomes averaged INR 7,460 per month (USD PPP 2.6 per person per day).10 Two-

thirds of households own agricultural land and about a quarter have a pukka house, constructed of

solid materials like brick. The average household has 3.3 adults living in 2.4 rooms. A quarter of

households have electricity from any source, with the most common source at baseline being diesel

electricity.

II.C. Electricity Source Characteristics

The solar micro-grid offered by HPS did not enter a vacuum but a marketplace in which diverse

sources of electricity were competing. Table 3 describes these alternative choices, which we discuss

in detail below. Each panel of the survey represents the source characteristics as of each of our

survey waves.

II.C.1. Grid electricity

In addition to solar micro-grids, formal electricity connections to the distribution grid were also

available in some villages in our sample. The availability of grid connections in a village depends

on how close the village is to the grid. If the grid reaches a village, households connect in two ways.

The formal way is to file an application with the state distribution company during government

‘connection camps’, or through a linesman, licensed revenue collector, or power contractor. The

informal way is to hire a local electrician to wire a connection, possibly paying bribes to utility

staff to look the other way.

In order to categorize households as having formal or informal connections, we collected house-

hold consumer IDs from grid customers during the follow-up survey conducted after the endline.

We matched these IDs with consumer IDs in three grid administrative data sets: a consumer base

dataset of all formal customers, a billing dataset containing monthly billing prices and quantities,

and a collection dataset of receipts for payments from customers to the distribution company.
10 Self-reported income can often be very imprecise in rural households with earnings from a combination of

livestock, agriculture, and intermittent or seasonal wage labor. For this reason, self-reported expenditure is fre-
quently preferred to income as a measure of household cash flow. In our sample, households report average monthly
expenditures of INR 10,733.

10
 We define households as formal if they were successfully found in the consumer base data.

For all such formal consumers, we can observe the bills they received as well as the payments

made. Informal consumers include households whose consumer IDs could not be found in the

administrative datasets, or were not provided to us at all. Some consumers also received electricity

bills but had made no payments over the last 12 months and we categorize these as informal

connections as well.

Table 3, column 1 describes some characteristics of grid electricity from our survey data. Grid

users pay the highest price of any electricity source at baseline, of INR 153 per month (Panel A).

For this price they are able to connect the highest average load amongst users of any source. The

grid is only available in 35% of villages, however, and when the grid is available the supply is not

continuous but for an average of 7.4 hours per day.

II.C.2. Diesel

At baseline diesel generators were a popular source of off-grid electricity. Diesel was supplied by

village residents who had purchased a generator and offered connection plans with monthly or

semi-monthly payments.

Table 3, Panel A, column 2 summarizes the characteristics of diesel connections at baseline.

Diesel generators are cheaper than the grid, at INR 100 per month on average, but support lower

average loads of 147 watts. While there was some variation in the plans offered, the modal option

consisted of a 100 watt connection for INR 100 per month through which customers could power

one or more light bulbs, or charge mobile phones in the evening.11 Generators run on a predictable

schedule in the evening and early night-time for an average daily supply of 3.4 hours. Because

diesel operators require a sufficient number of customers to support their fixed costs, diesel was

available in only 39% of villages at baseline.

II.C.3. Own solar

Households can buy and operate solar systems on their own, which we refer to as “Own Solar”

to distinguish it from the HPS micro-grid system. Households paid for these panels upfront and

would usually have to travel to a larger market town or city to make a purchase. Because these

markets were in principle accessible to all villages in our sample, we assume that households in

any village could have bought such a system if they wished.

Table 3, Panel A, column 3 summarizes the characteristics of own solar connections at baseline.

We translate the capital costs of own solar into an equivalent monthly payment using an assumed

lifetime of 7 years and interest rate of 20%. The characteristics of Own solar are strikingly similar

to diesel at baseline, with similar monthly prices and supported loads. Own solar users report they
11 The solar micro-grid provided similar lighting services at a much lower wattage because of the provision of
highly energy efficient LED lights with the connection.

11
can use their systems for 7 hours per day, significantly longer than users of diesel systems.

II.D. Electricity Source Market Shares

Over the course of the study from 2013 to 2017 there was a complete upheaval of the retail electricity

market in rural Bihar. The government dramatically expanded the grid and solar got cheaper and

better, so that these two sources of electricity overtook diesel generators.

In Figure 2, we plot the share of households using different electricity sources during the base-

line, endline 1 and endline 2 surveys. Each group of bars represents the households with electricity

from a given source, and each bar within a group the share of households with that source in

each of the three survey waves. The experiment is represented by the first group of bars: HPS

micro-grids were basically non-existent at baseline, increased to an 11% share of all households at

endline 1 (averaged across treatment arms), and then fell back.

The experiment occurred amidst a broader upheaval of the market. The share of households

with own solar systems rose from 5% to 6% and then 21% over three survey waves. Referring back

to Table 3, Panels A through C, in column 3, we can see that own solar systems got much cheaper

over the study period: the average system cost INR 100 per month at baseline and just INR 50

per month by the endline 2. The diesel share fell from 17%, the highest in the baseline, to 3%,

amidst increased competition, as many diesel operators exited the market. The number of grid

connected households skyrocketed from 5% at baseline to 25% at endline 1 and 41% at endline 2.

The government launched a concerted campaign to electrify villages and to connect households in

those villages through connection camps. Since our sample villages were chosen for being initially

un-electrified, our data collection bears witness to a massive increase in grid connections from this

campaign.

The market upheaval therefore appears to have involved several factors at once: falls in solar

price, a dramatic expansion of the grid by the government, and the endogenous exit of diesel

generators from the market. The following Section III reports on the demand for solar micro-grids

and their impacts in isolation. The Section V, afterwards, widens the scope of analysis to include

the entire demand system of competing electricity sources. With this demand system we can then

break down how much each change in the market has benefited households.

III Reduced-form Results

III.A. Demand for HPS Solar

Our experiment allows us to estimate the demand curve for solar micro-grid connections.

As we describe in Section II, households made monthly decisions on whether to pay for electric-

ity from Husk Power Systems. Households that did not pay would eventually be disconnected but

12
from month to month this risk was very low.12 Consequently the demand for micro-grids, defined

by the household decision to pay for this service, changes every month.

Recall that the micro-grid option was made available to households living in villages assigned

to the normal price and subsidized price treatment groups at INR 200 and INR 100 respectively.

Twelve months into the experiment, micro-grid connections were made available to households in

the regular price group at INR 160 per month. The random assignment of prices to villages allows

us to trace out the demand curve for solar micro-grid electricity.

In Figure 3, we plot the share of households choosing to pay for micro-grid electricity at different

prices and over different time intervals. Each separate line on the figure refers to payment at a

different horizon: at least once during the experiment, midway through during months 16-18, and

at the time of the endline 1 survey at month 29. Three features of consumer demand stand out.

First, the demand for micro-grid electricity at the normal price prevailing before the experiment

is practically zero: only about two percent of households paid even a single month in this arm.

Second, demand is highly elastic, and increases to a total of about 17% of households that paid

once during the experiment. Third, demand shifted in over the course of the experiment, such

that even at the subsidized price only 7% of households were paying by the time of the endline 1

survey.

Figure 7 shows the costs of supply for HPS per system per household. The total monthly cost,

in the left bar, is nearly USD 2 (INR 120) per month just for the system installation cost, excluding

all variables costs of billing, collection and system maintenance.13 Therefore the estimated demand

curve implies that demand is near zero at unsubsidized 2013 prices. The price sensitivity we find

for solar micro-grids agrees with other work that finds household demand is nearly zero when

off-grid solar is priced at cost (Aklin et al., 2017; Grimm et al., 2016).

III.B. Impacts of Access to Off-grid Solar

The demand curve reflects household willingness-to-pay for off-grid solar. This willingness-to-pay

reflects perceived household benefits to having a connection. There may also be benefits of solar

power that are not perceived by the household or not valued by the household decision-maker in

buying a connection. For example, improved lighting could lead to children having more time to

study at home, which may or may not be valued by parents. There may also be intra-household

spillovers from reduced kerosene consumption and indoor air pollution (Barron and Torero, 2017).

We estimate the impact of access to electricity on a battery of social outcomes. We begin by

12 This is an accurate description of how the contract between the micro-grid supplier and consumers played out
in practice. On paper, Husk Power Systems had an internal rule mandating that households who did not pay for
three consecutive months be disconnected, but this was sporadically enforced.
13 Capital costs as reported by our partner HPS are additionally net of capital subsidies provided by the govern-

ment, which were of the order of 60% in 2014.

13
estimating a reduced form model of the following type

yivt = α + βN N ormalP ricev + βS SubsidizedP ricev + δXi .

Here yivt is the outcome of interest for household i in village v at time t. N ormalP ricev and

SubsidizedP ricev are dummies that take the value 1 when village v was assigned to solar micro-

grids at regular and subsidized prices respectively. Xi is a vector of controls from the baseline

survey. The outcomes y may include measures of electricity access, adult and child respiratory

problems, reading and math test scores, and household income.

Households in the treatments got and used electricity micro-grids. The micro-grid powered two

low wattage LED lamps and one mobile charging point provided with every connection. Table 4

reports regressions of ownership of these appliances and use of electricity on treatment status. The

intent-to-treat effect of assignment to a subsidy treatment village is to increase light-bulb ownership

by 15 percentage points (standard error 4.7 pp) on a base of 32 percentage points in the control

(column 1). Subsidy treatment households increased hours of electricity use by an estimated 0.94

hours per day (standard error 0.24 hours per day) relative to 1.16 hours in the control (column 2).

The effects of being assigned to a normal price village are smaller but still statistically significant.

Households assigned to a subsidy treatment village are also more likely to own a mobile phone, by

3.4 pp on a high control ownership rate of 88 percent (column 3) (implying that control households

are 2.75 times as likely to own a mobile phone as a light bulb). Finally, assignment to a subsidy

treatment village also decreases the amount spent charging one’s mobile phone. This effect is

because households without electricity will typically charge their mobile phones at a shop for a

small fee, such as Rs. 5, which nonetheless implies a high unit cost of energy.

