Preliminary Draft: Please do not cite without author’s permission.

Household Responses to Food Subsidies:

Evidence from India∗

Tara Kaul†

University of Maryland, College Park

January 8, 2014

Abstract

This paper uses household survey data to examine the effect of food subsidies on thenutritional outcomes of poor households in India. The national food security program,known as the Public Distribution System (PDS), takes the form of a monthly quotaof cereals (rice and/or wheat) available for purchase at substantially discounted prices.The effect of the program is studied by exploiting the geographic and household sizespecific variations in the value of the subsidy that result from differences in state pro-gram rules and local market prices. The analysis focusses on the rice subsidy programand outcomes are consumption data from six years (2002-2008) of the NSSO Socio-Economic surveys. In agreement with other literature on food subsidies, this paperfinds small elasticities for cereal consumption and caloric intake with respect to thevalue of the subsidy. However, households benefit from the program in terms of overallfood intake and not just through cereals directly provided by the PDS. The elastici-ties for calories from all food groups are positive and significant. This is in contrastto studies on pure price subsidies for cereals that have found zero or negative effectson caloric intake. Thus, the results in this paper suggest that quotas may be moreeffective than price subsidies at improving nutrition. Taking into account state leveldifferences in the functioning of the program, a substantially smaller effect is found instates that have higher levels of corruption. Finally, the estimates from this paper areused to simulate the caloric impact of changes in the PDS under the provisions of theNational Food Security Bill of 2013.

JEL classification: I38; H42; O12; O15; Q18

∗I am grateful to Jessica Goldberg and Melissa Kearney for their invaluable guidance and support atevery stage of this project. I thank Sebastian Galiani, Parikshit Ghosh and Susan Godlonton for helpfulcomments and discussions. I also thank Ron Chan, Daisy Dai, Sonal Desai, Brinda Karat, Reetika Khera,Giordano Palloni, Brian Quistorff, Miguel Sarzosa, Suraj Shekhar and seminar participants at the Universityof Maryland, Delhi School of Economics, the 2013 MEA Annual Meeting, the 2013 Midwest InternationalEconomic Development Conference and the 2013 APPAM Fall Research Conference for useful feedback. Igratefully acknowledge use of data from the National Sample Survey Office, New Delhi and the India HumanDevelopment Survey. All errors and omissions are my own.

†Email: kaul@econ.umd.edu

1

1. Introduction

Provision of food security is one of the most basic forms of assistance to poor households

and many governments around the world provide aid in the form of food subsidies. Despite

their popularity and importance, there is a relatively small body of research studying the

impact of such subsidies on nutritional outcomes. While the primary rationale for all food

subsidies is reducing food insecurity (FAO 1997), their impact on nutrition (via food intake)

has generally been found to be quite small, in some cases even zero or negative.1

Food subsidies are implemented via food stamps, price subsidies, direct in-kind transfers

or through the operation of ration shops (price subsidies with a quantity cap). The mecha-

nism through which a food subsidy operates depends on the manner in which the program

affects income and prices (and the related elasticities of the demand for food), the specific

types of food groups it targets (and the income and cross price elasticities of other foods),

and how close the beneficiary households are to their ideal level of food consumption. This

paper examines India’s food security program, the Public Distribution System (PDS). The

PDS is one of the country’s biggest anti-poverty programs. It has a target population of

65.2 million households and the total cost of food subsidies amounts to almost 1% of GDP

(Planning Commission 2012).

1Butler, Ohls and Posner (1985) find a very limited nutritional impact of the food stamps program in theUnited States. The Indian food security program has also historically been found to have a small impacton nutrition (Kochar 2005, Tarozzi 2005, Khera 2011c). Jensen and Miller (2011) find a negative impact ofprice subsidies on caloric intake in China. There is also a literature comparing the effects of food subsidiesand income transfers. Hoynes and Schanzenbach (2009) find that the food stamps program in the UnitedStates has the same impact on food expenditure as cash. Steifel and Alderman (2006) and Laderchi (2001)find similar results in Peru where the impact of food subsidies on child nutrition is no different than that ofcash transfers.

2

The PDS provides a monthly quota of cereals (20-35 kg per household) at substantially

discounted prices to households that are below the poverty line (Government of India 2011).

The poverty line in India is determined by having an income deemed adequate to purchase a

basic minimum level of calories.2 Since cereals are relatively cheap and rich in energy, they

account for a substantial proportion of the total calories consumed by poor households.3

Thus the PDS aims to directly increase caloric intake by providing supplementary cereals to

poor households, who are, by assumption, food insecure. The program may, however, have

consequences in terms of consumption of other food groups which are critical to maintaining

a healthy and balanced diet. This is particularly important in India where the levels of child

malnutrition are alarmingly high and have been linked to severe micronutrient deficiencies

(Banerjee and Duflo 2011). Thus it is necessary to study the impact of the program not

just on the intake of cereals and total calories, but also on calories from different food groups.4

The value of the subsidy to a household is a function of the quantity cap and the discount

(difference between the market and PDS price of cereals) that the household is eligible for.

This paper uses two previously unexploited sources of variation in the value of the subsidy to

measure its impact on nutritional outcomes. First, the total quantity cap for each household

2Deaton and Dreze (2009) find that the average caloric intake in India has declined over time. However,they note that this declining trend does not undermine the importance of reducing calorie deficiencies forpoor households. In my sample, the caloric intake of the below poverty line (BPL) population is well belowthe minimum caloric norms (and the calories consumed by more affluent households). See Table 4 ahead.

3Cereals account for over 72% of the total calories consumed by BPL households in the sample understudy (Table 4) .

4While the links between improvements in caloric intake and health outcomes certainly exist, the rela-tionship is not necessarily linear and depends critically on other factors such as a balanced diet, activitylevels, sanitation, and basic medical facilities (Deaton and Dreze 2009). In the absence of data on healthoutcomes in my dataset, I am unable to predict the impact of the subsidy on health. However, assessing theimpact on caloric intake is important in itself. Reducing the incidence of hunger by improving caloric intakefor a food insecure population, such as BPL households in India, is among the major humanitarian goals ofmany governments and international organisations around the world.

3

varies by state since the program is jointly funded by the state and central governments.

Further, some states index the cap by family size and others offer a fixed quantity to every

household. This results in variation, even within state, in the per person quantity cap, due

to differences in family size. Second, the price charged for PDS goods is typically set for

the year and not linked in any way to fluctuations in market prices. Due to the absence of

perfectly integrated agricultural markets and controls on the movement of goods within the

country, the difference between the PDS and market price within a year varies substantially

across districts (within a state) and seasons.

State level program rules and differences in local market prices are used to compute the

value of the subsidy for each beneficiary household, based on its size and district-season-year

cell. This value is used to identify the effect of the subsidy, assuming that after controlling

for household characteristics, district, state-year and seasonal effects, the remaining variation

in the value of the subsidy (owing to unpredictable fluctuations in local market prices and

differences in the per person quantity cap due to family size) is exogenous. The consumption

data come from six years (2002-2008) of the nationally representative socio-economic surveys

conducted by the National Sample Survey Organisation (NSSO). The analysis focusses on

eight major states in India where rice is the primary staple. The main outcomes studied

are cereal consumption, caloric intake, calories from different food groups, and total food

expenditure. In order to compare the estimates with prior work on expenditure and income

elasticities, supplementary data on income and consumption from a smaller but much richer

dataset, the India Human Development Survey (2004-05), is used.

The estimated elasticities of cereal consumption and calories with respect to the value of

4

the subsidy are small, but higher than previous estimates – the elasticity of caloric intake

with respect to the value of the subsidy is 0.144, as compared to 0.06 in Kochar (2005).

An increase in the rupee value of the subsidy increases calories by more than twice of what

is implied by its impact on cereal consumption alone. This is an important indicator that

households benefit from the program in terms of overall caloric intake and not just through

the cereals directly provided by the PDS. Even though the program subsidizes only cereals,

it has a positive effect on the consumption of all types of food: the impact on calories for

all food groups is positive and significant. In contrast to Jensen and Miller (2011), who find

a zero or negative effect of pure price subsidies on overall calories and calories from differ-

ent food groups, the results in this paper suggest that quotas may be more effective than

price subsidies in improving nutrition via caloric intake. Results are robust to alternative

specifications of the value of the subsidy and the impact of the program on non beneficiary

households is examined and found to be small.

State level differences in the administration of the PDS are taken into account using ev-

idence on program fidelity from Khera (2011a). As expected, the subsidy has a substantially

smaller impact in those states where the illegal diversion of food grains away from their in-

tended beneficiaries is reported to be high. Despite serious implementation issues, the PDS

continues to be one of the government’s biggest anti-poverty programs and an expansion of

the system is imminent under the National Food Security Bill, which was passed in Septem-

ber 2013. The bill will expand both the reach, and the value of subsidies provided through

the PDS. The elasticity estimates from this paper, combined with current data on prices,

suggest that the increase in the value of the subsidy under the provisions of the bill will

lead to an increase of 66 - 72 kcal in the daily caloric intake of current beneficiaries of the

5

program. These gains will be much higher if states are able to successfully reduce leakages

from the PDS.

The paper is organized as follows: Section 2 describes the functioning, history and rules

of the Public Distribution System in India. The motivation is described in section 3, which

comprises the conceptual framework and a review of the related literature. The data, empir-

ical specification and descriptive statistics are presented in section 4 and section 5 discusses

the results. The final section concludes and describes avenues for future work.

2. The Public Distribution System in India

The Public Distribution System was established in India in 1939 by the then British govern-

ment to cope with rising food prices and food shortages in Bombay and other urban areas.

Over the years it expanded its scope dramatically and is currently one of the Indian gov-

ernments most significant welfare programs. In the 2012-13 central government budget, the

food subsidy was projected at Rs 750 billion (approximately US$13.8 billion, Government

of India 2012). The PDS provides a minimum support price to farmers and acts as a food

safety net for the rural and urban poor. It works alongside the free market and provides

rice, wheat, edible oils, sugar and kerosene at subsidized prices through 489,000 fair price

shops across the country (Planning Commission 2012).

Prior to 1997, any household could purchase a monthly quota of subsidized products at

fair price shops. In 1997, there was a major change in the PDS with the introduction of the

Targeted Public Distribution System (TPDS). Under the new system, families classified as

6

being below the poverty line (BPL) could get 10 kg of food grains per month at half the

economic cost to the central government of procuring them. This quota was increased in

2000 to 20 kg per month (rice and/or wheat) and in 2002, the quota was revised upwards to

35 kg for most states. The subsidy for above poverty line (APL) families was eliminated.5

Table 1 presents the details of state specific quotas for the period 2002-2008. Kochar (2005)

estimates that the value of the subsidy, defined as the product of the quantity entitlement

and the difference between the market and PDS price, increased from Rs 7 per household

per month in 1993 to Rs 48 per household per month in 2000 (an increase of over 500%).

Thus the TPDS introduced targeting and substantially increased the subsidy for BPL fam-

ilies. Currently, PDS prices are fixed by the central government and state governments are

only allowed to add limited transportation costs and taxes. PDS prices are not indexed in

any way to market prices. 24.3 million households classified as Antyodaya (poorest of the

poor) are entitled to a larger quota and still lower prices. The next big change in the PDS

is imminent as per the provisions of the National Food Security Bill, which was passed in

September 2013. This bill makes it a legal right for 67% of the population to obtain 5 kg of

foodgrains (per person per month) at prices between Re 1 and Rs 3 per kilogram. The eldest

woman in the household will be given access to the subsidy on behalf of her family (The

Hindu, July 2013). The resulting expansion in the reach and value of subsidies provided

through the PDS could make it the largest food security program in the world.

The PDS is an essential part of the social framework in many areas of the country, particu-

larly places with limited access to markets. However, the implementation of the program has

5Tamil Nadu continues to have a universal subsidy and, since June 2011, provides 20 kg of rice per monthfree of cost to BPL households.

7

been called into question on a number of dimensions.6 The first is the targeting accuracy of

the PDS post 1997. Local governments are responsible for periodically carrying out surveys

to assign poverty scores to households. The central government provides a cap on the number

of BPL households for each state. State governments translate these caps into cut offs for

the poverty score. All households that fall below their respective state- and district-specific

cut off are classified as being below the poverty line and are issued BPL cards (Karat 2013

and Dreze and Khera 2010). In 2002, the BPL survey comprised 13 questions related to

educational status, asset holdings, indebtedness, demographics etc. Targeting leads to er-

rors of inclusion and exclusion. The reported rate of exclusion (eligible households without

a BPL card) varies from 3% in Andhra Pradesh to 47% in Assam (Planning Commission

2005). In an independent study on Rajasthan, Khera (2008) finds an exclusion error of 44%.

Swaminathan (2008) also reports high rates of exclusion, particularly in states like Kerala

that switched from a universal PDS. Kochar (2005) argues that targeting reduced political

support for the program.

The second issue is corruption in the PDS, particularly through diversion of food grains

at different points along the distribution chain. This has been often highlighted in newspa-

per editorials, magazines, government reports and journals, though accurate estimates are

difficult to obtain. The Planning Commission (2005) estimates that the government of In-

dia spends Rs 3.65 to get Re 1 worth of subsidy to a BPL household. Based on a survey

conducted in 2001, the report concluded that 58% of food grains procured by PDS did not

reach their intended beneficiaries due to a combination of inaccurate targeting and diver-

sion. Khera (2011a) puts this number at 44% in 2007-08 by comparing official figures for

6A discussion of how the analysis deals with these implementation issues is presented in sections 3.3 and5.5.

8

food grains distributed through the PDS, with data on purchases from household surveys.

She divides states into those that are functioning, reforming, or languishing in terms of the

average quantity of food grains reaching households. Of late, significant state level differ-

ences in the working of the PDS have emerged and some states are improving delivery. A

nine state survey by Khera (2011b) conducted in May and June 2011 finds that over 85%

of the monthly entitlement was received by beneficiaries in these states. In November 2012,

the Indian government announced that a number of subsidy programs such as scholarships,

cooking fuel subsidies, pensions and unemployment benefits would be converted into direct

cash transfers in a phased manner starting in January 2013. This move is an attempt to

reduce inefficiencies and corruption in the implementation of various welfare schemes. Food

grains in the PDS are not yet a part of the proposed switch but some states, such as Bihar,

Madhya Pradesh and Delhi, are conducting pilot studies and have expressed an interest in

making that change.

3. Motivation

3.1 Conceptual Framework

A price discount with a quantity cap is typically represented by an outward shift of the

budget constraint. Figure 1 presents this shift in the household budget set with food on the

horizontal axis and the rupee value of non-food consumption on the vertical axis. Let the

price of non food purchases be normalized to Re 1. A subsidy comprising a discount of δ and

a quota Q shifts the budget set from NF to NCD. As demonstrated in Moffitt (1989) and

Deaton (1984), the household will choose segment I if its indifference curve is tangent at any

point (say, A) on the segment CD and a similar argument holds for tangency on segment

9

II. If tangency is not achieved on either segment, the household will choose the kink C. The

food subsidy program in India varies geographically and by family size in the amount of the

quota, the discount and the resulting value of the subsidy. Following Kochar (2005) and

Khera (2011c), these program rules are presented below by slightly modifying the standard

utility maximization problem.

Let the utility function for household i be Ui = f((1 − α)xi + yi, zi;Ti), where xi repre-

sents the subsidized food, yi represents food purchased from the market, zi is non-food

purchases and Ti denotes tastes which could depend on age, gender, composition and other

household characteristics. The subsidized and market food purchases are substitutes and

α (0 ≤ α ≤ 1) represents any (multiplicative) transaction costs associated with buying the

former. Let px, py and pz be the prices faced by the household where px= (1 − δ)py with

δ being the discount (0 ≤ δ ≤ 1). Let Q be the quota and Mi the household income. The

household’s maximization problem is:

maxxi,yi,zi

f((1− α)xi + yi, zi;Ti)

subject to:

pxxi + pyyi + pzzi ≤ Mi

xi ≤ Q

xi ≥ 0, yi ≥ 0, zi ≥ 0

10

The resulting Lagrangian is:

L = maxxi,yi,zi,λ,γ,µ1,µ2,µ3

f((1− α)xi + yi, zi;Ti) + λ(Mi − pxxi − pyyi − pzzi)

+γ(Q− xi) + µ1xi + µ2yi + µ3zi

Solving the first order conditions provides the demand functions for specific parameter val-

ues. The solution (x∗i , y

∗i , z

∗i ) will belong to one of the following four cases.

Case 1 : x∗i = 0, y∗i ≥ 0, z∗i ≥ 0

α > δ, MUy

MUz= py

pz(First order conditions)

Transaction costs (α) far outweigh the discount benefit (δ) and the household does not make

any purchases from the PDS (i.e. non participation).

Case 2 : 0 < x∗i < Q, y∗i = 0, z∗i ≥ 0

α < δ, MUxMUz

= pxpz

(First order conditions)

The household meets its entire food requirement through the PDS and does not need to

make any purchases from the market. This case is represented by point B in figure 1.

Case 3 : x∗i = Q, y∗i = 0, z∗i ≥ 0

α < δ, MUxMUz

> pxpz, MUy

MUz< py

pz(First order conditions)

The household would like to purchase more food at the discounted rate, but not at the mar-

ket price. This case refers to point C in figure 1, i.e. locating exactly at the kink.

11

Case 4 : x∗i = Q, y∗i > 0, z∗i ≥ 0

α < δ, MUy

MUz= py

pz(First order conditions)

The quota is binding and the household supplements its food with market purchases. This is

represented by point A in figure 1. Note that voluntary under-purchase, where 0 < x∗i < Q

and y∗i > 0 could take place if α = δ. However, this will never occur if there is even an

infinitely small (�) fixed cost associated with going to the fair price shop and thus, under-

purchase is likely driven by supply issues and collapses to case 4.

3.2 Expected Effect of the PDS

Post 1997, both the discount and the quantity cap in the PDS for BPL families were sub-

stantially increased. Khera (2011c) finds that participation in the program has dramatically

risen since the early 2000’s. She confirms that the conditions for case 1, where eligible house-

holds make no purchases from the PDS, are unlikely to be realized. Cases 2 and 3, where

households make purchases from the PDS but none from the market are also unlikely to

be relevant. The PDS is intended as a supplementary program and does not seek to meet

the entire food requirement of beneficiary households (Government of India, 2011). Kochar

(2005) and Khera (2011c) find that the quota is low enough to necessitate supplementary

purchases of food from the market for all households. The data from the NSSO surveys also

confirm that cases 1-3 are not supported. As shown ahead in section 4, all PDS beneficiary

households make food purchases in the market.7

The special case of households voluntarily under-purchasing from the PDS and buying addi-

tional food in the market is also not likely to hold. Khera (2011c) finds that the quantity of

7PDS rice expenditure accounts for less than 8% of the total food expenditure for a typical BPL household.

12

cereal purchased from the PDS does not respond to income, but overall cereal purchase does.

She also checks for the impact of other household characteristics on PDS purchases and finds

no effect. She concludes that under purchase is largely driven by supply side issues.8 Thus

based on program rules, past studies and summary statistics, the most relevant case is 4,

where the subsidy can be viewed as providing extra income to the household.9 This extra

income is given by the horizontal distance FD in figure 1: Si = (py − px)Q.

For a low enough quota and transaction costs less than the discount, the quota will bind

and the household will purchase food and other items in the market. To see this, consider a

Cobb-Douglas utility function of the form f = ((1− α)xi + yi)1/2z1/2i .10 The solution to the

household’s problem is:

y∗i =Mi −Q(2− α− δ)py

2py, z∗i =

Mi +Q(δ − α)py2pz

, x∗i = Q

conditional on:

Q < Mi(2−α−δ)py

(Quota < threshold value)

α < δ (Transaction costs < discount)

8In the IHDS (2005) data, of the poor households that had not accessed the PDS in the last 6 months,over 57% reported supply side constraints as the primary issue.

9This is also in agreement with opinions recently expressed in the popular press. Since all beneficiaryhouseholds purchase additional grain from the market, Arvind Panagariya believes that the proposed expan-sion of the PDS will have the same effect as a cash transfer of equal value (The Times of India, June 2013).Ashok Kotwal agrees that the PDS ’..is nothing but an income transfer.’ (The Economic Times, June 2013).

10Cobb-Douglas has the unique property that the proportion of the budget spent on food will be indepen-dent of the price of other commodities.

13

The total food consumption (F ∗i = y∗i +Q) is:

F ∗i =

Mi +Q(δ − α)py2py

where ∂Fi∗∂Q > 0 , ∂Fi∗

∂δ > 0 and ∂Fi∗∂α < 0.

Thus, the total food consumption is an increasing function of the quota and discount, and a

decreasing function of transaction costs. This motivates the use of a regression framework

taking into account the program rules, value of the subsidy, transaction costs and tastes.

Inefficiencies and corruption in the administration of the program would manifest in the

form of a higher α (poor quality, long waiting time, inconvenience, distance to FPS) or a

lower actual Q (supply side constraints).

3.3 Evidence on Food Subsidies and Nutrition

Food subsidies are an important policy tool and their impact on household outcomes such

as nutrition, caloric intake, early life development and income stabilization has been exam-

ined in both developed and developing countries. The national food stamps program (FSP)

in the US (renamed SNAP in 2008) has been studied extensively.11 Between 1961-75 the

program was phased in with different counties adopting it at different times. Hoynes and

Schanzenbach (2009) exploit this difference in timing to evaluate the impact of the program

and find that it led to an overall increase in food expenditures, as expected. Butler, Ohls

and Posner (1985) find very small effects of the FSP on the nutrient intake of the eligible

elderly, either through stamps or cash.

11See Hoynes and Schanzenbach (2009) for a review of the large literature on the food stamps program.

14

Despite lower per capita incomes and greater malnutrition, most studies in the context

of developing countries also find very small effects of food subsidies on nutritional outcomes.

Jensen and Miller (2011) conduct an experiment in rural China to estimate the impact of

price subsidies for rice and wheat and find no evidence that subsidies improve nutrition.

Laderchi (2001) finds that food transfers are no more successful at improving child nutrition

than other sources of income in Peru. In a survey of 300 households in Rajasthan, Khera

(2011c) finds that being eligible for PDS food grains does not lead to higher overall cereal

consumption. Kochar (2005) uses the change in 1997 from universal to targeted PDS to es-

timate the impact on calories consumed by rural households. She finds that the elasticity of

caloric intake with respect to the value of the subsidy is very low (0.06 on average.) Tarozzi

(2005) studies a sudden increase in the price of PDS rice in Andhra Pradesh and concludes

that it did not have a big impact on nutritional status and child anthropometrics.

While the effect of the PDS on nutrition has been found to be quite small, some limita-

tions of the identification strategies used in previous studies could cause their results to be

biased in the direction of finding no effect.12 Kochar (2005) uses variation in the value of the

subsidy that is determined by BPL status. The measure of eligibility for the newly targeted

program (BPL status) is imputed and not actually observed in the data which results in

errors of mis-classification. In practice, BPL status is determined using individual poverty

scores and region specific cut offs, both of which vary over time. Even within a particular

region, imputing BPL status using multi-dimensional indicators of poverty is problematic.

Niehaus et al. (forthcoming) survey households in Karnataka which were identified as po-

tential PDS beneficiaries in the government BPL survey. Using the criteria for 2007 (based

12Jensen and Miller (2011) raise similar concerns regarding the methodologies employed by Kochar (2005)and Tarozzi (2005).

15

on eight measures), they find that 13% of eligible households do not possess a BPL card,

while 70% of legally ineligible households do. Thus any exercise in imputing BPL status will

have significant mis-measurement resulting in biased estimates. Further, the time frame of

Kochar’s study is two to three years before and after the targeted program was introduced.

Take up rates were low (6%- 14%) during that time and the program was much less gener-

ous. Since 1997 and particularly during the early 2000’s, the PDS has undergone enormous

changes. There has been a push to increase the value of the subsidy to BPL households and

participation in the PDS has gone up.13 Finally, Kochar’s analysis also suffers from a bias

in the opposite direction. Since BPL status is a prerequisite for other types of government

assistance, using the BPL dummy could confound the effect of the PDS with the impact of

other government programs.14 Khera’s (2011c) study also uses variation due to BPL sta-

tus, and does not exploit differences in the value of the subsidy. Tarozzi’s (2005) analysis

focusses on anthropometrics for children below age 4; identification comes from variation

in the length of exposure to higher PDS prices but does not use differences in the quantity

entitlement. The length of exposure varies only from one to three months and the actual

receipt of benefits is not observed. His analysis focusses on the pre-targeted PDS in 1992,

when the program was much less generous. This paper adds to the literature by being the

first to compute a household specific value of the subsidy using state level program rules and

13Round 61 (2004-05) of the NSSO surveys provides information about the BPL status of a household.The data from this round suggest that participation is not driven by income or idiosyncratic householdcharacteristics that could potentially affect food consumption. 68% of BPL households in 2004-05 reporthaving accessed the PDS in the last 30 days. In data from the India Human Development Survey (2005),over 95% of BPL households report having accessed the PDS in the last 6 months. Thus, as the value ofthe program to BPL households has increased, there has been a corresponding increase in participation.Non-BPL participation in the PDS has fallen dramatically. In the 2004-05 round of the NSSO surveys,0.06% of PDS users were non-BPL. This decline has also been noted by Khera (2011a). Similar behaviourhas been observed in the context of other welfare programs – Brian Mc Call (1995) finds that take up ofunemployment benefits increases as the individual level recipient amounts increase.

