Jayaraman, Rajshri; Simroth, Dora
Working Paper
The impact of school lunches on primary school
enrollment: Evidence from India's midday meal scheme
ESMT Working Paper, No. 11-11
Provided in Cooperation with:
ESMT European School of Management and Technology, Berlin
Suggested Citation: Jayaraman, Rajshri; Simroth, Dora (2011) : The impact of school lunches on
primary school enrollment: Evidence from India's midday meal scheme, ESMT Working Paper, No.
11-11, European School of Management and Technology (ESMT), Berlin,
https://nbn-resolving.de/urn:nbn:de:101:1-201201176365
This Version is available at:
https://hdl.handle.net/10419/96555
Standard-Nutzungsbedingungen:
Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen
Zwecken und zum Privatgebrauch gespeichert und kopiert werden.
Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich
machen, vertreiben oder anderweitig nutzen.
Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten,
gelten abweichend von diesen Nutzungsbedingungen die in der dort
genannten Lizenz gewährten Nutzungsrechte.
Terms of use:
Documents in EconStor may be saved and copied for your personal
and scholarly purposes.
You are not to copy documents for public or commercial purposes, to
exhibit the documents publicly, to make them publicly available on the
internet, or to distribute or otherwise use the documents in public.
If the documents have been made available under an Open Content
Licence (especially Creative Commons Licences), you may exercise
further usage rights as specified in the indicated licence.
ESMT Working Paper
ESMT | European School of Management and Technology
ISSN 1866-3494
11–11
December 06, 2011
THE IMPACT OF SCHOOL
LUNCHES ON PRIMARY
SCHOOL ENROLLMENT
EVIDENCE FROM INDIA’S MIDDAY MEAL SCHEME
RAJSHRI JAYARAMAN, ESMT
DORA SIMROTH, ESMT
* Contact: Rajshri Jayaraman, ESMT, Schlossplatz 1, 10178 Berlin,
Phone: +49 (0) 30 21231-1293, rajshri.jayaraman@esmt.org.
+ We gratefully acknowledge funding from DFG Grant JA 1675/2-1 and thank Arun Metha
and Naveen Bhatia for kindly sharing their DISE data. We also thank Farzana Afridi, Jean
Drèze, Elena Nikolova, Imran Rasul, Debraj Ray, Matthias Schündeln, Rohini Somanathan,
Francis de Véricourt, various conference and seminar participants at ESMT, NEUDC,
Goethe University Frankfurt, IFMR Chennai, ISI Delhi, Paris School of Economics, CESifo,
and Oxford University for their comments and suggestions.
Copyright 2011 by ESMT European School of Management and Technology, Berlin, Germany,
www.esmt.org.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system, used in a spreadsheet, or transmitted in any form or by any means - electronic,
mechanical, photocopying, recording, or otherwise - without the permission of ESMT.
Abstract
The impact of school lunches on primary school enrollment:
Evidence from India’s midday meal scheme
+
Author(s):* Rajshri Jayaraman, ESMT
Dora Simroth, ESMT
At the end of 2001, the Indian Supreme Court issued a directive ordering states to
institute school lunches – known locally as “midday meals” – in government primary
schools. This paper provides a large-scale assessment of the enrollment effects of
India’s midday meal scheme, which offers warm lunches, free of cost, to 120
million primary school children across India and is the largest school feeding
program in the world. To isolate the causal effect of the policy, we make use of
staggered implementation across Indian states in government but not private
schools. Using a panel data set of almost 500,000 schools observed annually from
2002 to 2004, we find that midday meals result in substantial increases in primary
school enrollment, driven by early primary school responses to the program. Our
results are robust to a wide range of specification tests.
Keywords: primary school enrollment, school lunches, natural experiment, ITT
2
1. Introduction
Education is thought to be central to economic development. Beneficial in and of
itself, it is also viewed as a major contributor to human capital, leading to higher
productivity and living standards. Primary education is thought to be associ at ed with
especially high returns.
1
Its importance is enshrined in the Mill en n i um Development
Goals (MDGs), which call for universal primary education by 2015.
In fact, primary education is far from universal and this MDG remains elusive.
UNICEF (2008), the agency responsible for track i n g progress on this MDG , estimates
anetprimaryschoolenrollmentrateindevelopingcountriesof84percent;thisisalso
its estimated average for India. In view of this, governments across the developing
world have instituted a wide range of pol i ci e s aimed at encouraging school enrollment.
School lunches are one such policy. They are thought to increase enrollment through
two main channels.
2
First, they lower the cost of schooling, thereby providing an
implicit subsidy to parents. Second, by improving child nutrition school lunches are
thought to foster learning, thereby increasing the returns to education. School feeding
programs are popular in the developing world and beyond. Despite a large empirical
literature on the relationship between feeding programs and educational attainment,
reviewed in Bundy et al. (2009), there have, to the best of our knowledge, been no
large-scale assessm ents of their causal impact on enrollment (Adelman et al. 2007,
p.2).
This paper fills this gap by providing a large-scale impact assessment of India’s free
school lunch program – known locally as the “midday meal” scheme – on primary school
enrollment. India’s midday meal scheme is the lar gest school nutrit i on program in the
world. In 2006, it provided lunch to 120 million children in government primary schools
every school day (Kingdon 2007). We expl o i t a quasi-natural experim ent in order to
identify the causal impact of midday meals on primary school enrollment using a large
school-level panel data set, th e District Information System for Education (DISE).
Our sample cont ai n s almost 500,000 primary schools in 15 major states across India,
observed annually in academic years 2002/3, 2003/4 and 2004/5 (referred to hereafter
as 2002, 2003 and 2004).
1
Psacharopoulos and Patrinos (2002) estimate private returns to primary education of over 25%, while
Duflo (2001) finds in a developing country context between 6.8 and 10.6 % returns to e duc at ion from
primary school.
2
These ar e widely documented. See, for example, PROBE (1999), Dr`eze and Goyal (2003) and Kremer
and Vermeersch (2004).
3
Identification of a causal effect comes from state-level variation in the implementa-
tion of a 2001 Ind i an Supreme Court directive, which was instigated by public interest
litigation aimed at redressing starvation. The di r ec ti ve ordered states to institute mid-
day meals in government primary schools (referred to hereafter as public schools). Prior
to 2001, only two states had universal public primary school midday meal provision.
3
Over the subsequent three years, however, state governments across India introduced
midday meals.
Two main sources of variation are used in assessing the impact of midday meals: the
date on which states intro duced midday meals in primary schools, and the fact that
(in accordance wi t h the Supreme Court directive) they were introduced in public, but
not private primary schools. Since the directive was addressed nation-wide, concerns
regarding program placement bias are alleviated. Moreover, staggered implementation
at the state level in public but not private schools allows us to treat all private schools
as well as public schools in st at es not yet implementing the program, as a quasi-control
group for public schools in states which introduced midd ay meals.
We find that midday meals lead to large and statistically significant increases in
primary school enrollment. Our main triple difference intent to treat (ITT) estimates
point to a statistically significant 13% increase in primary school enrollment, amounting
to around 14 additional students in each primary school. If newly enrolled children were
all of primary-school age (6-10 years), th i s would imply that mi d d ay meals increased
the net primary school enrollment rate from 84% (in 2002) to 95%.
The enrol lm ent response to midday meals, althou gh positive across all grades, is
driven by a large and statisti cal l y significant response in grade 1. In grade 1 , enrollment
increases by approximately 21%. The magnitude of the estimate reflects the fact that
grade 1 absorbs all new enro l l m ents, which includes both under-aged children (typically
5-year-olds) as well as children over 6 years of age. In fact, since the net enrollment
rate in grade 1 is likely to have been close to 100% in 2002, older and younger children
are likely to account for most of the grade 1 enrollment increase.
In higher grades the response remains positive, with smaller poi nt estimates and
statistically insignificant coeffi ci ents across all specifications. In part, this pattern
reflects the fact that whereas grade 1 picks up all new enrollments, upper grades only
pick up dropouts: a child can only enroll in grades G =2, 3, 4, 5ifheorshewas
enrolled in grade G − 1. Since dropout rates in grades 2-5 are low (see Table 3), there
is limited scope for midday meals to increase enrollment in the first few years of the
programme exp osure studied here.
3
These two states were Tamil Nadu and Gujarat. A third state, Ke ral a, had an opt-in program.
4
The decreasing margi n al response to midday meals i n higher grades is also, however,
likely to reflect the fact that the relative value of the implicit subsidy declines with
grade, since direct costs (such as textbooks or uniforms) as well as opportuni ty costs
(in terms of the value of household production or wage income) are larger in higher
grades, while the value of the subsidy remains constant.
These results ar e robust to a wide range of specification tests. We demonstrate
that program timing is not associated with different initial schooling input levels, or
trends in enrollment outcomes. We further provide robustness checks which indicate
that our results are not driven by the timing of implementation. Our main results are
virtually unchanged for a matched sample of public and private schools; and we provide
some evidence th at enrollment in private schools did not respond to midday meal
introduction, su g ges ti n g that private schools are a legitimate control grou p . Neither
were there contemporaneous changes in relative inputs in public versus priv ate schools,
and this alleviates concerns regarding confounding policy changes.
In addition to DISE, we exploit cross-sectional household and school sur vey data
from the Indian Human Development Survey (IHDS) 2005. Although our results are
only suggestive, given the cross-sectional nature of the data as well as its timing (af-
ter midday meals were intro duced across India), these data nevertheless permit us to
extend our analysis in two ways. First, we explore whether the positive enrollment
response associated with midday meal provision is driven by more disadvantaged seg-
ments of the population, who are both least l i kely to be en ro l l ed , and most likely to
be responsive to a food subsidy. The data confirm that more disadvantaged socio-
economic groups display the largest enrollment responses.
Second, we examine whether midday meal provision is associated with improved
schooling outcomes on two additional dimensions, nam el y attendance, and learning
(as measured by separate test scores for reading, writing and mathematics among 8 to
11 year-olds.) We find that midday meals are associated with improved attendance.
This makes sense given that scho ol lunches are consumed on school premises at noon, so
children only benefit from this subsidy to the extent t h a t they actually attend school.
At the same time, midday m eal s are not associated with improved learning. This
indicates that the positive enrollment response we observe in our quasi-experimental
setting may be driven by the implicit subsidy channel rather than by nutrition-induced
improvements in learning.
5
This paper contributes to a growing l i t erat u r e which relies on natural experiments
to assess the impact of schooling policies on schooling outcomes.
4
Within the natural
experiments literature, th i s pa per most closely follows Duflo (2001), who examines th e
effect of a large school building program in Indonesia on educational attainment and
wages, and Chin (2005) who assesses India’s Operation Blackboard (which introduced
additional teachers), in that the natural experiment here directly concerns variation in
the policy variable.
Our paper also complements a recent literature, which uses randomized trials to
evaluate the effect of scho ol feeding programs on school participation. Powell et al.
(1998), Jacoby E. and E. (1996) and (in pre-scho o l ) Kremer and Vermeersch (2004)
each find increased participation resul t i n g from breakfast provision in Jamaica, Peru
and Kenya, respectively. And Kazianga et al. (2009) find that school lunches as well
as take home rations increase new enrollment for girls by 5 to 6 percentage points.
Identification in our quasi-experimental set t in g is unlikely t o be as clean as it is in
these carefully conducted randomized trials. Nevertheless, there are two strengths to
our approach. First, it enables an impact assessment of the world’s largest nutrition
program in a country which has the largest number of out-of-school children in the
world. Second, its large-scale nature allays concerns about generalizability , to which
smaller-scale studies are sometimes prey.
Finally, our findings generally corroborate the positive enrollment effect documented
in smaller-scale non-experimental survey-based assessments of midday meal provision
in India. Our grade 1 enrollment effect is similar to Dr`eze and Goyal (2003), who find
an 18%, 11%, and 14% increase in grade 1 enrollment in their Rajasthan, Chhattisgarh
and Karnataka villages, respectively; but substantially smaller than the 36% increase
in grade 1 enrollm e nt found by Jain and Shah (2005) in their 70 Madhya Pradesh
schools. The 13% primary school enrollment response we fi n d in our DISE data, is
also considerably smaller than the 23% increase in primary scho ol enrollment found
by Khera (2002) in her 63 Rajasthan schools. Since previous small-scale studies have
measured the effect of midday meal provision – which is likely to be an endogenous
outcome at the local level – often in relatively under-developed villages (such as Madhya
Pradesh and Rajasthan), our results suggest that t h e problem of purposi ve placement
may have resulted in an upward bias of previous estimates.