Therefore, the intervention had a meaningful first stage. We find that when solar micro-grids

are made available, households use more electricity, purchase more mobile phones, and spend less

money charging them. As we would expect, these effects are more pronounced when the monthly

costs of solar electricity connections are low.

Table 5 now turns to consider the effects of off-grid solar on social and economic outcomes, for

health, education and test scores. Panel A of the table is the reduced-form or intent-to-treat effect

for these outcomes, and Panel B is the instrumental variable estimate of the coefficient on hours

of electricity using the two treatment assignment dummies as instruments.

We find no evidence that respiratory problems decrease for adults or children (Panel B, columns

1 and 2). The predominant source of indoor air-pollution comes from cooking, which is unaffected

by the provision of micro-grids, and we do not find significant declines in kerosene expenditure (not

reported). Effects on reading test scores are positive but imprecisely estimated (columns 3 and

4). For example, we estimate that an hour of additional electricity use increase children’s reading

scores by 0.22 standard deviations (standard error 0.22 standard deviations). This is a fairly large

14
standardized effect but imprecise due to low first-stage take-up and the children tested being only

a subsample of the overall experiment. We cannot rule out a zero effect or a significant positive

effect of lighting on child test scores. Finally we find that electricity has a null effect on household

income of INR 150 per month (standard error 350), which is small compared to baseline income

of INR 7,500 per month.

Overall we find little evidence that the electricity provided by solar micro-grids had large welfare

impacts, with the caveat that our estimates for education outcomes are imprecise. Households still

value off-grid solar for lighting and other energy services. Of course, households may also see

greater benefits from other types or sources of electricity, in particular sources that support higher

loads. The next section introduces a model to understand how households value off-grid solar

relative to alternative sources of electricity.

IV Model of demand for electricity connections

We model consumer demand for electricity sources using a nested logit model. The model combines

several features to allow for a rich pattern of household demand. First, we estimate the model

using micro-data on household characteristics. Second, we allow for unobserved heterogeneity in

source quality at the technology-by-village-by-time level, and so can study the value of quality

improvements over time. Third, we estimate the household sensitivity to price in the model using

random variation in the price of solar micro-grids from the experiment.

IV.A. Specification

Household indirect utility from electricity connections is given by

X X
Uvtij = δvtj + ztir βrj + djg ζgi + (1 − σgj )ijt (1)
r g

where

X
δvtj = xvtjk β k + ξvtj (2)
k
X
Vij (β) = ztir βrj . (3)
r

Households have observable characteristics zitr indexed by r. These characteristics enter equation

1 with choice-specific household characteristic coefficients βrj . For example, a household with a

higher monthly income may have a greater preference for grid electricity. The average effects of

source covariates, such as price or hours of use, enter through equation 2 for δvtj , the mean indirect

utility of a product in a given village and survey wave.

Nests are indexed by g. The term ζgi is the utility of purchasing a product in group j ∈ Jg . For

15
this model to be a nested logit specification, we require that ijt is independently and identically

distributed with a Gumbel distribution and ζgi with the distribution described in Cardell (1991).

The term σg is a nest specific similarity parameter, which is equal to zero and not estimated for a

degenerate nest of only one choice. For small σg , the unobservable variance in product utilities is

mostly at the product level within nest, as σg approaches one the idiosyncratic variance in product

utilities is mostly at the nest level.

The probability that household i purchases product j is given by :

1−σ
e((δvtj +Vij (β))/(1−σg ) Dgj i g e((δvtj +Vij (β))/(1−σg )
P r(ij = j|δ, σ, β) = × P 1−σg = σ P 1−σ
Dgj i g Dgi Dgjgi g Dgi g

e((δvtj +Vij (β))/(1−σg ) . We sum across consumers to get the market share for a
P
for Dgi = j∈Jg

given technology and village

X e((δvtj +Vij (β))/(1−σg )

svtj (δ, σ, β) = σ P 1−σ .
i∈Ivt Dgjgi g Dgi g

The model gives straightforward analytic expressions for consumer surplus. For a fixed set of

options J in nests k = 1, . . . , K, we calculate the inclusive value of each nest g as

X
IVig = ln exp(Vbvtij /(1 − σg )).
j∈Jg

The value of the choice over nests is

!
X
E[CSi |J] = ln exp ((1 − σg )IVig ) .
g

To compare choice sets, the change in surplus under alternative choice sets J and J 0 is

X
− (E[CSi |J 0 ] − E[CSi |J]) /βprice .
i

We use this expression to place a value on changes like the removal of solar power from the choice

set or the expansion of the electric grid to villages where it is not now present.

IV.B. Estimation

We estimate the model in two stages. The first, nonlinear stage estimates the parameters of

equation 1 via maximum likelihood. The second, linear stage estimates uses the δ̂vtj from the first

stage as the dependent variable to estimate equation 2 using two-stage least squares.

This two-step procedure is common in the estimation of random coefficients logit models (Berry,

Levinsohn and Pakes, 1995, 2004). The key idea is to invert market shares to solve for unobserved

16
mean indirect utilities, allowing for linear IV estimates that are unbiased in the presence of the

endogeneity of price to quality (Berry, 1994) There are two ways in which our estimation differs

from common practice. First, since we have very rich household observable characteristics, we

use a simpler nested logit model rather than a mixed logit model, which is often essential with

aggregate data to model rich enough patterns of consumer substitution. Using nested logit also

means we can estimate the first stage by maximum likelihood for some gain in efficiency.14 Second,

in our application, we have much richer cross-market variation: we observe 100 villages in 3 survey

waves. That means that we can estimate the average effects of product characteristics far more

reliably than in applications that observe only regional or national markets.

IV.B.1. Nonlinear estimation of the first stage

The nonlinear stage recovers the parameters (β, δ, σ) from equation 1 via maximum likelihood. Let

cit ∈ J index the actual choice of household i in a given survey round t and let ij be the event

that household i chooses good j. The log likelihood function is

XX
log L(θ|c, z, x) = log P r(ij = cit |zti , xvt , β, σ, δ(β, σ)).
t i

We write δ(β, σ) because we concentrate the δ parameters out of the likelihood using a contraction

mapping (Berry, Levinsohn and Pakes, 1995). In our setting, there would be 4 inside sources ×

100 villages × 3 surveys = 1200 parameters in δ if all products were offered in all villages. The

contraction mapping greatly reduces the dimensionality of the nonlinear search by solving, for each

candidate parameter vector (β, σ), for the δ that exactly fit observed market shares.15

The maximum likelihood estimates are consistent without any adjustment for the possible

endogeneity of price to quality. The parameters β̂rj capture how household observables affect

indirect utility, relative to the mean effect. The average effects of price and other characteristics

are absorbed into the δvtj for each source. Therefore the possible endogeneity of price at the market

level does not affect the estimates in the first stage.

IV.B.2. Linear estimation of the second stage

b , ξˆ .
We now have estimates of the non-linear parameters (β̂, σ̂, δ̂vtj (β̂, σ̂)), and wish to recover β k vtj

We use δ̂vtj as the dependent variable in a linear regression specification of equation 2,

X
δ̂vtj = ξvtj + xvtjk β k + xvtj,price β price + ηvtj . (4)
k6=price

14 In principle, one could also estimate a mixed logit model using maximum likelihood, but this approach may be
biased by simulation error (Berry, Levinsohn and Pakes, 2004). In the nested logit model we specify, we solve for
market shares analytically, so maximum likelihood will be preferred to GMM on efficiency grounds.
15 In a nested logit model estimated with aggregate data, one could invert market shares analytically. In our model

with micro-data, the inclusive value of a given nest Igi = log Dgi varies from consumer to consumer. Therefore,
while choice probabilities for a given consumer can be inverted given that consumer’s observables, it is not possible
to analytically invert aggregate market shares to obtain δvtj .

17
Here ηvtj is the regression error term, assumed to have conditional mean of zero, and the β k

on average source characteristics are the coefficients of interest. We estimate this equation by

two-stage least squares. The set of instruments for price is made up of two dummies for the two

experimental treatment arms (normal price of Rs. 160 for micro-grid, subsidized price of Rs. 100

for micro-grid). Since the disutility of price is common across technologies the solar micro-grid

experiment allows us to estimate the price coefficient that applies to all sources.

We are also concerned that the hours of supply on the grid may be endogenous to demand and

hence the recovered δvtj in a given village. We therefore instrument, in some specifications, for

hours of supply in a village during both peak and off-peak hours using predicted supply during

those same periods, where the prediction is constructed based on the supply to nearby villages.

The logic of this leave-one-out style instrument is that supply to nearby villages may determine

supply in a given village, for example, because villages are served by the same power substation

following the same rationing rules. The exclusion restriction is that supply in nearby villages

does not depend on demand in a given village; this would be violated if there were, in addition

to endogeneity at the village level, common unobserved demand shocks across nearby villages,

conditional on our rich set of household observables. The Appendix lays out the construction of

the supply instrument in detail. Supply-side instruments, such as the characteristics of products in

other markets, under an assumption that there not common regional demand shocks, are common

in discrete choice applications (Nevo, 2000). The supply instruments in our setting may be more

credible because the structure of the distribution grid does physically connect supply decisions for

nearby villages.

Having estimated equation 4, we can solve for mean unobserved product qualities as

X
ξˆvtj = δ̂vtj − xvtjk β
b .
k
k

The quality terms ξvtj allow us to observe how unmodeled characteristics of electricity sources vary

across sources, villages and time. While electricity connections may appear homogenous, allowing

for unobserved quality is important to capture aspects such as the load a connection can serve, the

ease or difficulty of obtaining a connection, and the marketing and service associated with a given

source. These characteristics all vary widely across the sources that compete in the Bihar market.

V Model results: The value of solar innovation

This section reports estimates of household demand for electricity sources. We then use these

estimates to run counterfactual simulations to value changes in the technology, availability, and

pricing of various sources of electricity.