14Assistance programs include loans, scholarships, medical benefits, distribution of bicycles etc.

16

local market prices.15

4. Data and Empirical Strategy

4.1 Data and Sample

The National Sample Survey Organisation (Ministry of Statistics and Program Implemen-

tation) has conducted socio-economic surveys every one to two years since the 1950s. The

surveys are repeated cross sections and collect detailed expenditure information. They are

nationally representative, conducted year round to avoid seasonal biases and form the basis

for the government’s official poverty estimates. The NSSO data have been used to study

the PDS and other national programs such as the employment guarantee scheme as well as

issues of inequality, poverty and land relations (Kochar 2005, Khera 2011a, Dutta et.al 2012,

Deshpande 2000 and Sundaram and Tendulkar 2003). This paper uses six years of data from

July 2002 to June 2008.16

The NSSO surveys split expenditure data into several sub categories such as food items,

beverages, durable goods, medical expenditure, educational expenditure, conveyance and

rent. Other variables include family size, number of children, industry classification of the

head of the household, location, land owned, own production, type of dwelling, religion and

social group. Data are collected on the age, gender, education level, marital status and

15Using the value of the subsidy, as opposed to a dummy for eligibility, also enables a comparison withthe results for price subsidies in Jensen and Miller (2011).

16In 2002, PDS allocations were increased from 20 kg to 35 kg per family, per month in most states.Some states started making changes (increases and decreases) in the PDS quotas in 2008. In the absence ofcredible official records of all the changes that ensued, I focus on the relatively stable period between 2002and 2008 for which reliable state level program rules are available.

17

relation to the head of each household member. For goods that are available through the

PDS, households report spending in terms of cost and quantity separately for PDS and

non-PDS sources. This enables the classification of households into PDS beneficiaries and

non-beneficiaries based on actual utilization as opposed to imputing BPL status. The final

analytic sample for this paper comprises the current participants of the PDS.17

Regional preferences for cereals are distinct and strong in India. Atkin (forthcoming) pro-

poses climatic conditions and habit formation as possible reasons for these distinct tastes

and previous studies on the PDS (Kochar 2005, Khera 2011c) have focussed either on rice

or wheat.18 Following the literature, I consider the eight states (151 districts) that have rice

as their dominant staple and a generous rice subsidy.19 These states offer a very small or no

subsidy for wheat. Though the main focus is on rice, the wheat subsidy in these states is used

as a validity check and is found to have a very small impact. Due to data availability and

reliability issues, it is standard practice to focus on the major Indian states. Tamil Nadu, a

state in southern India, is distinct from the rest of the country in that it has a universal PDS.

All households, irrespective of their income, are eligible for the program, making it difficult

to compare their outcomes with the rest of the country since the impact at the bottom of the

income distribution could be very different from its impact at the top. Thus, in the absence

of a credible way to classify households into BPL and APL, and given that the program is

so different in this one state, I drop Tamil Nadu from the sample. The states in the final

17As discussed earlier, non participation in the PDS is not likely to be driven by household specific factors.To the extent that some fraction of households with BPL cards do not avail of the program, strictly speaking,my analysis estimates the effect on eligible households that participate in the program.

18Jensen and Miller’s (2011) results on the impact of price subsidies in China vary substantially by provincesuggesting that the two staples (rice and wheat) should be analysed separately. In this paper, I examine therice subsidy and a similar analysis can be undertaken for wheat.

19Honyes and Schanzenbach (2009) also restrict their sample to limited subgroups of the population thathave a higher probability of participation in the food stamps program.

18

sample are: Andhra Pradesh, Assam, Chattisgarh, Jharkhand, Karnataka, Kerala, Orissa

and West Bengal. As a robustness check, I estimate the impact of the program in all the

major Indian states. As expected, the rice subsidy has a much smaller impact in the non-rice

favoring states. Aside from regional preferences, the rice subsidy specifically is of interest

because there is some evidence that rice may be a giffen good for poor households (Jensen

and Miller 2008). Finally, the rice subsidy in India is cleaner to study because the amount

of illegal diversion of wheat grains is much higher (Khera 2011a).

Prices are computed using households’ reported expenditures and quantities for over 150

food items. For every item, prices can be determined by dividing expenditure by quantity.

Though technically speaking these are unit values, which can give rise to measurement error

and concerns of differentiated products in terms of quality, it is common in the literature to

use them as proxies for prices (Subramaium and Deaton 1996, Kochar 2005, Atkin 2013). In

the context of this study, these issues pose much less of a threat because prices for PDS and

market rice are computed by averaging the district-season-year specific price as reported by

BPL households.20 Each household is assigned this average price and not its self reported

unit value. Further, only prices reported by similar households, i.e. beneficiaries of the PDS,

are used to compute the local average price. Thus the local average does not include prices

paid by much wealthier households, who might be purchasing higher quality rice. Following

Atkin (forthcoming), median local prices are used to calculate the value of the subsidy as a

check. The results are robust to using median prices, which are not likely to be influenced

by quality and outliers. All prices are in 2005 rupees.

20Every NSSO survey period (one year) is subdivived into four or six sub-rounds/waves. This is done toensure that seasonal patterns in consumption can be studied.

19

The conversion of food purchases into per capita caloric intake is done using standard factors

(NSSO 1996) for each of the food items in the survey and total monthly household calories

are converted into per capita daily amounts.21 While this may not be ideal due to potential

differences between purchases and actual intake, it is an approximation used not just in the

literature, but also by the government to estimate the income poverty line based on caloric

intake. 22 Deaton (1997) discusses the merits and limitations of several types of equivalence

scales to convert household consumption into per person quantities. For practical applica-

tions, a simple weight is suggested as a reasonable approximation. All per capita values are

thus calculated using an equivalence scale that assigns 0.5 weight for children below 15 years.

This allows for a correction in the per capita amounts for larger families that are likely to

have more children (who consume fewer calories).23

4.2 Variation in Value of the Subsidy

The value of the subsidy for each household is calculated as the product of the local price

discount and the state specific per capita quota. The discount is the difference between

the average market and average PDS price, at the district-season-year level. The quota is

calculated based on program rules of the state of residence and household size.24

PerCapV alSubijswt = (Pmktjwt − P pds

jwt ) ∗Qis

21Subramanium and Deaton (1996) attempt to correct this conversion from the NSSO surveys to accountfor the number of meals given to guests. However, since there is no way to determine the composition ofthose meals, caloric consumption from different food groups is not adjusted and neither is the expenditure.

22As discussed in Jensen and Miller (2011), data from more detailed individual food intake diaries raisetheir own set of concerns of validity and accuracy.

23Results remain qualitatively the same if a simple per capita variable is used instead. Elasticities computedat the household level are also found to be similar.

24As discussed ahead, I use the national average family size to compute the value of the subsidy for everyhousehold as a robustness check.

20

where i= household, j= district, s= state, w= season, t= year and Qis= state-household

size specific quota for household i.

Prices

Recall that the central government sets a PDS price for the year and states are allowed to

add transport and distribution costs to the final price that households pay. The PDS price is

not indexed to market prices, which are in turn determined by demand and supply side fac-

tors.25 The diverse nature of India’s terrain and climate combined with its large area results

in substantial geographic variation in the prices of agricultural commodities. As discussed in

Jacoby (2013), controls on the inter state trade of food grains lead to substantial differences

in prices across states. Deaton (1997) finds important interregional differences in prices

using NSSO data. Kochar (2005) and Atkin (forthcoming) ascribe the within state, inter

district variation in market prices of food grains to transportation costs and state specific

market controls.26 Agricultural prices also vary by season and as a result of local conditions

and unpredictable weather phenomena.27 Thus, as a result of the government’s policy of

maintaining a fixed (within year) PDS price across regions in India, the difference between

the market and PDS price varies for every district-season-year cell.28

25In a study conducted in Delhi, SEWA (2009) finds that 31% of respondents report that PDS pricesremain low, but market prices keep rising.

26Wadhwa (2001) details the restrictions on the marketing and movement of agricultural goods in India.27Planning Commission’s (2001) extensive study on the price behaviour of rice and wheat confirms sea-

sonality of prices. Sekhar (2003) reports a higher intra year variability in domestic agricultural prices ascompared to international prices which suggests incomplete integration of markets within the country.

28Antyodaya households are identified as being the poorest of the poor and they pay even lower prices. Inround 61 (2004-05) of the NSSO surveys and the IHDS (2005), these households are identified and summarystatistics confirm that they pay 30% lower prices on average at the PDS shop (appendix Table C). Theycomprise less than 10 % of the BPL sample in my data. In the full dataset, it is not possible to identify thesehouseholds and this source of variation is not explicitly used in the analysis. As a check, I use data fromround 61 to include an interaction between the value of the subsidy and an Antyodaya dummy and find noevidence of a differential impact of the subsidy.

21

Table 2 presents the spread of the discount across and within states for different seasons.

The state level average discount varies between 38% (Assam in the monsoon season) and

67% (Karnataka in winter). These rates are in line with the government’s policy of offering

food grains at approximately half of the procurement cost. Within any particular season

and state, there is substantial variation in the district level discount. Figure 2 confirms

this variation graphically for each district-season-year cell in the sample. Panel A plots the

discount data from 2002 for each of the 151 districts in the sample. There is heterogeneity

within states in the discount across districts. For instance, in 2002, the discount in Assam

varies from 15% to about 45% in the monsoon season, with a state level average of 38%. The

average discount in Andhra Pradesh is 55% but the spread (47% to 65%) is much smaller as

compared to Assam. A similar pattern of within state variation in the seasonal discount is

seen for the other six years (panels B to G, Figure 2) .

Quotas

Different states set different quotas for rice and wheat ranging from 0 to 35 kg per household

per month, as shown in Table 1. Some states do not index the quota to household size and

for those that do, there is an upper limit on the maximum amount (except in West Bengal).

Thus the per person quota varies by household, depending on its size and state of residence.29

Table 3 combines the two sources of variation and presents the spread of the per capita

value of the subsidy for the sample under study. As expected, the spread of the per capita

value (standard deviation) is the smallest in West Bengal and Andhra Pradesh, both of

29An examination of the average quantities bought via the PDS confirms that this variation is realized inpractice (Appendix Table A).

22

which index the subsidy to household size. The average subsidy in West Bengal is less than

half of the subsidy in Andhra Pradesh. Overall, the value varies across seasons, states and

within states.

4.3 Descriptive Statistics

Table 4 presents descriptive statistics for PDS users and the full NSSO sample averaged

over all six years. The monthly per capita expenditure for PDS users is lower than that

for the full sample, as expected. The average per capita caloric intake for PDS users is

2190 kcal. Given that over 78% of PDS users live in rural areas, this average is well below

the minimum daily requirement (2100 kcal for urban and 2400 kcal for rural areas). PDS

users spend a higher fraction of their total monthly expenditure (58%) on food. Scheduled

Caste/Scheduled Tribe/Other Backward Castes comprise 76% of PDS users.

The PDS price is 50% lower than the market price. On average, each household buys

more rice in the market than what they receive through the program – the PDS contributes

41% of the total rice consumed. Thus, the PDS is an important source, but not one that

meets the entire food requirements of the household; almost three-fourths of the total food

expenditure is on items other than rice.30 Per capita cereal expenditure is a little over one

fourth of food expenditure, but cereals contribute over 72% of the total calories consumed.

30The fact that households also purchase rice in the market suggests that resale of PDS rice throughconsumers is not a serious concern.

23

4.4 Empirical Specification and Assumptions

The basic equation to estimate the impact of the subsidy on household outcomes is

Yijswt = α + βPerCapV alSubijswt +Xijswtγ + δj + χw + θst + εijswt (1)

Yijswt represents the outcome variable (such as per capita caloric intake) for household i in

district j, state s, season w and year t. PerCapV alSubijswt is the per capita value of the

subsidy to the household based on program rules and local prices. Xijswt is a vector of house-

hold characteristics, δj and χw control for district-specific and seasonal effects respectively,

and θst are state*year dummies. Standard errors are clustered at the district level, as all

households in a district are subject to the same rules determining the per person monthly

quota.

Household characteristics in X include education of the head of the household and the

spouse, a quadratic in the age of the household head, the proportion of females in the house-

hold, land holdings (proxy for income) and urban location. Behrman and Deolalikar (1988)

provide a comprehensive review of the literature on the demand for nutrients in developing

countries and find that education levels, size and demographic composition of the households

are important determinants of the demand for calories.31 While most studies on the PDS

have focussed on the rural poor, the program is also critical in urban areas and the nutri-

31Though the dependent and main explanatory variable in equation (1) are both in per capita terms, theremay still be economies of scale within a household that are not accounted for by this specification. I checkfor any effect of household size in the following ways. First, I use an alternative weighting scheme to generateper person variables for the household. Second, I use household level caloric intake and value of the subsidyto estimate the impact and explicitly control for family size. Third, I shut down the variation due to familysize by using quotas based on the national average family size. Finally, even though household size as aregressor in equation (1) is potentially endogeneous, as an additional check, I find that adding size to theregression does not qualitatively change the results.

24

tional needs of households in the two areas are very different. District and seasonal dummies

are included to take regional and climatic effects on the demand for calories into account.

The model is saturated by introducing state*year dummies that control for any differences

in state level effects and other assistance programs that BPL households in a particular state

might be eligible for. With district, season and state*year controls, any variation in the price

discount is due to unpredictable shocks to prices as a result of random weather phenomenon,

arbitrary controls on the movement of goods and imperfectly integrated agricultural markets.

The parameter of interest from equation (1) is β, the coefficient on the value of the sub-

sidy. The validity of the model rests on the assumption that after controlling for household

characteristics and geographic and seasonal effects, the value of the subsidy (price discount*

quota) is exogenous to unobservable factors that may affect the demand for food. If this

assumption holds, the value of the subsidy should have a negative (through market price) or

zero effect on the caloric consumption of households that do not use the program. To check

for this, I perform a falsification test using non-PDS users.32

The analysis makes two other substantive assumptions. First, the model assumes that

household size is exogenous to the state level quota, i.e. households do not adjust fam-

ily size according to more or less generous program rules. While the program is important,

it is not likely to affect the fertility or migration decisions of households. On average, the

value of the subsidy is 4% of the total monthly expenditure of the household. I check for the

validity of this assumption by using the national average family size instead of the actual

size of the household to compute the value of the subsidy and find that the results remain

32Tables 12-15 ahead present the results of multiple robustness checks.

25

qualitatively the same. I also check for the independent effect of family size by using total

household caloric intake as the outcome and explicitly controlling for family size. Second,

the analysis assumes that the demand for calories or rice by any one household does not

affect the market price that it faces i.e. the local price in its district-season-year cell. In the

absence of this assumption, the coefficients would suffer from simultaneity bias. Treating the

household as a price taker is a standard assumption from the theory of competitive markets.

The agricultural market in India has thousands of producers and consumers, which makes

this assumption fairly reasonable in the context of this study.

5. Results

5.1 Impact on Cereal Consumption and Caloric Intake

To study the impact of the PDS on nutrition, three outcome variables are considered: per

capita cereal consumption, per capita caloric intake and per capita calories from different

food groups. Cereals are a cheap and critical source of calories and the PDS is responsible

for providing a substantial amount of cereals to vulnerable households.33 Column 1 of Table

5 presents results for the main specification with per capita cereal intake as the outcome

variable. An increase in the monthly subsidy by Rs 10 leads to a 20.3 gm (60 kcal/day)

increase in the daily consumption of cereals. Based on average market prices, Rs 10 (per

month) can buy an extra 30.2 gm of cereals every day. This suggests that infra marginal

households prefer to spend part of the extra income on foods other than cereals or on non-

food items. The specification in column 1 assumes that an (absolute) increase in the value

33Most cereals have 3-3.5 kcal/gm. See Appendix Table B for a comparison of the price per calorie fordifferent food groups.

26

of the subsidy has the same effect, irrespective of the base value. Evaluated at the means,

the estimate from column 1 implies an elasticity of 0.12, which is similar to the estimate

from the log-log regression in column 2– 0.123. Given that the elasticity is small and the

sample comprises only BPL households, it is not surprising that the estimates are so close.34

Though small, the elasticity is positive and significant and confirms that the subsidy has a

real impact on cereal consumption, as expected. In order to examine the impact of each of

the components of the subsidy and determine their relative importance, column 3 splits the

value of the subsidy into quota, prices and interaction terms. The coefficient on quota is

insignificant, confirming that households are infra-marginal since the amount of the quota

doesn’t determine overall cereal consumption. The coefficient on market price is negative, as

expected. The interaction between the amount of the quota and the market price is positive

and significant, suggesting that a higher quota increases cereal consumption more, and is

thus more valuable, when the market price of rice is higher.

Though the estimates suggest that the program has a positive effect on cereal consumption,

households may be substituting some of the income gains away from cereals. Combining the

analysis for calories with the results on cereals helps to determine the impact on overall food

intake. Column 1 in Table 6 indicates that a Rs 10 increase in the value of the rice subsidy

results in an increase an increase of 126 kcal/day.35 This is more than double of the impact

34The levels estimate assumes that the impact of an absolute increase in the value of the subsidy isindependent of where in the distribution the household is located (linear relationship). In this sample,households are homogeneous to the extent that they all lie below the poverty line and receive a positive foodsubsidy. The log-log specification assumes a non-linear relationship and a constant elasticity. However, foran elasticity of 0.12, the resulting curve is very flat, i.e. almost linear, over the range of the value of thesubsidy, cereal consumption and caloric intake.

35In order to compare the estimates, I convert the gains in daily cereal consumption from Table 5 intocaloric gains. The 95 % confidence interval from Table 5 implies an increase between 51 kcal/day and 69kcal/day. This does not overlap with the confidence interval from Table 6 which is 104 kcal/day to 148kcal/day. Thus, the overall gains in caloric intake are greater than the increase in calories from cereals alone.

27

on calories through increased cereal consumption (60 kcal/day from by column 1 in Table 5),

suggesting that the subsidy works by increasing calories through not just cereals, but other

foods as well. This is an important indicator that households benefit from the program in

terms of overall food intake and not just through the food grains directly provided by the

PDS.

Table 6 also presents estimates of the cross food group elasticities.36 In China, Jensen

and Miller (2011) find negative effects of an experimental price subsidy for rice on fruits and

vegetables and on pulses (lentils). In contrast, all elasticities here are positive and significant.

Thus, the results suggest that provision of a food subsidy via quotas has a positive impact

on all food groups, similar to what would be expected from an income transfer whereas the

pure price subsidy causes households to consume less of the subsidized good and fewer overall

calories.

5.2 Comparison with the Expenditure Elasticity of Calories

The PDS subsidy is supplementary in nature and thus is expected to operate through the

income effect. The elasticity of calories with respect to the value of the subsidy is 0.144

from column 2 of Table 6. Though small, it lies between the historical estimates of income

elasticity of calories for India, which range from 0 to 0.34. Behrman and Deolalikar (1990)

find the income elasticity of calories for rural south India to be very low: 0.01. Subramanian

and Deaton (1996) use NSSO data and report an expenditure elasticity of calories of 0.34 for

rural households. To compare my results with these earlier studies, I estimate the elasticity

of calories with respect to total expenditure for rural PDS households in my sample. The

36For the sample in this paper: 5% of total calories come from lentils, 5% from fruits and vegetables andapproximately 1% from meat, fish and other animal products.

28

expenditure elasticity is close to the higher estimates in the previous literature: 0.4 from

column 1 in Table 7. Even though the elasticity of caloric intake with respect to the subsidy

is lower than the expenditure elasticity (0.140 for the rural sample), it is still positive and

significant.37

The fact that the PDS acts like an income transfer raises the question of its impact on

price paid per calorie i.e., the extent to which a rise in income induces households to sac-

rifice caloric intake for more expensive and less calorie rich foods. Behrman and Deolalikar

(1989) suggest that the income elasticity of calories is smaller than the income elasticity

of food expenditures because of changes in the shape of the food indifference curve as in-

come rises. They find that even relatively poor households value variety. Subramanian and

Deaton (1996) report the total expenditure elasticity of expenditure on food as 0.75 and the

elasticity of price paid per (1000) calories as 0.35. These numbers are very close to the esti-

mates for the sample used in this paper (columns 3 and 5 in Table 7). The elasticity of food

expenditure with respect to the value of the subsidy is lower at 0.146, but it is positive and

significant. The subsidy elasticity of price per calorie is not significantly different from zero.

Thus the value of the subsidy is not large enough to raise concerns of households sacrificing

calories for taste.

One issue with the NSSO data (and any estimates based on them) is that they do not

contain information on income. Table 8 presents alternate estimates of the elasticity of

37It is not surprising that the elasticity of caloric intake with respect to the value of the subsidy is lowerthan the expenditure elasticity. This is true for two reasons: First, transaction costs associated with goingto the fair price shop, such as irregular supply, inconvenient timings and long queues would make the impactof the subsidy smaller. A pure increase in income does not have any such costs associated with it. Second,this estimate takes into account leakages and inefficiencies in the system since it is based on actual programrules.

29

cereal consumption using income data from the India Human Development Survey 2005.38

The survey is nationally representative, covers 41,554 households and collects detailed house-

hold information on demographics, income, debt, insurance and consumption. The income

elasticity of cereals for PDS users is much lower (0.046 in column 1, panel A of Table 8)

than the elasticity with respect to the rice subsidy which is 0.295 in column 2. The income

elasticity of food expenditure is also much lower than the elasticity of food expenditure with

respect to the value of the subsidy from columns 3 and 4 of panel A. While income is a

reflection of more permanent characteristics of a household such as age and education of

the head, location and occupation, the value of the subsidy fluctuates from month to month

depending on market prices. Not all income is consumed, part of it saved or used to purchase

durable goods, thus it is not surprising that an increase in the value of the subsidy has a

bigger impact on cereal consumption than an increase in income. Further, income tends to

be under-reported and the resulting measurement error would bias the estimates downwards.

The IHDS dataset allows me to check if the household characteristics used in the main

regression are a valid control for income. Column 2 of panel A in Table 8 estimates the

elasticity with respect to the value of the subsidy controlling for income, location, district

and season effects. In column 3 of panel B, I run the same regression, but instead of explic-

itly including income, I add the education of the head of the household and the spouse, a

quadratic in the age of the household head, the proportion of females in the household and

land holdings. The estimates are similar, which validates the use of these characteristics in

place of income.

38This survey is jointly conducted by National Council of Applied Economic Research in Delhi and theUniversity of Maryland (Desai, Vanneman, & National Council of Applied Economic Research, 2005). I usethe IHDS dataset to construct variables that are identical to variables from the NSSO data. I also restrictthe sample to the eight rice favoring states used in the main analysis.

30

The food module of the IHDS surveys is similar to the NSSO, but is much less detailed.

To check for any effect of differences in survey methodologies, in panel B of Table 8, I es-

timate identical regressions for the IHDS and data from the NSSO covering the same time

period (November 2004 - October 2005). The estimates for expenditure elasticity of cereals

from the two datasets are similar: 0.247 and 0.269 in columns 1 and 2 of panel B.39 The

estimates of the elasticity of cereal consumption with respect to value of the subsidy across

the two samples (panel B columns, 3 and 4) are also similar, though the IHDS estimate

is slightly larger. The NSSO sample is larger than the IHDS, but the two don’t differ on

observable characteristics.40 PDS prices are lower, and consequently the subsidy value is

marginally higher on average (Rs 25.46) in the IHDS sample as compared to the NSSO data

(Rs 24.17).41 While it is difficult to say with certainty why the estimates differ, the IHDS

results do not qualitatively change the implications of the main analysis. At best, they

suggest that the estimates from the NSSO data may be a lower bound on the impact of the

program.

5.3 Heterogeneous Impacts

Households that benefit from the PDS are similar to the extent that they all lie below the

poverty line. However, the program could have differential impacts on certain sub-groups of

the population. Table 9 checks for heterogenous impacts of the subsidy on urban households,

39The expenditure elasticity of cereal consumption (column 2 of panel B, table 8) is lower than theexpenditure elasticity of calories (column 1, table 7). Deaton and Dreze (2009) find a similar pattern andattribute it to the fact that a higher fraction of the marginal rupee spent on food goes towards non-cerealfood items.