5
Anaivecomparisonof
4
See Kremer and Glewwe (2005) for a review of thi s literature and Hanushek (1995) for a c ri t i q ue of
earlier studies.
5
In this sense, our results corroborate Afridi (2007), who exploits staggered implementation using a
double difference strategy in 41 Madhya Pradesh villages, and finds a similarly muted response in
enrollment and attendance.
6
our estimate with these previous studies does suggest, however, that the magnitude of
the resp onse we observe here is entirely plausible.
The paper proceeds as follows. Section 2 provides background regarding the Supreme
Court d i r ect i ve and the midday meal scheme, together with a discussion of its imple-
mentation and content. Section 3 describ es the DISE data, and Section 4 presents our
empirical strategy. Our empir i cal resu l ts usin g DISE data, including specification tests,
are presented in Section 5. Section 6 presents an extension regarding heterogeneous
responses, attendance, and learning associated with midday meal provision using IHDS
data, and Section 7 concludes.
2. Midday Meals
2.1. Background. In India, pri m ar y school education typically covers grades 1-5, and
is the joint responsibility of central and state governments. The central government
generally issues guidelines and provides funding, but policy implementation is a state-
level decision. The central government has a long-standing commitment to the provi-
sion of midday meals. As early as August, 1995, The National Program of Nutritional
Support to Primary Education mandated cooked meals in all public primary schools.
Not a single state responded to this universal mandate. (Keral a responde d , but only
by offering an opt-in program which resulted in partial coverage in public primary
schools.) Two states had, by this time, long established universal midday meal provi-
sion in public primary sch ools. Tamil Nadu, a state in the Southeast, was a pioneer.
Its state-wide midday meal program was launched i n 1982 at th e personal behest of its
then-Chief Minister M.G. Ramachandran, who cited as his motivation early childhood
experiences with hun g er (Harriss 1991, p.10). Gujarat, a state in Central-west Indi a ,
followed suit in 1984.
6
Between 2002 and 2004, however, most Indian states instituted universal midday
meals in public primary schools. This wave was precipitated by a severe drought
that hit several states in 2001.
7
Reports of drought-related starvation deaths in the
press instigated a public interest litigation. In April, 2001 the People’s Union for Civil
6
Most other st at es provided “dry rat ion s” to enrolled children who attended school, which typically
comprised 3 kg. per month of raw wheat or rice grains (depending on local consumption habits).
By many accounts, the distribution of these dry rations was sporadic, of low quality and condit i on al
attendance r eq u i re ments went unenforced (see for example, PROBE (1999)). Moreover, there is
evidence of extensive leakage in this dry rations program, in the sense that children enrolled in private
schools also received dry rations (see, for example Muralidhar an (2006)).
7
There were 7 drought-affected states in 2001: Gujarat, Rajasthan, Maharashtra, Orissa, Madhya
Pradesh, Chhattisgarh, and Andhra P rad es h (Down t o Earth, Vol. 10, Issue 20010615, June 2001).
They include both early and late implementers of midday meals.
7
Liberties (PUCL), Rajasthan, submitted a writ petition to the Supreme Court pointing
out that “while on the o n e hand the stocks of food grains in the country are more th a n
the capacity of storage facilities, on t h e other there are reports from various states
alleging starvation deaths.”
8
The PUCL documented that, despite their protests to the
contrary, states could in fact afford to widen a number of statu t or y food and nutrition
programs, including the midday meal scheme in schools. The writ urged the court to
instruct the Government to release public food stocks, arguing that the right to life
under Article 21 of t h e Indian Constitution included the right to food.
9
The petition
has culminated in protracted public interest li ti g a ti o n which is yet to be concluded.
10
Nevertheless, on November 28, 2001 the Supr em e Court issued an interim ord er
directing states to introduce cooked midday meals, i.e. a warm school lunch, in all
public schools, but not in private schools. More specifically, the directive said, “Every
child in every government and government-assisted school should be given a prepared
midday meal”.
2.2. Implementation. Implementation of th i s and other Supr em e Court directives
are left to the relevant executive branch of government (Desai and Muralidhar 2000).
In this case, state governments were respo n si b l e for introducing midday meal s.
11
To ex-
amine the effects of this policy change on schooling outcomes, we gathered information
on the policy implementation in public schools from state documents and then cross-
checked this information using at least two (and usually, more) independent sources
(see Appendix A fo r meal contents by state and li st of state documents, independent
monitors, auditors, field surveys and news articles).
12
The result of this ex er ci s e is described in Table 1. Column 1 lists the 15 states which
are covered in the data for our school-level analysis, and Column 2 indicates the month
8
Rajasthan PUCL Writ in Supreme Court on Famine Deaths, PUCL Bulletin, November 2001.
9
Article 21 of the Constitution of India is entitled “Protection of life and personal liberty”. It states,
in its entirety, “No person shall be deprived of his life or personal liberty except according to procedure
established by law.”
10
PUCL vs Union of India and Others, Writ Petition [Civil] 196 of 2001. The Right To Food Campaign
has been closely monitoring the developments associated with this case and maintains an extremely
informative website at www.righttofoodindia.org.
11
As Gauri (2009, p.2) notes, “courts do not and cannot enforce many of their broad directives”.
For this reason, estimating the intent-to-treat by using the Supreme Court directive as a source of
exogenous variation at the national level is not particularly meaningful.
12
In the case of Andhra Pradesh , there was a discrepancy between independent sources and state
documents. The Comptroller and Auditor General of India claimed November 2003 implementation
and a best practice report of NUEPA put the date at 2001. We chose January 2003, as this was t he
date provided by the state documents, 6 reports of t h e Commissioner of Ind i a an d numerous press
reports. Dropping Andhra Pradesh from the sample or changing its imp l em entation year to 2001 or
2004, has no quali t at i ve bearing on our estimates.
8
and year in which the corresponding state is documented to have introduced a midday
meal. Note that th i s does not necessarily mean that midday meals were in fact on
the ground in every public school in the state.
13
Since, as we elaborate in Section 3,
enrollment figures are record ed as on September 30th of any given year, we regard a
state as having instituted a midday meal policy if its implementation took place before
September 30th in the corresponding academic year. The last column of the table
documents the year of initial treatment according to this criterion.
Data from three additi o n al states – Jharkhand, Ker al a and West Bengal – were
av ailable from DISE but are not used in our main analysis due to poor documentation
of partial implementation, and potential purposive placement.
14
Finally, also due to
worries of p u r posive placement, we dropped from our main sample 28 districts (in
2001 India had 593 districts) from Assam, Bihar, Karnataka and Orissa as well as all
tribal blocks from Madhya Pradesh in which the midday meal scheme was implemented
earlier. (We use the short hand “pilot” to refer to these tribal Madhya Prad esh regions
too.) Nevertheless, as we sh ow in our specificat i on checks, the ad d i ti o n of these pilot
regions, as well as Jharkhand, Kerala and West Ben g al does not change the results.
The wide geographic coverage o f our data an d state-level variation in th e date of
implementation, evident in Table 1 , ar e graphically displayed in Figure 1. Together,
the states covered in our data house over 80% of the Indian population according to
the 2001 Census of India. Pertinently, the geographic pattern in terms of timing of
implementation is mixed. For example, pioneers (Tamil Na d u and Gujarat) come from
South and Central India. Early implementers include not only the “usual suspects”
in Southern India, Andhra Pradesh and Karnataka, but also surprising candidates
13
Data limitations make it difficult to verify the proportion of public schools which actual l y provided
midday meals during our observation p e ri od. Hous eh old surveys are not con du ct e d annually and
rarely pose a midday meal consumption question. Deaton and Dr`eze (2006) assert, moreover, that
at least in the National Sample Surveys (NSS), midday meal consumption is underreported. The
school survey data from IHDS 2005 (which we descr i be in Section 6) indicate, however, that in states
which we classi fy as having been treated by 2005, 84% of schools covered in the pub li c school survey
are reported as providin g midday meals. This suggests that the vas t majority of schools which we
consider as treated in our ITT frame work are, in fact, treated.
14
Jharkhand i ns t i tut ed midday meals in November 2003 as a pilot project, but we are unable to
ascertain where these pilots were imp l eme nted. We could also not verify when full coverage was
announced as having been achieved. West Ben gal start ed a midday meal roll out in January 2003.
We could not find documentation for the placement, and full coverage is yet t o be achieved. Kerala,
as mentioned earlier, allows schools to opt-in to the midday meal scheme, and this raises concerns of
selection bias.
9
Note:!"#$%!&'(!)*($+,%!,#*!-*.-/'(#$+!+.0*/'-*!.1!2345!677686779:!³0LVF:´!/*1*/%!,.!($;.,!
/*-$.<%=!>*/';'=!?#'/@#'<)!'<)!A*%,!B*<-';:&
B$#'/!
C$&'+#';!
D/')*%#!
C'/E'<'!
F,,'/'<+#';!
F,,'/!D/')*%#!
G')#E'!D/')*%#!
G'#'/'%#,/'!
H<)#/'!
D/')*%#!
>'/<','@'!
"'&$;!!
I')J!
K/$%%'!
?#'/@#'<)!
A*%,!
B*<-';!
L'M'%,#'<!
NJM'/',!
677O!
6779!
677P!
QRST9=6776U!
G$%+:!
I.!2','!
Figure 1. DISE Data Coverage and Midday Meal Implementation
like Rajasthan and Chhattisgar h . The so-called “BIMARU” states include late im-
plementers (BIhar), middle implementers (MAdhya Prad es h and Uttar Pradesh) and
early implementers (Rajasthan).
15
Idiosyncratic timing in implementation has been attributed to successful p ressu r e
applied by civil soci ety. In particular, the initial 6-m o nth deadline set by the Supreme
Court was without exception breached, with states complaining that they did not have
sufficient fu n d in g to implement the policy. This excuse was widely dismissed by the
15
The acronym comes from its resemblance to the Hindi word “bimar”, meaning sick. These 4 states
have am ong the lowest domestic products in the country. The fact that Bi har and Assam, two “late”
implementers in our sample, also have rather poor economic educational characteristics does n ot
obviously detract from our claim of idiosyncratic timing in light of the fact that Punjab and West
Bengal – two states which are not marked on this map but have reasonably advanced economic and
educational outcomes – also had not fu ll y implemented midday meals by 2004.
10
media, two Supreme Court commissioners, and the activist community, who instead
blamed the “lack of political, bureaucratic and societal will” for state governments’
recalcitrance (Parikh and Yasmeen (2004); Dr`eze and Goyal (2003) and Zaidi (2005)
make similar claims.) State government inaction spurred grassroots activists, coordi-
nated by In d i a’ s Right to Food Campaign which had grown out of the PUCL’s S u p re m e
Court litigation efforts, to start public mobilization efforts. It was these efforts, sup-
ported by continued monitoring and chastisement on the part of two commissioners
as well as media, which compelled states to comply with the Supreme Court directive
(see Sharma et al. (2006) and Khera (2006)).
2.3. Financing and Content. The m i d d ay meal scheme is a joint undertaking of
central and state governments. During our observation period, the central government
provided financial assistance to cover the cost of food grains and their transport. In
particular, The Food Corporation of India (FCI), an institution set up in 1964 to
support the oper at i on of the central government’s food policies, provided states free
supply of food grains from the nearest of its warehouses. Provision for each student
with 100 grams of wheat or rice per day cost the central government app r oximately
Rs. 1.11 (NPNSPE 2004). In principle, fair average quality of the grains was also
guaranteed, with the FCI being responsible for replacing the grains otherwise. The
transport subsidy to carry the grains from the nearest FCI warehouse to the primary
school was set at a maximum of Rs. 50 per quintal, amounting to an average transport
subsidy of Rs. 0.05 per child per school day.
16
The total value of the central government
subsidy between 2002-2004 therefore amounted to Rs. 1.16 per child per school day.
The Supreme Court’s 2001 directive mandated that midday meals have “a minimum
content of 300 calories and 8-12 grams of protein each day of school; for a minimum
of 200 days a year.” The overall responsibility for implementation of this directive lies
with state governments, who supplem ent the central government’s contributions to
varying degrees.