18
V.A. Demand estimates

The full demand model has some 1033 parameters: 1000 technology-by-village-by-survey specific

mean indirect utility parameters backed out from the first-stage demand model, 29 parameters

governing household heterogeneity, 3 parameters on the average effects of source characteristics

and a parameter governing correlation of the source-specific utility shocks. We therefore report

select parameters to give a sense of how the model interprets household technology choices. First,

we present estimates from the non-linear estimation of how household characteristics affect their

choices. Second, we describe the linear estimates of the average effects of source characteristics.

Third, we present distributions of source quality to characterize quality changes over time.

Table 6 shows the estimated effects of household characteristics on choice probabilities in the

demand model. The choice probabilities are derived from the estimated coefficients of the demand

model, reported in Appendix Table 12. The effect of household characteristics on choice probabil-

ities are non-linear; we evaluate these effects for a poor household, which we define as a household

of two adults with a one-room house, without a solid roof or walls, that does not own agricultural

land. (The full profile of a poor household’s characteristics is in Appendix Table 11.) The table

shows how the household’s probability of choosing each electricity source (across columns) varies

with a change in characteristics, either from zero to one (for discrete variables) or of one standard

deviation (for continuous variables).

The main finding of the table is that richer households, by any measure, have stronger prefer-

ences for grid electricity over all other sources. For example, our representative poor household has

a 21 percent chance of choosing grid electricity. If the household had a solid roof, the probability of

choosing grid electricity would increase by 11 percentage points (standard error 2.5 pp). Similarly,

increases in the number of household adults, household income, land holdings, literacy, house qual-

ity or the number of household rooms all have positive, economically meaningful and statistically

significant effects on the household probability of choosing grid electricity, and all also reduce the

probability a household chooses no electricity (the outside option). Some household characteristics

also increase household choice probabilities for other inside goods; for example, households with

higher incomes are more likely to choose solar micro-grids. The effects of household characteristics

on demand for the other inside goods, however, is much less pronounced than on demand for the

grid. Table 3 offers a natural interpretation of this finding: grid electricity offers higher load, and

many more households on the grid can run a fan or a television. Richer households want the energy

services these devices bring.

The above estimates show how household characteristics change household mean indirect util-

ities, relative to the mean utility for a source for an average household. We now consider how

source characteristics change the mean household indirect utility.

Table 7 reports estimates of the linear part of the demand model, obtained by regressing the

19
mean household indirect utility, recovered from the model at the source-by-village-by-time level, on

source characteristics. We instrument for source price using the experimental variation in micro-

grid prices. In some specifications, we also instrument for hours of supply using the imputed hours

of supply based on the supply to other, nearby villages. The main finding from the linear part is

that price has a large, negative effect on mean household indirect utility (column 2). The estimated

coefficient on associated with a Rs. 100 price increase is -2.05 (standard error 0.77). To give a scale

to this number, the mean probability of choosing the grid is 24 percent, and the model estimates

imply that a Rs. 10 increase in the grid price (17 percent of the mean price of Rs. 59) decreases

grid market share by 3 pp (12.5 percent of the mean share). The elasticity of grid market share

with respect to price is therefore -0.73. We interpret this as high; Rs. 10 is enough money to buy

two cups of tea or three bananas, but raising the grid price by this amount in a month cuts market

share by a noticeable 3 pp. Similarly the elasticity of own solar market share with respect to price

is -2.9.

The experiment is critical to identifying the household price response. In an analogous ordinary

least squares specification, we find a small, negative and insignificant effect of price on mean indirect

utility (column 1). This non-sensical result suggests that the prices of some electricity sources are

endogenously set higher in villages with higher demand for those sources. By contrast, the first-

stage of the experimental instrumental variables estimates is strong (column 4). As expected, being

in the normal (unsubsidized) price group raises price, whereas being in the subsidized price group

lowers price. The instrument has a strong first stage despite that the price variation only applies

to one electricity source.

We also estimate coefficients on the hours of supply that a source offers both on peak (in the five

evening hours from five to ten) and off-peak (all other hours). We find a positive but statistically

insignificant effect of peak hours on supply and a smaller, negative coefficient of off-peak hours

(column 2). The estimate of the value of peak hours is not precise, but agrees with the idea that

agricultural households mainly value light in the evening hours. In the column 3 specification,

we additionally instrument for hours of supply using imputed hours of supply based on supply to

nearby villages. We find that the coefficient on price is unchanged and the coefficient on peak hours

is positive but not statistically different from zero. The point estimate is somewhat larger than

the estimate without instrumenting for hours, from column 2, but we could not reject that the

value of peak hours is the same in the two specifications. We proceed with the column 2 estimates,

instrumenting for price but not hours, as our main specification for counterfactuals and will report

results from the column 3 specification as a robustness check.

The demand model allows flexibly for changes in ξvtj , the unobserved mean quality of electricity

across villages and time. Figure 4 summarizes these source qualities by plotting histograms of

quality by source (each row is one source) and time (each column is a different survey wave).

Within each source and wave, the histogram shows the distribution of source quality across villages.

20
 The figure shows how the landscape of electrification in Bihar shifted, with the grid and own

solar systems gaining market share, in a relatively short period of time. Stagnant technologies do

not improve quality. The distribution of diesel generator quality, for example, is about the same

in all three survey waves (there is some truncation at the bottom, due to exit). The micro-grid

provider, HPS, in the experiment, did not offer their product in many villages at baseline (by

design), and did not change their product between our first and second endline surveys. The

stasis of the micro-grid product is apparent in the figure, as the distribution of quality is similar

between the second and third survey waves (HPS row, right-hand two panels). Contrast diesel

and HPS micro-grids with own solar systems, which shifted up in quality in each survey wave.

These improvements could be due to improvements in technical factors such as battery capacity

and load, which we do not model directly, or to a broader reach of marketing and distribution of

these systems. Finally, we see large improvements also in the quality of the grid, especially betwen

the second and third survey waves. These improvements show the results of a government drive

to increase household connections, which may have increased access and therefore the estimated

quality of the grid.

The increases in electrification over the period of study were therefore due in part to solar

innovation and improvements in quality, in part to falling prices for solar systems and in part to a

rapid expansion in the availability and quality of grid electricity. We now use the model to break

down the contribution of these factors and to value their contribution to household surplus.

V.B. Counterfactual simulations

The estimated demand system gives us the household willingness-to-pay for electricity from a

variety of sources as a function of their characteristics. These characteristics, in turn, are at least

partly set by policy, such as whether off-grid solar systems should be subsidized, the rationing

policy for supply on the grid, and the tolerance of the state grid for theft and non-payment of

customers. Given the close competition between different sources of electricity shown in the model

estimates, small changes in policy may cause large shifts in household electrification and surplus.

We use the model to address questions on three themes. First, what is the value of innovation

in solar for the poor? Second, what is the value of improving the quality of the grid? Third,

how would changes in grid policy to deter theft affect electrification and surplus? We lay out the

assumptions behind each counterfactual and then give the results.

V.B.1. Counterfactual scenarios

Value of solar innovation. During the sample period the market share of solar power increased

by 20 percentage points, or about four fold. The price of solar has continued to come down over

time and the quality has continued to improve. It is therefore relevant to ask what is the value

21
of innovation in solar in our data and whether, even if households prefer the grid today, further

innovation is likely to generate more surplus from off-grid solar.

Even rapid falls in the price of solar panels bring down the cost of an installed solar system only

so much, since an installed system has costs other than capital. To discipline the possible value of

innovation we obtained administrative estimates from our partner HPS on the cost of their solar

system. Appendix Figure 7 shows the installation cost of the system in units of amortized USD

per month broken out by the cost for each component of the system. The total cost is about INR

114 (USD 1.73) per month, which is above our lowest subsidized price of INR 100 per month. At

the normal price of INR 160, and without allowing for any variable costs of labor and collections,

which are certainly not negligible, this cost basis would imply a profit margin of 40%. The capital

component of the cost, which we consider the object of innovation, is USD 0.62 per month for the

panel (36% of the total) and USD 0.33 for the battery (19%), thus 55% in total.

Counterfactually, we consider reductions in cost for solar photovoltaics and for batteries. For

solar PV, we assume a 55% reduction in cost in line with the National Renewable Energy Labora-

tory’s projections for 2022 (Feldman, Margolis and Denholm, 2016). For batteries, we assume that

they fall in cost by 75% in accord with the US Department of Energy’s 2022 goal (Howell et al.,

2016). Since panel capital and batteries only make up part of a system, these changes imply a

reduction in price by 34% to USD 1.12, or 30% cheaper than our subsidy treatment. We apply the

same proportional reduction in price to own solar, conservatively assuming that the 55% capital

component of total cost observed for HPS applies to own solar as well. This is conservative since

the HPS product involves monthly recharge costs that do not apply to the own solar model. We

can also consider a range of other possible solar prices, but consider this feasible reduction in cost

as a benchmark.

Improving the quality of the grid. The experimental estimates are from a setting where the

grid is deficient: it is offered only in some villages and even there the average supply is only 11

hours per day. The effects of solar on electrification rates would have been different if the grid was

everywhere, so households had better substitutes, or if the grid has non-existent, as is the case in

some parts of India and much of rural Sub-Saharan Africa.

We predict demand under these scenarios by removing grid electricity from the choice set or

by extending it to all villages. When the grid is not present in a village, we cannot estimate an

unobserved quality for the grid in that village; we therefore impute the unobserved quality as

the mean unobserved quality for the grid in villages with the grid in that survey wave. We also

simulate improvements in grid supply that increase the duration of power available by two hours

per day (or to five hours, the number of peak hours per day, if a two hour increase would exceed

five peak hours in total).

Reducing theft. If a deficient grid might reduce the appeal of the grid and inflate the market

share of solar, the policy towards electricity pricing will do the opposite. Electricity in Bihar is

22
heavily subsidized and the state tolerates a high level of theft, which lowers effective prices further

still. If power on the grid was priced at cost improvements in solar may have been even more

valuable to households.