40Appendix Table D presents descriptive statistics for the two samples.41Price data in the IHDS is collected directly, by asking households for their estimate of the average market

price of various goods over the last 30 days.

31

traditionally marginalized communities, farming households and households in the lowest ex-

penditure quartile. While urban households face different patterns of price changes compared

to rural ones and may have different mechanisms to cope with food insecurity, (rice) farming

households may be more or less affected by market prices depending on whether they are net

buyers or sellers in the market. Traditionally marginalized communities may face discrimina-

tion in terms of access to the program and households in the lowest expenditure quartile may

lack sufficient income to fully leverage the subsidy provided through the program. Residing

in an urban area has a negative and significant impact on cereal consumption (column 1).

This is expected as urban calorie requirements are lower on average. However, the program

does not have a significantly different impact in urban areas. Columns 2 and 3 check for

the impact on cereal consumption of traditionally marginalized communities (by caste) and

the poorest quartile of the sample, respectively. There is no significantly different impact

for the SC/ST/OBC category. Households in the lowest expenditure quartile consume fewer

cereals, but the subsidy does not have a differential impact on them.

The program has a strong, positive effect on the cereal consumption and caloric intake

of farmers, i.e. households that report consuming some of their home grown rice (columns

4 and 8). Given that these households need to rely less on the market and are not likely

to make cereal purchases, the subsidy acts like a direct in-kind transfer of extra grains. As

there is no income gain, these household can not substitute cereals with other types of food

or other goods. For the sample of farmers, Table 10 splits the subsidy into its components

and confirms that market price is not a significant determinant of caloric intake (column 2)

and marginally significant for cereal consumption (column 1). This is in contrast to results

for the overall sample (column 3 of Table 5) where market price was seen to have a significant

32

and negative impact. The interaction between market price and per capita quota has the

expected positive sign and is significant.

From column 7 of Table 9, the impact of the subsidy on the caloric intake of the lowest

income quartile is significantly lower, compared to other households. This group represents

the poorest of the poor. For an increase in the value of the subsidy, these households appear

to substitute away from food towards other non food items. Table 11 presents elasticities

of cereal consumption and caloric intake for the four income quartiles of PDS users. The

elasticities increase with income, suggesting that the poorest households (also likely to be

the most credit constrained) use the extra income generated by the subsidy to purchase

non food items.42 This is consistent with the hypothesis in Sen (2005) that an increase in

the availability, demand and prices of non-food essential goods and services has resulted in

squeezing of the food budget of poor families.

5.4 Robustness and Sensitivity Checks

The identification strategy rests on the assumption that there are no underlying factors that

simultaneously affect the value of the subsidy and food consumption of households. To test

the validity of this assumption, I perform a falsification test by checking for the impact of

the local average value of the PDS subsidy on the cereal consumption and caloric intake of

households that do not receive subsidies from the PDS. From Table 12, it is clear that the

program has no impact on the consumption of Non-PDS beneficiaries. The only (marginally)

significant effect is the elasticity of cereal consumption. As expected, this is negative since

42There is no difference in terms of quantity of rice purchased from the PDS across expenditure quartiles,implying that even households in the lowest quartile purchase the full PDS entitlement and the lower impacton their cereal consumption is not driven by an inability to fully avail of the PDS discount.

33

a higher value of the subsidy (resulting from a higher market price of rice) would negatively

affect cereal consumption of households that purchase cereals only in the market.

Table 13 presents results from a series of robustness checks. Concerns about the possible

endogeneity of household size are addressed in columns 1 and 2. In column 1, the national

average family size is used to calculate the value of the subsidy instead of the actual size and

composition of each household. Column 2 presents the elasticity results at the household

level, with a control for household size. The estimates from both are very close to the main

result (0.144) which supports the assumption that household size is not influenced by the

program. To check that the results are not driven by the equivalence scale used in the main

analysis, column 3 uses a simple per capita estimate. The results remain qualitatively simi-

lar to those from the main analysis, though using a simple per capita without correcting for

the composition of the household makes the elasticity appear larger. Averaging across the

market price of rice within a district-season-year cell should not raise concerns of significant

differences in quality, since households in the sample all lie below the poverty line. However,

I check for this possibility by using the median prices to calculate the value of the subsidy

and find that the results remain the same (column 4). Finally, I check the sensitivity of my

results to seasonal price patterns by utilizing survey waves/sub-rounds. Replacing state*year

and season dummies with state*survey wave dummies does not change the estimates, as seen

in column 5.

Table 14 checks for the impact of the rice subsidy on three different samples – All India,

the rice favoring states and non-rice favoring states. As expected, the elasticity is much

smaller in states where rice is not the main staple, even though some of them offer a small

34

subsidy on rice. This justifies using the smaller sample of rice favoring states to get a more

valid estimate of the impact of the rice subsidy program. Conversely, Table 15 checks for

the impact of the (much smaller) wheat subsidy in the main sample of rice favoring states.

The effect of the wheat subsidy on both cereal consumption and caloric intake is extremely

small and its inclusion makes the effect of the rice subsidy a bit larger. To the extent that

the value of the subsidy is higher when the market price of rice (and possibly other goods)

is higher, this is expected because the wheat subsidy provides a control for the overall level

of market prices.

5.5 Issues in Implementation and Policy Changes

Corruption

The PDS has been criticized for various types of inefficiencies and corruption and there are

striking state-wise differences in implementation. Khera (2011a) estimates the extent of

diversion in food grains using a combination of NSSO data and administrative records on

grain allocation at the state level. She categorizes states into those that are performing (well)

and others that are reforming or languishing.43 Table 16 presents results for the impact on

cereal consumption and caloric intake. Columns 1 and 2 suggest that the impact on cereals

and caloric intake is almost 50% lower in states that are more corrupt.44 This is a huge cost

of inefficiency and is particularly troubling given the high levels of malnutrition that persist

in India.43Of the eight states in the sample, three are categorized as functioning well: Andhra Pradesh, Karnataka

and Kerala.44These estimates provide suggestive, not causal evidence of the effect of corruption as there could be

many other state level factors that make the program less (more) effective in more (less) corrupt states.

35

The National Food Security Bill

The PDS is the government’s flagship program for improving nutritional outcomes and the

National Food Security Bill (NFSB), passed in September 2013, plans to expand it further.

Unless the enormous leakages in the system are plugged, the government will have to procure

a much higher quantity of food grains (than the target) in order to have the intended effect

on nutritional outcomes. A rough estimate of the impact of the NFSB can be made using

price data, program rules and the elasticity estimates from this paper. The NFSB assures 5

kg of food grains per person per month to 67 % of the population at Rs 3 per kg for rice, Rs 2

per kg for wheat and Re 1 per kg for coarse grains. BPL households are currently eligible for

an average of 4 to 5 kg per person per month.45 At current prices, it is estimated that the per

kilogram subsidy will rise from Rs 13.5 to Rs 16.5 (The Financial Express, September 2013).

Thus the NFSB does not substantially increase the quantity of food grains assured to BPL

households, but it does entail a big increase in the price discount they receive. Combined

with average caloric intake and current price data, the elasticity estimate from this paper

suggests that the bill will lead to a per person increase of 72 kcal/day in rural areas and

66 kcal/day in urban areas for the current beneficiaries of the program.46 This estimate is

based on program rules and thus takes into account the leakages and inefficiencies in the

system. If the expansion of the PDS is accompanied by better enforcement, especially in the

more corrupt states, the impact will be even higher.

The bill will also expand the beneficiary pool of the PDS to include 67% of the popula-

45Many state governments such as Tamil Nadu, Andhra Pradesh and Chattisgarh already have a moregenerous PDS than the national average and provide food grains to a large fraction of their population atthese low prices. Further, Antyodaya households are eligible for 35 kg per month at even lower prices.

46This estimate uses the approximate caloric intake of PDS rice users in rural (2260.25 kcal/day) andurban areas (2076.5 kcal/day) from the 2009-10 round of the NSSO Surveys (Government of India, 2013).

36

tion.47 While it is difficult to precisely predict how newly eligible households will respond

to the subsidy, there is some evidence to suggest that there will be a positive effect (similar

to that for current beneficiaries) on their caloric intake. Given that the new beneficiaries

are likely to be better off, the subsidy will not be their primary source of food grains and

thus, should have a positive effect on caloric intake through the standard income effect. This

hypothesis is supported by the results in Table 11, which show that the elasticity of caloric

intake is higher for households in the higher expenditure quartiles. There is a concern that

expansion in the reach of the program could result in the inclusion of some households that

have an income high enough to make the relative value of the subsidy too insubstantial to

have an effect. These households are also likely to have have lower rates of participation,

which would further reduce the overall impact of the program on caloric intake. Neither of

these concerns is found to relevant in Tamil Nadu, which is the only state in India that has

a universal PDS. The participation rate for the entire population of the state (averaged over

the six years) is high (61%) and the elasticity of caloric intake with respect to the value of

the subsidy is 0.2, which is higher than the national average (0.144). Tamil Nadu has a well

functioning PDS and the data from this state suggest that if implemented well, the program

can have a substantial impact on caloric intake for a wide range of households.

6. Conclusion

This paper examines the Indian Public Distribution System and presents evidence of its

impact on nutrition using variation in state specific program rules and fluctuations in local

market prices. In agreement with the literature on food subsidies, the elasticities for cereal

consumption and calories with respect to the value of the subsidy are small. However the

4744.5% of the population currently participates in the PDS (The Indian Express, July 2013).

37

results indicate that households benefit from the program in terms of food intake (calories)

and not this is not just through the food grains (cereals) directly provided by the PDS.

Even though the program provides a subsidy only on cereals, it has a positive effect on the

consumption of different food groups. Thus, the PDS subsidy generates an income effect

for households and is effective in improving nutrition. The results also confirm state wise

differences in the functioning and impact of the PDS. Finally, the elasticity estimates suggest

that the implementation of the National Food Security Bill will lead to an increase of 66 -

72 kcal in the daily caloric intake of current beneficiaries of the program.

Some states and regions in India have had much more success in implementing the PDS

than others. The Indian Human Development Survey is a rich socio-economic dataset that

makes it possible to study these regional differences in more detail. In additional to details

on household expenditure, income and fertility, indicators for learning skills and anthropo-

metrics for children are also collected. Examining district level outcomes using participation

and average value of the subsidy would be an additional dimension to studying the PDS.

This would speak to the significant geographic differences that have been noted in the func-

tioning of the system. Another unexplored aspect of the PDS is the protection it provides

from fluctuations in market prices. This would be an interesting angle to assess its value as

a source of food security. Reduced exposure to market risk could also affect a households

investments and labour market choices.

The PDS has been the topic of a lot of debate recently. The National Food Security Bill,

which was passed in September 2013, will give two thirds of the population the legal right

to obtain 5 kg of food grains (per person per month) at prices between Re 1 and Rs 3 per

38

kg The bill also has implications for intra-household bargaining, since it makes the eldest

woman in the household responsible for accessing the PDS on behalf of her family. Given the

enormous scale of the PDS, its expansion and imminent overhaul, it is important to study

the impact that it has in its present form.

References

[1] Atkin, D. (Forthcoming). Trade, Tastes and Nutrition in India.The American Economic

Review.

[2] Atkin, D. (2013). The Caloric Costs of Culture: Evidence from Indian Migrants (No.

w19196). National Bureau of Economic Research.

[3] Banerjee, A. V., and Duflo, E. (2011). Poor economics A radical rethinking of the way

to fight global poverty. PublicAffairs.

[4] Behrman, J. R., and Deolalikar, A. B. (1990). The intrahousehold demand for nutrients

in rural south India: Individual estimates, fixed effects, and permanent income. Journal

of human resources, 665-696.

[5] Behrman, J. R., and Deolalikar, A. (1989). Is variety the spice of life? Implications for

calorie intake. The Review of Economics and Statistics, 666-672.

[6] Behrman, J. R. and Deolalikar, A. B. (1988). Health and nutrition. Handbook of devel-

opment economics, 1, 631-711.

[7] Butler, J. S., Ohls, J.C., and Posner, B. (1985). The Effect of the Food Stamp Program on

the Nutrient Intake of the Eligible Elderly. Journal of Human Resources, 20(3):405-420.

39

[8] Deaton, A. (1984). Household surveys as a data base for the analysis of optimality and

disequilibrium. Sankhya: The Indian Journal of Statistics, Series B, 228-246.

[9] Deaton, A. (1997). The analysis of household surveys: a microeconomic approach to

development policy. Johns Hopkins University Press.

[10] Deaton, A., and Dreze, J. (2009). Food and nutrition in India: facts and interpretations.

Economic and political weekly, 42-65.

[11] Desai, S., Vanneman, R., and National Council of Applied Economic Research. (2005).

India Human Development Survey (IHDS). Computer File, ICPSR22626-v5. Ann Arbor,

MI: Inter University Consortium for Political and Social Research [distributor], 2009-622.

http://dx.doi.org/10.3886/ICPSR22626.

[12] Deshpande, A. (2000). Does caste still define disparity? A look at inequality in Kerala,

India. The American Economic Review, Vol. 90, No.2, pp. 322-325.

[13] Dreze, J., and Khera, R. (2010). The Bpl census and a possible alternative. Economic

and Political Weekly, 45(9), 54-63.

[14] Dutta, P., Murgai, R., Ravallion, M., and Van de Walle, D. (2012). Does India’s

employment guarantee scheme guarantee employment? World Bank Policy Research

Working Paper, (6003).

[15] The Economic Times (June 14 2013). Why the Food Security Bill is neither populist

nor unaffordable. Viewed on July 11 2013, http://articles.economictimes.indiatimes.com

/2013-06-14/news/39976460 1 national-food-security-bill-cash-transfers-food-subsidy

40

[16] FAO (1997). Safety nets to protect consumers from possible adverse effects Economic

and Social Development Department, Food and Agriculture Organization of the United

Nations.

[17] The Financial Express (September 2 2013). Correct Costs of the Food Security Bill

Viewed on September 3 2013, http://www.financialexpress.com/news/correct-costs-of-

the-food-security-bill/1156251/0

[18] Government of India (2013). Public Distribution System and Other Sources of Household

Consumption. NSS 66th Round. Report No. 545 (66/1.0/6). National Sample Survey

Office (Ministry of Statistics and Programme Implementation). Government of India,

New Delhi.

[19] Government of India (2012). Union Budget 2012-13. Expenditure Budget Volume I

(Ministry of Finance). Government of India, New Delhi.

[20] Government of India (2011). Annual Report 2010-2011. Department of food and public

distribution (Ministry of Consumer Affairs, Food and Public Distribution). Government

of India, New Delhi.

[21] Hoynes, H. W. and Schanzenbach, D. W. (2009). Consumption Responses to In-Kind

Transfers: Evidence from the Introduction of the Food Stamp Program. American Eco-

nomic Journal: Applied Economics, American Economic Association, vol. 1(4), pages

109-39, October.

[22] The Hindu (July 5 2013). President agrees to Food Bill Ordinance. Viewed on

July 9 2013, http://www.thehindu.com/news/national /president-agrees-to-food-bill-

ordinance/article4884834.ece

41

[23] The Indian Express (July 6 2013). Manmonia’s FSB: 3% of GDP. Viewed on September

3 2013, http://www.indianexpress.com/news/manmonias-fsb-3–of-gdp/1138195/1

[24] Jacoby, H. G. (2013). Food prices, wages, and welfare in rural India. World Bank Policy

Research Working Paper, (6412).

[25] Jensen, R. T., and Miller, N.H. (2008). Giffen Behavior and Subsistence Consumption.

The American Economic Review, 98:4, 15531577.

[26] Jensen, R. T. and Miller, N.H. (2011). Do consumer price subsidies really improve

nutrition? Review of Economics and Statistics, 93(4), p. 1205-1223.

[27] Karat, B. (2013). Food Matters: Law, Policy and Hunger. Prajashakti Book House.

[28] Khera, R. (2008). Access to the Targeted Public Distribution System: A Case Study

in Rajasthan. Economic and Political Weekly, Vol - XLIII No. 44, November 01.

[29] Khera, R. (2011a). Trends In Diversion Of PDS Grain. Economic and Political Weekly,

Vol - XLVI No. 21, May 21.

[30] Khera, R. (2011b). Revival of the Public Distribution System: Evidence and Explana-

tions. Economic and Political Weekly, Vol - XLVI No. 44-45, November 05.

[31] Khera, R. (2011c). India’s Public Distribution System: Utilization and Impact. Journal

of Development Studies, March 07.

[32] Kochar, A. (2005). Can Targeted Programs Improve Nutrition? An Empirical Analysis

Of India’s Public Distribution System. Economic development and cultural change, 54

(1), 20335.

42

[33] Laderchi, C. R. (2001). Killing Two Birds With The Same Stone ? The Effectiveness

Of Food Transfers On Nutrition And Monetary Poverty. Working Paper no. 74. QEH

working paper series.

[34] McCall, B. P. (1995). The Impact of Unemployment Insurance Benefit Levels on Re-

cipiency. Journal of Business and Economic Statistics, American Statistical Association,

vol. 13(2), pages 189-98, April.

[35] Moffitt, B. Y. R. (1989). Estimating the value of an In-kind Transfer: The case of Food

stamps. Econometrica, 57(2), 385-409.

[36] National Sample Survey Organisation (1996). Nutritional Intake in India. Report No.

405. National Sample Survey Organisation, Delhi.

[37] Niehaus, P., Atanassova, A., Bertrand, M., and Mullainathan, S. (forthcoming). Tar-

geting with Agents. American Economic Journal: Economic Policy.

[38] Planning Commission (2001). An Analysis of the price behavior of selected commodities.

National Council of Applied Economic Research, India.

[39] Planning Commission (2005). Performance Evaluation of Targeted Public Distribution

System (TPDS). Programme Evaluation Organisation. Government of India.

[40] Planning Commission (2012). Twelfth Five Year Plan, 2012-17. Government of India.

[41] Sen, P. (2005). Of Calories and Things: Reflections on Nutritional Norms, Poverty

Lines and Consumption Behaviour in India. Economic and Political Weekly, 40(43), 22

October, 4611-8.

43

[42] Sekhar, C. S. C. (2003). Volatility of agricultural prices-an analysis of major inter-

national and domestic markets. Indian Council for International Economic Relations.

Working paper no 103.

[43] SEWA (2009). Do poor people in Delhi want to change from PDS to cash transfers?

Self Employed Womens Association.

[44] Stifel, D. and Alderman H. (2006). The Glass of Milk Subsidy Pro- gram and Malnu-

trition in Peru. World Bank Economic Review, 20:3 (2006), 421448.

[45] Subramanian, S., and Deaton, A. (1996). The demand for food and calories. Journal of

political economy, 133-162.

[46] Sundaram, K. and Tendulkar, S. D. (2003). Poverty has declined in the 1990s: A

resolution of comparability problems in NSS consumer expenditure data. Economic and

Political Weekly, 327-337.

[47] Swaminathan, M. (2008). Programmes to Protect the Hungry: Lessons from India.

DESA Working Paper No. 70.

[48] Tarozzi, A. (2005). The Indian Public Distribution System as Provider of Food Security:

Evidence from Child Nutrition in Andhra Pradesh. European Economic Review, 49, p.

1305-1330.

[49] The Times of India (June 1 2013). Why the food security Bill will not boost

foodgrain consumption for the poor. Viewed on July 11 2013, http://articles.

timesofindia.indiatimes.com/2013-06-01/edit-page/39674597 1 cereal-consumption-

food-security-bill-households

44

[50] Wadhwa, M. (2001). Parking Space for the Poor: Restrictions Imposed on Marketing

and Movement of Agricultural Goods in India. Centre for Civil Studies.

45

Figure 1: Impact of a food subsidy on the budget set

15

Food

No

n F

oo

d

C

Q

Segment II

Segment I

B

A

N

F D

46

Figure

2:Variation

inPDSrice

discount

(2002-2008)

Pan

elA

020406080 020406080

Monso

on

Post M

onso

on Monso

on

Post M

onso

on Monso

on

Post M

onso

on Monso

on

Post M

onso

on

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

002

Pan

elB

020406080 020406080 Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

003

Pan

elC

020406080 020406080 Winter

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

004

Pan

elD

020406080 020406080 Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer

Post M

onso

on

Monso

on

Summer

Winter

Monso

onPos

t Mon

soon

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

005

Notes:1.

Rou

nd58

surveyed

hou

seholdsbetweenJu

lyan

dDecem

ber

2002,resultingin

noob

servationsfrom

thewinteran

dsummer

season

forthis

year.2.

Discount

calculatedas

(Mkt.price

-PDSprice)/Mkt.price*100

3.Averagesbased

onPDSan

dmarket

pricesreportedby

PDSusers

inthesample.

47

Figure

2:Variation

inPDSrice

discount

(2002-2008)

Pan

elE

020406080 020406080 Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

006

Pan

elF

020406080 020406080 Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Winter

Summer Monso

onPos

t Mon

soon

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

007

Pan

elG

020406080 020406080 Winter

Summer

Monso

onWint

er

Summer

Monso

onWint

er

Summer

Monso

onWint

er

Summer

Monso

on

Assa

mW

est B

enga

lJh

arkh

and

Oris

sa

Chat

tisga

rhAn

dhra

Pra

desh

Karn

atak

aKe

rala

Rice discount ( % of Mkt. price )

Year

= 2

008

Notes:1.

Rou

nd63

surveyed

hou

seholdsbetweenJa

nuaryan

dJu

ne2008,resultingin

noob

servationsfrom

thepostmon

soon

season

forthis

year.2.

Discount

calculatedas

(Mkt.price

-PDSprice)/Mkt.price*100

3.Averagesbased

onPDSan

dmarketprices

reportedby

PDSusers

inthesample.

48

Table 1: State Specific PDS Quotas for BPL households

State Rice (kg) Wheat (kg)

Andhra Pradesh 4 per person (20 hh max) 5 (at APL price*)Assam 20 0Bihar 15 15Chattisgarh 25 0Gujarat 1 per person (3.5 hh max) 1.5 per person (9 hh max)Haryana 10 25Himachal Pradesh 15 20Jharkhand 35 0Karnataka 16 4Kerala 8 per adult 4 per child (20 hh max) 5 (at APL price*)Madhya Pradesh 6 17Maharashtra 5 15Meghalaya 2 per person 0Orissa 16 0Punjab 10 25Rajasthan 5 25Uttar Pradesh 20 15West Bengal 2 per person 2 per person

Sources: Planning Commission (2005), Khera (2011b) and Simplifying the food security bill at

http : //bit.ly/PMNFSB. Notes: 1. * denotes all households pay Above Poverty Line prices i.e. no price

discount. 2. PDS entitlements are for the period under study: 2002-2008. 3. Tamil Nadu is excluded as it

follows a universal PDS.

49

Tab

le2:

Variation

inPDSRicediscount

(%)

Winter(J

anuary

-March)

Summer

(April-M

ay)

State

Mean

Std.Dev

p10

p50

p90

NMean

Std.Dev

p10

p50

p90

N

Assam

41.46

10.59

2040.04

56.45

233

41.78

11.11

6.17

41.82

54.17

163

WestBengal

47.79

1023.51

47.94

60.32

445

48.56

11.28

11.12

50.77

61.18

320

Jharkh

and

54.75

13.57

20.45

51.52

72.73

7356.62

11.6

2062.86

7068

Orissa

42.88

11.76

18.33

44.12

59.42

544

44.58

12.01

13.33

44.57

57.92

427

Chattisgarh

45.15

13.83

17.65

41.04

67.09

265

49.14

17.95

1649.53

73.33

178

Andhra

Pradesh

53.36

5.52

39.91

53.52

60.24

2952

56.13

9.96

37.39

54.53

72.83

2134

Karnataka

66.92

11.6

23.39

69.82

77.32

1239

66.86

11.63

35.56

69.67

78.18

791

Kerala

45.16

10.17

27.39

42.8

59.89

970

47.03

12.42

18.18

45.82

63.74

620

Monso

on

(June-Sep

tember)

Post

Monso

on

(October-D

ecem

ber)

State

Mean

Std.Dev

p10

p50

p90

NMean

Std.Dev

p10

p50

p90

N

Assam

38.1

13.92

4.55

35.48

60283

39.44

13.41

17.65

36.93

59.13

230

WestBengal

43.29

11.9

7.82

43.56

57.19

618

43.35

12.77

6.41

43.13

58.5

531

Jharkh

and

52.73

17.17

6.25

54.83

7576

54.47

14.22

16.67

54.69

71.33

52Orissa

40.63

10.11

16.59

40.42

53.48

693

40.3

10.9

16.67

3955.48

623

Chattisgarh

42.48

12.94

22.5

41.98

63.69

398

41.21

13.67

11.33

39.48

58.43

324

Andhra

Pradesh

52.94

8.01

36.13

53.05

58.99

3952

52.4

5.53

40.98

52.15

60.23

3307

Karnataka

58.22

15.45

24.38

62.13

75.07

1697

60.2

15.04

14.67

65.5

751335

Kerala

42.77

10.55

18.7

42.32

58.22

1452

41.65

10.71

18.26

39.27

58.66

1217

Sou

rce:

Calculation

susing2002-2008NSSO

Socio-E

conom

icSurveys.