17
Day-to-day operations lie in the hands of local government bodies,
16
This figure is calculated from NPNSPE (2004, Section 3.4) which states that at the end of 2004, i.e.
after our period of observation, the transport subsidy grew by one third, namely to Rs . 75 per quintal,
which amounted to an average of Rs. 0.08 per child per school day. Following our observation period,
an additional Rs. 1 per child per school day was contributed by the central government towards
cooking costs, comprising cost of i n gre di e nts other than grains, including vegetables, cooking oil, and
condiments, as well as the cost of fuel and wages for personnel.
17
These supplements are non-transparent and poorly documented, but available evidence suggests
that there is no obvious correlation between suppl eme nts and timing of midday meal implementation.
For example, Tamil Nadu (an early implementer) and Andhr a Pradesh (which implemented in 2003)
both contributed Rs. 1 pe r child per day towards cooking costs in 2005, whereas Rajasthan and
Chattisgarh, which implemented earlier than Andra Pradesh, contributed little towards cook i ng cost s
(Secretariat of the Right to Food Campaign 2005).
11
typically village governments (panchayat s) , who sometimes delegate implementation
to local Parent Teacher Associati on s (PTAs) or NGOs.
In practice, the meal itself tends to be a simple affair. At ar ou n d midday children sit
at their plates, which are typically set on the ground, where they are served a cooked
meal prepared on site, usually by a cook who is hired for this purpose. The meal
comprises cooked rice or wheat (depending on the local staple), mixed with lentils or
jaggery, and sometimes supplemented with oil , vegetables, fruits, nuts, eggs or dessert
at the local level (see Appendix A for details on meal content by state) . Eye-witness
accounts (from present company included) note that, although the quality and variety
of the meal varies from district to district or even school to school, children seem to
enjoy their lunch (see, for example, Dr`eze and Goyal (2003)).
3. Da t a
In order to execute a large-scal e evaluati on of the midday meal program we use the
District Information System for Education (DISE), which is the “most comprehensive
information system in the education sector” in India (Ward 2007, p. 291). DISE is
aschool-leveldatasetcoveringgovernment-recognizedelementaryinstitutions.Itisa
joint initiative of the Government of India, UNICEF and the National University of
Educational Planning and Administration (NUEPA), and came into bein g explicitly
because of a lack of reliable statistical databases for education in India (Mehta 2007).
Initiated on a pilot basis in 1995 to monitor schooling inputs and enrollment out-
comes for those distri cts covered by the District Primary School Education Programme
(DPEP), DISE was gradually rolled out to cover non-DPEP districts. Starting from
2002, DISE achi eved coverage of all districts of the 18 states mentioned in Sect io n 2,
where it was initially launched (DISE 2008).
Data is collected annually, and reflects primary scho ol characteristics (such as in-
frastructure and staff) as well as student enrollment as on September 30th of the
respective year.
18
School headmasters answer a nationally standardized school survey
questionnaire. The data is verified and manually checked at various stages from lower
to higher levels of administration. At the cluster level, responses are verified for com-
pleteness and accuracy. The data is then aggregated at the district level, where it
is checked for computational and consistency errors. Further consistency checks take
18
During our observati on period, enrollment data was consistently collected only for grades 1 - 5, and
not for secondary or upper-secondary school. Although age - dis aggregated enrollment, as well as non-
enrollment outcomes such as exam results, attendance, failure, drop-out and readmission questions
were posed in the DISE survey, these data are missing for the vast majority of schools, and are riddled
with measurement error and inconsistencies even when they exist .
12
place at the state level. In addition to these measures, the NUEPA has commissioned
post-enumeration audits through external agencies, so as to verify the accuracy of the
data provided by the school headmasters. In these audits, 5% of schools chosen ran-
domly from at least 10% of districts from each state were thus cross-checked with site
visits (Kaushal 2009). The major findings of these surveys is that the total enroll-
ment figures for primary school a r e overwhelmingly accurate. Systematic errors were,
however, found in responses to questions which were either unclear, or open to subjec-
tive interpretation. Hence, we refrain from using variables whi ch capture qualitative
assessments. For example, rather than construct a variable capturing the quality of
classroom infrastructure, we use the total number of classroom s in the school.
We exploit a three year balanced panel of 491,253 schools over the academic years
2002/03 to 2004/05.
19
We consider public and private primary scho ols. Private schools
in Indian school system parlance are, in the context of our data, “unaided schools”.
What we call publ i c schools in o u r sample are government owned and operated schools;
they are not so-called “government aided” schools. Government aided schools were
dropped since the documentation is opaque as to when and whether these schools were
covered by the midday meal program at the state level. They constituted 4.90% of
the full 2002-2004 data set, and including them in the analysis as either part of the
treatment or quasi-control groups does not alter the results.
Private school s constitute 6.53% of our sample. The distribu t i on of public and
private schools among states in our sample can be seen in Table 2. The former closely
follows the state population distribution.
We estimate enrollment responses separately for grades 1 to 5, as well as for pri-
mary school as a whole. Table 3 furnishes 2002 means of enrollment and of schooling
inputs, which we use in our sp ec i fi cat i o n tests. It indicates that average enroll me nt in
primary scho ols is just above 122 students, with a low average attrition of between 2-3
students per year between grades 2-5. On average, a pri m ary school has about 3 class-
rooms, 1 additional room, 2 teachers, 0.4 non-teaching staff (including para-teachers),
4 blackboard s and 1.6 “trunks” of teaching materials. Just half of the scho o l s have
aplayground,onefifthhaveelectricity,80%ofschoolshavewater,andthemajority
does not have toilets; 97% teach in the vernacular. In our estimations, we control for
19
These are the only years for which data for all DISE districts were made available to us. Prior data
would, however, not have been representative at the state leve l, since survey coverage in previous
years was substantially more limited, and restrict e d overwhelmingly to educationally underdeveloped
districts within each state vis `a vis education.
13
these inputs and also create a matched sample based on these observable schooling
characteristics.
4. Empirical Strategy
4.1. Approach. To study the impact of the midday meal policy on primary school
enrollment, we expl oi t the variation created by its staggered intro d u ct ion in public pri-
mary schools throughout India.
20
We employ an intent-to-treat (ITT) analysis through-
out (see for example, Imbens and Rubin (1997)). In particular, all public schools lo-
cated in a state which has been documented as having impl em e nted the Supreme Court
directive at time t and thereafter (see Table 1) are considered as treated.
This approach has three related merits. First, it is a natural way to analyze a
policy which may be characterized by non-random compliance at the school or village
level. Second, it is useful from a policy perspective since state governments’ budgetary
allocations to midday meals are typically associated with their decision to introduce
the policy even if these allotments are not spent at the local level by non-compliers.
Finally, since DISE does not include information on midday meal implementation at
the school level, we are unable to verify compliance. (In Section 6, we ex p l o i t hous eh ol d
survey data containing information on schools’ midday meal compliance.)
Our aim is to identify the effect of midday meals instituted in public schools (treat-
ment group) by certain states (experimental states). In order to accomplish this, we
need to control for systematic shocks in enrollment outcomes of the treatment group in
experimental states th a t are correlated with, but not due to, the institution of midday
meals. We accomplish this by estimating the following triple d i ffer en ce equation, which
uses private schools as an additional control group:
21
(4.1) Y
ist
= βM DM
ist
+γ
t
+λ
s
+αP ub
i
+δ
1s
(Pub
i
·λ
s
)+δ
2t
(Pub
i
·γ
t
)+δ
3st
(λ
s
·γ
t
)+�
i
20
Broadly speak i ng, our use of staggered implementation as an identification strategy follows Gruber
and Hungerman (2008), who assess the impact on religious participation of the repeal of “Blue Laws”
in U.S. states, and Field (2007) who studies a nation- wi de titling program in Peru.
21
Note that the approach used here is in fact an extension of the triple difference method, in that there
are more than just 3 treatment and control groups (public schools from 15 states and the respective
groups of private schools) and more than just 2 time periods. When extending the triple difference to
the case of multiple groups and time periods, the policy variable is no longer a triple interaction term,
but a policy dummy set to unity for groups and time periods when t h e policy was in place (see Imbens
& Wooldridge, Lecture Notes 10, NBER, 2007). Therefore, the result of the following estimation is
not equivalent to the difference in estimates from two separate double differences, as would be the
case for the standard triple difference. For simplicity, we will refer to our estimates as being triple
difference estimates.
14
where Y
ist
is the log of enrollment, for school i,instates,attimet =2002, 2003, 2004.
In various speci fi cat i on s it pertains to enrollment in grades 1-5 separately, as well as
to total primary school enrollment.
The policy variable MDM
ist
is equal to 1 if the midday meal program was in place
in public school i from state s prior to the September 30th enrollment deadline in
year t,asdescribedinTable1. Thecoefficientβ is the triple difference estimate. It
captures changes in enrollment in public school s following the institution of a midday
meal program.
National trends in enrollment a r e captured throu gh year fixed effects, γ
t
.Statefixed
effects, λ
s
,accountforenrollmentdifferencesacrossstates. ThedummyvariablePub
i
,
which is equal to 1 if school i is a public school and 0 if it is a private school, allows
for different average enrollments in public relative to private schools. The interaction
term Pub
i
· λ
s
permits this average to vary by state, and Pub
i
· γ
t
captures a national
trend in public school enrollment.
The key advantage of this approach is t h at it allows us to account for state specific
shocks over the observation period through state-by-year effects, λ
s
· γ
t
. This is impor-
tant in a federal country such as India, where scho oling policy is largely governed by
states which have not only different levels, but also different trends in economic and
demographic development.
22
There has been much discussion in the literature about the calculation of standard
errors for differences-in-differences estimates, so this is worth commenting on upfront .
Following Bertrand et al. (2004), we cluster standard errors at the state level. How-
ever, as Cameron et al. (2008) point out, this may not resol ve the problem of serial
correlation if the number of clusters is not large, as in our case with 15 states. To a c-
count for this we follow the recommendat i on of Cameron et al. (2008) and wil d cluster
bootstrap the standard errors with 1000 replications (a cluster generalization of the
wild bootstrap for hetero sced a st i c models with equal weights and probability.) The
results are qualitatively identical. For simplicity, we therefore use the cluster-robust
standard errors in all estimations.
22
A double difference strategy would not allow us to dis t in gu is h state-by-year effects from the midday
meal effect. Given state-time heterogeneity in India, where time-var yi n g state level variables are
likely to vary between states pre and post treatment, this i s likely to result in biased treatment effect
estimates. Double difference estimates in these data (not reported) are never statistically significant.
Inconsistent with extant statistical as well as anecdotal evidence, this is likely to be a reflection of
confounding state-by-year effects. Moreover, the triple difference approach is a way of dealing with
worries of potential endogeneity of treatment, by including an additional control group that is also
affected by the same time-varying state level variables.
15
4.2. Identification. Th e main ident i fy i n g assumption in this triple difference specifi-
cation is that ther e were no contemporaneous shocks in states at the t i m e of midday
meal introduction, which impacted relative outcomes of the treatment group. At the
state level, such a change may occur in pu b l i c schools if ther e is a contemp or aneou s
change in state school policy , and in Section 5.2.2 we provide a detailed discussion of
possible candidates. Additionally, private schools may have respond ed to the intro-
duction of a midday meal in public schools by strategically improving school quality
in the hope of attr act i ng or retai n i n g students. Such confounding changes are likely t o
be reflected in relative changes in schooling inputs (including teachers, teaching aids
and physical infrastructure). We test this by putting these variables on the left hand
side of our triple difference equation (4.1). Our r esu l t s indicate that there were no
contemporaneous changes in the relative inputs between treatment and control groups
at the time of midday meal introduction.
There are two pre-conditions for the validity of our quasi-exper i mental approach.
The first is that control group outcomes are unaffected by treatment. In our specifi -
cation tests, we try to verify this by showing that private school enrollment did n ot
change in response to the introduction of midday meals. The second pre-condition is
that there was no purposive placement of the midday meal policy.
As discussed in section 2, the timin g of m i d d ay meal introduction was idiosyncratic.
This is supported by Figure 2, which depicts mean inputs (and their 95% confidence
intervals) for schools, grouped by the year in which the midday meal was implemented.
(So, for example, the top left-hand graph indicates that schools located in states which
implemented midday meals in 2002 or earlier on average had about 4 classrooms.)