Counterfactually we raise the price of grid electricity from Rs. 59 per month, the sample mean

in the endline 1 survey wave, to Rs. 140, which we calculate would be sufficient to cover the variable

costs of supply. This price change will shift surplus from consumers to the government and change

household take-up of electricity connections from the grid and other sources. We also consider a

budget neutral reform that would increase supply during the peak hours and pay for that increase

by increasing prices. We calculate that a price increase to Rs. 95 per month would be sufficient to

pay for the increase in supply.

V.B.2. Counterfactual results

Figure 5 shows how the market shares of all electricity sources respond to changes in the price of

solar. The prices, on the horizontal axis, range from INR 70 per month in the reduced capital cost

scenario, up through the range of our experimental treatments, to a near-choke-price of Rs. 300

per month. Own solar capital costs are varied proportionally with HPS costs since these sources

have similar cost structures. The vertical axis shows market shares for each source technology.

Further reductions in the price of solar would increase its market share moderately. At the

lowest subsidised price from the experiment, 8 percent of households adopt HPS in the model and

11 adopt own solar, thus solar has a total market share of 19 percent. Setting solar prices at 2022

projections, solar adoption increases to 32 percent. Most of this gain in market share comes from

households that would not otherwise be electrified (solid line at top) rather than substitution from

other sources. For example, considering an HPS price cut from INR 200 to INR 70, HPS market

share increase by 27 pp, and the share of households without electricity declines by 26 pp. Thus

most the new customers are coming not from other sources of power but from non-electrification.

If the price of solar were raised to the choke price of Rs. 300 per month, the share of households

without any source of electricity would rise from the 57 observed to 62 percent. Solar power

therefore makes a significant contribution to electrification in the status quo.

The market share and value of solar depends on the extent to which the grid is available as

a substitute. Figure 6, panel A shows how the market shares of solar and the grid change as the

extent of the grid varies. Solar achieves a 33% market share when the grid is absent from all

villages, which is cut in half if the grid is present in all villages. There is a wide range of grid

extent spanned in our surveys, from 43% of villages having grid in the baseline to 57% in endline

one and 61% in endline two. In the model, holding constant household and source characteristics,

the increase in grid presence between baseline and endline one alone accounts for an 11 pp (40%)

drop in solar market share.

Figure 6, panel B shows that the presence of the grid similarly diminishes the value of solar

23
power. We calculate the consumer surplus from the bundle of offered electricity sources in our

model, recalculate this surplus without solar power in the choice set, and then normalize the differ-

ence in surplus by the estimated household price coefficient to get household-specific willingness-to-

pay. The figure reports the value of household WTP averaged over all households for four different

extents of the grid. Without the grid, household willingness-to-pay for the presence of solar in the

choice set is Rs. 974 per year; if the grid is everywhere this drops to around Rs. 520. The presence

of grid electricity cuts the surplus due to solar power by about half. The value of solar would be

further diminished by improvements in the quality of the grid, as we observed in the last wave of

our survey.

Table 8 summarizes the results of a broader set of counterfactuals on solar innovation, grid

extension and electricity distribution policy. Every row is a different counterfactual corresponding

to a scenario described in the prior subsection. Columns 1 through 5 give the market shares of each

technology in the row scenario and columns 6 through 8 measure consumer, producer and total

surplus. We take consumer and source characteristics as of the first endline survey. Consumer

surplus in each scenario is measured relative to the surplus consumers derive from having no

electricity, assumed to be zero. Producer surplus is the absolute surplus we estimate for the grid

only, based on its cost structure. The producer surplus can be taken as comprehensive if we assume

that the other sources are competitively supplied; this is probably accurate for own solar but not

for diesel (which, in any case, generally has a small share). All surplus measures are per sample

household per year.

Table 8, panel A shows that the model fits sample market shares. That fit is by design since

the source-by-village-by-wave specific quality measures are estimated to fit market shares exactly.

In both the data and the model, 57 percent of households have no electricity connection. The grid

is the most popular source of electricity, with a 24% share, but the two types of solar together are

not far behind, with a 17% share. Consumers have a mean surplus of Rs. 1918 (USD 29) per year

in this scenario and the state grid a surplus of negative Rs. 497 (USD 36.5). The grid is heavily

subsidized and many consumers steal; the average payment for the grid of Rs. 57 per month is

about 40% of the variable cost of supply. The grid therefore loses a significant amount of money

on every customer connected.

We organize our main findings from the counterfactual simulations to consider solar innovation,

grid extension and changes in grid pricing. Our first main finding, from Panels A and B, is that

solar makes a meaningful contribution to electrification and reduces the financial losses of the state

grid. From the benchmark scenario, if we remove solar altogether (Panel B, row 1), the share of

households without electricity increase by 12 percentage points, from 57% to 69%. That increase is

less than solar market share since households substitute to diesel and the grid. As the grid market

share increases by five percentage points the losses of the grid increase in magnitude to negative

Rs. 599 per year (20% larger).

24
 One can rightly say that a twelve percentage point swing in electrification is not transformative.

What would we expect from further innovation in solar? Taking a longer view with solar nowhere

as the baseline, we can state two values for solar innovation. The first value is from no solar (Panel

B, row 1) to the first endline observed in our data. We estimate that solar increased household

surplus by Rs. 775 (USD 12) per household in our sample. The second considers further solar

innovation if prices were cut along our optimistic path for 2022 innovation. The total solar market

share increases to 33 percentage points, surpassing the grid and raising consumer a further Rs. 176

per year, for a total gain, relative to the absence of distributed solar, of Rs. 951 per year (USD 14).

To put this gain in perspective, the average monthly income in our sample is Rs. 7,576. Therefore,

even the entry of solar gaining market share equal to one third of all households only raises surplus

by 1% of mean household income or 3% of the median.

Our second main finding is that improvements to the availability and quality of the grid would

increase electrification, with an effect similar in size to that of further innovation from off-grid

solar. By the time of our first endline the grid had reached only 57% of sample villages. If the

grid were removed from all villages, the share of households without electricity, from any source,

would rise from 57 to 64%, and household consumer surplus from all sources would fall by 46%. If

the grid were extended to all villages, it would increase electrification by 11 percentage points and

household surplus from all electricity sources by 18%. Even after this grid extension, we estimate

that 46% of households would remain without any source of electricity. This finding shows the

challenge of achieving universal electrification in a poor population. The model estimates also

allow us to compare intensive and extensive margin policy changes. If the grid were to offer two

more hours of power during the peak period, this would bring 7 pp more households onto the grid,

and have a similar effect on overall electrification rates as extending the grid to all villages.

Our third main finding is that raising the price of the grid to cover the cost of supply would

devastate grid market share. In Panel D, row 1, we simulate raising the grid price from Rs. 57

to Rs. 140 per month, which would cut demand for the grid from 24 pp to only 4 pp, with many

household substituting towards solar power instead, and cut consumer surplus by 36%. The fall

in consumer surplus is from pricing power at cost is only slightly lower than the fall in surplus due

to removing the grid altogether (46%). Producer surplus, on the contrary, would surge from a loss

of roughly Rs. 500 per household per month to nearly break-even. Total surplus from this price

change is therefore only somewhat (17%) lower than in the status quo, though there is a dramatic

shift in surplus from consumers to the state and many consumers fall off the grid. This large

swing in market share may seem extreme, but it is consistent with our experimental estimates, in

particular the high price sensitivity observed for solar around the price (at Rs. 100 per month) of

competing options like diesel power (Figure 3).

Could the government do better than such a pure transfer, by trying to improve prices and

quality at the same time? Our fourth main finding is that the scope for this kind of grand bargain

25
appears limited. We estimate that increasing the supply on the grid up to two hours in the peak,

and raising prices enough to just cover the cost of this additional supply, would slightly reduce

household surplus and total surplus. However, this bargain would increase the electrification rate

by 6 pp. The logic, in the model, is that the increase in peak hours brings some consumers onto

the grid, even as the increase in price drives more price sensitive consumers off the grid towards

solar. Therefore there is almost zero net flow of consumers to the grid but the total electrification

rate rises.

The interpretation that consumers are not willing to pay much for better quality depends on

the model specification we use to a greater extent than do the other counterfactual findings. We

estimate the value of price to consumers using our experiment, but do not have as strong an

instrument for hours of supply. Because our estimate of the value of peak supply is imprecise,

the conclusion depends on the exact specification of the demand system. For example, if we

instrument for both price and hours in the second, linear stage of the demand model, then the

“grand bargain” increase in quality and price would increase consumer surplus by 12% and total

surplus by 6%, due to the higher household valuation, in that specificaiton, for additional peak

hours of supply. Appendix Table 16 shows all of the counterfactual results under this alternative

model specifications.

V.B.3. Interpretation

The broad finding of the counterfactual simulations is therefore that the value of solar power

depends with great sensitivity on grid policy, namely availability and pricing. In particular, the

value of solar is much greater the weaker is the grid on availability and price. On price, we may

consider that competition from the grid, which is heavily subsidized through both explicitly low

prices and a tolerance for theft, is unfair in this setting. A fairly extensive grid with cheap power

and connection fees contrast with some African countries where the grid is largely absent and

prices are higher (World Bank, 2017).

A second policy angle that emerges from the model results is that the government may have

a rationale to subsidize solar, wholly aside from environmental concerns. Tolerance for theft, as

much as being able to serve higher loads, is a large part of the grid’s appeal in this setting. Each

customer here does not consume much but loses a lot of money for the state. Every customer

that better solar power takes from the money-losing grid increases the state’s producer surplus by

reducing losses. These changes are large, with the entry of solar saving the government a sum of

money about equal to what the households that take-up solar themselves pay.

26
VI Conclusion

Many large developing countries in Asia and Africa have made universal electrification a high-

priority policy goal. Our estimates show that the economics of serving the rural poor will make

achieving this goal costly. Off-grid solar systems are adequate for the needs of the rural poor, but

not for relatively richer households, who prefer the grid. Moreover, households are extremely price

sensitive and so extending the grid and achieving high rates of household access may require large

government subsidies.