Notes:1.

Discount

calculatedas

(Mkt.price

-PDSprice)/Mkt.price*100

2.Averagesbased

onPDSan

dmarketpricesreportedby

PDSusers

inthesample.

50

Tab

le3:

Variation

intheper

capitavalueof

thePDSRiceSubsidy(R

s)

Winter(J

anuary

-March)

Summer

(April-M

ay)

State

Mean

Std.Dev

p10

p50

p90

NMean

Std.Dev

p10

p50

p90

N

Assam

23.52

11.11

8.64

20.94

38.67

233

26.56

14.4

3.86

23.47

43.42

163

WestBengal

11.44

3.27

4.62

11.01

16.13

445

11.9

3.37

2.21

1215.63

320

Jharkh

and

52.11

29.69

14.35

44.38

8473

60.51

38.41

15.12

53.2

9868

Orissa

18.89

12.55

4.14

15.98

31.97

544

18.95

11.43

416.48

29.99

427

Chattisgarh

33.15

27.58

8.28

25.55

52.71

265

35.66

29.53

5.19

29.39

61.99

178

Andhra

Pradesh

25.36

6.23

12.21

25.18

33.44

2952

26.95

7.81

11.94

25.96

36.35

2134

Karnataka

35.89

21.56

6.93

31.74

58.24

1239

37.93

25.26

8.32

31.49

63.76

791

Kerala

31.11

14.14

7.27

29.72

49.81

970

33.07

15.31

7.8

29.86

53.1

620

Monso

on

(June-Sep

tember)

Post

Monso

on

(October-D

ecem

ber)

State

Mean

Std.Dev

p10

p50

p90

NMean

Std.Dev

p10

p50

p90

N

Assam

24.38

13.98

2.7

2241.81

283

26.12

18.65

7.2

21.89

41.63

230

WestBengal

10.76

3.55

210.55

14.74

618

10.88

3.83

1.52

10.69

15.61

531

Jharkh

and

45.45

29.08

7.17

38.04

7776

54.03

31.11

7.7

45.53

79.8

52Orissa

17.07

9.69

4.34

15.18

28693

17.94

11.33

4.85

14.87

28.8

623

Chattisgarh

32.51

23.28

7.02

24.72

59.79

398

31.2

26.64

3.73

24.55

54.93

324

Andhra

Pradesh

25.87

6.99

12.09

25.27

34.72

3952

26.1

6.55

12.43

25.6

34.94

3307

Karnataka

32.62

20.37

6.63

28.71

55.37

1697

33.71

21.71

4.79

29.1

56.25

1335

Kerala

29.89

13.48

7.7

28.28

48.89

1452

29.12

12.81

6.34

27.64

46.27

1217

Sou

rce:

Calculation

susing2002-2008NSSO

Socio-E

conom

icSurveys.

Notes:1.

Valueof

thesubsidycalculatedas

Per

capitaQuota*(M

kt.price

-PDSprice)2.

Averagesbased

onPDSan

dmarketprices

reportedby

PDSusers

inthesample.

51

Table 4: Descriptive statistics for the full sample and PDS users

Sample: Full Sample PDS users

Mean Standard Mean Standarddeviation deviation

Monthly expenditure per capita (Rs) 1011.0 (1085.3) 636.8 (393.4)Daily calories per capita (kcal) 2334.2 (1300.9) 2190.9 (623.7)Proportion spent on food 0.547 (0.142) 0.577 (0.116)

Size of the household 4.570 (2.382) 4.736 (1.891)Number of children below 15 1.411 (1.428) 1.538 (1.339)Proportion of women 0.515 (0.207) 0.512 (0.152)Age of household head 46.58 (13.61) 45.38 (12.17)

Urban dummy 0.363 (0.481) 0.219 (0.414)Electricity 0.724 (0.447) 0.725 (0.446)Permanent home 0.338 (0.473) 0.314 (0.464)SC/ST/OBC 0.592 (0.491) 0.765 (0.424)

PDS rice price (Rs/kg) 5.271 (1.855)Mkt rice price (Rs/kg) 10.80 (2.193)

PDS rice qty (kg) 18.64 (9.392)Market rice qty (kg) 26.11 (20.24)

Food expenditure per capita (Rs) 400.5 (168.9)Cereal expenditure per capita (Rs) 116.0 (43.85)Rice subsidy per capita (Rs) 25.71 (11.90)

Rice proportion of food expenditure 0.260 (0.130)

Proportion of calories from rice 0.615 (0.175)Proportion of calories from cereals 0.727 (0.0987)

Observations 124228 22564

Notes: 1. Rural Poverty line is Rs 497.6, Urban Poverty line is Rs 635.7 (Planning Commission,

Government of India). 2. Average daily minimum calorie requirements are 2400 kcal for rural

and 2100 kcal for urban areas. 3. All prices in 2005 Rupees. (Rs 45.3 = 1 USD in 2005). 4. The

category cereals includes rice, wheat, semolina, jowar, bajra, millets, corn, barley and ragi. 5. The

sample comprises households in the following states: Andhra Pradesh, Assam, Karnataka, Kerala,

Orissa, Jharkhand, Chattisgarh and West Bengal.

52

Table 5: Impact of the subsidy on cereal consumption

Dependent variable: Cereal consumption Log cereal consumption Cereal consumption(1) (2) (3)

Rice subsidy per capita 2.030∗∗∗

(0.158)

Log rice subsidy per capita 0.123∗∗∗

(0.00963)

Rice quota per capita -1.968(4.442)

Market price* quota per capita 1.697∗∗∗

(0.436)

PDS price* quota per capita 0.157(0.463)

PDS price -1.607(2.849)

Market price -6.131∗∗

(2.395)

Observations 22564 22564 22564Adjusted R2 0.250 0.270 0.258

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head, proportion

of females, land owned) and urban, state*year, district and season dummies. 3. Dependent variable in

columns (1) and (3) is daily cereal consumption per capita (grams), dependent variable in column (2)

is log of daily cereal consumption per capita.

53

Tab

le6:

Impactof

thesubsidyon

calories

from

differentfood

grou

ps

Dep

endentvariab

le:

Caloric

intake

Log

caloricintake

Log

caloricintake

from

food

grou

pCereals

Lentils

Fruits&

Vegetab

les

Meat

(1)

(2)

(3)

(4)

(5)

(6)

Ricesubsidyper

capita

12.58∗

∗∗

(1.116)

Log

rice

subsidyper

capita

0.144∗

∗∗0.123∗

∗∗0.154∗

∗∗0.234∗

∗∗0.170∗

∗∗

(0.0103)

(0.00963)

(0.0177)

(0.0160)

(0.0187)

Observations

22564

22564

22564

22118

22562

19833

Adjusted

R2

0.124

0.166

0.270

0.215

0.441

0.426

Standarderrors

inparentheses.

∗p<

0.10,∗∗

p<

0.05,∗∗

∗p<

0.01

Notes:1.

Allequationspresent

resultsclustered

atthedistrictlevel.2.

Allequationsincludehou

seholdcharacteristics(education

ofhou

sehold

headan

dspou

se,agean

dagesquared

ofhou

seholdhead,proportion

offemales

landow

ned)an

durban

,state*year,districtan

dseason

dummies.

3.Dep

endentvariab

lein

column(1)is

daily

caloricintake

per

capita(kcal),dep

endentvariab

lein

column(2)is

logof

daily

caloricintake

per

capita,

dep

endentvariab

lein

column(3)is

logof

calories

from

cerealsper

capita,

dep

endentvariab

lein

column(4)is

logof

calories

from

lentilsper

capita,

dep

endentvariab

lein

column(5)is

logof

calories

from

fruitsan

dvegetablesper

capita,

anddep

endentvariab

lein

column(6)is

logof

calories

from

meatper

capita.

54

Tab

le7:

Exp

enditure

elasticity

ofcalories

andfood

expenditure:Ruralsample

Dep

endentvariab

le:

Log

caloricintake

Log

food

expenditure

Log

Rupeesper

calorie

(1)

(2)

(3)

(4)

(5)

(6)

Log

mon

thly

expenditure

per

capita

0.406∗

∗∗0.751∗

∗∗0.345∗

∗∗

(0.0103)

(0.00883)

(0.0102)

Log

rice

subsidyper

capita

0.140∗

∗∗0.146∗

∗∗0.00575

(0.0135)

(0.0153)

(0.00632)

Observations

13333

13333

13333

13333

13333

13333

Adjusted

R2

0.437

0.157

0.820

0.404

0.675

0.527

Standarderrors

inparentheses.

∗p<

0.10,∗∗

p<

0.05,∗∗

∗p<

0.01

Notes:1.

Allequationspresent

resultsclustered

atthedistrictlevel.2.

Allequationsincludehou

seholdcharacteristics

(education

ofhou

seholdheadan

dspou

se,agean

dagesquared

ofhou

seholdhead,proportion

offemales,landow

ned)an

d

state*year,districtan

dseason

dummies.

3.Dep

endentvariab

lein

columns(1)an

d(2)is

logof

daily

caloricintake

per

capita,

dep

endentvariab

lein

columns(3)an

d(4)is

logof

food

expenditure

per

capita,

dep

endentvariab

lein

columns(5)an

d(6)is

logof

price

paidper

1000

calories.5.

Asin

Subraman

ianan

dDeaton(1996),thesample

comprises

only

ruralhou

seholdsan

dmon

thly

expenditure

iscalculatednet

ofpurchases

ofdurable

good

s.

55

Table 8: Income elasticity of cereal and food consumption

Panel A: IHDS data

Dependent variable: Log cereal consumption Log food expenditure

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

Log monthly income per capita 0.0462∗∗∗ 0.0304∗∗∗ 0.0966∗∗∗ 0.0827∗∗∗

(0.00972) (0.00959) (0.0108) (0.0105)

Log rice subsidy per capita 0.295∗∗∗ 0.259∗∗∗

(0.0323) (0.0323)

Observations 3962 3962 3962 3962Adjusted R2 0.306 0.354 0.402 0.429

Panel B: IHDS and NSSO data

Dependent variable: Log cereal consumptionData: IHDS NSSO IHDS NSSO

Log monthly expenditure per capita 0.247∗∗∗ 0.269∗∗∗

(0.0178) (0.0264)

Log rice subsidy per capita 0.320∗∗∗ 0.179∗∗∗

(0.0314) (0.0390)

Observations 3962 4255 3962 4255Adjusted R2 0.388 0.357 0.358 0.286

Standard errors in parentheses.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations in panel A include

urban, district and season dummies. All equations in panel B include household characteristics (education

of household head and spouse, age and age squared of household head, proportion of females, land owned)

and urban, district and season dummies. 3. Dependent variable in columns (1) and (2) of panel A and

columns (1)-(4) of panel B is log of cereal consumption per capita, dependent variable in columns (3) and

(4) of panel A is log of food expenditure per capita. 4. The data come from the India Human Development

Survey 2004-05 and the NSSO surveys covering November 2004 - October 2005.

56

Tab

le9:

Impactof

thesubsidyon

cereal

consumption

andcaloricintake:Heterogeneity

bysocial

grou

p

Dep

endentvariab

le:

Cerealconsumption

per

capita

Caloric

intake

per

capita

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Ricesubsidyper

capita

1.979∗

∗∗1.892∗

∗∗1.830∗

∗∗1.960∗

∗∗12.41∗

∗∗11.61∗

∗∗11.55∗

∗∗12.28∗

∗∗

(0.172)

(0.209)

(0.173)

(0.159)

(1.214)

(1.046)

(1.309)

(1.142)

Urban

dummy

-31.61

∗∗∗

-24.97

∗∗∗

-30.90

∗∗∗

-24.86

∗∗∗

-82.32

∗∗∗

-63.93

∗∗∗

-104.0

∗∗∗

-59.30

∗∗∗

(6.412)

(3.026)

(3.084)

(3.030)

(27.98)

(13.92)

(13.52)

(13.91)

Urban

*Ricesubsidy

0.240

0.772

(0.241)

(1.054)

SC/S

T/O

BC

1.488

-76.34

∗∗∗

(5.767)

(27.20)

SC/S

T/O

BC*R

icesubsidy

0.186

1.242

(0.205)

(1.001)

Low

estexpenditure

quartile

-52.27

∗∗∗

-334.8

∗∗∗

(6.058)

(29.14)

Low

estqu

artile*R

icesubsidy

-0.106

-2.961

∗∗

(0.205)

(1.169)

Hom

egrow

nrice

-15.73

∗-41.21

(8.247)

(39.68)

Hom

egrow

n*R

icesubsidy

1.300∗

∗∗5.893∗

∗∗

(0.328)

(1.609)

Observations

22564

22564

22564

22564

22564

22564

22564

22564

Adjusted

R2

0.250

0.250

0.274

0.251

0.124

0.124

0.183

0.126

Standarderrors

inparentheses.

∗p<

0.10,∗∗

p<

0.05,∗∗

∗p<

0.01

Notes:1.

Allequationspresent

resultsclustered

atthedistrictlevel.2.

Allequationsincludehou

seholdcharacteristics(education

ofhou

sehold

headan

dspou

se,agean

dagesquared

ofhou

seholdhead,proportion

offemales,landow

ned)an

durban

,state*year,districtan

dseason

dummies.

3.Dep

endentvariab

lein

columns(1)-(4)is

daily

cereal

consumption

per

capita(grams),dep

endentvariab

lein

columns(5)-(8)is

daily

caloricintake

per

capita(kcal).

57

Table 10: Impact of the subsidy on rice producing households

Dependent variable: Cereal consumption Caloric intake

(1) (2)

Rice quota per capita -14.15 -38.18(18.56) (84.41)

Market price* quota per capita 4.483∗∗ 19.52∗∗

(1.792) (7.896)

PDS price* quota per capita -0.989 -5.916(1.976) (8.667)

PDS price 3.226 38.62(10.98) (49.72)

Market price -17.15∗ -52.60(8.760) (40.03)

Observations 2111 2111Adjusted R2 0.231 0.241

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. The equation present results clustered at the district level. 2. The equation includes household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state*year, district and season dummies. 3. Dependent

variable in column (1) is daily cereal consumption per capita (grams), dependent variable in column (2)

is daily caloric intake per capita (kcal). 4. The sample comprises PDS households that report

own (produced) rice as one of their sources of supplementary grains.

58

Table 11: Elasticity of cereal consumption and caloric intake by expenditure quartile

Panel A: Cereal consumption

Dependent variable: Log cereal consumption per capitaExpenditure quartile: Lowest Highest

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

Log rice subsidy per capita 0.0689∗∗∗ 0.0895∗∗∗ 0.109∗∗∗ 0.175∗∗∗

(0.0138) (0.0147) (0.0137) (0.0180)

Observations 5677 5709 5696 5482Adjusted R2 0.320 0.362 0.330 0.272

Panel B: Caloric intake

Dependent variable: Log caloric intake per capitaExpenditure quartile: Lowest Highest

Log rice subsidy per capita 0.0682∗∗∗ 0.0933∗∗∗ 0.111∗∗∗ 0.196∗∗∗

(0.0135) (0.0132) (0.0129) (0.0182)

Observations 5677 5709 5696 5482Adjusted R2 0.251 0.294 0.279 0.177

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state*year, district and season dummies. 3. In going

from columns (1) to (4), the sample comprises households in the lowest, second lowest, second

highest and highest expenditure quartile of PDS rice users, respectively. 4. Dependent variable in

panel A is log of cereal consumption per capita, dependent variable in panel B is log of daily

caloric intake per capita.

59

Table 12: Impact on Non-PDS users

Dependent variable: Cereal cons. Log cereal cons. Caloric intake Log caloric intake(1) (2) (3) (4)

Rice subsidy per capita -0.0706 0.462(0.115) (0.772)

Log rice subsidy per capita -0.0141∗ -0.00863(0.00750) (0.00686)

Observations 26494 26494 26494 26494Adjusted R2 0.256 0.261 0.045 0.152

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state*year, district and season dummies. 3. Dependent

variable in column (1) is daily cereal consumption per capita (grams), dependent variable in

column (2) is log of daily cereal consumption per capita, dependent variable in column (3) is daily caloric

intake per capita (kcal), dependent variable in column (4) is log of daily caloric intake per capita.

4. The sample comprises households that do not receive any subsidies from the PDS. They are assigned

their respective local average value of the PDS subsidy.

60

Table 13: Alternative specifications for value of the subsidy

Dependent variable: Log caloric intake(1) (2) (3) (4) (5)

Log rice subsidy (avg. family size) 0.134∗∗∗

(0.00763)

Log rice subsidy (household level) 0.131∗∗∗

(0.0147)

Size of the household 0.134∗∗∗

(0.00243)

Log rice subsidy (per person) 0.202∗∗∗

(0.00992)

Log rice subsidy (median prices) 0.119∗∗∗

(0.0103)

Log rice subsidy per capita 0.148∗∗∗

(state*survey wave) (0.0107)

Observations 22564 22564 22564 22543 22564Adjusted R2 0.173 0.613 0.200 0.159 0.168

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include

household characteristics (education of household head and spouse, age and age squared of household

head, proportion of females land owned) and urban and district dummies. Equations (1)-(4) include

state*year and season dummies. Equation (5) includes state*survey-wave dummies. 3. Dependent

variable in columns (1), (4) and (5) is log of daily caloric intake per capita, dependent variable

in column (2) is log of daily caloric intake at the household level, dependent variable in

column (3) is log of daily caloric intake at the household level divided by the total number

of household members.

61

Table 14: Impact on different sub-samples of states

Dependent variable: Log cereal consumption Log caloric intakeStates: All Rice Non Rice All Rice Non Rice

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

Log rice subsidy per capita 0.0796∗∗∗ 0.123∗∗∗ 0.0439∗∗∗ 0.101∗∗∗ 0.144∗∗∗ 0.0666∗∗∗

(0.00677) (0.00963) (0.00664) (0.00701) (0.0103) (0.00675)

Observations 33231 22564 10667 33231 22564 10667Adjusted R2 0.263 0.270 0.255 0.197 0.166 0.255

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state*year, district and season dummies. 3. Dependent

variable in columns (1) - (3) is log of daily cereal consumption per capita, dependent variable in columns

(4)-(6) is log of daily caloric intake per capita. 4. The rice favoring states are: Andhra Pradesh, Assam,

Chattisgarh, Jharkhand, Karnataka, Kerela, Orissa and West Bengal. The non-rice favoring states are:

Bihar, Gujarat, Haryana, Himachal Pradesh, Madhya Pradesh, Maharashtra, Punjab, Rajasthan

and Uttar Pradesh.

Table 15: Impact of the wheat subsidy

Dependent variable: Log cereal consumption Log caloric intake(1) (2)

Log rice subsidy per capita 0.138∗∗∗ 0.156∗∗∗

(0.0110) (0.0125)

Log wheat subsidy per capita 0.0172∗∗∗ 0.0270∗∗∗

(0.00549) (0.00519)

Observations 12235 12235Adjusted R2 0.275 0.193

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level. 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state*year, district and season dummies. 3. Dependent

variable in column (1) is log of daily cereal consumption per capita, dependent variable in

column is log of daily caloric intake per capita.

62

Table 16: Impact of the subsidy on cereal consumption and caloric intake: State levelfunctioning

Dependent variable: Cereal consumption per capita Caloric intake per capita(1) (2)

Rice subsidy per capita 2.359∗∗∗ 12.04∗∗∗

(0.180) (1.148)

Corrupt*Rice subsidy -1.207∗∗∗ -7.060∗∗∗

(0.304) (1.480)

Observations 22564 22564Adjusted R2 0.251 0.117

Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Notes: 1. All equations present results clustered at the district level 2. All equations include household

characteristics (education of household head and spouse, age and age squared of household head,

proportion of females, land owned) and urban, state-year, district and season dummies. 3. Dependent

variable in column (1) is daily cereal consumption per capita (grams), dependent variable in column (2)

is daily caloric intake per capita (kcal).

63

Appendix Tables

A. Per capita purchase of PDS Rice

Mean (kg) Std. Dev. N

Assam 7.29 3.5 909

West Bengal 2.82 2.46 1914

Orissa 6.60 3.63 2287

Jharkhand 4.74 2.41 269

Chattisgarh 9.09 4.07 1169

Andhra Pradesh 5.07 2.50 12345

Karnataka 5.05 2.49 5062

Kerala 5.50 3.49 4259

Full Sample 5.34 3.1 28214

Note: The sample comprises PDS rice users.

B. Price paid per calorie by food group

Foodgroup: Cereals Lentils Vegetables Meat

Mean (Rs/1000 kcal) 2.40 8.46 18.89 41.45Std. Dev. 0.63 2.80 6.48 19.42

Note: Price calculations based on quantity and value reported by PDS rice users.

1

C. BPL and Antyodaya Households - IHDS and NSSO 61st round data

Sample: BPL AntyodayaData: IHDS NSSO IHDS NSSO

Monthly expenditure per capita (Rs) 660.3 558.5 485.3 478.9(539.4) (306.4) (481.2) (309.5)

Monthly income per capita (Rs) 606.5 471.0(785.6) (370.1)

PDS rice price (Rs/kg) 4.475 5.399 3.329 3.461(1.493) (1.290) (0.954) (0.983)

Market rice price (Rs/kg) 10.54 10.53 9.753 10.09(2.227) (2.016) (1.634) (2.049)

PDS rice Qty (kg) 18.60 17.95 24.71 23.73(6.766) (8.513) (9.438) (10.81)

Market rice Qty (kg) 23.08 25.29 25.67 16.68(22.63) (20.88) (26.31) (20.38)

Daily cereal consumption per capita (kg) 0.441 0.467 0.538 0.496(0.170) (0.136) (0.195) (0.186)

Rice subsidy per capita (Rs) 27.23 24.97 32.47 30.85(15.42) (14.29) (23.30) (23.90)

Observations 4196 7405 283 788

Notes: 1. Means with standard deviations in parentheses. 2. The sample comprises PDS rice users

from the 61st round (2004-05) of the NSSO surveys and the India Human Development Survey 2005.

2

D. PDS Rice users - IHDS and NSSO

Data: IHDS NSSO

Monthly expenditure per capita (Rs) 625.1 596.7(442.2) (362.4)

Monthly income per capita (Rs) 587.5(567.3)

PDS rice price (Rs/kg) 4.546 5.237(1.612) (1.627)

Market rice price (Rs/kg) 10.48 10.56(1.906) (2.122)

PDS rice qty (kg) 19.16 18.54(7.253) (8.785)

Market rice qty (kg) 24.52 27.97(23.03) (21.37)

Daily cereal consumption per capita (kg) 0.434 0.475(0.158) (0.129)

Rice subsidy per capita (Rs) 25.46 24.17(11.18) (11.37)

Observations 3962 4255

Notes: 1. Means with standard deviations in parentheses. 2. The sample comprises PDS rice users

(November 2004 to October 2005) from the NSSO surveys and the India Human Development Survey 2005.

3

Removing  Observation  Bias:  using  “big  data”  to  create  cooking  metrics  from  stove  usage  monitors  and  minimal  observation  

 Andrew  M.  Simons,  Theresa  Beltramo,  Garrick  Blalock,  David  I.  Levine*  

 "Big   data"   holds   the   promise   of   improving   social   science   research.   One  application  of  big  data   is   in  understanding  the  adoption  behaviors  of  new  cooking   technologies   in   the   developing   world.   Stove   Usage  Monitors,   for  example,   can   generate   millions   of   observations   with   lower   Hawthorne  effects  than  traditional  surveys  or  direct  observation.  This  paper  describes  field  methods   and   proposes   a   new   algorithm   for   converting   temperature  data   generated   from   stove   use   monitors   into   usage   metrics   for   both  traditional  and  nontraditional  stoves.  Central  to  our  technique  is  recording  the  visual  on/off  status  of  a  stove  anytime  research  staff  observes  the  stove.  The   observations   are   then   combined   with   temperature   readings   and   a  predictive   logistic   regression   to   predict   the   probability   a   stove   is   in   use.  Using   this   algorithm   we   correctly   predict   89%   of   three   stone   fire  observations   and   94%   of   Envirofit   observations.   The   logistic   regression  correctly  classifies  more  observations  than  published  temperature  analysis  algorithms.   Further,   we   analyze   stove   usage   data   when   multi-­‐person  measurement  teams  are  present  in  households  compared  to  the  same  times  in  the  previous  and  subsequent  weeks  when  only  the  unobtrusive  sensors  are   collecting   temperature   data.  We   find   a   statistically   and   economically  significant  Hawthorne  effect  whereby  study  participants  increase  the  use  of  the  introduced  stoves  by  approximately  one  third  (2  hours  per  day),  while  simultaneously   decreasing   three   stone   fire   usage   by   approximately   the  same  amount  when  measurement  teams  are  present.  