The fact that these confidence intervals overl ap indicates that differences in means,
by timing of implementation, are not statistically significant. At the same time, 2005
implementers, Bih a r and Assam, do seem to have consistently worse schooling quality.
We account for this in our sp ecification tests by showing that our results are robust to
the exclusion of late (as well as early) implementers.
There may also be lingering concer n that the timing of midday meal adoption is
related to state policies or preferences which are correlated with state-level trends in
educational outcomes. Figure 3, which presents literacy data from India’s decennial
censuses, su ggest s that th i s is not the case. It shows that literacy rates, in states from
the sample grouped by timing of implementation over our period of observation, have
developed in a largely parallel fashion over the last twenty years.
Additionally, enrollment in publi c and private schools also developed in a parallel
manner two years before program implementation, for the states for which there is
16
Note:&!"#$%&#'()*%+*,#-.$%/00/%$-"112#3'%#3,(.$%&)14%56789%')1(,*+%:;%.#4#3'%1&%4#++<;%4*<2%#4,2*4*3.<.#13=%!"*%
')1(,$%,*).<#3%.1%>""<..#$'<)"9%?(@<)<.9%!<4#2%A<+(%<3+%B<@<$."<3%#3%/00/%1)%*<)2#*)%C')1(, %DEF%G3+")<%H)<+*$"9%
I<)3<.<J<9%K<"<)<$".)<%<3+%L..<)<3-"<2%#3%/00M%C')1(,%/EF%N<);<3<9%N#4<-"<2%H)<+*$"9%K<+";<%H)<+*$"9%O)#$$<%
<3+%L..<)% H)<+*$"%#3% /00P%C')1(,%ME F%<3+% G$$<4% <3+% Q#"<)% #3% /00R%C')1(,% PE=% !"*%+<.<%,1#3.$%)*,)*$*3.%')1(,%
4*<3$%<3+%."*%:<)$%)*,)*$*3.%."*%SRT%-13&#+*3-*%#3.*)U<2$&
>2<$$)114$% O."*)%)114$% !*<-"*)$% 7.<&&%
V<.*)% 82*-.)#-#.;% *LUOVµ!1#2*.%
>14413%!1#2*.% H2<;')1(3+% Q2<-J:1<)+$% !*<-"#3'%!)(3J%
W*'*3+X%
D%±%/00/%1)%*<)2#*)%
/%±%/00M%
M%±%/00P%
P%Y%/00R%
K*<3$%1&%63,(.$%
!#4*%1&%64,2*4*3.<.#13%
Figure 2. 2002 School Inputs by Timing of Midday Meal Implementation
sufficient pre-treatment data. Table 4 show s that for 2004 implementers there is no
statistically significant difference i n the enrollment trend between 2002 -2003 for pub li c
and private schools.
23
However, there do exist observable differ en ce s in schoolin g in-
puts between public and private sch ools, as do cu m ented recently in Muralidharan and
23
The results hold as well for the 2005 implementers (table not reported).
17
Note:&This%figure%depicts%trends%in%literacy%rates%in%states, %which%are%grouped%according%to%the%timing%of%midday%meal%
implementation.% The% groups% pertain% to:% Gujarat,% Tamil% Nadu% and% Rajasthan% in% 2002% or% earlier?% Andhra% Pradesh,%
Karnataka,%Madhya%Pradesh%and%Maharashtra%in%2003?%and%Assam%and%Bihar%in%2005.%Chhattisgarh%and%Uttaranchal%
are%not%separately%included%since%they%became%states%only%in%2000.%Source:%Census%of%India.&
Figure 3. Literacy Rates by Timing of Midday Meal Implementation
Kremer (2006) and Kingdon (2007).
24
As the first two columns of Table 5 indicate, pri-
vate schools have larger student bodies; have more rooms, staff and equipment; better
24
Muralidharan and Kremer (2006) and Kingdon (2007) have also noted a growth in private school
enrollment, driven primarily by the entry of private unrecognized schools. Since DISE only surveys
recognized schools and our sample constitutes a balanced panel, our results are not directly driven by
births in the sample. There may, however, be an indirect effect if new entrant s draw e nrol l me nt away
from ex t ant public or private schools. To the extent that new private entrants (whether recognized
or unrecognized) draw proportionately from enrollment in extant public and private schools at the
state level over the period of observation, this should not compromise the identification strategy in our
balanced panel. If, on the other hand, private unrecognized scho ol s enter strategically where there has
been a failure in public schools, then our treatment effect estimates may be biased downward. This
seems unlikely for two reasons . First, there is no reason to believe that private entry is corr el at e d with
idiosyncratic midday meal introduction. Second, in a narrow, high-frequency window of ob se rvation,
parallel trends between private and public school enrollment within a state seems like a reasonable
assumption even with entry. Table 4 provides c orr oborat i ve evidence in th i s regard.
18
schooling infrastructure; and are les s li kely to te ach in the vernacular (likely, reflecting
more English language instruction).
The main concern arising from these observed differences is tha t characteristics which
differentiate private and public schools may be associated with different trends in en-
rollment between the two groups within a given state. We account for these concerns
by applying our triple difference estimation described in Equation (4.1) to a matched
sample of public and p r i vat e schools. Remaining concerns pertaining to standard omit-
ted variable bias are accounted for by extending the empirical model to include a vector
of potentially time-varying school-level inputs X
it
.
The goal of the matching exercise is to find a grou p of private schools that is as
similar as possible to the public schools in our sample.
25
To achieve this, we first
estimate for each school th e propensity score with a standard probit regression mod el
in which the independent variables are from the base year 2002. We match on basic
infrastructure (classro oms, other rooms, toilets, water, electricity, playgrounds), staff
(teachers and other staff), teaching learnin g materials (blackb o ar d s and trunks that
contain learning materials), language o f instruction (vernacular) and on primary school
size. In the common support regi on, for each public school we find a comp ar ab l e private
school located in the same st ate with the closest propensity score. The propensity score
matching is done to the first nearest neighbor without replacement so as to obtain a
sample of public schools as similar as possible to that of privat e schools. Unmatched
schools are discarded and not used in estimating the treatment impact.
26
As the last two columns of Table 5 indicate, the matched sample of public and private
schools are indistinguishable in terms of observable characteristics. The residual dif-
ferences in average school characteristics after matching are close to zero and therefore
economically trivial.
5. Results
5.1. Main Results . We begin by estimating Equation (4.1) using pooled OLS. Table
6presentsourmainresult: thetripledifferenceestimateβ,whichcapturestheeffect
of midd ay meals (MDM
ist
)onschoolenrollment. Eachcolumnrepresentsadifferent
25
A group of public and private schools that is similar on observable school characteristics will also
be more likely similar on unobservable characterist i cs such as the quality of schooling. Suggestive
evidence from the IHDS 2005 data set shows that, only once observable schooling characteristics are
accounted for , there is no statistical significant difference in learning results between public and private
schools in areas where the midday meal program was not yet implemented.
26
Matching with replacement does not eliminate the differences in observable average characteristics
between p r ivate and public schools. Our analysis was performed using the user-writt en Stata program
‘psmatch2’ (described in Leuven and Sianesi http://ideas.repec.org/c/boc/bocode/s432001.html).
19
regression. In this and all other triple difference estimations, we control for state,
time, and public school dummies as well as their pair-wise interactions as presented in
Equation 4.1 (although, in the interest of space, coefficient estimates are not reported).
In col u m n s 1-5, the dependent variable is log of enrollment in grades 1-5, respectively,
and in column 6 the dependent variable is the log of total primary school enrollment.
Following Bertrand et al. (2004), in this and all subsequent tables, standard errors
clustered at the state level are presented in parentheses.
The positive coefficients for β in row 1 indicate that midday meals incre ase primary
school en r ol l m ent. The response is largest in grade 1 (col u m n 1), where enrollment
increases by a large and statistically significant 20.8%.
27
The magnitude of this point
estimate reflects the fact that all new enrollments – 6-year-olds, older chi l d r en , as well
as under-aged children – are mainstreamed in first grade.
In grades 2, 3, 4 and 5 the point estimate for β falls an d is statistical l y insignificant.
As mentioned earlier, this is likely to reflect two th i n g s. First, enrollment in grades
2-5 can only be increased by reducing dropouts and baseline dropout rates in grades
2-5 are low to begin with. Second, the relative value of the implicit midday subsidy
decreases with grades. Hence, midday meals are likely to be less effective at spurring
(re)enrollment in upper grades.
Overall, m id d ay meals engender a statistically significant 13.3% increase in primary
school enrollment (column 6). The level results (not report ed ) underscore the economic
significance of this percentage increase: it cor r esponds to around 14 additional students
per primary school, 6 of whom enter gr ad e 1. (The fact that the level results closely
resemble the log results in table 6 suggests, moreover, that this main result is not
sensitive to functional form.)
This translates i nto almost 6.3 million (≈ 450,000 public schools × 14 additional
students) children entering school on account of midday meal introduction in our sam-
ple states. If 27 million primary school-aged children in the country were out of scho ol
in 2002 (UNESCO 2006), and 20 million (≈ 80%) of these resided in the states we
study, this would mean that midday meals are responsible for absorbing a striking 30%
(≈ 6.3/20) of out-o f-scho o l child re n. Even if half of the 6 additional children entering
grade 1 were below 6 years of age, our estimates still suggest that midday meals would
still account for a 25% reduction in out of scho ol 6 to 10 year-olds.
27
This and other percentage incre ase s in enrollment following from the binary explanatory variable,
MDM, are calculated in the following manner: 0.208 = exp(0.189) − 1.
20
5.2. Specification Tests & Extensions. In this section we run a number of spec-
ification checks to ascertain the robustness of our main results and validity of our
empirical strategy.
5.2.1. School-level Heterogeneity. Our research design allows for different average en-
rollments at the state level between public and privat e schools. However, we may
still be concerned that secular differences in scho ol characteristics are correlated with
different trends in enrollment between public and private schools at the state level.
We account for this possibility in Table 7, whi ch presents triple differen c e estimates
analogous to those in Table 6 using the matched sample of public and private schools
described earlier.
The results described in the top half of Table 7 closely resemble our main results.
The 13.1% increase in p r i m a ry school enrollment presented in column 6 is strikingly
similar to the 13.3% increase estim a t ed for the full sam p l e. Also the point estimates at
the in d i v i d u al grade level (columns 1-5) are not statisti ca l l y significant different from
the estimates on the main sample. The magnitudes as well as the pattern of the point
estimates are qualitati vely identical to those pr esented in Table 6. In particular, the
overall increase in primary school is d r i ven by statistically significant increases in grade
1 enrollment, and enrollment responses are positive throughout.
The bottom half of Table 7 extends this exer cise to account for omitted var iab l e bias
by including a vector of potentially time -va ry i n g schooling input s X
it
,summarizedin
Table 3. The coefficient estimates on the schooling inputs (not reported) are consistent
with our priors: more classrooms, teachers, other staff, blackboards, and physical in-
frastructure are associated with higher enrollment. The triple difference point estimates
in this specification are very similar to those in the top half of the table, suggesting the
our simple triple difference estimates does not suffer from significant omitted variable
bias. However t h i s interpretation needs to be t r eat ed with caution since, to the extent
that schooling inputs are endogenous, all the coefficients in this table wi l l be biased. In
general, however, the magnitude of the point estimates are not significantly different
and the overall pattern of the estimates is qualitatively identical to the triple difference
estimates in both the full and the matched sample.
Together these robustness checks alleviate concerns that heterogeneity across (private
and public) schools is driving our main results. Given the loss in sample size entailed
in this matching exercise, we conduct further specification tests on the full sample,
although the results are qualitatively similar when using the matched sample.
21
5.2.2. Confounding Changes. State governments have discretion over the implementa-
tion of school policies. This could be problematic for our triple difference model if there
were confounding policy changes at the state level contemporaneous to the institution
of midday meals, which affected treatment and control groups differentially. In this
respect, the main public policy contender is the S a rva Shiksha Abhiyan (SSA).
Targeted at the 6-14 age group, the SSA’s stated aims were to achieve universal en-
rollment and retention, bridge gender and caste gaps, and improve education quality.
It was launched by the Government of India in 2001-02, before our observation per i od.
To this extent, the observed effect of the introduction of school lunches cannot be con-
founded with any effect associated with cross state-time differences in the introduction
(or withdrawal) of the SSA per se. The SSA merged all previous investments in el-
ementary education, including th e District Primary Educati on Programme (DPEP),
from the state or from the central government (SSA 2008).