Our analysis of the demand for electricity offers an explanation for why the supply of electricity

to the rural poor is so abysmal. Poor households are not willing to pay much for power or for

increases in supply. If the state values access, then it must heavily subsidize power, either explicitly

or through a tolerance for theft, to keep these households on the grid. The state adopts rationing

of supply in order to limit the losses it makes in serving these households. The combination of

massive subsidies and rationing make the competition between the grid and off-grid power lopsided

and limit solar market share in particular.

A broader question that our static analysis cannot answer is what the cost of this dysfunctional

electricity supply sector is in the longer run. There is some evidence that electrification has large

external returns (Lipscomb, Mobarak and Barham, 2013). It is hard to imagine a large business,

for example in manufacturing, opening in an area with eleven hours of electricity supply (Allcott,

Collard-Wexler and O’Connell, 2016). The combination of these facts implies that, even if rationing

electricity is a statically necessary policy to support electricity access, it may limit rural growth.

Off-grid solar systems cannot replace the scale economies of a well-functioning grid.

27
VII Figures

Figure 1: Map of sample villages in India

Kilometers
0 15 30 60

Treatment Groups
Subsidy Price
" Normal Price
" D
D"" "
" "D" D Control
D
D"
D"DD D
"
D "" D D"
D D
"
"" "" ""
D
DD D
D "
" "
"" D
DD
"
"D
" DD D "
D
D "

"
" DD
D
DD
"

28
 Figure 2: Take-up of electricity sources across household surveys

0.43
Baseline
Endline 1

.4
Endline 2

.3
Share of Households

0.25

0.22
.2

0.17
.1

0.06
0.05 0.05 0.05
0.04 0.04
0.03
0.01
0

Diesel Grid HPS Own Solar

The figure shows the take-up of electricity sources across household surveys.
Each group of bars represents the households with electricity from a given
source, and each bar within a group the share of households with that source
in each of the three survey waves. The experiment is represented by the
first group of bars: HPS micro-grids were almost non-existent at baseline,
increased to a 5% share of all households at endline 1 (averaged across
treatment arms), and then fell slightly in endline 2. The share of households
with own solar systems rose from 5% to 6% and then 22% over three survey
waves. The diesel share fell from 17%, the highest in the baseline, to 3%,
amidst increased competition, as many diesel operators exited the market.
Owing to the government’s village electrification campaign, the number
of grid connected households skyrocketed from 5% at baseline to 25% at
endline 1 and 43% at endline 2.

Figure 3: Demand curve for HPS solar

160 200
Price

Capital costs only = Rs. 115

100

Paid at endline Paid at midline Paid once

0

0 .1 .2 .3
Quantity: Share of Customers
The figure plots the share of households choosing to pay for micro-grid electricity
at different prices and over different time intervals. Each separate line on the fig-
ure refers to payment at a different horizon: at least once during the experiment,
midway through during months 16-18, and at the time of the endline 1 survey at
month 29.

29
 Figure 4: Distribution of unobserved mean indirect utility (ξ) by source and wave

Grid, Baseline Grid, Endline 1 Grid, Endline 2

5
5

3
3

1
1
ξ
-1

-1

-1
-3

-3

-3
-5

-5

-5
0 .1 .2 .3 .4 0 .1 .2 .3 .4 0 .1 .2 .3 .4
Diesel, Baseline Diesel, Endline 1 Diesel, Endline 2
5

5
5

3
3

1
1
ξ
-1

-1

-1
-3

-3

-3
-5

-5

-5

0 .1 .2 .3 .4 0 .1 .2 .3 .4 0 .1 .2 .3 .4
HPS, Baseline HPS, Endline 1 HPS, Endline 2
5

5
5

3
3

1
1
ξ
-1

-1

-1
-3

-3

-3
-5

-5

-5

0 .1 .2 .3 .4 0 .1 .2 .3 .4 0 .1 .2 .3 .4
OwnSolar, Baseline OwnSolar, Endline 1 OwnSolar, Endline 2
5

5
5

3
3

1
1
ξ

-1

-1
-1

-3

-3
-3
-5

-5

-5

0 .1 .2 .3 .4 0 .1 .2 .3 .4 0 .1 .2 .3 .4
Fraction Fraction Fraction
Our demand model allows flexibility for changes in ξvtj , the unobserved mean quality of electricity
across villages and time. The figure summarizes these source qualities by plotting histograms of quality
by source (each row is one source) and time (each column is a different survey wave. Within each source
and wave, the histogram shows the distribution of source quality across villages.

30
 Figure 5: Market shares under varying solar prices

.7
No Elec
.6

Experiment market price

Solar cost falls
.5
Market Share
.3 .4

Grid

Own Solar
.2
.1

Diesel HPS Solar

0

0 100 200 300

Solar Price (INR)
Based on the first and second stage estimated parameters, the figure shows market shares
under varying prices for solar technology. Household and source characteristics and the
availability of all sources are assumed to be at their endline 1 levels.

Figure 6: Changing Availability of Grid

1000 0.4
Solar
0.35 Grid

800
Annual INR per Household

0.3
Market Share

0.25
600

0.2

400
0.15

0.1
200
0.05

0 0
Baseline Endline 1 Full Grid No Grid Baseline Endline 1 Full Grid No Grid

(A.) WTP for solar (B.) Market shares of Grid and Solar
Based on the first and second stage estimated parameters, the figure shows willingness to pay for solar
under varying grid availability scenarios. WTP is calculated as described in Section ?? Household and
source characteristics and the availability of all sources (except grid) are assumed to be at their endline
1 levels.

31
VIII Tables

Table 1: Electrification context

Location US India Sub-Saharan Africa Bihar

(1) (2) (3) (4)
GDP per capita (USD) 57,467 1,709 1,449 420
kWh per capita 12,985 765 481 122
Electricity access (% of population) 100 79 37 25
kWh per capita / US kWh per capita 1 0.059 0.037 0.009
Source: World Bank (2015, 2017)

32
 Table 2: Baseline covariate balance

Control Normal Subsidy Diff(N-C) Diff(S-C) FTest

(1) (2) (3) (4) (5) (6)
Panel A. Demographics
Literacy of household head (1-8) 2.44 2.69 2.60 0.25 0.16 1.33
[2.04] [2.15] [2.10] (0.16) (0.15) (0.27)
1031 971 989 2002 2020
Number of adults 3.31 3.50 3.49 0.20∗ 0.18∗ 2.19
[1.58] [1.75] [1.78] (0.11) (0.11) (0.12)
1052 983 1001 2035 2053
Panel B. Wealth Proxies
Income (Rs. ’000s/month) 7.46 7.32 7.28 -0.14 -0.18 0.068
[6.91] [6.93] [7.09] (0.57) (0.51) (0.93)
1041 963 983 2004 2024
Number of rooms 2.40 2.55 2.53 0.15 0.13 1.29
[1.32] [1.45] [1.45] (0.10) (0.098) (0.28)
1052 981 999 2033 2051
House type (pukka = 1) 0.24 0.27 0.31 0.035 0.074∗∗ 2.79∗
[0.43] [0.45] [0.46] (0.037) (0.031) (0.066)
1052 983 1001 2035 2053
Owns agricultural land 0.67 0.69 0.67 0.015 0.0022 0.045
[0.47] [0.46] [0.47] (0.056) (0.053) (0.96)
1052 983 1001 2035 2053
Solid Roof (=1) 0.42 0.46 0.51 0.042 0.095∗∗ 3.08∗
[0.49] [0.50] [0.50] (0.043) (0.039) (0.050)
1052 983 1001 2035 2053
Panel C. Energy Access
Any elec source (=1) 0.25 0.31 0.27 0.061 0.022 0.63
[0.43] [0.46] [0.44] (0.055) (0.050) (0.54)
1052 983 1001 2035 2053
Uses gov. elec (=1) 0.030 0.036 0.091 0.0052 0.060∗∗ 2.53∗
[0.17] [0.19] [0.29] (0.017) (0.028) (0.085)
1052 983 1001 2035 2053
Uses diesel elec (=1) 0.17 0.21 0.11 0.039 -0.063 1.70
[0.38] [0.41] [0.31] (0.058) (0.046) (0.19)
1052 983 1001 2035 2053
Uses own panel (=1) 0.034 0.050 0.061 0.016 0.027∗ 1.81
[0.18] [0.22] [0.24] (0.014) (0.015) (0.17)
1052 983 1001 2035 2053
Uses HPS solar (=1) 0.0067 0.0081 0.0050 0.0015 -0.0017 0.14
[0.081] [0.090] [0.071] (0.0078) (0.0054) (0.87)
1052 983 1001 2035 2053
The table reports the balance of covariates in our baseline survey across treatment arms for demographic variables
(Panel A), wealth or demand proxy variables (Panel B) and energy access (Panel C). The first three columns show
the mean values of each variable in the control, normal price and subsidized price arms, with standard deviations
in brackets. The next two columns show the differences between the normal price and control arms and subsidized
price and control arms, respectively, with the standard error of the difference. The final column shows the F-stat
and p-value from a test of the null that the treatment dummies are jointly zero at baseline. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0 : 01

33
 Table 3: Summary of electricity sources, all panels

Grid Diesel Other Solar HPS

(1) (2) (3) (4)
Panel A. Baseline
Monthly price (Rs.) 153.0 99.76 100.5 -
Load (watts) 278.1 146.8 178.8 -
Supply hours 7.41 3.38 7.47 -
Source available 0.35 0.39 1 0
Panel B. Endline 1
Monthly price (Rs.) 118.6 104.8 89.45 163.8
Load (watts) 154.7 118.3 51.57 37.07
Supply hours 6.32 3.08 5.86 6
Source available 0.59 0.13 1 0.65
Ownership of assets
Mobile and/or bulb 1 1 1 0.93
Fan 0.34 0.04 0.10 0.03
TV 0.11 0.01 0.04 0.02
Panel C. Endline 2
Monthly price (Rs.) - - 50.58 200
Load (watts) - - 66.11 -
Supply hours - - - 6
Source available 0.77 - 1 -
The table shows characteristics of the various sources of electricity that constitute
the rural electricity market we study. Load reported here is based on household
survey appliance ownership, and household survey reports of own solar watt rat-
ings. In the model, we apply diesel generator survey data to assign load available
to households served by each generator, as well as technical specifications from
HPS for panel capacity.