 Keywords:   observation   bias,   Hawthorne   effect,   improved   cookstoves,   technology  adoption,  biomass  fuel,  monitoring  and  evaluation,  stove  use  monitors    *  Simons:  Charles  H.  Dyson  School  of  Applied  Economics  and  Management,  Cornell  University  (email:  ams727@cornell.edu).  Beltramo:   Impact  Carbon,   San  Francisco   (email:   tbeltramo@impactcarbon.org).  Blalock:  Charles  H.  Dyson  School  of  Applied  Economics   and   Management,   Cornell   University   (email:   garrick.blalock@cornell.edu).   Levine:   Haas   School   of   Business,  University  of  California,  Berkeley  (email:  levine@haas.berkeley.edu).    

Acknowledgments:   This   study   was   funded   by   the   United   States   Agency   for   International  Development   under   Translating   Research   into   Action,   Cooperative   Agreement   No.   GHS-­‐A-­‐00-­‐09-­‐00015-­‐00.   The   recipient   of   the   grant   was   Impact   Carbon   who   co-­‐funded   and   managed   the  project.  Juliet   Kyaesimira   and   Stephen   Harrell   expertly   oversaw   field   operations   and   Amy   Gu  provided   excellent   research   support.  We   thank   Impact   Carbon   partners  Matt   Evans,   Evan  Haigler,  Jimmy   Tran,   Caitlyn   Toombs,   and   Johanna   Young;   Berkeley   Air   project   partners   including   Dana  Charron,   Dr.  David  Pennise,   Dr.   Michael   Johnson,   and   Erin   Milner,   the   USAID   TRAction   Technical  Advisory   Group,   and   seminar   participants   at   Cornell   University   for   valuable   comments.   Special  thanks   to   Kirk   Smith,   Ilse   Ruiz-­‐Mercado,   Ajay   Pillarisetti   and   colleagues   from   the   U.C.   Berkeley  Household   Energy,   Climate,   and   Health   Research   Group   for   the   development   of   the   Stove   Use  Monitoring  system  (SUMs)  concept  and  consultations  on  optimal  SUMs  placement  and  holder  design  for   this   project.   Data   collection   was   carried   out   by   the   Center   for   Integrated   Research   and  Community   Development   (CIRCODU),   and   the   project’s   success   relied   on   expert   oversight   by  CIRCODU's   Director   General   Joseph  Ndemere   Arineitwe   and   field   supervisors  Moreen   Akankunda,  

  2  

Innocent   Byaruhanga,   Fred   Isabirye,   Noah   Kirabo,   and   Michael   Mukembo.  This   study   was   made  possible  by  the  support  of   the  American  people  through  the  United  States  Agency   for   International  Development  (USAID).  We  thank  the  Atkinson  Center  for  a  Sustainable  Future  at  Cornell  University  for  additional  funding  of  related  expenses.  The  findings  of  this  study  are  the  sole  responsibility  of  the  authors,   and  do  not  necessarily   reflect   the   views  of   their   respective   institutions,   nor  USAID  or   the  United  States  Government.    Introduction  “Big   data”   is   transforming   economic   policy   and   the   process   of   economic   research  (Einav  and  Levin  2014).  Prominent  examples   include  using   internet  search  results  to  provide  real-­‐time  estimates  of  economic  activity  (Choi  and  Varian  2012;  Goel  et  al.   2010),   analyzing   administrative   records   to   evaluate   the   long-­‐term   effects   of  better  teachers  and  the  effects  of  providing  health  insurance  (Chetty,  Friedman,  and  Rockoff  2011;  Finkelstein  et  al.  2012),  and  utilizing  shopping  data  from  millions  of  online  shoppers   to  understand   their   sensitivity   to   in-­‐state   sales   taxes   (Einav  et  al.  2012).      The  developing  world  also  has  opportunities   for  the  analysis  of  “big  data.”  We  use  millions  of  data  points  collected  by  non-­‐invasive,  relatively  inexpensive  temperature  monitors  to  create  metrics  of  stove  use  in  a  developing  country  setting.  Traditional  wood-­‐and   charcoal-­‐burning   stoves   both   burn   inefficiently   and   produce   a   great  amount  of   smoke.  The  smoke   leads   to   respiratory,  heart,   and  other  disorders   that  kill  approximately  4.0  million  people  per  year  (Lim  et  al.  2012).  Much  of  the  health  burden   is   concentrated   on   women   and   children,   as   is   the   time   burden   collecting  biomass   fuel   (Kammen,   Bailis,   and   Herzog   2002).   Environmental   damage   can   be  significant;  by  one  estimate,  household  energy  use  in  Africa  will  produce  6.7  billion  tons   of   carbon   by   2050   (Bailis,   Ezzati,   &   Kammen,   2005).   Furthermore,   the  incomplete  combustion  of  biomass  fuels  leads  to  the  release  of  black  carbon  (soot)  that  contributes  to  current  global  warming  (Bond,  Venkataraman,  and  Masera  2004;  Ramanathan  and  Carmichael  2008).  These  inefficiencies  imply  deeper  poverty,  loss  of   biodiversity,   rapid   deforestation,   and   climate   change.   These   problems   have   led  both   policy-­‐makers   and   practitioners   to   search   for   safer   and   more   fuel-­‐efficient  stoves  that  are  attractive  to  consumers.      One   challenge   to   understanding   stove   use   substitution   behaviors   is   to   measure  stove  use  precisely,  economically  and  with  minimal  intrusion  into  the  lives  of  study  participants.  Measuring  stove  usage  with  minimal  observation  bias   is  necessary   to  recover   unbiased   estimates   of   stove   use   because   study   participants   potentially  behave   differently   when   research   staff   is   present.   Scientists   studying   the   health  effects  of  cooking  technologies  need  to  understand  how  many  hours  a  day  a  stove  is  used   to   calculate   exposure   to   household   air   pollution   (Martin   II   et   al.   2011;  McCracken  et  al.  2007).  Project  financiers  need  to  measure  the  time  cooked  on  both  new  and  old   stoves   to  allocate   the   carbon  credits   that   can   fund  stove   subsidies   in  developing   countries   (Gold   Standard   Foundation   2013;   GIZ   2011;   Freeman   and  Zerriffi  2012).  Further,  “stove  stacking”  (the  use  of  multiple  fuels  and  stoves)  often  occurs  when  multiple  cooking  options  exist  (Masera,  Saatkamp,  and  Kammen  2000;  Ruiz-­‐Mercado   et   al.   2013).   To   understand   any   of   these   issues   researchers   must  

  3  

know  the  baseline  amount  of  cooking  on  existing  stoves,  and  then  the  time  cooked  on  both  the  existing  and  introduced  stove.    Pioneering  stove  studies  have  highlighted  the  harmful  effects  of  indoor  air  pollution  and  the  health  benefits  of  adopting  fuel-­‐efficient  nontraditional  stoves  (Ezzati,  Saleh,  and  Kammen  2000;  Ezzati  and  Kammen  2002;  Ezzati  and  Kammen  2001;  Smith  et  al.  2010;  Smith  et  al.  2011;  Smith  et  al.  2006;  Smith-­‐Sivertsen  et  al.  2009).  However  these   studies   were   characterized   by   high   levels   of   interactions   between   research  staff  and  study  participants;  for  example,  in  Ezzati  et  al.  (2000)  research  staff  noted  the   locations  of  each  household  member  and   the   status  of  how  active   the  cooking  fire  is  every  5-­‐10  minutes.  Observational  studies  of  stove  use  raise  the  possibility  of  observation   bias   (or   Hawthorne   effect),   which   arises   when   participants   act  differently   than   they  normally  would  because   they  know   they   are  being  observed  (as  noted  by  Smith-­‐Sivertsen  et  al.,  2009).    To  minimize  observation  bias,  stove  use  monitor  systems  (SUMs)  can  record  stove  temperatures  without   the  need   for   an  observer   to  be  present.   The   SUMs  used   for  our  project,   iButtons™  manufactured  by  Maxim  Integrated  Products,  Inc.,  are  small  stainless  steel  temperature  sensors  about  the  size  of  a  small  coin  and  the  thickness  of   a   watch   battery   which   can   be   affixed   to   any   stove   type.   Our   SUMs   record  temperatures  with  an  accuracy  of  +/-­‐  1.3  degrees  C  up  to  85°C  (see  Figure  1   for  a  photograph).  1  Stove  Usage  Monitors  that  log  stove  temperatures  was  first  suggested  by   Ruiz-­‐Mercado   et   al.   (2008).   SUMs   offer   an   unobtrusive,   precise,   relatively  inexpensive,  and  objective  measure  of  stove  usage  (Ruiz-­‐Mercado  et  al.  2008).2    We  know  of  two  published  algorithms  for  determining  stove  usage  with  SUMs.  Both  studies   use   SUMs   to   gather   continuous   temperature   data   for   nontraditional  cookstoves,   and   then   calculate   temperature   changes   between   subsequent  temperature   readings.   In   Ruiz-­‐Mercado,   Canuz,   and   Smith   (2012)   the   algorithm  considers   an   increase   in   temperature   above   a   certain   threshold   (1.52°C   increase  over  40  minutes)  as  the  start  of  a  candidate  cooking  or  refueling  event.  If  the  post-­‐increase   stove   temperature   is   also   above   the   ambient   temperature,   the   algorithm  counts   the   passage   of   time   until   a   temperature   drop   indicates   that   the   stove   is  cooling  down  (indicated  by  a  2.28°C  decrease  over  60  minutes),  signaling  the  end  of  the   cooking   event.   These   slope   thresholds   are   chosen   by   taking   the   99th   and   1st  percentiles  of  slope  values  from  the  distribution  of  ambient  air  temperatures  in  the  research  area.  Multiple  temperature  peaks  within  a  2  hour  period  of  each  other  are  considered   the   same   cooking   event   (Ruiz-­‐Mercado,   Canuz,   and   Smith   2012).  Mukhopadhyay  et  al.  (2012)  use  a  similar  technique  combining  both  change  in  slope  and  a   temperature   threshold   to   classify  when  a   stove   is  being  used.  However,   the  

                                                                                                               1 For   additional   details   concerning   the   SUMs   see   the   product   description   website   at:  http://www.berkeleyair.com/products-­‐and-­‐services/instrument-­‐services/78-­‐sums  2  The  SUMs  used  in  our  project  cost  approximately  USD$16  each  and  could  record  temperature  data  for   24   hours   a   day   for   four   to   six  weeks   in   a   household   before   needing  minimal   servicing   from   a  technician  (after  the  data  download  they  can  be  reset  and  re-­‐used).    

  4  

researchers  choose  different   thresholds  based  on  observed   local  cooking  practices  and  local  ambient  temperatures.      While  the  decision  of  what  temperature  thresholds  to  use  is  informed  by  data  from  the  experiment,  it  requires  the  assumption  that  the  slope  thresholds  for  starting  and  ending   cooking   are   the   1st   and   99th   percentiles   of   the   slope   of   the   ambient   air  temperatures.   If   these   slope   thresholds   are   not   the   true   indicator   of   starting   and  ending  cooking,   this   introduces  error   into  the  specification.  Further,  a  challenge  of  the   existing   algorithms   is   that   different   thresholds   are   needed   for   different  geographic  areas  and  potentially  seasons.      We  propose  a  methodology  that  uses  SUMs  temperature  data  gathered  in  the  same  way   as   previous   studies,   but   combines   this   temperature   data   with   observations  made  throughout  the  experiment  of  when  stoves  were  seen  as  on  or  off.  We  run  a  logistic  regression  and  use  SUMs  temperature  data  to  predict  observations  of  stoves  being   on   or   off.   We   then   take   the   coefficients   from   this   equation   to   predict   the  probability   of   cooking   over   the   much   larger   set   of   SUM   temperature   readings  without   any   observational   data.   The   researcher   can   sum   the   predicted   minutes  cooked  in  a  24-­‐hour  period  to  estimate  minutes  of  cooking  that  day.  This  measure  of  cooking  can  then  be  used  as  the  basis  for  other  analysis  such  as  comparing  cooking  behaviors   between   groups   of   households   with   certain   characteristics   or   groups  differentiated  by  experimental  design.      Methods  Our   technique   requires   continuous   SUMs   temperature   data   for   a   given   stove,   and  recorded  instances  of  whether  that  particular  stove  is  seen  in  use  or  not  (henceforth  called  “observations”).  We  matched  observations  of  stove  use  to  SUMs  temperature  data  by  time  and  date  stamps.  The  core  of  our  method  is  a  logistic  regression  using  the   lags   and   leads   of   the   SUMs   temperature   data   to   predict   observations   of   stove  usage.      Continuous  SUMs  temperature  data  A  stove  usage  monitor   takes  a   temperature  reading  at  an   interval  specified  by   the  user.   The   SUMs   used   in   this   experiment   hold   approximately   2,000   data   points  before   the  data  needs   to   be  downloaded   and   the  device   reset.  We   set   the   SUM   to  record   temperature   every   30   minutes,   giving   them   six   weeks   between   required  servicing.      Observations  of  stove  use  Each  time  the  data  collection  team  visited  a  household  they  observed  which  stoves  were  in  use  (along  with  the  date  and  time).  Enumerators  visited  a  house  numerous  times   during   a   “measurement   week,”   when   we   also   enumerated   a   survey   and  weighed   wood   for   a   kitchen   performance   test   (Rob   Bailis,   Smith,   and   Edwards  

  5  

2007).   Another   enumerator   visited   once   every   4-­‐6   weeks   to   download   data   and  reset  the  SUM.3      We  originally  intended  to  use  the  entire  sample  of  observations.  We  found,  however,  a  number  of  anomalies.  For  example,  3.0%  (10  out  of  329)  of   “lit”   three  stone   fire  observations  had  SUM  readings  below  the  typical  daily  mean  temperature  (23.8°C).  At   the  same  time  7.8%  (105  out  of  1339)  of   “not   lit”   three  stone   fire  observations  had   SUM   readings   over   40°C,   which   was   the   highest   ambient   air   temperature  recorded   in   the   observation   period.   While   suspicious,   these   readings   may   not  necessarily  be  erroneous,  as  SUMs  can  be  cool  if  the  stove  was  just  lit  in  the  last  few  minutes  and  the  SUM  has  not  yet  registered  the  heat  increase,  or  a  SUM  could  still  be  hot  on  an  unlit  stove  if  the  stove  was  just  extinguished.      We   suspect   a   substantial   share   of   such   cases  were   due   to   errors   in   observing   or  recording  the  lit  status  of  stoves  or  in  errors  in  matching  the  time  stamps,  stoves,  or  homes  of  observations  and  SUMs  measurements.  Therefore  we  trim  the  observation  sample   based   on   the   hourly   ambient   temperatures   (with   a   conservative  adjustment).  We  drop  observations  observed  as  “lit”  below  the  mean  minus  2.0°C  of  the  ambient  temperature  for  that  hour,  and  we  drop  observations  observed  as  “not  lit”  that  are  above  the  maximum  ambient  temperature  plus  2.0°C  for  that  hour.  The  mean   and   maximum   hourly   ambient   temperatures   are   taken   across   the   entire  sample   of   ambient   temperature   readings.   This   removes   11.0%   (184   out   of   1668)  and   3.6%   (28   out   of   782)   of   the   observation   samples   for   three   stone   fires   and  Envirofits,  respectively.    Regression  specification  There   is   no   theory   of   what   combinations   or   transforms   of   current,   lagged,   and  leading  temperature  readings  should  be  used  to  predict  observed  cooking.  We  test  ten  specifications  and  compare  the  goodness  of  fit  for  each  specification.  (We  do  not  also  include  slope  terms,  because  they  are  perfectly  collinear  with  multiple   lags  or  leads.)  For  simplicity,  our  desire   is  to  use  the  same  specification  on  both  the  three  stone  fire  sample  and  the  Envirofit  sample.    Data  A  series  of   randomized  control   trials  were  executed   in   rural   areas  of   the  Mbarara  District   in   southwestern   Uganda   (Figure   2)   from   February   to   September   2012,  which   focused   on   the   use,   and   adoption   of   fuel-­‐efficient   stoves.4  In   our   study   the                                                                                                                  3  One  caveat  with  observations  used  in  this  study,  is  that  at  times  when  there  are  multiple  stoves  of  a  given  type  (for  example  two  Envirofits  or  two  three  stone  fires)  in  a  household,  the  observation  was  the   answer   to   the   following   question,   “Is   an   [Envirofit   stove/Three   Stone   Fire]   on   in   the   given  household?”   In   cases   where   the   answer   was   “Yes,   one   Envirofit   stove   is   on   and   in   use”   but   two  Envirofits  were  present   in   the  household,   the  data   cleaning  process   required   comparing   the   SUMs  temperature  reading  of  both  Envirofits   that  were   in   the  house  at   the   timestamp  of   the  observation  and  then  assigning  the  “on”  value  to  the  hotter  one,  and  the  “off”  value  to  the  cooler  one.  The  same  process  was  applied  to  three  stone  fires  as  well.  This  process  is  described  in  detail  in  Appendix  A.  4  See  Harrell,  Beltramo,  Levine,  Blalock,  &  Simons  (2014)  and  Levine,  Beltramo,  Blalock,  &  Cotterman  (2013)  for  more  details  on  the  experimental  design  and  an  overview  of  the  study  area.  

  6  

traditional   stove  was   a   three   stone   fire,5  and   the   nontraditional   stove  was   a   fuel-­‐efficient  Envirofit  G-­‐3300.6  The  study  area   is   characterized  by  agrarian   livelihoods  including  farming  of  matooke  (starchy  cooking  banana),  potatoes,  and  millet  as  well  as  raising  livestock.  At  baseline  almost  all  families  cook  on  a  traditional  three-­‐stone  fire   (97%),  usually   located  within  a  separate  cooking  hut   (62%  of  households  had  totally  enclosed  kitchens  with  no  windows,  while  38%  had  semi-­‐enclosed  kitchens  with   at   least   one   window).   Most   stove   usage   occurs   during   lunch   and   dinner  preparation,  with  matooke  and  beans  the  most  common  and  most  time  consuming  foods  cooked.  Matooke,   a  main   food   for   lunch  and  dinner,   is  unripe  plantain  eaten  after   steaming   for  3-­‐5  hours.  Beans,   another   common  dish,   is  prepared  by  boiling  and   simmering   for   generally   2-­‐4   hours.   Thus   for   the   main   meals,   it   is   common  cooking  practice  to  simmer  and/or  steam  foods  for  several  hours  in  a  row.      The   study   tracked   stove   usage   (before   and   after   the   purchase   of   a   fuel   efficient  stove)   amongst   twelve   households   in   each   of   fourteen   rural   parishes   in   Mbarara  (168   total   households).   Upon   arriving   in   a   new   parish,   staff   displayed   the   new  Envirofit   and   offered   it   for   sale   to   anyone   who   wanted   to   purchase   at   40,000  Ugandan  Shillings  (approximately  USD  $16,  see  Levine  et  al.  (2013)  for  an  overview  of   the   sales   contract).   Consumers   who   wanted   to   buy   the   stove   were   randomly  assigned   into   two  groups  (early  buyers,   late  buyers).  The  project  asked  both  early  buyers  and  late  buyers  if  they  would  agree  to  have  SUMs  placed  on  their  traditional  stoves   immediately.   Then   approximately   four  weeks   later   the   early   buyers   group  received  their  first  Envirofit  stove,  and  approximately  four  weeks  after  that  the  late  buyers   received   their   first   Envirofit   stove.   In   each   parish,   more   than   twelve  households   agreed   to   join   the   study;   therefore   among   those   that   agreed,   we  randomly  selected  twelve  households  per  parish  for  the  usage  study  with  the  SUMs.  Stove  temperatures  were  tracked  for  approximately  six  months  (April  –  September  2012).    Approximately   four  weeks   after   late  buyers   received   their  Envirofits,   both   groups  were  surprised  with  a  second  Envirofit  stove.  Common  cooking  practices  in  the  area  require  two  simultaneous  cooking  pots  (for  example  rice  and  beans,  or  matooke  and  some  type  of  sauce);  therefore,  because  the  Envirofit  is  sized  for  one  cooking  pot,  we  gave  a  second  Envirofit  to  mimic  normal  cooking  behavior  as  much  as  possible.  Each  household  had  as  many  as  two  three  stone  fires  and  two  Envirofit  stoves  monitored  with   SUMs   throughout   the   study.   The   results   of   stove   use   will   be   presented   in   a  subsequent  analysis.    

                                                                                                               5  A  three  stone  fire  is  simply  three  large  stones,  approximately  the  same  height,  on  which  a  cooking  pot  can  be  balanced  over  a  fire.  Almost  all  cooks  burned  firewood  or  other  solid  biomass.  6  Product  manufacturer   features  state   that   the  Envirofit  G-­‐3300  reduces  smoke  and  harmful  gasses  by  up  to  80%,  reduces  biomass  fuel  use  by  up  to  60%  and  reduces  cooking  time  by  up  to  50%.  For  more  details  see:  http://www.envirofit.org/products/?sub=cookstoves&pid=10  

  7  

 Results  Descriptive  statistics    The  data  consist  of  approximately  1.7  million  temperature  readings  from  more  than  34,000  stove-­‐days  of  use  recorded  across  580  stoves  at  168  households.  The  sample  sizes  by  stove  type  are  in  Table  1.      SUMs   must   be   placed   close   enough   to   the   heat   source   to   capture   changes   in  temperatures,   but   not   so   close   that   they   exceed   85°C   or   they   overheat   and  malfunction.  We  do  not  need  to  recover  the  exact  temperature  of  the  hottest  part  of  the   fire   to   learn   about   cooking   behaviors.   Even   with   SUMs   that   are   reading  temperatures   20-­‐30   centimeters   from   the   center   of   the   fire,   as   long   as   the  temperature   readings   for   times  when   stoves   are   in   use   are   largely   different   than  times   when   stoves   are   not   used   the   logistic   regression   will   be   able   to   predict   a  probability  of  usage.  We  did  not  have  strict  control  over  the  location  of  our  SUM  in  a  three   stone   fire,   while   we   did   for   an   Envirofit.   SUMs   for   three   stone   fires   were  placed   in  a  SUM  holder  (Figure  3)7  and  then  placed  under  one  of   the  stones   in  the  three  stone  fire  (left  panel,  Figure  4).  The  SUMs  for  Envirofits  were  attached  using  duct  tape  and  wire  and  placed  at  the  base  of  the  stove  behind  the  intake  location  for  the   firewood   (right   panel,   Figure   4).   Over   multiple   weeks   users   likely   move   the  locations  of  the  stones  in  their  three  stone  fires  depending  on  their  cooking  needs,  pushing  the  stones  closer  together  for  a  smaller  fire  and  wider  to  fit  more  fuel  wood  for  a  larger  fire.  Therefore  we  expect  the  tracking  of  three  stone  fire  temperatures  to  be  more  difficult  with  SUMs  than  the  tracking  of  Envirofit  stove  usage.      Figure  5  shows  an  example  of  SUMs  temperature  data  for  a  household  across  about  three  weeks.  The  left  panel  shows  the  temperatures  registered  in  a  three  stone  fire  versus  the  ambient  temperature  also  recorded  with  SUMs  in  this  household,  while  the   right   panel   compares   the   temperature   of   the   Envirofit   to   the   ambient  temperature  reading.  These   two   figures   illustrate  a  point   that   is   largely  consistent  throughout   our   wider   data   set.   SUMs   readings   of   the   three   stone   fires   are   less  responsive   than   Envirofit   SUMs   readings   to   temperature   changes.   It   seems   that  residual  heat  stays  in  the  cooking  area  of  a  three  stone  fire  preventing  temperatures  from  dropping  all  the  way  to  the  ambient  temperature.  The  design  of  the  Envirofit,  which  is  engineered  to  be  much  more  heat  efficient,  shows  up  in  these  SUMs  graphs  as   less   residual   heat  when   a   stove   is   not   in   use   by   the   temperatures   dropping   to  room  temperature  or  below.      A  final  difficulty  is  found  July  2-­‐4th  (left  panel,  Figure  5)  where  temperatures  for  the  three   stone   fire   remain   above   40°C   for   about   a   48-­‐hour   period.  While   it   could   be  that   the   fire  was  on   for   this  entire  period,   it   could  also  be   that   the  household  was  

                                                                                                               7  The  location  for  optimal  SUM  placement  on  three  stone  fires  and  Envirofit  stoves  was  field  tested  in  Uganda  prior   to   starting   the  experiment.  Additionally,  we  are   thankful   for  numerous   consultations  with  Dr.   Kirk   Smith   and   colleagues   at   the  University   of   California,   Berkeley  who  have   successfully  used  SUMs  and  designed  the  SUM  holder  (Figure  3)  to  track  stove  temperatures  in  previous  studies  (Ruiz-­‐Mercado  et  al.  2008;  Ruiz-­‐Mercado,  Canuz,  and  Smith  2012;  Mukhopadhyay  et  al.  2012).    