28
Under the SSA, new schools were opened in habitations with no schooling facilities
and the basic infrastructure of exi st i n g schools was strengthen ed . New teachers were
hired and grants were given for the development of teaching learning materials. The
interventions for out of school children focused mainly on alternative schooling models
(Alternative and Innovative Education (AIE) s cho o l s, residential bridge courses, tent
schools, mobile school i n g or home based ed u cat i o n ) and on the buil d i n g of Education
Guarantee Scheme (EGS) schools. These types of schools are not included in our panel
of schools. Therefore, as long as there is no differential impact of these interventions
on our treatment and control groups, they should not affect our estimates.
Still, the concern remains that changes in schooling inputs introduced under the
auspices of the SSA may have coincided with midday meal implementation. We ex-
amine this possibility by estimating a triple difference with different schooling inputs
(instead of enrol l ment) on the left hand side of Eq u a t i on (4.1), focusing on the set
of schooling in p u t s that could have been changed under the SSA: bas i c infrastructure
of the schools (classrooms, other rooms, toilets, water, electricity, playgrounds), staff
(teachers and other staff) and teaching materials (blackboards and trunks that contain
learning materials).
28
The DPEP was conceptualize d in t h e early 1990s in respon se to Ind i a’ s low literacy rates. Its stated
aims were to provide primary scho ol acc es s f or al l children, reduce dropout rates, increase learning
achievements, and r ed u ce gender and caste gaps in educational attainment (DOE 1995). (S ee World
Bank (2003) for a review of the evidence regarding the impact of this program .) Exter nal funding for
the DPEP expired in 2001-02; only in Andhra Pradesh and West Bengal did the DPEP continue to
be funde d (in this case, by the UK Government) u ntil 2003 (Krishna Kumar and Saxena 2001). In
the case of West Bengal, this does not pose a threat to identification since West Bengal is not in our
main sample, and dropping Andhra Pradesh from our sample does not change the results.
22
Table 8 furnishes the results of this exercise. Each column has, as a dependent
variable, a different schooling input on the left hand side. With only one exception (a
common toilet, whi ch is significant only at the 10% level, and of the “wrong” sign),
the triple difference estimates for these inputs are statistically insignificant, indicating
that schooling inputs in public versus private schools within each state did not change
differentially at the same time of midday meal introduction. This is likely to be a
reflection of the fact that there was little change in public or private scho ol inputs over
time during our three-year obser vation period, whether contemporaneous to midday
meal introduction or otherwise; this is immediately evident from a cursory glance at
descriptive statistics of schooling inputs by academic year (not shown). This feature
further alleviates worries regarding potentially confou n d i n g changes.
5.2.3. Contamination. In principal the increased enrollment in public schools can come
from two potential sources: children who would not have otherwise been in school (new
enrollments), or children who would otherwise be enrolled in private schools and may
be switching from private to public schools. In the latter case, our control group would
be contaminated an d the triple di ffer en ce estimates presented i n Tab l e 6 would be
upward bias estimates of the general equilibrium enrollment effects of midday meals.
We explor e t h i s possibility by estimating the following double differ en ce (DD) model
for our sample of private schools:
(5.1) Y
ist
= λ
s
+ γ
t
+ φm
st
+ �
i
,
where Y
ist
, λ
s
and γ
t
are defined as in Equation (4.1). The policy variable m
st
is equal
to 1 for all schools if the midday meal program was in place in public schools in state
s at time t.
The DD coeffi ci ent, φ,isonlysuggestiveofpotentialcontamination,sincewelack
acontrolgroupforprivateschools(i.e. thisisadouble-andnotatriple-difference.)
Nevertheless, if increased public scho o l enrollment in g r ad e 1 and pr i m a ry school as a
whole reflected transfers, then we should expect to see a statistically significant negative
coefficient for our estimate of φ at these levels. Table 9 suggests that this is not the
case: coefficients for grades and primary school as a whole are statistically insi gn ifi cant.
This allays fears of contamin at i o n and provides some validation for the use of private
schools as a control group in the triple difference model.
5.2.4. Timing of Implementation. Our empirical strategy relies on the staggered timing
of implementation of the midday meal scheme. We argued ear l i er that the timin g of
23
implementation during our observation period is idiosyncratic. But there may still
be concern that early or late implementers have p olicies and preferences which are
correlated wit h trends in enrollment that are different from o t h er s in our sample. One
way of addressing this concern is to examine whether our results are being driven by
these states.
In Table 10 we estim a te th e t ri p l e d i ffer en ce model in Equation (4.1) on four different
samples of public a n d private schools, depending upon ea rl y or late implementation.
The point estimates are v i r tu a l l y identical when we drop laggards Assam and Bihar
(first quarter of Table 10), pioneers Tamil Nadu and Gujarat (s eco n d quarter), or both
laggards and pioneers (third quarter). In addition, when we exclude one state at a
time from the sample our results are also unchanged (not r eported), indicating that n o
single state is driving our results.
Finally, as related in Section 2, we did not include pilot regions, Kerala, Jhark-
hand or West Bengal in our sample, because of both poor documentation regarding
implementation and worries of bias introduced by purposive placement. In the bottom
quarter of Table 10, we include schools in Kerala, Jharkhand, West Bengal as well as
schools covered in these pilot regions, treating each pilot region in a given state as a
“new” state, with the MDM
ist
variable defined accordingly. The bottom quarter of
Table 10 reports our tri p l e difference estimates for th i s extended sample. Th e picture
remains the same (the p-value for the primary school coefficient estimate is 0.104).
Together, these robustness checks indicate that our results are not driven by poten-
tially non-random timing of implementation.
6. Heterogeneous responses, Attendance and Learning
In this section, we use a recent household and school-level survey from the Indian
Human Development Survey (IHDS) 2005 in order to extend our main results in three
ways. Fi r st , one would expect that children from r el a ti vely disadvantaged backgrounds
comprise the bulk of the observed enrollment respo n se, both becau se t h ey are the most
likely to be out of school in the first place, and because they are likely to be most
responsive to a food subsidy. We explore this by allowing for heterogeneous “responses”
to midday meal provision by caste, income and gender.
Second, proponents claim that on account of its on-site consumption after morning
lessons, one of the chief merits of midday meals is that it boosts school attendance,
which can be quite different from enrollment, particularly in the In d i an context. We
therefore explore whether midday meal provision is associated with higher attendance.
24
Third, the positive enrollment response to midd ay meals documented in Section 5
reflects the sum of two effects, alluded to in the introduction. The first is the im-
plicit subsidy effect, which is thought to be positive as school lunches lower the cost
of schooling. The second is the learning effect whose sign is, in general, ambiguous see
(Kremer and Vermeersch (2004) and Kazianga et al. (2009) for detailed discussions.)
On the one hand, there is a positive direct effect, as improved nutrition from midday
meal consumption leads to more learni ng, and commensurately high e r r et u r n s to ed -
ucation, and thereby higher enrollment. But there is also a negative indirect learning
effect. This arises from the possibility that limited resources in ter m s of personnel,
teaching tools, and infrastructure may have to be stretched over a larger number of
enrolled children; or from the prospect of teachers being distracted from teaching due
to meal-related administration. As a final ext en si o n , therefore, we explore whether
there is any net learning effect associated with midday meal provision.
Each of these outcomes are import ant policy issues in t hei r own right, and therefore
worthy of investigation. It is worth emphasizing up front, however, that due to the
cross-sectional nature of the data as well as to the timing of the survey, we cannot rule
out endogeneity concerns, so the results presented here are on l y sugg est i ve.
6.1. Data. IHDS 2005 is a nationally representative survey conducted in 41,554 house-
holds during 2004-2005 across all states and u n io n territories of India wit h the exception
of the Andaman & Nicobar and Lakshadweep islands (see IHDS (2008 ) for complete
documentation). The survey covers 1,504 v i l lag es and 97 0 u r b an neighborhoods. In
addition to careful data collection and quality contr o l (Desai et al. 20 08) , this survey
has 3 features which are useful for our p u rposes. First, income and demographic data
from the household survey allow us to examine heterogeneous programme responses.
The second unique feature o f the household s ur vey is that it includes not on l y stan-
dard enrollment data, but also inform a ti o n regarding each child’s school attendance,
as well as assessments of reading, writing and arithmetic skills for children aged 8-11
(developed in conjunction with Pratham, an NGO with extensive knowledge in this
area.)
29
Third, in addition to the household survey, IHDS includes a prim a r y school survey
which covered at least one public and (where present) one private school in each village
29
The income and demographic quest i ons are answered by the head of the household. Questions
pertaining to children in the household are answered typically by the mother. Tests were administere d
in the household. Although all 8-11 year-olds in the sample households were suppose d to take the test,
only about 72% of them actually did so, and we cannot r ul e out the p oss i bi l i ty that missing scores
are non-random. Non-response is much higher for non-enrolled children, than for enrolled children.
However, non-response is not correlated with the degree of midday meal implementation.
25
or urban block, the primary sampling units (PSU). Where there was no school faci l i ty
within the selected village, the nearest school was surveyed. Importantly, this school
survey inclu d e d a question regarding whether a midday meal was o ffer ed in the school.
We use this response to construct a dummy variable equal to one if at least one public
school in the PSU provided midday meals.
30
Since tests were only administered to 8-11 year-olds, our core sample comprises
children in this age group who are either out of school or are currently enrolled in a
public primary school. Table 11 presents summary statistics for the 9,224 observations
in our sampl e . It indicates that 77% of children in this age group have access to a
midday meal offered at a l ocal public school. On average, 88% are currently enrolled
and, in the past week attended school for 30.9 hours. 34% belong to Other Backward
Castes (OBC), 36% are eit h er Scheduled Castes or Scheduled Tribes (SC\ST); 15%
belong to an upper caste; and the remainder (Other) belong to minority religions (86%
of this category are Muslim). The vast majority of children come from households
where parents have completed only 5 years of school in g or fewer.
Three dummy variables, Reading, Math and Writing,areconstructedtocapture
learning. Of the children who were administered learning tests, 72% can read at least
words; 40% of the children that took the math test can solve at least a simple addition
problem; and 61% can write a simple sentence with at most one mistake.
6.2. Empirical Mode l & Results. In contrast to our empirical strategy u si n g DISE’s
panel data structure, we cannot use an ITT strategy exploiting staggered implemen-
tation of th e poli cy. The simple reason for t h i s is that by 2005 when the IHDS was
conducted, the vast majority of the Indian states had introduced the midday meal
scheme. Furthermore, because IHDS is a cross-section, and midday meals are only
offered in public schools, we cannot use private schools as a control group since this
would not permit us to distinguish the midday meal effect from secular differences in
enrollment between private and public schools.
We estimate the following baseline mo del:
(6.1) R
ihj
= λMDM
j
+ νZ
ih
+ �,
30
The choice of school was non-random – where more than one of either facility was present, inter-
viewers were asked to select the facility which was p r ed omi nantly used by residents. However, since
there is no variat i on in midday meal implementation across public schools within a given village, we
do not believe t hi s introduces any bias in our estimation.
26
where our unit of observation is child i,livinginhouseholdh and PSU j.Theleft
hand side variable R
ihj
pertains, in various specifications to, (i) a dummy variable equal
to 1 if child i is enrolled in school (enrollm ent); (ii) the number of hours spent attending
school in the previous week (attendance); and (in t hr ee separate specifications) whether
(=1) or not (=0) the child can read, write or do math.
The dummy variab le MDM
j
indicates whether, in PSU j where child i resides, mid-
day meals are served in public schools (MDM
j
=1)ornot(MDM
j
=0). Thevector
Z
ih
contains individual characteristics such as gender and age, and household char-
acteristics including caste and parents’ education. In order to capture heterogeneous
treatment effect, we interact MDM
j
in three separate specifications with dummies for
caste/religious group, income quartile and gender. Effectively, this means rep l a ci n g
MDM
j
with the corresponding interaction terms.
Table 12 presents OLS estimates for equation 6.1. (Probit estimations produce
qualitatively identical results.) The sampl e in column 1 pertains to all children between
the ages of 8 and 11 who are either non-enrolled or currently enrolled in public primary
school. The point estimate in row 1 indicates that midday meals are associated with
10.8% higher enrollment in this age group. This estimate is similar to our DISE
estimates for primary school, but much larger relati ve to the responses in grades 3-
5 (wher e 8-11 year-olds are typically enrolled). While we cannot rule out bias, this
would be consistent with student retention in upper grades after 3-5 years of program
exposure (in 2005) following a large grade 1 response in the initial years of exposure.