Table 4: First stage electricity, bulb, and mobile phone outcomes

Light Bulb Daily Hours of Mobile Phone Price of

Ownership (=1) Electricity Use Ownership Full Charge (Rs.)
(1) (2) (3) (4)
Subsidy treat village (=1) 0.15∗∗∗ 0.94∗∗∗ 0.034∗∗ -0.67∗∗∗
(0.047) (0.24) (0.014) (0.24)
No subsidy treat village (=1) 0.098∗∗ 0.52∗∗ 0.022 -0.46∗
(0.044) (0.20) (0.013) (0.23)
Baseline Controls Yes Yes Yes Yes
Control mean 0.32 1.16 0.88 4.72
Observations 3001 2868 3001 964
The table shows regressions of ownership of LED lamps and one mobile charging point and the use of electricity
on treatment status. Households in the treatments got and used electricity micro-grids and these appliances. the
specifications include baseline electricity source indicators, baseline monthly income, and baseline equivalent of
outcome variable as controls. Standard errors clustered at the village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0.01

34
 Table 5: Household welfare impacts

Respiratory problems (=1) Standardized test score Monthly income

Adults Children Reading Math (INR ’000s)
(1) (2) (3) (4) (5)
Panel A. Reduced Form
Subsidy treat village (=1) 0.026 0.012 0.11∗ 0.095 0.18
(0.021) (0.0082) (0.061) (0.065) (0.31)
Normal treat village (=1) 0.017 0.0041 0.020 0.071 0.63∗
(0.018) (0.0082) (0.061) (0.062) (0.33)
Panel B. Instrumental Variables
Hours of electricity 0.027 0.014 0.22 0.21 0.15
(0.027) (0.012) (0.24) (0.23) (0.35)
Baseline Controls Yes Yes Yes Yes Yes
Control mean 0.14 0.024 0 0 7.5
Observations 2710 2669 646 637 2692
The table shows the effects of off-grid solar on social and economic outcomes, for health, education and test scores.
Panel A of the table is the reduced-form or intent-to-treat effect for these outcomes, and Panel B is the instru-
mental variable estimate of the coefficient on hours of electricity using the two treatment assignment dummies as
instruments. Test score results are at the child level. The regressions include baseline electricity source indicators,
baseline monthly income, and baseline equivalent of outcome variable controls. Standard errors clustered at the
village level in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Table 6: Impact of Household Characteristics on Choice Probabilities

Grid Diesel Own Solar HPS None

(1) (2) (3) (4) (5)
∆ Adults 0.0363 0.0023 0.0014 0.0049 -0.0449
(0.0090) (0.0055) (0.0012) (0.0041) (0.0081)
∆ Income 0.0161 0.0025 0.0010 0.0143 -0.0338
(0.0075) (0.0047) (0.0010) (0.0045) (0.0076)
∆ Land 0.0488 -0.0233 0.0028 0.0085 -0.0367
(0.0183) (0.0096) (0.0028) (0.0090) (0.0159)
∆ Literacy 0.0256 0.0084 -0.0006 0.0023 -0.0356
(0.0079) (0.0050) (0.0014) (0.0037) (0.0074)
∆ Pukka 0.0771 -0.0037 -0.0021 -0.0060 -0.0652
(0.0234) (0.0127) (0.0027) (0.0078) (0.0188)
∆ Roof 0.1074 -0.0072 0.0047 -0.0071 -0.0979
(0.0249) (0.0134) (0.0033) (0.0072) (0.0185)
∆ Rooms 0.0259 0.0106 0.0033 0.0031 -0.0429
(0.0083) (0.0055) (0.0014) (0.0043) (0.0083)
The table presents the marginal impact of change in household covari-
ates in the probability of choosing a particular source of electricity for
”poor” households, who are in the first quartile of asset ownership,
income, literacy, and family size. The profile of a poor household is
described in Table ??. The standard errors are constructed using the
delta method.

35
 Table 7: Two-Stage Least Squares Estimates for Demand for Electricity

Price & Price

OLS Price IV Hours IV First Stage
(1) (2) (3) (4)
Price (Rs. 100) -0.17 -2.05∗∗∗ -2.04∗∗∗
(0.11) (0.77) (0.76)
Hours of supply on peak 0.21 0.26 0.42 0.032
(0.21) (0.24) (0.33) (0.065)
Hours of supply off peak -0.095∗∗ -0.11∗∗ -0.16∗∗ -0.0091
(0.047) (0.053) (0.073) (0.016)
Normal Price 0.057∗
(0.034)
Subsidy Price -0.13∗∗∗
(0.029)
ξtj mean effects Yes Yes Yes Yes
Observations 1000 1000 1000 1000
F 168.4 .
The table presents 2SLS estimates of our demand system. The dependent variable is mean indirect
utility at the market × survey wave level retrived from the non-linear first stage. Peak hours refers
to supply of electrricity during the evening (5pm-10pm). The second column presents two-stage least
squares estimates where instrument price with the experimentally varied HPS treatment assignment.
In the third column we instrument for price and peak, off-peak hours. for grid hours, we use predictions
from a random forest model (see appendix for details) as instruments. For off-grid sources, we use the
data out instrument matrix. The last column provides first stage regrressions for the specification in
column 2. First stage regressions for the specification in column 3 are presented in the appendix Table
∗ ∗∗ ∗∗∗
14. All regressions control for wave × source mean effects. p < 0.10, p < 0.05, p < 0.01.
Standard errors cluster at the village level in parentheses.

36
 Table 8: Grid and Solar Policy Counterfactuals

Market Shares Social surplus (INR)

Grid Diesel Own solar HPS None Consumer Producer Total
Panel A: Fit of model
Data 24 3 7 10 57 - - -
Model 24 3 7 10 57 1918 -497 1421
Panel B: Value of solar innovation
Solar nowhere 29 3 0 0 69 1143 -599 544
Solar everywhere 23 2 13 10 52 1822 -489 1333
Solar cost falls 21 1 16 16 46 2171 -435 1659
Further solar innovation 20 1 17 16 46 2094 -422 1749
Panel C: Grid Extension
37

Grid nowhere 0 3 20 13 64 1032 0 1032

Grid everywhere 37 1 7 9 46 2269 -766 1503
Extra 2 Hours 31 2 11 9 47 2167 -784 1383
Panel D: Theft Reductions
Remove theft by raising grid price 4 2 18 13 62 1219 -50 1169
Increase peak supply and raise prices 23 2 13 10 52 1829 -497 1332
Note: We currently calculate annual losses to the distribution company per household using the following formula:
12×Number of grid customers×(60 kWh ×3.88 INR −xvt,grid,price ) kWh
month kWh . 60 month is the
mean consumption among Bihar households in billing
N
INR
data, 3.88 kWh power purchase cost (NBPDCL Tariff Order FY 2014-15). The price of grid is the mean price in the endline 1 sample
(INR 57.7). Number of grid customers is computed by multiplying the grid market share of each counterfactual case by the total
number of HHs in the endline 1 survey. Panel A demonstrates that market shares predicted by our model match the market shares
observed in the data. Panel B considers three counterfactual solar policy scenarios. In particular for “further solar innovation”,
we assume a 55% reduction in cost of solar panels in line with the National Renewable Energy Laboratory’s projections for 2022.
For batteries, we assume a cost reduction of 75% in accordance with the US Department of Energy’s 2022 goal. Panel C considers
counterfactual grid policy scenarios and Panel D considers theft removal from the grid. Grid price is defined as the reported bill
value in our surveys multiplied by payment rate, where payment rate is the mean of the responses to ‘Do you pay your bill?’ in
the endline 2 survey. We therefore define the ‘remove theft by raising grid price’ counterfactual by using the reported bill value
as full price, unscaled by payment rate, which yields INR 140. In the last row, we have set peak hours equal to its maximum of 5
and raise the grid price until annual losses are approximately equal to actual annual losses, i.e., to INR 95.
Appendix

VIII.A. Additional results

Figure 7: HPS price under current and counterfactual capital costs

2
1.5

Base Structure
Labor
Monthly Price ($)

Transport
Other
MeterCharger
1

Wiring
Battery
Panel
.5
0

Actual Reduced Panel/Battery Price

In our counterfactuals we consider reductions in cost for solar photovoltaics and for
batteries. We assume a 55% reduction in cost of solar PV in line with the National
Renewable Energy Laboratory’s projections for 2022. For batteries, we assume a cost
reduction of 75% in accordance with the US Department of Energy’s 2022 goal.