  8  

banking  embers   (keeping   coals  warm,  but  not   cooking)  and   the  SUM  was   close   to  those   embers.   This   figure   highlights   some   of   the   potential   challenges   of   using  temperatures  as  a  predictive  mechanism  for  cooking  metrics,  especially  with  three  stone  fires.      Ambient  temperature  sensors  were  placed  in  one  household  in  each  of  the  fourteen  parishes.8  Figure   6   is   a   box   plot   of   the   ambient   temperatures   by   hour   of   the   day  (averaged  over  each  day  of  the  experiment).      There  are  1,668  visual  observations  for  three  stone  fires,  and  782  for  the  Envirofit  stove   (see   Appendix   A   for   a   description   of   the   process   to   match   observations   to  temperature   data).   The   summary   statistics   for   the   entire   sample   and   for   the  trimmed   sample   (dropping   observations   that  were   implausible   given   the   ambient  temperature)  are  in  Table  2.  Note  the  spread  of  mean  temperatures  (“off”  vs.  “on”)  is  wider  for  observations  for  Envirofits  (26.2°C  vs.  41.5°C  in  full  sample,  and  25.3°C  vs.  42.0°C  in  trimmed  sample)  as  compared  to  the  three  stone  fires  (30.2°C  vs.  38.3°C  in  full  sample,  and  28.2°C  vs.  39.0°C  in  trimmed  sample).      Comparison  of  fit:  full  vs.  trimmed  sample  We   first   examine   one   regression   with   two   leads   and   lags   on   both   the   full   and  trimmed  samples.  Table  3  shows  the  logistic  regression  predicting  the  observation  that  the  stove  was  recorded  as  lit.  The  percent  correctly  classified  is  good  (83.1%),  but  the  pseudo-­‐R2  is  a  modest  0.17.  Figure  7  shows  the  predicted  probability  a  stove  is  observed   in  use  at  each  SUM  temperature.  Contrary  to  our  priors,   the  predicted  probability  is  fairly  linear  for  the  full  sample  (especially  for  the  three  stone  fire).      We   originally   hoped   to   let   the   dataset   “speak   for   itself.”   However,   the   predicted  probability   a   three   stone   fire   with   a   temperature   reading   of   50°C   is   “lit”   is   only  about  50%.  Because  50°C  is  much  higher  than  the  highest  ambient  temperature  we  recorded  (40°C),  it  should  have  a  predicted  probability  of  cooking  much  higher  than  50%.      The   predictive   power   rises   substantially   when   we   discard   outliers,   defined   as  observations   coded   “lit”   and   more   than   two   degrees   below   the   hourly   mean   air  temperatures   (Figure   6),   or   “not   lit”   and   more   than   two   degrees   above   the  maximum  air  temperature  we  observed  that  hour  of  the  day  across  the  duration  of  the   experiment   (Figure   6).   This   increase   in   fit   holds   for   the   three   stone   fires  (pseudo-­‐R2   increases   from   0.17   to   0.40,   percentage   correctly   classified   increases  from   83.1%   to   89.3%)   and   Envirofits   (pseudo-­‐R2   increases   from   0.45   to   0.70,  percentage  correctly  classified  increases  from  90.7%  to  93.8%).  (Note  that  trimming  these  outliers  mechanically  improves  the  fit  and  percent  correctly  classified.)  When  we   drop   outliers   the   predicted   probability   now   has   the   expected   logistic   S-­‐shape  (Figure  7).                                                                                                                      8  Ambient   temperature   readings   were   only   recovered   in   13   out   of   14   districts.   In   total   there   are  about  56,000  ambient  temperature  readings.    

  9  

Regression  Specifications    We   test   ten   regression   specifications   on   the   trimmed   sample   for   both   the   three  stone  fires  and  the  Envirofit  (Table  4).  Initially,  we  include  only  the  temperature  at  time  t  (specification  1);  it  has  a  pseudo-­‐R2  of  37%  and  60%,  and  correctly  classifies  87.5%  and  96.1%  of  the  three  stone  fire  and  Envirofit  sample,  respectively.  Adding  a  lead  and  lag  of  one  period  (spec.  2)  increases  the  pseudo  R2  to  41%  and  68%,  and  increases   the   percentage   correctly   classified   to   89.2%   for   three   stone   fires,   but  decreases   the   percentage   correctly   classified   to   94.9%   in   the   Envirofits.   Adding   a  second  (spec.  4),  third  (spec.  6)  or  fourth  (spec.  8)  lead  and  lag  does  not  improve  fit  by  much.  Adding  a  slope-­‐squared  term  when  the  slope  is  positive  for  the  time  step  from   t-­‐1   to   t   (coding   slope-­‐squared   as   zero   otherwise)   measures   steep   positive  increases  in  slope,  as  would  be  the  case  when  a  fire  is  lit.  However  adding  this  slope-­‐squared  term  (col.  3,  5,  and  7)  minimally  changes  the  results.  Only   including   leads  (spec.   9)   or   only   including   lags   (spec.   10)   reduces   pseudo-­‐R2,   but   increases  percentage  correctly  classified  to  89.4%  and  95.3%  (spec.  10)  for  three  stone  fires  and  Envirofits,  respectively.      In   order   to   determine   which   of   the   models   is   most   appropriate   we   test   the   ten  specifications   with   the   Akaike   information   criterion   (AIC)   (Akaike   1974;   Akaike  1981).   The  AIC   trades   off   goodness   of   fit   of   the  model  with   the   complexity   of   the  model  to  guard  against  over  fitting.  The  AIC  is  defined  as:       !"# =  2! − 2  !"(!)   (0)    where  k  is  the  number  of  parameters  in  the  statistical  model,  and  L  is  the  maximized  value  of  the  likelihood  function  of  the  model.  Given  a  set  of  candidate  models  for  a  set   of   data,   the   preferred   model   is   the   one   with   the   lowest   AIC.   Burnham   and  Anderson   (2002)   argue   that   the  AIC   is   appropriate   in   cases  where   the  number   of  observations  is  many  times  larger  that  the  square  of  the  number  of  predictors,  as  in  this  case.      Among  both  the  three  stone  fire  and  Envirofit  sample,  the  AIC  is  a  much  higher  value  for  the  specifications  with  temperature  at  t  (spec.  1),  temperature  at  t  and  four  leads  (spec.  9),  temperature  at  t  and  4  lags  (spec.  10)  than  for  the  other  specifications.  The  remaining  seven  specifications  all  have  similar  AICs.  The  specification  that  includes  two   leads   and   two   lags   (spec.   4)   has   the   lowest   AIC   among   the   three   stone   fire  sample.   The   lowest   AIC   among   the   Envirofit   sample   is   the   specification  with   four  leads  and  four  lags  (spec.  8).  For  simplicity,  we  use  the  same  specification  on  both  the   three   stone   fire   sample   and   the   Envirofit   sample;   therefore,   we   rely   on   the  specification  with  two  leads  and  lags  (spec.  4)  for  both  samples.        Logistic  Regression  Specification  According  to  the  AIC  for  three  stone  fires,  the  preferred  specification  is:       !!" =  !(γ!!!" +  γ!!!"!! +  γ!!!"!! + γ!!!"!! + γ!!!"!! + γ!!!" + !!")   (0)    

  10  

where  F(.)   is   the   logistic   functional   form,  !!"  is  a  dummy  variable   for  observations  for  stove  i  at  time  t.  !!"  is  the  SUMs  temperature  reading  for  stove  i  at  time  t.  !!"!!  is  the  SUMs  temperature  reading  for  stove  i  at  time  t+τ,  for  τ  =  -­‐2,  -­‐1,  1,  and  2  (that  is  30  and  60  minutes  prior   to   the  observation,  and  30  and  60  minutes  after).  !!"  is  a  dummy   variable   for   the   hour   of   the   day   for   stove   i   at   time   t.   The   terms  γ!, γ!, γ!, γ!, γ!, γ!  are  the  parameters  and  !!"  is  an  error  term.    Estimated  odds  ratios  of  the  logistic  regression  for  both  the  three  stone  fires  and  the  Envirofit   are   presented   in   Table   3   (columns   2   and   4).   An   increase   of   one   degree  Celsius  at  time  t  is  associated  with  an  odds  ratio  of  1.37  and  1.46  (both  statistically  significant  at  the  1%  level)  for  the  three  stone  fire  and  Envirofit,  respectively.  This  means,   holding   all   else   constant,   that   an   increase   of   one   degree   Celsius   at   time   t  means  the  odds  of  a  stove  being  “on”  are  approximately  37%  higher  for  three  stone  fires,   and   46%   higher   for   Envirofits   than   the   odds   of   the   stove   being   “on”   at   the  initial   time   t   temperature.  The   three  stone   fire   temperature   reading  at   time   t+2   is  associated  with  an  odds  ratio  of  1.07  (significant  at  5%  level),  meaning  the  odds  of  a  three   stone   fire  being  on  at   time   t   are  7%  higher   in   the  presence  of   a  one  degree  increase   in   temperature   at   time   t+2   holding   all   else   constant.   The   Envirofit  temperature  reading  at  time  t+1  has  an  odds  ratio  of  1.24  (significant  at  1%  level)  and  at  time  t+2  an  odds  ratio  of  0.90  (significant  at  5%  level).  This  means  the  odds  of  an  Envirofit  being  on  at   time   t   are  24%  higher   in   the  presence  of  a  one  degree  increase  in  temperature  at  time  t+1,  and  10%  lower  in  the  presence  of  a  one  degree  increase  in  temperature  at  time  t+2,  holding  all  else  constant.  9      As   a   robustness   check   Table   5   shows   the   in   sample   and   out   of   sample   predictive  accuracy   of   the   preferred   specification.   We   randomly   split   the   sample   of  observations   of   each   stove   in   half   and   ran   the   logistic   regression   on   half   of   the  observations.  We   then  predict   the  probability  of  use   to  both  halves  of   the  sample.  For   the   three   stone   fires   the   in-­‐sample  predictive   accuracy   is  89.8%  while  out-­‐of-­‐sample   predictive   accuracy   is   87.8%.   For   the   Envirofits   the   in-­‐sample   predictive  accuracy  is  95.5%  while  out–of-­‐sample  predictive  accuracy  is  91.1%.  Because  the  in-­‐sample  and  out-­‐of-­‐sample  predictive  accuracy  is  very  similar  for  each  stove  type,  it  is  unlikely  that  we  are  over-­‐fitting  the  data.      A  classification  table  for  the  entire  sample  (Table  6)  shows  how  well  the  regression  predicts  the  entire  observation  sample  for  the  three  stone  fires  and  for  the  Envirofit.  The   sensitivity   (probability   predicted   “lit”   when   observation   observed   as   “lit”)   is  only  56.5%  for  the  three  stone  fires  and  77.6%  for  the  Envirofits,  showing  that  the  technique  struggles  when  only  predicting  the  sample  that  was  observed  as  “lit.”  The                                                                                                                  9  We  also  tested  probit  and  the  linear  probability  model  (LPM)  in  addition  to  the  logistic  regression.  Predictions  of  the  probability  of  stove  usage  from  the  probit  and  logistic  regression  were  correlated  at  levels  higher  than  0.99  for  both  the  three  stone  fire  and  Envirofit  sample.  Among  the  predictions  of  probability   of   stove  use   from   the   linear  probability  model,   28.8%  and  36.2%  of   the  prediction   fell  outside  of  the  range  [0  to  1]  for  the  three  stone  fire  and  Envirofit,  respectively.  As  the  proportion  of  linear   probability  model   predicted   probabilities   that   fall   outside   of   the   unit   interval   increases,   the  potential   bias   of   using   LPM   also   increases   (Horrace   and   Oaxaca   2006).   Because   we   have   a   high  proportion  of  LPM  predictions  that  fall  outside  of  this  range,  using  LPM  would  likely  result  in  biases.  

  11  

specificity   (probability   predicted   “not   lit”   when   the   observation   observed   as   “not  lit”)  is  much  better,  97.6%  for  the  three  stone  fire,  and  97.0%  for  the  Envirofits.  This  technique  generally  holds  more  predictive  power  with  Envirofits   than   three   stone  fires   as   evidenced   by   the   Envirofit   sample   having   a   higher   percentage   of   overall  correctly  classified  observations  (93.8%  vs.  89.2%).      Applying  published  SUMs  algorithms  to  our  data  We   replicate   the   algorithm   described   in   (Ruiz-­‐Mercado,   Canuz,   and   Smith   2012;  Mukhopadhyay  et  al.  2012)  to  convert  temperature  readings  to  hours  of  usage.  We  first   determined   a   threshold   increase   and  decrease   in   stove   temperature   equal   to  the   99th   and   1st   percentiles   of   ambient   temperature   slopes.   In   our   sample   of  approximately  56,000  ambient  temperature  readings  the  99th  percentile  of  slopes  is  0.0833°C/minute  (triggering  slope  of  an  increase  of  2.5°C  in  30  minutes),  and  the  1st  percentile   of   slopes   is   -­‐0.05°C/minute   (exit   slope   of   a   decrease   of   1.5°C   in   30  minutes).    The   algorithm   counts   a   stove   as   turning   “on”   when   the   slope   is   larger   than   a  triggering  slope,  then  it  continues  to  count  each  subsequent  temperature  reading  as  “on”   (as   long   as   temperatures   are   higher   than   the   ambient   temperature)   until   it  reaches  a  negative  slope  that  is  steeper  than  the  exit  slope.  We  calculate  the  on/off  status  of  our  entire  data  set  using   this  algorithm.  We  apply   this  algorithm  to  both  the  traditional  three  stone  fires  and  the  Envirofit  stoves;  however  Mukhopadhyay  et  al.  (2012)  and  Ruiz-­‐Mercado,  Canuz,  and  Smith  (2012)  designed  their  algorithms  for  nontraditional  stoves  and  do  not  endorse  this  technique  for  three  stone  fires.    We   observe   some   irregularities   in   this   algorithm  when   applied   to   our   data,  most  notably  a  very  hot  stove  could  be  marked  as  “off”  too  quickly  if  the  temperature  falls  steeper  than  a  decrease  of  1.5°C  in  30  minutes  even  if  the  absolute  temperature  is  still  hot  enough  to  likely  indicate  cooking.  Take  for  example  the  following  series  of  temperature  data  {26,  51,  52,  47,  49,  46,  41,  37,  33,  31°C}  with  each  reading  thirty  minutes  apart.  The  algorithm  would  return  {off,  on,  on,  off,  off,  off,  off,  off,  off,  off}.  When  the  temperature  falls  from  52°C  to  47°C  between  the  third  and  fourth  reading  the  slope  is  steep  enough  and  the  algorithm  switches  to  off.  However,  in  reality  the  subsequent   temperature   readings   above   40°C   are   likely   to   indicate   stove   usage  (recall  no  ambient  air  temperatures  were  ever  that  high).  Another  irregularity  with  the  basic  algorithm  is  that  stoves  can  cool  too  slowly  to  trigger  an  exit  slope  so  the  algorithm   returns   an   excessively   long   period   of   cooking   (at   times   multiple   days)  despite   temperatures   low   enough   that   cooking   is   unlikely.   For   example   the   data  series  {26,  33,  32,  31,  30,  29,  28,  27,  26,  25,  24,  23,  22,  21,  20°C}  would  return  {off,  on,   on,   on,   on,   on,   on,   on,   on,   on,   on,   on,   on,   on,   on}  despite  20°C  being  below   the  mean  ambient   temperature   in   our  period,   and   almost   certainly  not  hot   enough   to  indicate  stove  usage.    We   created   an   adjusted   algorithm   to   account   for   these   two   cases.   The   first  adjustment   is   that  once   the  algorithm  has   a   triggering   slope  marking   the   stove  as  “on”   it   does   not   consider   an   exit   slope   to   trigger   as   “off”   until   the   absolute  temperature  is  less  than  40°C.  The  second  adjustment  is  that  once  the  algorithm  has  

  12  

a   triggering   slope  marking   the   stove   as   “on”   it  will   switch   to   “off”   even  without   a  sufficiently  steep  exit  slope  if  it  has  negative  slope  readings  for  an  hour  straight  and  the  absolute  temperature  is  below  35°C.    Comparison  of  fit  Next   we   test   which   algorithm   correctly   predicts   the   highest   percentage   of  observations.   In   addition   to   testing   the   Ruiz-­‐Mercado-­‐Mukhopadhyay   et   al.  algorithm,  and  the  adjusted  Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm,  we  also  test   various   cutoff   points   in   relation   to   mean   ambient   temperature.   For   example  considering  all  data  points  above  temperature  thresholds  such  as  32°C,  34°C,  36°C,  and   38°C   as   “on”   (these   correspond   to   approximately   8°C,   10°C,   12°C,   and   14°C  above  mean  ambient  temperature,  respectively).      Table   7   reports   the   classification   table   of   how   well   the   algorithms   predict   the  observations   of   three   stone   fire   use.   The   strict   threshold   cutoff   of   36°C   (correctly  classified  85.0%  of  the  observations.  This  was  the  best  performing  strict  threshold,  correctly  predicting  more   than   the   thresholds  of  32°C   (76.8%),  34°C   (81.9%),  and  38°C  (84.8%).  The  adjusted  Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm  (77.7%)  correctly   predicted   better   than   the   Ruiz-­‐Mercado-­‐Mukhopadhyay   et   al.   algorithm  (58.9%),  but  not  as  well  as  the  strict  threshold  of  36°C  (85.0%).      The  logistic  regression  with  temperature  readings  at  times  t-­‐2,  t-­‐1,  t,  t+1,  t+2  (that  is  at  time  t,  and  30  and  60  minutes  prior  to,  and  30  and  60  minutes  after)  and  hour  of  the  day  predicts  the  most  observations  of  three  stone  fire  use,  predicting  89.3%  of  observations   correctly.   The   pseudo   R-­‐squared   of   the   logistic   regression   explains  41.8%  of  the  variation  in  observations,  whereas  each  of  the  temperature  thresholds  has   pseudo   R-­‐squared   in   the   range   of   10.3%   to   18.4%.   The   Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm  and  the  adjusted  Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm  have  a  pseudo  R-­‐squared  of  0.1%  and  7.7%,  respectively.      Table   8   shows   how  well   the   algorithms   predict   the   observations   of   Envirofit   use.  Among  Envirofits,  the  strict  threshold  cutoff  of  36°C  correctly  classified  94.7%  of  the  observations.   This   was   the   best   performing   strict   threshold,   correctly   predicting  more   than   the   thresholds   of   32°C   (91.6%),   34°C   (93.9%),   and   38°C   (94.3%).   The  strict   threshold   of   36°C   also   correctly   classified   more   observations   (94.7%)   than  both   the   Ruiz-­‐Mercado-­‐Mukhopadhyay   et  al.   algorithm   (84.4%),   and   the   adjusted  Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm  (85.9%).      The  logistic  regression  with  temperature  readings  at  times  t-­‐2,  t-­‐1,  t,  t+1,  t+2  (that  is  at  time  t,  and  30  and  60  minutes  prior  to,  and  30  and  60  minutes  after)  and  hour  of  the   day   correctly   predicts   slightly   fewer   (93.8%)   than   the   best   performing  temperature   threshold   of   36°C   (94.7%).   However,   considering   the   pseudo   R-­‐squared   the   logistic   regression   explains   67.1%   of   the   variation   in   trimmed  observations,  whereas  each  of  the  temperature  thresholds  has  a  pseudo  R-­‐squared  in  the  range  of  27.0%  to  36.7%.  The  Ruiz-­‐Mercado-­‐Mukhopadhyay  et  al.  algorithm  and   the   adjusted   Ruiz-­‐Mercado-­‐Mukhopadhyay   et   al.   algorithm   have   a   pseudo   R-­‐squared   of   14.4%   and   19.6%,   respectively.   Because   the   performance   of   a   strict  

  13  

temperature  threshold  would  vary  based  on  the  local  climate  the  logistic  regression  is  a  more  robust  algorithm.      Evidence  of  a  Hawthorne  Effect    Our  kitchen  performance  tests  (KPT)  involve  visits  on  four  days  in  a  row  to  weigh  wood  and  install  measurement  devices,  all  while  asking  participants  to  record  their  cooking   on   a   food   diary.   A  major   risk   is   that   direct   observational   process   altered  participants’   behavior   (as   noted   by   (Ezzati,   Saleh,   and   Kammen   2000;   Smith-­‐Sivertsen   et   al.   2009)).   Specifically,   it   is   plausible   that   participants   used   their  Envirofit   stoves  more  (and   their   three-­‐stone   fire   less)  during   the  KPT  than  during  other  times.        To   test   for   such  Hawthorne  effects,  we  examine   stove  usage  during   the   final  KPT,  when  all  participants  have  two  Envirofit.  We  compare  stove  usage  during  the  KPT  with   stove   usage   over   the   same   duration   and   days   of   the   week   in   the   prior   and  following   weeks.   Our   measurement   periods   are   roughly   72   hours   (generally   late  Monday   morning   to   early   Thursday   afternoon).   Table   9   shows   the   summary  statistics  for  predicted  hours  cooked  per  day  among  the  four  stoves.  By  using  paired  t-­‐tests   we   compare   stove   usage   on   a   given   stove   within   the   same   household.  Households  used  their  first  Envirofit  about  3.8  hours  per  day  in  the  week  before  the  KPT  and  3.4  hours  in  the  following  week.  Their  usage  was  about  4.6  hours10  during  the  KPT  week.  The  differences  between  the  KPT  measurement  period  with  previous  week   (0.75   hours,   95%   CI   =   [0.12   to   1.37],   p<0.05)   and  with   the   following  week  (1.21  hours,  95%  CI  =  [0.58  to  1.84],  p<0.01)  were  statistically  significant  at  the  5%  and   1%   levels,   respectively.   The   difference   in   daily   hours   cooked   on   the   first  Envirofit   stove   in   the   week   prior   to   and   following   the   KPT   was   not   statistically  significantly  different   than  zero.  Households  used   their   second  Envirofit   about  2.7  hours  per  day  in  the  week  before  the  KPT  and  2.4  hours  in  the  following  week.  Their  usage  was  about  3.4  hours11  during  the  KPT  week.  The  differences  between  the  KPT  measurement   period   with   previous   week   (0.83   hours,   95%   CI   =   [0.27   to   1.38],  p<0.01)  and  with  the  following  week  (0.81  hours,  95%  CI  =  [0.34  to  1.27],  p<0.01)  were  both  highly  statistically  significant.  The  difference  in  daily  hours  cooked  on  the  second   Envirofit   stove   in   the   week   prior   to   and   following   the   KPT   was   not  statistically  significantly  different  than  zero.      While  usage  on  each  Envirofit  stove  rose,  three  stone  fire  usage  declined  during  the  KPT.  Households  used  their  primary  three  stone  fire  about  8.9  hours  per  day  in  the  

                                                                                                               10  To  be  precise,  the  cooking  during  the  KPT  measurement  week  for  the  first  Envirofit  was  4.59  hours  per   day   among   the   105   pairs   of   households   that   had   SUMs   temperature   readings   on   their   first  Envirofit   stove   for  both   the  KPT  week  and   the  prior  week,  and  4.56  hours  among   the  111  pairs  of  households  that  had  SUMs  temperature  readings  on  their  first  Envirofit  stove  for  the  KPT  week  and  the  following  week.  11  To  be  precise,   the  cooking  during   the  KPT  measurement  week   for   the  second  Envirofit  was  3.48  hours   per   day   among   the   88   pairs   of   households   that   had   SUMs   temperature   readings   on   their  second  Envirofit   for  both   the  KPT  week  and  the  prior  week,  and  3.24  hours  among  the  95  pairs  of  households  that  had  SUMs  temperature  readings  for  their  second  Envirofit  for  the  KPT  week  and  the  following  week.  