The next three columns permi t this average enrollment response to vary by caste and
religion (column 2), income quartile (column 3) and gender (column 4). Enrollment
increases across the board, but is largest for relatively disadvantaged children. With
respect to social group, colum n 2 indicates the response is hig hest for SC\STs and
the Other (predominantly Muslim) category; column 3 indicates that it is largest for
the bottom three-quarters of the income distribution; and column 4 indicates that it
is larger, although not significantly so, for girls than for boys. Al t h ou g h this may be
indicative of purposive placement, it is nevertheless consistent with our priors that
disadvantaged children are likely to be most responsive to this food subsidy.
In column 5, the d ependent variable is attendance and our sample is restricted
accordingly to children who are actually enrolled in school. The result suggests that
midday meal provision is associated with 2.6 ad d i t i on a l hours of school attendance
a week, which corresponds to a one-third of a standard deviation i n c re ase . As with
enrollment, this may reflect purposive placement. However, it is consistent with the
fact that children have to attend at least morning classes to get lunch at noon. It
27
is also suppo r te d by anecdotal evidence (PROBE 1999) that shows that with midday
meals in place children themselves like to come to school.
The dependent variables in Columns 6, 7 and 8 are dummies indicating a child’s
ability to read, solve math problems, and write, respectively. The co efficients in row 1
indicate that midday meals are not associated with any learning effect: the estimates
are statistically insignificant and close to zero in each of the three categories. This
weak correlation may reflect purposive placement if implementation occurs in more
disadvantaged regions. It is also likely to reflect a selection effect, since th e esti m a t e
captures an average effect of students from (as columns 2-4 seem to suggest) weaker
socio-economic backgrounds who are r esponding to the programme an d stronger stu-
dents who are already enrolled.
Nevertheless, it is consistent with evidence from studies in other geographies, re-
viewed in Kazianga et al. (2009), that school feeding programs are often ineffectual at
raising acad em i c achievement. It is also consist ent with lower average schooling inputs,
resulting from a large enrollment response and an absence of any concomitant increase
in staff or infrastructure. If midday meals are associated with higher enrollment but,
as these r esu l t s suggest, no increase in learning, these data seem to suggest t h a t the
implicit subsidy channel is driving the positive enrollment response to midday meals.
7. Conclusion
This pap er provides evid ence that India’s midday meal scheme has led to large in-
creases in primary school enrollment. Our mai n triple difference estimates indicate
that primary school enrollment increased by 13%. Back-of-the envelope calculations
(described in section 5.1) suggest that this corresponds to about 6.3 million additional
children in school, which is likely to amount to a substantial reduction in the esti-
mated 20 million 6-10 year-olds who were out of school in the states we study in 2002.
Household survey data also indicat e that many of new enr ol l m ents may be chil d r en
from disadvantaged socio -ec on o m i c backgrounds, suggesting that the policy may be
successful in reaching segments of the popul a t i on which have otherwise proved difficult
to enroll.
The largest and most robust overall increases are in grade 1, where enrollment rose
by 21%. Enrollment responses in grades 2, 3, 4 and 5 are, by contrast, more muted.
The magnitude of the grade 1 effect is consistent with the fact that never-enrolled
children are mainstreamed in grade 1, regardless of age. By contrast, enrollment in
later grades can only be boosted by lowering dropouts from the previous year, and
the scope for this is limited given the low dropout rate in higher grades. Effectively,
28
therefore, in order to boost enrollment in (for example) grade 5, a state would need to
have had midday meals in place for 5 years and retained the large grade 1 intake up
into grade 5.
The fact that we don’t observe this is partly a reflection of the fact that most states
in our sample were exposed to the programme for 1 to 2 years: hence, the response
in grade 1 and not thereafter. Even in the long-run, however, midday meals are likely
to be more effective at encouraging school participation among children in the lowest
grades than the highest grades in primary school. This is because the cash value of the
meal is constant while costs of schooling increase with grade, due to the higher direct
costs associated wit h school materials (uniforms, books, etc.) and opportunity costs
(value of home and labor market product i o n ). This means that, in relative terms, the
implicit meal subsidy is higher in lower grades.
The main advantages of the data we exploit in our mai n analysis are its wide coverage,
timing, and panel data structure, which allow for a large-scale impact assessment of
this important school lunch policy. The disadvantage of the data is that it only has
reliable data on e n ro l l m ent. Although this is an important and commonly utilized
metric for school attainment, it is arguably not as important as attendance or learning.
Starting in 2005 and continuing annually since then, ASER has initiated a rich large-
scale household and school survey. An interesting avenue of future research will be to
exploit exogenous variation in exposure to the midday meal program to identify its
effects on attendance and learning.
Results from the household cross-section we exploit in this paper are only sugges-
tive of there being increased attend an ce b u t no significant learning effects associated
with the program. However, the former finding is intuitive given the administration
of lunches on site at midday. The latter finding does seem to be substantiated by
anecdotal evidence that the administration of midday meals distracts from teaching,
and that the enrollment response to the program has stretched limited resources, both
of which compromise learning. This is further corroborated by the fact that our DISE
data indicate little change in complementary staff, materials and infrastructure. Given
the magnitu d e of the enrollment response engendered by midday meals, such invest-
ments seem necessary prerequisite if learning is to be promoted. Still, the absence of
any evidence of increasing learning coupled with large enrollment effects suggests that
the implicit subsidy channel is responsi b l e for the latter effect .
Given the wide coverage of the data we exploit, we believe our main DISE enrollm ent
results to be representative for India. This is policy relevant given both the scale
of the midday meal program, and the fact that India houses the largest number of
29
out-of-school children in the world (UNICEF 2008). It seems fair to specul a te that
the magnitude of the r es ponse that we document here is larger than it would be,
were a similar school feeding program to be instituted in Latin America or East Asia,
where primary school enrollment is already considerably ad vanced. Quite apart from
enrollment effects, however, there may be important nutritional or school attendance
benefits which may still speak for the introduction of similar school feeding programs
in these regions. At the same time the enrollment effects we document in this paper
may be generalizable to parts of Su b -S a h ar a n Africa, where primary school enrollment
rates are compa r abl e to those of India, and decentralized government institutions have
the capacity to implement this logistically demanding policy.
30
Table 1. Sample of states and time of implementation
State Name Implementation Treatment Year
Andhra Pradesh January 2003 2003
Assam
∗
January 2005 2005
Bihar
∗
January 2005 2005
Chhattisgarh April 2002 2002
Gujarat November 1984 1986
Haryana August 2004 2004
Himachal Pradesh September 2004 2004
Karnataka
∗
July 2003 2003
Madhya Pradesh
∗
July 2003 2003
Maharashtra January 2003 2003
Orissa
∗
September 2004 2004
Rajasthan July 2002 2002
Tamil Nadu July 1982 1982
Uttar Pradesh September 2004 2004
Uttaranchal July 2003 2003
Note. a. The second column contains the month and year when the
midday meal scheme was implemented with full coverage throughout
the state; these dates were collected from state midday meal scheme
audit and budget reports. The third column contains the academic
year starting from which a state is considered to have implemented
the midday meal sch em e; an academic year is considered to start on
the 30th of September. States marked with
∗
implemented the midday
meal scheme in pilot districts as follows: Assam Pilot in December 2004
(treatment year 2005), Bihar Pilot in Septe mb e r 2004 (treatment year
2004), Karnataka Pilot in June 2002 (treatment year 2002), Mad hya
Pradesh Pilot in October 1995 (treatment year 1996) and Orissa Pilot
in June 2001 ( tr e atm ent year 2001).
b. States or districts excluded from the main DISE sample due to par-
tial implementation, lack of information regarding where the scheme
was implemented or due to potential purposive placement: Jharkhand,
Kerala, West Bengal, Assam Pilot, Bihar Pilot, Karnataka Pilot, Mad-
hya Pradesh Pilot and Orissa Pilot. The main regressions in the paper
are similar if these districts and blocks are included (see text). All other
states are not covered by DISE.
31
Table 2. School Distribution among States in Sample
Schools
State Name Population Public Private
Andhra Pradesh 9.24 7.67 1.98
Assam 3.23 5.55 0.09
Bihar 10.06 8.16 0.08
Chhattisgarh 2.53 5.07 3.07
Gujarat 6.14 2.11 1.46
Haryana 2.56 0.60 0.02
Himachal Pradesh 0.74 2.50 1.47
Karnataka 6.41 6.58 7.46
Madhya Pradesh 7.31 9.36 18.40
Maharashtra 11.74 7.53 3.59
Orissa 4.46 6.32 1.55
Rajasthan 6.85 10.30 16.56
Tamil Nadu 7.56 5.49 8.02
Uttar Pradesh 20.14 20.45 33.45
Uttaranchal 1.03 2.30 2.82
Total 100.00 100.00 100.00
Note. In percentages. The second column figures are cal-
culated from Census of India 2001 data. T he figures in the
third column are calculated from our main sample of public
schools. The figures in the fourth column are calculated from
our main sample of private scho ol s .
32
Table 3. Means of 2002 variables
Enrollment
a
Grade 1 35.02
(37.89)
Grade 2 26.20
(27.35)
Grade 3 23.38
(24.99)
Grade 4 20.40
(22.71)
Grade 5 17.66
(22.54)
Primary school 122.65
(118.61)
Schooling Inputs
b
Number of classrooms 3.27
(2.89)
Number of other rooms 0.96
(1.69)
Number of teachers 1.97
(1.93)
Number of other staff 0.37
(1.06)
Dummy for water 0.80
(0.40)
Dummy for electricity 0.20
(0.40)
Dummy for girls’ toilet 0.23
(0.42)
Dummy for common toilet 0.35
(0.48)
Dummy for playground 0.51
(0.50)
Number of blackbo a rd s 4.41
(3.85)
Number of teaching trunks 1.62
(2.52)
Dummy for teaching in vernacular 0.97
(0.17)
Note. Standard deviation in parentheses. All regressions omit observations in 3
states and 29 pilot districts due to partial implementation, lack of information re-
garding where the scheme was implemented or due to potential purposive placement.
Data are from DISE 2002. Observations: a:489,125b:428,491.
33
Table 4. Double Difference: Parallel Trends between Public and Pri-
vate Schools
(1) (2) (3) (4) (5) (6)
Grade 1 Grade 2 Grade 3 Grade 4 Grade 5 Primary
time x public -0.079 0.026 0.006 0.001 -0.000 -0.025
(0.049) (0.016) (0.014) (0.015) (0.016) (0.036)
time 0.041
∗
0.088
∗∗∗
0.110
∗∗∗
0.132
∗∗∗
0.122
∗∗∗
0.097
∗∗∗
(0.013) (0.008) (0.013) (0.017) (0.017) (0.013)
public 0.056 0.055 -0.010 -0.118 -0.264
∗∗
0.083
(0.151) (0.129) (0.105) (0.072) (0.058) (0.145)
Obs. 297,635 297,635 297,635 297,635 297,635 297,635
Adj. R
2
0.00 0.00 0.00 0.00 0.01 0.00
Note. Robust st an d ard errors in parentheses cl us t er ed at the state level. The dependent
variables are log of yearly primary school enrollment, total and disaggregated by grade.
The time dummy is set t o unity for the year 2003. The public dummy is set to unity for
public schools. All regressions include only pu bl i c and private primary schools from the
two years prior to midday meal implementation in the group of states that implemented
in 2004: Haryana, Himachal Pradesh, Orissa ( e xc l ud i ng the pilot districts), Uttar Pradesh.
Data are from D IS E 2002-2003.