Table 9: Definition of Household Characteristics and Magnitude of Marginal Change

Characteristic Definition Marginal Change

Income Monthly income 1 SD (INR 6486)

Land Indicator for agricultural land 0 to 1

Roof Indicator for solid roof 0 to 1

Pukka Indicator for pukka house 0 to 1

Rooms Number of rooms in the house 1 SD (1.32 rooms)

Adults Adults in the HH 1 SD (1.82 persons)

Literacy Literacy of household head (1-8) 1 SD (2.04 years)

The table shows the magnitude of the change in household covarates for which
the marginal impact on household choice probabilities is estimated in Table ??.
Literacy classification: 1 =not literate, 2= Aanganwadi, 3 = literate but be-
low primary, 4 = literate till primary, 5 = literate till middle, 6 = literate till
secondary, 7= literate till higher secondary, 8 = graduate and above

38
 Table 10: Summary Statistics of Household Characteristics

Mean Median Q1 Q3 SD Min Max

Number of rooms in the house 2.45 2 2 3 1.32 1 11
Indicator for pukka house 0.32 0 0 1 0.47 0 1
Indicator for agricultural land 0.63 1 0 1 0.48 0 1
Indicator for solid roof 0.51 1 0 1 0.50 0 1
Literacy of household head (1-8) 2.48 1 1 4 2.04 1 8
Adults in the HH 3.67 3 2 5 1.83 1 15
Monthly income (INR) 7575.5 6000 4000 8500 6486.2 0 65000.0
Observations 8822

Table 11: Household profile for marginal effects

Profile Rooms Pukka Land Roof Literacy Adults Income (INR)

Poor 1 0 0 0 1 2 3750

39
 Table 12: First step results choice-specific household characteristics (βrj )
and nest-similarity parameter (σg )

(Grid, diesel, HPS) (Grid/Diesel) (Grid) &

βrj Multinomial Logit
& (Own Solar) & (Solar) (Off-grid)
Grid × Income 0.211 0.187 0.204 0.195
(0.064) (0.054) (0.063)
Diesel × Income 0.140 0.162 0.143 0.212
(0.085) (0.063) (0.083)
Own solar × Income 0.167 0.178 0.189 0.209
(0.073) (0.057) (0.08)
HPS × Income 0.492 0.482 0.432 0.220
(0.115) (0.114) (0.145)
Grid × Land 0.243 0.202 0.243 0.237
(0.087) (0.077) (0.085)
Diesel × Land -0.149 -0.085 -0.148 0.039
(0.121) (0.107) (0.116)
Own solar × Land 0.145 0.180 0.153 0.071
(0.114) (0.086) (0.107)
HPS × Land 0.223 0.219 0.162 0.064
(0.172) (0.172) (0.144)
Grid × Adults 0.123 0.116 0.123 0.123
(0.023) (0.02) (0.023)
Diesel × Adults 0.086 0.086 0.088 0.090
(0.035) (0.025) (0.034)
Own solar × Adults 0.092 0.098 0.091 0.095
(0.027) (0.022) (0.025)
HPS × Adults 0.090 0.091 0.093 0.096
(0.043) (0.043) (0.034)
Grid × Pukka 0.423 0.365 0.429 0.423
(0.096) (0.089) (0.097)
Diesel × Pukka 0.133 0.199 0.172 0.096
(0.143) (0.108) (0.147)
Own solar × Pukka 0.120 0.163 0.113 0.099
(0.124) (0.1) (0.121)
HPS × Pukka -0.032 -0.030 0.066 0.106
(0.193) (0.194) (0.177)
Grid × Lit 0.096 0.082 0.095 0.097
(0.019) (0.017) (0.019)
Diesel × Lit 0.094 0.083 0.094 0.051
(0.027) (0.02) (0.026)
Own solar × Lit 0.017 0.048 0.018 0.047
(0.024) (0.022) (0.023)
HPS × Lit 0.048 0.050 0.047 0.051
(0.038) (0.038) (0.032)
Grid × Roof 0.580 0.515 0.568 0.580
(0.097) (0.087) (0.097)
Diesel × Roof 0.201 0.280 0.201 0.256
(0.134) (0.11) (0.132)
Own solar × Roof 0.415 0.444 0.358 0.282
(0.118) (0.09) (0.128)
HPS × Roof 0.000 0.0003 0.065 0.255
(0.184) (0.183) (0.216)
Grid × Rooms 0.133 0.144 0.133 0.135
(0.034) (0.029) (0.033)
Diesel × Rooms 0.149 0.155 0.149 0.158
(0.046) (0.034) (0.044)
Own solar × Rooms 0.183 0.167 0.178 0.160
(0.039) (0.032) (0.039)
HPS × Rooms 0.100 0.098 0.119 0.157
(0.067) (0.067) (0.061)
σ1 0.554 0.168 0.950
(0.201) (0.341)
σ2 0.415
(0.54)
Number of observations 8822 8822 8822 8822
Log Likelihood -5791.3539 -5789.3465 -5791.3155 -5789.2562
Likelihood Ratio Test Statistic - 4.015 0.077 4.195
LR test p-value - 0.045 0.96 0.0405
a
Note: The likelihood Ratio test statistic: LR = −2 {LL(θconstrained ) − LL(θunconstrained )}.
Each of the nested-logit specifications (columns 2, 3, and 4) are tested against the constrained
multinomial logit specification in column 1 of Table ??. LR is distributed χ2 with degrees of
freedom equal to the number of constraints on θ. LL is the negative of the optimized objective
function in MATLAB (which is defined as the negative of the sum of the individual household
contributions to log of the likelihood function).

40
 Table 13: Two-Stage Least Squares Estimates for Demand for Electricity

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

(Grid, Diesel, HPS) (Grid, Diesel) (Grid) Multinomial
(Own Solar) (Solar) (Off-Grid) Logit
Price (Rs. 100) -2.051∗∗∗ -2.082∗∗∗ -2.116∗∗ -1.918∗∗
(0.769) (0.775) (0.876) (0.746)
Hours of supply on peak 0.263 0.401 0.439∗ 0.437∗
(0.243) (0.259) (0.261) (0.255)
Hours of supply off peak -0.110∗∗ -0.139∗∗ -0.145∗∗ -0.144∗∗∗
(0.0534) (0.0567) (0.0571) (0.0555)
ξtj mean effects Yes Yes Yes Yes
Observations 1000 1000 1000 1000
The table presents 2SLS estimates of our demand system for different first-stge nest specifications. The
dependent variable is mean indirect utility at the market × survey wave level retrived from the non-linear
first stage. The first column, uses our preferred first stage nest-specification of grouping grid, diesel, and
HPS in one nest and own-solar in its own nest. The estimates in the first column are the same as column
2 of Table 7. The second column uses a first-stage model with grid and diesel in one nest and both solar
technologies in another. In the third column, we group grid in its own nest and all off-grid technologies in
a second nest. In the last column, we use the mean indirect utilities derived from a multinomial logit first
stage. We instrument price with our experimentally varied HPS treatment assignment. Peak hours refers to
supply of electrricity during the evening (5pm-10pm). All regressions control for wave × source mean effects.
∗ ∗∗ ∗∗∗
p < 0.10, p < 0.05, p < 0.01. Standard errors cluster at the village level in parentheses.

Table 14: First stage regressions of the 2SLS estimates

Price Peak hours Off-peak hours

First Stage First Stage First Stage
(1) (2) (3)
Normal Price 0.058∗ 0.0050 -0.0046
(0.034) (0.0050) (0.030)
Subsidy Price -0.13∗∗∗ 0.0075 0.014
(0.029) (0.0063) (0.031)
Peak Hours Instrument 0.068 0.94∗∗∗ 0.19
(0.059) (0.063) (0.15)
Peak Hours Instrument -0.014 0.032∗∗ 0.88∗∗∗
(0.012) (0.013) (0.030)
ξtj mean effects Yes Yes Yes
Observations 1000 1000 1000
F . . .
The table presents the first stage estimates of the 2SLS estimates provided in column 3 of Table 7. The
∗ ∗∗
construction of the instrument for hours of supply is outlined in Section ??. p < 0.10, p < 0.05,
∗∗∗
p < 0.01. Standard errors cluster at the village level in parentheses.

Table 15: 10 INR price increase from endline 1 price

Market share percentage point change

Grid -3.00
Diesel -0.37
Own solar -2.29
HPS -1.60

41
 Table 16: Grid and Solar Policy Counterfactuals for 2SLS second stage with IV for price and hours

Market Shares Social surplus (INR)

Grid Diesel Own solar HPS None Consumer Producer Total
Panel A: Fit of model
Data 24 3 7 10 57 NaN NaN NaN
Model 24 3 7 10 57 1909 -497 1412
Panel B: Value of solar innovation
Solar nowhere 29 2 0 0 69 1129 -601 528
Solar everywhere 23 1 13 10 52 1799 -491 1308
Solar cost falls 21 0 16 16 47 2143 -439 1628
Further solar innovation 20 0 18 16 46 2066 -426 1716
Panel C: Grid Extension
42

Grid nowhere 0 2 20 13 65 995 0 995

Grid everywhere 37 0 7 9 46 2255 -774 1480
Extra 2 Hours 36 1 10 8 45 2400 -900 1500
Panel D: Theft Reductions
Remove theft by raising grid price 5 2 19 12 63 1190 -54 1135
Increase peak supply and raise prices 30 1 12 9 48 2145 -651 1494
Note: We currently calculate annual losses to the distribution company per household using the following formula:
12×Number of grid customers×(60 kWh ×3.88 INR −xvt,grid,price ) kWh
month kWh . 60 month is the
mean consumption among Bihar households in billing
N
INR
data, 3.88 kWh power purchase cost (NBPDCL Tariff Order FY 2014-15). The price of grid is the mean price in the endline 1 sample
(INR 57.7). Number of grid customers is computed by multiplying the grid market share of each counterfactual case by the total
number of HHs in the endline 1 survey. Panel A demonstrates that market shares predicted by our model match the market shares
observed in the data. Panel B considers three counterfactual solar policy scenarios. In particular for “further solar innovation”,
we assume a 55% reduction in cost of solar panels in line with the National Renewable Energy Laboratory’s projections for 2022.
For batteries, we assume a cost reduction of 75% in accordance with the US Department of Energy’s 2022 goal. Panel C considers
counterfactual grid policy scenarios and Panel D considers theft removal from the grid. Grid price is defined as the reported bill
value in our surveys multiplied by payment rate, where payment rate is the mean of the responses to ‘Do you pay your bill?’ in
the endline 2 survey. We therefore define the ‘remove theft by raising grid price’ counterfactual by using the reported bill value
as full price, unscaled by payment rate, which yields INR 140. In the last row, we have set peak hours equal to its maximum of 5
and raise the grid price until annual losses are approximately equal to actual annual losses, i.e., to INR 95.
 Table 17: Counterfactual Analysis: Assumptions

Scenario Source availability Source hours (peak) Source price

Data
Model Endline 1
Endline 1 for grid and diesel,
Solar nowhere
solar nowhere
Endline 1 for grid and diesel,
Solar everywhere
solar everywhere
Endline 1 for grid and diesel,
Solar cost falls
solar everywhere
Endline 1 for grid and diesel,
Further solar innovation
solar everywhere
43
Panel A: Theft Reductions
Endline 1 Endline 1
Panel B: Value of solar innovation
Endline 1 Endline 1
Endline 1 Endline 1
Reduction in HPS price from INR 170 to INR 120, according to the ”solar
Endline 1
cost falls” scenario in Figure 3. Own solar price is proportionally decreased.
Own solar: 55% reduction in solar panel cost (NREL Fig 8, low estimates, pp.
17) and a 75% reduction in batteries (DOE, pp. 2). Panel is 36% of total cost
and batteries is 19% of total cost. These numbers imply a reduction in total
Endline 1 own solar price by 34%. Mean own solar price is reduced by 34%. All other
solar costs (meter charger, wiring, labour, transport, other are assumed
constant). Mean HPS price is similarly reduced by 34%. Prices of all other

sources are according to endline 1.