  14  

week  before  the  KPT  and  10.1  hours  in  the  following  week.  Their  usage  was  about  7.4   hours12  during   the  KPT  week.   The   differences   between   the  KPT  measurement  period  with  previous  week  (1.44  hours,  95%  CI  =  [0.43  to  2.46],  p<0.01)  and  with  the  following  week  (2.74  hours,  95%  CI  =  [1.56  to  3.93],  p<0.01)  were  both  highly  statistically   significant.  The  difference   in  daily  hours   cooked  on   the  primary   three  stone   fire   in   the   week   prior   to   and   following   the   KPT   was   weakly   statistically  significantly   different   than   zero   (p<0.10).   Households   used   their   secondary   three  stone  fire  about  9.8  hours  per  day  in  the  week  before  the  KPT  and  9.8  hours  in  the  following   week.   Their   usage   was   9.0   during   the   KPT   week.   Due   to   a   very   small  sample  size,13  none  of  these  differences  are  significantly  different  than  zero.        Ultimately,  our  interest  is  to  know  how  combined  cooking  on  Envirofits  compares  to  combined   cooking   on   three   stone   fires   during   this   observation   period.   Due   to   a  constraint  in  the  number  of  functioning  SUM  devices  at  the  end  of  our  experiment,  we  have  a  very  small   sample   (N=22)  of   temperature  readings   for  secondary   three  stone   fires.   Therefore   to   analyze   total   combined   stove   usage   we   will   use   the  balanced  sample  in  which  we  successfully  recovered  SUMs  data  for  both  Envirofits  and   for   the   primary   three   stone   fire   (Table   10)   in   a   given  household.  Among   this  sample  (N=48)  households  used  their  Envirofits  about  6.1  hours  per  day  in  the  week  before   the   KPT   and   5.6   hours   in   the   following   week.   Their   usage   was   8.1   hours  during   the   KPT   week.   (Recall   we   match   the   same   measurement   period   in   each  week.)  The  differences  between  the  KPT  measurement  period  with  previous  week  (2.06   hours,   95%   CI   =   [1.07   to   3.05],   p<0.01)   and  with   the   following  week   (2.54  hours,  95%  CI  =  [1.66  to  3.41],  p<0.01)  were  both  highly  statistically  significant.  The  difference  in  daily  hours  cooked  in  the  week  prior  to  and  following  the  KPT  was  not  statistically  significantly  different  than  zero.      While  Envirofit  use  rose,   three  stone   fire  use  declined  during   the  KPT.  Among   the  same  sample   (Table  10)  households  used   their  primary   three   stone   fire  about  8.1  hours  per  day  in  the  week  prior  to  the  KPT  and  10.5  hours  in  the  week  following  the  KPT.   The   usage   was   7.0   hours   per   day   during   the   KPT   week.   The   differences  between  the  KPT  measurement  periods  with  previous  week  (1.13  hours,  95%  CI  =  [-­‐0.13  to  2.38],  p<0.10)  and  with  the   following  week  (3.50  hours,  95%  CI  =   [1.92  to  5.09],  p<0.01)  were  statistically  significant  at   the  10%  and  1%   levels   respectively.  The   difference   in   daily   hours   cooked   on   the   primary   three   stone   fire   in   the  week                                                                                                                  12  To  be  precise,  the  cooking  during  the  KPT  measurement  week  for  the  primary  three  stone  fire  was  7.43  hours  per  day  among  the  98  pairs  of  households  that  had  SUMs  temperature  readings  on  their  primary  three  stone  fire  for  both  the  KPT  week  and  the  prior  week,  and  7.40  hours  among  the  100  pairs   of   households   that  had  SUMs   temperature   readings  on   their  primary   three   stone   fire   for   the  KPT  week  and  the  following  week.  13  SUMs  devices  malfunction  and  can  no  longer  be  used  once  they  reach  85°C.  Our  study  is  similar  to  others  (Ruiz-­‐Mercado,  Canuz,  and  Smith  2012;  Ruiz-­‐Mercado  et  al.  2008;  Mukhopadhyay  et  al.  2012)  in   that   our   experiment   was   no   exception   to   the   losses   of   SUMs   devices.   Due   to   losses   from  malfunction   or   the   inability   to   recover   all   the   devices  we   initially   placed,  we   did   not   have   enough  SUMs  to  place  on  all  stoves  by  the  end  of  the  experiment.  When  the  final  KPT  took  place  (near  the  end  of   our   experiment)  we   instructed   enumerators   to   exclude   the   second   three   stone   fire   if   they  were  constrained  in  the  number  of  SUMs  devices.  Table  9  shows  few  SUMs  readings  from  the  second  three  stone  fire  (N=22)  compared  to  the  other  three  stoves  types.    

  15  

prior   to   and   following   the   KPT   was   significantly   different   (p<.01).   All   parishes  overlapped  in  time,  and  therefore  robust  to  time  effects.    While  balancing  the  sample  and  focusing  on  combined  Envirofit  usage  compared  to  primary   three   stone   fire   usage   within   the   same   household   shows   evidence   of   a  Hawthorne   effect,   it   also   drastically   reduces   the   sample   size.   Another   option   that  utilizes  more  data  is  regression  analysis  to  estimate  the  effect  of  the  presence  of  the  KPT  measurement  team  on  stove  usage.  Assign  the  subscripts  t=-­‐1  to  the  week  prior  to  measurement  week,  t=0  to  the  measurement  week,  and  t=1  to  the  week  after  the  measurement  week.   Let  !!"!"#  be   total   hours   cooked  per   day   on   the   primary   three  stone  fire   for  household   i  during  week  t.  Let  !"#$%&'(_!""#  be  a  dummy  variable  for  an  adjacent  week  (t=-­‐1  or  t=1).  The  regression  is  modeled  using  Ordinary  Least  Squares  (OLS)  as:       !!"!"# =  !!"# ∗ !"#$%&'(_!""# + !! + !!"   (1)    where  !!  is   fixed   effects   for   the   individual   household   (this   controls   for   household  level   characteristics   that   don’t   change   over   these   three   weeks   like   family   size,  income,   etc.),   and  !!"  is   an   error   term.   The   coefficient  !!"#  is   the   estimate   of   how  different  (in  hours  cooked  per  day)  an  adjacent  week  is  compared  to  a  measurement  week  (this  specification  imposes  that  the  previous  and  subsequent  week  are  equal).  Standard  errors  are  clustered  at  the  household  level.    The  regression  for  the  combined  Envirofit  usage  is  similar:       !!"!"# =  !!"# ∗ !"#$%&'(_!""# + !! + !!"   (2)    where  !!"!"#  is   the   total   hours   cooked   per   day   on   both   Envirofits   and   all   other  variables   are   the   same   as   in   the   three   stone   fire   case.   The   coefficient  !!"#  is   the  estimate  of  how  different  (in  hours  cooked  per  day)  an  adjacent  week  is  compared  to   the   measurement   week.   This   also   imposes   equality   of   the   previous   and  subsequent  week.    To  test  the  weeks  separately,  we  use  a  slightly  different  specification.  Let  Week_Prior  be  a  dummy  variable  equal  to  1  for  the  week  before  the  measurement  week  (when  t=-­‐1)   and  0  otherwise,   and   let  Week_After   be  a  dummy  variable  equal   to  1   for   the  week  after  the  measurement  week  (when  t=1)  and  0  otherwise.  Then  the  regression  is  modeled  using  OLS  as:       !!"!"# =  !!!"#!""#_!"#$" + !!!"#!""#_!"#$%  + !! + !!"   (3)    where  !!  is  fixed  effects  for  the  individual  household  and  !!!"#  is  the  estimate  of  the  difference   (in   hours   cooked   per   day)   of   the   week   before   the  measurement   week  compared   to   the  measurement  week.  The   coefficient  of  !!!"#  is   the   estimate  of   the  difference   (in   hours   cooked   per   day)   of   the   week   after   the   measurement   week  compared   to   the   measurement   week.   The   standard   errors   are   clustered   at   the  

  16  

household   level.   The   specification   for   the   combined   total   daily   cooking   with  Envirofits  is  similar:       !!"!"# =  !!!"#!""#_!"#$" + !!!"#!""#_!"#$%  + !! + !!"   (4)    where  !!"!"#  is   the   total   hours   cooked   per   day   on   both   Envirofits   and   all   other  variables   are   the   same   as   in   the   three   stone   fire   case.   The   coefficient  !!!"#  is   the  estimate   of   how   different   (in   hours   cooked   per   day)   the   week   prior   to   the  measurement  week  is  compared  to  the  measurement  week.  The  coefficient  !!!"#  is  the   estimate   of   how   different   (in   hours   cooked   per   day)   the   week   after   the  measurement  week  is  compared  to  the  measurement  week.      Based   on   the   regression   estimates   (Table   11)   the   patterns   of   stove   use   are  essentially  the  same  (as  Tables  9-­‐10)  as  Envirofit  usage  increases  while  three  stone  fire   usage   decreases   in   the   presence   of   the   measurement   team.   These   effects  disappear  once  the  measurement  team  departs.  When  constraining  the  week  prior  to  and  after  the  measurement  week  to  be  equal,  estimated  primary  three  stone  fire  usage   is   higher   by   2.26   hours   per   day   (95%   CI   =   [1.13   to   3.39],   p<0.01)   when  comparing  the  adjacent  weeks  to  the  measurement  week  (col.  1)  and  the  estimated  total  Envirofit  usage  is  lower  by  2.40  hours  per  day  (95%  CI  =  [1.48  to  3.32],  p<0.01)  when  comparing  the  adjacent  weeks  to  the  measurement  week  (col.  3).      Running  the  analysis  weekly,  the  estimated  use  of  primary  three  stone  fires  (col.  2)  as   compared   to  usage   levels  during   the  measurement  week   is   1.62  hours  per  day  higher  (95%  CI  =  [0.32  to  2.92],  p<0.05)  and  2.88  hours  per  day  higher  (95%  CI  =  [1.37   to   4.39],   p<0.01)   in   the   week   prior   to   and   after   the   measurement   week,  respectively.  These  coefficients  are   jointly   significantly  different   than  zero   (joint  F  test   significant   at   1%   level).   The   total   usage  on  Envirofits   (col.   4)   as   compared   to  Envirofit  usage   levels  during   the  measurement  week   is  estimated  as  a  decrease  of  2.02  hours  per  day  (95%  CI  =  [0.92  to  3.13],  p<0.01)  and  a  decrease  of  2.71  hours  per   day   (95%   CI   =   [1.72   to   3.71],   p<0.01)   in   the   week   prior   to   and   after   the  measurement   week,   respectively.   These   coefficients   are   jointly   significantly  different  from  zero  (joint  F  test  significant  at  1%  level).      Importance  of  a  Hawthorne  Effect    To  see  the  importance  of  high  Envirofit  usage  during  the  KPT,  consider  the  following  illustrative  example.  Assume:  

o Households   without   Envirofits   use   three   stone   fires   twelve  hours  per  day  

o Households  with  Envirofits  normally  use  three  stone  fires  nine  hours  a  day  and  Envirofits  four  hours  per  day  

o During   the   KPT,   households   with   Envirofits   use   three   stone  fires  seven  hours  a  day  and  Envirofits  six  hours  per  day  

o Three  stone  fires  create  one  unit  and  Envirofits  create  ½  unit  of  pollution  per  hour  

 

  17  

With   these   illustrative   assumptions,   pollution   per   day   declines   from   twelve   units  prior  to  the  introduction  of  an  Envirofit  to  eleven  units  of  pollution  (9  +  (½  *4)  =  11)  once   the  Envirofit   is  present,   a  decline  of  8.33%.  However,   if   instead  we  used   the  data   from   the   kitchen  performance   test,  we  would   estimate   a   decline   in   pollution  from  twelve  units   to   ten  units   (7  +   (½*6)  =  10),  a  decline  of  16.67%,  or   twice   the  true   decline.   While   these   figures   are   merely   illustrative,   they   illustrate   the  importance  of  observer  effects  in  observational  studies.      The   kitchen   performance   test   is   one   of   the  workhorse   tests   in   determining   field-­‐based  use  of  stoves.  Our  results  suggest  the  need  to  correct  for  the  Hawthorne  effect  prior   to  publishing   results.   For   example  Bailis   et   al.   (2007)   find,  when   comparing  the   results   of   pre   and  post   intervention  kitchen  performance   tests,   that   improved  cookstoves   statistically   significantly   reduce   average   per   capita   fuel   use   and   this  reduction  ranges  from  19%-­‐67%  in  rural  Mexico.  Reductions  of  the  emissions  of  CO2  equivalents  is  estimated  at  3.9  tons  per  home  per  year  (Johnson  et  al.,  2009)  in  the  same   study   area   in   rural   Mexico.   In   these   studies   daily   biomass   consumption  patterns  were   calculated  based  on   the   results  of   kitchen  performance   tests.  Other  important  studies  focusing  on  the  health  benefits  of  improved  stoves  (Ezzati,  Saleh,  and  Kammen  2000;  Smith-­‐Sivertsen  et  al.  2009)  also  use  kitchen  performance  tests  or  other  direct  observational  methods  when  measuring   indoor  air  pollution.  Given  our  findings  demonstrate  that  stove  users  change  their  behaviors,  favoring  the  clean  stove   technology  when   they   are   directly   observed,   it   is   possible   that   estimates   of  health  benefits  would  be  incorrectly  estimated  if  the  underlying  behaviors  in  those  studies  were  influenced  in  a  similar  manner.      Conclusion  We  describe  an  algorithm  that  combines  observations  of  stove  use  with  continuous  temperature  data  from  a  stove  usage  monitor  to  predict  the  on/off  status  of  a  stove.  We  test  various  regression  specifications  and  use  the  Akaike  Information  Criterion  for   choosing   a   logistic   regression   where   observations   are   regressed   against  temperature  at   times  t-­‐2,  t-­‐1,  t,  t+1,  t+2  and  a  dummy  variable   for  hour  of   the  day.  This   regression   specification   correctly   predicts   89.3%   of   three   stone   fire  observations   and   93.8%   of   Envirofit   observations   of   stove   usage.   The   predictive  accuracy  is  almost  the  same  in  and  out-­‐of-­‐sample.      We   continue   by   comparing   the   performance   of   this   algorithm   to   previously  published  algorithms  (Mukhopadhyay  et  al.  2012;  Ruiz-­‐Mercado,  Canuz,  and  Smith  2012)   and   strict   temperature   threshold   cutoffs.   Our   logistic   regression   correctly  classifies   more   observations,   with   a   higher   pseudo   R-­‐squared,   than   any   other  algorithm   for   three   stone   fires.   Among   the   Envirofit   observations   our   logistic  regression   correctly   predicts   a   higher   percentage   of   observations   than   both   the  Ruiz-­‐Mercado-­‐Mukhopadhyay   et   al.   algorithm   and   the   adjusted   Ruiz-­‐Mercado-­‐Mukhopadhyay   et   al.   algorithm.   The   strict   36°C   threshold   predicts   slightly   more  observations   (0.4%)   than   our   logistic   regression,   however   the   logistic   regression  has   a   much   higher   pseudo   R-­‐squared   than   all   other   algorithms.  We   argue   that   a  predictive   logistic   regression   with   observations   of   actual   use   and   temperature  readings  at  times  t-­‐2,  t-­‐1,  t,  t+1,  t+2  and  hour  of  the  day  is  the  best  algorithm  for  both  

  18  

traditional  and  nontraditional  stove  types.  Further,  this  algorithm  is  transferable  to  many  climatic  zones  (both  hot  and  cold)  whereas  strict  temperature  thresholds  and  algorithms  based  on  entry  and  exit  slopes  need  adjustment  depending  on  the  local  climate  and  local  cooking  practices.    We   find   evidence   of   a   large   Hawthorne   effect   during   the   72   hour-­‐long   kitchen  performance   tests.   Households   appear   to   increase   cooking   on   Envirofit   stoves   by  32.4%  (2.02  hours  per  day)  during  the  measurement  week;  followed  by  a  reduction  in  Envirofit  use  of  33.6%  (2.71  hours  per  day)  in  the  same  period  the  week  after  the  measurement   team   has   departed.   Further,   use   of   the   primary   three   stone   fires  decreases  by  18.7%  (1.62  hours  per  day)  during  the  measurement  week,   followed  by   an   increase   of   39.9%   (2.88   hours   per   day)   in   the   week   after   the   kitchen  performance  test.    Attention  is  warranted  in  interpreting  results  of  observational  studies,  our  finding  of  a  large  and  statistically  significant  Hawthorne  effect  suggests  that  reported  benefits  due  to  fuel  efficient  and  clean  burning  stoves  could  be  overstated  if  researchers  are  measuring  the  benefits  of  an  improved  stove  when  behavior  is  biased  in  favor  of  the  cleaner   technology   due   to   a   Hawthorne   effect.   By   using   SUMs   and   a   machine-­‐learning   algorithm   to   predict   the   probability   of   a   stove   in   use,  we   show   one  way  around  this  obstacle.    Bibliography  

Akaike,   Hirotugu.   1974.   “A   New   Look   at   the   Statistical   Model   Identification.”  Automatic  Control,  IEEE  Transactions  on  19  (6):  716–723.  

———.   1981.   “Likelihood   of   a   Model   and   Information   Criteria.”   Journal   of  Econometrics  16:  3–14.  

Bailis,  Rob,  Victor  Berrueta,  Chaya  Chengappa,  Karabi  Dutta,  Rufus  Edwards,  Omar  Masera,   Dean   Still,   and   Kirk   R.   Smith.   2007.   “Performance   Testing   for  Monitoring   Improved   Biomass   Stove   Interventions:   Experiences   of   the  Household  Energy  and  Health  Project.”  Energy  for  Sustainable  Development  11  (2)   (June):   57–70.   doi:10.1016/S0973-­‐0826(08)60400-­‐7.  http://linkinghub.elsevier.com/retrieve/pii/S0973082608604007.  

Bailis,   Rob,   Kirk   R.   Smith,   and   Rufus   Edwards.   2007.   “Kitchen   Performance   Test  (KPT)”.  Univerisity  of  California,  Berkeley,  CA.  

Bailis,   Robert,   Majid   Ezzati,   and   Daniel   M   Kammen.   2005.   “Mortality   and  Greenhouse  Gas  Impacts  of  Biomass  and  Petroleum  Energy  Futures   in  Africa.”  Science   308   (5718)   (April   1):   98–103.   doi:10.1126/science.1106881.  http://www.ncbi.nlm.nih.gov/pubmed/15802601.  

Bond,  Tami,  Chandra  Venkataraman,  and  Omar  Masera.  2004.  “Global  Atmospheric  Impacts   of   Residential   Fuels.”   Energy   for   Sustainable   Development   8   (3)  

  19  

(September):   20–32.   doi:10.1016/S0973-­‐0826(08)60464-­‐0.  http://linkinghub.elsevier.com/retrieve/pii/S0973082608604640.  

Burnham,  Kenneth  P.,  and  David  R.  Anderson.  2002.  Model  Selection  and  Multimodel  Inference:  A  Practical   Information-­‐Theoretic  Approach.   2nd   ed.   New   York,   NY:  Springer-­‐Verlag.  

Chetty,  Raj,  John  N.  Friedman,  and  Jonah  E.  Rockoff.  2011.  “The  Long-­‐Term  Impacts  of   Teachers:   Teacher   Value-­‐Added   and   Student   Outcomes   in   Adulthood”.  Cambridge,  MA,  NBER.  

Choi,   Hyunyoung,   and   Hal   Varian.   2012.   “Predicting   the   Present   with   Google  Trends.”   Economic   Record   88   (June   27):   2–9.   doi:10.1111/j.1475-­‐4932.2012.00809.x.  http://doi.wiley.com/10.1111/j.1475-­‐4932.2012.00809.x.  

Einav,   Liran,   Dan   Knoepfle,   Jonathan   D   Levin,   and   Neel   Sundaresan.   2012.   “Sales  Taxes  and  Internet  Commerce”.  Cambridge,  MA,  NBER.  

Einav,   Liran,   and   Jonathan   D   Levin.   2014.   “The   Data   Revolution   and   Economic  Analysis.”  In  Innovation  Policy  and  the  Economy,  edited  by  Josh  Lerner  and  Scott  Stern,  Vol.  14.  University  of  Chicago  Press.  

Ezzati,   Majid,   and   Daniel   M   Kammen.   2001.   “Indoor   Air   Pollution   from   Biomass  Combustion  and  Acute  Respiratory  Infections  in  Kenya:  An  Exposure-­‐Response  Study.”   Lancet   358   (9282)   (August   25):   619–24.  http://www.ncbi.nlm.nih.gov/pubmed/11530148.  

———.  2002.   “Evaluating   the  Health  Benefits   of   Transitions   in  Household  Energy  Technologies   in   Kenya.”   Energy   Policy   30   (10)   (August):   815–826.  doi:10.1016/S0301-­‐4215(01)00125-­‐2.  http://linkinghub.elsevier.com/retrieve/pii/S0301421501001252.  

Ezzati,  Majid,  Homayoun  Saleh,  and  Daniel  M  Kammen.  2000.  “The  Contributions  of  Emissions  and  Spatial  Microenvironments  to  Exposure  to  Indoor  Air  Pollution  from   Biomass   Combustion   in   Kenya.”   Environmental   Health   Perspectives   108  (9)  (September):  833–839.  

Finkelstein,   Amy,   Sarah   Taubman,   Bill   Wright,   Mira   Bernstein,   Jonathan   Gruber,  Joseph  P  Newhouse,  Heidi  Allen,  Katherine  Baicker,   and  Oregon  Health   Study  Group.   2012.   “The   Oregon   Health   Insurance   Experiment:   Evidence   from   the  First   Year.”   The   Quarterly   Journal   of   Economics   127   (3):   1057–1106.  doi:10.1093/qje/qjs020.Advance.  

Freeman,   Olivia   E.,   and   Hisham   Zerriffi.   2012.   “Carbon   Credits   for   Cookstoves:  Trade-­‐Offs  in  Climate  and  Health  Benefits.”  The  Forestry  Chronicle  88  (5):  600–608.  

  20  

GIZ.  2011.   “Carbon  Markets   for   Improved  Cooking  Stoves:  A  GIZ  Guide   for  Project  Operators”.  Poverty-­‐oriented  Basic  Energy  Services,  Eschborn,  Germany.  

Goel,   Sharad,   Jake   M   Hofman,   Sébastien   Lahaie,   David   M   Pennock,   and   Duncan   J  Watts.  2010.  “Predicting  Consumer  Behavior  with  Web  Search.”  Proceedings  of  the   National   Academy   of   Sciences   of   the   United   States   of   America   107   (41)  (October   12):   17486–90.   doi:10.1073/pnas.1005962107.  http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=2955127&tool=pmcentrez&rendertype=abstract.  

Gold   Standard   Foundation.   2013.   “The  Gold   Standard:   Simplified  Methodology   for  Efficient  Cookstoves”.  Geneva-­‐Cointrin,  Switzerland.  

Harrell,  Stephen,  Theresa  Beltramo,  David  I.  Levine,  Garrick  Blalock,  and  Andrew  M.  Simons.   2014.   “What   Is   a   ‘Meal’?   Comparing   Methods   to   Determine   Cooking  Events”.  Mimeo.  

Horrace,   William   C.,   and   Ronald   L.   Oaxaca.   2006.   “Results   on   the   Bias   and  Inconsistency   of   Ordinary   Least   Squares   for   the   Linear   Probability   Model.”  Economics  Letters  90  (3)  (March):  321–327.  doi:10.1016/j.econlet.2005.08.024.  http://linkinghub.elsevier.com/retrieve/pii/S0165176505003150.  

Johnson,   Michael,   Rufus   Edwards,   Adrián   Ghilardi,   Victor   Berrueta,   Dan   Gillen,  Claudio   Alatorre   Frenk,   and   Omar   Masera.   2009.   “Quantification   of   Carbon  Savings  from  Improved  Biomass  Cookstove  Projects.”  Environmental  Science  &  Technology   43   (7)   (April):   2456–2462.   doi:10.1021/es801564u.  http://pubs.acs.org/doi/abs/10.1021/es801564u.  

Kammen,   Daniel   M,   Robert   Bailis,   and   Antonia   Herzog.   2002.   “Clean   Energy   for  Development   and   Economic   Growth:   Biomass   and   Other   Renewable   Energy  Options   to  Meet  Energy  and  Development  Needs   in  Poor  Nations”.  New  York:  UNDP  and  Government  of  Morocco.  

Levine,   David   I,   Theresa   Beltramo,   Garrick   Blalock,   and   Carolyn   Cotterman.   2013.  “What   Impedes   Efficient   Adoption   of   Products?   Evidence   from   Randomized  Variation  in  Sales  Offers  for  Improved  Cookstoves  in  Uganda”.  Mimeo.  