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
34
Table 5. Mean s of 2002 Variables: Before and After Matching
Before
a
After
b
Public Pri vate Public Private
School size 122.83 163.58 151.77 162.34
(112.47) (188.12) (170.55) (183.06)
Number of classrooms 3.02 6.98 5.81 6.92
(2.40) (5.69) (4.43) (5.45)
Number of other rooms 0.89 1.87 1.64 1.86
(1.60) (2.55) (2.37) (2.52)
Number of teachers 1.90 2.95 2.54 2.90
(1.73) (3.70) (2.91) (3.37)
Number of other staff 0.34 0.75 0.54 0.74
(0.95) (2.07) (1.47) (2.00)
Dummy for water 0.79 0.96 0.96 0.96
(0.41) (0.19) (0.19) (0.19)
Dummy for electricity 0.17 0.66 0.61 0.66
(0.37) (0.47) (0.49) (0.47)
Dummy for girls’ toilet 0.20 0.67 0.64 0.67
(0.40) (0.47) (0.48) (0.47)
Dummy for common toilet 0.33 0.73 0.71 0.73
(0.47) (0.44) (0.45) (0.44)
Dummy for playground 0.49 0.81 0.82 0.81
(0.50) (0.40) (0.38) (0.40)
Number of blackbo a rd s 4.20 7.57 6.47 7.52
(3.49) (6.64) (5.56) (6.47)
Number of teaching trunks 1.64 1.28 1.51 1.28
(2.49) (3.03) (1.97) (3.03)
Dummy for teaching in vernacular 0.98 0.83 0.94 0.84
(0.15) (0.37) (0.24) (0.37)
Note. Standard deviation in parentheses. Means are calculated on the
basis of 2002 values for a. full sample comprising 428,491 observations
and b matched sample comprising 53,954 observations. Propensity score
matching uses the nearest neighbor without replacement.
35
Table 6. Triple Difference: Primary School Enrollment
(1) (2) (3) (4) (5) (6)
Grade 1 Grade 2 Grade 3 Grade 4 Grade 5 Primary
MDM (β)0.189
∗∗∗
0.052 0.011 0.009 0.014 0.125
∗∗∗
(0.054) (0.033) (0.071) (0.072) (0.067) (0.031)
Obs. 1,473,759 1,473,759 1,473,759 1,473,759 1,473,759 1,473,759
Adj. R
2
0.11 0.10 0.08 0.07 0.15 0.07
Note. Robust standard errors in parentheses clustered at the state level. All regressions
include state dummies, year dummies, a public scho ol dummy PUB, and state x time,
state x PUB, time x PUB interaction terms. The dependent variables are log of yearly
primary school enrollment, total and disaggregated by grade. The MDM dummy is set to
unity for public schools once a state implements the midday meal scheme. Sample: All
regressions include public primary schools and private primary schools. All regressions omit
observations in 3 states and pilot regions from 5 states due to partial implementation, lack
of information r egar d in g where the scheme was implem ented or due to potential pur posi ve
placement. Data are from DISE 2002 - 2004.
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
36
Table 7. Pri m a r y Scho ol Enr ol l m ent on Matched Sample
(1) (2) (3) (4) (5) (6)
Grade 1 Grade 2 Grade 3 Grade 4 Grade 5 Primary
Triple Difference
a
MDM (β)0.142
∗∗
0.081 0.059 0.044 0.057 0.123
∗∗
(0.058) (0.050) (0.039) (0.044) (0.043) (0.048)
Obs. 155,766 155,766 155,766 155,766 155,766 155,766
Adj. R
2
0.05 0.06 0.06 0.06 0.06 0.06
Triple Difference with Covariates
b
MDM (β)0.143
∗∗
0.078 0.054 0.036 0.047 0.125
∗
(0.051) (0.057) (0.062) (0.063) (0.068) (0.060)
Schooling Inputs YES YES YES YES YES YES
Obs. 150,241 150,241 150,241 150,241 150,241 150,241
Adj. R
2
0.25 0.25 0.26 0.26 0.24 0.25
Note. Robust standard errors in parentheses clustered at the state level. All regressions
include state dummies, year dummies, a public school dummy PUB , and state x time, state
x PUB, time x PUB interaction terms. Regressions b include as covari at es the schooling
inputs l i st e d in part b of Table 3. The dependent variables are log of yearly primary school
enrollment, total and disaggregate d by grade. The MDM dummy is set to uni ty for public
schools once a state implements the midday meal scheme. From the sample in Tabl e 6 a
sub-sample was created through a propensity score first nearest neighbor match withou t
replacement on the common s up port, based on the 2002 values of the schooling inputs
described in Table 5, by stat e between public and private schools.
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
37
Table 8. Triple Difference: Schooling Inputs
(1) (2) (3) (4) (5) (6) (7) ( 8) (9) (10) (11)
Classrooms Otherrooms Teachers Staff Water Electricity Gtoilet Ctoilet Playground Blackboard Trunk
MDM (β) -0.103 -0.056 -0. 206 -0. 046 -0. 014 0.001 -0.023 -0.039
∗
-0.015 0.205 0.037
(0.378) (0.197) (0.209) (0.143) (0.012) (0.010) (0.017) (0.019) (0.013) (0.379) (0.093)
Obs. 1,473,759 1,473,759 1,458,615 1,458,595 1,420,100 1,437,599 1,429,051 1,431,237 1,432,754 1,473,759 1,473,759
Adj. R
2
0.19 0.05 0.10 0.10 0.09 0.29 0.17 0.15 0.08 0.17 0.03
Note. Robust standard errors in parentheses clustered at the state l e vel. All regressions i nc lu d e state du mmi e s, year dummies, a public school dummy
PUB, state x time, state x PUB, time x PUB interaction t er ms . The dependent variables are various schooling inputs as noted in the column title. The
MDM dummy is s et to unity for public schools only once a state i m pl em ents the midday meal scheme. Sample is as in Table 6.
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
38
Table 9. Double Difference: Private School Enrollment
(1) (2) (3) (4) (5) (6)
Grade 1 Grade 2 Grade 3 Grade 4 Grade 5 Primary
MDMstate (φ)-0.053 -0.049 -0.034 -0.045 -0.068
∗
-0.076
(0.056) (0.051) (0.053) (0.051) (0.034) (0.064)
Obs. 101,120 101,120 101,120 101,120 101,120 101,120
Adj. R
2
0.06 0.06 0.06 0.06 0.08 0.06
Note. Robust standard errors in parentheses clustered at the state level. All regressions
include state dummies and year dummies. The dependent variables are log of yearly primary
school enrollment, total and disaggregated by grade. The MDMstate dummy is set to uni ty
once a state implements the midday meal scheme in public schools. All regressions includ e
recognized private unaided pr im ary schools only. All regress ion s omit observations in 3
states and pilot regions from 5 states due to partial implementation, lack of information
regarding where the scheme was implemented or due to potential purposive placement. Data
are from DISE 2002-2004.
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
39
Table 10. Triple Differe n ce: Primary School Enrollment, Various Samples
(1) (2) (3) (4) (5) (6)
Grade 1 Grade 2 Grade 3 Grade 4 Grade 5 Primary
Without Late Implementers
a
MDM (β)0.189
∗∗∗
0.052 0.011 0.009 0.014 0.125
∗∗∗
(0.055) (0.033) (0.072) (0.072) (0.068) (0.032)
Obs. 1,312,917 1,312,917 1,312,917 1,312,917 1,312,917 1,312,917
Adj. R
2
0.08 0.08 0.07 0.06 0.09 0.05
Without Early Implementers
b
MDM (β)0.182
∗∗∗
0.026 -0.030 -0.030 - 0. 0 1 3 0.108
∗∗∗
(0.055) (0.029) (0.067) (0.067) (0.068) (0.027)
Obs. 1,352,485 1,352,485 1,352,485 1,352,485 1,352,485 1,352,485
Adj. R
2
0.12 0.10 0.09 0.07 0.14 0.07
Without Early or Late Implementers
c
MDM (β)0.182
∗∗∗
0.025 -0.031 -0.030 - 0. 0 1 3 0.108
∗∗∗
(0.057) (0.030) (0.067) (0.068) (0.069) (0.027)
Obs. 1,191,643 1,191,643 1,191,643 1,191,643 1,191,643 1,191,643
Adj. R
2
0.08 0.08 0.07 0.06 0.09 0.05
With Pilots, Kerala, Jharkhand and West Bengal
d
MDM (β)0.149
∗∗
0.023 -0.004 -0.013 0.001 0.083
(0.064) (0.043) (0.070) (0.070) (0.065) (0.049)
Obs. 1,751,224 1,751,224 1,751,224 1,751,224 1,751,224 1,751,224
Adj. R
2
0.12 0.11 0.10 0.09 0.21 0.09
Note. Robust standard errors in parentheses clustered at the state level. All regressions
include state dummies, year dummies, a public scho ol dummy PUB, and state x time,
state x PUB, time x PUB interaction terms. The dependent variables are log of yearly
primary school enrol l ment, total and disaggregated by grade. The MDM dummy is set
to unity once a state imp le me nts t h e midday meal scheme. All regressions include public
primary schools and private unaided primary schools only. From the sample in Table 6 new
samples are created in the following way: In regressions a Assam and Bihar are excluded; In
regressions b Tamil Nadu and Gujarat are ex cl u de d; In regressions c Tamil Nadu, Gujarat,
Assam and Bihar are excluded; In regressions d the pilot distri ct s Assam Pi l ot , Bihar Pilot,
Karnataka Pilot, Madhya Pradesh Pilot and Orissa Pilot are included as well as Kerala
(with implement at ion year 1995), Jharkhand (with i mp le mentation year 2004) and West
Bengal with (implementation year 2003).
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
40
Table 11. IHDS: Means of Selected Variables
MDM 0.77
(0.42)
Dependent Variables
Currently enrolled 0.88
(0.33)
Attendance 30.91
(8.47)
Reading 0.72
(0.45)
Math 0.40
(0.49)
Writing 0.61
(0.49)
Individual and Household Characteristics
Male 0.50
(0.50)
Age 9.37
(1.04)
Upper Castes 0.15
(0.35)
OBC 0.34
(0.47)
SC\ST 0.36
(0.48)
Other 0.15
(0.36)
Mother no education 0.64
(0.48)
Mother completed primary school 0.19
(0.39)
Mother completed more than 5 years of schooling 0.05
(0.22)
Father no education 0.36
(0.48)
Father completed primary school 0.30
(0.46)
Father completed more than 5 years of schooling 0.14
(0.35)
Note. Standard deviations in parentheses. Sample: children be-
tween 8 and 11 years of age, either out of school or enrolled i n
public primary schools. 9,224 observations. Mean for attendance
is calcu la t ed on a 87% sub-sample of children enrolled in public
primary school. Me a n s for reading, ma t h and writing are calcu-
lated on a 78% sub-sample of children that took a learning test.
41
Table 12. OLS: Enrol l m ent, Heterogeneous Treatment Effects, Atten-
dance and Learning
(1) (2) (3) (4) (5) (6) (7) (8)
Enrollment Caste Income Gender Attendan c e Reading Math Writing
MDM (λ)0.108
∗∗∗
2.592
∗∗∗
0.004 0.015 0.006
(0.03) (0.69) (0.04) (0.04) (0.05)
MDM x Upper Castes 0.057
∗∗
(0.02)
MDM x OBC 0.076
∗∗
(0.04)
MDM x SC\ST 0.115
∗∗∗
(0.03)
MDM x Other 0.193
∗∗∗
(0.06)
MDM x Top Income 0.030
∗∗
(0.01)
MDM x Highmid Income 0.128
∗∗∗
(0.03)
MDM x Lowmid Income 0.123
∗∗∗
(0.02)
MDM x Low Income 0.106
∗∗
(0.05)
MDM x Female 0.123
∗∗∗
(0.04)
MDM x Male 0.093
∗∗∗
(0.02)
Controls YES YES YES YES YES YES YES YES
Obs. 9,224 9,224 9,224 9,224 7,984 6,644 6,631 6,594
Adj. R
2
0.12 0.12 0.12 0.12 0.04 0.08 0.11 0.09
Note. Robust standard errors in parentheses clustered at the state level. The dependent variable in Columns 1-4 is a dummy
equal to unity if a child is enrolled in school. Attendance refers to how many hours in school a child was present in the past
week. In Columns 6-7 the dependent v ariables are dummy v ariables equal to unity if a child can read, do simple math or
write. Controls included are gender, age, household size, caste dummies, income, dummies for mother and father’s education.
Sample: 8-11 year-olds that are either out of school or enrolled in a public primary school (Columns 1-4), sub-sample of
Column 1 sample of children that are enrolled in a public primary school (Column 5), sub-sample of Column 1 sample of
children that took the learning test (columns 6-8).
∗
p<0.10,
∗∗
p<0.05,
∗∗∗
p<0.01
42
References
Adelman, S., D. O. Gilliag an , and K. Lehrer (2007, December). How effective are food
for education programs? A critical assessm ent of the evidence from developing
countries. International Food Policy Research I n st i tu t e.