Panel C: Grid Extension
Endline 1 for solar and
Grid nowhere Endline 1 Endline 1
diesel, grid nowhere
Grid everywhere, endline 1
Grid Everywhere Endline 1 Endline 1
for diesel, solar everywhere
Two additional peak
Endline 1 for grid and diesel, hours for grid, endline
Grid 2 extra peak hours Endline 1
solar everywhere 1 peak hours for all
other sources.
Panel D: Theft Reductions
Grid at INR 140, all else at endline 1. The grid price was derived as follows:
grid price in the model estimation is presently defined as the reported bill
Remove theft by raising Endline 1 for grid and diesel, value in the surveys multiplied by payment rate, where payment rate is the
Endline 1
grid price solar everywhere mean of the responses to “Do you pay your bill?” in the endline 2 survey. We
therefore define the “remove theft” counterfactual by using the reported bill
value as full price un-scaled by payment rate. This yields INR 140.”
Grid at INR 95 everywhere, all else at endline 1. INR 95 is derived as follows:
Grid peak hour = 5
Increase peak supply and Endline 1 for grid and diesel, grid peak supply is set to 5 hours everywhere. Price is set so that total loss
hours everywhere, all
raise prices solar everywhere per HH in the endline sample is equal to that obtained in row 3 (model with
else at endline 1
solar everywhere).
Note: Household characteristics and source off-peak hours are unchanged (at their endline 1 levels) across all counterfactual cases and hence omitted from the table.
VIII.B. Random Forest for endline 1 survey
The imputation of missing feeder supply data is done in two steps.

1. If supply data for a village is missing for a given survey wave but some data is available for that village in the +/ − 6
month window, we replace the missing observation with the temporal mean of available data in this time window.
2. If a village has missing data for all months in the +/ − 6 month time window, we use a random forest (RF) algorithm
to impute missing hours of supply for that village.

RF has the advantage that it necessarily yields internal predictions and so imputed hours of supply are sensibly bounded.
We include the following predictor variables (features) in the RF model: (1) hours of supply of the three nearest villages for
which we do have data, (2) division fixed effects, (3) polynomials upto degree 5 of district-demeaned latitude and longitude
of each village, and (4) interactions of division fixed effects with each of the demeaned lat-lon polynomials. Hence, all the
features that go into the RF model are plausibly exogenous to our demand model. We exclude unelectrified villages (supply
for these is replaced zero) from the sample because we don’t want to use data of unelectrified villages to impute missing
supply data for electrified villages. For instance, there are 56 electrified villages in the endline survey which have non-mising
data. This is our master sample for the RF model. We randomly select 80% of this sample (45 obs) as the training sample
and the remaining 20% as the testing sample (11 obs). The RF model is fit on the training sample.
Figures 8 and 9 describe our prediction model. The main parameter to tune in a random Forest model is the number
of candidate variables to select from at each split. To do this, we start with 2 variables and increase by a step factor of 1.5
until the improvement in out-of-box (OOB) error is less than one percent. As shown in panel A of Figure 8, for the endline
1 data, this yields 6 variables. Figure 9 shows the most relevant variables chosen by the model.

Figure 8: Random Forest Algorithm: endline 1 data

4.5

●
6.0
4.4

5.5
4.3
OOB Error

Error

5.0

●
4.2

4.5
4.1

4.0

2 3 4 5 6 0 200 400 600 800 1000

Number of features at each node trees

(A.) OOB error minimized at m = 6 (B.) OOB stabilizes as number of trees rises
Panel A of the figure shows how we tune on the number of variables to split the data during the creation of
each tree. Panel B shows that the out of box error of the RF model stabilizes as the number of trees rise.

44
 Figure 9: Variable Importance Plot: Top 15 Features

div_3_lat_demeaned4 ● lat_demeaned4 ●

elec_ss_feeder_wave21 ● elec_ss_feeder_wave22 ●

div_3_lat_demeaned3 ● lat_demeaned2 ●

div_2_lon_demeaned5 ● div_3_lat_demeaned3 ●

div_2_lon_demeaned3 ● lat_demeaned3 ●

div_2_lat_demeaned3 ● div_3_lat_demeaned ●

lon_demeaned ● lat_demeaned ●

lat_demeaned2 ● div_3_lon_demeaned ●

div_2_lon_demeaned ● lat_demeaned5 ●

elec_ss_feeder_wave22 ● div_3_lat_demeaned4 ●

elec_ss_feeder_wave23 ● div_3_lon_demeaned5 ●

div_3_lat_demeaned ● elec_ss_feeder_wave21 ●

div_2_lat_demeaned ● div_3_lat_demeaned5 ●

div_2_lat_demeaned5 ● div_3_lat_demeaned2 ●

div_3_lat_demeaned2 ● div_3_lon_demeaned3 ●

5.5 6.0 6.5 7.0 0 2 4 6 8

%IncMSE IncNodePurity

The figure measures relative importance of the variables in our model based on standard metrics of variable impurity.

The RMSE of our prediction model is 1.9 hours. We take the predictions from this model and use them to impute
missing observations in administrative grid supply data.

VIII.C. Construction of instrument for hours of supply

In one our second-stage two-stage least squares specifications (column 3, table 7), we instrument for peak and off-peak
hours of supply of electricity, in addition to price. we instrument HPS, diesel, and own solar supply by a constant. For HPS
and diesel this is 0 for off-peak hours and 5 (maximum) for peak hours. For own solar, it is the global median of the peak
and off-peak hours for that source. For grid, we use our predictions from the above random forest model as instrument.

References
Aklin, Michaël, Patrick Bayer, SP Harish, and Johannes Urpelainen. 2017. “Does basic energy access generate
socioeconomic benefits? A field experiment with off-grid solar power in India.” Science Advances, 3(5): e1602153.

Allcott, Hunt, Allan Collard-Wexler, and Stephen D O’Connell. 2016. “How do electricity shortages affect indus-
try? Evidence from India.” The American Economic Review, 106(3): 587–624.

45
Barron, Manuel, and Maximo Torero. 2017. “Household electrification and indoor air pollution.” Journal of Environ-
mental Economics and Management.

Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. “Automobile prices in market equilibrium.” Econometrica:
Journal of the Econometric Society, 841–890.

Berry, Steven, James Levinsohn, and Ariel Pakes. 2004. “Differentiated products demand systems from a combina-
tion of micro and macro data: The new car market.” Journal of Political Economy, 112(1): 68–105.

Berry, Steven T. 1994. “Estimating discrete-choice models of product differentiation.” The RAND Journal of Economics,
242–262.

Bryan, Gharad, Shyamal Chowdhury, and Ahmed Mushfiq Mobarak. 2014. “Underinvestment in a profitable
technology: The case of seasonal migration in Bangladesh.” Econometrica, 82(5): 1671–1748.

De Mel, Suresh, David McKenzie, and Christopher Woodruff. 2013. “The demand for, and consequences of,
formalization among informal firms in Sri Lanka.” American Economic Journal: Applied Economics, 5(2): 122–150.

Dinkelman, Taryn. 2011. “The effects of rural electrification on employment: New evidence from South Africa.” The
American Economic Review, 101(7): 3078–3108.

Duflo, Esther, Rema Hanna, and Stephen P Rya. 2012. “Incentives work: Getting teachers to come to school.” The
American Economic Review, 102(4): 1241–1278.

Dupas, Pascaline, and Jonathan Robinsona. 2013. “Why don’t the poor save more? Evidence from health savings
experiments.” The American Economic Review, 103(4): 1138–1171.

Feldman, David, Robert Margolis, and Paul Denholm. 2016. “Exploring the potential competitiveness of utility-scale
photovoltaics plus batteries with concentrating solar power, 2015–2030.” National Renewable Energy Laboratory.

Grimm, Michael, Luciane Lenz, Jorg Peters, and Maximiliane Sievert. 2016. “Demand for Off-grid Solar Elec-
tricity: Experimental Evidence from Rwanda.” IZA Discussion Paper Series, , (10427).

Howell, David, Brian Cunningham, Tien Duong, and Peter Faguy. 2016. “Overview of the DOE VTO advanced
battery R&D pogram.” Department of Energy.

International Energy Agency. 2017. “From Poverty to Prosperity, a World Energy Outlook-2017 Special Report.”

Kremer, Michael, Jessica Leino, Edward Miguel, and Alix Peterson Zwane. 2011. “Spring cleaning: Rural water
impacts, valuation, and property rights institutions.” The Quarterly Journal of Economics, 126(1): 145–205.

Lee, Kenneth, Edward Miguel, and Catherine Wolfram. 2016. “Experimental Evidence on the Demand for and
Costs of Rural electrification.” National Bureau of Economic Research Working Paper 22292.

Lipscomb, Molly, Mushfiq A Mobarak, and Tania Barham. 2013. “Development effects of electrification: Evidence
from the topographic placement of hydropower plants in Brazil.” American Economic Journal: Applied Economics,
5(2): 200–231.

Nevo, Aviv. 2000. “A practitioner’s guide to estimation of random-coefficients logit models of demand.” Journal of
economics & management strategy, 9(4): 513–548.

Rud, Juan Pablo. 2012. “Electricity provision and industrial development: Evidence from India.” Journal of Development
Economics, 97(2): 352–367.

World Bank. 2017. “State of electricity access report 2017 (Vol. 2.” World Bank.

46