Lim,   Stephen   S,   Theo   Vos,   Abraham   D   Flaxman,   Goodarz   Danaei,   Kenji   Shibuya,  Heather   Adair-­‐Rohani,   Markus   Amann,   et   al.   2012.   “A   Comparative   Risk  Assessment  of  Burden  of  Disease  and  Injury  Attributable  to  67  Risk  Factors  and  Risk   Factor  Clusters   in   21  Regions,   1990-­‐2010:  A   Systematic  Analysis   for   the  Global   Burden   of   Disease   Study   2010.”   Lancet   380   (9859)   (December   15):  2224–60.  

Martin   II,  William   J,   Roger   I   Glass,   John  M  Balbus,   and   Francis   S   Collins.   2011.   “A  Major  Environmental  Cause  of  Death.”  Science  334:  180–181.  

  21  

Masera,  Omar  R,  Barbara  D  Saatkamp,  and  Daniel  M  Kammen.  2000.   “From  Linear  Fuel  Switching  to  Multiple  Cooking  Strategies:  A  Critique  and  Alternative  to  the  Energy   Ladder   Model.”  World  Development   28   (12)   (December):   2083–2103.  doi:10.1016/S0305-­‐750X(00)00076-­‐0.  http://linkinghub.elsevier.com/retrieve/pii/S0305750X00000760.  

McCracken,  John,  J  Schwartz,  M  Mittleman,  L  Ryan,  A  Diaz  Artiga,  and  Kirk  R  Smith.  2007.   “Biomass   Smoke   Exposure   and   Acute   Lower   Respiratory   Infections  Among  Guatemalan  Children.”  Epidemiology  18  (5):  S185.  

Mukhopadhyay,   Rupak,   Sankar   Sambandam,   Ajay   Pillarisetti,   Darby   Jack,  Krishnendu  Mukhopadhyay,  Kalpana  Balakrishnan,  Mayur  Vaswani,  et  al.  2012.  “Cooking  Practices,  Air  Quality,  and   the  Acceptability  of  Advanced  Cookstoves  in  Haryana,   India:  An  Exploratory  Study   to   Inform  Large-­‐Scale   Interventions.”  Global  Health  Action  5:  19016.  

Ramanathan,  V.,  and  G.  Carmichael.  2008.  “Global  and  Regional  Climate  Changes  due  to  Black  Carbon.”  Nature  Geoscience  1  (4):  221–227.  

Ruiz-­‐Mercado,   Ilse,   Eduardo   Canuz,   and   Kirk   R.   Smith.   2012.   “Temperature  Dataloggers  as  Stove  Use  Monitors  (SUMs):  Field  Methods  and  Signal  Analysis.”  Biomass  and  Bioenergy  47:  459–468.  doi:10.1016/j.biombioe.2012.09.003.  

Ruiz-­‐Mercado,   Ilse,   Eduardo   Canuz,   Joan   L.   Walker,   and   Kirk   R.   Smith.   2013.  “Quantitative   Metrics   of   Stove   Adoption   Using   Stove   Use   Monitors   (SUMs).”  Biomass   and   Bioenergy   57   (August):   136–148.  doi:10.1016/j.biombioe.2013.07.002.  http://linkinghub.elsevier.com/retrieve/pii/S096195341300319X.  

Ruiz-­‐Mercado,   Ilse,  Nick  L  Lam,  Eduardo  Canuz,  Gilberto  Davila,  and  Kirk  R  Smith.  2008.  “Low-­‐Cost  Temperature  Loggers  as  Stove  Use  Monitors  (SUMs).”  Boiling  Point  55:  16–18.  

Smith,   Kirk   R,   Nigel   Bruce,   Byron   Arana,   Alisa   Jenny,   Asheena   Khalakdina,   John  McCracken,  Anaite  Diaz,  Morten  Schei,  and  Sandy  Gove.  2006.  “Conducting  The  First   Randomized   Control   Trial   of   Acute   Lower   Respiratory   Infections   and  Indoor  Air  Pollution:  Description  of  Process  and  Methods.”  Epidemiology  17  (6):  S44.  

Smith,  Kirk  R,   John  P  McCracken,   Lisa  Thompson,  Rufus  Edwards,  Kyra  N   Shields,  Eduardo   Canuz,   and   Nigel   Bruce.   2010.   “Personal   Child   and   Mother   Carbon  Monoxide   Exposures   and   Kitchen   Levels:   Methods   and   Results   from   a  Randomized  Trial  of  Woodfired  Chimney  Cookstoves  in  Guatemala  (RESPIRE).”  Journal  of  Exposure  Science  and  Environmental  Epidemiology  20  (5)  (July):  406–16.   doi:10.1038/jes.2009.30.  http://www.ncbi.nlm.nih.gov/pubmed/19536077.  

  22  

Smith,  Kirk  R,  John  P  McCracken,  Martin  W  Weber,  Alan  Hubbard,  Alisa  Jenny,  Lisa  M  Thompson,   John   Balmes,   Anaité   Diaz,   Byron   Arana,   and   Nigel   Bruce.   2011.  “Effect   of   Reduction   in   Household   Air   Pollution   on   Childhood   Pneumonia   in  Guatemala   (RESPIRE):   A   Randomised   Controlled   Trial.”   Lancet   378   (9804)  (November   12):   1717–26.   doi:10.1016/S0140-­‐6736(11)60921-­‐5.  http://www.ncbi.nlm.nih.gov/pubmed/22078686.  

Smith-­‐Sivertsen,   Tone,   Esperanza   Díaz,   Dan   Pope,   Rolv   T   Lie,   Anaite   Díaz,   John  McCracken,   Per   Bakke,   Byron   Arana,   Kirk   R   Smith,   and   Nigel   Bruce.   2009.  “Effect   of   Reducing   Indoor   Air   Pollution   on  Women’s   Respiratory   Symptoms  and   Lung   Function:   The   RESPIRE   Randomized   Trial,   Guatemala.”   American  Journal   of   Epidemiology   170   (2)   (July   15):   211–20.   doi:10.1093/aje/kwp100.  http://www.ncbi.nlm.nih.gov/pubmed/19443665.  

Figure 1Photograph of stove use monitoring system (SUMs)

Source: http://www.berkeleyair.com/products-and-services/instrument-services/78-sums

Figure 2Map of Uganda with Mbarara district highlighted

1

Figure 3SUM holder designed to encase the stove use monitor to protect it from malfunctions when exceeding

temperatures of 85 degrees Celsius

Figure 4Arrows mark the placement of SUMs on three stone fire and Envirofit

(a) Three Stone Fire (b) Envirofit

2

Figure 5Temperature readings on a three stone fire, Envirofit, and ambient air in one household over 20 days

16jun2012 22jun2012 04jul2012Date

28jun2012

(a) Three Stone Fire

16jun2012 22jun2012 04jul2012Date

28jun2012

(b) Envirofit

Figure 6Distribution of hourly ambient temperatures

1520

2530

3540

SUM

s Te

mpe

ratu

re, [

C]

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

when time is formatted 00:00-23:59Box Plot of Ambient Temperature by Hour of the Day

Hour of Day

Box and whisker plot: The bottom and top of each box represent the first and the third quartiles of the hourly ambient temperaturedistribution, the horizontal band inside each box represents the median hourly ambient temperature. The end of the whiskers marked with a

horizontal line represent the lowest datum within 1.5 times the interquartile range of the lower quartile and the highest datum within 1.5 timesthe interquartile range of the upper quartile. The dots represent individual observations that are outliers outside of this range.

3

Figure 7Comparison of predicted probability of stove in use: full verses trimmed sample

0.2

.4.6

.81

Prob

abilit

y St

ove

in U

se

20 30 40 50Temperature [C]

TSF (full sample) TSF (trimmed sample)

(a) Three Stone Fire

0.2

.4.6

.81

Prob

abilit

y St

ove

in U

se

20 30 40 50Temperature [C]

ENV (full sample) ENV (trimmed sample)

(b) Envirofit

Note: The trimmed sample dropped observations when observed as ‘lit’ and more than two degrees below hourly mean air temperature, or ifobserved as ‘not lit’ but more than two degrees above the hourly maximum air temperature.

4

Table 1Temperature readings, days, stoves observed by stove type

STOVE TYPE READINGS DAYS STOVESEnvirofit 1st 432,033 8,945 153Envirofit 2nd 137,413 2,974 107Three Stone Fire 1st 708,048 14,081 168Three Stone Fire 2nd 417,149 8,204 152Totals 1,694,643 34,204 580Note: data collected March - September 2012

Table 2Summary statistics: observations of stove usage

Observations Mean Std. Dev. Min. Max. NTemp [C] when three stone fire observed as off 30.17 7.27 19.5 85 1339Temp [C] when three stone fire observed as on 38.30 10.63 20 85 329Temp [C] when Envirofit observed as off 26.17 6.34 18.5 76 704Temp [C] when Envirofit observed as on 41.52 11.7 20 69.5 78

Observations - Trimmed Sample* Mean Std. Dev. Min. Max. NTemp [C] when three stone fire observed as off 28.21 4.60 19.5 41.5 1169Temp [C] when three stone fire observed as on 38.97 10.36 22 85 315Temp [C] when Envirofit observed as off 25.27 3.86 18.5 41 678Temp [C] when Envirofit observed as on 42.04 11.39 20 69.5 76

Note*: In the trimmed sample, observations are dropped if ‘lit’ and more than two degrees below hourly meanair temperature, or if observed as ‘not lit’ but more than two degrees above the hourly maximum air temperature.The trimmed sample removes 11.0% (184 out of 1668) and 3.6% (28 out of 782) of the observation samples for thethree stone fire and Envirofits, respectively.

5

Table 3Probability (odds ratio) that three stone fire (TSF) or Envirofit (ENV) is lit: full versus trimmed samples

(1) (2) (3) (4)TSF TSF trimmed ENV ENV trimmed

SUMs Temperature [C], t 1.13*** 1.37*** 1.08 1.46***(0.03) (0.06) (0.06) (0.13)

SUMs Temperature [C], t-1 0.99 0.95 0.99 1.09(0.06) (0.07) (0.06) (0.08)

SUMs Temperature [C], t+1 0.98 0.98 1.16*** 1.24***(0.03) (0.04) (0.05) (0.07)

SUMs Temperature [C], t-2 0.95 1.03 1.05 0.94(0.05) (0.07) (0.05) (0.05)

SUMs Temperature [C], t+2 1.06** 1.07** 0.96 0.90**(0.03) (0.03) (0.03) (0.05)

Control: Hour of Day Yes Yes Yes Yes

Observations 1,184 1,052 366 356Pseudo R-squared 0.167 0.403 0.445 0.696Correctly Classified 83.11% 89.26% 90.71% 93.82%

Robust standard error exponentiated form in parentheses*** p<0.01, ** p<0.05, * p<0.1

Note: Columns (1) and (3) use the full observation sample. Columns (2) and (4) are trimmed foroutliers, meaning observations are dropped if ‘lit’ and more than two degrees below hourly meanair temperature, or if observed as ‘not lit’ but more than two degrees above the hourly maximumair temperature.

Note: Standard errors clustered at stove level.

Table 4Candidate regression specifications: all specifications include temperature at time [t] and control for

hour of the day

Three Stone Fire EnvirofitSpec Leads Lags Slope2 if

Slope>0Pseudo R2 AIC % Correctly

PredictedN Pseudo R2 AIC % Correctly

PredictedN

(1) 0 0 - 0.37 984.01 87.49 1479 0.60 212.05 96.05 735(2) 1 1 - 0.41 649.04 89.16 1052 0.68 120.29 94.94 356(3) 1 1 Yes 0.42 649.44 89.07 1052 0.69 121.55 95.22 356(4) 2 2 - 0.42 648.29 89.26 1052 0.70 120.14 93.82 356(5) 2 2 Yes 0.42 648.49 89.35 1052 0.70 121.61 94.10 356(6) 3 3 - 0.42 650.87 89.15 1051 0.72 117.33 94.93 355(7) 3 3 Yes 0.42 651.19 89.34 1051 0.72 118.49 94.93 355(8) 4 4 - 0.42 653.48 89.05 1050 0.73 116.33 94.93 355(9) 4 0 - 0.39 865.75 86.98 1260 0.69 149.97 94.57 516(10) 0 4 - 0.40 772.18 89.36 1269 0.61 183.79 95.30 574

Note: Classified as correctly predicted if predicted Pr(D) >= 0.5: True D defined when stove use observed as ‘on’

Note: Control for hour of the day is a dummy variable for each hour of the day stoves were observed in use.

Akaike information criterion: defined as AIC = 2k-2ln(L), where k is number of parameters and L is the maximized value of thelikelihood function. Given a set of candidate models for a set of data, the preferred model is the one with the minimum AIC.

6

Table 5Out of sample predictive accuracy: three stone fire and envirofit classification tables for in-sample

predictions and out-of-sample predictions

Three Stone Fire Envirofit

In-Sample Out-Sample In-Sample Out-SampleSensitivity: classified as ‘on’ when observed as ‘on’ 62.04 49.07 80.00 67.86Specificity: classified as ‘off’ when observed as ‘off’ 97.09 97.45 98.65 95.09Positive predictive value: observed ‘on’ when classifed ‘on’ 84.81 82.81 92.31 70.37Negative predictive value: observed ‘off’ when classified ‘off’ 90.70 88.45 96.05 94.51False positive rate given observed as ‘off’ 2.91 2.55 1.35 4.91False negative rate given observed as ‘on’ 37.96 50.93 20.00 32.14False positive rate given classified as ‘on’ 15.19 17.19 7.69 29.63False negative rate given classified as ‘off’ 9.30 11.55 3.95 5.49Overall correctly classified (percentage) 89.81 87.78 95.51 91.10Sample size 520 540 178 191

Note: Unless noted cell contents to be read as probabilities out of one hundredNote: Classified as ‘on’ if predicted probability >= 0.5Note: For this exercise, the samples of observations are randomly split into two groups, the logistic regression is run on one of the twogroups, then based on the coefficients of that regression, the prediction is made to the other (out-of-sample) group. Running a classificationtable on the in-sample and out-of-sample groups shows very little difference between the two samples, suggesting the technique is robustout-of-sample. When the samples were randomly split in half the groups were equally sized (three stone fire groups sized 742/742 andEnvirofit groups sized 377/377), however the sample size of the groups presented in this table is different because various observations dropout of the regression if lacking temperature data for any of the leads and lags. The observation sample is trimmed for outliers, observationsare dropped if ‘lit’ and more than two degrees below hourly mean air temperature, or if observed as ‘not lit’ but more than two degreesabove the hourly maximum air temperature.

Table 6How well do temperature readings generated from stove use monitors predict observations of stove use

Three Stone Fire Envirofit

Sensitivity: classified as ‘on’ when observed as ‘on’ 56.54 77.59Specificity: classified as ‘off’ when observed as ‘off’ 97.61 96.98Positive predictive value: observed ‘on’ when classifed ‘on’ 85.82 83.33Negative predictive value: observed ‘off’ when classified ‘off’ 89.79 95.70False positive rate given observed as ‘off’ 2.39 3.02False negative rate given observed as ‘on’ 43.46 22.41False positive rate given classified as ‘on’ 14.18 16.67False negative rate given classified as ‘off’ 10.21 4.30Overall correctly classified (percentage) 89.26 93.82Sample size 1052 356

Note: Unless noted cell contents to be read as probabilities out of one hundredNote: Classified as ‘on’ if predicted probability >= 0.5Note: The observation sample is trimmed for outliers, observations are dropped if ‘lit’ and more than two degreesbelow hourly mean air temperature, or if observed as ‘not lit’ but more than two degrees above the hourly maximum airtemperature.

7

Table 7How well do temperature readings generated from stove use monitors predict observations of three stone

fires in use: classification tables comparing algorithms

Three Stone Fire

Over 32 Over 34 Over 36 Over 38 RMM adj RMM LogitSensitivity: classified as ‘on’ when obs as ‘on’ 53.0 44.1 36.8 31.4 40.0 41.3 56.5Specificity: classified as ‘off’ when obs as ‘off’ 83.2 92.0 98.0 99.1 64.0 87.5 97.6Positive predictive: obs ‘on’ when classifed ‘on’ 46.0 59.9 83.5 90.8 23.0 47.1 85.8Negative predictive: obs ‘off’ when classified ‘off’ 86.8 85.9 85.2 84.3 79.8 84.7 89.8False positive rate given observed as ‘off’ 16.8 8.0 2.0 0.9 36.0 12.5 2.4False negative rate given observed as ‘on’ 47.0 55.9 63.2 68.6 60.0 58.7 43.5False positive rate given classified as ‘on’ 54.0 40.1 16.5 9.2 77.0 52.9 14.2False negative rate given classified as ‘off’ 13.2 14.1 14.8 15.7 20.2 15.3 10.2Overall correctly classified (percentage) 76.8 81.9 85.0 84.8 58.9 77.7 89.3Pseudo R-squared 10.3 13.4 18.4 17.7 0.1 7.7 41.8

Note: Each column represents a logistic regression with the trimmed sample of observations as the dependent variable and the column header asthe independent variable, cell contents are classified as ‘on’ if predicted probability >= 0.5, the psuedo R-squared is reported for each regression.As an example, the first column is the observation sample regressed on a binary variable registering ‘on’ when the temperature is greater than 32Celsius, and ‘off’ when 32 Celsius or below. The same is true for the ‘Over 34,’ ‘Over 36,’ and ‘Over 38’ columns. The column titled ‘RMM’ usesthe ‘on’/‘off’ status as determined by the Ruiz-Mercado-Mukhopadhyay et al. algorithm as the independent variable for the logistic regression.The column titled ‘adj RMM’ uses the ‘on’/‘off’ status as determined by the adjusted Ruiz-Mercado-Mukhopadhyay et al. algorithm (allowingthe stove to remain ‘on’ if temperatures are very hot even if an exit slope has triggered, and switching ‘off’ if cooling and approaching ambienttemperature even if exit slope not triggered) as the independent variable. The column titled ‘Logit’ uses the logistic regression developed in thispaper where the observations of stove use are regressed against temperature at time t, and two lead and lag temperature readings, and the hourof day. In cases where the regression does not run due to insufficient variability or colinearity, the classification tables are calculated manually,and the pseudo R-squared is reported as n/a. All cells are to be read as probabilities out of one hundred.Note: The observation sample is trimmed for outliers, observations are dropped if ‘lit’ and more than two degrees below hourly mean airtemperature, or if observed as ‘not lit’ but more than two degrees above the hourly maximum air temperature.

Table 8How well do temperature readings generated from stove use monitors predict visual observations of

Envirofit use: classification tables comparing algorithms

Envirofit

Over 32 Over 34 Over 36 Over 38 RMM adj RMM LogitSensitivity: classified as ‘on’ when obs as ‘on’ 57.9 52.6 48.7 43.4 56.6 63.2 77.6Specificity: classified as ‘off’ when obs as ‘off’ 95.4 98.5 99.9 100.0 87.5 88.5 97.0Positive predictive: obs ‘on’ when classifed ‘on’ 58.7 80.0 97.4 100.0 33.6 38.1 83.3Negative predictive: obs ‘off’ when classified ‘off’ 95.3 94.9 94.6 94.0 94.7 95.5 95.7False positive rate given observed as ‘off’ 4.6 1.5 0.1 0.0 12.5 11.5 3.0False negative rate given observed as ‘on’ 42.1 47.4 51.3 56.6 43.4 36.8 22.4False positive rate given classified as ‘on’ 41.3 20.0 2.6 0.0 66.4 61.9 16.7False negative rate given classified as ‘off’ 4.7 5.1 5.4 6.0 5.3 4.5 4.3Overall correctly classified (percentage) 91.6 93.9 94.7 94.3 84.4 85.9 93.8Pseudo R-squared 27.0 32.2 36.7 n/a 14.4 19.6 69.6

Note: Each column represents a logistic regression with the trimmed sample of observations as the dependent variable and the column header asthe independent variable, cell contents are classified as ‘on’ if predicted probability >= 0.5, the psuedo R-squared is reported for each regression.As an example, the first column is the observation sample regressed on a binary variable registering ‘on’ when the temperature is greater than 32Celsius, and ‘off’ when 32 Celsius or below. The same is true for the ‘Over 34,’ ‘Over 36,’ and ‘Over 38’ columns. The column titled ‘RMM’ usesthe ‘on’/‘off’ status as determined by the Ruiz-Mercado-Mukhopadhyay et al. algorithm as the independent variable for the logistic regression.The column titled ‘adj RMM’ uses the ‘on’/‘off’ status as determined by the adjusted Ruiz-Mercado-Mukhopadhyay et al. algorithm (allowingthe stove to remain ‘on’ if temperatures are very hot even if an exit slope has triggered, and switching ‘off’ if cooling and approaching ambienttemperature even if exit slope not triggered) as the independent variable. The column titled ‘Logit’ uses the logistic regression developed in thispaper where the observations of stove use are regressed against temperature at time t, and two lead and lag temperature readings, and the hourof day. In cases where the regression does not run due to insufficient variability or colinearity, the classification tables are calculated manually,and the pseudo R-squared is reported as n/a. All cells are to be read as probabilities out of one hundred.Note: The observation sample is trimmed for outliers, observations are dropped if ‘lit’ and more than two degrees below hourly mean airtemperature, or if observed as ‘not lit’ but more than two degrees above the hourly maximum air temperature.

8

Table 9Summary statistics: predicted hours cooked per day for each three stone fire (TSF) and each Envirofit in

the time period one week prior to and after the approximately 72 hour long kitchen performance test(KPT)

Predicted Daily Hours Cooked

Primary TSF Secondary TSF Envirofit 1st Envirofit 2ndMean SD N Mean SD N Mean SD N Mean SD N

Week Before Final KPT 8.87*** 7.54 98 9.82 8.93 22 3.84** 3.50 105 2.65*** 2.84 88Final KPT Week 7.43 7.62 98 8.99 8.19 22 4.59 3.27 105 3.48 3.07 88

Final KPT Week 7.40 7.56 100 8.99 8.19 22 4.56 3.24 111 3.24 2.95 95Week After Final KPT 10.14*** 8.41 100 9.79 8.34 22 3.35*** 3.21 111 2.44*** 2.15 95

Note: Stove usage during KPT verses stove usage over the same duration and days of the week for the same stove types in the prior and following

weeks. We use paired t-tests relative to the measurement (KPT) week to examine statistical significance. The difference in use is statistically

different than zero at the following levels *** p<0.01, ** p<0.05, * p<0.10. The differences in hours cooked per day in the week prior to and

after the KPT are weakly statistically significantly different than zero for the primary three stone fire (p<0.10). There is no statistically significant

difference in daily hours cooked in the week prior to and after the KPT for neither the secondary three stone fire, the first Envirofit, nor the second

Envirofit.

Table 10Test for Hawthorne effect: primary three stone fire verses combined hours cooked on two Envirofits

during the approximately 72 hour long kitchen performance test (KPT) “week”

Predicted Daily Hours CookedBoth Envirofit Stoves Mean SD N

One Week Before Final KPT 6.08*** 3.91 48Final KPT Measurement Week 8.14 3.67 48One Week After Final KPT 5.60*** 3.05 48

Primary Three Stone Fire Mean SD NOne Week Before Final KPT 8.07* 7.39 48Final KPT Measurement Week 6.95 7.14 48One Week After Final KPT 10.45*** 8.40 48

Note: Stove usage during KPT verses stove usage over the same duration and days of the

week in the prior and following weeks. We use paired t-tests relative to the measurement

(KPT) week to examine statistical significance. The difference in use is statistically dif-

ferent than zero at the following levels *** p<0.01, ** p<0.05, * p<0.10. Each of the 48

households in this sample has SUMs readings for two Envirofit stoves and SUMs readings

for the primary three stone fire. The week prior to and after the KPT are not statistically

significantly different than zero for Envirofit, however they are statistically significantly

different than zero for the three stone fire (at p<0.01).

9

Table 11Regressions testing for Hawthorne effect: estimates of effects of the presence of measurement team in

primary three stone fire (TSF) usage and combined Envirofit usage, the coefficients represent the changein hours cooked per day compared to hours cooked per day in the measurement week

Primary TSF Combined Envirofit(1) (2) (3) (4)

Week prior to and after measurement week 2.26*** -2.40***constrained to be equal (0.57) (0.46)

Week prior to measurement week 1.62** -2.02***(0.66) (0.56)

Week after measurement week 2.88*** -2.71***(0.76) (0.50)

Household fixed effects Yes Yes Yes YesObservations 316 316 229 229R-squared 0.82 0.82 0.79 0.80Household clusters 118 118 89 89

Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1

Note: The unit of analysis is a measurement “week” (approximately 72 hours) at a household. Thespecification in columns 1 and 3 imposes that the weeks prior to and after the measurement week areequal. The specification in columns 2 and 4 tests usage in the week prior to and after the measurementweek separately. The coefficient estimates in columns 2 and 4 are both jointly significantly not equalto zero (joint F test significant at 1% level in both cases).

Note: Standard errors clustered at household level.

10