Afridi, F. (2007 ). The impact of school meals on schoo l participation: Evidence from
rural India. Syracuse University.
Bertrand, M., E. Duflo, an d S. Mullainath a n (2004) . How much should we trust
differences-in-differences estimates? Quarterly Journal of Economics 119 (1),
249–275.
Bundy, D., C. Burbano, M. Grosh, A. Gelli, M. Jukes, and L. Drake (200 9) . Re-
thinking school fee d i n g. Social safety n et s, child development and the education
sector. Directions in Development. Human Development, World Bank.
Cameron, C. A., J. B. Gelbach, and D. L. Miller (2008) . Bootstrap-based improve-
ments for inference with clustered errors. The Review of Economics and Statis-
tics 90 (3), 414–427.
Chin, A. (2005). Can redistributing teachers across scho o l s raise educa t i ona l attain-
ment? Evidence from Operation Blackboard in India. Journal of Development
Economics 78,384–405.
Deaton, A. and J. Dr`eze (2006). Tracking mid-day meals in Indi a n schools. Draft
notes.
Desai, A. H. and S. Mura l i d h ar (2000). Public interest litigation: Potential an d
problems.
Desai, S. et al. (2008, October). Indian Human Development Survey: Design and
data quality. Ind i a n Huma n Development Survey.
DISE (2008). Analytical Report. DISE Annual Publications. Human Development
Sector, South Asia Region.
Dr`eze, J. and A. Goyal (2003). The future of mid-day meal s. Economic and Political
Weekly November 1,4673–4683.
Duflo, E. (2001). Schooling and labor market consequences of scho ol construction
in Indonesia: Evidence from an unusual policy experiment. American Economic
Review 91 (4), 795–813.
Field, E. (2007). Entitled to wo rk : Urban tenure security and labor supply in Peru.
Quarterly Journal of Economics 122 (4), 1561–1602.
Gauri, V. (2009). Public interest litigation in India. Overreaching or u n d er achieving?
Policy Research Working Paper 5109, World Bank.
References 43
Gruber, J. and D. M. Hungerman (2008). The church versus the mall: What happens
when religion faces increased secular competition? The Quarterly Journal of
Economics 123,831–862.
Hanushek, E. (1995). Interpreting recent research on schooli n g in developing coun-
tries. World Bank Research Observer 10 (2), 227–246.
Harriss, B. (1991 ). Child nutrition and poverty in South I ndia: Noon meals in Tamil
Nadu. New Delhi: Concept Publ i sh i n g .
IHDS (2008). Indian human development survey. Release 01, July.
Imbens, G. and D. Rub i n (1997). Bayesian inference for causal effects in randomized
experiments with noncomplia n ce. The Annals of Statistics 25 (1), 305–327.
Jacoby E., C. S. and P. E. (1996). Benefits o f a school breakfast program among
Andean children in Huaraz, Peru. Food and Nutrition Bulletin 17 (1), 54–64.
Jain, J. and M. Shah (2005, November 26). Antyo d aya Anna Yojana and mi d d ay
meals in MP. Economic and Political Weekly.
Kaushal, S. (2009). Post-enumeration survey of dise data: Major findin gs 2008-09.
Technical report, National University of Educational Planning and Administra-
tion.
Kazianga, H., D. de Walque, and H. Alde rm a n (2009, June). Educational and healt h
impacts of two school feeding schemes: Evidence from a randomized trial in rural
burkina faso. Policy research Work i n g Paper No. 4976.
Khera, R. (2002). Mid-day meals in Rajasthan. The Hindu November 13.
Khera, R. (2006). Mid-day meals in primary schools: Achievements and challenges.
Economic and Political Weekly November 18,4742–4750.
Kingdon, G. G. (2007). The progress of scho ol education in India. Oxford Review of
Economic Policy 23 (2), 163–195.
Kremer, M. an d P. Glewwe (2005). Schools, teachers and education outcomes in
developing countries. Second draft chapter for Handbook on the Economics of
Education.
Kremer, M. and C. Verm eer sch ( 2004). School meals, educational achievement and
school competit i on : Evidence from a randomized evalua t i on . World Bank Policy
Research Working Paper No. 2523.
Krishna Kumar , M. P. and S. Saxena (2001). Looking beyond the smokescreen: Dpep
and primary education in india. Economic and Political Weekly 36 (7), 560–568.
Mehta, A. (2007). S tu d e nt flow at pri m a r y level: An analysis based on dise data.
Technical report, National University of Educational Planning and Administra-
tion.
44 References
Muralidharan, K. (2006, September). Public-private partnership s fo r q u al i ty educa-
tion in India. Seminar 565.
Muralidharan, K. and M. Kremer (2006, March). Public and private schools in rural
India. MIT Mimeo.
NPNSPE (2004). Guidelines of revised National Programme of N utritional Support
to Primary Education. India: Ministry of Human Resource Development, Gov-
ernment of India.
Parikh, K. and S. Yasmeen (2004, January). Groundswell for midday meal scheme.
India Together.
Powell, C., S. P. Walker, S. M. Chang, and S. M. Grantham-McGregor (1998).
Nutrition and education: A randomized trial of the effects of breakfast in rural
primary school children. American Journal of Clinical Nutrition 68,873–879.
PROBE ( 1999) . Public Report on Basic Education. New Delhi: Oxford University
Press.
Psacharopoulos, G. and H. A. Patrinos (September, 2002 ). Returns to investment in
education: A further update. World Bank Pol i cy Research Working Paper 2881.
Secretariat of the Right to Food Campaig n (2005, November). Midday Mea l s: A
primer. Right to Food Campaign Materials.
Sharma, S., S. J. Passi, L. Thomas, and H. S. Gopala n (2006). Evaluation of midday
meal program in MCD schools. Nutrition Foundation of India, Scientific Report
18.
SSA (2008). Framework of Sarva Shiksha Abhiyan. India: Ministry of Human Re-
source Development, Government of India.
UNESCO (2006). Educatio for All Global Monitoring Report 2007: Strong Founda-
tions. Paris: UNESCO Publishin g .
UNICEF (2008). State of the World’s Children. New York: UNICEF.
Ward, M. (2007). Rural education. In India Infrastructure Report. Oxford University
Press.
World Bank (2003). A Review of Eucational Progress and Reform in the District
Primary Education Program (Phases I and II). World Bank. Human Development
Sector, South Asia Region.
Zaidi, A. (2005, Feb. -M ar . ) . Food, for education. Frontline 22 (5).
References 45
Appendix A. Midday Meal Imple m e ntation in Public Primary Schools
State Implementation Date Midday Meal Content
Andhra Pradesh January 2003 Rice, sa mbhar, egg/banana twice a week
Assam January 2005 Rice, dal, vegetables
Bihar September 2004(Pilot)
January 2005
Rice with sabji, dal, pulao, karhi or khichri
Chhattisgarh April 2002 Rice with dal or vegetables
Gujarat November 1 9 8 4 Wheat, rice, pulses, oil, spices
Haryana August 2004 Mitha rice, vegetbale pulao, dalia, paushtic khichri or bakli by rotation
Himachal Pradesh September 2004 Grains, seasonal vegetables, fruit, eggs
Karnataka July 2002(Pilot)
June 2003
Rice, pulses, oil, salt, vegetables
Madhya Pradesh July 20 0 3 Dal-roti/d a l -s abji (in wheat predomi n a nt areas) or dal-rice/dal-rice-
sabji (in rice predominant areas)
Maharashtra January 2003 Rice, d a l, vegetables, spices, oil, banana/egg at least once a week
Orissa June 2001(Pilot)
September 2004
Rice, dal, egg/soya twice a week
Rajast h a n July 2002 Ghooghari (mixture o f gur/jaggery and boiled wheat), dalia
Tamil Nadu July 1982 Rice, eggs, boiled potatoes, cooked black bengal, vegetables with vari-
ation
Uttar Pradesh September 2004 Food grains, pulses, oil, salt, spices
46 References
Uttaranchal November 2 0 0 2 - July 2003 Rice, dal, kheer, fruits and eggs alternately
The information provided in this table was drawn from state g overnment documents listed in a ,andthen
verified and cross-checked using more than one independent source listed in b-e). Sources of information are:
a. state government documents: The National Programme of Midday M ea l in Schools, Annual Work Plan and
Budget, 2009-10’; b. planning comm iss io n : Program Evaluation Organ iz a t io n (2010): ‘Performance Evaluation
of Cooked Mid-Day Meal’, Plann in g Commission; independent monitors: the 6 reports of the Commissioner of
India on the Writ Petition 196 of 2001 (PUCL vs. Union of India and Others); c. independent auditors: Civil
Performance Audit Reports from 2007 and 2008 of the Comptroller and Au d it o r General of India (for Andhra
Pradesh, Assam, Bihar, Chhattisgarh, Gujarat, Haryana, Kerala, Madhya Pradesh, Orissa, Uttar Pradesh,
Uttaranchal); National University of Educational Planning and Administration, New Delhi, Study of best
practices in: Andhra Pradesh by Y.Josephine, Assam by VPRS. Raju, Haryana by M. Narul a , Karnataka
by K. Srinivas, Maharashtra by S. Chugh, Orissa by S.K. Malik, Rajasthan by S. Kaushal, Uttar Pradesh
by K. Wizarat ; d. field surveys: Kumar P. and Sood T. (20 0 5 ): ‘Bihar: Mid-day Meal Survey Report’.
Right to food campaign , Afridi F. (2005): ‘Mid-day Meals: A Comparison of the Financial and Institut i on a l
Organization of the Program in Two States (Madhya Pradesh and Karnataka). Economic and Political
Weekly, Robinson F. (2007) ‘The Mid-Day Meal Scheme In Four Districts of Madhya Pradesh’. Jawaharlal
Nehru University The Hunger Project, CUTS Center for Co n su me r Action, Research & Train in g (CART) and
World Bank (2007): ‘An assessment of the Mid-Day Meal Scheme in Chittorgarh District (Rajasthan)’; e.
selected news articles and reports : Chettiparambil-Rajan A. (2007): ‘India: A desk review of the Midday Meal
Programme’ World Food Programme, Khera R. (2006): ‘Mid-Day Meals in Primary Schools: Achievements
and Ch a l len g es ’ Economic and Political Weekly, Parikh K. and Yasmeen S. (2004): ‘Groundswell for mid-day
meal scheme’ India Together, Dreze J. and Goyal A. (2003): ‘The Future of Mid-day Meals’ Economic and
Political Weekly, R. Anuradha (2003): ‘Nutriti o n Schemes in Tamil Nadu’ UNDP, Khera R. (2002): ‘Mid-day
Meals in Rajasthan’ The Hindu.
Recent ESMT Working Paper
ESMT No.
The impact of school lunches on primary school enrollment:
Evidence from India’s midday meal scheme
Rajshri Jayaraman, ESMT
Dora Simroth, ESMT
11-11
Dynamic coordination via organizational routines
Andreas Blume, University of Pittsburgh
April M. Franco, Rotman School of Management, University of
Toronto
Paul Heidhues, ESMT
11-10
Naïve and capricious: Stumbling into the ring of self-control
conflict
Kristian Ove R. Myrseth, ESMT
Conny Wollbrant, University of Gothenburg
11-09
Decomposing first mover advantages in the mobile
telecommunications industry
J.P. Eggers, NYU Stern School of Business
Michał Grajek, ESMT
Tobias Kretschmer, LMU Munich
11-03 (R1)
Vertical coordination through renegotiation
Özlem Bedre-Defolie, ESMT
11-08
Merger efficiency and welfare implications of buyer power
Özlem Bedre-Defolie, ESMT
Stéphane Caprice, Toulouse School of Economics
11-07
Patent examination at the State Intellectual Property Office in
China?
Johannes Liegsalz, BMW AG
Stefan Wagner, ESMT
11-06
When and how does corporate social responsibility encourage
customer orientation?
Daniel Korschun, LeBow College of Business, Drexel University
C.B. Bhattacharya, ESMT
Scott D. Swain, Northeastern University
11-05
ESMT
European School of Management and Technology
Faculty Publications
Schlossplatz 1
10178 Berlin
Germany
Phone: +49 (0) 30 21231-1279
www.esmt.org
ESMT. The business school founded by business.
